* [PATCH ipsec-next v8 01/16] xfrm: config: add CONFIG_XFRM_IPTFS
2024-08-04 20:33 [PATCH ipsec-next v8 00/16] Add IP-TFS mode to xfrm Christian Hopps
@ 2024-08-04 20:33 ` Christian Hopps
2024-08-04 20:33 ` [PATCH ipsec-next v8 02/16] include: uapi: add ip_tfs_*_hdr packet formats Christian Hopps
` (14 subsequent siblings)
15 siblings, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-04 20:33 UTC (permalink / raw)
To: devel; +Cc: Steffen Klassert, netdev, Christian Hopps, Christian Hopps
From: Christian Hopps <chopps@labn.net>
Add new Kconfig option to enable IP-TFS (RFC9347) functionality.
Signed-off-by: Christian Hopps <chopps@labn.net>
---
net/xfrm/Kconfig | 16 ++++++++++++++++
1 file changed, 16 insertions(+)
diff --git a/net/xfrm/Kconfig b/net/xfrm/Kconfig
index d7b16f2c23e9..f0157702718f 100644
--- a/net/xfrm/Kconfig
+++ b/net/xfrm/Kconfig
@@ -135,6 +135,22 @@ config NET_KEY_MIGRATE
If unsure, say N.
+config XFRM_IPTFS
+ tristate "IPsec IP-TFS/AGGFRAG (RFC 9347) encapsulation support"
+ depends on XFRM
+ help
+ Information on the IP-TFS/AGGFRAG encapsulation can be found
+ in RFC 9347. This feature supports demand driven (i.e.,
+ non-constant send rate) IP-TFS to take advantage of the
+ AGGFRAG ESP payload encapsulation. This payload type
+ supports aggregation and fragmentation of the inner IP
+ packet stream which in turn yields higher small-packet
+ bandwidth as well as reducing MTU/PMTU issues. Congestion
+ control is unimplementated as the send rate is demand driven
+ rather than constant.
+
+ If unsure, say N.
+
config XFRM_ESPINTCP
bool
--
2.46.0
^ permalink raw reply related [flat|nested] 40+ messages in thread* [PATCH ipsec-next v8 02/16] include: uapi: add ip_tfs_*_hdr packet formats
2024-08-04 20:33 [PATCH ipsec-next v8 00/16] Add IP-TFS mode to xfrm Christian Hopps
2024-08-04 20:33 ` [PATCH ipsec-next v8 01/16] xfrm: config: add CONFIG_XFRM_IPTFS Christian Hopps
@ 2024-08-04 20:33 ` Christian Hopps
2024-08-04 20:33 ` [PATCH ipsec-next v8 03/16] include: uapi: add IPPROTO_AGGFRAG for AGGFRAG in ESP Christian Hopps
` (13 subsequent siblings)
15 siblings, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-04 20:33 UTC (permalink / raw)
To: devel; +Cc: Steffen Klassert, netdev, Christian Hopps, Christian Hopps
From: Christian Hopps <chopps@labn.net>
Add the on-wire basic and congestion-control IP-TFS packet headers.
Signed-off-by: Christian Hopps <chopps@labn.net>
---
include/uapi/linux/ip.h | 16 ++++++++++++++++
1 file changed, 16 insertions(+)
diff --git a/include/uapi/linux/ip.h b/include/uapi/linux/ip.h
index 283dec7e3645..5bd7ce934d74 100644
--- a/include/uapi/linux/ip.h
+++ b/include/uapi/linux/ip.h
@@ -137,6 +137,22 @@ struct ip_beet_phdr {
__u8 reserved;
};
+struct ip_iptfs_hdr {
+ __u8 subtype; /* 0*: basic, 1: CC */
+ __u8 flags;
+ __be16 block_offset;
+};
+
+struct ip_iptfs_cc_hdr {
+ __u8 subtype; /* 0: basic, 1*: CC */
+ __u8 flags;
+ __be16 block_offset;
+ __be32 loss_rate;
+ __be64 rtt_adelay_xdelay;
+ __be32 tval;
+ __be32 techo;
+};
+
/* index values for the variables in ipv4_devconf */
enum
{
--
2.46.0
^ permalink raw reply related [flat|nested] 40+ messages in thread* [PATCH ipsec-next v8 03/16] include: uapi: add IPPROTO_AGGFRAG for AGGFRAG in ESP
2024-08-04 20:33 [PATCH ipsec-next v8 00/16] Add IP-TFS mode to xfrm Christian Hopps
2024-08-04 20:33 ` [PATCH ipsec-next v8 01/16] xfrm: config: add CONFIG_XFRM_IPTFS Christian Hopps
2024-08-04 20:33 ` [PATCH ipsec-next v8 02/16] include: uapi: add ip_tfs_*_hdr packet formats Christian Hopps
@ 2024-08-04 20:33 ` Christian Hopps
2024-08-04 20:33 ` [PATCH ipsec-next v8 04/16] xfrm: netlink: add config (netlink) options Christian Hopps
` (12 subsequent siblings)
15 siblings, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-04 20:33 UTC (permalink / raw)
To: devel; +Cc: Steffen Klassert, netdev, Christian Hopps, Christian Hopps
From: Christian Hopps <chopps@labn.net>
Add the RFC assigned IP protocol number for AGGFRAG.
Signed-off-by: Christian Hopps <chopps@labn.net>
---
include/uapi/linux/in.h | 2 ++
1 file changed, 2 insertions(+)
diff --git a/include/uapi/linux/in.h b/include/uapi/linux/in.h
index d358add1611c..268086e85d04 100644
--- a/include/uapi/linux/in.h
+++ b/include/uapi/linux/in.h
@@ -79,6 +79,8 @@ enum {
#define IPPROTO_MPLS IPPROTO_MPLS
IPPROTO_ETHERNET = 143, /* Ethernet-within-IPv6 Encapsulation */
#define IPPROTO_ETHERNET IPPROTO_ETHERNET
+ IPPROTO_AGGFRAG = 144, /* AGGFRAG in ESP (RFC 9347) */
+#define IPPROTO_AGGFRAG IPPROTO_AGGFRAG
IPPROTO_RAW = 255, /* Raw IP packets */
#define IPPROTO_RAW IPPROTO_RAW
IPPROTO_SMC = 256, /* Shared Memory Communications */
--
2.46.0
^ permalink raw reply related [flat|nested] 40+ messages in thread* [PATCH ipsec-next v8 04/16] xfrm: netlink: add config (netlink) options
2024-08-04 20:33 [PATCH ipsec-next v8 00/16] Add IP-TFS mode to xfrm Christian Hopps
` (2 preceding siblings ...)
2024-08-04 20:33 ` [PATCH ipsec-next v8 03/16] include: uapi: add IPPROTO_AGGFRAG for AGGFRAG in ESP Christian Hopps
@ 2024-08-04 20:33 ` Christian Hopps
2024-08-04 20:33 ` [PATCH ipsec-next v8 05/16] xfrm: add mode_cbs module functionality Christian Hopps
` (11 subsequent siblings)
15 siblings, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-04 20:33 UTC (permalink / raw)
To: devel; +Cc: Steffen Klassert, netdev, Christian Hopps, Christian Hopps
From: Christian Hopps <chopps@labn.net>
Add netlink options for configuring IP-TFS SAs.
Signed-off-by: Christian Hopps <chopps@labn.net>
---
include/uapi/linux/xfrm.h | 9 ++++++-
net/xfrm/xfrm_compat.c | 10 ++++++--
net/xfrm/xfrm_user.c | 52 +++++++++++++++++++++++++++++++++++++++
3 files changed, 68 insertions(+), 3 deletions(-)
diff --git a/include/uapi/linux/xfrm.h b/include/uapi/linux/xfrm.h
index f28701500714..042ebf94bb3d 100644
--- a/include/uapi/linux/xfrm.h
+++ b/include/uapi/linux/xfrm.h
@@ -158,7 +158,8 @@ enum {
#define XFRM_MODE_ROUTEOPTIMIZATION 2
#define XFRM_MODE_IN_TRIGGER 3
#define XFRM_MODE_BEET 4
-#define XFRM_MODE_MAX 5
+#define XFRM_MODE_IPTFS 5
+#define XFRM_MODE_MAX 6
/* Netlink configuration messages. */
enum {
@@ -322,6 +323,12 @@ enum xfrm_attr_type_t {
XFRMA_MTIMER_THRESH, /* __u32 in seconds for input SA */
XFRMA_SA_DIR, /* __u8 */
XFRMA_NAT_KEEPALIVE_INTERVAL, /* __u32 in seconds for NAT keepalive */
+ XFRMA_IPTFS_DROP_TIME, /* __u32 in: usec to wait for next seq */
+ XFRMA_IPTFS_REORDER_WINDOW, /* __u16 in: reorder window size (pkts) */
+ XFRMA_IPTFS_DONT_FRAG, /* out: don't use fragmentation */
+ XFRMA_IPTFS_INIT_DELAY, /* __u32 out: initial packet wait delay (usec) */
+ XFRMA_IPTFS_MAX_QSIZE, /* __u32 out: max ingress queue size (octets) */
+ XFRMA_IPTFS_PKT_SIZE, /* __u32 out: size of outer packet, 0 for PMTU */
__XFRMA_MAX
#define XFRMA_OUTPUT_MARK XFRMA_SET_MARK /* Compatibility */
diff --git a/net/xfrm/xfrm_compat.c b/net/xfrm/xfrm_compat.c
index 91357ccaf4af..5c55e07f3d10 100644
--- a/net/xfrm/xfrm_compat.c
+++ b/net/xfrm/xfrm_compat.c
@@ -282,9 +282,15 @@ static int xfrm_xlate64_attr(struct sk_buff *dst, const struct nlattr *src)
case XFRMA_MTIMER_THRESH:
case XFRMA_SA_DIR:
case XFRMA_NAT_KEEPALIVE_INTERVAL:
+ case XFRMA_IPTFS_DROP_TIME:
+ case XFRMA_IPTFS_REORDER_WINDOW:
+ case XFRMA_IPTFS_DONT_FRAG:
+ case XFRMA_IPTFS_INIT_DELAY:
+ case XFRMA_IPTFS_MAX_QSIZE:
+ case XFRMA_IPTFS_PKT_SIZE:
return xfrm_nla_cpy(dst, src, nla_len(src));
default:
- BUILD_BUG_ON(XFRMA_MAX != XFRMA_NAT_KEEPALIVE_INTERVAL);
+ BUILD_BUG_ON(XFRMA_MAX != XFRMA_IPTFS_PKT_SIZE);
pr_warn_once("unsupported nla_type %d\n", src->nla_type);
return -EOPNOTSUPP;
}
@@ -439,7 +445,7 @@ static int xfrm_xlate32_attr(void *dst, const struct nlattr *nla,
int err;
if (type > XFRMA_MAX) {
- BUILD_BUG_ON(XFRMA_MAX != XFRMA_NAT_KEEPALIVE_INTERVAL);
+ BUILD_BUG_ON(XFRMA_MAX != XFRMA_IPTFS_PKT_SIZE);
NL_SET_ERR_MSG(extack, "Bad attribute");
return -EOPNOTSUPP;
}
diff --git a/net/xfrm/xfrm_user.c b/net/xfrm/xfrm_user.c
index 55f039ec3d59..f6ed019192f3 100644
--- a/net/xfrm/xfrm_user.c
+++ b/net/xfrm/xfrm_user.c
@@ -297,6 +297,16 @@ static int verify_newsa_info(struct xfrm_usersa_info *p,
NL_SET_ERR_MSG(extack, "TFC padding can only be used in tunnel mode");
goto out;
}
+ if ((attrs[XFRMA_IPTFS_DROP_TIME] ||
+ attrs[XFRMA_IPTFS_REORDER_WINDOW] ||
+ attrs[XFRMA_IPTFS_DONT_FRAG] ||
+ attrs[XFRMA_IPTFS_INIT_DELAY] ||
+ attrs[XFRMA_IPTFS_MAX_QSIZE] ||
+ attrs[XFRMA_IPTFS_PKT_SIZE]) &&
+ p->mode != XFRM_MODE_IPTFS) {
+ NL_SET_ERR_MSG(extack, "IP-TFS options can only be used in IP-TFS mode");
+ goto out;
+ }
break;
case IPPROTO_COMP:
@@ -417,6 +427,18 @@ static int verify_newsa_info(struct xfrm_usersa_info *p,
goto out;
}
+ if (attrs[XFRMA_IPTFS_DROP_TIME]) {
+ NL_SET_ERR_MSG(extack, "IP-TFS drop time should not be set for output SA");
+ err = -EINVAL;
+ goto out;
+ }
+
+ if (attrs[XFRMA_IPTFS_REORDER_WINDOW]) {
+ NL_SET_ERR_MSG(extack, "IP-TFS reorder window should not be set for output SA");
+ err = -EINVAL;
+ goto out;
+ }
+
if (attrs[XFRMA_REPLAY_VAL]) {
struct xfrm_replay_state *replay;
@@ -454,6 +476,30 @@ static int verify_newsa_info(struct xfrm_usersa_info *p,
}
}
+
+ if (attrs[XFRMA_IPTFS_DONT_FRAG]) {
+ NL_SET_ERR_MSG(extack, "IP-TFS don't fragment should not be set for input SA");
+ err = -EINVAL;
+ goto out;
+ }
+
+ if (attrs[XFRMA_IPTFS_INIT_DELAY]) {
+ NL_SET_ERR_MSG(extack, "IP-TFS initial delay should not be set for input SA");
+ err = -EINVAL;
+ goto out;
+ }
+
+ if (attrs[XFRMA_IPTFS_MAX_QSIZE]) {
+ NL_SET_ERR_MSG(extack, "IP-TFS max queue size should not be set for input SA");
+ err = -EINVAL;
+ goto out;
+ }
+
+ if (attrs[XFRMA_IPTFS_PKT_SIZE]) {
+ NL_SET_ERR_MSG(extack, "IP-TFS packet size should not be set for input SA");
+ err = -EINVAL;
+ goto out;
+ }
}
out:
@@ -3176,6 +3222,12 @@ const struct nla_policy xfrma_policy[XFRMA_MAX+1] = {
[XFRMA_MTIMER_THRESH] = { .type = NLA_U32 },
[XFRMA_SA_DIR] = NLA_POLICY_RANGE(NLA_U8, XFRM_SA_DIR_IN, XFRM_SA_DIR_OUT),
[XFRMA_NAT_KEEPALIVE_INTERVAL] = { .type = NLA_U32 },
+ [XFRMA_IPTFS_DROP_TIME] = { .type = NLA_U32 },
+ [XFRMA_IPTFS_REORDER_WINDOW] = { .type = NLA_U16 },
+ [XFRMA_IPTFS_DONT_FRAG] = { .type = NLA_FLAG },
+ [XFRMA_IPTFS_INIT_DELAY] = { .type = NLA_U32 },
+ [XFRMA_IPTFS_MAX_QSIZE] = { .type = NLA_U32 },
+ [XFRMA_IPTFS_PKT_SIZE] = { .type = NLA_U32 },
};
EXPORT_SYMBOL_GPL(xfrma_policy);
--
2.46.0
^ permalink raw reply related [flat|nested] 40+ messages in thread* [PATCH ipsec-next v8 05/16] xfrm: add mode_cbs module functionality
2024-08-04 20:33 [PATCH ipsec-next v8 00/16] Add IP-TFS mode to xfrm Christian Hopps
` (3 preceding siblings ...)
2024-08-04 20:33 ` [PATCH ipsec-next v8 04/16] xfrm: netlink: add config (netlink) options Christian Hopps
@ 2024-08-04 20:33 ` Christian Hopps
2024-08-04 20:33 ` [PATCH ipsec-next v8 06/16] xfrm: add generic iptfs defines and functionality Christian Hopps
` (10 subsequent siblings)
15 siblings, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-04 20:33 UTC (permalink / raw)
To: devel; +Cc: Steffen Klassert, netdev, Christian Hopps, Christian Hopps
From: Christian Hopps <chopps@labn.net>
Add a set of callbacks xfrm_mode_cbs to xfrm_state. These callbacks
enable the addition of new xfrm modes, such as IP-TFS to be defined
in modules.
Signed-off-by: Christian Hopps <chopps@labn.net>
---
include/net/xfrm.h | 43 +++++++++++++++++++++++++
net/xfrm/xfrm_device.c | 3 +-
net/xfrm/xfrm_input.c | 18 +++++++++--
net/xfrm/xfrm_output.c | 2 ++
net/xfrm/xfrm_policy.c | 18 +++++++----
net/xfrm/xfrm_state.c | 72 ++++++++++++++++++++++++++++++++++++++++++
net/xfrm/xfrm_user.c | 13 ++++++++
7 files changed, 159 insertions(+), 10 deletions(-)
diff --git a/include/net/xfrm.h b/include/net/xfrm.h
index 54cef89f6c1e..323e768ce4f3 100644
--- a/include/net/xfrm.h
+++ b/include/net/xfrm.h
@@ -209,6 +209,7 @@ struct xfrm_state {
u16 family;
xfrm_address_t saddr;
int header_len;
+ int enc_hdr_len;
int trailer_len;
u32 extra_flags;
struct xfrm_mark smark;
@@ -299,6 +300,9 @@ struct xfrm_state {
* interpreted by xfrm_type methods. */
void *data;
u8 dir;
+
+ const struct xfrm_mode_cbs *mode_cbs;
+ void *mode_data;
};
static inline struct net *xs_net(struct xfrm_state *x)
@@ -451,6 +455,45 @@ struct xfrm_type_offload {
int xfrm_register_type_offload(const struct xfrm_type_offload *type, unsigned short family);
void xfrm_unregister_type_offload(const struct xfrm_type_offload *type, unsigned short family);
+/**
+ * struct xfrm_mode_cbs - XFRM mode callbacks
+ * @owner: module owner or NULL
+ * @create_state: Add mode specific state to new `xfrm_state *x`
+ * @delete_state: Cleanup mode specific state from `xfrm_state *x`
+ * @user_init: Process mode specific netlink attributes from user
+ * @copy_to_user: Add netlink attributes to `attrs` based on state in `x`
+ * @clone: Copy mode specific values from `orig` to new state `x`
+ * @sa_len: Return space required to store mode specific netlink attributes
+ * @get_inner_mtu: Return avail payload space after removing encap overhead
+ * @input: Process received packet from SA using mode
+ * @output: Output given packet using mode
+ * @prepare_output: Add mode specific encapsulation to packet in skb. On return
+ * `transport_header` should point at ESP header, `network_header` should
+ * point at outer IP header and `mac_header` should opint at the
+ * protocol/nexthdr field of the outer IP.
+ *
+ * One should examine and understand the specific uses of these callbacks in
+ * xfrm for further detail on how and when these functions are called. RTSL.
+ */
+struct xfrm_mode_cbs {
+ struct module *owner;
+ int (*create_state)(struct xfrm_state *x);
+ void (*delete_state)(struct xfrm_state *x);
+ int (*user_init)(struct net *net, struct xfrm_state *x,
+ struct nlattr **attrs,
+ struct netlink_ext_ack *extack);
+ int (*copy_to_user)(struct xfrm_state *x, struct sk_buff *skb);
+ int (*clone)(struct xfrm_state *x, struct xfrm_state *orig);
+ unsigned int (*sa_len)(const struct xfrm_state *x);
+ u32 (*get_inner_mtu)(struct xfrm_state *x, int outer_mtu);
+ int (*input)(struct xfrm_state *x, struct sk_buff *skb);
+ int (*output)(struct net *net, struct sock *sk, struct sk_buff *skb);
+ int (*prepare_output)(struct xfrm_state *x, struct sk_buff *skb);
+};
+
+int xfrm_register_mode_cbs(u8 mode, const struct xfrm_mode_cbs *mode_cbs);
+void xfrm_unregister_mode_cbs(u8 mode);
+
static inline int xfrm_af2proto(unsigned int family)
{
switch(family) {
diff --git a/net/xfrm/xfrm_device.c b/net/xfrm/xfrm_device.c
index 9a44d363ba62..e412e4afb169 100644
--- a/net/xfrm/xfrm_device.c
+++ b/net/xfrm/xfrm_device.c
@@ -42,7 +42,8 @@ static void __xfrm_mode_tunnel_prep(struct xfrm_state *x, struct sk_buff *skb,
skb->transport_header = skb->network_header + hsize;
skb_reset_mac_len(skb);
- pskb_pull(skb, skb->mac_len + x->props.header_len);
+ pskb_pull(skb,
+ skb->mac_len + x->props.header_len - x->props.enc_hdr_len);
}
static void __xfrm_mode_beet_prep(struct xfrm_state *x, struct sk_buff *skb,
diff --git a/net/xfrm/xfrm_input.c b/net/xfrm/xfrm_input.c
index 749e7eea99e4..b7b5cda986fb 100644
--- a/net/xfrm/xfrm_input.c
+++ b/net/xfrm/xfrm_input.c
@@ -446,6 +446,9 @@ static int xfrm_inner_mode_input(struct xfrm_state *x,
WARN_ON_ONCE(1);
break;
default:
+ if (x->mode_cbs && x->mode_cbs->input)
+ return x->mode_cbs->input(x, skb);
+
WARN_ON_ONCE(1);
break;
}
@@ -453,6 +456,10 @@ static int xfrm_inner_mode_input(struct xfrm_state *x,
return -EOPNOTSUPP;
}
+/* NOTE: encap_type - In addition to the normal (non-negative) values for
+ * encap_type, a negative value of -1 or -2 can be used to resume/restart this
+ * function after a previous invocation early terminated for async operation.
+ */
int xfrm_input(struct sk_buff *skb, int nexthdr, __be32 spi, int encap_type)
{
const struct xfrm_state_afinfo *afinfo;
@@ -489,6 +496,10 @@ int xfrm_input(struct sk_buff *skb, int nexthdr, __be32 spi, int encap_type)
family = x->props.family;
+ /* An encap_type of -2 indicates reconstructed inner packet */
+ if (encap_type == -2)
+ goto resume_decapped;
+
/* An encap_type of -1 indicates async resumption. */
if (encap_type == -1) {
async = 1;
@@ -679,11 +690,14 @@ int xfrm_input(struct sk_buff *skb, int nexthdr, __be32 spi, int encap_type)
XFRM_MODE_SKB_CB(skb)->protocol = nexthdr;
- if (xfrm_inner_mode_input(x, skb)) {
+ err = xfrm_inner_mode_input(x, skb);
+ if (err == -EINPROGRESS)
+ return 0;
+ else if (err) {
XFRM_INC_STATS(net, LINUX_MIB_XFRMINSTATEMODEERROR);
goto drop;
}
-
+resume_decapped:
if (x->outer_mode.flags & XFRM_MODE_FLAG_TUNNEL) {
decaps = 1;
break;
diff --git a/net/xfrm/xfrm_output.c b/net/xfrm/xfrm_output.c
index e5722c95b8bb..ef81359e4038 100644
--- a/net/xfrm/xfrm_output.c
+++ b/net/xfrm/xfrm_output.c
@@ -472,6 +472,8 @@ static int xfrm_outer_mode_output(struct xfrm_state *x, struct sk_buff *skb)
WARN_ON_ONCE(1);
break;
default:
+ if (x->mode_cbs && x->mode_cbs->prepare_output)
+ return x->mode_cbs->prepare_output(x, skb);
WARN_ON_ONCE(1);
break;
}
diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c
index c56c61b0c12e..f764b1409175 100644
--- a/net/xfrm/xfrm_policy.c
+++ b/net/xfrm/xfrm_policy.c
@@ -2719,13 +2719,17 @@ static struct dst_entry *xfrm_bundle_create(struct xfrm_policy *policy,
dst1->input = dst_discard;
- rcu_read_lock();
- afinfo = xfrm_state_afinfo_get_rcu(inner_mode->family);
- if (likely(afinfo))
- dst1->output = afinfo->output;
- else
- dst1->output = dst_discard_out;
- rcu_read_unlock();
+ if (xfrm[i]->mode_cbs && xfrm[i]->mode_cbs->output) {
+ dst1->output = xfrm[i]->mode_cbs->output;
+ } else {
+ rcu_read_lock();
+ afinfo = xfrm_state_afinfo_get_rcu(inner_mode->family);
+ if (likely(afinfo))
+ dst1->output = afinfo->output;
+ else
+ dst1->output = dst_discard_out;
+ rcu_read_unlock();
+ }
xdst_prev = xdst;
diff --git a/net/xfrm/xfrm_state.c b/net/xfrm/xfrm_state.c
index 37478d36a8df..e7b656bb6c0d 100644
--- a/net/xfrm/xfrm_state.c
+++ b/net/xfrm/xfrm_state.c
@@ -515,6 +515,60 @@ static const struct xfrm_mode *xfrm_get_mode(unsigned int encap, int family)
return NULL;
}
+static const struct xfrm_mode_cbs __rcu *xfrm_mode_cbs_map[XFRM_MODE_MAX];
+static DEFINE_SPINLOCK(xfrm_mode_cbs_map_lock);
+
+int xfrm_register_mode_cbs(u8 mode, const struct xfrm_mode_cbs *mode_cbs)
+{
+ if (mode >= XFRM_MODE_MAX)
+ return -EINVAL;
+
+ spin_lock_bh(&xfrm_mode_cbs_map_lock);
+ rcu_assign_pointer(xfrm_mode_cbs_map[mode], mode_cbs);
+ spin_unlock_bh(&xfrm_mode_cbs_map_lock);
+
+ return 0;
+}
+EXPORT_SYMBOL(xfrm_register_mode_cbs);
+
+void xfrm_unregister_mode_cbs(u8 mode)
+{
+ if (mode >= XFRM_MODE_MAX)
+ return;
+
+ spin_lock_bh(&xfrm_mode_cbs_map_lock);
+ RCU_INIT_POINTER(xfrm_mode_cbs_map[mode], NULL);
+ spin_unlock_bh(&xfrm_mode_cbs_map_lock);
+ synchronize_rcu();
+}
+EXPORT_SYMBOL(xfrm_unregister_mode_cbs);
+
+static const struct xfrm_mode_cbs *xfrm_get_mode_cbs(u8 mode)
+{
+ const struct xfrm_mode_cbs *cbs;
+ bool try_load = true;
+
+ if (mode >= XFRM_MODE_MAX)
+ return NULL;
+
+retry:
+ rcu_read_lock();
+
+ cbs = rcu_dereference(xfrm_mode_cbs_map[mode]);
+ if (cbs && !try_module_get(cbs->owner))
+ cbs = NULL;
+
+ rcu_read_unlock();
+
+ if (mode == XFRM_MODE_IPTFS && !cbs && try_load) {
+ request_module("xfrm-iptfs");
+ try_load = false;
+ goto retry;
+ }
+
+ return cbs;
+}
+
void xfrm_state_free(struct xfrm_state *x)
{
kmem_cache_free(xfrm_state_cache, x);
@@ -523,6 +577,8 @@ EXPORT_SYMBOL(xfrm_state_free);
static void ___xfrm_state_destroy(struct xfrm_state *x)
{
+ if (x->mode_cbs && x->mode_cbs->delete_state)
+ x->mode_cbs->delete_state(x);
hrtimer_cancel(&x->mtimer);
del_timer_sync(&x->rtimer);
kfree(x->aead);
@@ -680,6 +736,7 @@ struct xfrm_state *xfrm_state_alloc(struct net *net)
x->replay_maxage = 0;
x->replay_maxdiff = 0;
spin_lock_init(&x->lock);
+ x->mode_data = NULL;
}
return x;
}
@@ -1806,6 +1863,12 @@ static struct xfrm_state *xfrm_state_clone(struct xfrm_state *orig,
x->new_mapping_sport = 0;
x->dir = orig->dir;
+ x->mode_cbs = orig->mode_cbs;
+ if (x->mode_cbs && x->mode_cbs->clone) {
+ if (x->mode_cbs->clone(x, orig))
+ goto error;
+ }
+
return x;
error:
@@ -2845,6 +2908,9 @@ u32 xfrm_state_mtu(struct xfrm_state *x, int mtu)
case XFRM_MODE_TUNNEL:
break;
default:
+ if (x->mode_cbs && x->mode_cbs->get_inner_mtu)
+ return x->mode_cbs->get_inner_mtu(x, mtu);
+
WARN_ON_ONCE(1);
break;
}
@@ -2945,6 +3011,12 @@ int __xfrm_init_state(struct xfrm_state *x, bool init_replay, bool offload,
}
}
+ x->mode_cbs = xfrm_get_mode_cbs(x->props.mode);
+ if (x->mode_cbs) {
+ if (x->mode_cbs->create_state)
+ err = x->mode_cbs->create_state(x);
+ module_put(x->mode_cbs->owner);
+ }
error:
return err;
}
diff --git a/net/xfrm/xfrm_user.c b/net/xfrm/xfrm_user.c
index f6ed019192f3..419bbeea6b20 100644
--- a/net/xfrm/xfrm_user.c
+++ b/net/xfrm/xfrm_user.c
@@ -918,6 +918,12 @@ static struct xfrm_state *xfrm_state_construct(struct net *net,
goto error;
}
+ if (x->mode_cbs && x->mode_cbs->user_init) {
+ err = x->mode_cbs->user_init(net, x, attrs, extack);
+ if (err)
+ goto error;
+ }
+
return x;
error:
@@ -1331,6 +1337,10 @@ static int copy_to_user_state_extra(struct xfrm_state *x,
if (ret)
goto out;
}
+ if (x->mode_cbs && x->mode_cbs->copy_to_user)
+ ret = x->mode_cbs->copy_to_user(x, skb);
+ if (ret)
+ goto out;
if (x->mapping_maxage) {
ret = nla_put_u32(skb, XFRMA_MTIMER_THRESH, x->mapping_maxage);
if (ret)
@@ -3540,6 +3550,9 @@ static inline unsigned int xfrm_sa_len(struct xfrm_state *x)
if (x->nat_keepalive_interval)
l += nla_total_size(sizeof(x->nat_keepalive_interval));
+ if (x->mode_cbs && x->mode_cbs->sa_len)
+ l += x->mode_cbs->sa_len(x);
+
return l;
}
--
2.46.0
^ permalink raw reply related [flat|nested] 40+ messages in thread* [PATCH ipsec-next v8 06/16] xfrm: add generic iptfs defines and functionality
2024-08-04 20:33 [PATCH ipsec-next v8 00/16] Add IP-TFS mode to xfrm Christian Hopps
` (4 preceding siblings ...)
2024-08-04 20:33 ` [PATCH ipsec-next v8 05/16] xfrm: add mode_cbs module functionality Christian Hopps
@ 2024-08-04 20:33 ` Christian Hopps
2024-08-04 20:33 ` [PATCH ipsec-next v8 07/16] xfrm: iptfs: add new iptfs xfrm mode impl Christian Hopps
` (9 subsequent siblings)
15 siblings, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-04 20:33 UTC (permalink / raw)
To: devel; +Cc: Steffen Klassert, netdev, Christian Hopps, Christian Hopps
From: Christian Hopps <chopps@labn.net>
Define `XFRM_MODE_IPTFS` and `IPSEC_MODE_IPTFS` constants, and add these to
switch case and conditionals adjacent with the existing TUNNEL modes.
Signed-off-by: Christian Hopps <chopps@labn.net>
---
include/net/xfrm.h | 1 +
include/uapi/linux/ipsec.h | 3 ++-
include/uapi/linux/snmp.h | 3 +++
net/ipv4/esp4.c | 3 ++-
net/ipv6/esp6.c | 3 ++-
net/netfilter/nft_xfrm.c | 3 ++-
net/xfrm/xfrm_device.c | 1 +
net/xfrm/xfrm_output.c | 4 ++++
net/xfrm/xfrm_policy.c | 8 ++++++--
net/xfrm/xfrm_proc.c | 3 +++
net/xfrm/xfrm_state.c | 12 ++++++++++++
net/xfrm/xfrm_user.c | 12 ++++++++++++
12 files changed, 50 insertions(+), 6 deletions(-)
diff --git a/include/net/xfrm.h b/include/net/xfrm.h
index 323e768ce4f3..f06e1d76e16a 100644
--- a/include/net/xfrm.h
+++ b/include/net/xfrm.h
@@ -37,6 +37,7 @@
#define XFRM_PROTO_COMP 108
#define XFRM_PROTO_IPIP 4
#define XFRM_PROTO_IPV6 41
+#define XFRM_PROTO_IPTFS IPPROTO_AGGFRAG
#define XFRM_PROTO_ROUTING IPPROTO_ROUTING
#define XFRM_PROTO_DSTOPTS IPPROTO_DSTOPTS
diff --git a/include/uapi/linux/ipsec.h b/include/uapi/linux/ipsec.h
index 50d8ee1791e2..696b790f4346 100644
--- a/include/uapi/linux/ipsec.h
+++ b/include/uapi/linux/ipsec.h
@@ -14,7 +14,8 @@ enum {
IPSEC_MODE_ANY = 0, /* We do not support this for SA */
IPSEC_MODE_TRANSPORT = 1,
IPSEC_MODE_TUNNEL = 2,
- IPSEC_MODE_BEET = 3
+ IPSEC_MODE_BEET = 3,
+ IPSEC_MODE_IPTFS = 4
};
enum {
diff --git a/include/uapi/linux/snmp.h b/include/uapi/linux/snmp.h
index adf5fd78dd50..77eb078f06a6 100644
--- a/include/uapi/linux/snmp.h
+++ b/include/uapi/linux/snmp.h
@@ -339,6 +339,9 @@ enum
LINUX_MIB_XFRMACQUIREERROR, /* XfrmAcquireError */
LINUX_MIB_XFRMOUTSTATEDIRERROR, /* XfrmOutStateDirError */
LINUX_MIB_XFRMINSTATEDIRERROR, /* XfrmInStateDirError */
+ LINUX_MIB_XFRMNOSKBERROR, /* XfrmNoSkbError */
+ LINUX_MIB_XFRMINIPTFSERROR, /* XfrmInIptfsError */
+ LINUX_MIB_XFRMOUTNOQSPACE, /* XfrmOutNoQueueSpace */
__LINUX_MIB_XFRMMAX
};
diff --git a/net/ipv4/esp4.c b/net/ipv4/esp4.c
index 47378ca41904..9a0165e3ceba 100644
--- a/net/ipv4/esp4.c
+++ b/net/ipv4/esp4.c
@@ -815,7 +815,8 @@ int esp_input_done2(struct sk_buff *skb, int err)
}
skb_pull_rcsum(skb, hlen);
- if (x->props.mode == XFRM_MODE_TUNNEL)
+ if (x->props.mode == XFRM_MODE_TUNNEL ||
+ x->props.mode == XFRM_MODE_IPTFS)
skb_reset_transport_header(skb);
else
skb_set_transport_header(skb, -ihl);
diff --git a/net/ipv6/esp6.c b/net/ipv6/esp6.c
index 3920e8aa1031..5f85435de722 100644
--- a/net/ipv6/esp6.c
+++ b/net/ipv6/esp6.c
@@ -858,7 +858,8 @@ int esp6_input_done2(struct sk_buff *skb, int err)
skb_postpull_rcsum(skb, skb_network_header(skb),
skb_network_header_len(skb));
skb_pull_rcsum(skb, hlen);
- if (x->props.mode == XFRM_MODE_TUNNEL)
+ if (x->props.mode == XFRM_MODE_TUNNEL ||
+ x->props.mode == XFRM_MODE_IPTFS)
skb_reset_transport_header(skb);
else
skb_set_transport_header(skb, -hdr_len);
diff --git a/net/netfilter/nft_xfrm.c b/net/netfilter/nft_xfrm.c
index 1c866757db55..620238c6ef4c 100644
--- a/net/netfilter/nft_xfrm.c
+++ b/net/netfilter/nft_xfrm.c
@@ -112,7 +112,8 @@ static bool xfrm_state_addr_ok(enum nft_xfrm_keys k, u8 family, u8 mode)
return true;
}
- return mode == XFRM_MODE_BEET || mode == XFRM_MODE_TUNNEL;
+ return mode == XFRM_MODE_BEET || mode == XFRM_MODE_TUNNEL ||
+ mode == XFRM_MODE_IPTFS;
}
static void nft_xfrm_state_get_key(const struct nft_xfrm *priv,
diff --git a/net/xfrm/xfrm_device.c b/net/xfrm/xfrm_device.c
index e412e4afb169..d4905796e9ab 100644
--- a/net/xfrm/xfrm_device.c
+++ b/net/xfrm/xfrm_device.c
@@ -69,6 +69,7 @@ static void __xfrm_mode_beet_prep(struct xfrm_state *x, struct sk_buff *skb,
static void xfrm_outer_mode_prep(struct xfrm_state *x, struct sk_buff *skb)
{
switch (x->outer_mode.encap) {
+ case XFRM_MODE_IPTFS:
case XFRM_MODE_TUNNEL:
if (x->outer_mode.family == AF_INET)
return __xfrm_mode_tunnel_prep(x, skb,
diff --git a/net/xfrm/xfrm_output.c b/net/xfrm/xfrm_output.c
index ef81359e4038..b5025cf6136e 100644
--- a/net/xfrm/xfrm_output.c
+++ b/net/xfrm/xfrm_output.c
@@ -677,6 +677,10 @@ static void xfrm_get_inner_ipproto(struct sk_buff *skb, struct xfrm_state *x)
return;
}
+ if (x->outer_mode.encap == XFRM_MODE_IPTFS) {
+ xo->inner_ipproto = IPPROTO_AGGFRAG;
+ return;
+ }
/* non-Tunnel Mode */
if (!skb->encapsulation)
diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c
index f764b1409175..928c3ed79d21 100644
--- a/net/xfrm/xfrm_policy.c
+++ b/net/xfrm/xfrm_policy.c
@@ -2473,6 +2473,7 @@ xfrm_tmpl_resolve_one(struct xfrm_policy *policy, const struct flowi *fl,
struct xfrm_tmpl *tmpl = &policy->xfrm_vec[i];
if (tmpl->mode == XFRM_MODE_TUNNEL ||
+ tmpl->mode == XFRM_MODE_IPTFS ||
tmpl->mode == XFRM_MODE_BEET) {
remote = &tmpl->id.daddr;
local = &tmpl->saddr;
@@ -3264,7 +3265,8 @@ struct dst_entry *xfrm_lookup_with_ifid(struct net *net,
ok:
xfrm_pols_put(pols, drop_pols);
if (dst && dst->xfrm &&
- dst->xfrm->props.mode == XFRM_MODE_TUNNEL)
+ (dst->xfrm->props.mode == XFRM_MODE_TUNNEL ||
+ dst->xfrm->props.mode == XFRM_MODE_IPTFS))
dst->flags |= DST_XFRM_TUNNEL;
return dst;
@@ -4509,6 +4511,7 @@ static int migrate_tmpl_match(const struct xfrm_migrate *m, const struct xfrm_tm
switch (t->mode) {
case XFRM_MODE_TUNNEL:
case XFRM_MODE_BEET:
+ case XFRM_MODE_IPTFS:
if (xfrm_addr_equal(&t->id.daddr, &m->old_daddr,
m->old_family) &&
xfrm_addr_equal(&t->saddr, &m->old_saddr,
@@ -4551,7 +4554,8 @@ static int xfrm_policy_migrate(struct xfrm_policy *pol,
continue;
n++;
if (pol->xfrm_vec[i].mode != XFRM_MODE_TUNNEL &&
- pol->xfrm_vec[i].mode != XFRM_MODE_BEET)
+ pol->xfrm_vec[i].mode != XFRM_MODE_BEET &&
+ pol->xfrm_vec[i].mode != XFRM_MODE_IPTFS)
continue;
/* update endpoints */
memcpy(&pol->xfrm_vec[i].id.daddr, &mp->new_daddr,
diff --git a/net/xfrm/xfrm_proc.c b/net/xfrm/xfrm_proc.c
index eeb984be03a7..e851b388995a 100644
--- a/net/xfrm/xfrm_proc.c
+++ b/net/xfrm/xfrm_proc.c
@@ -43,6 +43,9 @@ static const struct snmp_mib xfrm_mib_list[] = {
SNMP_MIB_ITEM("XfrmAcquireError", LINUX_MIB_XFRMACQUIREERROR),
SNMP_MIB_ITEM("XfrmOutStateDirError", LINUX_MIB_XFRMOUTSTATEDIRERROR),
SNMP_MIB_ITEM("XfrmInStateDirError", LINUX_MIB_XFRMINSTATEDIRERROR),
+ SNMP_MIB_ITEM("XfrmNoSkbError", LINUX_MIB_XFRMNOSKBERROR),
+ SNMP_MIB_ITEM("XfrmInIptfsError", LINUX_MIB_XFRMINIPTFSERROR),
+ SNMP_MIB_ITEM("XfrmOutNoQueueSpace", LINUX_MIB_XFRMOUTNOQSPACE),
SNMP_MIB_SENTINEL
};
diff --git a/net/xfrm/xfrm_state.c b/net/xfrm/xfrm_state.c
index e7b656bb6c0d..1d4b884f82d0 100644
--- a/net/xfrm/xfrm_state.c
+++ b/net/xfrm/xfrm_state.c
@@ -467,6 +467,11 @@ static const struct xfrm_mode xfrm4_mode_map[XFRM_MODE_MAX] = {
.flags = XFRM_MODE_FLAG_TUNNEL,
.family = AF_INET,
},
+ [XFRM_MODE_IPTFS] = {
+ .encap = XFRM_MODE_IPTFS,
+ .flags = XFRM_MODE_FLAG_TUNNEL,
+ .family = AF_INET,
+ },
};
static const struct xfrm_mode xfrm6_mode_map[XFRM_MODE_MAX] = {
@@ -488,6 +493,11 @@ static const struct xfrm_mode xfrm6_mode_map[XFRM_MODE_MAX] = {
.flags = XFRM_MODE_FLAG_TUNNEL,
.family = AF_INET6,
},
+ [XFRM_MODE_IPTFS] = {
+ .encap = XFRM_MODE_IPTFS,
+ .flags = XFRM_MODE_FLAG_TUNNEL,
+ .family = AF_INET6,
+ },
};
static const struct xfrm_mode *xfrm_get_mode(unsigned int encap, int family)
@@ -2194,6 +2204,7 @@ static int __xfrm6_state_sort_cmp(const void *p)
#endif
case XFRM_MODE_TUNNEL:
case XFRM_MODE_BEET:
+ case XFRM_MODE_IPTFS:
return 4;
}
return 5;
@@ -2220,6 +2231,7 @@ static int __xfrm6_tmpl_sort_cmp(const void *p)
#endif
case XFRM_MODE_TUNNEL:
case XFRM_MODE_BEET:
+ case XFRM_MODE_IPTFS:
return 3;
}
return 4;
diff --git a/net/xfrm/xfrm_user.c b/net/xfrm/xfrm_user.c
index 419bbeea6b20..40c79bd14a7e 100644
--- a/net/xfrm/xfrm_user.c
+++ b/net/xfrm/xfrm_user.c
@@ -379,6 +379,16 @@ static int verify_newsa_info(struct xfrm_usersa_info *p,
case XFRM_MODE_ROUTEOPTIMIZATION:
case XFRM_MODE_BEET:
break;
+ case XFRM_MODE_IPTFS:
+ if (p->id.proto != IPPROTO_ESP) {
+ NL_SET_ERR_MSG(extack, "IP-TFS mode only supported with ESP");
+ goto out;
+ }
+ if (sa_dir == 0) {
+ NL_SET_ERR_MSG(extack, "IP-TFS mode requires in or out direction attribute");
+ goto out;
+ }
+ break;
default:
NL_SET_ERR_MSG(extack, "Unsupported mode");
@@ -1984,6 +1994,8 @@ static int validate_tmpl(int nr, struct xfrm_user_tmpl *ut, u16 family,
return -EINVAL;
}
break;
+ case XFRM_MODE_IPTFS:
+ break;
default:
if (ut[i].family != prev_family) {
NL_SET_ERR_MSG(extack, "Mode in template doesn't support a family change");
--
2.46.0
^ permalink raw reply related [flat|nested] 40+ messages in thread* [PATCH ipsec-next v8 07/16] xfrm: iptfs: add new iptfs xfrm mode impl
2024-08-04 20:33 [PATCH ipsec-next v8 00/16] Add IP-TFS mode to xfrm Christian Hopps
` (5 preceding siblings ...)
2024-08-04 20:33 ` [PATCH ipsec-next v8 06/16] xfrm: add generic iptfs defines and functionality Christian Hopps
@ 2024-08-04 20:33 ` Christian Hopps
2024-08-04 20:33 ` [PATCH ipsec-next v8 08/16] xfrm: iptfs: add user packet (tunnel ingress) handling Christian Hopps
` (8 subsequent siblings)
15 siblings, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-04 20:33 UTC (permalink / raw)
To: devel; +Cc: Steffen Klassert, netdev, Christian Hopps, Christian Hopps
From: Christian Hopps <chopps@labn.net>
Add a new xfrm mode implementing AggFrag/IP-TFS from RFC9347.
This utilizes the new xfrm_mode_cbs to implement demand-driven IP-TFS
functionality. This functionality can be used to increase bandwidth
utilization through small packet aggregation, as well as help solve PMTU
issues through it's efficient use of fragmentation.
Link: https://www.rfc-editor.org/rfc/rfc9347.txt
Multiple commits follow to build the functionality into xfrm_iptfs.c
Signed-off-by: Christian Hopps <chopps@labn.net>
---
net/xfrm/Makefile | 1 +
net/xfrm/xfrm_iptfs.c | 210 ++++++++++++++++++++++++++++++++++++++++++
2 files changed, 211 insertions(+)
create mode 100644 net/xfrm/xfrm_iptfs.c
diff --git a/net/xfrm/Makefile b/net/xfrm/Makefile
index 512e0b2f8514..5a1787587cb3 100644
--- a/net/xfrm/Makefile
+++ b/net/xfrm/Makefile
@@ -21,5 +21,6 @@ obj-$(CONFIG_XFRM_USER) += xfrm_user.o
obj-$(CONFIG_XFRM_USER_COMPAT) += xfrm_compat.o
obj-$(CONFIG_XFRM_IPCOMP) += xfrm_ipcomp.o
obj-$(CONFIG_XFRM_INTERFACE) += xfrm_interface.o
+obj-$(CONFIG_XFRM_IPTFS) += xfrm_iptfs.o
obj-$(CONFIG_XFRM_ESPINTCP) += espintcp.o
obj-$(CONFIG_DEBUG_INFO_BTF) += xfrm_state_bpf.o
diff --git a/net/xfrm/xfrm_iptfs.c b/net/xfrm/xfrm_iptfs.c
new file mode 100644
index 000000000000..201406175d17
--- /dev/null
+++ b/net/xfrm/xfrm_iptfs.c
@@ -0,0 +1,210 @@
+// SPDX-License-Identifier: GPL-2.0
+/* xfrm_iptfs: IPTFS encapsulation support
+ *
+ * April 21 2022, Christian Hopps <chopps@labn.net>
+ *
+ * Copyright (c) 2022, LabN Consulting, L.L.C.
+ *
+ */
+
+#include <linux/kernel.h>
+#include <linux/icmpv6.h>
+#include <net/gro.h>
+#include <net/icmp.h>
+#include <net/ip6_route.h>
+#include <net/inet_ecn.h>
+#include <net/xfrm.h>
+
+#include <crypto/aead.h>
+
+#include "xfrm_inout.h"
+
+/**
+ * struct xfrm_iptfs_config - configuration for the IPTFS tunnel.
+ * @pkt_size: size of the outer IP packet. 0 to use interface and MTU discovery,
+ * otherwise the user specified value.
+ */
+struct xfrm_iptfs_config {
+ u32 pkt_size; /* outer_packet_size or 0 */
+};
+
+/**
+ * struct xfrm_iptfs_data - mode specific xfrm state.
+ * @cfg: IPTFS tunnel config.
+ * @x: owning SA (xfrm_state).
+ * @payload_mtu: max payload size.
+ */
+struct xfrm_iptfs_data {
+ struct xfrm_iptfs_config cfg;
+
+ /* Ingress User Input */
+ struct xfrm_state *x; /* owning state */
+ u32 payload_mtu; /* max payload size */
+};
+
+/* ========================== */
+/* State Management Functions */
+/* ========================== */
+
+/**
+ * iptfs_get_inner_mtu() - return inner MTU with no fragmentation.
+ * @x: xfrm state.
+ * @outer_mtu: the outer mtu
+ */
+static u32 iptfs_get_inner_mtu(struct xfrm_state *x, int outer_mtu)
+{
+ struct crypto_aead *aead;
+ u32 blksize;
+
+ aead = x->data;
+ blksize = ALIGN(crypto_aead_blocksize(aead), 4);
+ return ((outer_mtu - x->props.header_len - crypto_aead_authsize(aead)) &
+ ~(blksize - 1)) - 2;
+}
+
+/**
+ * iptfs_user_init() - initialize the SA with IPTFS options from netlink.
+ * @net: the net data
+ * @x: xfrm state
+ * @attrs: netlink attributes
+ * @extack: extack return data
+ *
+ * Return: 0 on success or a negative error code on failure
+ */
+static int iptfs_user_init(struct net *net, struct xfrm_state *x,
+ struct nlattr **attrs,
+ struct netlink_ext_ack *extack)
+{
+ struct xfrm_iptfs_data *xtfs = x->mode_data;
+ struct xfrm_iptfs_config *xc;
+
+ xc = &xtfs->cfg;
+
+ if (attrs[XFRMA_IPTFS_PKT_SIZE]) {
+ xc->pkt_size = nla_get_u32(attrs[XFRMA_IPTFS_PKT_SIZE]);
+ if (!xc->pkt_size) {
+ xtfs->payload_mtu = 0;
+ } else if (xc->pkt_size > x->props.header_len) {
+ xtfs->payload_mtu = xc->pkt_size - x->props.header_len;
+ } else {
+ NL_SET_ERR_MSG(extack,
+ "Packet size must be 0 or greater than IPTFS/ESP header length");
+ return -EINVAL;
+ }
+ }
+ return 0;
+}
+
+static unsigned int iptfs_sa_len(const struct xfrm_state *x)
+{
+ struct xfrm_iptfs_data *xtfs = x->mode_data;
+ struct xfrm_iptfs_config *xc = &xtfs->cfg;
+ unsigned int l = 0;
+
+ if (x->dir == XFRM_SA_DIR_OUT)
+ l += nla_total_size(sizeof(xc->pkt_size));
+
+ return l;
+}
+
+static int iptfs_copy_to_user(struct xfrm_state *x, struct sk_buff *skb)
+{
+ struct xfrm_iptfs_data *xtfs = x->mode_data;
+ struct xfrm_iptfs_config *xc = &xtfs->cfg;
+ int ret = 0;
+
+ if (x->dir == XFRM_SA_DIR_OUT)
+ ret = nla_put_u32(skb, XFRMA_IPTFS_PKT_SIZE, xc->pkt_size);
+
+ return ret;
+}
+
+static void __iptfs_init_state(struct xfrm_state *x,
+ struct xfrm_iptfs_data *xtfs)
+{
+ /* Modify type (esp) adjustment values */
+
+ if (x->props.family == AF_INET)
+ x->props.header_len += sizeof(struct iphdr) + sizeof(struct ip_iptfs_hdr);
+ else if (x->props.family == AF_INET6)
+ x->props.header_len += sizeof(struct ipv6hdr) + sizeof(struct ip_iptfs_hdr);
+ x->props.enc_hdr_len = sizeof(struct ip_iptfs_hdr);
+
+ /* Always keep a module reference when x->mode_data is set */
+ __module_get(x->mode_cbs->owner);
+
+ x->mode_data = xtfs;
+ xtfs->x = x;
+}
+
+static int iptfs_clone(struct xfrm_state *x, struct xfrm_state *orig)
+{
+ struct xfrm_iptfs_data *xtfs;
+
+ xtfs = kmemdup(orig->mode_data, sizeof(*xtfs), GFP_KERNEL);
+ if (!xtfs)
+ return -ENOMEM;
+
+ __iptfs_init_state(x, xtfs);
+
+ return 0;
+}
+
+static int iptfs_create_state(struct xfrm_state *x)
+{
+ struct xfrm_iptfs_data *xtfs;
+
+ xtfs = kzalloc(sizeof(*xtfs), GFP_KERNEL);
+ if (!xtfs)
+ return -ENOMEM;
+
+ __iptfs_init_state(x, xtfs);
+
+ return 0;
+}
+
+static void iptfs_delete_state(struct xfrm_state *x)
+{
+ struct xfrm_iptfs_data *xtfs = x->mode_data;
+
+ if (!xtfs)
+ return;
+
+ kfree_sensitive(xtfs);
+
+ module_put(x->mode_cbs->owner);
+}
+
+static const struct xfrm_mode_cbs iptfs_mode_cbs = {
+ .owner = THIS_MODULE,
+ .create_state = iptfs_create_state,
+ .delete_state = iptfs_delete_state,
+ .user_init = iptfs_user_init,
+ .copy_to_user = iptfs_copy_to_user,
+ .sa_len = iptfs_sa_len,
+ .clone = iptfs_clone,
+ .get_inner_mtu = iptfs_get_inner_mtu,
+};
+
+static int __init xfrm_iptfs_init(void)
+{
+ int err;
+
+ pr_info("xfrm_iptfs: IPsec IP-TFS tunnel mode module\n");
+
+ err = xfrm_register_mode_cbs(XFRM_MODE_IPTFS, &iptfs_mode_cbs);
+ if (err < 0)
+ pr_info("%s: can't register IP-TFS\n", __func__);
+
+ return err;
+}
+
+static void __exit xfrm_iptfs_fini(void)
+{
+ xfrm_unregister_mode_cbs(XFRM_MODE_IPTFS);
+}
+
+module_init(xfrm_iptfs_init);
+module_exit(xfrm_iptfs_fini);
+MODULE_LICENSE("GPL");
+MODULE_DESCRIPTION("IP-TFS support for xfrm ipsec tunnels");
--
2.46.0
^ permalink raw reply related [flat|nested] 40+ messages in thread* [PATCH ipsec-next v8 08/16] xfrm: iptfs: add user packet (tunnel ingress) handling
2024-08-04 20:33 [PATCH ipsec-next v8 00/16] Add IP-TFS mode to xfrm Christian Hopps
` (6 preceding siblings ...)
2024-08-04 20:33 ` [PATCH ipsec-next v8 07/16] xfrm: iptfs: add new iptfs xfrm mode impl Christian Hopps
@ 2024-08-04 20:33 ` Christian Hopps
2024-08-05 17:10 ` Simon Horman
2024-08-04 20:33 ` [PATCH ipsec-next v8 09/16] xfrm: iptfs: share page fragments of inner packets Christian Hopps
` (7 subsequent siblings)
15 siblings, 1 reply; 40+ messages in thread
From: Christian Hopps @ 2024-08-04 20:33 UTC (permalink / raw)
To: devel; +Cc: Steffen Klassert, netdev, Christian Hopps, Christian Hopps
From: Christian Hopps <chopps@labn.net>
Add tunnel packet output functionality. This is code handles
the ingress to the tunnel.
Signed-off-by: Christian Hopps <chopps@labn.net>
---
net/xfrm/xfrm_iptfs.c | 577 +++++++++++++++++++++++++++++++++++++++++-
1 file changed, 574 insertions(+), 3 deletions(-)
diff --git a/net/xfrm/xfrm_iptfs.c b/net/xfrm/xfrm_iptfs.c
index 201406175d17..9c48c15cbed0 100644
--- a/net/xfrm/xfrm_iptfs.c
+++ b/net/xfrm/xfrm_iptfs.c
@@ -19,29 +19,553 @@
#include "xfrm_inout.h"
+/* ------------------------------------------------ */
+/* IPTFS default SA values (tunnel ingress/dir-out) */
+/* ------------------------------------------------ */
+
+/**
+ * define IPTFS_DEFAULT_INIT_DELAY_USECS - default initial output delay
+ *
+ * The initial output delay is the amount of time prior to servicing the output
+ * queue after queueing the first packet on said queue. This applies anytime the
+ * output queue was previously empty.
+ *
+ * Default 0.
+ */
+#define IPTFS_DEFAULT_INIT_DELAY_USECS 0
+
+/**
+ * define IPTFS_DEFAULT_MAX_QUEUE_SIZE - default max output queue size.
+ *
+ * The default IPTFS max output queue size in octets. The output queue is where
+ * received packets destined for output over an IPTFS tunnel are stored prior to
+ * being output in aggregated/fragmented form over the IPTFS tunnel.
+ *
+ * Default 1M.
+ */
+#define IPTFS_DEFAULT_MAX_QUEUE_SIZE (1024 * 10240)
+
+#define NSECS_IN_USEC 1000
+
+#define IPTFS_HRTIMER_MODE HRTIMER_MODE_REL_SOFT
+
/**
* struct xfrm_iptfs_config - configuration for the IPTFS tunnel.
* @pkt_size: size of the outer IP packet. 0 to use interface and MTU discovery,
* otherwise the user specified value.
+ * @max_queue_size: The maximum number of octets allowed to be queued to be sent
+ * over the IPTFS SA. The queue size is measured as the size of all the
+ * packets enqueued.
*/
struct xfrm_iptfs_config {
u32 pkt_size; /* outer_packet_size or 0 */
+ u32 max_queue_size; /* octets */
};
/**
* struct xfrm_iptfs_data - mode specific xfrm state.
* @cfg: IPTFS tunnel config.
* @x: owning SA (xfrm_state).
+ * @queue: queued user packets to send.
+ * @queue_size: number of octets on queue (sum of packet sizes).
+ * @ecn_queue_size: octets above with ECN mark.
+ * @init_delay_ns: nanoseconds to wait to send initial IPTFS packet.
+ * @iptfs_timer: output timer.
* @payload_mtu: max payload size.
*/
struct xfrm_iptfs_data {
struct xfrm_iptfs_config cfg;
/* Ingress User Input */
- struct xfrm_state *x; /* owning state */
+ struct xfrm_state *x; /* owning state */
+ struct sk_buff_head queue; /* output queue */
+
+ u32 queue_size; /* octets */
+ u32 ecn_queue_size; /* octets above which ECN mark */
+ u64 init_delay_ns; /* nanoseconds */
+ struct hrtimer iptfs_timer; /* output timer */
u32 payload_mtu; /* max payload size */
};
+static u32 iptfs_get_inner_mtu(struct xfrm_state *x, int outer_mtu);
+static enum hrtimer_restart iptfs_delay_timer(struct hrtimer *me);
+
+/* ================================= */
+/* IPTFS Sending (ingress) Functions */
+/* ================================= */
+
+/* ------------------------- */
+/* Enqueue to send functions */
+/* ------------------------- */
+
+/**
+ * iptfs_enqueue() - enqueue packet if ok to send.
+ * @xtfs: xtfs state
+ * @skb: the packet
+ *
+ * Return: true if packet enqueued.
+ */
+static bool iptfs_enqueue(struct xfrm_iptfs_data *xtfs, struct sk_buff *skb)
+{
+ u64 newsz = xtfs->queue_size + skb->len;
+ struct iphdr *iph;
+
+ assert_spin_locked(&xtfs->x->lock);
+
+ if (newsz > xtfs->cfg.max_queue_size)
+ return false;
+
+ /* Set ECN CE if we are above our ECN queue threshold */
+ if (newsz > xtfs->ecn_queue_size) {
+ iph = ip_hdr(skb);
+ if (iph->version == 4)
+ IP_ECN_set_ce(iph);
+ else if (iph->version == 6)
+ IP6_ECN_set_ce(skb, ipv6_hdr(skb));
+ }
+
+ __skb_queue_tail(&xtfs->queue, skb);
+ xtfs->queue_size += skb->len;
+ return true;
+}
+
+static int iptfs_get_cur_pmtu(struct xfrm_state *x,
+ struct xfrm_iptfs_data *xtfs, struct sk_buff *skb)
+{
+ struct xfrm_dst *xdst = (struct xfrm_dst *)skb_dst(skb);
+ u32 payload_mtu = xtfs->payload_mtu;
+ u32 pmtu = iptfs_get_inner_mtu(x, xdst->child_mtu_cached);
+
+ if (payload_mtu && payload_mtu < pmtu)
+ pmtu = payload_mtu;
+
+ return pmtu;
+}
+
+static int iptfs_is_too_big(struct sock *sk, struct sk_buff *skb, u32 pmtu)
+{
+ if (skb->len <= pmtu)
+ return 0;
+
+ /* We only send ICMP too big if the user has configured us as
+ * dont-fragment.
+ */
+ if (skb->dev)
+ XFRM_INC_STATS(dev_net(skb->dev), LINUX_MIB_XFRMOUTERROR);
+
+ if (sk) {
+ xfrm_local_error(skb, pmtu);
+ } else if (ip_hdr(skb)->version == 4) {
+ icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
+ htonl(pmtu));
+ } else {
+ WARN_ON_ONCE(ip_hdr(skb)->version != 6);
+ icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, pmtu);
+ }
+ return 1;
+}
+
+/* IPv4/IPv6 packet ingress to IPTFS tunnel, arrange to send in IPTFS payload
+ * (i.e., aggregating or fragmenting as appropriate).
+ * This is set in dst->output for an SA.
+ */
+static int iptfs_output_collect(struct net *net, struct sock *sk,
+ struct sk_buff *skb)
+{
+ struct dst_entry *dst = skb_dst(skb);
+ struct xfrm_state *x = dst->xfrm;
+ struct xfrm_iptfs_data *xtfs = x->mode_data;
+ struct sk_buff *segs, *nskb;
+ u32 pmtu = 0;
+ bool ok = true;
+ bool was_gso;
+
+ /* We have hooked into dst_entry->output which means we have skipped the
+ * protocol specific netfilter (see xfrm4_output, xfrm6_output).
+ * when our timer runs we will end up calling xfrm_output directly on
+ * the encapsulated traffic.
+ *
+ * For both cases this is the NF_INET_POST_ROUTING hook which allows
+ * changing the skb->dst entry which then may not be xfrm based anymore
+ * in which case a REROUTED flag is set. and dst_output is called.
+ *
+ * For IPv6 we are also skipping fragmentation handling for local
+ * sockets, which may or may not be good depending on our tunnel DF
+ * setting. Normally with fragmentation supported we want to skip this
+ * fragmentation.
+ */
+
+ BUG_ON(!xtfs);
+
+ pmtu = iptfs_get_cur_pmtu(x, xtfs, skb);
+
+ /* Break apart GSO skbs. If the queue is nearing full then we want the
+ * accounting and queuing to be based on the individual packets not on the
+ * aggregate GSO buffer.
+ */
+ was_gso = skb_is_gso(skb);
+ if (!was_gso) {
+ segs = skb;
+ } else {
+ segs = skb_gso_segment(skb, 0);
+ if (IS_ERR_OR_NULL(segs)) {
+ XFRM_INC_STATS(net, LINUX_MIB_XFRMOUTERROR);
+ kfree_skb(skb);
+ if (IS_ERR(segs))
+ return PTR_ERR(segs);
+ return -EINVAL;
+ }
+ consume_skb(skb);
+ skb = NULL;
+ }
+
+ /* We can be running on multiple cores and from the network softirq or
+ * from user context depending on where the packet is coming from.
+ */
+ spin_lock_bh(&x->lock);
+
+ skb_list_walk_safe(segs, skb, nskb) {
+ skb_mark_not_on_list(skb);
+
+ /* Once we drop due to no queue space we continue to drop the
+ * rest of the packets from that GRO.
+ */
+ if (!ok) {
+nospace:
+ XFRM_INC_STATS(net, LINUX_MIB_XFRMOUTNOQSPACE);
+ kfree_skb_reason(skb, SKB_DROP_REASON_FULL_RING);
+ continue;
+ }
+
+ /* Fragmenting handled in following commits. */
+ if (iptfs_is_too_big(sk, skb, pmtu)) {
+ kfree_skb_reason(skb, SKB_DROP_REASON_PKT_TOO_BIG);
+ continue;
+ }
+
+ /* Enqueue to send in tunnel */
+ ok = iptfs_enqueue(xtfs, skb);
+ if (!ok)
+ goto nospace;
+ }
+
+ /* Start a delay timer if we don't have one yet */
+ if (!hrtimer_is_queued(&xtfs->iptfs_timer))
+ hrtimer_start(&xtfs->iptfs_timer, xtfs->init_delay_ns,
+ IPTFS_HRTIMER_MODE);
+
+ spin_unlock_bh(&x->lock);
+ return 0;
+}
+
+/* -------------------------- */
+/* Dequeue and send functions */
+/* -------------------------- */
+
+static void iptfs_output_prepare_skb(struct sk_buff *skb, u32 blkoff)
+{
+ struct ip_iptfs_hdr *h;
+ size_t hsz = sizeof(*h);
+
+ /* now reset values to be pointing at the rest of the packets */
+ h = skb_push(skb, hsz);
+ memset(h, 0, hsz);
+ if (blkoff)
+ h->block_offset = htons(blkoff);
+
+ /* network_header current points at the inner IP packet
+ * move it to the iptfs header
+ */
+ skb->transport_header = skb->network_header;
+ skb->network_header -= hsz;
+
+ IPCB(skb)->flags |= IPSKB_XFRM_TUNNEL_SIZE;
+}
+
+static struct sk_buff **iptfs_rehome_fraglist(struct sk_buff **nextp,
+ struct sk_buff *child)
+{
+ u32 fllen = 0;
+
+ /* It might be possible to account for a frag list in addition to page
+ * fragment if it's a valid state to be in. The page fragments size
+ * should be kept as data_len so only the frag_list size is removed,
+ * this must be done above as well.
+ */
+ BUG_ON(skb_shinfo(child)->nr_frags);
+ *nextp = skb_shinfo(child)->frag_list;
+ while (*nextp) {
+ fllen += (*nextp)->len;
+ nextp = &(*nextp)->next;
+ }
+ skb_frag_list_init(child);
+ BUG_ON(fllen > child->data_len);
+ child->len -= fllen;
+ child->data_len -= fllen;
+
+ return nextp;
+}
+
+static void iptfs_output_queued(struct xfrm_state *x, struct sk_buff_head *list)
+{
+ struct xfrm_iptfs_data *xtfs = x->mode_data;
+ struct sk_buff *skb, *skb2, **nextp;
+ struct skb_shared_info *shi;
+
+ while ((skb = __skb_dequeue(list))) {
+ u32 mtu = iptfs_get_cur_pmtu(x, xtfs, skb);
+ int remaining;
+
+ /* protocol comes to us cleared sometimes */
+ skb->protocol = x->outer_mode.family == AF_INET ?
+ htons(ETH_P_IP) :
+ htons(ETH_P_IPV6);
+
+ if (skb->len > mtu) {
+ /* We handle this case before enqueueing so we are only
+ * here b/c MTU changed after we enqueued before we
+ * dequeued, just drop these.
+ */
+ XFRM_INC_STATS(xs_net(x), LINUX_MIB_XFRMOUTERROR);
+
+ kfree_skb_reason(skb, SKB_DROP_REASON_PKT_TOO_BIG);
+ continue;
+ }
+
+ /* If we don't have a cksum in the packet we need to add one
+ * before encapsulation.
+ */
+ if (skb->ip_summed == CHECKSUM_PARTIAL) {
+ if (skb_checksum_help(skb)) {
+ XFRM_INC_STATS(dev_net(skb_dst(skb)->dev),
+ LINUX_MIB_XFRMOUTERROR);
+ kfree_skb(skb);
+ continue;
+ }
+ }
+
+ /* Convert first inner packet into an outer IPTFS packet */
+ iptfs_output_prepare_skb(skb, 0);
+
+ /* The space remaining to send more inner packet data is `mtu` -
+ * (skb->len - sizeof iptfs header). This is b/c the `mtu` value
+ * has the basic IPTFS header len accounted for, and we added
+ * that header to the skb so it is a part of skb->len, thus we
+ * subtract it from the skb length.
+ */
+ remaining = mtu - (skb->len - sizeof(struct ip_iptfs_hdr));
+
+ /* Re-home (un-nest) nested fragment lists. We need to do this
+ * b/c we will simply be appending any following aggregated
+ * inner packets to the frag list.
+ */
+ shi = skb_shinfo(skb);
+ nextp = &shi->frag_list;
+ while (*nextp) {
+ if (skb_has_frag_list(*nextp))
+ nextp = iptfs_rehome_fraglist(&(*nextp)->next,
+ *nextp);
+ else
+ nextp = &(*nextp)->next;
+ }
+
+ /* See if we have enough space to simply append.
+ *
+ * NOTE: Maybe do not append if we will be mis-aligned,
+ * SW-based endpoints will probably have to copy in this
+ * case.
+ */
+ while ((skb2 = skb_peek(list))) {
+ if (skb2->len > remaining)
+ break;
+
+ __skb_unlink(skb2, list);
+
+ /* If we don't have a cksum in the packet we need to add
+ * one before encapsulation.
+ */
+ if (skb2->ip_summed == CHECKSUM_PARTIAL) {
+ if (skb_checksum_help(skb2)) {
+ XFRM_INC_STATS(xs_net(x),
+ LINUX_MIB_XFRMOUTERROR);
+ kfree_skb(skb2);
+ continue;
+ }
+ }
+
+ /* Do accounting */
+ skb->data_len += skb2->len;
+ skb->len += skb2->len;
+ remaining -= skb2->len;
+
+ /* Append to the frag_list */
+ *nextp = skb2;
+ nextp = &skb2->next;
+ BUG_ON(*nextp);
+ if (skb_has_frag_list(skb2))
+ nextp = iptfs_rehome_fraglist(nextp, skb2);
+ skb->truesize += skb2->truesize;
+ }
+
+ xfrm_output(NULL, skb);
+ }
+}
+
+static enum hrtimer_restart iptfs_delay_timer(struct hrtimer *me)
+{
+ struct sk_buff_head list;
+ struct xfrm_iptfs_data *xtfs;
+ struct xfrm_state *x;
+
+ xtfs = container_of(me, typeof(*xtfs), iptfs_timer);
+ x = xtfs->x;
+
+ /* Process all the queued packets
+ *
+ * softirq execution order: timer > tasklet > hrtimer
+ *
+ * Network rx will have run before us giving one last chance to queue
+ * ingress packets for us to process and transmit.
+ */
+
+ spin_lock(&x->lock);
+ __skb_queue_head_init(&list);
+ skb_queue_splice_init(&xtfs->queue, &list);
+ xtfs->queue_size = 0;
+ spin_unlock(&x->lock);
+
+ /* After the above unlock, packets can begin queuing again, and the
+ * timer can be set again, from another CPU either in softirq or user
+ * context (not from this one since we are running at softirq level
+ * already).
+ */
+
+ iptfs_output_queued(x, &list);
+
+ return HRTIMER_NORESTART;
+}
+
+/**
+ * iptfs_encap_add_ipv4() - add outer encaps
+ * @x: xfrm state
+ * @skb: the packet
+ *
+ * This was originally taken from xfrm4_tunnel_encap_add. The reason for the
+ * copy is that IP-TFS/AGGFRAG can have different functionality for how to set
+ * the TOS/DSCP bits. Sets the protocol to a different value and doesn't do
+ * anything with inner headers as they aren't pointing into a normal IP
+ * singleton inner packet.
+ *
+ * Return: 0 on success or a negative error code on failure
+ */
+static int iptfs_encap_add_ipv4(struct xfrm_state *x, struct sk_buff *skb)
+{
+ struct dst_entry *dst = skb_dst(skb);
+ struct iphdr *top_iph;
+
+ skb_reset_inner_network_header(skb);
+ skb_reset_inner_transport_header(skb);
+
+ skb_set_network_header(skb,
+ -(x->props.header_len - x->props.enc_hdr_len));
+ skb->mac_header =
+ skb->network_header + offsetof(struct iphdr, protocol);
+ skb->transport_header = skb->network_header + sizeof(*top_iph);
+
+ top_iph = ip_hdr(skb);
+ top_iph->ihl = 5;
+ top_iph->version = 4;
+ top_iph->protocol = IPPROTO_AGGFRAG;
+
+ /* As we have 0, fractional, 1 or N inner packets there's no obviously
+ * correct DSCP mapping to inherit. ECN should be cleared per RFC9347
+ * 3.1.
+ */
+ top_iph->tos = 0;
+
+ top_iph->frag_off = htons(IP_DF);
+ top_iph->ttl = ip4_dst_hoplimit(xfrm_dst_child(dst));
+ top_iph->saddr = x->props.saddr.a4;
+ top_iph->daddr = x->id.daddr.a4;
+ ip_select_ident(dev_net(dst->dev), skb, NULL);
+
+ return 0;
+}
+
+/**
+ * iptfs_encap_add_ipv6() - add outer encaps
+ * @x: xfrm state
+ * @skb: the packet
+ *
+ * This was originally taken from xfrm6_tunnel_encap_add. The reason for the
+ * copy is that IP-TFS/AGGFRAG can have different functionality for how to set
+ * the flow label and TOS/DSCP bits. It also sets the protocol to a different
+ * value and doesn't do anything with inner headers as they aren't pointing into
+ * a normal IP singleton inner packet.
+ *
+ * Return: 0 on success or a negative error code on failure
+ */
+static int iptfs_encap_add_ipv6(struct xfrm_state *x, struct sk_buff *skb)
+{
+ struct dst_entry *dst = skb_dst(skb);
+ struct ipv6hdr *top_iph;
+ int dsfield;
+
+ skb_reset_inner_network_header(skb);
+ skb_reset_inner_transport_header(skb);
+
+ skb_set_network_header(skb,
+ -x->props.header_len + x->props.enc_hdr_len);
+ skb->mac_header =
+ skb->network_header + offsetof(struct ipv6hdr, nexthdr);
+ skb->transport_header = skb->network_header + sizeof(*top_iph);
+
+ top_iph = ipv6_hdr(skb);
+ top_iph->version = 6;
+ top_iph->priority = 0;
+ memset(top_iph->flow_lbl, 0, sizeof(top_iph->flow_lbl));
+ top_iph->nexthdr = IPPROTO_AGGFRAG;
+
+ /* As we have 0, fractional, 1 or N inner packets there's no obviously
+ * correct DSCP mapping to inherit. ECN should be cleared per RFC9347
+ * 3.1.
+ */
+ dsfield = 0;
+ ipv6_change_dsfield(top_iph, 0, dsfield);
+
+ top_iph->hop_limit = ip6_dst_hoplimit(xfrm_dst_child(dst));
+ top_iph->saddr = *(struct in6_addr *)&x->props.saddr;
+ top_iph->daddr = *(struct in6_addr *)&x->id.daddr;
+
+ return 0;
+}
+
+/**
+ * iptfs_prepare_output() - prepare the skb for output
+ * @x: xfrm state
+ * @skb: the packet
+ *
+ * Return: Error value, if 0 then skb values should be as follows:
+ * - transport_header should point at ESP header
+ * - network_header should point at Outer IP header
+ * - mac_header should point at protocol/nexthdr of the outer IP
+ */
+static int iptfs_prepare_output(struct xfrm_state *x, struct sk_buff *skb)
+{
+ if (x->outer_mode.family == AF_INET)
+ return iptfs_encap_add_ipv4(x, skb);
+ if (x->outer_mode.family == AF_INET6) {
+#if IS_ENABLED(CONFIG_IPV6)
+ return iptfs_encap_add_ipv6(x, skb);
+#else
+ WARN_ON_ONCE(1);
+ return -EAFNOSUPPORT;
+#endif
+ }
+ WARN_ON_ONCE(1);
+ return -EOPNOTSUPP;
+}
+
/* ========================== */
/* State Management Functions */
/* ========================== */
@@ -77,8 +601,11 @@ static int iptfs_user_init(struct net *net, struct xfrm_state *x,
{
struct xfrm_iptfs_data *xtfs = x->mode_data;
struct xfrm_iptfs_config *xc;
+ u64 q;
xc = &xtfs->cfg;
+ xc->max_queue_size = IPTFS_DEFAULT_MAX_QUEUE_SIZE;
+ xtfs->init_delay_ns = IPTFS_DEFAULT_INIT_DELAY_USECS * NSECS_IN_USEC;
if (attrs[XFRMA_IPTFS_PKT_SIZE]) {
xc->pkt_size = nla_get_u32(attrs[XFRMA_IPTFS_PKT_SIZE]);
@@ -92,6 +619,17 @@ static int iptfs_user_init(struct net *net, struct xfrm_state *x,
return -EINVAL;
}
}
+ if (attrs[XFRMA_IPTFS_MAX_QSIZE])
+ xc->max_queue_size = nla_get_u32(attrs[XFRMA_IPTFS_MAX_QSIZE]);
+ if (attrs[XFRMA_IPTFS_INIT_DELAY])
+ xtfs->init_delay_ns =
+ (u64)nla_get_u32(attrs[XFRMA_IPTFS_INIT_DELAY]) *
+ NSECS_IN_USEC;
+
+ q = (u64)xc->max_queue_size * 95;
+ (void)do_div(q, 100);
+ xtfs->ecn_queue_size = (u32)q;
+
return 0;
}
@@ -101,8 +639,11 @@ static unsigned int iptfs_sa_len(const struct xfrm_state *x)
struct xfrm_iptfs_config *xc = &xtfs->cfg;
unsigned int l = 0;
- if (x->dir == XFRM_SA_DIR_OUT)
+ if (x->dir == XFRM_SA_DIR_OUT) {
+ l += nla_total_size(sizeof(u32)); /* init delay usec */
+ l += nla_total_size(sizeof(xc->max_queue_size));
l += nla_total_size(sizeof(xc->pkt_size));
+ }
return l;
}
@@ -112,9 +653,22 @@ static int iptfs_copy_to_user(struct xfrm_state *x, struct sk_buff *skb)
struct xfrm_iptfs_data *xtfs = x->mode_data;
struct xfrm_iptfs_config *xc = &xtfs->cfg;
int ret = 0;
+ u64 q;
+
+ if (x->dir == XFRM_SA_DIR_OUT) {
+ q = xtfs->init_delay_ns;
+ (void)do_div(q, NSECS_IN_USEC);
+ ret = nla_put_u32(skb, XFRMA_IPTFS_INIT_DELAY, q);
+ if (ret)
+ return ret;
+
+ ret = nla_put_u32(skb, XFRMA_IPTFS_MAX_QSIZE,
+ xc->max_queue_size);
+ if (ret)
+ return ret;
- if (x->dir == XFRM_SA_DIR_OUT)
ret = nla_put_u32(skb, XFRMA_IPTFS_PKT_SIZE, xc->pkt_size);
+ }
return ret;
}
@@ -122,6 +676,10 @@ static int iptfs_copy_to_user(struct xfrm_state *x, struct sk_buff *skb)
static void __iptfs_init_state(struct xfrm_state *x,
struct xfrm_iptfs_data *xtfs)
{
+ __skb_queue_head_init(&xtfs->queue);
+ hrtimer_init(&xtfs->iptfs_timer, CLOCK_MONOTONIC, IPTFS_HRTIMER_MODE);
+ xtfs->iptfs_timer.function = iptfs_delay_timer;
+
/* Modify type (esp) adjustment values */
if (x->props.family == AF_INET)
@@ -166,10 +724,21 @@ static int iptfs_create_state(struct xfrm_state *x)
static void iptfs_delete_state(struct xfrm_state *x)
{
struct xfrm_iptfs_data *xtfs = x->mode_data;
+ struct sk_buff_head list;
+ struct sk_buff *skb;
if (!xtfs)
return;
+ spin_lock_bh(&xtfs->x->lock);
+ hrtimer_cancel(&xtfs->iptfs_timer);
+ __skb_queue_head_init(&list);
+ skb_queue_splice_init(&xtfs->queue, &list);
+ spin_unlock_bh(&xtfs->x->lock);
+
+ while ((skb = __skb_dequeue(&list)))
+ kfree_skb(skb);
+
kfree_sensitive(xtfs);
module_put(x->mode_cbs->owner);
@@ -184,6 +753,8 @@ static const struct xfrm_mode_cbs iptfs_mode_cbs = {
.sa_len = iptfs_sa_len,
.clone = iptfs_clone,
.get_inner_mtu = iptfs_get_inner_mtu,
+ .output = iptfs_output_collect,
+ .prepare_output = iptfs_prepare_output,
};
static int __init xfrm_iptfs_init(void)
--
2.46.0
^ permalink raw reply related [flat|nested] 40+ messages in thread* Re: [PATCH ipsec-next v8 08/16] xfrm: iptfs: add user packet (tunnel ingress) handling
2024-08-04 20:33 ` [PATCH ipsec-next v8 08/16] xfrm: iptfs: add user packet (tunnel ingress) handling Christian Hopps
@ 2024-08-05 17:10 ` Simon Horman
2024-08-06 10:19 ` [devel-ipsec] " Christian Hopps
0 siblings, 1 reply; 40+ messages in thread
From: Simon Horman @ 2024-08-05 17:10 UTC (permalink / raw)
To: Christian Hopps; +Cc: devel, Steffen Klassert, netdev, Christian Hopps
On Sun, Aug 04, 2024 at 04:33:37PM -0400, Christian Hopps wrote:
> From: Christian Hopps <chopps@labn.net>
>
> Add tunnel packet output functionality. This is code handles
> the ingress to the tunnel.
>
> Signed-off-by: Christian Hopps <chopps@labn.net>
...
> +/**
> + * iptfs_prepare_output() - prepare the skb for output
> + * @x: xfrm state
> + * @skb: the packet
> + *
> + * Return: Error value, if 0 then skb values should be as follows:
> + * - transport_header should point at ESP header
> + * - network_header should point at Outer IP header
> + * - mac_header should point at protocol/nexthdr of the outer IP
> + */
> +static int iptfs_prepare_output(struct xfrm_state *x, struct sk_buff *skb)
> +{
> + if (x->outer_mode.family == AF_INET)
> + return iptfs_encap_add_ipv4(x, skb);
> + if (x->outer_mode.family == AF_INET6) {
> +#if IS_ENABLED(CONFIG_IPV6)
> + return iptfs_encap_add_ipv6(x, skb);
iptfs_encap_add_ipv6 is flagged as unused when IPV6 is not enabled.
Perhaps it should also be wrapped in a CONFIG_IPV6 check.
> +#else
> + WARN_ON_ONCE(1);
> + return -EAFNOSUPPORT;
> +#endif
> + }
> + WARN_ON_ONCE(1);
> + return -EOPNOTSUPP;
> +}
> +
...
^ permalink raw reply [flat|nested] 40+ messages in thread* Re: [devel-ipsec] [PATCH ipsec-next v8 08/16] xfrm: iptfs: add user packet (tunnel ingress) handling
2024-08-05 17:10 ` Simon Horman
@ 2024-08-06 10:19 ` Christian Hopps
2024-08-06 15:24 ` Simon Horman
0 siblings, 1 reply; 40+ messages in thread
From: Christian Hopps @ 2024-08-06 10:19 UTC (permalink / raw)
To: Simon Horman
Cc: Christian Hopps, Steffen Klassert, netdev, Christian Hopps, devel
Simon Horman via Devel <devel@linux-ipsec.org> writes:
> On Sun, Aug 04, 2024 at 04:33:37PM -0400, Christian Hopps wrote:
>> From: Christian Hopps <chopps@labn.net>
>>
>> Add tunnel packet output functionality. This is code handles
>> the ingress to the tunnel.
>>
>> Signed-off-by: Christian Hopps <chopps@labn.net>
>
> ...
>
>> +static int iptfs_prepare_output(struct xfrm_state *x, struct sk_buff *skb)
>> +{
>> + if (x->outer_mode.family == AF_INET)
>> + return iptfs_encap_add_ipv4(x, skb);
>> + if (x->outer_mode.family == AF_INET6) {
>> +#if IS_ENABLED(CONFIG_IPV6)
>> + return iptfs_encap_add_ipv6(x, skb);
>
> iptfs_encap_add_ipv6 is flagged as unused when IPV6 is not enabled.
> Perhaps it should also be wrapped in a CONFIG_IPV6 check.
Done, and tested with CONFIG_IPV6=n.
Thanks,
Chris.
^ permalink raw reply [flat|nested] 40+ messages in thread* Re: [devel-ipsec] [PATCH ipsec-next v8 08/16] xfrm: iptfs: add user packet (tunnel ingress) handling
2024-08-06 10:19 ` [devel-ipsec] " Christian Hopps
@ 2024-08-06 15:24 ` Simon Horman
0 siblings, 0 replies; 40+ messages in thread
From: Simon Horman @ 2024-08-06 15:24 UTC (permalink / raw)
To: Christian Hopps; +Cc: Steffen Klassert, netdev, Christian Hopps, devel
On Tue, Aug 06, 2024 at 06:19:28AM -0400, Christian Hopps wrote:
>
> Simon Horman via Devel <devel@linux-ipsec.org> writes:
>
> > On Sun, Aug 04, 2024 at 04:33:37PM -0400, Christian Hopps wrote:
> > > From: Christian Hopps <chopps@labn.net>
> > >
> > > Add tunnel packet output functionality. This is code handles
> > > the ingress to the tunnel.
> > >
> > > Signed-off-by: Christian Hopps <chopps@labn.net>
> >
> > ...
> >
> > > +static int iptfs_prepare_output(struct xfrm_state *x, struct sk_buff *skb)
> > > +{
> > > + if (x->outer_mode.family == AF_INET)
> > > + return iptfs_encap_add_ipv4(x, skb);
> > > + if (x->outer_mode.family == AF_INET6) {
> > > +#if IS_ENABLED(CONFIG_IPV6)
> > > + return iptfs_encap_add_ipv6(x, skb);
> >
> > iptfs_encap_add_ipv6 is flagged as unused when IPV6 is not enabled.
> > Perhaps it should also be wrapped in a CONFIG_IPV6 check.
>
> Done, and tested with CONFIG_IPV6=n.
Thanks!
^ permalink raw reply [flat|nested] 40+ messages in thread
* [PATCH ipsec-next v8 09/16] xfrm: iptfs: share page fragments of inner packets
2024-08-04 20:33 [PATCH ipsec-next v8 00/16] Add IP-TFS mode to xfrm Christian Hopps
` (7 preceding siblings ...)
2024-08-04 20:33 ` [PATCH ipsec-next v8 08/16] xfrm: iptfs: add user packet (tunnel ingress) handling Christian Hopps
@ 2024-08-04 20:33 ` Christian Hopps
2024-08-04 20:33 ` [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets Christian Hopps
` (6 subsequent siblings)
15 siblings, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-04 20:33 UTC (permalink / raw)
To: devel; +Cc: Steffen Klassert, netdev, Christian Hopps, Christian Hopps
From: Christian Hopps <chopps@labn.net>
When possible rather than appending secondary (aggregated) inner packets
to the fragment list, share their page fragments with the outer IPTFS
packet. This allows for more efficient packet transmission.
Signed-off-by: Christian Hopps <chopps@labn.net>
---
net/xfrm/xfrm_iptfs.c | 88 ++++++++++++++++++++++++++++++++++++++-----
1 file changed, 79 insertions(+), 9 deletions(-)
diff --git a/net/xfrm/xfrm_iptfs.c b/net/xfrm/xfrm_iptfs.c
index 9c48c15cbed0..20c19894720e 100644
--- a/net/xfrm/xfrm_iptfs.c
+++ b/net/xfrm/xfrm_iptfs.c
@@ -9,6 +9,7 @@
#include <linux/kernel.h>
#include <linux/icmpv6.h>
+#include <linux/skbuff_ref.h>
#include <net/gro.h>
#include <net/icmp.h>
#include <net/ip6_route.h>
@@ -90,6 +91,24 @@ struct xfrm_iptfs_data {
static u32 iptfs_get_inner_mtu(struct xfrm_state *x, int outer_mtu);
static enum hrtimer_restart iptfs_delay_timer(struct hrtimer *me);
+/* ================= */
+/* SK_BUFF Functions */
+/* ================= */
+
+/**
+ * skb_head_to_frag() - initialize a skb_frag_t based on skb head data
+ * @skb: skb with the head data
+ * @frag: frag to initialize
+ */
+static void skb_head_to_frag(const struct sk_buff *skb, skb_frag_t *frag)
+{
+ struct page *page = virt_to_head_page(skb->data);
+ unsigned char *addr = (unsigned char *)page_address(page);
+
+ BUG_ON(!skb->head_frag);
+ skb_frag_fill_page_desc(frag, page, skb->data - addr, skb_headlen(skb));
+}
+
/* ================================= */
/* IPTFS Sending (ingress) Functions */
/* ================================= */
@@ -306,14 +325,44 @@ static struct sk_buff **iptfs_rehome_fraglist(struct sk_buff **nextp,
return nextp;
}
+static void iptfs_consume_frags(struct sk_buff *to, struct sk_buff *from)
+{
+ struct skb_shared_info *fromi = skb_shinfo(from);
+ struct skb_shared_info *toi = skb_shinfo(to);
+ unsigned int new_truesize;
+
+ /* If we have data in a head page, grab it */
+ if (!skb_headlen(from)) {
+ new_truesize = SKB_TRUESIZE(skb_end_offset(from));
+ } else {
+ skb_head_to_frag(from, &toi->frags[toi->nr_frags]);
+ skb_frag_ref(to, toi->nr_frags++);
+ new_truesize = SKB_DATA_ALIGN(sizeof(struct sk_buff));
+ }
+
+ /* Move any other page fragments rather than copy */
+ memcpy(&toi->frags[toi->nr_frags], fromi->frags,
+ sizeof(fromi->frags[0]) * fromi->nr_frags);
+ toi->nr_frags += fromi->nr_frags;
+ fromi->nr_frags = 0;
+ from->data_len = 0;
+ from->len = 0;
+ to->truesize += from->truesize - new_truesize;
+ from->truesize = new_truesize;
+
+ /* We are done with this SKB */
+ consume_skb(from);
+}
+
static void iptfs_output_queued(struct xfrm_state *x, struct sk_buff_head *list)
{
struct xfrm_iptfs_data *xtfs = x->mode_data;
struct sk_buff *skb, *skb2, **nextp;
- struct skb_shared_info *shi;
+ struct skb_shared_info *shi, *shi2;
while ((skb = __skb_dequeue(list))) {
u32 mtu = iptfs_get_cur_pmtu(x, xtfs, skb);
+ bool share_ok = true;
int remaining;
/* protocol comes to us cleared sometimes */
@@ -357,7 +406,7 @@ static void iptfs_output_queued(struct xfrm_state *x, struct sk_buff_head *list)
/* Re-home (un-nest) nested fragment lists. We need to do this
* b/c we will simply be appending any following aggregated
- * inner packets to the frag list.
+ * inner packets using the frag list.
*/
shi = skb_shinfo(skb);
nextp = &shi->frag_list;
@@ -369,6 +418,9 @@ static void iptfs_output_queued(struct xfrm_state *x, struct sk_buff_head *list)
nextp = &(*nextp)->next;
}
+ if (shi->frag_list || skb_cloned(skb) || skb_shared(skb))
+ share_ok = false;
+
/* See if we have enough space to simply append.
*
* NOTE: Maybe do not append if we will be mis-aligned,
@@ -393,18 +445,36 @@ static void iptfs_output_queued(struct xfrm_state *x, struct sk_buff_head *list)
}
}
+ /* skb->pp_recycle is passed to __skb_flag_unref for all
+ * frag pages so we can only share pages with skb's who
+ * match ourselves.
+ */
+ shi2 = skb_shinfo(skb2);
+ if (share_ok &&
+ (shi2->frag_list ||
+ (!skb2->head_frag && skb_headlen(skb)) ||
+ skb->pp_recycle != skb2->pp_recycle ||
+ skb_zcopy(skb2) ||
+ (shi->nr_frags + shi2->nr_frags + 1 > MAX_SKB_FRAGS)))
+ share_ok = false;
+
/* Do accounting */
skb->data_len += skb2->len;
skb->len += skb2->len;
remaining -= skb2->len;
- /* Append to the frag_list */
- *nextp = skb2;
- nextp = &skb2->next;
- BUG_ON(*nextp);
- if (skb_has_frag_list(skb2))
- nextp = iptfs_rehome_fraglist(nextp, skb2);
- skb->truesize += skb2->truesize;
+ if (share_ok) {
+ iptfs_consume_frags(skb, skb2);
+ } else {
+ /* Append to the frag_list */
+ *nextp = skb2;
+ nextp = &skb2->next;
+ BUG_ON(*nextp);
+ if (skb_has_frag_list(skb2))
+ nextp = iptfs_rehome_fraglist(nextp,
+ skb2);
+ skb->truesize += skb2->truesize;
+ }
}
xfrm_output(NULL, skb);
--
2.46.0
^ permalink raw reply related [flat|nested] 40+ messages in thread* [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-04 20:33 [PATCH ipsec-next v8 00/16] Add IP-TFS mode to xfrm Christian Hopps
` (8 preceding siblings ...)
2024-08-04 20:33 ` [PATCH ipsec-next v8 09/16] xfrm: iptfs: share page fragments of inner packets Christian Hopps
@ 2024-08-04 20:33 ` Christian Hopps
2024-08-04 22:25 ` Sabrina Dubroca
2024-08-04 20:33 ` [PATCH ipsec-next v8 11/16] xfrm: iptfs: add basic receive packet (tunnel egress) handling Christian Hopps
` (5 subsequent siblings)
15 siblings, 1 reply; 40+ messages in thread
From: Christian Hopps @ 2024-08-04 20:33 UTC (permalink / raw)
To: devel; +Cc: Steffen Klassert, netdev, Christian Hopps, Christian Hopps
From: Christian Hopps <chopps@labn.net>
Add support for tunneling user (inner) packets that are larger than the
tunnel's path MTU (outer) using IP-TFS fragmentation.
Signed-off-by: Christian Hopps <chopps@labn.net>
---
net/xfrm/xfrm_iptfs.c | 407 +++++++++++++++++++++++++++++++++++++++---
1 file changed, 381 insertions(+), 26 deletions(-)
diff --git a/net/xfrm/xfrm_iptfs.c b/net/xfrm/xfrm_iptfs.c
index 20c19894720e..38735e2d64c3 100644
--- a/net/xfrm/xfrm_iptfs.c
+++ b/net/xfrm/xfrm_iptfs.c
@@ -46,12 +46,23 @@
*/
#define IPTFS_DEFAULT_MAX_QUEUE_SIZE (1024 * 10240)
+/* 1) skb->head should be cache aligned.
+ * 2) when resv is for L2 headers (i.e., ethernet) we want the cacheline to
+ * start -16 from data.
+ * 3) when resv is for L3+L2 headers IOW skb->data points at the IPTFS payload
+ * we want data to be cache line aligned so all the pushed headers will be in
+ * another cacheline.
+ */
+#define XFRM_IPTFS_MIN_L3HEADROOM 128
+#define XFRM_IPTFS_MIN_L2HEADROOM (64 + 16)
+#define IPTFS_FRAG_COPY_MAX 256 /* max for copying to create iptfs frags */
#define NSECS_IN_USEC 1000
#define IPTFS_HRTIMER_MODE HRTIMER_MODE_REL_SOFT
/**
* struct xfrm_iptfs_config - configuration for the IPTFS tunnel.
+ * @dont_frag: true to inhibit fragmenting across IPTFS outer packets.
* @pkt_size: size of the outer IP packet. 0 to use interface and MTU discovery,
* otherwise the user specified value.
* @max_queue_size: The maximum number of octets allowed to be queued to be sent
@@ -59,6 +70,7 @@
* packets enqueued.
*/
struct xfrm_iptfs_config {
+ bool dont_frag : 1;
u32 pkt_size; /* outer_packet_size or 0 */
u32 max_queue_size; /* octets */
};
@@ -88,13 +100,71 @@ struct xfrm_iptfs_data {
u32 payload_mtu; /* max payload size */
};
-static u32 iptfs_get_inner_mtu(struct xfrm_state *x, int outer_mtu);
+static u32 __iptfs_get_inner_mtu(struct xfrm_state *x, int outer_mtu);
static enum hrtimer_restart iptfs_delay_timer(struct hrtimer *me);
/* ================= */
/* SK_BUFF Functions */
/* ================= */
+/**
+ * iptfs_alloc_skb() - Allocate a new `skb` using a meta-data template.
+ * @tpl: the template to copy the new `skb`s meta-data from.
+ * @len: the linear length of the head data, zero is fine.
+ * @l3resv: true if reserve needs to support pushing L3 headers
+ *
+ * A new `skb` is allocated and it's meta-data is initialized from `tpl`, the
+ * head data is sized to `len` + reserved space set according to the @l3resv
+ * boolean. When @l3resv is false, resv is XFRM_IPTFS_MIN_L2HEADROOM which
+ * arranges for `skb->data - 16` (etherhdr space) to be the start of a cacheline.
+ * Otherwise, @l3resv is true and resv is either the size of headroom from `tpl` or
+ * XFRM_IPTFS_MIN_L3HEADROOM whichever is greater, which tries to align
+ * skb->data to a cacheline as all headers will be pushed on the previous
+ * cacheline bytes.
+ *
+ * When copying meta-data from the @tpl, the sk_buff->headers are not copied.
+ *
+ * Zero length skbs are allocated when we only need a head skb to hold new
+ * packet headers (basically the mac header) that sit on top of existing shared
+ * packet data.
+ *
+ * Return: the new skb or NULL.
+ */
+static struct sk_buff *iptfs_alloc_skb(struct sk_buff *tpl, u32 len,
+ bool l3resv)
+{
+ struct sk_buff *skb;
+ u32 resv;
+
+ if (!l3resv) {
+ resv = XFRM_IPTFS_MIN_L2HEADROOM;
+ } else {
+ resv = skb_headroom(tpl);
+ if (resv < XFRM_IPTFS_MIN_L3HEADROOM)
+ resv = XFRM_IPTFS_MIN_L3HEADROOM;
+ }
+
+ skb = alloc_skb(len + resv, GFP_ATOMIC);
+ if (!skb) {
+ XFRM_INC_STATS(dev_net(tpl->dev), LINUX_MIB_XFRMNOSKBERROR);
+ return NULL;
+ }
+
+ skb_reserve(skb, resv);
+
+ /* We do not want any of the tpl->headers copied over, so we do
+ * not use `skb_copy_header()`.
+ */
+ skb->tstamp = tpl->tstamp;
+ skb->dev = tpl->dev;
+ memcpy(skb->cb, tpl->cb, sizeof(skb->cb));
+ skb_dst_copy(skb, tpl);
+ __skb_ext_copy(skb, tpl);
+ __nf_copy(skb, tpl, false);
+
+ return skb;
+}
+
/**
* skb_head_to_frag() - initialize a skb_frag_t based on skb head data
* @skb: skb with the head data
@@ -109,6 +179,41 @@ static void skb_head_to_frag(const struct sk_buff *skb, skb_frag_t *frag)
skb_frag_fill_page_desc(frag, page, skb->data - addr, skb_headlen(skb));
}
+/**
+ * skb_copy_bits_seq - copy bits from a skb_seq_state to kernel buffer
+ * @st: source skb_seq_state
+ * @offset: offset in source
+ * @to: destination buffer
+ * @len: number of bytes to copy
+ *
+ * Copy @len bytes from @offset bytes into the source @st to the destination
+ * buffer @to. `offset` should increase (or be unchanged) with each subsequent
+ * call to this function. If offset needs to decrease from the previous use `st`
+ * should be reset first.
+ *
+ * Return: 0 on success or a negative error code on failure
+ */
+static int skb_copy_bits_seq(struct skb_seq_state *st, int offset, void *to,
+ int len)
+{
+ const u8 *data;
+ u32 sqlen;
+
+ for (;;) {
+ sqlen = skb_seq_read(offset, &data, st);
+ if (sqlen == 0)
+ return -ENOMEM;
+ if (sqlen >= len) {
+ memcpy(to, data, len);
+ return 0;
+ }
+ memcpy(to, data, sqlen);
+ to += sqlen;
+ offset += sqlen;
+ len -= sqlen;
+ }
+}
+
/* ================================= */
/* IPTFS Sending (ingress) Functions */
/* ================================= */
@@ -153,7 +258,7 @@ static int iptfs_get_cur_pmtu(struct xfrm_state *x,
{
struct xfrm_dst *xdst = (struct xfrm_dst *)skb_dst(skb);
u32 payload_mtu = xtfs->payload_mtu;
- u32 pmtu = iptfs_get_inner_mtu(x, xdst->child_mtu_cached);
+ u32 pmtu = __iptfs_get_inner_mtu(x, xdst->child_mtu_cached);
if (payload_mtu && payload_mtu < pmtu)
pmtu = payload_mtu;
@@ -216,7 +321,8 @@ static int iptfs_output_collect(struct net *net, struct sock *sk,
BUG_ON(!xtfs);
- pmtu = iptfs_get_cur_pmtu(x, xtfs, skb);
+ if (xtfs->cfg.dont_frag)
+ pmtu = iptfs_get_cur_pmtu(x, xtfs, skb);
/* Break apart GSO skbs. If the queue is nearing full then we want the
* accounting and queuing to be based on the individual packets not on the
@@ -256,8 +362,10 @@ static int iptfs_output_collect(struct net *net, struct sock *sk,
continue;
}
- /* Fragmenting handled in following commits. */
- if (iptfs_is_too_big(sk, skb, pmtu)) {
+ /* If the user indicated no iptfs fragmenting check before
+ * enqueue.
+ */
+ if (xtfs->cfg.dont_frag && iptfs_is_too_big(sk, skb, pmtu)) {
kfree_skb_reason(skb, SKB_DROP_REASON_PKT_TOO_BIG);
continue;
}
@@ -301,6 +409,219 @@ static void iptfs_output_prepare_skb(struct sk_buff *skb, u32 blkoff)
IPCB(skb)->flags |= IPSKB_XFRM_TUNNEL_SIZE;
}
+/**
+ * iptfs_copy_create_frag() - create an inner fragment skb.
+ * @st: The source packet data.
+ * @offset: offset in @st of the new fragment data.
+ * @copy_len: the amount of data to copy from @st.
+ *
+ * Create a new skb holding a single IPTFS inner packet fragment. @copy_len must
+ * not be greater than the max fragment size.
+ *
+ * Return: the new fragment skb or an ERR_PTR().
+ */
+static struct sk_buff *iptfs_copy_create_frag(struct skb_seq_state *st,
+ u32 offset, u32 copy_len)
+{
+ struct sk_buff *src = st->root_skb;
+ struct sk_buff *skb;
+ int err;
+
+ skb = iptfs_alloc_skb(src, copy_len, true);
+ if (!skb)
+ return ERR_PTR(-ENOMEM);
+
+ /* Now copy `copy_len` data from src */
+ err = skb_copy_bits_seq(st, offset, skb_put(skb, copy_len), copy_len);
+ if (err) {
+ kfree_skb(skb);
+ return ERR_PTR(err);
+ }
+
+ return skb;
+}
+
+/**
+ * iptfs_copy_create_frags() - create and send N-1 fragments of a larger skb.
+ * @skbp: the source packet skb (IN), skb holding the last fragment in
+ * the fragment stream (OUT).
+ * @xtfs: IPTFS SA state.
+ * @mtu: the max IPTFS fragment size.
+ *
+ * This function is responsible for fragmenting a larger inner packet into a
+ * sequence of IPTFS payload packets. The last fragment is returned rather than
+ * being sent so that the caller can append more inner packets (aggregation) if
+ * there is room.
+ *
+ * Return: 0 on success or a negative error code on failure
+ */
+static int iptfs_copy_create_frags(struct sk_buff **skbp,
+ struct xfrm_iptfs_data *xtfs, u32 mtu)
+{
+ struct skb_seq_state skbseq;
+ struct list_head sublist;
+ struct sk_buff *skb = *skbp;
+ struct sk_buff *nskb = *skbp;
+ u32 copy_len, offset;
+ u32 to_copy = skb->len - mtu;
+ int err = 0;
+
+ INIT_LIST_HEAD(&sublist);
+
+ BUG_ON(skb->len <= mtu);
+ skb_prepare_seq_read(skb, 0, skb->len, &skbseq);
+
+ /* A trimmed `skb` will be sent as the first fragment, later. */
+ offset = mtu;
+ to_copy = skb->len - offset;
+ while (to_copy) {
+ /* Send all but last fragment to allow agg. append */
+ list_add_tail(&nskb->list, &sublist);
+
+ /* FUTURE: if the packet has an odd/non-aligning length we could
+ * send less data in the penultimate fragment so that the last
+ * fragment then ends on an aligned boundary.
+ */
+ copy_len = min(to_copy, mtu);
+ nskb = iptfs_copy_create_frag(&skbseq, offset, copy_len);
+ if (IS_ERR(nskb)) {
+ XFRM_INC_STATS(xs_net(xtfs->x),
+ LINUX_MIB_XFRMOUTERROR);
+ skb_abort_seq_read(&skbseq);
+ err = PTR_ERR(nskb);
+ nskb = NULL;
+ break;
+ }
+ iptfs_output_prepare_skb(nskb, to_copy);
+ offset += copy_len;
+ to_copy -= copy_len;
+ }
+ skb_abort_seq_read(&skbseq);
+
+ /* return last fragment that will be unsent (or NULL) */
+ *skbp = nskb;
+
+ /* trim the original skb to MTU */
+ if (!err)
+ err = pskb_trim(skb, mtu);
+
+ if (err) {
+ /* Free all frags. Don't bother sending a partial packet we will
+ * never complete.
+ */
+ kfree_skb(nskb);
+ list_for_each_entry_safe(skb, nskb, &sublist, list) {
+ skb_list_del_init(skb);
+ kfree_skb(skb);
+ }
+ return err;
+ }
+
+ /* prepare the initial fragment with an iptfs header */
+ iptfs_output_prepare_skb(skb, 0);
+
+ /* Send all but last fragment, if we fail to send a fragment then free
+ * the rest -- no point in sending a packet that can't be reassembled.
+ */
+ list_for_each_entry_safe(skb, nskb, &sublist, list) {
+ skb_list_del_init(skb);
+ if (!err)
+ err = xfrm_output(NULL, skb);
+ else
+ kfree_skb(skb);
+ }
+ if (err)
+ kfree_skb(*skbp);
+ return err;
+}
+
+/**
+ * iptfs_first_should_copy() - determine if we should copy packet data.
+ * @first_skb: the first skb in the packet
+ * @mtu: the MTU.
+ *
+ * Determine if we should create subsequent skbs to hold the remaining data from
+ * a large inner packet by copying the packet data, or cloning the original skb
+ * and adjusting the offsets.
+ *
+ * Return: true if we should copy the data out of the skb.
+ */
+static bool iptfs_first_should_copy(struct sk_buff *first_skb, u32 mtu)
+{
+ u32 frag_copy_max;
+
+ /* If we have less than frag_copy_max for remaining packet we copy
+ * those tail bytes as it is more efficient.
+ */
+ frag_copy_max = min(mtu, IPTFS_FRAG_COPY_MAX);
+ if ((int)first_skb->len - (int)mtu < (int)frag_copy_max)
+ return true;
+
+ /* If we have non-linear skb just use copy */
+ if (skb_is_nonlinear(first_skb))
+ return true;
+
+ /* So we have a simple linear skb, easy to clone and share */
+ return false;
+}
+
+/**
+ * iptfs_first_skb() - handle the first dequeued inner packet for output
+ * @skbp: the source packet skb (IN), skb holding the last fragment in
+ * the fragment stream (OUT).
+ * @xtfs: IPTFS SA state.
+ * @mtu: the max IPTFS fragment size.
+ *
+ * This function is responsible for fragmenting a larger inner packet into a
+ * sequence of IPTFS payload packets. If it needs to fragment into subsequent
+ * skb's, it will either do so by copying or cloning.
+ *
+ * The last fragment is returned rather than being sent so that the caller can
+ * append more inner packets (aggregation) if there is room.
+ *
+ * Return: 0 on success or a negative error code on failure
+ */
+static int iptfs_first_skb(struct sk_buff **skbp, struct xfrm_iptfs_data *xtfs,
+ u32 mtu)
+{
+ struct sk_buff *skb = *skbp;
+ int err;
+
+ /* Classic ESP skips the don't fragment ICMP error if DF is clear on
+ * the inner packet or ignore_df is set. Otherwise it will send an ICMP
+ * or local error if the inner packet won't fit it's MTU.
+ *
+ * With IPTFS we do not care about the inner packet DF bit. If the
+ * tunnel is configured to "don't fragment" we error back if things
+ * don't fit in our max packet size. Otherwise we iptfs-fragment as
+ * normal.
+ */
+
+ /* The opportunity for HW offload has ended */
+ if (skb->ip_summed == CHECKSUM_PARTIAL) {
+ err = skb_checksum_help(skb);
+ if (err)
+ return err;
+ }
+
+ /* We've split these up before queuing */
+ BUG_ON(skb_is_gso(skb));
+
+ /* Simple case -- it fits. `mtu` accounted for all the overhead
+ * including the basic IPTFS header.
+ */
+ if (skb->len <= mtu) {
+ iptfs_output_prepare_skb(skb, 0);
+ return 0;
+ }
+
+ if (iptfs_first_should_copy(skb, mtu))
+ return iptfs_copy_create_frags(skbp, xtfs, mtu);
+
+ /* For now we always copy */
+ return iptfs_copy_create_frags(skbp, xtfs, mtu);
+}
+
static struct sk_buff **iptfs_rehome_fraglist(struct sk_buff **nextp,
struct sk_buff *child)
{
@@ -360,6 +681,15 @@ static void iptfs_output_queued(struct xfrm_state *x, struct sk_buff_head *list)
struct sk_buff *skb, *skb2, **nextp;
struct skb_shared_info *shi, *shi2;
+ /* If we are fragmenting due to a large inner packet we will output all
+ * the outer IPTFS packets required to contain the fragments of the
+ * single large inner packet. These outer packets need to be sent
+ * consecutively (ESP seq-wise). Since this output function is always
+ * running from a timer we do not need a lock to provide this guarantee.
+ * We will output our packets consecutively before the timer is allowed
+ * to run again on some other CPU.
+ */
+
while ((skb = __skb_dequeue(list))) {
u32 mtu = iptfs_get_cur_pmtu(x, xtfs, skb);
bool share_ok = true;
@@ -370,7 +700,7 @@ static void iptfs_output_queued(struct xfrm_state *x, struct sk_buff_head *list)
htons(ETH_P_IP) :
htons(ETH_P_IPV6);
- if (skb->len > mtu) {
+ if (skb->len > mtu && xtfs->cfg.dont_frag) {
/* We handle this case before enqueueing so we are only
* here b/c MTU changed after we enqueued before we
* dequeued, just drop these.
@@ -381,26 +711,22 @@ static void iptfs_output_queued(struct xfrm_state *x, struct sk_buff_head *list)
continue;
}
- /* If we don't have a cksum in the packet we need to add one
- * before encapsulation.
+ /* Convert first inner packet into an outer IPTFS packet,
+ * dealing with any fragmentation into multiple outer packets
+ * if necessary.
*/
- if (skb->ip_summed == CHECKSUM_PARTIAL) {
- if (skb_checksum_help(skb)) {
- XFRM_INC_STATS(dev_net(skb_dst(skb)->dev),
- LINUX_MIB_XFRMOUTERROR);
- kfree_skb(skb);
- continue;
- }
- }
-
- /* Convert first inner packet into an outer IPTFS packet */
- iptfs_output_prepare_skb(skb, 0);
+ if (iptfs_first_skb(&skb, xtfs, mtu))
+ continue;
- /* The space remaining to send more inner packet data is `mtu` -
- * (skb->len - sizeof iptfs header). This is b/c the `mtu` value
- * has the basic IPTFS header len accounted for, and we added
- * that header to the skb so it is a part of skb->len, thus we
- * subtract it from the skb length.
+ /* If fragmentation was required the returned skb is the last
+ * IPTFS fragment in the chain, and it's IPTFS header blkoff has
+ * been set just past the end of the fragment data.
+ *
+ * In either case the space remaining to send more inner packet
+ * data is `mtu` - (skb->len - sizeof iptfs header). This is b/c
+ * the `mtu` value has the basic IPTFS header len accounted for,
+ * and we added that header to the skb so it is a part of
+ * skb->len, thus we subtract it from the skb length.
*/
remaining = mtu - (skb->len - sizeof(struct ip_iptfs_hdr));
@@ -641,11 +967,13 @@ static int iptfs_prepare_output(struct xfrm_state *x, struct sk_buff *skb)
/* ========================== */
/**
- * iptfs_get_inner_mtu() - return inner MTU with no fragmentation.
+ * __iptfs_get_inner_mtu() - return inner MTU with no fragmentation.
* @x: xfrm state.
* @outer_mtu: the outer mtu
+ *
+ * Return: Correct MTU taking in to account the encap overhead.
*/
-static u32 iptfs_get_inner_mtu(struct xfrm_state *x, int outer_mtu)
+static u32 __iptfs_get_inner_mtu(struct xfrm_state *x, int outer_mtu)
{
struct crypto_aead *aead;
u32 blksize;
@@ -656,6 +984,23 @@ static u32 iptfs_get_inner_mtu(struct xfrm_state *x, int outer_mtu)
~(blksize - 1)) - 2;
}
+/**
+ * iptfs_get_inner_mtu() - return the inner MTU for an IPTFS xfrm.
+ * @x: xfrm state.
+ * @outer_mtu: Outer MTU for the encapsulated packet.
+ *
+ * Return: Correct MTU taking in to account the encap overhead.
+ */
+static u32 iptfs_get_inner_mtu(struct xfrm_state *x, int outer_mtu)
+{
+ struct xfrm_iptfs_data *xtfs = x->mode_data;
+
+ /* If not dont-frag we have no MTU */
+ if (!xtfs->cfg.dont_frag)
+ return x->outer_mode.family == AF_INET ? IP_MAX_MTU : IP6_MAX_MTU;
+ return __iptfs_get_inner_mtu(x, outer_mtu);
+}
+
/**
* iptfs_user_init() - initialize the SA with IPTFS options from netlink.
* @net: the net data
@@ -677,6 +1022,8 @@ static int iptfs_user_init(struct net *net, struct xfrm_state *x,
xc->max_queue_size = IPTFS_DEFAULT_MAX_QUEUE_SIZE;
xtfs->init_delay_ns = IPTFS_DEFAULT_INIT_DELAY_USECS * NSECS_IN_USEC;
+ if (attrs[XFRMA_IPTFS_DONT_FRAG])
+ xc->dont_frag = true;
if (attrs[XFRMA_IPTFS_PKT_SIZE]) {
xc->pkt_size = nla_get_u32(attrs[XFRMA_IPTFS_PKT_SIZE]);
if (!xc->pkt_size) {
@@ -710,6 +1057,8 @@ static unsigned int iptfs_sa_len(const struct xfrm_state *x)
unsigned int l = 0;
if (x->dir == XFRM_SA_DIR_OUT) {
+ if (xc->dont_frag)
+ l += nla_total_size(0); /* dont-frag flag */
l += nla_total_size(sizeof(u32)); /* init delay usec */
l += nla_total_size(sizeof(xc->max_queue_size));
l += nla_total_size(sizeof(xc->pkt_size));
@@ -726,6 +1075,12 @@ static int iptfs_copy_to_user(struct xfrm_state *x, struct sk_buff *skb)
u64 q;
if (x->dir == XFRM_SA_DIR_OUT) {
+ if (xc->dont_frag) {
+ ret = nla_put_flag(skb, XFRMA_IPTFS_DONT_FRAG);
+ if (ret)
+ return ret;
+ }
+
q = xtfs->init_delay_ns;
(void)do_div(q, NSECS_IN_USEC);
ret = nla_put_u32(skb, XFRMA_IPTFS_INIT_DELAY, q);
--
2.46.0
^ permalink raw reply related [flat|nested] 40+ messages in thread* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-04 20:33 ` [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets Christian Hopps
@ 2024-08-04 22:25 ` Sabrina Dubroca
2024-08-05 2:33 ` Christian Hopps
2024-08-06 11:32 ` Steffen Klassert
0 siblings, 2 replies; 40+ messages in thread
From: Sabrina Dubroca @ 2024-08-04 22:25 UTC (permalink / raw)
To: Christian Hopps; +Cc: devel, Steffen Klassert, netdev, Christian Hopps
Please CC the reviewers from previous versions of the patchset. It's
really hard to keep track of discussions and reposts otherwise.
2024-08-04, 16:33:39 -0400, Christian Hopps wrote:
> From: Christian Hopps <chopps@labn.net>
>
> Add support for tunneling user (inner) packets that are larger than the
> tunnel's path MTU (outer) using IP-TFS fragmentation.
>
> Signed-off-by: Christian Hopps <chopps@labn.net>
> ---
> net/xfrm/xfrm_iptfs.c | 407 +++++++++++++++++++++++++++++++++++++++---
> 1 file changed, 381 insertions(+), 26 deletions(-)
>
> diff --git a/net/xfrm/xfrm_iptfs.c b/net/xfrm/xfrm_iptfs.c
> index 20c19894720e..38735e2d64c3 100644
> --- a/net/xfrm/xfrm_iptfs.c
> +++ b/net/xfrm/xfrm_iptfs.c
> @@ -46,12 +46,23 @@
> */
> #define IPTFS_DEFAULT_MAX_QUEUE_SIZE (1024 * 10240)
>
> +/* 1) skb->head should be cache aligned.
> + * 2) when resv is for L2 headers (i.e., ethernet) we want the cacheline to
> + * start -16 from data.
> + * 3) when resv is for L3+L2 headers IOW skb->data points at the IPTFS payload
> + * we want data to be cache line aligned so all the pushed headers will be in
> + * another cacheline.
> + */
> +#define XFRM_IPTFS_MIN_L3HEADROOM 128
> +#define XFRM_IPTFS_MIN_L2HEADROOM (64 + 16)
How did you pick those values?
> +static struct sk_buff *iptfs_alloc_skb(struct sk_buff *tpl, u32 len,
> + bool l3resv)
> +{
> + struct sk_buff *skb;
> + u32 resv;
> +
> + if (!l3resv) {
> + resv = XFRM_IPTFS_MIN_L2HEADROOM;
> + } else {
> + resv = skb_headroom(tpl);
> + if (resv < XFRM_IPTFS_MIN_L3HEADROOM)
> + resv = XFRM_IPTFS_MIN_L3HEADROOM;
> + }
> +
> + skb = alloc_skb(len + resv, GFP_ATOMIC);
> + if (!skb) {
> + XFRM_INC_STATS(dev_net(tpl->dev), LINUX_MIB_XFRMNOSKBERROR);
Hmpf, so we've gone from incrementing the wrong counter to
incrementing a new counter that doesn't have a precise meaning.
> + return NULL;
> + }
> +
> + skb_reserve(skb, resv);
> +
> + /* We do not want any of the tpl->headers copied over, so we do
> + * not use `skb_copy_header()`.
> + */
This is a bit of a bad sign for the implementation. It also worries
me, as this may not be updated when changes are made to
__copy_skb_header().
(c/p'd from v1 review since this was still not answered)
> +/**
> + * skb_copy_bits_seq - copy bits from a skb_seq_state to kernel buffer
> + * @st: source skb_seq_state
> + * @offset: offset in source
> + * @to: destination buffer
> + * @len: number of bytes to copy
> + *
> + * Copy @len bytes from @offset bytes into the source @st to the destination
> + * buffer @to. `offset` should increase (or be unchanged) with each subsequent
> + * call to this function. If offset needs to decrease from the previous use `st`
> + * should be reset first.
> + *
> + * Return: 0 on success or a negative error code on failure
> + */
> +static int skb_copy_bits_seq(struct skb_seq_state *st, int offset, void *to,
> + int len)
Probably belongs in net/core/skbuff.c, although I'm really not
convinced copying data around is the right way to implement the type
of packet splitting IPTFS does (which sounds a bit like a kind of
GSO). And there are helpers in net/core/skbuff.c (such as
pskb_carve/pskb_extract) that seem to do similar things to what you
need here, without as much data copying.
> +static int iptfs_first_skb(struct sk_buff **skbp, struct xfrm_iptfs_data *xtfs,
> + u32 mtu)
> +{
> + struct sk_buff *skb = *skbp;
> + int err;
> +
> + /* Classic ESP skips the don't fragment ICMP error if DF is clear on
> + * the inner packet or ignore_df is set. Otherwise it will send an ICMP
> + * or local error if the inner packet won't fit it's MTU.
> + *
> + * With IPTFS we do not care about the inner packet DF bit. If the
> + * tunnel is configured to "don't fragment" we error back if things
> + * don't fit in our max packet size. Otherwise we iptfs-fragment as
> + * normal.
> + */
> +
> + /* The opportunity for HW offload has ended */
> + if (skb->ip_summed == CHECKSUM_PARTIAL) {
> + err = skb_checksum_help(skb);
> + if (err)
> + return err;
> + }
> +
> + /* We've split these up before queuing */
> + BUG_ON(skb_is_gso(skb));
As I've said previously, I don't think that's a valid reason to
crash. BUG_ON should be used very rarely:
https://elixir.bootlin.com/linux/v6.10/source/Documentation/process/coding-style.rst#L1230
Dropping a bogus packet is an easy way to recover from this situation,
so we should not crash here (and probably in all of IPTFS).
--
Sabrina
^ permalink raw reply [flat|nested] 40+ messages in thread* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-04 22:25 ` Sabrina Dubroca
@ 2024-08-05 2:33 ` Christian Hopps
2024-08-05 4:19 ` Christian Hopps
2024-08-06 8:47 ` Sabrina Dubroca
2024-08-06 11:32 ` Steffen Klassert
1 sibling, 2 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-05 2:33 UTC (permalink / raw)
To: Sabrina Dubroca
Cc: Christian Hopps, devel, Steffen Klassert, netdev, Christian Hopps
Sabrina Dubroca <sd@queasysnail.net> writes:
> Please CC the reviewers from previous versions of the patchset. It's
> really hard to keep track of discussions and reposts otherwise.
Wasn't aware of this requirement, will try and add all the reviewers in the future.
> 2024-08-04, 16:33:39 -0400, Christian Hopps wrote:
>> From: Christian Hopps <chopps@labn.net>
>>
>> Add support for tunneling user (inner) packets that are larger than the
>> tunnel's path MTU (outer) using IP-TFS fragmentation.
>>
>> Signed-off-by: Christian Hopps <chopps@labn.net>
>> ---
>> net/xfrm/xfrm_iptfs.c | 407 +++++++++++++++++++++++++++++++++++++++---
>> 1 file changed, 381 insertions(+), 26 deletions(-)
>>
>> diff --git a/net/xfrm/xfrm_iptfs.c b/net/xfrm/xfrm_iptfs.c
>> index 20c19894720e..38735e2d64c3 100644
>> --- a/net/xfrm/xfrm_iptfs.c
>> +++ b/net/xfrm/xfrm_iptfs.c
>> @@ -46,12 +46,23 @@
>> */
>> #define IPTFS_DEFAULT_MAX_QUEUE_SIZE (1024 * 10240)
>>
>> +/* 1) skb->head should be cache aligned.
>> + * 2) when resv is for L2 headers (i.e., ethernet) we want the cacheline to
>> + * start -16 from data.
>> + * 3) when resv is for L3+L2 headers IOW skb->data points at the IPTFS payload
>> + * we want data to be cache line aligned so all the pushed headers will be in
>> + * another cacheline.
>> + */
>> +#define XFRM_IPTFS_MIN_L3HEADROOM 128
>> +#define XFRM_IPTFS_MIN_L2HEADROOM (64 + 16)
>
> How did you pick those values?
That's what the comment is talking to. When reserving space for L2 headers we pick 64 + 16 (a 2^(<=6) cacheline + 16 bytes so the the cacheline should start -16 from where skb->data will point at. For L3 we reserve double the power of 2 space we reserved for L2 only.
We have to reserve some amount of space for pushed headers, so the above made sense to me for good performance/cache locality.
>> +static struct sk_buff *iptfs_alloc_skb(struct sk_buff *tpl, u32 len,
>> + bool l3resv)
>> +{
>> + struct sk_buff *skb;
>> + u32 resv;
>> +
>> + if (!l3resv) {
>> + resv = XFRM_IPTFS_MIN_L2HEADROOM;
>> + } else {
>> + resv = skb_headroom(tpl);
>> + if (resv < XFRM_IPTFS_MIN_L3HEADROOM)
>> + resv = XFRM_IPTFS_MIN_L3HEADROOM;
>> + }
>> +
>> + skb = alloc_skb(len + resv, GFP_ATOMIC);
>> + if (!skb) {
>> + XFRM_INC_STATS(dev_net(tpl->dev), LINUX_MIB_XFRMNOSKBERROR);
>
> Hmpf, so we've gone from incrementing the wrong counter to
> incrementing a new counter that doesn't have a precise meaning.
The new "No SKB" counter is supposed to mean "couldn't get an SKB", given plenty of other errors are logged under "OutErr" or "InErr" i'm not sure what level of precision you're looking for here. :)
>> + return NULL;
>> + }
>> +
>> + skb_reserve(skb, resv);
>> +
>> + /* We do not want any of the tpl->headers copied over, so we do
>> + * not use `skb_copy_header()`.
>> + */
>
> This is a bit of a bad sign for the implementation. It also worries
> me, as this may not be updated when changes are made to
> __copy_skb_header().
> (c/p'd from v1 review since this was still not answered)
I don't agree that this is a bad design at all, I'm curious what you think a good design to be.
I did specifically state why we are not re-using skb_copy_header(). The functionality is different. We are not trying to make a copy of an skb we are using an skb as a template for new skbs.
>> +/**
>> + * skb_copy_bits_seq - copy bits from a skb_seq_state to kernel buffer
>> + * @st: source skb_seq_state
>> + * @offset: offset in source
>> + * @to: destination buffer
>> + * @len: number of bytes to copy
>> + *
>> + * Copy @len bytes from @offset bytes into the source @st to the destination
>> + * buffer @to. `offset` should increase (or be unchanged) with each subsequent
>> + * call to this function. If offset needs to decrease from the previous use `st`
>> + * should be reset first.
>> + *
>> + * Return: 0 on success or a negative error code on failure
>> + */
>> +static int skb_copy_bits_seq(struct skb_seq_state *st, int offset, void *to,
>> + int len)
>
> Probably belongs in net/core/skbuff.c, although I'm really not
> convinced copying data around is the right way to implement the type
> of packet splitting IPTFS does (which sounds a bit like a kind of
> GSO). And there are helpers in net/core/skbuff.c (such as
> pskb_carve/pskb_extract) that seem to do similar things to what you
> need here, without as much data copying.
I don't have an issue with moving more general skb functionality into skbuff.c; however, I do not want to gate IP-TFS on this change to the general net infra, it is appropriate for a patchset of it's own.
Re copying: Let's be clear here, we are not always copying data, there are sharing code paths as well; however, there are times when it is the best (and even fastest) way to accomplish things (e.g., b/c the packet is small or the data is arranged in skbs in a way to make sharing ridiculously complex and thus slow).
This is in fact a very elegant and efficient sequential walk solution to a complex problem.
>> +static int iptfs_first_skb(struct sk_buff **skbp, struct xfrm_iptfs_data *xtfs,
>> + u32 mtu)
>> +{
>> + struct sk_buff *skb = *skbp;
>> + int err;
>> +
>> + /* Classic ESP skips the don't fragment ICMP error if DF is clear on
>> + * the inner packet or ignore_df is set. Otherwise it will send an ICMP
>> + * or local error if the inner packet won't fit it's MTU.
>> + *
>> + * With IPTFS we do not care about the inner packet DF bit. If the
>> + * tunnel is configured to "don't fragment" we error back if things
>> + * don't fit in our max packet size. Otherwise we iptfs-fragment as
>> + * normal.
>> + */
>> +
>> + /* The opportunity for HW offload has ended */
>> + if (skb->ip_summed == CHECKSUM_PARTIAL) {
>> + err = skb_checksum_help(skb);
>> + if (err)
>> + return err;
>> + }
>> +
>> + /* We've split these up before queuing */
>> + BUG_ON(skb_is_gso(skb));
>
> As I've said previously, I don't think that's a valid reason to
> crash. BUG_ON should be used very rarely:
>
> https://elixir.bootlin.com/linux/v6.10/source/Documentation/process/coding-style.rst#L1230
>
> Dropping a bogus packet is an easy way to recover from this situation,
> so we should not crash here (and probably in all of IPTFS).
This is basically following a style of coding that aims to simplify overall code by eliminating multiple checks for the same condition over and over in code. It does this by arranging for a single variant at the beginning of an operation and then counting on that from then on in the code. Asserts are the way to document this, if no assert then nothing b/c using a conditional is exactly against the design principle.
An existing example of this in the kernel is `assert_spin_locked()`.
Anyway, I will just remove it if this is going to block adoption of the patch.
Thanks,
Chris.
^ permalink raw reply [flat|nested] 40+ messages in thread* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-05 2:33 ` Christian Hopps
@ 2024-08-05 4:19 ` Christian Hopps
2024-08-06 8:47 ` Sabrina Dubroca
1 sibling, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-05 4:19 UTC (permalink / raw)
To: Sabrina Dubroca
Cc: Christian Hopps, devel, Steffen Klassert, netdev, Christian Hopps
Christian Hopps <chopps@chopps.org> writes:
> Sabrina Dubroca <sd@queasysnail.net> writes:
>>> + /* The opportunity for HW offload has ended */
>>> + if (skb->ip_summed == CHECKSUM_PARTIAL) {
>>> + err = skb_checksum_help(skb);
>>> + if (err)
>>> + return err;
>>> + }
>>> +
>>> + /* We've split these up before queuing */
>>> + BUG_ON(skb_is_gso(skb));
>>
>> As I've said previously, I don't think that's a valid reason to
>> crash. BUG_ON should be used very rarely:
>>
>> https://elixir.bootlin.com/linux/v6.10/source/Documentation/process/coding-style.rst#L1230
>>
>> Dropping a bogus packet is an easy way to recover from this situation,
>> so we should not crash here (and probably in all of IPTFS).
>
> This is basically following a style of coding that aims to simplify overall code
> by eliminating multiple checks for the same condition over and over in code. It
> does this by arranging for a single variant at the beginning of an operation and
> then counting on that from then on in the code. Asserts are the way to document
> this, if no assert then nothing b/c using a conditional is exactly against the
> design principle.
>
> An existing example of this in the kernel is `assert_spin_locked()`.
>
> Anyway, I will just remove it if this is going to block adoption of the patch.
Actually, I'll just convert all BUG_ON() to WARN_ON().
Thanks,
Chris.
> Thanks,
> Chris.
^ permalink raw reply [flat|nested] 40+ messages in thread* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-05 2:33 ` Christian Hopps
2024-08-05 4:19 ` Christian Hopps
@ 2024-08-06 8:47 ` Sabrina Dubroca
2024-08-06 8:54 ` Christian Hopps
1 sibling, 1 reply; 40+ messages in thread
From: Sabrina Dubroca @ 2024-08-06 8:47 UTC (permalink / raw)
To: Christian Hopps; +Cc: devel, Steffen Klassert, netdev, Christian Hopps
2024-08-04, 22:33:05 -0400, Christian Hopps wrote:
> > > +/* 1) skb->head should be cache aligned.
> > > + * 2) when resv is for L2 headers (i.e., ethernet) we want the cacheline to
> > > + * start -16 from data.
> > > + * 3) when resv is for L3+L2 headers IOW skb->data points at the IPTFS payload
> > > + * we want data to be cache line aligned so all the pushed headers will be in
> > > + * another cacheline.
> > > + */
> > > +#define XFRM_IPTFS_MIN_L3HEADROOM 128
> > > +#define XFRM_IPTFS_MIN_L2HEADROOM (64 + 16)
> >
> > How did you pick those values?
>
> That's what the comment is talking to. When reserving space for L2 headers we pick 64 + 16 (a 2^(<=6) cacheline + 16 bytes so the the cacheline should start -16 from where skb->data will point at.
Hard-coding the x86 cacheline size is not a good idea. And what's the
16B for? You don't know that it's enough for the actual L2 headers.
> For L3 we reserve double the power of 2 space we reserved for L2 only.
But that's the core of my question. Why is that correct/enough?
>
> We have to reserve some amount of space for pushed headers, so the above made sense to me for good performance/cache locality.
>
> > > +static struct sk_buff *iptfs_alloc_skb(struct sk_buff *tpl, u32 len,
> > > + bool l3resv)
> > > +{
> > > + struct sk_buff *skb;
> > > + u32 resv;
> > > +
> > > + if (!l3resv) {
> > > + resv = XFRM_IPTFS_MIN_L2HEADROOM;
> > > + } else {
> > > + resv = skb_headroom(tpl);
> > > + if (resv < XFRM_IPTFS_MIN_L3HEADROOM)
> > > + resv = XFRM_IPTFS_MIN_L3HEADROOM;
> > > + }
> > > +
> > > + skb = alloc_skb(len + resv, GFP_ATOMIC);
> > > + if (!skb) {
> > > + XFRM_INC_STATS(dev_net(tpl->dev), LINUX_MIB_XFRMNOSKBERROR);
> >
> > Hmpf, so we've gone from incrementing the wrong counter to
> > incrementing a new counter that doesn't have a precise meaning.
>
> The new "No SKB" counter is supposed to mean "couldn't get an SKB",
> given plenty of other errors are logged under "OutErr" or "InErr"
> i'm not sure what level of precision you're looking for here. :)
OutErr and InErr would be better than that new counter IMO.
> > > + return NULL;
> > > + }
> > > +
> > > + skb_reserve(skb, resv);
> > > +
> > > + /* We do not want any of the tpl->headers copied over, so we do
> > > + * not use `skb_copy_header()`.
> > > + */
> >
> > This is a bit of a bad sign for the implementation. It also worries
> > me, as this may not be updated when changes are made to
> > __copy_skb_header().
> > (c/p'd from v1 review since this was still not answered)
>
> I don't agree that this is a bad design at all, I'm curious what you think a good design to be.
Strange skb manipulations hiding in a protocol module is not good
design.
c/p bits of core code into a module (where they will never get fixed
up when the core code gets updated) is always a bad idea.
> I did specifically state why we are not re-using
> skb_copy_header(). The functionality is different. We are not trying
> to make a copy of an skb we are using an skb as a template for new
> skbs.
I saw that. That doesn't mean it's a good thing to do.
> > > +/**
> > > + * skb_copy_bits_seq - copy bits from a skb_seq_state to kernel buffer
> > > + * @st: source skb_seq_state
> > > + * @offset: offset in source
> > > + * @to: destination buffer
> > > + * @len: number of bytes to copy
> > > + *
> > > + * Copy @len bytes from @offset bytes into the source @st to the destination
> > > + * buffer @to. `offset` should increase (or be unchanged) with each subsequent
> > > + * call to this function. If offset needs to decrease from the previous use `st`
> > > + * should be reset first.
> > > + *
> > > + * Return: 0 on success or a negative error code on failure
> > > + */
> > > +static int skb_copy_bits_seq(struct skb_seq_state *st, int offset, void *to,
> > > + int len)
> >
> > Probably belongs in net/core/skbuff.c, although I'm really not
> > convinced copying data around is the right way to implement the type
> > of packet splitting IPTFS does (which sounds a bit like a kind of
> > GSO). And there are helpers in net/core/skbuff.c (such as
> > pskb_carve/pskb_extract) that seem to do similar things to what you
> > need here, without as much data copying.
>
> I don't have an issue with moving more general skb functionality
> into skbuff.c; however, I do not want to gate IP-TFS on this change
> to the general net infra, it is appropriate for a patchset of it's
> own.
If you need helpers that don't exist, it's part of your job to make
the core changes that are required to implement the functionality.
> Re copying: Let's be clear here, we are not always copying data,
> there are sharing code paths as well; however, there are times when
> it is the best (and even fastest) way to accomplish things (e.g.,
> b/c the packet is small or the data is arranged in skbs in a way to
> make sharing ridiculously complex and thus slow).
I'm not finding the sharing code. You mean iptfs_first_should_copy
returning false?
--
Sabrina
^ permalink raw reply [flat|nested] 40+ messages in thread* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-06 8:47 ` Sabrina Dubroca
@ 2024-08-06 8:54 ` Christian Hopps
2024-08-06 10:03 ` Florian Westphal
` (2 more replies)
0 siblings, 3 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-06 8:54 UTC (permalink / raw)
To: Sabrina Dubroca
Cc: Christian Hopps, devel, Steffen Klassert, netdev, Christian Hopps
Sabrina Dubroca <sd@queasysnail.net> writes:
> 2024-08-04, 22:33:05 -0400, Christian Hopps wrote:
>> > > +/* 1) skb->head should be cache aligned.
>> > > + * 2) when resv is for L2 headers (i.e., ethernet) we want the cacheline to
>> > > + * start -16 from data.
>> > > + * 3) when resv is for L3+L2 headers IOW skb->data points at the IPTFS payload
>> > > + * we want data to be cache line aligned so all the pushed headers will be in
>> > > + * another cacheline.
>> > > + */
>> > > +#define XFRM_IPTFS_MIN_L3HEADROOM 128
>> > > +#define XFRM_IPTFS_MIN_L2HEADROOM (64 + 16)
>> >
>> > How did you pick those values?
>>
>> That's what the comment is talking to. When reserving space for L2 headers we
>> pick 64 + 16 (a 2^(<=6) cacheline + 16 bytes so the the cacheline should start
>> -16 from where skb->data will point at.
>
> Hard-coding the x86 cacheline size is not a good idea. And what's the
> 16B for? You don't know that it's enough for the actual L2 headers.
I am not hard coding the x86 cacheline. I am picking 64 as the largest cacheline that this is optimized for, it also works for smaller cachelines.
16B is to allow for the incredibly common 14B L2 header to fit.
>> For L3 we reserve double the power of 2 space we reserved for L2 only.
>
> But that's the core of my question. Why is that correct/enough?
I have to pick a value. There is no magically perfect number that I can pick. I've given you technical reasons and justifications for the numbers I have chosen -- not sure what else I can say. Do you have better suggestions for the sizes which would be more optimal on more architectures? If not then let's use the numbers that I have given technical reasons for choosing.
Put this another way. I could just pick 128 b/c it's 2 cachelines and fits lots of different headers and would be "good enough". That's plenty justification too. I think you looking for too much here -- this isn't a precision thing, it's a "Good Enough" thing.
>> We have to reserve some amount of space for pushed headers, so the above made sense to me for good performance/cache locality.
>>
>> > > +static struct sk_buff *iptfs_alloc_skb(struct sk_buff *tpl, u32 len,
>> > > + bool l3resv)
>> > > +{
>> > > + struct sk_buff *skb;
>> > > + u32 resv;
>> > > +
>> > > + if (!l3resv) {
>> > > + resv = XFRM_IPTFS_MIN_L2HEADROOM;
>> > > + } else {
>> > > + resv = skb_headroom(tpl);
>> > > + if (resv < XFRM_IPTFS_MIN_L3HEADROOM)
>> > > + resv = XFRM_IPTFS_MIN_L3HEADROOM;
>> > > + }
>> > > +
>> > > + skb = alloc_skb(len + resv, GFP_ATOMIC);
>> > > + if (!skb) {
>> > > + XFRM_INC_STATS(dev_net(tpl->dev), LINUX_MIB_XFRMNOSKBERROR);
>> >
>> > Hmpf, so we've gone from incrementing the wrong counter to
>> > incrementing a new counter that doesn't have a precise meaning.
>>
>> The new "No SKB" counter is supposed to mean "couldn't get an SKB",
>> given plenty of other errors are logged under "OutErr" or "InErr"
>> i'm not sure what level of precision you're looking for here. :)
>
> OutErr and InErr would be better than that new counter IMO.
Why?
My counter tracks the SKB depletion failure that is actually happening. Would you have me now pass in the direction argument just so I can tick the correct overly general MIB counter that provides less value to the user in identifying the actual problem? How is that good design?
I'm inclined to just delete the thing altogether rather than block on this thing that will almost never happen.
>> > > + return NULL;
>> > > + }
>> > > +
>> > > + skb_reserve(skb, resv);
>> > > +
>> > > + /* We do not want any of the tpl->headers copied over, so we do
>> > > + * not use `skb_copy_header()`.
>> > > + */
>> >
>> > This is a bit of a bad sign for the implementation. It also worries
>> > me, as this may not be updated when changes are made to
>> > __copy_skb_header().
>> > (c/p'd from v1 review since this was still not answered)
>>
>> I don't agree that this is a bad design at all, I'm curious what you think a good design to be.
>
> Strange skb manipulations hiding in a protocol module is not good
> design.
It's a fragmentation and aggregation protocol, it's needs work with skbs by design. It's literally the function of the protocol to manipulate packet content.
I would appreciate it if you could provide technical reasons to justify referring to things as "bad" or "strange" -- it's not helpful otherwise.
> c/p bits of core code into a module (where they will never get fixed
> up when the core code gets updated) is always a bad idea.
I need some values from the SKB, so I copy them -- it's that simple.
>> I did specifically state why we are not re-using
>> skb_copy_header(). The functionality is different. We are not trying
>> to make a copy of an skb we are using an skb as a template for new
>> skbs.
>
> I saw that. That doesn't mean it's a good thing to do.
Please suggest an alternative.
>> > > +/**
>> > > + * skb_copy_bits_seq - copy bits from a skb_seq_state to kernel buffer
>> > > + * @st: source skb_seq_state
>> > > + * @offset: offset in source
>> > > + * @to: destination buffer
>> > > + * @len: number of bytes to copy
>> > > + *
>> > > + * Copy @len bytes from @offset bytes into the source @st to the destination
>> > > + * buffer @to. `offset` should increase (or be unchanged) with each subsequent
>> > > + * call to this function. If offset needs to decrease from the previous use `st`
>> > > + * should be reset first.
>> > > + *
>> > > + * Return: 0 on success or a negative error code on failure
>> > > + */
>> > > +static int skb_copy_bits_seq(struct skb_seq_state *st, int offset, void *to,
>> > > + int len)
>> >
>> > Probably belongs in net/core/skbuff.c, although I'm really not
>> > convinced copying data around is the right way to implement the type
>> > of packet splitting IPTFS does (which sounds a bit like a kind of
>> > GSO). And there are helpers in net/core/skbuff.c (such as
>> > pskb_carve/pskb_extract) that seem to do similar things to what you
>> > need here, without as much data copying.
>>
>> I don't have an issue with moving more general skb functionality
>> into skbuff.c; however, I do not want to gate IP-TFS on this change
>> to the general net infra, it is appropriate for a patchset of it's
>> own.
>
> If you need helpers that don't exist, it's part of your job to make
> the core changes that are required to implement the functionality.
This is part of a new code protocol and feature addition and it's a single use. Another patchset can present this code to the general network community to see if they think it *also* has value outside of IPTFS. There is *no* reason to delay IPTFS on general network infrastructure improvements. Please don't do this.
>> Re copying: Let's be clear here, we are not always copying data,
>> there are sharing code paths as well; however, there are times when
>> it is the best (and even fastest) way to accomplish things (e.g.,
>> b/c the packet is small or the data is arranged in skbs in a way to
>> make sharing ridiculously complex and thus slow).
>
> I'm not finding the sharing code. You mean iptfs_first_should_copy
> returning false?
/* Try share then copy. */
if (fragwalk && skb_can_add_frags(newskb, fragwalk, data, copylen)) {
...
leftover = skb_add_frags(newskb, fragwalk, data, copylen);
} else {
/* copy fragment data into newskb */
if (skb_copy_bits_seq(st, data, skb_put(newskb, copylen),
...
}
Thanks,
Chris.
^ permalink raw reply [flat|nested] 40+ messages in thread* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-06 8:54 ` Christian Hopps
@ 2024-08-06 10:03 ` Florian Westphal
2024-08-06 10:05 ` Christian Hopps
2024-08-06 11:05 ` Sabrina Dubroca
2024-08-06 11:07 ` Steffen Klassert
2 siblings, 1 reply; 40+ messages in thread
From: Florian Westphal @ 2024-08-06 10:03 UTC (permalink / raw)
To: Christian Hopps
Cc: Sabrina Dubroca, devel, Steffen Klassert, netdev, Christian Hopps
Christian Hopps <chopps@chopps.org> wrote:
> > > > > + if (!l3resv) {
> > > > > + resv = XFRM_IPTFS_MIN_L2HEADROOM;
> > > > > + } else {
> > > > > + resv = skb_headroom(tpl);
> > > > > + if (resv < XFRM_IPTFS_MIN_L3HEADROOM)
> > > > > + resv = XFRM_IPTFS_MIN_L3HEADROOM;
> > > > > + }
> > > > > +
> > > > > + skb = alloc_skb(len + resv, GFP_ATOMIC);
> > > > > + if (!skb) {
> > > > > + XFRM_INC_STATS(dev_net(tpl->dev), LINUX_MIB_XFRMNOSKBERROR);
> > > >
> > > > Hmpf, so we've gone from incrementing the wrong counter to
> > > > incrementing a new counter that doesn't have a precise meaning.
> > >
> > > The new "No SKB" counter is supposed to mean "couldn't get an SKB",
> > > given plenty of other errors are logged under "OutErr" or "InErr"
> > > i'm not sure what level of precision you're looking for here. :)
> >
> > OutErr and InErr would be better than that new counter IMO.
>
> Why?
>
> My counter tracks the SKB depletion failure that is actually happening. Would you have me now pass in the direction argument just so I can tick the correct overly general MIB counter that provides less value to the user in identifying the actual problem? How is that good design?
>
> I'm inclined to just delete the thing altogether rather than block on this thing that will almost never happen.
Makes sense to me, skb allocation failure is transient anyway, there is
no action that could be taken if this error counter is incrementing.
You might want to pass GFP_ATOMIC | __GFP_NOWARN to alloc_skb() to avoid
any splats given this is a high-volume allocation.
^ permalink raw reply [flat|nested] 40+ messages in thread* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-06 10:03 ` Florian Westphal
@ 2024-08-06 10:05 ` Christian Hopps
0 siblings, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-06 10:05 UTC (permalink / raw)
To: Florian Westphal
Cc: Christian Hopps, Sabrina Dubroca, devel, Steffen Klassert, netdev,
Christian Hopps
Florian Westphal <fw@strlen.de> writes:
> Christian Hopps <chopps@chopps.org> wrote:
>> > > > > + if (!l3resv) {
>> > > > > + resv = XFRM_IPTFS_MIN_L2HEADROOM;
>> > > > > + } else {
>> > > > > + resv = skb_headroom(tpl);
>> > > > > + if (resv < XFRM_IPTFS_MIN_L3HEADROOM)
>> > > > > + resv = XFRM_IPTFS_MIN_L3HEADROOM;
>> > > > > + }
>> > > > > +
>> > > > > + skb = alloc_skb(len + resv, GFP_ATOMIC);
>> > > > > + if (!skb) {
>> > > > > + XFRM_INC_STATS(dev_net(tpl->dev), LINUX_MIB_XFRMNOSKBERROR);
>> > > >
>> > > > Hmpf, so we've gone from incrementing the wrong counter to
>> > > > incrementing a new counter that doesn't have a precise meaning.
>> > >
>> > > The new "No SKB" counter is supposed to mean "couldn't get an SKB",
>> > > given plenty of other errors are logged under "OutErr" or "InErr"
>> > > i'm not sure what level of precision you're looking for here. :)
>> >
>> > OutErr and InErr would be better than that new counter IMO.
>>
>> Why?
>>
>> My counter tracks the SKB depletion failure that is actually happening. Would
>> you have me now pass in the direction argument just so I can tick the correct
>> overly general MIB counter that provides less value to the user in identifying
>> the actual problem? How is that good design?
>>
>> I'm inclined to just delete the thing altogether rather than block on this thing that will almost never happen.
>
> Makes sense to me, skb allocation failure is transient anyway, there is
> no action that could be taken if this error counter is incrementing.
>
> You might want to pass GFP_ATOMIC | __GFP_NOWARN to alloc_skb() to avoid
> any splats given this is a high-volume allocation.
Will do.
Thanks,
Chris.
^ permalink raw reply [flat|nested] 40+ messages in thread
* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-06 8:54 ` Christian Hopps
2024-08-06 10:03 ` Florian Westphal
@ 2024-08-06 11:05 ` Sabrina Dubroca
2024-08-06 11:07 ` Christian Hopps
2024-08-08 11:30 ` Christian Hopps
2024-08-06 11:07 ` Steffen Klassert
2 siblings, 2 replies; 40+ messages in thread
From: Sabrina Dubroca @ 2024-08-06 11:05 UTC (permalink / raw)
To: Christian Hopps; +Cc: devel, Steffen Klassert, netdev, Christian Hopps
2024-08-06, 04:54:53 -0400, Christian Hopps wrote:
>
> Sabrina Dubroca <sd@queasysnail.net> writes:
>
> > 2024-08-04, 22:33:05 -0400, Christian Hopps wrote:
> > > > > +/* 1) skb->head should be cache aligned.
> > > > > + * 2) when resv is for L2 headers (i.e., ethernet) we want the cacheline to
> > > > > + * start -16 from data.
> > > > > + * 3) when resv is for L3+L2 headers IOW skb->data points at the IPTFS payload
> > > > > + * we want data to be cache line aligned so all the pushed headers will be in
> > > > > + * another cacheline.
> > > > > + */
> > > > > +#define XFRM_IPTFS_MIN_L3HEADROOM 128
> > > > > +#define XFRM_IPTFS_MIN_L2HEADROOM (64 + 16)
> > > >
> > > > How did you pick those values?
> > >
> > > That's what the comment is talking to. When reserving space for L2 headers we
> > > pick 64 + 16 (a 2^(<=6) cacheline + 16 bytes so the the cacheline should start
> > > -16 from where skb->data will point at.
> >
> > Hard-coding the x86 cacheline size is not a good idea. And what's the
> > 16B for? You don't know that it's enough for the actual L2 headers.
>
> I am not hard coding the x86 cacheline. I am picking 64 as the largest cacheline that this is optimized for, it also works for smaller cachelines.
At least use SMP_CACHE_BYTES then?
> 16B is to allow for the incredibly common 14B L2 header to fit.
Why not use skb->dev->needed_headroom, like a bunch of tunnels are
already doing? No guessing required. ethernet is the most common, but
there's no reason to penalize other protocols when the information is
available.
> > > For L3 we reserve double the power of 2 space we reserved for L2 only.
> >
> > But that's the core of my question. Why is that correct/enough?
>
> I have to pick a value. There is no magically perfect number that I can pick. I've given you technical reasons and justifications for the numbers I have chosen -- not sure what else I can say. Do you have better suggestions for the sizes which would be more optimal on more architectures? If not then let's use the numbers that I have given technical reasons for choosing.
Yes, now you've spelled it out, and we can evaluate your choices.
> Put this another way. I could just pick 128 b/c it's 2 cachelines
> and fits lots of different headers and would be "good
> enough". That's plenty justification too. I think you looking for
> too much here -- this isn't a precision thing, it's a "Good Enough"
> thing.
I'm asking questions. That's kind of the reviewer's job, understanding
how the thing they're reviewing works. ¯\_(ツ)_/¯
> > > We have to reserve some amount of space for pushed headers, so the above made sense to me for good performance/cache locality.
> > >
> > > > > +static struct sk_buff *iptfs_alloc_skb(struct sk_buff *tpl, u32 len,
> > > > > + bool l3resv)
> > > > > +{
> > > > > + struct sk_buff *skb;
> > > > > + u32 resv;
> > > > > +
> > > > > + if (!l3resv) {
> > > > > + resv = XFRM_IPTFS_MIN_L2HEADROOM;
> > > > > + } else {
> > > > > + resv = skb_headroom(tpl);
> > > > > + if (resv < XFRM_IPTFS_MIN_L3HEADROOM)
> > > > > + resv = XFRM_IPTFS_MIN_L3HEADROOM;
> > > > > + }
> > > > > +
> > > > > + skb = alloc_skb(len + resv, GFP_ATOMIC);
> > > > > + if (!skb) {
> > > > > + XFRM_INC_STATS(dev_net(tpl->dev), LINUX_MIB_XFRMNOSKBERROR);
> > > >
> > > > Hmpf, so we've gone from incrementing the wrong counter to
> > > > incrementing a new counter that doesn't have a precise meaning.
> > >
> > > The new "No SKB" counter is supposed to mean "couldn't get an SKB",
> > > given plenty of other errors are logged under "OutErr" or "InErr"
> > > i'm not sure what level of precision you're looking for here. :)
> >
> > OutErr and InErr would be better than that new counter IMO.
>
> Why?
>
> My counter tracks the SKB depletion failure that is actually happening. Would you have me now pass in the direction argument just so I can tick the correct overly general MIB counter that provides less value to the user in identifying the actual problem? How is that good design?
>
> I'm inclined to just delete the thing altogether rather than block on this thing that will almost never happen.
Fine.
> > > > > + return NULL;
> > > > > + }
> > > > > +
> > > > > + skb_reserve(skb, resv);
> > > > > +
> > > > > + /* We do not want any of the tpl->headers copied over, so we do
> > > > > + * not use `skb_copy_header()`.
> > > > > + */
> > > >
> > > > This is a bit of a bad sign for the implementation. It also worries
> > > > me, as this may not be updated when changes are made to
> > > > __copy_skb_header().
> > > > (c/p'd from v1 review since this was still not answered)
> > >
> > > I don't agree that this is a bad design at all, I'm curious what you think a good design to be.
> >
> > Strange skb manipulations hiding in a protocol module is not good
> > design.
>
> It's a fragmentation and aggregation protocol, it's needs work with skbs by design. It's literally the function of the protocol to manipulate packet content.
packet content != cherry-picked parts of sk_buff
> I would appreciate it if you could provide technical reasons to justify referring to things as "bad" or "strange" -- it's not helpful otherwise.
I did say it's a bad sign, not a blocking issue on its own. But that
bad sign, combined with the unusual use of skb_seq and a lot of
copying data around, indicates that this is not the right way to
implement this part of the protocol.
> > c/p bits of core code into a module (where they will never get fixed
> > up when the core code gets updated) is always a bad idea.
>
> I need some values from the SKB, so I copy them -- it's that simple.
>
> > > I did specifically state why we are not re-using
> > > skb_copy_header(). The functionality is different. We are not trying
> > > to make a copy of an skb we are using an skb as a template for new
> > > skbs.
> >
> > I saw that. That doesn't mean it's a good thing to do.
>
> Please suggest an alternative.
A common helper in a location where people are going to know that they
need to fix it up when they modify things about sk_buff would be a
good start.
> > > > > +/**
> > > > > + * skb_copy_bits_seq - copy bits from a skb_seq_state to kernel buffer
> > > > > + * @st: source skb_seq_state
> > > > > + * @offset: offset in source
> > > > > + * @to: destination buffer
> > > > > + * @len: number of bytes to copy
> > > > > + *
> > > > > + * Copy @len bytes from @offset bytes into the source @st to the destination
> > > > > + * buffer @to. `offset` should increase (or be unchanged) with each subsequent
> > > > > + * call to this function. If offset needs to decrease from the previous use `st`
> > > > > + * should be reset first.
> > > > > + *
> > > > > + * Return: 0 on success or a negative error code on failure
> > > > > + */
> > > > > +static int skb_copy_bits_seq(struct skb_seq_state *st, int offset, void *to,
> > > > > + int len)
> > > >
> > > > Probably belongs in net/core/skbuff.c, although I'm really not
> > > > convinced copying data around is the right way to implement the type
> > > > of packet splitting IPTFS does (which sounds a bit like a kind of
> > > > GSO). And there are helpers in net/core/skbuff.c (such as
> > > > pskb_carve/pskb_extract) that seem to do similar things to what you
> > > > need here, without as much data copying.
> > >
> > > I don't have an issue with moving more general skb functionality
> > > into skbuff.c; however, I do not want to gate IP-TFS on this change
> > > to the general net infra, it is appropriate for a patchset of it's
> > > own.
> >
> > If you need helpers that don't exist, it's part of your job to make
> > the core changes that are required to implement the functionality.
>
> This is part of a new code protocol and feature addition and it's a single use.
Of course the helper would be single use when it's introduced. You
don't know if it will remain single use. And pskb_extract is single
use, it's fine.
> Another patchset can present this code to the general network
> community to see if they think it *also* has value outside of
> IPTFS. There is *no* reason to delay IPTFS on general network
> infrastructure improvements. Please don't do this.
Sorry, I don't think that's how it works.
> > > Re copying: Let's be clear here, we are not always copying data,
> > > there are sharing code paths as well; however, there are times when
> > > it is the best (and even fastest) way to accomplish things (e.g.,
> > > b/c the packet is small or the data is arranged in skbs in a way to
> > > make sharing ridiculously complex and thus slow).
> >
> > I'm not finding the sharing code. You mean iptfs_first_should_copy
> > returning false?
>
>
> /* Try share then copy. */
> if (fragwalk && skb_can_add_frags(newskb, fragwalk, data, copylen)) {
> ...
> leftover = skb_add_frags(newskb, fragwalk, data, copylen);
> } else {
> /* copy fragment data into newskb */
> if (skb_copy_bits_seq(st, data, skb_put(newskb, copylen),
> ...
> }
You're talking about reassembly now. This patch is fragmentation/TX.
--
Sabrina
^ permalink raw reply [flat|nested] 40+ messages in thread* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-06 11:05 ` Sabrina Dubroca
@ 2024-08-06 11:07 ` Christian Hopps
2024-08-08 11:30 ` Christian Hopps
1 sibling, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-06 11:07 UTC (permalink / raw)
To: Sabrina Dubroca
Cc: Christian Hopps, devel, Steffen Klassert, netdev, Christian Hopps
Sabrina Dubroca <sd@queasysnail.net> writes:
> 2024-08-06, 04:54:53 -0400, Christian Hopps wrote:
>>
>> Sabrina Dubroca <sd@queasysnail.net> writes:
>>
>> > 2024-08-04, 22:33:05 -0400, Christian Hopps wrote:
>> > > > > +/* 1) skb->head should be cache aligned.
>> > > > > + * 2) when resv is for L2 headers (i.e., ethernet) we want the cacheline to
>> > > > > + * start -16 from data.
>> > > > > + * 3) when resv is for L3+L2 headers IOW skb->data points at the IPTFS payload
>> > > > > + * we want data to be cache line aligned so all the pushed headers will be in
>> > > > > + * another cacheline.
>> > > > > + */
>> > > > > +#define XFRM_IPTFS_MIN_L3HEADROOM 128
>> > > > > +#define XFRM_IPTFS_MIN_L2HEADROOM (64 + 16)
>> > > >
>> > > > How did you pick those values?
>> > >
>> > > That's what the comment is talking to. When reserving space for L2 headers we
>> > > pick 64 + 16 (a 2^(<=6) cacheline + 16 bytes so the the cacheline should start
>> > > -16 from where skb->data will point at.
>> >
>> > Hard-coding the x86 cacheline size is not a good idea. And what's the
>> > 16B for? You don't know that it's enough for the actual L2 headers.
>>
>> I am not hard coding the x86 cacheline. I am picking 64 as the largest cacheline that this is optimized for, it also works for smaller cachelines.
>
> At least use SMP_CACHE_BYTES then?
Ok.
>> 16B is to allow for the incredibly common 14B L2 header to fit.
>
> Why not use skb->dev->needed_headroom, like a bunch of tunnels are
> already doing? No guessing required. ethernet is the most common, but
> there's no reason to penalize other protocols when the information is
> available.
>> > > For L3 we reserve double the power of 2 space we reserved for L2 only.
>> >
>> > But that's the core of my question. Why is that correct/enough?
>>
>> I have to pick a value. There is no magically perfect number that I can pick.
>> I've given you technical reasons and justifications for the numbers I have
>> chosen -- not sure what else I can say. Do you have better suggestions for the
>> sizes which would be more optimal on more architectures? If not then let's use
>> the numbers that I have given technical reasons for choosing.
>
> Yes, now you've spelled it out, and we can evaluate your choices.
>
>> Put this another way. I could just pick 128 b/c it's 2 cachelines
>> and fits lots of different headers and would be "good
>> enough". That's plenty justification too. I think you looking for
>> too much here -- this isn't a precision thing, it's a "Good Enough"
>> thing.
>
> I'm asking questions. That's kind of the reviewer's job, understanding
> how the thing they're reviewing works. ¯\_(ツ)_/¯
>
>> > > We have to reserve some amount of space for pushed headers, so the above made sense to me for good performance/cache locality.
>> > >
>> > > > > +static struct sk_buff *iptfs_alloc_skb(struct sk_buff *tpl, u32 len,
>> > > > > + bool l3resv)
>> > > > > +{
>> > > > > + struct sk_buff *skb;
>> > > > > + u32 resv;
>> > > > > +
>> > > > > + if (!l3resv) {
>> > > > > + resv = XFRM_IPTFS_MIN_L2HEADROOM;
>> > > > > + } else {
>> > > > > + resv = skb_headroom(tpl);
>> > > > > + if (resv < XFRM_IPTFS_MIN_L3HEADROOM)
>> > > > > + resv = XFRM_IPTFS_MIN_L3HEADROOM;
>> > > > > + }
>> > > > > +
>> > > > > + skb = alloc_skb(len + resv, GFP_ATOMIC);
>> > > > > + if (!skb) {
>> > > > > + XFRM_INC_STATS(dev_net(tpl->dev), LINUX_MIB_XFRMNOSKBERROR);
>> > > >
>> > > > Hmpf, so we've gone from incrementing the wrong counter to
>> > > > incrementing a new counter that doesn't have a precise meaning.
>> > >
>> > > The new "No SKB" counter is supposed to mean "couldn't get an SKB",
>> > > given plenty of other errors are logged under "OutErr" or "InErr"
>> > > i'm not sure what level of precision you're looking for here. :)
>> >
>> > OutErr and InErr would be better than that new counter IMO.
>>
>> Why?
>>
>> My counter tracks the SKB depletion failure that is actually happening. Would
>> you have me now pass in the direction argument just so I can tick the correct
>> overly general MIB counter that provides less value to the user in identifying
>> the actual problem? How is that good design?
>>
>> I'm inclined to just delete the thing altogether rather than block on this thing that will almost never happen.
>
> Fine.
>
>> > > > > + return NULL;
>> > > > > + }
>> > > > > +
>> > > > > + skb_reserve(skb, resv);
>> > > > > +
>> > > > > + /* We do not want any of the tpl->headers copied over, so we do
>> > > > > + * not use `skb_copy_header()`.
>> > > > > + */
>> > > >
>> > > > This is a bit of a bad sign for the implementation. It also worries
>> > > > me, as this may not be updated when changes are made to
>> > > > __copy_skb_header().
>> > > > (c/p'd from v1 review since this was still not answered)
>> > >
>> > > I don't agree that this is a bad design at all, I'm curious what you think a good design to be.
>> >
>> > Strange skb manipulations hiding in a protocol module is not good
>> > design.
>>
>> It's a fragmentation and aggregation protocol, it's needs work with skbs by design. It's literally the function of the protocol to manipulate packet content.
>
> packet content != cherry-picked parts of sk_buff
>
>> I would appreciate it if you could provide technical reasons to justify referring to things as "bad" or "strange" -- it's not helpful otherwise.
>
> I did say it's a bad sign, not a blocking issue on its own. But that
> bad sign, combined with the unusual use of skb_seq and a lot of
> copying data around, indicates that this is not the right way to
> implement this part of the protocol.
The seq walk is not a bad design, I am literally using an existing walk API to walk the possibly complex nested chain of skbs.
All I've added is a useful utility function using that API to copy chunks of data from the chain. Its doing this using a single seq walk through the chain. It's the most obvious clean solution I can imagine to extract the fragments -- hardly bad design, the opposite really.
>> > c/p bits of core code into a module (where they will never get fixed
>> > up when the core code gets updated) is always a bad idea.
>>
>> I need some values from the SKB, so I copy them -- it's that simple.
>>
>> > > I did specifically state why we are not re-using
>> > > skb_copy_header(). The functionality is different. We are not trying
>> > > to make a copy of an skb we are using an skb as a template for new
>> > > skbs.
>> >
>> > I saw that. That doesn't mean it's a good thing to do.
>>
>> Please suggest an alternative.
>
> A common helper in a location where people are going to know that they
> need to fix it up when they modify things about sk_buff would be a
> good start.
What I am copying is specific to IP-TFS use case, that's what I am trying to convey here. I am copying some data from the SKB, it is not a generalized cloning operation that should be shared by anyone.
Is the advice that I should move this IPTFS functionality into skbuff.c?
>> > > > > +/**
>> > > > > + * skb_copy_bits_seq - copy bits from a skb_seq_state to kernel buffer
>> > > > > + * @st: source skb_seq_state
>> > > > > + * @offset: offset in source
>> > > > > + * @to: destination buffer
>> > > > > + * @len: number of bytes to copy
>> > > > > + *
>> > > > > + * Copy @len bytes from @offset bytes into the source @st to the destination
>> > > > > + * buffer @to. `offset` should increase (or be unchanged) with each subsequent
>> > > > > + * call to this function. If offset needs to decrease from the previous use `st`
>> > > > > + * should be reset first.
>> > > > > + *
>> > > > > + * Return: 0 on success or a negative error code on failure
>> > > > > + */
>> > > > > +static int skb_copy_bits_seq(struct skb_seq_state *st, int offset, void *to,
>> > > > > + int len)
>> > > >
>> > > > Probably belongs in net/core/skbuff.c, although I'm really not
>> > > > convinced copying data around is the right way to implement the type
>> > > > of packet splitting IPTFS does (which sounds a bit like a kind of
>> > > > GSO). And there are helpers in net/core/skbuff.c (such as
>> > > > pskb_carve/pskb_extract) that seem to do similar things to what you
>> > > > need here, without as much data copying.
>> > >
>> > > I don't have an issue with moving more general skb functionality
>> > > into skbuff.c; however, I do not want to gate IP-TFS on this change
>> > > to the general net infra, it is appropriate for a patchset of it's
>> > > own.
>> >
>> > If you need helpers that don't exist, it's part of your job to make
>> > the core changes that are required to implement the functionality.
>>
>> This is part of a new code protocol and feature addition and it's a single use.
>
> Of course the helper would be single use when it's introduced. You
> don't know if it will remain single use. And pskb_extract is single
> use, it's fine.
That is what a nice targeted patch and review process on netdev would reveal.
>> Another patchset can present this code to the general network
>> community to see if they think it *also* has value outside of
>> IPTFS. There is *no* reason to delay IPTFS on general network
>> infrastructure improvements. Please don't do this.
>
> Sorry, I don't think that's how it works.
I guess I disagree. Trying to boil the ocean here is what this feels like. Let's introduce this major new feature IPTFS. That's enough to handle for this patchset and review.
>> > > Re copying: Let's be clear here, we are not always copying data,
>> > > there are sharing code paths as well; however, there are times when
>> > > it is the best (and even fastest) way to accomplish things (e.g.,
>> > > b/c the packet is small or the data is arranged in skbs in a way to
>> > > make sharing ridiculously complex and thus slow).
>> >
>> > I'm not finding the sharing code. You mean iptfs_first_should_copy
>> > returning false?
>>
>>
>> /* Try share then copy. */
>> if (fragwalk && skb_can_add_frags(newskb, fragwalk, data, copylen)) {
>> ...
>> leftover = skb_add_frags(newskb, fragwalk, data, copylen);
>> } else {
>> /* copy fragment data into newskb */
>> if (skb_copy_bits_seq(st, data, skb_put(newskb, copylen),
>> ...
>> }
>
> You're talking about reassembly now. This patch is fragmentation/TX.
Correct.
This code has been tested and performs quite good, especially for an initial implementation. It keeps up with existing ESP in the general IPmix case, and even outperforms the existing IPsec/ESP for small packets flows when aggregation kicks in. We also gain all the other benefits from IPTFS framing.
Adding to the code path you're referring to is possible enhancement task, it will be complex, and as it is not required to achieve on-par and even better performance than the existing code, it should not block or be required for the initial IPTFS implementation.
Thanks,
Chris.
^ permalink raw reply [flat|nested] 40+ messages in thread* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-06 11:05 ` Sabrina Dubroca
2024-08-06 11:07 ` Christian Hopps
@ 2024-08-08 11:30 ` Christian Hopps
2024-08-08 13:28 ` Sabrina Dubroca
1 sibling, 1 reply; 40+ messages in thread
From: Christian Hopps @ 2024-08-08 11:30 UTC (permalink / raw)
To: Sabrina Dubroca
Cc: Christian Hopps, devel, Steffen Klassert, netdev, Christian Hopps
[-- Attachment #1: Type: text/plain, Size: 10528 bytes --]
Sabrina Dubroca <sd@queasysnail.net> writes:
> 2024-08-06, 04:54:53 -0400, Christian Hopps wrote:
>>
>> Sabrina Dubroca <sd@queasysnail.net> writes:
>>
>> > 2024-08-04, 22:33:05 -0400, Christian Hopps wrote:
>> > > > > +/* 1) skb->head should be cache aligned.
>> > > > > + * 2) when resv is for L2 headers (i.e., ethernet) we want the cacheline to
>> > > > > + * start -16 from data.
>> > > > > + * 3) when resv is for L3+L2 headers IOW skb->data points at the IPTFS payload
>> > > > > + * we want data to be cache line aligned so all the pushed headers will be in
>> > > > > + * another cacheline.
>> > > > > + */
>> > > > > +#define XFRM_IPTFS_MIN_L3HEADROOM 128
>> > > > > +#define XFRM_IPTFS_MIN_L2HEADROOM (64 + 16)
>> > > >
>> > > > How did you pick those values?
>> > >
>> > > That's what the comment is talking to. When reserving space for L2 headers we
>> > > pick 64 + 16 (a 2^(<=6) cacheline + 16 bytes so the the cacheline should start
>> > > -16 from where skb->data will point at.
>> >
>> > Hard-coding the x86 cacheline size is not a good idea. And what's the
>> > 16B for? You don't know that it's enough for the actual L2 headers.
>>
>> I am not hard coding the x86 cacheline. I am picking 64 as the largest cacheline that this is optimized for, it also works for smaller cachelines.
>
> At least use SMP_CACHE_BYTES then?
Right, I have changed this work with L1_CACHE_BYTES value.
>> 16B is to allow for the incredibly common 14B L2 header to fit.
>
> Why not use skb->dev->needed_headroom, like a bunch of tunnels are
> already doing? No guessing required. ethernet is the most common, but
> there's no reason to penalize other protocols when the information is
> available.
We can't use `skb->dev->needed_headroom`, b/c `skb->dev` is not correct for the new packets. `skb->dev` is from the received IPTFS tunnel packet. The skb being created here are the inner user packets leaving the tunnel, so they have an L3 header (thus why we are only making room for L2 header). They are being handed to gro receive and still have to be routed to their correct destination interface/dev.
16 handles the general common case an ethernet device being the destination, if it's not correct after routing, nothing is broken, it just means that we may or may not achieve this maximal cache locality (but we still might e.g., if its destined to a GRE tunnel then we are looking at a bunch more headers so they and the existing L3 header will occupy 2 cachelines anyway). Again, this is a best effort thing.
Thanks,
Chris.
>
>> > > For L3 we reserve double the power of 2 space we reserved for L2 only.
>> >
>> > But that's the core of my question. Why is that correct/enough?
>>
>> I have to pick a value. There is no magically perfect number that I can pick.
>> I've given you technical reasons and justifications for the numbers I have
>> chosen -- not sure what else I can say. Do you have better suggestions for the
>> sizes which would be more optimal on more architectures? If not then let's use
>> the numbers that I have given technical reasons for choosing.
>
> Yes, now you've spelled it out, and we can evaluate your choices.
>
>> Put this another way. I could just pick 128 b/c it's 2 cachelines
>> and fits lots of different headers and would be "good
>> enough". That's plenty justification too. I think you looking for
>> too much here -- this isn't a precision thing, it's a "Good Enough"
>> thing.
>
> I'm asking questions. That's kind of the reviewer's job, understanding
> how the thing they're reviewing works. ¯\_(ツ)_/¯
>
>> > > We have to reserve some amount of space for pushed headers, so the above made sense to me for good performance/cache locality.
>> > >
>> > > > > +static struct sk_buff *iptfs_alloc_skb(struct sk_buff *tpl, u32 len,
>> > > > > + bool l3resv)
>> > > > > +{
>> > > > > + struct sk_buff *skb;
>> > > > > + u32 resv;
>> > > > > +
>> > > > > + if (!l3resv) {
>> > > > > + resv = XFRM_IPTFS_MIN_L2HEADROOM;
>> > > > > + } else {
>> > > > > + resv = skb_headroom(tpl);
>> > > > > + if (resv < XFRM_IPTFS_MIN_L3HEADROOM)
>> > > > > + resv = XFRM_IPTFS_MIN_L3HEADROOM;
>> > > > > + }
>> > > > > +
>> > > > > + skb = alloc_skb(len + resv, GFP_ATOMIC);
>> > > > > + if (!skb) {
>> > > > > + XFRM_INC_STATS(dev_net(tpl->dev), LINUX_MIB_XFRMNOSKBERROR);
>> > > >
>> > > > Hmpf, so we've gone from incrementing the wrong counter to
>> > > > incrementing a new counter that doesn't have a precise meaning.
>> > >
>> > > The new "No SKB" counter is supposed to mean "couldn't get an SKB",
>> > > given plenty of other errors are logged under "OutErr" or "InErr"
>> > > i'm not sure what level of precision you're looking for here. :)
>> >
>> > OutErr and InErr would be better than that new counter IMO.
>>
>> Why?
>>
>> My counter tracks the SKB depletion failure that is actually happening. Would
>> you have me now pass in the direction argument just so I can tick the correct
>> overly general MIB counter that provides less value to the user in identifying
>> the actual problem? How is that good design?
>>
>> I'm inclined to just delete the thing altogether rather than block on this thing that will almost never happen.
>
> Fine.
>
>> > > > > + return NULL;
>> > > > > + }
>> > > > > +
>> > > > > + skb_reserve(skb, resv);
>> > > > > +
>> > > > > + /* We do not want any of the tpl->headers copied over, so we do
>> > > > > + * not use `skb_copy_header()`.
>> > > > > + */
>> > > >
>> > > > This is a bit of a bad sign for the implementation. It also worries
>> > > > me, as this may not be updated when changes are made to
>> > > > __copy_skb_header().
>> > > > (c/p'd from v1 review since this was still not answered)
>> > >
>> > > I don't agree that this is a bad design at all, I'm curious what you think a good design to be.
>> >
>> > Strange skb manipulations hiding in a protocol module is not good
>> > design.
>>
>> It's a fragmentation and aggregation protocol, it's needs work with skbs by design. It's literally the function of the protocol to manipulate packet content.
>
> packet content != cherry-picked parts of sk_buff
>
>> I would appreciate it if you could provide technical reasons to justify referring to things as "bad" or "strange" -- it's not helpful otherwise.
>
> I did say it's a bad sign, not a blocking issue on its own. But that
> bad sign, combined with the unusual use of skb_seq and a lot of
> copying data around, indicates that this is not the right way to
> implement this part of the protocol.
>
>> > c/p bits of core code into a module (where they will never get fixed
>> > up when the core code gets updated) is always a bad idea.
>>
>> I need some values from the SKB, so I copy them -- it's that simple.
>>
>> > > I did specifically state why we are not re-using
>> > > skb_copy_header(). The functionality is different. We are not trying
>> > > to make a copy of an skb we are using an skb as a template for new
>> > > skbs.
>> >
>> > I saw that. That doesn't mean it's a good thing to do.
>>
>> Please suggest an alternative.
>
> A common helper in a location where people are going to know that they
> need to fix it up when they modify things about sk_buff would be a
> good start.
>
>> > > > > +/**
>> > > > > + * skb_copy_bits_seq - copy bits from a skb_seq_state to kernel buffer
>> > > > > + * @st: source skb_seq_state
>> > > > > + * @offset: offset in source
>> > > > > + * @to: destination buffer
>> > > > > + * @len: number of bytes to copy
>> > > > > + *
>> > > > > + * Copy @len bytes from @offset bytes into the source @st to the destination
>> > > > > + * buffer @to. `offset` should increase (or be unchanged) with each subsequent
>> > > > > + * call to this function. If offset needs to decrease from the previous use `st`
>> > > > > + * should be reset first.
>> > > > > + *
>> > > > > + * Return: 0 on success or a negative error code on failure
>> > > > > + */
>> > > > > +static int skb_copy_bits_seq(struct skb_seq_state *st, int offset, void *to,
>> > > > > + int len)
>> > > >
>> > > > Probably belongs in net/core/skbuff.c, although I'm really not
>> > > > convinced copying data around is the right way to implement the type
>> > > > of packet splitting IPTFS does (which sounds a bit like a kind of
>> > > > GSO). And there are helpers in net/core/skbuff.c (such as
>> > > > pskb_carve/pskb_extract) that seem to do similar things to what you
>> > > > need here, without as much data copying.
>> > >
>> > > I don't have an issue with moving more general skb functionality
>> > > into skbuff.c; however, I do not want to gate IP-TFS on this change
>> > > to the general net infra, it is appropriate for a patchset of it's
>> > > own.
>> >
>> > If you need helpers that don't exist, it's part of your job to make
>> > the core changes that are required to implement the functionality.
>>
>> This is part of a new code protocol and feature addition and it's a single use.
>
> Of course the helper would be single use when it's introduced. You
> don't know if it will remain single use. And pskb_extract is single
> use, it's fine.
>
>> Another patchset can present this code to the general network
>> community to see if they think it *also* has value outside of
>> IPTFS. There is *no* reason to delay IPTFS on general network
>> infrastructure improvements. Please don't do this.
>
> Sorry, I don't think that's how it works.
>
>> > > Re copying: Let's be clear here, we are not always copying data,
>> > > there are sharing code paths as well; however, there are times when
>> > > it is the best (and even fastest) way to accomplish things (e.g.,
>> > > b/c the packet is small or the data is arranged in skbs in a way to
>> > > make sharing ridiculously complex and thus slow).
>> >
>> > I'm not finding the sharing code. You mean iptfs_first_should_copy
>> > returning false?
>>
>>
>> /* Try share then copy. */
>> if (fragwalk && skb_can_add_frags(newskb, fragwalk, data, copylen)) {
>> ...
>> leftover = skb_add_frags(newskb, fragwalk, data, copylen);
>> } else {
>> /* copy fragment data into newskb */
>> if (skb_copy_bits_seq(st, data, skb_put(newskb, copylen),
>> ...
>> }
>
> You're talking about reassembly now. This patch is fragmentation/TX.
[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 857 bytes --]
^ permalink raw reply [flat|nested] 40+ messages in thread* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-08 11:30 ` Christian Hopps
@ 2024-08-08 13:28 ` Sabrina Dubroca
2024-08-08 13:35 ` Christian Hopps
0 siblings, 1 reply; 40+ messages in thread
From: Sabrina Dubroca @ 2024-08-08 13:28 UTC (permalink / raw)
To: Christian Hopps; +Cc: devel, Steffen Klassert, netdev, Christian Hopps
2024-08-08, 07:30:13 -0400, Christian Hopps wrote:
>
> Sabrina Dubroca <sd@queasysnail.net> writes:
>
> > 2024-08-06, 04:54:53 -0400, Christian Hopps wrote:
> > >
> > > Sabrina Dubroca <sd@queasysnail.net> writes:
> > >
> > > > 2024-08-04, 22:33:05 -0400, Christian Hopps wrote:
> > > > > > > +/* 1) skb->head should be cache aligned.
> > > > > > > + * 2) when resv is for L2 headers (i.e., ethernet) we want the cacheline to
> > > > > > > + * start -16 from data.
> > > > > > > + * 3) when resv is for L3+L2 headers IOW skb->data points at the IPTFS payload
> > > > > > > + * we want data to be cache line aligned so all the pushed headers will be in
> > > > > > > + * another cacheline.
> > > > > > > + */
> > > > > > > +#define XFRM_IPTFS_MIN_L3HEADROOM 128
> > > > > > > +#define XFRM_IPTFS_MIN_L2HEADROOM (64 + 16)
> > > > > >
> > > > > > How did you pick those values?
> > > > >
> > > > > That's what the comment is talking to. When reserving space for L2 headers we
> > > > > pick 64 + 16 (a 2^(<=6) cacheline + 16 bytes so the the cacheline should start
> > > > > -16 from where skb->data will point at.
> > > >
> > > > Hard-coding the x86 cacheline size is not a good idea. And what's the
> > > > 16B for? You don't know that it's enough for the actual L2 headers.
> > >
> > > I am not hard coding the x86 cacheline. I am picking 64 as the largest cacheline that this is optimized for, it also works for smaller cachelines.
> >
> > At least use SMP_CACHE_BYTES then?
>
> Right, I have changed this work with L1_CACHE_BYTES value.
>
> > > 16B is to allow for the incredibly common 14B L2 header to fit.
> >
> > Why not use skb->dev->needed_headroom, like a bunch of tunnels are
> > already doing? No guessing required. ethernet is the most common, but
> > there's no reason to penalize other protocols when the information is
> > available.
>
> We can't use `skb->dev->needed_headroom`, b/c `skb->dev` is not
> correct for the new packets. `skb->dev` is from the received IPTFS
> tunnel packet. The skb being created here are the inner user packets
> leaving the tunnel, so they have an L3 header (thus why we are only
> making room for L2 header). They are being handed to gro receive and
> still have to be routed to their correct destination interface/dev.
You're talking about RX now. You're assuming the main use-case is an
IPsec GW that's going to send the decapped packets out on another
ethernet interface? (or at least, that's that's a use-case worth
optimizing for)
What about TX? Is skb->dev->needed_headroom also incorrect there?
Is iptfs_alloc_skb's l3resv argument equivalent to a RX/TX switch?
> 16 handles the general common case an ethernet device being the
> destination, if it's not correct after routing, nothing is broken,
> it just means that we may or may not achieve this maximal cache
> locality (but we still might e.g., if its destined to a GRE tunnel
> then we are looking at a bunch more headers so they and the existing
> L3 header will occupy 2 cachelines anyway). Again, this is a best
> effort thing.
Ok.
--
Sabrina
^ permalink raw reply [flat|nested] 40+ messages in thread
* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-08 13:28 ` Sabrina Dubroca
@ 2024-08-08 13:35 ` Christian Hopps
2024-08-08 14:01 ` Sabrina Dubroca
0 siblings, 1 reply; 40+ messages in thread
From: Christian Hopps @ 2024-08-08 13:35 UTC (permalink / raw)
To: Sabrina Dubroca; +Cc: Christian Hopps, devel, Steffen Klassert, netdev
> On Aug 8, 2024, at 09:28, Sabrina Dubroca <sd@queasysnail.net> wrote:
>
> 2024-08-08, 07:30:13 -0400, Christian Hopps wrote:
>>
>> Sabrina Dubroca <sd@queasysnail.net> writes:
>>
>>> 2024-08-06, 04:54:53 -0400, Christian Hopps wrote:
>>>>
>>>> Sabrina Dubroca <sd@queasysnail.net> writes:
>>>>
>>>>> 2024-08-04, 22:33:05 -0400, Christian Hopps wrote:
>>>>>>>> +/* 1) skb->head should be cache aligned.
>>>>>>>> + * 2) when resv is for L2 headers (i.e., ethernet) we want the cacheline to
>>>>>>>> + * start -16 from data.
>>>>>>>> + * 3) when resv is for L3+L2 headers IOW skb->data points at the IPTFS payload
>>>>>>>> + * we want data to be cache line aligned so all the pushed headers will be in
>>>>>>>> + * another cacheline.
>>>>>>>> + */
>>>>>>>> +#define XFRM_IPTFS_MIN_L3HEADROOM 128
>>>>>>>> +#define XFRM_IPTFS_MIN_L2HEADROOM (64 + 16)
>>>>>>>
>>>>>>> How did you pick those values?
>>>>>>
>>>>>> That's what the comment is talking to. When reserving space for L2 headers we
>>>>>> pick 64 + 16 (a 2^(<=6) cacheline + 16 bytes so the the cacheline should start
>>>>>> -16 from where skb->data will point at.
>>>>>
>>>>> Hard-coding the x86 cacheline size is not a good idea. And what's the
>>>>> 16B for? You don't know that it's enough for the actual L2 headers.
>>>>
>>>> I am not hard coding the x86 cacheline. I am picking 64 as the largest cacheline that this is optimized for, it also works for smaller cachelines.
>>>
>>> At least use SMP_CACHE_BYTES then?
>>
>> Right, I have changed this work with L1_CACHE_BYTES value.
>>
>>>> 16B is to allow for the incredibly common 14B L2 header to fit.
>>>
>>> Why not use skb->dev->needed_headroom, like a bunch of tunnels are
>>> already doing? No guessing required. ethernet is the most common, but
>>> there's no reason to penalize other protocols when the information is
>>> available.
>>
>> We can't use `skb->dev->needed_headroom`, b/c `skb->dev` is not
>> correct for the new packets. `skb->dev` is from the received IPTFS
>> tunnel packet. The skb being created here are the inner user packets
>> leaving the tunnel, so they have an L3 header (thus why we are only
>> making room for L2 header). They are being handed to gro receive and
>> still have to be routed to their correct destination interface/dev.
>
> You're talking about RX now. You're assuming the main use-case is an
> IPsec GW that's going to send the decapped packets out on another
> ethernet interface? (or at least, that's that's a use-case worth
> optimizing for)
>
> What about TX? Is skb->dev->needed_headroom also incorrect there?
>
> Is iptfs_alloc_skb's l3resv argument equivalent to a RX/TX switch?
Exactly right. When we are generating IPTFS tunnel packets we need to add all the L3+l2 headers, and in that case we pass l3resv = true.
Thanks,
Chris.
>
>> 16 handles the general common case an ethernet device being the
>> destination, if it's not correct after routing, nothing is broken,
>> it just means that we may or may not achieve this maximal cache
>> locality (but we still might e.g., if its destined to a GRE tunnel
>> then we are looking at a bunch more headers so they and the existing
>> L3 header will occupy 2 cachelines anyway). Again, this is a best
>> effort thing.
>
> Ok.
>
> --
> Sabrina
^ permalink raw reply [flat|nested] 40+ messages in thread
* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-08 13:35 ` Christian Hopps
@ 2024-08-08 14:01 ` Sabrina Dubroca
2024-08-08 21:42 ` Christian Hopps
0 siblings, 1 reply; 40+ messages in thread
From: Sabrina Dubroca @ 2024-08-08 14:01 UTC (permalink / raw)
To: Christian Hopps; +Cc: devel, Steffen Klassert, netdev
2024-08-08, 09:35:04 -0400, Christian Hopps wrote:
>
>
> > On Aug 8, 2024, at 09:28, Sabrina Dubroca <sd@queasysnail.net> wrote:
> >
> > 2024-08-08, 07:30:13 -0400, Christian Hopps wrote:
> >>
> >> Sabrina Dubroca <sd@queasysnail.net> writes:
> >>
> >>> 2024-08-06, 04:54:53 -0400, Christian Hopps wrote:
> >>>>
> >>>> Sabrina Dubroca <sd@queasysnail.net> writes:
> >>>>
> >>>>> 2024-08-04, 22:33:05 -0400, Christian Hopps wrote:
> >>>>>>>> +/* 1) skb->head should be cache aligned.
> >>>>>>>> + * 2) when resv is for L2 headers (i.e., ethernet) we want the cacheline to
> >>>>>>>> + * start -16 from data.
> >>>>>>>> + * 3) when resv is for L3+L2 headers IOW skb->data points at the IPTFS payload
> >>>>>>>> + * we want data to be cache line aligned so all the pushed headers will be in
> >>>>>>>> + * another cacheline.
> >>>>>>>> + */
> >>>>>>>> +#define XFRM_IPTFS_MIN_L3HEADROOM 128
> >>>>>>>> +#define XFRM_IPTFS_MIN_L2HEADROOM (64 + 16)
> >>>>>>>
> >>>>>>> How did you pick those values?
> >>>>>>
> >>>>>> That's what the comment is talking to. When reserving space for L2 headers we
> >>>>>> pick 64 + 16 (a 2^(<=6) cacheline + 16 bytes so the the cacheline should start
> >>>>>> -16 from where skb->data will point at.
> >>>>>
> >>>>> Hard-coding the x86 cacheline size is not a good idea. And what's the
> >>>>> 16B for? You don't know that it's enough for the actual L2 headers.
> >>>>
> >>>> I am not hard coding the x86 cacheline. I am picking 64 as the largest cacheline that this is optimized for, it also works for smaller cachelines.
> >>>
> >>> At least use SMP_CACHE_BYTES then?
> >>
> >> Right, I have changed this work with L1_CACHE_BYTES value.
> >>
> >>>> 16B is to allow for the incredibly common 14B L2 header to fit.
> >>>
> >>> Why not use skb->dev->needed_headroom, like a bunch of tunnels are
> >>> already doing? No guessing required. ethernet is the most common, but
> >>> there's no reason to penalize other protocols when the information is
> >>> available.
> >>
> >> We can't use `skb->dev->needed_headroom`, b/c `skb->dev` is not
> >> correct for the new packets. `skb->dev` is from the received IPTFS
> >> tunnel packet. The skb being created here are the inner user packets
> >> leaving the tunnel, so they have an L3 header (thus why we are only
> >> making room for L2 header). They are being handed to gro receive and
> >> still have to be routed to their correct destination interface/dev.
> >
> > You're talking about RX now. You're assuming the main use-case is an
> > IPsec GW that's going to send the decapped packets out on another
> > ethernet interface? (or at least, that's that's a use-case worth
> > optimizing for)
> >
> > What about TX? Is skb->dev->needed_headroom also incorrect there?
> >
> > Is iptfs_alloc_skb's l3resv argument equivalent to a RX/TX switch?
>
> Exactly right. When we are generating IPTFS tunnel packets we need
> to add all the L3+l2 headers, and in that case we pass l3resv =
> true.
Could you add a little comment alongside iptfs_alloc_skb? It would
help make sense of the sizes you're choosing and how they fit the use
of those skbs (something like "l3resv=true is used on TX, because we
need to reserve L2+L3 headers. On RX, we only need L2 headers because
[reason why we need L2 headers].").
And if skb->dev->needed_headroom is correct in the TX case, I'd still
prefer (skb->dev->needed_headroom + <some space for l3>) to a fixed 128.
--
Sabrina
^ permalink raw reply [flat|nested] 40+ messages in thread
* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-08 14:01 ` Sabrina Dubroca
@ 2024-08-08 21:42 ` Christian Hopps
0 siblings, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-08 21:42 UTC (permalink / raw)
To: Sabrina Dubroca; +Cc: Christian Hopps, devel, Steffen Klassert, netdev
> On Aug 8, 2024, at 10:01, Sabrina Dubroca <sd@queasysnail.net> wrote:
>
> 2024-08-08, 09:35:04 -0400, Christian Hopps wrote:
>>
>>
>>> On Aug 8, 2024, at 09:28, Sabrina Dubroca <sd@queasysnail.net> wrote:
>>>
>>> 2024-08-08, 07:30:13 -0400, Christian Hopps wrote:
>>>>
>>>> Sabrina Dubroca <sd@queasysnail.net> writes:
>>>>
>>>>> 2024-08-06, 04:54:53 -0400, Christian Hopps wrote:
>>>>>>
>>>>>> Sabrina Dubroca <sd@queasysnail.net> writes:
>>>>>>
>>>>>>> 2024-08-04, 22:33:05 -0400, Christian Hopps wrote:
>>>>>>>>>> +/* 1) skb->head should be cache aligned.
>>>>>>>>>> + * 2) when resv is for L2 headers (i.e., ethernet) we want the cacheline to
>>>>>>>>>> + * start -16 from data.
>>>>>>>>>> + * 3) when resv is for L3+L2 headers IOW skb->data points at the IPTFS payload
>>>>>>>>>> + * we want data to be cache line aligned so all the pushed headers will be in
>>>>>>>>>> + * another cacheline.
>>>>>>>>>> + */
>>>>>>>>>> +#define XFRM_IPTFS_MIN_L3HEADROOM 128
>>>>>>>>>> +#define XFRM_IPTFS_MIN_L2HEADROOM (64 + 16)
>>>>>>>>>
>>>>>>>>> How did you pick those values?
>>>>>>>>
>>>>>>>> That's what the comment is talking to. When reserving space for L2 headers we
>>>>>>>> pick 64 + 16 (a 2^(<=6) cacheline + 16 bytes so the the cacheline should start
>>>>>>>> -16 from where skb->data will point at.
>>>>>>>
>>>>>>> Hard-coding the x86 cacheline size is not a good idea. And what's the
>>>>>>> 16B for? You don't know that it's enough for the actual L2 headers.
>>>>>>
>>>>>> I am not hard coding the x86 cacheline. I am picking 64 as the largest cacheline that this is optimized for, it also works for smaller cachelines.
>>>>>
>>>>> At least use SMP_CACHE_BYTES then?
>>>>
>>>> Right, I have changed this work with L1_CACHE_BYTES value.
>>>>
>>>>>> 16B is to allow for the incredibly common 14B L2 header to fit.
>>>>>
>>>>> Why not use skb->dev->needed_headroom, like a bunch of tunnels are
>>>>> already doing? No guessing required. ethernet is the most common, but
>>>>> there's no reason to penalize other protocols when the information is
>>>>> available.
>>>>
>>>> We can't use `skb->dev->needed_headroom`, b/c `skb->dev` is not
>>>> correct for the new packets. `skb->dev` is from the received IPTFS
>>>> tunnel packet. The skb being created here are the inner user packets
>>>> leaving the tunnel, so they have an L3 header (thus why we are only
>>>> making room for L2 header). They are being handed to gro receive and
>>>> still have to be routed to their correct destination interface/dev.
>>>
>>> You're talking about RX now. You're assuming the main use-case is an
>>> IPsec GW that's going to send the decapped packets out on another
>>> ethernet interface? (or at least, that's that's a use-case worth
>>> optimizing for)
>>>
>>> What about TX? Is skb->dev->needed_headroom also incorrect there?
>>>
>>> Is iptfs_alloc_skb's l3resv argument equivalent to a RX/TX switch?
>>
>> Exactly right. When we are generating IPTFS tunnel packets we need
>> to add all the L3+l2 headers, and in that case we pass l3resv =
>> true.
>
> Could you add a little comment alongside iptfs_alloc_skb? It would
> help make sense of the sizes you're choosing and how they fit the use
> of those skbs (something like "l3resv=true is used on TX, because we
> need to reserve L2+L3 headers. On RX, we only need L2 headers because
> [reason why we need L2 headers].").
Sure.
> And if skb->dev->needed_headroom is correct in the TX case, I'd still
> prefer (skb->dev->needed_headroom + <some space for l3>) to a fixed 128.
So dev->needed_headroom is defined as possible extra needed headroom.For ethernet it is 12. It's not what we want.
The actual MAC size value in this case is: dev->hard_header_len which is 14..
(gdb) p st->root_skb
$2 = (struct sk_buff *) 0xffff888012969a00
(gdb) p (struct dst_entry *)$2->_skb_refdst
$3 = (struct dst_entry *) 0xffff888012ef7180
(gdb) p $3->deb
$5 = (struct net_device *) 0xffff88800dca8000
(gdb) p $5->needed_headroom
$6 = 12
(gdb) p $5->hard_header_len
$7 = 14
(gdb) p $5->min_header_len
$8 = 14 '\016'
We also have access to the IPTFS required header space by looking in the tpl->dst...
# L3 header space requirement for xfrm
(gdb) p $3->header_len
$4 = 40
which in this case is 40 == IP (20) + ESP (8) + GCM-IV (8) + IPTFS (4)
So what I think you're looking for is this:
struct dst_entry *dst = skb_dst(tpl);
resv = LL_RESERVED_SPACE(dst->dev) + dst->header_len;
resv = L1_CACHE_ALIGN(resv);
FWIW (not that much) this is 128 in the ethernet dev case :)
# LL_RESERVED_SPACE()
(gdb) p (14 + 12 + 16) & ~15
$9 = 32
# above + dst->header_len
(gdb) p 40 + 32
$10 = 72
# aligned to L1_CACHE_BYTES
(gdb) p (72 + 64) & ~63
$12 = 128
Thanks,
Chris.
>
> --
> Sabrina
^ permalink raw reply [flat|nested] 40+ messages in thread
* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-06 8:54 ` Christian Hopps
2024-08-06 10:03 ` Florian Westphal
2024-08-06 11:05 ` Sabrina Dubroca
@ 2024-08-06 11:07 ` Steffen Klassert
2024-08-07 16:23 ` Christian Hopps
2 siblings, 1 reply; 40+ messages in thread
From: Steffen Klassert @ 2024-08-06 11:07 UTC (permalink / raw)
To: Christian Hopps; +Cc: Sabrina Dubroca, devel, netdev, Christian Hopps
On Tue, Aug 06, 2024 at 04:54:53AM -0400, Christian Hopps wrote:
>
> Sabrina Dubroca <sd@queasysnail.net> writes:
>
> > 2024-08-04, 22:33:05 -0400, Christian Hopps wrote:
> > > > > +/* 1) skb->head should be cache aligned.
> > > > > + * 2) when resv is for L2 headers (i.e., ethernet) we want the cacheline to
> > > > > + * start -16 from data.
> > > > > + * 3) when resv is for L3+L2 headers IOW skb->data points at the IPTFS payload
> > > > > + * we want data to be cache line aligned so all the pushed headers will be in
> > > > > + * another cacheline.
> > > > > + */
> > > > > +#define XFRM_IPTFS_MIN_L3HEADROOM 128
> > > > > +#define XFRM_IPTFS_MIN_L2HEADROOM (64 + 16)
> > > >
> > > > How did you pick those values?
> > >
> > > That's what the comment is talking to. When reserving space for L2 headers we
> > > pick 64 + 16 (a 2^(<=6) cacheline + 16 bytes so the the cacheline should start
> > > -16 from where skb->data will point at.
> >
> > Hard-coding the x86 cacheline size is not a good idea. And what's the
> > 16B for? You don't know that it's enough for the actual L2 headers.
>
> I am not hard coding the x86 cacheline. I am picking 64 as the largest cacheline that this is optimized for, it also works for smaller cachelines.
Maybe use L1_CACHE_BYTES instead of 64? This will give you
the actual size of the cacheline.
> > > > > +
> > > > > + skb_reserve(skb, resv);
> > > > > +
> > > > > + /* We do not want any of the tpl->headers copied over, so we do
> > > > > + * not use `skb_copy_header()`.
> > > > > + */
> > > >
> > > > This is a bit of a bad sign for the implementation. It also worries
> > > > me, as this may not be updated when changes are made to
> > > > __copy_skb_header().
> > > > (c/p'd from v1 review since this was still not answered)
> > >
> > > I don't agree that this is a bad design at all, I'm curious what you think a good design to be.
> >
> > Strange skb manipulations hiding in a protocol module is not good
> > design.
>
> It's a fragmentation and aggregation protocol, it's needs work with skbs by design. It's literally the function of the protocol to manipulate packet content.
>
> I would appreciate it if you could provide technical reasons to justify referring to things as "bad" or "strange" -- it's not helpful otherwise.
>
> > c/p bits of core code into a module (where they will never get fixed
> > up when the core code gets updated) is always a bad idea.
>
> I need some values from the SKB, so I copy them -- it's that simple.
>
> > > I did specifically state why we are not re-using
> > > skb_copy_header(). The functionality is different. We are not trying
> > > to make a copy of an skb we are using an skb as a template for new
> > > skbs.
> >
> > I saw that. That doesn't mean it's a good thing to do.
>
> Please suggest an alternative.
Maybe create a helper like this:
void ___copy_skb_header(struct sk_buff *new, const struct sk_buff *old)
{
new->tstamp = old->tstamp;
/* We do not copy old->sk */
new->dev = old->dev;
memcpy(new->cb, old->cb, sizeof(old->cb));
skb_dst_copy(new, old);
__skb_ext_copy(new, old);
__nf_copy(new, old, false);
}
and change __copy_skb_header() to use this too. That way it gets
updated whenever something changes here.
It also might make sense to split out the generic infrastructure changes
into a separate pachset wih netdev maintainers Cced on. That would make
the changes more visible.
^ permalink raw reply [flat|nested] 40+ messages in thread* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-06 11:07 ` Steffen Klassert
@ 2024-08-07 16:23 ` Christian Hopps
0 siblings, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-07 16:23 UTC (permalink / raw)
To: Steffen Klassert
Cc: Christian Hopps, Sabrina Dubroca, devel, netdev, Christian Hopps
[-- Attachment #1: Type: text/plain, Size: 4262 bytes --]
Steffen Klassert <steffen.klassert@secunet.com> writes:
> On Tue, Aug 06, 2024 at 04:54:53AM -0400, Christian Hopps wrote:
>>
>> Sabrina Dubroca <sd@queasysnail.net> writes:
>>
>> > 2024-08-04, 22:33:05 -0400, Christian Hopps wrote:
>> > > > > +/* 1) skb->head should be cache aligned.
>> > > > > + * 2) when resv is for L2 headers (i.e., ethernet) we want the cacheline to
>> > > > > + * start -16 from data.
>> > > > > + * 3) when resv is for L3+L2 headers IOW skb->data points at the IPTFS payload
>> > > > > + * we want data to be cache line aligned so all the pushed headers will be in
>> > > > > + * another cacheline.
>> > > > > + */
>> > > > > +#define XFRM_IPTFS_MIN_L3HEADROOM 128
>> > > > > +#define XFRM_IPTFS_MIN_L2HEADROOM (64 + 16)
>> > > >
>> > > > How did you pick those values?
>> > >
>> > > That's what the comment is talking to. When reserving space for L2 headers we
>> > > pick 64 + 16 (a 2^(<=6) cacheline + 16 bytes so the the cacheline should start
>> > > -16 from where skb->data will point at.
>> >
>> > Hard-coding the x86 cacheline size is not a good idea. And what's the
>> > 16B for? You don't know that it's enough for the actual L2 headers.
>>
>> I am not hard coding the x86 cacheline. I am picking 64 as the largest cacheline that this is optimized for, it also works for smaller cachelines.
>
> Maybe use L1_CACHE_BYTES instead of 64? This will give you
> the actual size of the cacheline.
Yes, although a bit more than just a swap:
#define XFRM_IPTFS_MIN_L2HEADROOM (L1_CACHE_BYTES > 64 ? 64 : 64 + 16)
Here's the new comment text which explains this:
/*
* L2 Header resv: Arrange for cacheline to start at skb->data - 16 to keep the
* to-be-pushed L2 header in the same cacheline as resulting `skb->data` (i.e.,
* the L3 header). If cacheline size is > 64 then skb->data + pushed L2 will all
* be in a single cacheline if we simply reserve 64 bytes.
*/
I'm simply protecting against some new arch that decides to have 256 byte cacheline since we do not need to reserve 256 bytes for L2 headers.
>> > > > > + skb_reserve(skb, resv);
>> > > > > +
>> > > > > + /* We do not want any of the tpl->headers copied over, so we do
>> > > > > + * not use `skb_copy_header()`.
>> > > > > + */
>> > > >
>> > > > This is a bit of a bad sign for the implementation. It also worries
>> > > > me, as this may not be updated when changes are made to
>> > > > __copy_skb_header().
>> > > > (c/p'd from v1 review since this was still not answered)
>> > >
>> > > I don't agree that this is a bad design at all, I'm curious what you think a good design to be.
>> >
>> > Strange skb manipulations hiding in a protocol module is not good
>> > design.
>>
>> It's a fragmentation and aggregation protocol, it's needs work with skbs by design. It's literally the function of the protocol to manipulate packet content.
>>
>> I would appreciate it if you could provide technical reasons to justify referring to things as "bad" or "strange" -- it's not helpful otherwise.
>>
>> > c/p bits of core code into a module (where they will never get fixed
>> > up when the core code gets updated) is always a bad idea.
>>
>> I need some values from the SKB, so I copy them -- it's that simple.
>>
>> > > I did specifically state why we are not re-using
>> > > skb_copy_header(). The functionality is different. We are not trying
>> > > to make a copy of an skb we are using an skb as a template for new
>> > > skbs.
>> >
>> > I saw that. That doesn't mean it's a good thing to do.
>>
>> Please suggest an alternative.
>
> Maybe create a helper like this:
>
> void ___copy_skb_header(struct sk_buff *new, const struct sk_buff *old)
> {
> new->tstamp = old->tstamp;
> /* We do not copy old->sk */
> new->dev = old->dev;
> memcpy(new->cb, old->cb, sizeof(old->cb));
> skb_dst_copy(new, old);
> __skb_ext_copy(new, old);
> __nf_copy(new, old, false);
> }
>
> and change __copy_skb_header() to use this too. That way it gets
> updated whenever something changes here.
Ok.
Thanks,
Chris.
> It also might make sense to split out the generic infrastructure changes
> into a separate pachset wih netdev maintainers Cced on. That would make
> the changes more visible.
[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 857 bytes --]
^ permalink raw reply [flat|nested] 40+ messages in thread
* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-04 22:25 ` Sabrina Dubroca
2024-08-05 2:33 ` Christian Hopps
@ 2024-08-06 11:32 ` Steffen Klassert
2024-08-07 19:40 ` Christian Hopps
1 sibling, 1 reply; 40+ messages in thread
From: Steffen Klassert @ 2024-08-06 11:32 UTC (permalink / raw)
To: Sabrina Dubroca; +Cc: Christian Hopps, devel, netdev, Christian Hopps
On Mon, Aug 05, 2024 at 12:25:57AM +0200, Sabrina Dubroca wrote:
>
> > +/**
> > + * skb_copy_bits_seq - copy bits from a skb_seq_state to kernel buffer
> > + * @st: source skb_seq_state
> > + * @offset: offset in source
> > + * @to: destination buffer
> > + * @len: number of bytes to copy
> > + *
> > + * Copy @len bytes from @offset bytes into the source @st to the destination
> > + * buffer @to. `offset` should increase (or be unchanged) with each subsequent
> > + * call to this function. If offset needs to decrease from the previous use `st`
> > + * should be reset first.
> > + *
> > + * Return: 0 on success or a negative error code on failure
> > + */
> > +static int skb_copy_bits_seq(struct skb_seq_state *st, int offset, void *to,
> > + int len)
>
> Probably belongs in net/core/skbuff.c, although I'm really not
> convinced copying data around is the right way to implement the type
> of packet splitting IPTFS does (which sounds a bit like a kind of
> GSO).
I tried to come up with a 'GSO like' variant of this when I did the
initial review last year at the IPsec workshop. But it turned out
that things will get even more complicated as they are now.
We did some performance tests and it was quite compareable to
tunnel mode, so for a first implementation I'd be ok with the
copy variant.
> And there are helpers in net/core/skbuff.c (such as
> pskb_carve/pskb_extract) that seem to do similar things to what you
> need here, without as much data copying.
In case we have helpers that will fit here, we should use them of
course.
^ permalink raw reply [flat|nested] 40+ messages in thread
* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-06 11:32 ` Steffen Klassert
@ 2024-08-07 19:40 ` Christian Hopps
2024-08-08 9:26 ` Sabrina Dubroca
0 siblings, 1 reply; 40+ messages in thread
From: Christian Hopps @ 2024-08-07 19:40 UTC (permalink / raw)
To: Steffen Klassert
Cc: Sabrina Dubroca, Christian Hopps, devel, netdev, Christian Hopps
Steffen Klassert <steffen.klassert@secunet.com> writes:
> On Mon, Aug 05, 2024 at 12:25:57AM +0200, Sabrina Dubroca wrote:
>>
>> > +/**
>> > + * skb_copy_bits_seq - copy bits from a skb_seq_state to kernel buffer
>> > + * @st: source skb_seq_state
>> > + * @offset: offset in source
>> > + * @to: destination buffer
>> > + * @len: number of bytes to copy
>> > + *
>> > + * Copy @len bytes from @offset bytes into the source @st to the destination
>> > + * buffer @to. `offset` should increase (or be unchanged) with each subsequent
>> > + * call to this function. If offset needs to decrease from the previous use `st`
>> > + * should be reset first.
>> > + *
>> > + * Return: 0 on success or a negative error code on failure
>> > + */
>> > +static int skb_copy_bits_seq(struct skb_seq_state *st, int offset, void *to,
>> > + int len)
>>
>> Probably belongs in net/core/skbuff.c, although I'm really not
>> convinced copying data around is the right way to implement the type
>> of packet splitting IPTFS does (which sounds a bit like a kind of
>> GSO).
>
> I tried to come up with a 'GSO like' variant of this when I did the
> initial review last year at the IPsec workshop. But it turned out
> that things will get even more complicated as they are now.
> We did some performance tests and it was quite compareable to
> tunnel mode, so for a first implementation I'd be ok with the
> copy variant.
>
>
>> And there are helpers in net/core/skbuff.c (such as
>> pskb_carve/pskb_extract) that seem to do similar things to what you
>> need here, without as much data copying.
>
> In case we have helpers that will fit here, we should use them of
> course.
FWIW, The reason I didn't use pskb_extract() rather than the simple iptfs_copy_create_frag() is because pskb_extract uses skb_clone on the original skb then pskb_carve() to narrow the (copied) data pointers to a subset of the original. The new skb data is read-only which does not work for us.
Each of these new skbs are IP-TFS tunnel packets and as such we need to push and write IPTFS+ESP+IP+Ethernet headers on them. In order to make pskb_extract()s skbs writable we would have to allocate new buffer space and copy the data turning them into a writeable skb buffer, and now we're doing something more complex and more cpu intensive to arrive back to what iptfs_copy_create_frag() did simply and straight-forwardly to begin with.
Thanks,
Chris.
^ permalink raw reply [flat|nested] 40+ messages in thread
* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-07 19:40 ` Christian Hopps
@ 2024-08-08 9:26 ` Sabrina Dubroca
2024-08-08 11:23 ` Christian Hopps
0 siblings, 1 reply; 40+ messages in thread
From: Sabrina Dubroca @ 2024-08-08 9:26 UTC (permalink / raw)
To: Christian Hopps; +Cc: Steffen Klassert, devel, netdev, Christian Hopps
2024-08-07, 15:40:14 -0400, Christian Hopps wrote:
>
> Steffen Klassert <steffen.klassert@secunet.com> writes:
>
> > On Mon, Aug 05, 2024 at 12:25:57AM +0200, Sabrina Dubroca wrote:
> > >
> > > > +/**
> > > > + * skb_copy_bits_seq - copy bits from a skb_seq_state to kernel buffer
> > > > + * @st: source skb_seq_state
> > > > + * @offset: offset in source
> > > > + * @to: destination buffer
> > > > + * @len: number of bytes to copy
> > > > + *
> > > > + * Copy @len bytes from @offset bytes into the source @st to the destination
> > > > + * buffer @to. `offset` should increase (or be unchanged) with each subsequent
> > > > + * call to this function. If offset needs to decrease from the previous use `st`
> > > > + * should be reset first.
> > > > + *
> > > > + * Return: 0 on success or a negative error code on failure
> > > > + */
> > > > +static int skb_copy_bits_seq(struct skb_seq_state *st, int offset, void *to,
> > > > + int len)
> > >
> > > Probably belongs in net/core/skbuff.c, although I'm really not
> > > convinced copying data around is the right way to implement the type
> > > of packet splitting IPTFS does (which sounds a bit like a kind of
> > > GSO).
> >
> > I tried to come up with a 'GSO like' variant of this when I did the
> > initial review last year at the IPsec workshop. But it turned out
> > that things will get even more complicated as they are now.
> > We did some performance tests and it was quite compareable to
> > tunnel mode, so for a first implementation I'd be ok with the
> > copy variant.
Ok.
> > > And there are helpers in net/core/skbuff.c (such as
> > > pskb_carve/pskb_extract) that seem to do similar things to what you
> > > need here, without as much data copying.
> >
> > In case we have helpers that will fit here, we should use them of
> > course.
>
> FWIW, The reason I didn't use pskb_extract() rather than the simple
> iptfs_copy_create_frag() is because pskb_extract uses skb_clone on
> the original skb then pskb_carve() to narrow the (copied) data
> pointers to a subset of the original. The new skb data is read-only
> which does not work for us.
>
> Each of these new skbs are IP-TFS tunnel packets and as such we need
> to push and write IPTFS+ESP+IP+Ethernet headers on them. In order to
> make pskb_extract()s skbs writable we would have to allocate new
> buffer space and copy the data turning them into a writeable skb
> buffer, and now we're doing something more complex and more cpu
> intensive to arrive back to what iptfs_copy_create_frag() did simply
> and straight-forwardly to begin with.
That only requires the header to be writeable, not the full packet,
right? I doubt it would actually be more cpu/memory intensive.
--
Sabrina
^ permalink raw reply [flat|nested] 40+ messages in thread
* Re: [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets
2024-08-08 9:26 ` Sabrina Dubroca
@ 2024-08-08 11:23 ` Christian Hopps
0 siblings, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-08 11:23 UTC (permalink / raw)
To: Sabrina Dubroca
Cc: Christian Hopps, Steffen Klassert, devel, netdev, Christian Hopps
[-- Attachment #1: Type: text/plain, Size: 3032 bytes --]
Sabrina Dubroca <sd@queasysnail.net> writes:
> 2024-08-07, 15:40:14 -0400, Christian Hopps wrote:
>>
>> Steffen Klassert <steffen.klassert@secunet.com> writes:
>>
>> > On Mon, Aug 05, 2024 at 12:25:57AM +0200, Sabrina Dubroca wrote:
>> > >
>> > > > +/**
>> > > > + * skb_copy_bits_seq - copy bits from a skb_seq_state to kernel buffer
>> > > > + * @st: source skb_seq_state
>> > > > + * @offset: offset in source
>> > > > + * @to: destination buffer
>> > > > + * @len: number of bytes to copy
>> > > > + *
>> > > > + * Copy @len bytes from @offset bytes into the source @st to the destination
>> > > > + * buffer @to. `offset` should increase (or be unchanged) with each subsequent
>> > > > + * call to this function. If offset needs to decrease from the previous use `st`
>> > > > + * should be reset first.
>> > > > + *
>> > > > + * Return: 0 on success or a negative error code on failure
>> > > > + */
>> > > > +static int skb_copy_bits_seq(struct skb_seq_state *st, int offset, void *to,
>> > > > + int len)
>> > >
>> > > Probably belongs in net/core/skbuff.c, although I'm really not
>> > > convinced copying data around is the right way to implement the type
>> > > of packet splitting IPTFS does (which sounds a bit like a kind of
>> > > GSO).
>> >
>> > I tried to come up with a 'GSO like' variant of this when I did the
>> > initial review last year at the IPsec workshop. But it turned out
>> > that things will get even more complicated as they are now.
>> > We did some performance tests and it was quite compareable to
>> > tunnel mode, so for a first implementation I'd be ok with the
>> > copy variant.
>
> Ok.
>
>> > > And there are helpers in net/core/skbuff.c (such as
>> > > pskb_carve/pskb_extract) that seem to do similar things to what you
>> > > need here, without as much data copying.
>> >
>> > In case we have helpers that will fit here, we should use them of
>> > course.
>>
>> FWIW, The reason I didn't use pskb_extract() rather than the simple
>> iptfs_copy_create_frag() is because pskb_extract uses skb_clone on
>> the original skb then pskb_carve() to narrow the (copied) data
>> pointers to a subset of the original. The new skb data is read-only
>> which does not work for us.
>>
>> Each of these new skbs are IP-TFS tunnel packets and as such we need
>> to push and write IPTFS+ESP+IP+Ethernet headers on them. In order to
>> make pskb_extract()s skbs writable we would have to allocate new
>> buffer space and copy the data turning them into a writeable skb
>> buffer, and now we're doing something more complex and more cpu
>> intensive to arrive back to what iptfs_copy_create_frag() did simply
>> and straight-forwardly to begin with.
>
> That only requires the header to be writeable, not the full packet,
> right? I doubt it would actually be more cpu/memory intensive.
pskb_extract() function leaves `skb->head = skb->data` there is no room left to push headers that we need to push.
FWIW, pskb_extract() is used in one place for TCP; it's not suited for IPTFS.
Thanks,
Chris.
[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 857 bytes --]
^ permalink raw reply [flat|nested] 40+ messages in thread
* [PATCH ipsec-next v8 11/16] xfrm: iptfs: add basic receive packet (tunnel egress) handling
2024-08-04 20:33 [PATCH ipsec-next v8 00/16] Add IP-TFS mode to xfrm Christian Hopps
` (9 preceding siblings ...)
2024-08-04 20:33 ` [PATCH ipsec-next v8 10/16] xfrm: iptfs: add fragmenting of larger than MTU user packets Christian Hopps
@ 2024-08-04 20:33 ` Christian Hopps
2024-08-04 20:33 ` [PATCH ipsec-next v8 12/16] xfrm: iptfs: handle received fragmented inner packets Christian Hopps
` (4 subsequent siblings)
15 siblings, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-04 20:33 UTC (permalink / raw)
To: devel; +Cc: Steffen Klassert, netdev, Christian Hopps, Christian Hopps
From: Christian Hopps <chopps@labn.net>
Add handling of packets received from the tunnel. This implements
tunnel egress functionality.
Signed-off-by: Christian Hopps <chopps@labn.net>
---
net/xfrm/xfrm_iptfs.c | 268 ++++++++++++++++++++++++++++++++++++++++++
1 file changed, 268 insertions(+)
diff --git a/net/xfrm/xfrm_iptfs.c b/net/xfrm/xfrm_iptfs.c
index 38735e2d64c3..ea0f47af345d 100644
--- a/net/xfrm/xfrm_iptfs.c
+++ b/net/xfrm/xfrm_iptfs.c
@@ -20,6 +20,10 @@
#include "xfrm_inout.h"
+/* IPTFS encap (header) values. */
+#define IPTFS_SUBTYPE_BASIC 0
+#define IPTFS_SUBTYPE_CC 1
+
/* ------------------------------------------------ */
/* IPTFS default SA values (tunnel ingress/dir-out) */
/* ------------------------------------------------ */
@@ -214,6 +218,269 @@ static int skb_copy_bits_seq(struct skb_seq_state *st, int offset, void *to,
}
}
+/* ================================== */
+/* IPTFS Receiving (egress) Functions */
+/* ================================== */
+
+/**
+ * iptfs_pskb_extract_seq() - Create and load data into a new sk_buff.
+ * @skblen: the total data size for `skb`.
+ * @st: The source for the rest of the data to copy into `skb`.
+ * @off: The offset into @st to copy data from.
+ * @len: The length of data to copy from @st into `skb`. This must be <=
+ * @skblen.
+ *
+ * Create a new sk_buff `skb` with @skblen of packet data space. If non-zero,
+ * copy @rlen bytes of @runt into `skb`. Then using seq functions copy @len
+ * bytes from @st into `skb` starting from @off.
+ *
+ * It is an error for @len to be greater than the amount of data left in @st.
+ *
+ * Return: The newly allocated sk_buff `skb` or NULL if an error occurs.
+ */
+static struct sk_buff *
+iptfs_pskb_extract_seq(u32 skblen, struct skb_seq_state *st, u32 off, int len)
+{
+ struct sk_buff *skb = iptfs_alloc_skb(st->root_skb, skblen, false);
+
+ if (!skb)
+ return NULL;
+ if (skb_copy_bits_seq(st, off, skb_put(skb, len), len)) {
+ XFRM_INC_STATS(dev_net(st->root_skb->dev),
+ LINUX_MIB_XFRMINERROR);
+ kfree_skb(skb);
+ return NULL;
+ }
+ return skb;
+}
+
+/**
+ * iptfs_complete_inner_skb() - finish preparing the inner packet for gro recv.
+ * @x: xfrm state
+ * @skb: the inner packet
+ *
+ * Finish the standard xfrm processing on the inner packet prior to sending back
+ * through gro_cells_receive. We do this separately b/c we are building a list
+ * of packets in the hopes that one day a list will be taken by
+ * xfrm_input.
+ */
+static void iptfs_complete_inner_skb(struct xfrm_state *x, struct sk_buff *skb)
+{
+ skb_reset_network_header(skb);
+
+ /* The packet is going back through gro_cells_receive no need to
+ * set this.
+ */
+ skb_reset_transport_header(skb);
+
+ /* Packet already has checksum value set. */
+ skb->ip_summed = CHECKSUM_NONE;
+
+ /* Our skb will contain the header data copied when this outer packet
+ * which contained the start of this inner packet. This is true
+ * when we allocate a new skb as well as when we reuse the existing skb.
+ */
+ if (ip_hdr(skb)->version == 0x4) {
+ struct iphdr *iph = ip_hdr(skb);
+
+ if (x->props.flags & XFRM_STATE_DECAP_DSCP)
+ ipv4_copy_dscp(XFRM_MODE_SKB_CB(skb)->tos, iph);
+ if (!(x->props.flags & XFRM_STATE_NOECN))
+ if (INET_ECN_is_ce(XFRM_MODE_SKB_CB(skb)->tos))
+ IP_ECN_set_ce(iph);
+
+ skb->protocol = htons(ETH_P_IP);
+ } else {
+ struct ipv6hdr *iph = ipv6_hdr(skb);
+
+ if (x->props.flags & XFRM_STATE_DECAP_DSCP)
+ ipv6_copy_dscp(XFRM_MODE_SKB_CB(skb)->tos, iph);
+ if (!(x->props.flags & XFRM_STATE_NOECN))
+ if (INET_ECN_is_ce(XFRM_MODE_SKB_CB(skb)->tos))
+ IP6_ECN_set_ce(skb, iph);
+
+ skb->protocol = htons(ETH_P_IPV6);
+ }
+}
+
+/**
+ * iptfs_input() - handle receipt of iptfs payload
+ * @x: xfrm state
+ * @skb: the packet
+ *
+ * Process the IPTFS payload in `skb` and consume it afterwards.
+ *
+ * Returns 0.
+ */
+static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
+{
+ u8 hbytes[sizeof(struct ipv6hdr)];
+ struct ip_iptfs_cc_hdr iptcch;
+ struct skb_seq_state skbseq;
+ struct list_head sublist; /* rename this it's just a list */
+ struct sk_buff *first_skb, *next;
+ const unsigned char *old_mac;
+ struct ip_iptfs_hdr *ipth;
+ struct iphdr *iph;
+ struct net *net;
+ u32 remaining, iplen, iphlen, data, tail;
+ u32 blkoff;
+
+ net = xs_net(x);
+ first_skb = NULL;
+
+ /* Large enough to hold both types of header */
+ ipth = (struct ip_iptfs_hdr *)&iptcch;
+
+ /* Save the old mac header if set */
+ old_mac = skb_mac_header_was_set(skb) ? skb_mac_header(skb) : NULL;
+
+ skb_prepare_seq_read(skb, 0, skb->len, &skbseq);
+
+ /* Get the IPTFS header and validate it */
+
+ if (skb_copy_bits_seq(&skbseq, 0, ipth, sizeof(*ipth))) {
+ XFRM_INC_STATS(net, LINUX_MIB_XFRMINBUFFERERROR);
+ goto done;
+ }
+ data = sizeof(*ipth);
+
+ /* Set data past the basic header */
+ if (ipth->subtype == IPTFS_SUBTYPE_CC) {
+ /* Copy the rest of the CC header */
+ remaining = sizeof(iptcch) - sizeof(*ipth);
+ if (skb_copy_bits_seq(&skbseq, data, ipth + 1, remaining)) {
+ XFRM_INC_STATS(net, LINUX_MIB_XFRMINBUFFERERROR);
+ goto done;
+ }
+ data += remaining;
+ } else if (ipth->subtype != IPTFS_SUBTYPE_BASIC) {
+ XFRM_INC_STATS(net, LINUX_MIB_XFRMINHDRERROR);
+ goto done;
+ }
+
+ if (ipth->flags != 0) {
+ XFRM_INC_STATS(net, LINUX_MIB_XFRMINHDRERROR);
+ goto done;
+ }
+
+ INIT_LIST_HEAD(&sublist);
+
+ /* Fragment handling in following commits */
+ blkoff = ntohs(ipth->block_offset);
+ data += blkoff;
+
+ /* New packets */
+ tail = skb->len;
+ while (data < tail) {
+ __be16 protocol = 0;
+
+ /* Gather information on the next data block.
+ * `data` points to the start of the data block.
+ */
+ remaining = tail - data;
+
+ /* try and copy enough bytes to read length from ipv4/ipv6 */
+ iphlen = min_t(u32, remaining, 6);
+ if (skb_copy_bits_seq(&skbseq, data, hbytes, iphlen)) {
+ XFRM_INC_STATS(net, LINUX_MIB_XFRMINBUFFERERROR);
+ goto done;
+ }
+
+ iph = (struct iphdr *)hbytes;
+ if (iph->version == 0x4) {
+ /* must have at least tot_len field present */
+ if (remaining < 4)
+ break;
+
+ iplen = be16_to_cpu(iph->tot_len);
+ iphlen = iph->ihl << 2;
+ protocol = cpu_to_be16(ETH_P_IP);
+ XFRM_MODE_SKB_CB(skbseq.root_skb)->tos = iph->tos;
+ } else if (iph->version == 0x6) {
+ /* must have at least payload_len field present */
+ if (remaining < 6)
+ break;
+
+ iplen = be16_to_cpu(((struct ipv6hdr *)hbytes)->payload_len);
+ iplen += sizeof(struct ipv6hdr);
+ iphlen = sizeof(struct ipv6hdr);
+ protocol = cpu_to_be16(ETH_P_IPV6);
+ XFRM_MODE_SKB_CB(skbseq.root_skb)->tos =
+ ipv6_get_dsfield((struct ipv6hdr *)iph);
+ } else if (iph->version == 0x0) {
+ /* pad */
+ break;
+ } else {
+ XFRM_INC_STATS(net, LINUX_MIB_XFRMINBUFFERERROR);
+ goto done;
+ }
+
+ if (unlikely(skbseq.stepped_offset)) {
+ /* We need to reset our seq read, it can't backup at
+ * this point.
+ */
+ struct sk_buff *save = skbseq.root_skb;
+
+ skb_abort_seq_read(&skbseq);
+ skb_prepare_seq_read(save, data, tail, &skbseq);
+ }
+
+ if (!first_skb)
+ first_skb = skb;
+
+ /* Fragment handling in following commits */
+ if (iplen > remaining)
+ break;
+
+ skb = iptfs_pskb_extract_seq(iplen, &skbseq, data, iplen);
+ if (!skb) {
+ /* skip to next packet or done */
+ data += iplen;
+ continue;
+ }
+
+ skb->protocol = protocol;
+ if (old_mac) {
+ /* rebuild the mac header */
+ skb_set_mac_header(skb, -first_skb->mac_len);
+ memcpy(skb_mac_header(skb), old_mac,
+ first_skb->mac_len);
+ eth_hdr(skb)->h_proto = skb->protocol;
+ }
+
+ data += iplen;
+ iptfs_complete_inner_skb(x, skb);
+ list_add_tail(&skb->list, &sublist);
+ }
+
+ /* Send the packets! */
+ list_for_each_entry_safe(skb, next, &sublist, list) {
+ skb_list_del_init(skb);
+ if (xfrm_input(skb, 0, 0, -2))
+ kfree_skb(skb);
+ }
+
+done:
+ skb = skbseq.root_skb;
+ skb_abort_seq_read(&skbseq);
+
+ if (first_skb) {
+ consume_skb(first_skb);
+ } else {
+ /* skb is the original passed in skb, but we didn't get far
+ * enough to process it as the first_skb.
+ */
+ kfree_skb(skb);
+ }
+
+ /* We always have dealt with the input SKB, either we are re-using it,
+ * or we have freed it. Return EINPROGRESS so that xfrm_input stops
+ * processing it.
+ */
+ return -EINPROGRESS;
+}
+
/* ================================= */
/* IPTFS Sending (ingress) Functions */
/* ================================= */
@@ -1178,6 +1445,7 @@ static const struct xfrm_mode_cbs iptfs_mode_cbs = {
.sa_len = iptfs_sa_len,
.clone = iptfs_clone,
.get_inner_mtu = iptfs_get_inner_mtu,
+ .input = iptfs_input,
.output = iptfs_output_collect,
.prepare_output = iptfs_prepare_output,
};
--
2.46.0
^ permalink raw reply related [flat|nested] 40+ messages in thread* [PATCH ipsec-next v8 12/16] xfrm: iptfs: handle received fragmented inner packets
2024-08-04 20:33 [PATCH ipsec-next v8 00/16] Add IP-TFS mode to xfrm Christian Hopps
` (10 preceding siblings ...)
2024-08-04 20:33 ` [PATCH ipsec-next v8 11/16] xfrm: iptfs: add basic receive packet (tunnel egress) handling Christian Hopps
@ 2024-08-04 20:33 ` Christian Hopps
2024-08-04 20:33 ` [PATCH ipsec-next v8 13/16] xfrm: iptfs: add reusing received skb for the tunnel egress packet Christian Hopps
` (3 subsequent siblings)
15 siblings, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-04 20:33 UTC (permalink / raw)
To: devel; +Cc: Steffen Klassert, netdev, Christian Hopps, Christian Hopps
From: Christian Hopps <chopps@labn.net>
Add support for handling receipt of partial inner packets that have
been fragmented across multiple outer IP-TFS tunnel packets.
Signed-off-by: Christian Hopps <chopps@labn.net>
---
net/xfrm/xfrm_iptfs.c | 488 ++++++++++++++++++++++++++++++++++++++++--
1 file changed, 474 insertions(+), 14 deletions(-)
diff --git a/net/xfrm/xfrm_iptfs.c b/net/xfrm/xfrm_iptfs.c
index ea0f47af345d..5f9159260319 100644
--- a/net/xfrm/xfrm_iptfs.c
+++ b/net/xfrm/xfrm_iptfs.c
@@ -24,6 +24,21 @@
#define IPTFS_SUBTYPE_BASIC 0
#define IPTFS_SUBTYPE_CC 1
+/* ----------------------------------------------- */
+/* IP-TFS default SA values (tunnel egress/dir-in) */
+/* ----------------------------------------------- */
+
+/**
+ * define IPTFS_DEFAULT_DROP_TIME_USECS - default drop time
+ *
+ * The default IPTFS drop time in microseconds. The drop time is the amount of
+ * time before a missing out-of-order IPTFS tunnel packet is considered lost.
+ * See also the reorder window.
+ *
+ * Default 1s.
+ */
+#define IPTFS_DEFAULT_DROP_TIME_USECS 1000000
+
/* ------------------------------------------------ */
/* IPTFS default SA values (tunnel ingress/dir-out) */
/* ------------------------------------------------ */
@@ -89,6 +104,13 @@ struct xfrm_iptfs_config {
* @init_delay_ns: nanoseconds to wait to send initial IPTFS packet.
* @iptfs_timer: output timer.
* @payload_mtu: max payload size.
+ * @drop_lock: lock to protect reorder queue.
+ * @drop_timer: timer for considering next packet lost.
+ * @drop_time_ns: timer intervan in nanoseconds.
+ * @ra_newskb: new pkt being reassembled.
+ * @ra_wantseq: expected next sequence for reassembly.
+ * @ra_runt: last pkt bytes from very end of last skb.
+ * @ra_runtlen: size of ra_runt.
*/
struct xfrm_iptfs_data {
struct xfrm_iptfs_config cfg;
@@ -102,10 +124,33 @@ struct xfrm_iptfs_data {
u64 init_delay_ns; /* nanoseconds */
struct hrtimer iptfs_timer; /* output timer */
u32 payload_mtu; /* max payload size */
+
+ /* Tunnel egress */
+ spinlock_t drop_lock;
+ struct hrtimer drop_timer;
+ u64 drop_time_ns;
+
+ /* Tunnel egress reassembly */
+ struct sk_buff *ra_newskb; /* new pkt being reassembled */
+ u64 ra_wantseq; /* expected next sequence */
+ u8 ra_runt[6]; /* last pkt bytes from last skb */
+ u8 ra_runtlen; /* count of ra_runt */
};
static u32 __iptfs_get_inner_mtu(struct xfrm_state *x, int outer_mtu);
static enum hrtimer_restart iptfs_delay_timer(struct hrtimer *me);
+static enum hrtimer_restart iptfs_drop_timer(struct hrtimer *me);
+
+/* ================= */
+/* Utility Functions */
+/* ================= */
+
+static u64 __esp_seq(struct sk_buff *skb)
+{
+ u64 seq = ntohl(XFRM_SKB_CB(skb)->seq.input.low);
+
+ return seq | (u64)ntohl(XFRM_SKB_CB(skb)->seq.input.hi) << 32;
+}
/* ================= */
/* SK_BUFF Functions */
@@ -254,6 +299,67 @@ iptfs_pskb_extract_seq(u32 skblen, struct skb_seq_state *st, u32 off, int len)
return skb;
}
+/**
+ * iptfs_input_save_runt() - save data in xtfs runt space.
+ * @xtfs: xtfs state
+ * @seq: the current sequence
+ * @buf: packet data
+ * @len: length of packet data
+ *
+ * Save the small (`len`) start of a fragmented packet in `buf` in the xtfs data
+ * runt space.
+ */
+static void iptfs_input_save_runt(struct xfrm_iptfs_data *xtfs, u64 seq,
+ u8 *buf, int len)
+{
+ BUG_ON(xtfs->ra_newskb); /* we won't have a new SKB yet */
+
+ memcpy(xtfs->ra_runt, buf, len);
+
+ xtfs->ra_runtlen = len;
+ xtfs->ra_wantseq = seq + 1;
+}
+
+/**
+ * __iptfs_iphlen() - return the v4/v6 header length using packet data.
+ * @data: pointer at octet with version nibble
+ *
+ * The version data is expected to be valid (i.e., either 4 or 6).
+ *
+ * Return: the IP header size based on the IP version.
+ */
+static u32 __iptfs_iphlen(u8 *data)
+{
+ struct iphdr *iph = (struct iphdr *)data;
+
+ if (iph->version == 0x4)
+ return sizeof(*iph);
+ BUG_ON(iph->version != 0x6);
+ return sizeof(struct ipv6hdr);
+}
+
+/**
+ * __iptfs_iplen() - return the v4/v6 length using packet data.
+ * @data: pointer to ip (v4/v6) packet header
+ *
+ * Grab the IPv4 or IPv6 length value in the start of the inner packet header
+ * pointed to by `data`. Assumes data len is enough for the length field only.
+ *
+ * The version data is expected to be valid (i.e., either 4 or 6).
+ *
+ * Return: the length value.
+ */
+static u32 __iptfs_iplen(u8 *data)
+{
+ struct iphdr *iph = (struct iphdr *)data;
+
+ if (iph->version == 0x4)
+ return ntohs(iph->tot_len);
+ BUG_ON(iph->version != 0x6);
+ return ntohs(((struct ipv6hdr *)iph)->payload_len) +
+ sizeof(struct ipv6hdr);
+}
+
/**
* iptfs_complete_inner_skb() - finish preparing the inner packet for gro recv.
* @x: xfrm state
@@ -303,6 +409,239 @@ static void iptfs_complete_inner_skb(struct xfrm_state *x, struct sk_buff *skb)
}
}
+static void __iptfs_reassem_done(struct xfrm_iptfs_data *xtfs, bool free)
+{
+ assert_spin_locked(&xtfs->drop_lock);
+
+ /* We don't care if it works locking takes care of things */
+ hrtimer_try_to_cancel(&xtfs->drop_timer);
+ if (free)
+ kfree_skb(xtfs->ra_newskb);
+ xtfs->ra_newskb = NULL;
+}
+
+/**
+ * iptfs_reassem_abort() - In-progress packet is aborted free the state.
+ * @xtfs: xtfs state
+ */
+static void iptfs_reassem_abort(struct xfrm_iptfs_data *xtfs)
+{
+ __iptfs_reassem_done(xtfs, true);
+}
+
+/**
+ * iptfs_reassem_done() - In-progress packet is complete, clear the state.
+ * @xtfs: xtfs state
+ */
+static void iptfs_reassem_done(struct xfrm_iptfs_data *xtfs)
+{
+ __iptfs_reassem_done(xtfs, false);
+}
+
+/**
+ * iptfs_reassem_cont() - Continue the reassembly of an inner packets.
+ * @xtfs: xtfs state
+ * @seq: sequence of current packet
+ * @st: seq read stat for current packet
+ * @skb: current packet
+ * @data: offset into sequential packet data
+ * @blkoff: packet blkoff value
+ * @list: list of skbs to enqueue completed packet on
+ *
+ * Process an IPTFS payload that has a non-zero `blkoff` or when we are
+ * expecting the continuation b/c we have a runt or in-progress packet.
+ *
+ * Return: the new data offset to continue processing from.
+ */
+static u32 iptfs_reassem_cont(struct xfrm_iptfs_data *xtfs, u64 seq,
+ struct skb_seq_state *st, struct sk_buff *skb,
+ u32 data, u32 blkoff, struct list_head *list)
+{
+ struct sk_buff *newskb = xtfs->ra_newskb;
+ u32 remaining = skb->len - data;
+ u32 runtlen = xtfs->ra_runtlen;
+ u32 copylen, fraglen, ipremain, iphlen, iphremain, rrem;
+
+ /* Handle packet fragment we aren't expecting */
+ if (!runtlen && !xtfs->ra_newskb)
+ return data + min(blkoff, remaining);
+
+ /* Important to remember that input to this function is an ordered
+ * packet stream (unless the user disabled the reorder window). Thus if
+ * we are waiting for, and expecting the next packet so we can continue
+ * assembly, a newer sequence number indicates older ones are not coming
+ * (or if they do should be ignored). Technically we can receive older
+ * ones when the reorder window is disabled; however, the user should
+ * have disabled fragmentation in this case, and regardless we don't
+ * deal with it.
+ *
+ * blkoff could be zero if the stream is messed up (or it's an all pad
+ * insertion) be careful to handle that case in each of the below
+ */
+
+ /* Too old case: This can happen when the reorder window is disabled so
+ * ordering isn't actually guaranteed.
+ */
+ if (seq < xtfs->ra_wantseq)
+ return data + remaining;
+
+ /* Too new case: We missed what we wanted cleanup. */
+ if (seq > xtfs->ra_wantseq) {
+ XFRM_INC_STATS(xs_net(xtfs->x), LINUX_MIB_XFRMINIPTFSERROR);
+ goto abandon;
+ }
+
+ if (blkoff == 0) {
+ if ((*skb->data & 0xF0) != 0) {
+ XFRM_INC_STATS(xs_net(xtfs->x),
+ LINUX_MIB_XFRMINIPTFSERROR);
+ goto abandon;
+ }
+ /* Handle all pad case, advance expected sequence number.
+ * (RFC 9347 S2.2.3)
+ */
+ xtfs->ra_wantseq++;
+ /* will end parsing */
+ return data + remaining;
+ }
+
+ if (runtlen) {
+ BUG_ON(xtfs->ra_newskb);
+
+ /* Regardless of what happens we're done with the runt */
+ xtfs->ra_runtlen = 0;
+
+ /* The start of this inner packet was at the very end of the last
+ * iptfs payload which didn't include enough for the ip header
+ * length field. We must have *at least* that now.
+ */
+ rrem = sizeof(xtfs->ra_runt) - runtlen;
+ if (remaining < rrem || blkoff < rrem) {
+ XFRM_INC_STATS(xs_net(xtfs->x),
+ LINUX_MIB_XFRMINIPTFSERROR);
+ goto abandon;
+ }
+
+ /* fill in the runt data */
+ if (skb_copy_bits_seq(st, data, &xtfs->ra_runt[runtlen],
+ rrem)) {
+ XFRM_INC_STATS(xs_net(xtfs->x),
+ LINUX_MIB_XFRMINBUFFERERROR);
+ goto abandon;
+ }
+
+ /* We have enough data to get the ip length value now,
+ * allocate an in progress skb
+ */
+ ipremain = __iptfs_iplen(xtfs->ra_runt);
+ if (ipremain < sizeof(xtfs->ra_runt)) {
+ /* length has to be at least runtsize large */
+ XFRM_INC_STATS(xs_net(xtfs->x),
+ LINUX_MIB_XFRMINIPTFSERROR);
+ goto abandon;
+ }
+
+ /* For the runt case we don't attempt sharing currently. NOTE:
+ * Currently, this IPTFS implementation will not create runts.
+ */
+
+ newskb = iptfs_alloc_skb(skb, ipremain, false);
+ if (!newskb) {
+ XFRM_INC_STATS(xs_net(xtfs->x), LINUX_MIB_XFRMINERROR);
+ goto abandon;
+ }
+ xtfs->ra_newskb = newskb;
+
+ /* Copy the runt data into the buffer, but leave data
+ * pointers the same as normal non-runt case. The extra `rrem`
+ * recopied bytes are basically cacheline free. Allows using
+ * same logic below to complete.
+ */
+ memcpy(skb_put(newskb, runtlen), xtfs->ra_runt,
+ sizeof(xtfs->ra_runt));
+ }
+
+ /* Continue reassembling the packet */
+ ipremain = __iptfs_iplen(newskb->data);
+ iphlen = __iptfs_iphlen(newskb->data);
+
+ /* Sanity check, we created the newskb knowing the IP length so the IP
+ * length can't now be shorter.
+ */
+ BUG_ON(newskb->len > ipremain);
+
+ ipremain -= newskb->len;
+ if (blkoff < ipremain) {
+ /* Corrupt data, we don't have enough to complete the packet */
+ XFRM_INC_STATS(xs_net(xtfs->x), LINUX_MIB_XFRMINIPTFSERROR);
+ goto abandon;
+ }
+
+ /* We want the IP header in linear space */
+ if (newskb->len < iphlen) {
+ iphremain = iphlen - newskb->len;
+ if (blkoff < iphremain) {
+ XFRM_INC_STATS(xs_net(xtfs->x),
+ LINUX_MIB_XFRMINIPTFSERROR);
+ goto abandon;
+ }
+ fraglen = min(blkoff, remaining);
+ copylen = min(fraglen, iphremain);
+ BUG_ON(skb_tailroom(newskb) < copylen);
+ if (skb_copy_bits_seq(st, data, skb_put(newskb, copylen),
+ copylen)) {
+ XFRM_INC_STATS(xs_net(xtfs->x),
+ LINUX_MIB_XFRMINBUFFERERROR);
+ goto abandon;
+ }
+ /* this is a silly condition that might occur anyway */
+ if (copylen < iphremain) {
+ xtfs->ra_wantseq++;
+ return data + fraglen;
+ }
+ /* update data and things derived from it */
+ data += copylen;
+ blkoff -= copylen;
+ remaining -= copylen;
+ ipremain -= copylen;
+ }
+
+ fraglen = min(blkoff, remaining);
+ copylen = min(fraglen, ipremain);
+
+ /* We verified this was true in the main receive routine */
+ BUG_ON(skb_tailroom(newskb) < copylen);
+
+ /* copy fragment data into newskb */
+ if (skb_copy_bits_seq(st, data, skb_put(newskb, copylen), copylen)) {
+ XFRM_INC_STATS(dev_net(skb->dev), LINUX_MIB_XFRMINBUFFERERROR);
+ goto abandon;
+ }
+
+ if (copylen < ipremain) {
+ xtfs->ra_wantseq++;
+ } else {
+ /* We are done with packet reassembly! */
+ BUG_ON(copylen != ipremain);
+ iptfs_reassem_done(xtfs);
+ iptfs_complete_inner_skb(xtfs->x, newskb);
+ list_add_tail(&newskb->list, list);
+ }
+
+ /* will continue on to new data block or end */
+ return data + fraglen;
+
+abandon:
+ if (xtfs->ra_newskb) {
+ iptfs_reassem_abort(xtfs);
+ } else {
+ xtfs->ra_runtlen = 0;
+ xtfs->ra_wantseq = 0;
+ }
+ /* skip past fragment, maybe to end */
+ return data + min(blkoff, remaining);
+}
+
/**
* iptfs_input() - handle receipt of iptfs payload
* @x: xfrm state
@@ -320,15 +659,20 @@ static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
struct list_head sublist; /* rename this it's just a list */
struct sk_buff *first_skb, *next;
const unsigned char *old_mac;
+ struct xfrm_iptfs_data *xtfs;
struct ip_iptfs_hdr *ipth;
struct iphdr *iph;
struct net *net;
u32 remaining, iplen, iphlen, data, tail;
- u32 blkoff;
+ u32 blkoff, capturelen;
+ u64 seq;
+ xtfs = x->mode_data;
net = xs_net(x);
first_skb = NULL;
+ seq = __esp_seq(skb);
+
/* Large enough to hold both types of header */
ipth = (struct ip_iptfs_hdr *)&iptcch;
@@ -366,12 +710,27 @@ static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
INIT_LIST_HEAD(&sublist);
- /* Fragment handling in following commits */
+ /* Handle fragment at start of payload, and/or waiting reassembly. */
+
blkoff = ntohs(ipth->block_offset);
- data += blkoff;
+ /* check before locking i.e., maybe */
+ if (blkoff || xtfs->ra_runtlen || xtfs->ra_newskb) {
+ spin_lock(&xtfs->drop_lock);
+
+ /* check again after lock */
+ if (blkoff || xtfs->ra_runtlen || xtfs->ra_newskb) {
+ data = iptfs_reassem_cont(xtfs, seq, &skbseq, skb, data,
+ blkoff, &sublist);
+ }
+
+ spin_unlock(&xtfs->drop_lock);
+ }
/* New packets */
+
tail = skb->len;
+ BUG_ON(xtfs->ra_newskb && data < tail);
+
while (data < tail) {
__be16 protocol = 0;
@@ -390,8 +749,13 @@ static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
iph = (struct iphdr *)hbytes;
if (iph->version == 0x4) {
/* must have at least tot_len field present */
- if (remaining < 4)
+ if (remaining < 4) {
+ /* save the bytes we have, advance data and exit */
+ iptfs_input_save_runt(xtfs, seq, hbytes,
+ remaining);
+ data += remaining;
break;
+ }
iplen = be16_to_cpu(iph->tot_len);
iphlen = iph->ihl << 2;
@@ -399,8 +763,13 @@ static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
XFRM_MODE_SKB_CB(skbseq.root_skb)->tos = iph->tos;
} else if (iph->version == 0x6) {
/* must have at least payload_len field present */
- if (remaining < 6)
+ if (remaining < 6) {
+ /* save the bytes we have, advance data and exit */
+ iptfs_input_save_runt(xtfs, seq, hbytes,
+ remaining);
+ data += remaining;
break;
+ }
iplen = be16_to_cpu(((struct ipv6hdr *)hbytes)->payload_len);
iplen += sizeof(struct ipv6hdr);
@@ -410,6 +779,7 @@ static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
ipv6_get_dsfield((struct ipv6hdr *)iph);
} else if (iph->version == 0x0) {
/* pad */
+ data = tail;
break;
} else {
XFRM_INC_STATS(net, LINUX_MIB_XFRMINBUFFERERROR);
@@ -429,16 +799,14 @@ static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
if (!first_skb)
first_skb = skb;
- /* Fragment handling in following commits */
- if (iplen > remaining)
- break;
-
- skb = iptfs_pskb_extract_seq(iplen, &skbseq, data, iplen);
+ capturelen = min(iplen, remaining);
+ skb = iptfs_pskb_extract_seq(iplen, &skbseq, data, capturelen);
if (!skb) {
/* skip to next packet or done */
- data += iplen;
+ data += capturelen;
continue;
}
+ BUG_ON(skb->len != capturelen);
skb->protocol = protocol;
if (old_mac) {
@@ -449,11 +817,38 @@ static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
eth_hdr(skb)->h_proto = skb->protocol;
}
- data += iplen;
+ data += capturelen;
+
+ if (skb->len < iplen) {
+ BUG_ON(data != tail);
+ BUG_ON(xtfs->ra_newskb);
+
+ /* Start reassembly */
+ spin_lock(&xtfs->drop_lock);
+
+ xtfs->ra_newskb = skb;
+ xtfs->ra_wantseq = seq + 1;
+ if (!hrtimer_is_queued(&xtfs->drop_timer)) {
+ /* softirq blocked lest the timer fire and interrupt us */
+ BUG_ON(!in_interrupt());
+ hrtimer_start(&xtfs->drop_timer,
+ xtfs->drop_time_ns,
+ IPTFS_HRTIMER_MODE);
+ }
+
+ spin_unlock(&xtfs->drop_lock);
+
+ break;
+ }
+
iptfs_complete_inner_skb(x, skb);
list_add_tail(&skb->list, &sublist);
}
+ if (data != tail)
+ /* this should not happen from the above code */
+ XFRM_INC_STATS(net, LINUX_MIB_XFRMINIPTFSERROR);
+
/* Send the packets! */
list_for_each_entry_safe(skb, next, &sublist, list) {
skb_list_del_init(skb);
@@ -481,6 +876,47 @@ static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
return -EINPROGRESS;
}
+/**
+ * iptfs_drop_timer() - Handle drop timer expiry.
+ * @me: the timer
+ *
+ * This is similar to our input function.
+ *
+ * The drop timer is set when we start an in progress reassembly, and also when
+ * we save a future packet in the window saved array.
+ *
+ * NOTE packets in the save window are always newer WRT drop times as
+ * they get further in the future. i.e. for:
+ *
+ * if slots (S0, S1, ... Sn) and `Dn` is the drop time for slot `Sn`,
+ * then D(n-1) <= D(n).
+ *
+ * So, regardless of why the timer is firing we can always discard any inprogress
+ * fragment; either it's the reassembly timer, or slot 0 is going to be
+ * dropped as S0 must have the most recent drop time, and slot 0 holds the
+ * continuation fragment of the in progress packet.
+ *
+ * Returns HRTIMER_NORESTART.
+ */
+static enum hrtimer_restart iptfs_drop_timer(struct hrtimer *me)
+{
+ struct xfrm_iptfs_data *xtfs;
+ struct xfrm_state *x;
+
+ xtfs = container_of(me, typeof(*xtfs), drop_timer);
+ x = xtfs->x;
+
+ /* Drop any in progress packet */
+ spin_lock(&xtfs->drop_lock);
+ if (xtfs->ra_newskb) {
+ kfree_skb(xtfs->ra_newskb);
+ xtfs->ra_newskb = NULL;
+ }
+ spin_unlock(&xtfs->drop_lock);
+
+ return HRTIMER_NORESTART;
+}
+
/* ================================= */
/* IPTFS Sending (ingress) Functions */
/* ================================= */
@@ -1287,6 +1723,7 @@ static int iptfs_user_init(struct net *net, struct xfrm_state *x,
xc = &xtfs->cfg;
xc->max_queue_size = IPTFS_DEFAULT_MAX_QUEUE_SIZE;
+ xtfs->drop_time_ns = IPTFS_DEFAULT_DROP_TIME_USECS * NSECS_IN_USEC;
xtfs->init_delay_ns = IPTFS_DEFAULT_INIT_DELAY_USECS * NSECS_IN_USEC;
if (attrs[XFRMA_IPTFS_DONT_FRAG])
@@ -1305,6 +1742,10 @@ static int iptfs_user_init(struct net *net, struct xfrm_state *x,
}
if (attrs[XFRMA_IPTFS_MAX_QSIZE])
xc->max_queue_size = nla_get_u32(attrs[XFRMA_IPTFS_MAX_QSIZE]);
+ if (attrs[XFRMA_IPTFS_DROP_TIME])
+ xtfs->drop_time_ns =
+ (u64)nla_get_u32(attrs[XFRMA_IPTFS_DROP_TIME]) *
+ NSECS_IN_USEC;
if (attrs[XFRMA_IPTFS_INIT_DELAY])
xtfs->init_delay_ns =
(u64)nla_get_u32(attrs[XFRMA_IPTFS_INIT_DELAY]) *
@@ -1323,7 +1764,9 @@ static unsigned int iptfs_sa_len(const struct xfrm_state *x)
struct xfrm_iptfs_config *xc = &xtfs->cfg;
unsigned int l = 0;
- if (x->dir == XFRM_SA_DIR_OUT) {
+ if (x->dir == XFRM_SA_DIR_IN) {
+ l += nla_total_size(sizeof(u32)); /* drop time usec */
+ } else {
if (xc->dont_frag)
l += nla_total_size(0); /* dont-frag flag */
l += nla_total_size(sizeof(u32)); /* init delay usec */
@@ -1341,7 +1784,11 @@ static int iptfs_copy_to_user(struct xfrm_state *x, struct sk_buff *skb)
int ret = 0;
u64 q;
- if (x->dir == XFRM_SA_DIR_OUT) {
+ if (x->dir == XFRM_SA_DIR_IN) {
+ q = xtfs->drop_time_ns;
+ (void)do_div(q, NSECS_IN_USEC);
+ ret = nla_put_u32(skb, XFRMA_IPTFS_DROP_TIME, q);
+ } else {
if (xc->dont_frag) {
ret = nla_put_flag(skb, XFRMA_IPTFS_DONT_FRAG);
if (ret)
@@ -1372,6 +1819,10 @@ static void __iptfs_init_state(struct xfrm_state *x,
hrtimer_init(&xtfs->iptfs_timer, CLOCK_MONOTONIC, IPTFS_HRTIMER_MODE);
xtfs->iptfs_timer.function = iptfs_delay_timer;
+ spin_lock_init(&xtfs->drop_lock);
+ hrtimer_init(&xtfs->drop_timer, CLOCK_MONOTONIC, IPTFS_HRTIMER_MODE);
+ xtfs->drop_timer.function = iptfs_drop_timer;
+
/* Modify type (esp) adjustment values */
if (x->props.family == AF_INET)
@@ -1395,6 +1846,8 @@ static int iptfs_clone(struct xfrm_state *x, struct xfrm_state *orig)
if (!xtfs)
return -ENOMEM;
+ xtfs->ra_newskb = NULL;
+
__iptfs_init_state(x, xtfs);
return 0;
@@ -1431,6 +1884,13 @@ static void iptfs_delete_state(struct xfrm_state *x)
while ((skb = __skb_dequeue(&list)))
kfree_skb(skb);
+ spin_lock_bh(&xtfs->drop_lock);
+ hrtimer_cancel(&xtfs->drop_timer);
+ spin_unlock_bh(&xtfs->drop_lock);
+
+ if (xtfs->ra_newskb)
+ kfree_skb(xtfs->ra_newskb);
+
kfree_sensitive(xtfs);
module_put(x->mode_cbs->owner);
--
2.46.0
^ permalink raw reply related [flat|nested] 40+ messages in thread* [PATCH ipsec-next v8 13/16] xfrm: iptfs: add reusing received skb for the tunnel egress packet
2024-08-04 20:33 [PATCH ipsec-next v8 00/16] Add IP-TFS mode to xfrm Christian Hopps
` (11 preceding siblings ...)
2024-08-04 20:33 ` [PATCH ipsec-next v8 12/16] xfrm: iptfs: handle received fragmented inner packets Christian Hopps
@ 2024-08-04 20:33 ` Christian Hopps
2024-08-04 20:33 ` [PATCH ipsec-next v8 14/16] xfrm: iptfs: add skb-fragment sharing code Christian Hopps
` (2 subsequent siblings)
15 siblings, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-04 20:33 UTC (permalink / raw)
To: devel; +Cc: Steffen Klassert, netdev, Christian Hopps, Christian Hopps
From: Christian Hopps <chopps@labn.net>
Add an optimization of re-using the tunnel outer skb re-transmission
of the inner packet to avoid skb allocation and copy.
Signed-off-by: Christian Hopps <chopps@labn.net>
---
net/xfrm/xfrm_iptfs.c | 126 +++++++++++++++++++++++++++++++++++-------
1 file changed, 105 insertions(+), 21 deletions(-)
diff --git a/net/xfrm/xfrm_iptfs.c b/net/xfrm/xfrm_iptfs.c
index 5f9159260319..b05b72cd3ba5 100644
--- a/net/xfrm/xfrm_iptfs.c
+++ b/net/xfrm/xfrm_iptfs.c
@@ -657,19 +657,20 @@ static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
struct ip_iptfs_cc_hdr iptcch;
struct skb_seq_state skbseq;
struct list_head sublist; /* rename this it's just a list */
- struct sk_buff *first_skb, *next;
+ struct sk_buff *first_skb, *defer, *next;
const unsigned char *old_mac;
struct xfrm_iptfs_data *xtfs;
struct ip_iptfs_hdr *ipth;
struct iphdr *iph;
struct net *net;
- u32 remaining, iplen, iphlen, data, tail;
+ u32 remaining, first_iplen, iplen, iphlen, data, tail;
u32 blkoff, capturelen;
u64 seq;
xtfs = x->mode_data;
net = xs_net(x);
first_skb = NULL;
+ defer = NULL;
seq = __esp_seq(skb);
@@ -796,25 +797,94 @@ static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
skb_prepare_seq_read(save, data, tail, &skbseq);
}
- if (!first_skb)
+ if (first_skb) {
+ skb = NULL;
+ } else {
first_skb = skb;
+ first_iplen = iplen;
+
+ /* We are going to skip over `data` bytes to reach the
+ * start of the IP header of `iphlen` len for `iplen`
+ * inner packet.
+ */
+
+ if (skb_has_frag_list(skb)) {
+ defer = skb;
+ skb = NULL;
+ } else if (data + iphlen <= skb_headlen(skb) &&
+ /* make sure our header is 32-bit aligned? */
+ /* ((uintptr_t)(skb->data + data) & 0x3) == 0 && */
+ skb_tailroom(skb) + tail - data >= iplen) {
+ /* Reuse the received skb.
+ *
+ * We have enough headlen to pull past any
+ * initial fragment data, leaving at least the
+ * IP header in the linear buffer space.
+ *
+ * For linear buffer space we only require that
+ * linear buffer space is large enough to
+ * eventually hold the entire reassembled
+ * packet (by including tailroom in the check).
+ *
+ * For non-linear tailroom is 0 and so we only
+ * re-use if the entire packet is present
+ * already.
+ *
+ * NOTE: there are many more options for
+ * sharing, KISS for now. Also, this can produce
+ * skb's with the IP header unaligned to 32
+ * bits. If that ends up being a problem then a
+ * check should be added to the conditional
+ * above that the header lies on a 32-bit
+ * boundary as well.
+ */
+ skb_pull(skb, data);
+
+ /* our range just changed */
+ data = 0;
+ tail = skb->len;
+ remaining = skb->len;
+
+ skb->protocol = protocol;
+ skb_mac_header_rebuild(skb);
+ if (skb->mac_len)
+ eth_hdr(skb)->h_proto = skb->protocol;
+
+ /* all pointers could be changed now reset walk */
+ skb_abort_seq_read(&skbseq);
+ skb_prepare_seq_read(skb, data, tail, &skbseq);
+ } else {
+ /* We couldn't reuse the input skb so allocate a
+ * new one.
+ */
+ defer = skb;
+ skb = NULL;
+ }
+
+ /* Don't trim `first_skb` until the end as we are
+ * walking that data now.
+ */
+ }
capturelen = min(iplen, remaining);
- skb = iptfs_pskb_extract_seq(iplen, &skbseq, data, capturelen);
if (!skb) {
- /* skip to next packet or done */
- data += capturelen;
- continue;
- }
- BUG_ON(skb->len != capturelen);
-
- skb->protocol = protocol;
- if (old_mac) {
- /* rebuild the mac header */
- skb_set_mac_header(skb, -first_skb->mac_len);
- memcpy(skb_mac_header(skb), old_mac,
- first_skb->mac_len);
- eth_hdr(skb)->h_proto = skb->protocol;
+ skb = iptfs_pskb_extract_seq(iplen, &skbseq, data,
+ capturelen);
+ if (!skb) {
+ /* skip to next packet or done */
+ data += capturelen;
+ continue;
+ }
+ BUG_ON(skb->len != capturelen);
+
+ skb->protocol = protocol;
+ if (old_mac) {
+ /* rebuild the mac header */
+ skb_set_mac_header(skb, -first_skb->mac_len);
+ memcpy(skb_mac_header(skb), old_mac,
+ first_skb->mac_len);
+ eth_hdr(skb)->h_proto = skb->protocol;
+ }
}
data += capturelen;
@@ -849,8 +919,19 @@ static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
/* this should not happen from the above code */
XFRM_INC_STATS(net, LINUX_MIB_XFRMINIPTFSERROR);
+ if (first_skb && first_iplen && !defer && first_skb != xtfs->ra_newskb) {
+ /* first_skb is queued b/c !defer and not partial */
+ if (pskb_trim(first_skb, first_iplen)) {
+ /* error trimming */
+ list_del(&first_skb->list);
+ defer = first_skb;
+ }
+ first_skb->ip_summed = CHECKSUM_NONE;
+ }
+
/* Send the packets! */
list_for_each_entry_safe(skb, next, &sublist, list) {
+ BUG_ON(skb == defer);
skb_list_del_init(skb);
if (xfrm_input(skb, 0, 0, -2))
kfree_skb(skb);
@@ -860,12 +941,15 @@ static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
skb = skbseq.root_skb;
skb_abort_seq_read(&skbseq);
- if (first_skb) {
- consume_skb(first_skb);
- } else {
+ if (defer) {
+ consume_skb(defer);
+ } else if (!first_skb) {
/* skb is the original passed in skb, but we didn't get far
- * enough to process it as the first_skb.
+ * enough to process it as the first_skb, if we had it would
+ * either be save in ra_newskb, trimmed and sent on as an skb or
+ * placed in defer to be freed.
*/
+ BUG_ON(!skb);
kfree_skb(skb);
}
--
2.46.0
^ permalink raw reply related [flat|nested] 40+ messages in thread* [PATCH ipsec-next v8 14/16] xfrm: iptfs: add skb-fragment sharing code
2024-08-04 20:33 [PATCH ipsec-next v8 00/16] Add IP-TFS mode to xfrm Christian Hopps
` (12 preceding siblings ...)
2024-08-04 20:33 ` [PATCH ipsec-next v8 13/16] xfrm: iptfs: add reusing received skb for the tunnel egress packet Christian Hopps
@ 2024-08-04 20:33 ` Christian Hopps
2024-08-04 20:33 ` [PATCH ipsec-next v8 15/16] xfrm: iptfs: handle reordering of received packets Christian Hopps
2024-08-04 20:33 ` [PATCH ipsec-next v8 16/16] xfrm: iptfs: add tracepoint functionality Christian Hopps
15 siblings, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-04 20:33 UTC (permalink / raw)
To: devel; +Cc: Steffen Klassert, netdev, Christian Hopps, Christian Hopps
From: Christian Hopps <chopps@labn.net>
Avoid copying the inner packet data by sharing the skb data fragments
from the output packet skb into new inner packet skb.
Signed-off-by: Christian Hopps <chopps@labn.net>
---
net/xfrm/xfrm_iptfs.c | 305 ++++++++++++++++++++++++++++++++++++++++--
1 file changed, 297 insertions(+), 8 deletions(-)
diff --git a/net/xfrm/xfrm_iptfs.c b/net/xfrm/xfrm_iptfs.c
index b05b72cd3ba5..598dc88b1408 100644
--- a/net/xfrm/xfrm_iptfs.c
+++ b/net/xfrm/xfrm_iptfs.c
@@ -75,6 +75,7 @@
#define XFRM_IPTFS_MIN_L3HEADROOM 128
#define XFRM_IPTFS_MIN_L2HEADROOM (64 + 16)
#define IPTFS_FRAG_COPY_MAX 256 /* max for copying to create iptfs frags */
+#define IPTFS_PKT_SHARE_MIN 129 /* min to try to share vs copy pkt data */
#define NSECS_IN_USEC 1000
#define IPTFS_HRTIMER_MODE HRTIMER_MODE_REL_SOFT
@@ -228,6 +229,207 @@ static void skb_head_to_frag(const struct sk_buff *skb, skb_frag_t *frag)
skb_frag_fill_page_desc(frag, page, skb->data - addr, skb_headlen(skb));
}
+/**
+ * struct skb_frag_walk - use to track a walk through fragments
+ * @fragi: current fragment index
+ * @past: length of data in fragments before @fragi
+ * @total: length of data in all fragments
+ * @nr_frags: number of fragments present in array
+ * @initial_offset: the value passed in to skb_prepare_frag_walk()
+ * @pp_recycle: copy of skb->pp_recycle
+ * @frags: the page fragments inc. room for head page
+ */
+struct skb_frag_walk {
+ u32 fragi;
+ u32 past;
+ u32 total;
+ u32 nr_frags;
+ u32 initial_offset;
+ bool pp_recycle;
+ skb_frag_t frags[MAX_SKB_FRAGS + 1];
+};
+
+/**
+ * skb_prepare_frag_walk() - initialize a frag walk over an skb.
+ * @skb: the skb to walk.
+ * @initial_offset: start the walk @initial_offset into the skb.
+ * @walk: the walk to initialize
+ *
+ * Future calls to skb_add_frags() will expect the @offset value to be at
+ * least @initial_offset large.
+ */
+static void skb_prepare_frag_walk(struct sk_buff *skb, u32 initial_offset,
+ struct skb_frag_walk *walk)
+{
+ struct skb_shared_info *shinfo = skb_shinfo(skb);
+ skb_frag_t *frag, *from;
+ u32 i;
+
+ walk->initial_offset = initial_offset;
+ walk->fragi = 0;
+ walk->past = 0;
+ walk->total = 0;
+ walk->nr_frags = 0;
+ walk->pp_recycle = skb->pp_recycle;
+
+ if (skb->head_frag) {
+ if (initial_offset >= skb_headlen(skb)) {
+ initial_offset -= skb_headlen(skb);
+ } else {
+ frag = &walk->frags[walk->nr_frags++];
+ skb_head_to_frag(skb, frag);
+ frag->offset += initial_offset;
+ frag->len -= initial_offset;
+ walk->total += frag->len;
+ initial_offset = 0;
+ }
+ } else {
+ BUG_ON(skb_headlen(skb) > initial_offset);
+ initial_offset -= skb_headlen(skb);
+ }
+
+ for (i = 0; i < shinfo->nr_frags; i++) {
+ from = &shinfo->frags[i];
+ if (initial_offset >= from->len) {
+ initial_offset -= from->len;
+ continue;
+ }
+ frag = &walk->frags[walk->nr_frags++];
+ *frag = *from;
+ if (initial_offset) {
+ frag->offset += initial_offset;
+ frag->len -= initial_offset;
+ initial_offset = 0;
+ }
+ walk->total += frag->len;
+ }
+ BUG_ON(initial_offset != 0);
+}
+
+static u32 __skb_reset_frag_walk(struct skb_frag_walk *walk, u32 offset)
+{
+ /* Adjust offset to refer to internal walk values */
+ BUG_ON(offset < walk->initial_offset);
+ offset -= walk->initial_offset;
+
+ /* Get to the correct fragment for offset */
+ while (offset < walk->past) {
+ walk->past -= walk->frags[--walk->fragi].len;
+ if (offset >= walk->past)
+ break;
+ BUG_ON(walk->fragi == 0);
+ }
+ while (offset >= walk->past + walk->frags[walk->fragi].len)
+ walk->past += walk->frags[walk->fragi++].len;
+
+ /* offset now relative to this current frag */
+ offset -= walk->past;
+ return offset;
+}
+
+/**
+ * skb_can_add_frags() - check if ok to add frags from walk to skb
+ * @skb: skb to check for adding frags to
+ * @walk: the walk that will be used as source for frags.
+ * @offset: offset from beginning of original skb to start from.
+ * @len: amount of data to add frag references to in @skb.
+ *
+ * Return: true if ok to add frags.
+ */
+static bool skb_can_add_frags(const struct sk_buff *skb,
+ struct skb_frag_walk *walk, u32 offset, u32 len)
+{
+ struct skb_shared_info *shinfo = skb_shinfo(skb);
+ u32 fragi, nr_frags, fraglen;
+
+ if (skb_has_frag_list(skb) || skb->pp_recycle != walk->pp_recycle)
+ return false;
+
+ /* Make offset relative to current frag after setting that */
+ offset = __skb_reset_frag_walk(walk, offset);
+
+ /* Verify we have array space for the fragments we need to add */
+ fragi = walk->fragi;
+ nr_frags = shinfo->nr_frags;
+ while (len && fragi < walk->nr_frags) {
+ skb_frag_t *frag = &walk->frags[fragi];
+
+ fraglen = frag->len;
+ if (offset) {
+ fraglen -= offset;
+ offset = 0;
+ }
+ if (++nr_frags > MAX_SKB_FRAGS)
+ return false;
+ if (len <= fraglen)
+ return true;
+ len -= fraglen;
+ fragi++;
+ }
+ /* We may not copy all @len but what we have will fit. */
+ return true;
+}
+
+/**
+ * skb_add_frags() - add a range of fragment references into an skb
+ * @skb: skb to add references into
+ * @walk: the walk to add referenced fragments from.
+ * @offset: offset from beginning of original skb to start from.
+ * @len: amount of data to add frag references to in @skb.
+ *
+ * skb_can_add_frags() should be called before this function to verify that the
+ * destination @skb is compatible with the walk and has space in the array for
+ * the to be added frag references.
+ *
+ * Return: The number of bytes not added to @skb b/c we reached the end of the
+ * walk before adding all of @len.
+ */
+static int skb_add_frags(struct sk_buff *skb, struct skb_frag_walk *walk,
+ u32 offset, u32 len)
+{
+ struct skb_shared_info *shinfo = skb_shinfo(skb);
+ u32 fraglen;
+
+ BUG_ON(skb->pp_recycle != walk->pp_recycle);
+ if (!walk->nr_frags || offset >= walk->total + walk->initial_offset)
+ return len;
+
+ /* make offset relative to current frag after setting that */
+ offset = __skb_reset_frag_walk(walk, offset);
+ BUG_ON(shinfo->nr_frags >= MAX_SKB_FRAGS);
+
+ while (len && walk->fragi < walk->nr_frags) {
+ skb_frag_t *frag = &walk->frags[walk->fragi];
+ skb_frag_t *tofrag = &shinfo->frags[shinfo->nr_frags];
+
+ *tofrag = *frag;
+ if (offset) {
+ tofrag->offset += offset;
+ tofrag->len -= offset;
+ offset = 0;
+ }
+ __skb_frag_ref(tofrag);
+ shinfo->nr_frags++;
+ BUG_ON(shinfo->nr_frags > MAX_SKB_FRAGS);
+
+ /* see if we are done */
+ fraglen = tofrag->len;
+ if (len < fraglen) {
+ tofrag->len = len;
+ skb->len += len;
+ skb->data_len += len;
+ return 0;
+ }
+ /* advance to next source fragment */
+ len -= fraglen; /* careful, use dst bv_len */
+ skb->len += fraglen; /* careful, " " " */
+ skb->data_len += fraglen; /* careful, " " " */
+ walk->past += frag->len; /* careful, use src bv_len */
+ walk->fragi++;
+ }
+ return len;
+}
+
/**
* skb_copy_bits_seq - copy bits from a skb_seq_state to kernel buffer
* @st: source skb_seq_state
@@ -267,6 +469,53 @@ static int skb_copy_bits_seq(struct skb_seq_state *st, int offset, void *to,
/* IPTFS Receiving (egress) Functions */
/* ================================== */
+/**
+ * iptfs_pskb_add_frags() - Create and add frags into a new sk_buff.
+ * @tpl: template to create new skb from.
+ * @walk: The source for fragments to add.
+ * @off: The offset into @walk to add frags from, also used with @st and
+ * @copy_len.
+ * @len: The length of data to add covering frags from @walk into @skb.
+ * This must be <= @skblen.
+ * @st: The sequence state to copy from into the new head skb.
+ * @copy_len: Copy @copy_len bytes from @st at offset @off into the new skb
+ * linear space.
+ *
+ * Create a new sk_buff `skb` using the template @tpl. Copy @copy_len bytes from
+ * @st into the new skb linear space, and then add shared fragments from the
+ * frag walk for the remaining @len of data (i.e., @len - @copy_len bytes).
+ *
+ * Return: The newly allocated sk_buff `skb` or NULL if an error occurs.
+ */
+static struct sk_buff *iptfs_pskb_add_frags(struct sk_buff *tpl,
+ struct skb_frag_walk *walk, u32 off,
+ u32 len, struct skb_seq_state *st,
+ u32 copy_len)
+{
+ struct sk_buff *skb;
+
+ skb = iptfs_alloc_skb(tpl, copy_len, false);
+ if (!skb)
+ return NULL;
+
+ /* this should not normally be happening */
+ if (!skb_can_add_frags(skb, walk, off + copy_len, len - copy_len)) {
+ kfree_skb(skb);
+ return NULL;
+ }
+
+ if (copy_len &&
+ skb_copy_bits_seq(st, off, skb_put(skb, copy_len), copy_len)) {
+ XFRM_INC_STATS(dev_net(st->root_skb->dev),
+ LINUX_MIB_XFRMINERROR);
+ kfree_skb(skb);
+ return NULL;
+ }
+
+ skb_add_frags(skb, walk, off + copy_len, len - copy_len);
+ return skb;
+}
+
/**
* iptfs_pskb_extract_seq() - Create and load data into a new sk_buff.
* @skblen: the total data size for `skb`.
@@ -457,6 +706,8 @@ static u32 iptfs_reassem_cont(struct xfrm_iptfs_data *xtfs, u64 seq,
struct skb_seq_state *st, struct sk_buff *skb,
u32 data, u32 blkoff, struct list_head *list)
{
+ struct skb_frag_walk _fragwalk;
+ struct skb_frag_walk *fragwalk = NULL;
struct sk_buff *newskb = xtfs->ra_newskb;
u32 remaining = skb->len - data;
u32 runtlen = xtfs->ra_runtlen;
@@ -609,13 +860,31 @@ static u32 iptfs_reassem_cont(struct xfrm_iptfs_data *xtfs, u64 seq,
fraglen = min(blkoff, remaining);
copylen = min(fraglen, ipremain);
- /* We verified this was true in the main receive routine */
- BUG_ON(skb_tailroom(newskb) < copylen);
+ /* If we may have the opportunity to share prepare a fragwalk. */
+ if (!skb_has_frag_list(skb) && !skb_has_frag_list(newskb) &&
+ (skb->head_frag || skb->len == skb->data_len) &&
+ skb->pp_recycle == newskb->pp_recycle) {
+ fragwalk = &_fragwalk;
+ skb_prepare_frag_walk(skb, data, fragwalk);
+ }
- /* copy fragment data into newskb */
- if (skb_copy_bits_seq(st, data, skb_put(newskb, copylen), copylen)) {
- XFRM_INC_STATS(dev_net(skb->dev), LINUX_MIB_XFRMINBUFFERERROR);
- goto abandon;
+ /* Try share then copy. */
+ if (fragwalk && skb_can_add_frags(newskb, fragwalk, data, copylen)) {
+ u32 leftover;
+
+ leftover = skb_add_frags(newskb, fragwalk, data, copylen);
+ BUG_ON(leftover != 0);
+ } else {
+ /* We verified this was true in the main receive routine */
+ BUG_ON(skb_tailroom(newskb) < copylen);
+
+ /* copy fragment data into newskb */
+ if (skb_copy_bits_seq(st, data, skb_put(newskb, copylen),
+ copylen)) {
+ XFRM_INC_STATS(xs_net(xtfs->x),
+ LINUX_MIB_XFRMINBUFFERERROR);
+ goto abandon;
+ }
}
if (copylen < ipremain) {
@@ -656,6 +925,8 @@ static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
u8 hbytes[sizeof(struct ipv6hdr)];
struct ip_iptfs_cc_hdr iptcch;
struct skb_seq_state skbseq;
+ struct skb_frag_walk _fragwalk;
+ struct skb_frag_walk *fragwalk = NULL;
struct list_head sublist; /* rename this it's just a list */
struct sk_buff *first_skb, *defer, *next;
const unsigned char *old_mac;
@@ -802,6 +1073,7 @@ static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
} else {
first_skb = skb;
first_iplen = iplen;
+ fragwalk = NULL;
/* We are going to skip over `data` bytes to reach the
* start of the IP header of `iphlen` len for `iplen`
@@ -853,6 +1125,13 @@ static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
/* all pointers could be changed now reset walk */
skb_abort_seq_read(&skbseq);
skb_prepare_seq_read(skb, data, tail, &skbseq);
+ } else if (skb->head_frag &&
+ /* We have the IP header right now */
+ remaining >= iphlen) {
+ fragwalk = &_fragwalk;
+ skb_prepare_frag_walk(skb, data, fragwalk);
+ defer = skb;
+ skb = NULL;
} else {
/* We couldn't reuse the input skb so allocate a
* new one.
@@ -868,8 +1147,18 @@ static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
capturelen = min(iplen, remaining);
if (!skb) {
- skb = iptfs_pskb_extract_seq(iplen, &skbseq, data,
- capturelen);
+ if (!fragwalk ||
+ /* Large enough to be worth sharing */
+ iplen < IPTFS_PKT_SHARE_MIN ||
+ /* Have IP header + some data to share. */
+ capturelen <= iphlen ||
+ /* Try creating skb and adding frags */
+ !(skb = iptfs_pskb_add_frags(first_skb, fragwalk,
+ data, capturelen,
+ &skbseq, iphlen))) {
+ skb = iptfs_pskb_extract_seq(iplen, &skbseq,
+ data, capturelen);
+ }
if (!skb) {
/* skip to next packet or done */
data += capturelen;
--
2.46.0
^ permalink raw reply related [flat|nested] 40+ messages in thread* [PATCH ipsec-next v8 15/16] xfrm: iptfs: handle reordering of received packets
2024-08-04 20:33 [PATCH ipsec-next v8 00/16] Add IP-TFS mode to xfrm Christian Hopps
` (13 preceding siblings ...)
2024-08-04 20:33 ` [PATCH ipsec-next v8 14/16] xfrm: iptfs: add skb-fragment sharing code Christian Hopps
@ 2024-08-04 20:33 ` Christian Hopps
2024-08-04 20:33 ` [PATCH ipsec-next v8 16/16] xfrm: iptfs: add tracepoint functionality Christian Hopps
15 siblings, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-04 20:33 UTC (permalink / raw)
To: devel; +Cc: Steffen Klassert, netdev, Christian Hopps, Christian Hopps
From: Christian Hopps <chopps@labn.net>
Handle the receipt of the outer tunnel packets out-of-order. Pointers to
the out-of-order packets are saved in a window (array) awaiting needed
prior packets. When the required prior packets are received the now
in-order packets are then passed on to the regular packet receive code.
A timer is used to consider missing earlier packet as lost so the
algorithm will advance.
Signed-off-by: Christian Hopps <chopps@labn.net>
---
net/xfrm/xfrm_iptfs.c | 507 +++++++++++++++++++++++++++++++++++++++++-
1 file changed, 495 insertions(+), 12 deletions(-)
diff --git a/net/xfrm/xfrm_iptfs.c b/net/xfrm/xfrm_iptfs.c
index 598dc88b1408..6d022935c9dc 100644
--- a/net/xfrm/xfrm_iptfs.c
+++ b/net/xfrm/xfrm_iptfs.c
@@ -39,6 +39,17 @@
*/
#define IPTFS_DEFAULT_DROP_TIME_USECS 1000000
+/**
+ * define IPTFS_DEFAULT_REORDER_WINDOW - default reorder window size
+ *
+ * The default IPTFS reorder window size. The reorder window size dictates the
+ * maximum number of IPTFS tunnel packets in a sequence that may arrive out of
+ * order.
+ *
+ * Default 3. (tcp folks suggested)
+ */
+#define IPTFS_DEFAULT_REORDER_WINDOW 3
+
/* ------------------------------------------------ */
/* IPTFS default SA values (tunnel ingress/dir-out) */
/* ------------------------------------------------ */
@@ -83,6 +94,8 @@
/**
* struct xfrm_iptfs_config - configuration for the IPTFS tunnel.
* @dont_frag: true to inhibit fragmenting across IPTFS outer packets.
+ * @reorder_win_size: the number slots in the reorder window, thus the number of
+ * packets that may arrive out of order.
* @pkt_size: size of the outer IP packet. 0 to use interface and MTU discovery,
* otherwise the user specified value.
* @max_queue_size: The maximum number of octets allowed to be queued to be sent
@@ -91,10 +104,16 @@
*/
struct xfrm_iptfs_config {
bool dont_frag : 1;
+ u16 reorder_win_size;
u32 pkt_size; /* outer_packet_size or 0 */
u32 max_queue_size; /* octets */
};
+struct skb_wseq {
+ struct sk_buff *skb;
+ u64 drop_time;
+};
+
/**
* struct xfrm_iptfs_data - mode specific xfrm state.
* @cfg: IPTFS tunnel config.
@@ -105,6 +124,10 @@ struct xfrm_iptfs_config {
* @init_delay_ns: nanoseconds to wait to send initial IPTFS packet.
* @iptfs_timer: output timer.
* @payload_mtu: max payload size.
+ * @w_seq_set: true after first seq received.
+ * @w_wantseq: waiting for this seq number as next to process (in order).
+ * @w_saved: the saved buf array (reorder window).
+ * @w_savedlen: the saved len (not size).
* @drop_lock: lock to protect reorder queue.
* @drop_timer: timer for considering next packet lost.
* @drop_time_ns: timer intervan in nanoseconds.
@@ -126,12 +149,16 @@ struct xfrm_iptfs_data {
struct hrtimer iptfs_timer; /* output timer */
u32 payload_mtu; /* max payload size */
- /* Tunnel egress */
+ /* Tunnel input reordering */
+ bool w_seq_set; /* true after first seq received */
+ u64 w_wantseq; /* expected next sequence */
+ struct skb_wseq *w_saved; /* the saved buf array */
+ u32 w_savedlen; /* the saved len (not size) */
spinlock_t drop_lock;
struct hrtimer drop_timer;
u64 drop_time_ns;
- /* Tunnel egress reassembly */
+ /* Tunnel input reassembly */
struct sk_buff *ra_newskb; /* new pkt being reassembled */
u64 ra_wantseq; /* expected next sequence */
u8 ra_runt[6]; /* last pkt bytes from last skb */
@@ -912,15 +939,13 @@ static u32 iptfs_reassem_cont(struct xfrm_iptfs_data *xtfs, u64 seq,
}
/**
- * iptfs_input() - handle receipt of iptfs payload
+ * iptfs_input_ordered() - handle next in order IPTFS payload.
* @x: xfrm state
- * @skb: the packet
+ * @skb: current packet
*
* Process the IPTFS payload in `skb` and consume it afterwards.
- *
- * Returns 0.
*/
-static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
+static void iptfs_input_ordered(struct xfrm_state *x, struct sk_buff *skb)
{
u8 hbytes[sizeof(struct ipv6hdr)];
struct ip_iptfs_cc_hdr iptcch;
@@ -1241,12 +1266,368 @@ static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
BUG_ON(!skb);
kfree_skb(skb);
}
+}
- /* We always have dealt with the input SKB, either we are re-using it,
- * or we have freed it. Return EINPROGRESS so that xfrm_input stops
- * processing it.
+/* ------------------------------- */
+/* Input (Egress) Re-ordering Code */
+/* ------------------------------- */
+
+static void __vec_shift(struct xfrm_iptfs_data *xtfs, u32 shift)
+{
+ u32 savedlen = xtfs->w_savedlen;
+
+ if (shift > savedlen)
+ shift = savedlen;
+ if (shift != savedlen)
+ memcpy(xtfs->w_saved, xtfs->w_saved + shift,
+ (savedlen - shift) * sizeof(*xtfs->w_saved));
+ memset(xtfs->w_saved + savedlen - shift, 0,
+ shift * sizeof(*xtfs->w_saved));
+ xtfs->w_savedlen -= shift;
+}
+
+static void __reorder_past(struct xfrm_iptfs_data *xtfs, struct sk_buff *inskb,
+ struct list_head *freelist)
+{
+ list_add_tail(&inskb->list, freelist);
+}
+
+static u32 __reorder_drop(struct xfrm_iptfs_data *xtfs, struct list_head *list)
+
+{
+ struct skb_wseq *s, *se;
+ const u32 savedlen = xtfs->w_savedlen;
+ time64_t now = ktime_get_raw_fast_ns();
+ u32 count = 0;
+ u32 scount = 0;
+
+ BUG_ON(!savedlen);
+ if (xtfs->w_saved[0].drop_time > now)
+ goto set_timer;
+
+ ++xtfs->w_wantseq;
+
+ /* Keep flushing packets until we reach a drop time greater than now. */
+ s = xtfs->w_saved;
+ se = s + savedlen;
+ do {
+ /* Walking past empty slots until we reach a packet */
+ for (; s < se && !s->skb; s++)
+ if (s->drop_time > now)
+ goto outerdone;
+ /* Sending packets until we hit another empty slot. */
+ for (; s < se && s->skb; scount++, s++)
+ list_add_tail(&s->skb->list, list);
+ } while (s < se);
+outerdone:
+
+ count = s - xtfs->w_saved;
+ if (count) {
+ xtfs->w_wantseq += count;
+
+ /* Shift handled slots plus final empty slot into slot 0. */
+ __vec_shift(xtfs, count);
+ }
+
+ if (xtfs->w_savedlen) {
+set_timer:
+ /* Drifting is OK */
+ hrtimer_start(&xtfs->drop_timer,
+ xtfs->w_saved[0].drop_time - now,
+ IPTFS_HRTIMER_MODE);
+ }
+ return scount;
+}
+
+static void __reorder_this(struct xfrm_iptfs_data *xtfs, struct sk_buff *inskb,
+ struct list_head *list)
+{
+ struct skb_wseq *s, *se;
+ const u32 savedlen = xtfs->w_savedlen;
+ u32 count = 0;
+
+ /* Got what we wanted. */
+ list_add_tail(&inskb->list, list);
+ ++xtfs->w_wantseq;
+ if (!savedlen)
+ return;
+
+ /* Flush remaining consecutive packets. */
+
+ /* Keep sending until we hit another missed pkt. */
+ for (s = xtfs->w_saved, se = s + savedlen; s < se && s->skb; s++)
+ list_add_tail(&s->skb->list, list);
+ count = s - xtfs->w_saved;
+ if (count)
+ xtfs->w_wantseq += count;
+
+ /* Shift handled slots plus final empty slot into slot 0. */
+ __vec_shift(xtfs, count + 1);
+}
+
+/* Set the slot's drop time and all the empty slots below it until reaching a
+ * filled slot which will already be set.
+ */
+static void iptfs_set_window_drop_times(struct xfrm_iptfs_data *xtfs, int index)
+{
+ const u32 savedlen = xtfs->w_savedlen;
+ struct skb_wseq *s = xtfs->w_saved;
+ time64_t drop_time;
+
+ assert_spin_locked(&xtfs->drop_lock);
+
+ if (savedlen > index + 1) {
+ /* we are below another, our drop time and the timer are already set */
+ BUG_ON(xtfs->w_saved[index + 1].drop_time !=
+ xtfs->w_saved[index].drop_time);
+ return;
+ }
+ /* we are the most future so get a new drop time. */
+ drop_time = ktime_get_raw_fast_ns();
+ drop_time += xtfs->drop_time_ns;
+
+ /* Walk back through the array setting drop times as we go */
+ s[index].drop_time = drop_time;
+ while (index-- > 0 && !s[index].skb)
+ s[index].drop_time = drop_time;
+
+ /* If we walked all the way back, schedule the drop timer if needed */
+ if (index == -1 && !hrtimer_is_queued(&xtfs->drop_timer))
+ hrtimer_start(&xtfs->drop_timer, xtfs->drop_time_ns,
+ IPTFS_HRTIMER_MODE);
+}
+
+static void __reorder_future_fits(struct xfrm_iptfs_data *xtfs,
+ struct sk_buff *inskb,
+ struct list_head *freelist)
+{
+ const u32 nslots = xtfs->cfg.reorder_win_size + 1;
+ const u64 inseq = __esp_seq(inskb);
+ const u64 wantseq = xtfs->w_wantseq;
+ const u64 distance = inseq - wantseq;
+ const u32 savedlen = xtfs->w_savedlen;
+ const u32 index = distance - 1;
+
+ BUG_ON(distance >= nslots);
+
+ /* Handle future sequence number received which fits in the window.
+ *
+ * We know we don't have the seq we want so we won't be able to flush
+ * anything.
*/
- return -EINPROGRESS;
+
+ /* slot count is 4, saved size is 3 savedlen is 2
+ *
+ * "window boundary" is based on the fixed window size
+ * distance is also slot number
+ * index is an array index (i.e., - 1 of slot)
+ * : : - implicit NULL after array len
+ *
+ * +--------- used length (savedlen == 2)
+ * | +----- array size (nslots - 1 == 3)
+ * | | + window boundary (nslots == 4)
+ * V V | V
+ * |
+ * 0 1 2 3 | slot number
+ * --- 0 1 2 | array index
+ * [-] [b] : :| array
+ *
+ * "2" "3" "4" *5*| seq numbers
+ *
+ * We receive seq number 5
+ * distance == 3 [inseq(5) - w_wantseq(2)]
+ * index == 2 [distance(6) - 1]
+ */
+
+ if (xtfs->w_saved[index].skb) {
+ /* a dup of a future */
+ list_add_tail(&inskb->list, freelist);
+ return;
+ }
+
+ xtfs->w_saved[index].skb = inskb;
+ xtfs->w_savedlen = max(savedlen, index + 1);
+ iptfs_set_window_drop_times(xtfs, index);
+}
+
+static void __reorder_future_shifts(struct xfrm_iptfs_data *xtfs,
+ struct sk_buff *inskb,
+ struct list_head *list,
+ struct list_head *freelist)
+{
+ const u32 nslots = xtfs->cfg.reorder_win_size + 1;
+ const u64 inseq = __esp_seq(inskb);
+ u32 savedlen = xtfs->w_savedlen;
+ u64 wantseq = xtfs->w_wantseq;
+ struct sk_buff *slot0 = NULL;
+ struct skb_wseq *wnext;
+ u32 beyond, shifting, slot;
+ u64 distance;
+
+ BUG_ON(inseq <= wantseq);
+ distance = inseq - wantseq;
+ BUG_ON(distance <= nslots - 1);
+ beyond = distance - (nslots - 1);
+
+ /* Handle future sequence number received.
+ *
+ * IMPORTANT: we are at least advancing w_wantseq (i.e., wantseq) by 1
+ * b/c we are beyond the window boundary.
+ *
+ * We know we don't have the wantseq so that counts as a drop.
+ */
+
+ /* ex: slot count is 4, array size is 3 savedlen is 2, slot 0 is the
+ * missing sequence number.
+ *
+ * the final slot at savedlen (index savedlen - 1) is always occupied.
+ *
+ * beyond is "beyond array size" not savedlen.
+ *
+ * +--------- array length (savedlen == 2)
+ * | +----- array size (nslots - 1 == 3)
+ * | | +- window boundary (nslots == 4)
+ * V V | V
+ * |
+ * 0 1 2 3 | slot number
+ * --- 0 1 2 | array index
+ * [b] [c] : :| array
+ * |
+ * "2" "3" "4" "5"|*6* seq numbers
+ *
+ * We receive seq number 6
+ * distance == 4 [inseq(6) - w_wantseq(2)]
+ * newslot == distance
+ * index == 3 [distance(4) - 1]
+ * beyond == 1 [newslot(4) - lastslot((nslots(4) - 1))]
+ * shifting == 1 [min(savedlen(2), beyond(1)]
+ * slot0_skb == [b], and should match w_wantseq
+ *
+ * +--- window boundary (nslots == 4)
+ * 0 1 2 3 | 4 slot number
+ * --- 0 1 2 | 3 array index
+ * [b] : : : :| array
+ * "2" "3" "4" "5" *6* seq numbers
+ *
+ * We receive seq number 6
+ * distance == 4 [inseq(6) - w_wantseq(2)]
+ * newslot == distance
+ * index == 3 [distance(4) - 1]
+ * beyond == 1 [newslot(4) - lastslot((nslots(4) - 1))]
+ * shifting == 1 [min(savedlen(1), beyond(1)]
+ * slot0_skb == [b] and should match w_wantseq
+ *
+ * +-- window boundary (nslots == 4)
+ * 0 1 2 3 | 4 5 6 slot number
+ * --- 0 1 2 | 3 4 5 array index
+ * [-] [c] : :| array
+ * "2" "3" "4" "5" "6" "7" *8* seq numbers
+ *
+ * savedlen = 2, beyond = 3
+ * iter 1: slot0 == NULL, missed++, lastdrop = 2 (2+1-1), slot0 = [-]
+ * iter 2: slot0 == NULL, missed++, lastdrop = 3 (2+2-1), slot0 = [c]
+ * 2 < 3, extra = 1 (3-2), missed += extra, lastdrop = 4 (2+2+1-1)
+ *
+ * We receive seq number 8
+ * distance == 6 [inseq(8) - w_wantseq(2)]
+ * newslot == distance
+ * index == 5 [distance(6) - 1]
+ * beyond == 3 [newslot(6) - lastslot((nslots(4) - 1))]
+ * shifting == 2 [min(savedlen(2), beyond(3)]
+ *
+ * slot0_skb == NULL changed from [b] when "savedlen < beyond" is true.
+ */
+
+ /* Now send any packets that are being shifted out of saved, and account
+ * for missing packets that are exiting the window as we shift it.
+ */
+
+ /* If savedlen > beyond we are shifting some, else all. */
+ shifting = min(savedlen, beyond);
+
+ /* slot0 is the buf that just shifted out and into slot0 */
+ slot0 = NULL;
+ wnext = xtfs->w_saved;
+ for (slot = 1; slot <= shifting; slot++, wnext++) {
+ /* handle what was in slot0 before we occupy it */
+ if (slot0)
+ list_add_tail(&slot0->list, list);
+ slot0 = wnext->skb;
+ wnext->skb = NULL;
+ }
+
+ /* slot0 is now either NULL (in which case it's what we now are waiting
+ * for, or a buf in which case we need to handle it like we received it;
+ * however, we may be advancing past that buffer as well..
+ */
+
+ /* Handle case where we need to shift more than we had saved, slot0 will
+ * be NULL iff savedlen is 0, otherwise slot0 will always be
+ * non-NULL b/c we shifted the final element, which is always set if
+ * there is any saved, into slot0.
+ */
+ if (savedlen < beyond) {
+ if (savedlen == 0) {
+ BUG_ON(slot0);
+ } else {
+ BUG_ON(!slot0);
+ list_add_tail(&slot0->list, list);
+ }
+ slot0 = NULL;
+ /* slot0 has had an empty slot pushed into it */
+ }
+
+ /* Remove the entries */
+ __vec_shift(xtfs, beyond);
+
+ /* Advance want seq */
+ xtfs->w_wantseq += beyond;
+
+ /* Process drops here when implementing congestion control */
+
+ /* We've shifted. plug the packet in at the end. */
+ xtfs->w_savedlen = nslots - 1;
+ xtfs->w_saved[xtfs->w_savedlen - 1].skb = inskb;
+ iptfs_set_window_drop_times(xtfs, xtfs->w_savedlen - 1);
+
+ /* if we don't have a slot0 then we must wait for it */
+ if (!slot0)
+ return;
+
+ /* If slot0, seq must match new want seq */
+ BUG_ON(xtfs->w_wantseq != __esp_seq(slot0));
+
+ /* slot0 is valid, treat like we received expected. */
+ __reorder_this(xtfs, slot0, list);
+}
+
+/* Receive a new packet into the reorder window. Return a list of ordered
+ * packets from the window.
+ */
+static void iptfs_input_reorder(struct xfrm_iptfs_data *xtfs,
+ struct sk_buff *inskb, struct list_head *list,
+ struct list_head *freelist)
+{
+ const u32 nslots = xtfs->cfg.reorder_win_size + 1;
+ u64 inseq = __esp_seq(inskb);
+ u64 wantseq;
+
+ assert_spin_locked(&xtfs->drop_lock);
+
+ if (unlikely(!xtfs->w_seq_set)) {
+ xtfs->w_seq_set = true;
+ xtfs->w_wantseq = inseq;
+ }
+ wantseq = xtfs->w_wantseq;
+
+ if (likely(inseq == wantseq))
+ __reorder_this(xtfs, inskb, list);
+ else if (inseq < wantseq)
+ __reorder_past(xtfs, inskb, freelist);
+ else if ((inseq - wantseq) < nslots)
+ __reorder_future_fits(xtfs, inskb, freelist);
+ else
+ __reorder_future_shifts(xtfs, inskb, list, freelist);
}
/**
@@ -1273,23 +1654,92 @@ static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
*/
static enum hrtimer_restart iptfs_drop_timer(struct hrtimer *me)
{
+ struct sk_buff *skb, *next;
+ struct list_head freelist, list;
struct xfrm_iptfs_data *xtfs;
struct xfrm_state *x;
+ u32 count;
xtfs = container_of(me, typeof(*xtfs), drop_timer);
x = xtfs->x;
- /* Drop any in progress packet */
spin_lock(&xtfs->drop_lock);
+
+ INIT_LIST_HEAD(&list);
+ INIT_LIST_HEAD(&freelist);
+
+ /* Drop any in progress packet */
+
if (xtfs->ra_newskb) {
kfree_skb(xtfs->ra_newskb);
xtfs->ra_newskb = NULL;
}
+
+ /* Now drop as many packets as we should from the reordering window
+ * saved array
+ */
+ count = xtfs->w_savedlen ? __reorder_drop(xtfs, &list) : 0;
+
spin_unlock(&xtfs->drop_lock);
+ if (count) {
+ list_for_each_entry_safe(skb, next, &list, list) {
+ skb_list_del_init(skb);
+ iptfs_input_ordered(x, skb);
+ }
+ }
return HRTIMER_NORESTART;
}
+/**
+ * iptfs_input() - handle receipt of iptfs payload
+ * @x: xfrm state
+ * @skb: the packet
+ *
+ * We have an IPTFS payload order it if needed, then process newly in order
+ * packets.
+ *
+ * Return: -EINPROGRESS to inform xfrm_input to stop processing the skb.
+ */
+static int iptfs_input(struct xfrm_state *x, struct sk_buff *skb)
+{
+ struct list_head freelist, list;
+ struct xfrm_iptfs_data *xtfs = x->mode_data;
+ struct sk_buff *next;
+
+ /* Fast path for no reorder window. */
+ if (xtfs->cfg.reorder_win_size == 0) {
+ iptfs_input_ordered(x, skb);
+ goto done;
+ }
+
+ /* Fetch list of in-order packets from the reordering window as well as
+ * a list of buffers we need to now free.
+ */
+ INIT_LIST_HEAD(&list);
+ INIT_LIST_HEAD(&freelist);
+
+ spin_lock(&xtfs->drop_lock);
+ iptfs_input_reorder(xtfs, skb, &list, &freelist);
+ spin_unlock(&xtfs->drop_lock);
+
+ list_for_each_entry_safe(skb, next, &list, list) {
+ skb_list_del_init(skb);
+ iptfs_input_ordered(x, skb);
+ }
+
+ list_for_each_entry_safe(skb, next, &freelist, list) {
+ skb_list_del_init(skb);
+ kfree_skb(skb);
+ }
+done:
+ /* We always have dealt with the input SKB, either we are re-using it,
+ * or we have freed it. Return EINPROGRESS so that xfrm_input stops
+ * processing it.
+ */
+ return -EINPROGRESS;
+}
+
/* ================================= */
/* IPTFS Sending (ingress) Functions */
/* ================================= */
@@ -2096,11 +2546,24 @@ static int iptfs_user_init(struct net *net, struct xfrm_state *x,
xc = &xtfs->cfg;
xc->max_queue_size = IPTFS_DEFAULT_MAX_QUEUE_SIZE;
+ xc->reorder_win_size = IPTFS_DEFAULT_REORDER_WINDOW;
xtfs->drop_time_ns = IPTFS_DEFAULT_DROP_TIME_USECS * NSECS_IN_USEC;
xtfs->init_delay_ns = IPTFS_DEFAULT_INIT_DELAY_USECS * NSECS_IN_USEC;
if (attrs[XFRMA_IPTFS_DONT_FRAG])
xc->dont_frag = true;
+ if (attrs[XFRMA_IPTFS_REORDER_WINDOW])
+ xc->reorder_win_size =
+ nla_get_u16(attrs[XFRMA_IPTFS_REORDER_WINDOW]);
+ /* saved array is for saving 1..N seq nums from wantseq */
+ if (xc->reorder_win_size) {
+ xtfs->w_saved = kcalloc(xc->reorder_win_size,
+ sizeof(*xtfs->w_saved), GFP_KERNEL);
+ if (!xtfs->w_saved) {
+ NL_SET_ERR_MSG(extack, "Cannot alloc reorder window");
+ return -ENOMEM;
+ }
+ }
if (attrs[XFRMA_IPTFS_PKT_SIZE]) {
xc->pkt_size = nla_get_u32(attrs[XFRMA_IPTFS_PKT_SIZE]);
if (!xc->pkt_size) {
@@ -2139,6 +2602,7 @@ static unsigned int iptfs_sa_len(const struct xfrm_state *x)
if (x->dir == XFRM_SA_DIR_IN) {
l += nla_total_size(sizeof(u32)); /* drop time usec */
+ l += nla_total_size(sizeof(xc->reorder_win_size));
} else {
if (xc->dont_frag)
l += nla_total_size(0); /* dont-frag flag */
@@ -2161,6 +2625,11 @@ static int iptfs_copy_to_user(struct xfrm_state *x, struct sk_buff *skb)
q = xtfs->drop_time_ns;
(void)do_div(q, NSECS_IN_USEC);
ret = nla_put_u32(skb, XFRMA_IPTFS_DROP_TIME, q);
+ if (ret)
+ return ret;
+
+ ret = nla_put_u16(skb, XFRMA_IPTFS_REORDER_WINDOW,
+ xc->reorder_win_size);
} else {
if (xc->dont_frag) {
ret = nla_put_flag(skb, XFRMA_IPTFS_DONT_FRAG);
@@ -2220,6 +2689,14 @@ static int iptfs_clone(struct xfrm_state *x, struct xfrm_state *orig)
return -ENOMEM;
xtfs->ra_newskb = NULL;
+ if (xtfs->cfg.reorder_win_size) {
+ xtfs->w_saved = kcalloc(xtfs->cfg.reorder_win_size,
+ sizeof(*xtfs->w_saved), GFP_KERNEL);
+ if (!xtfs->w_saved) {
+ kfree_sensitive(xtfs);
+ return -ENOMEM;
+ }
+ }
__iptfs_init_state(x, xtfs);
@@ -2243,6 +2720,7 @@ static void iptfs_delete_state(struct xfrm_state *x)
{
struct xfrm_iptfs_data *xtfs = x->mode_data;
struct sk_buff_head list;
+ struct skb_wseq *s, *se;
struct sk_buff *skb;
if (!xtfs)
@@ -2264,6 +2742,11 @@ static void iptfs_delete_state(struct xfrm_state *x)
if (xtfs->ra_newskb)
kfree_skb(xtfs->ra_newskb);
+ for (s = xtfs->w_saved, se = s + xtfs->w_savedlen; s < se; s++)
+ if (s->skb)
+ kfree_skb(s->skb);
+
+ kfree_sensitive(xtfs->w_saved);
kfree_sensitive(xtfs);
module_put(x->mode_cbs->owner);
--
2.46.0
^ permalink raw reply related [flat|nested] 40+ messages in thread* [PATCH ipsec-next v8 16/16] xfrm: iptfs: add tracepoint functionality
2024-08-04 20:33 [PATCH ipsec-next v8 00/16] Add IP-TFS mode to xfrm Christian Hopps
` (14 preceding siblings ...)
2024-08-04 20:33 ` [PATCH ipsec-next v8 15/16] xfrm: iptfs: handle reordering of received packets Christian Hopps
@ 2024-08-04 20:33 ` Christian Hopps
15 siblings, 0 replies; 40+ messages in thread
From: Christian Hopps @ 2024-08-04 20:33 UTC (permalink / raw)
To: devel; +Cc: Steffen Klassert, netdev, Christian Hopps, Christian Hopps
From: Christian Hopps <chopps@labn.net>
Add tracepoints to the IP-TFS code.
Signed-off-by: Christian Hopps <chopps@labn.net>
---
net/xfrm/trace_iptfs.h | 218 +++++++++++++++++++++++++++++++++++++++++
net/xfrm/xfrm_iptfs.c | 70 ++++++++++++-
2 files changed, 287 insertions(+), 1 deletion(-)
create mode 100644 net/xfrm/trace_iptfs.h
diff --git a/net/xfrm/trace_iptfs.h b/net/xfrm/trace_iptfs.h
new file mode 100644
index 000000000000..74391ba24445
--- /dev/null
+++ b/net/xfrm/trace_iptfs.h
@@ -0,0 +1,218 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/* xfrm_trace_iptfs.h
+ *
+ * August 12 2023, Christian Hopps <chopps@labn.net>
+ *
+ * Copyright (c) 2023, LabN Consulting, L.L.C.
+ */
+
+#undef TRACE_SYSTEM
+#define TRACE_SYSTEM iptfs
+
+#if !defined(_TRACE_IPTFS_H) || defined(TRACE_HEADER_MULTI_READ)
+#define _TRACE_IPTFS_H
+
+#include <linux/kernel.h>
+#include <linux/skbuff.h>
+#include <linux/tracepoint.h>
+#include <net/ip.h>
+
+struct xfrm_iptfs_data;
+
+TRACE_EVENT(iptfs_egress_recv,
+ TP_PROTO(struct sk_buff *skb, struct xfrm_iptfs_data *xtfs, u16 blkoff),
+ TP_ARGS(skb, xtfs, blkoff),
+ TP_STRUCT__entry(__field(struct sk_buff *, skb)
+ __field(void *, head)
+ __field(void *, head_pg_addr)
+ __field(void *, pg0addr)
+ __field(u32, skb_len)
+ __field(u32, data_len)
+ __field(u32, headroom)
+ __field(u32, tailroom)
+ __field(u32, tail)
+ __field(u32, end)
+ __field(u32, pg0off)
+ __field(u8, head_frag)
+ __field(u8, frag_list)
+ __field(u8, nr_frags)
+ __field(u16, blkoff)),
+ TP_fast_assign(__entry->skb = skb;
+ __entry->head = skb->head;
+ __entry->skb_len = skb->len;
+ __entry->data_len = skb->data_len;
+ __entry->headroom = skb_headroom(skb);
+ __entry->tailroom = skb_tailroom(skb);
+ __entry->tail = (u32)skb->tail;
+ __entry->end = (u32)skb->end;
+ __entry->head_frag = skb->head_frag;
+ __entry->frag_list = (bool)skb_shinfo(skb)->frag_list;
+ __entry->nr_frags = skb_shinfo(skb)->nr_frags;
+ __entry->blkoff = blkoff;
+ __entry->head_pg_addr = page_address(virt_to_head_page(skb->head));
+ __entry->pg0addr = (__entry->nr_frags
+ ? page_address(netmem_to_page(skb_shinfo(skb)->frags[0].netmem))
+ : NULL);
+ __entry->pg0off = (__entry->nr_frags
+ ? skb_shinfo(skb)->frags[0].offset
+ : 0);
+ ),
+ TP_printk("EGRESS: skb=%p len=%u data_len=%u headroom=%u head_frag=%u frag_list=%u nr_frags=%u blkoff=%u\n\t\ttailroom=%u tail=%u end=%u head=%p hdpgaddr=%p pg0->addr=%p pg0->data=%p pg0->off=%u",
+ __entry->skb, __entry->skb_len, __entry->data_len, __entry->headroom,
+ __entry->head_frag, __entry->frag_list, __entry->nr_frags, __entry->blkoff,
+ __entry->tailroom, __entry->tail, __entry->end, __entry->head,
+ __entry->head_pg_addr, __entry->pg0addr, __entry->pg0addr + __entry->pg0off,
+ __entry->pg0off)
+ )
+
+DECLARE_EVENT_CLASS(iptfs_ingress_preq_event,
+ TP_PROTO(struct sk_buff *skb, struct xfrm_iptfs_data *xtfs,
+ u32 pmtu, u8 was_gso),
+ TP_ARGS(skb, xtfs, pmtu, was_gso),
+ TP_STRUCT__entry(__field(struct sk_buff *, skb)
+ __field(u32, skb_len)
+ __field(u32, data_len)
+ __field(u32, pmtu)
+ __field(u32, queue_size)
+ __field(u32, proto_seq)
+ __field(u8, proto)
+ __field(u8, was_gso)
+ ),
+ TP_fast_assign(__entry->skb = skb;
+ __entry->skb_len = skb->len;
+ __entry->data_len = skb->data_len;
+ __entry->queue_size =
+ xtfs->cfg.max_queue_size - xtfs->queue_size;
+ __entry->proto = __trace_ip_proto(ip_hdr(skb));
+ __entry->proto_seq = __trace_ip_proto_seq(ip_hdr(skb));
+ __entry->pmtu = pmtu;
+ __entry->was_gso = was_gso;
+ ),
+ TP_printk("INGRPREQ: skb=%p len=%u data_len=%u qsize=%u proto=%u proto_seq=%u pmtu=%u was_gso=%u",
+ __entry->skb, __entry->skb_len, __entry->data_len,
+ __entry->queue_size, __entry->proto, __entry->proto_seq,
+ __entry->pmtu, __entry->was_gso));
+
+DEFINE_EVENT(iptfs_ingress_preq_event, iptfs_enqueue,
+ TP_PROTO(struct sk_buff *skb, struct xfrm_iptfs_data *xtfs, u32 pmtu, u8 was_gso),
+ TP_ARGS(skb, xtfs, pmtu, was_gso));
+
+DEFINE_EVENT(iptfs_ingress_preq_event, iptfs_no_queue_space,
+ TP_PROTO(struct sk_buff *skb, struct xfrm_iptfs_data *xtfs, u32 pmtu, u8 was_gso),
+ TP_ARGS(skb, xtfs, pmtu, was_gso));
+
+DEFINE_EVENT(iptfs_ingress_preq_event, iptfs_too_big,
+ TP_PROTO(struct sk_buff *skb, struct xfrm_iptfs_data *xtfs, u32 pmtu, u8 was_gso),
+ TP_ARGS(skb, xtfs, pmtu, was_gso));
+
+DECLARE_EVENT_CLASS(iptfs_ingress_postq_event,
+ TP_PROTO(struct sk_buff *skb, u32 mtu, u16 blkoff, struct iphdr *iph),
+ TP_ARGS(skb, mtu, blkoff, iph),
+ TP_STRUCT__entry(__field(struct sk_buff *, skb)
+ __field(u32, skb_len)
+ __field(u32, data_len)
+ __field(u32, mtu)
+ __field(u32, proto_seq)
+ __field(u16, blkoff)
+ __field(u8, proto)),
+ TP_fast_assign(__entry->skb = skb;
+ __entry->skb_len = skb->len;
+ __entry->data_len = skb->data_len;
+ __entry->mtu = mtu;
+ __entry->blkoff = blkoff;
+ __entry->proto = iph ? __trace_ip_proto(iph) : 0;
+ __entry->proto_seq = iph ? __trace_ip_proto_seq(iph) : 0;
+ ),
+ TP_printk("INGRPSTQ: skb=%p len=%u data_len=%u mtu=%u blkoff=%u proto=%u proto_seq=%u",
+ __entry->skb, __entry->skb_len, __entry->data_len, __entry->mtu,
+ __entry->blkoff, __entry->proto, __entry->proto_seq));
+
+DEFINE_EVENT(iptfs_ingress_postq_event, iptfs_first_dequeue,
+ TP_PROTO(struct sk_buff *skb, u32 mtu, u16 blkoff,
+ struct iphdr *iph),
+ TP_ARGS(skb, mtu, blkoff, iph));
+
+DEFINE_EVENT(iptfs_ingress_postq_event, iptfs_first_fragmenting,
+ TP_PROTO(struct sk_buff *skb, u32 mtu, u16 blkoff,
+ struct iphdr *iph),
+ TP_ARGS(skb, mtu, blkoff, iph));
+
+DEFINE_EVENT(iptfs_ingress_postq_event, iptfs_first_final_fragment,
+ TP_PROTO(struct sk_buff *skb, u32 mtu, u16 blkoff,
+ struct iphdr *iph),
+ TP_ARGS(skb, mtu, blkoff, iph));
+
+DEFINE_EVENT(iptfs_ingress_postq_event, iptfs_first_toobig,
+ TP_PROTO(struct sk_buff *skb, u32 mtu, u16 blkoff,
+ struct iphdr *iph),
+ TP_ARGS(skb, mtu, blkoff, iph));
+
+TRACE_EVENT(iptfs_ingress_nth_peek,
+ TP_PROTO(struct sk_buff *skb, u32 remaining),
+ TP_ARGS(skb, remaining),
+ TP_STRUCT__entry(__field(struct sk_buff *, skb)
+ __field(u32, skb_len)
+ __field(u32, remaining)),
+ TP_fast_assign(__entry->skb = skb;
+ __entry->skb_len = skb->len;
+ __entry->remaining = remaining;
+ ),
+ TP_printk("INGRPSTQ: NTHPEEK: skb=%p len=%u remaining=%u",
+ __entry->skb, __entry->skb_len, __entry->remaining));
+
+TRACE_EVENT(iptfs_ingress_nth_add, TP_PROTO(struct sk_buff *skb, u8 share_ok),
+ TP_ARGS(skb, share_ok),
+ TP_STRUCT__entry(__field(struct sk_buff *, skb)
+ __field(u32, skb_len)
+ __field(u32, data_len)
+ __field(u8, share_ok)
+ __field(u8, head_frag)
+ __field(u8, pp_recycle)
+ __field(u8, cloned)
+ __field(u8, shared)
+ __field(u8, nr_frags)
+ __field(u8, frag_list)
+ ),
+ TP_fast_assign(__entry->skb = skb;
+ __entry->skb_len = skb->len;
+ __entry->data_len = skb->data_len;
+ __entry->share_ok = share_ok;
+ __entry->head_frag = skb->head_frag;
+ __entry->pp_recycle = skb->pp_recycle;
+ __entry->cloned = skb_cloned(skb);
+ __entry->shared = skb_shared(skb);
+ __entry->nr_frags = skb_shinfo(skb)->nr_frags;
+ __entry->frag_list = (bool)skb_shinfo(skb)->frag_list;
+ ),
+ TP_printk("INGRPSTQ: NTHADD: skb=%p len=%u data_len=%u share_ok=%u head_frag=%u pp_recycle=%u cloned=%u shared=%u nr_frags=%u frag_list=%u",
+ __entry->skb, __entry->skb_len, __entry->data_len, __entry->share_ok,
+ __entry->head_frag, __entry->pp_recycle, __entry->cloned, __entry->shared,
+ __entry->nr_frags, __entry->frag_list));
+
+DECLARE_EVENT_CLASS(iptfs_timer_event,
+ TP_PROTO(struct xfrm_iptfs_data *xtfs, u64 time_val),
+ TP_ARGS(xtfs, time_val),
+ TP_STRUCT__entry(__field(u64, time_val)
+ __field(u64, set_time)),
+ TP_fast_assign(__entry->time_val = time_val;
+ __entry->set_time = xtfs->iptfs_settime;
+ ),
+ TP_printk("TIMER: set_time=%llu time_val=%llu",
+ __entry->set_time, __entry->time_val));
+
+DEFINE_EVENT(iptfs_timer_event, iptfs_timer_start,
+ TP_PROTO(struct xfrm_iptfs_data *xtfs, u64 time_val),
+ TP_ARGS(xtfs, time_val));
+
+DEFINE_EVENT(iptfs_timer_event, iptfs_timer_expire,
+ TP_PROTO(struct xfrm_iptfs_data *xtfs, u64 time_val),
+ TP_ARGS(xtfs, time_val));
+
+#endif /* _TRACE_IPTFS_H */
+
+/* This part must be outside protection */
+#undef TRACE_INCLUDE_PATH
+#define TRACE_INCLUDE_PATH ../../net/xfrm
+#undef TRACE_INCLUDE_FILE
+#define TRACE_INCLUDE_FILE trace_iptfs
+#include <trace/define_trace.h>
diff --git a/net/xfrm/xfrm_iptfs.c b/net/xfrm/xfrm_iptfs.c
index 6d022935c9dc..c2fd8e1fc79d 100644
--- a/net/xfrm/xfrm_iptfs.c
+++ b/net/xfrm/xfrm_iptfs.c
@@ -19,6 +19,7 @@
#include <crypto/aead.h>
#include "xfrm_inout.h"
+#include "trace_iptfs.h"
/* IPTFS encap (header) values. */
#define IPTFS_SUBTYPE_BASIC 0
@@ -123,6 +124,7 @@ struct skb_wseq {
* @ecn_queue_size: octets above with ECN mark.
* @init_delay_ns: nanoseconds to wait to send initial IPTFS packet.
* @iptfs_timer: output timer.
+ * @iptfs_settime: time the output timer was set.
* @payload_mtu: max payload size.
* @w_seq_set: true after first seq received.
* @w_wantseq: waiting for this seq number as next to process (in order).
@@ -147,6 +149,7 @@ struct xfrm_iptfs_data {
u32 ecn_queue_size; /* octets above which ECN mark */
u64 init_delay_ns; /* nanoseconds */
struct hrtimer iptfs_timer; /* output timer */
+ time64_t iptfs_settime; /* time timer was set */
u32 payload_mtu; /* max payload size */
/* Tunnel input reordering */
@@ -173,6 +176,39 @@ static enum hrtimer_restart iptfs_drop_timer(struct hrtimer *me);
/* Utility Functions */
/* ================= */
+static u32 __trace_ip_proto(struct iphdr *iph)
+{
+ if (iph->version == 4)
+ return iph->protocol;
+ return ((struct ipv6hdr *)iph)->nexthdr;
+}
+
+static u32 __trace_ip_proto_seq(struct iphdr *iph)
+{
+ void *nexthdr;
+ u32 protocol = 0;
+
+ if (iph->version == 4) {
+ nexthdr = (void *)(iph + 1);
+ protocol = iph->protocol;
+ } else if (iph->version == 6) {
+ nexthdr = (void *)(((struct ipv6hdr *)(iph)) + 1);
+ protocol = ((struct ipv6hdr *)(iph))->nexthdr;
+ }
+ switch (protocol) {
+ case IPPROTO_ICMP:
+ return ntohs(((struct icmphdr *)nexthdr)->un.echo.sequence);
+ case IPPROTO_ICMPV6:
+ return ntohs(((struct icmp6hdr *)nexthdr)->icmp6_sequence);
+ case IPPROTO_TCP:
+ return ntohl(((struct tcphdr *)nexthdr)->seq);
+ case IPPROTO_UDP:
+ return ntohs(((struct udphdr *)nexthdr)->source);
+ default:
+ return 0;
+ }
+}
+
static u64 __esp_seq(struct sk_buff *skb)
{
u64 seq = ntohl(XFRM_SKB_CB(skb)->seq.input.low);
@@ -492,6 +528,13 @@ static int skb_copy_bits_seq(struct skb_seq_state *st, int offset, void *to,
}
}
+/* ================================== */
+/* IPTFS Trace Event Definitions */
+/* ================================== */
+
+#define CREATE_TRACE_POINTS
+#include "trace_iptfs.h"
+
/* ================================== */
/* IPTFS Receiving (egress) Functions */
/* ================================== */
@@ -986,6 +1029,8 @@ static void iptfs_input_ordered(struct xfrm_state *x, struct sk_buff *skb)
}
data = sizeof(*ipth);
+ trace_iptfs_egress_recv(skb, xtfs, be16_to_cpu(ipth->block_offset));
+
/* Set data past the basic header */
if (ipth->subtype == IPTFS_SUBTYPE_CC) {
/* Copy the rest of the CC header */
@@ -1883,6 +1928,7 @@ static int iptfs_output_collect(struct net *net, struct sock *sk,
*/
if (!ok) {
nospace:
+ trace_iptfs_no_queue_space(skb, xtfs, pmtu, was_gso);
XFRM_INC_STATS(net, LINUX_MIB_XFRMOUTNOQSPACE);
kfree_skb_reason(skb, SKB_DROP_REASON_FULL_RING);
continue;
@@ -1892,6 +1938,7 @@ static int iptfs_output_collect(struct net *net, struct sock *sk,
* enqueue.
*/
if (xtfs->cfg.dont_frag && iptfs_is_too_big(sk, skb, pmtu)) {
+ trace_iptfs_too_big(skb, xtfs, pmtu, was_gso);
kfree_skb_reason(skb, SKB_DROP_REASON_PKT_TOO_BIG);
continue;
}
@@ -1900,12 +1947,17 @@ static int iptfs_output_collect(struct net *net, struct sock *sk,
ok = iptfs_enqueue(xtfs, skb);
if (!ok)
goto nospace;
+
+ trace_iptfs_enqueue(skb, xtfs, pmtu, was_gso);
}
/* Start a delay timer if we don't have one yet */
- if (!hrtimer_is_queued(&xtfs->iptfs_timer))
+ if (!hrtimer_is_queued(&xtfs->iptfs_timer)) {
hrtimer_start(&xtfs->iptfs_timer, xtfs->init_delay_ns,
IPTFS_HRTIMER_MODE);
+ xtfs->iptfs_settime = ktime_get_raw_fast_ns();
+ trace_iptfs_timer_start(xtfs, xtfs->init_delay_ns);
+ }
spin_unlock_bh(&x->lock);
return 0;
@@ -1990,6 +2042,7 @@ static int iptfs_copy_create_frags(struct sk_buff **skbp,
struct sk_buff *nskb = *skbp;
u32 copy_len, offset;
u32 to_copy = skb->len - mtu;
+ u32 blkoff = 0;
int err = 0;
INIT_LIST_HEAD(&sublist);
@@ -2002,6 +2055,7 @@ static int iptfs_copy_create_frags(struct sk_buff **skbp,
to_copy = skb->len - offset;
while (to_copy) {
/* Send all but last fragment to allow agg. append */
+ trace_iptfs_first_fragmenting(nskb, mtu, to_copy, NULL);
list_add_tail(&nskb->list, &sublist);
/* FUTURE: if the packet has an odd/non-aligning length we could
@@ -2021,11 +2075,14 @@ static int iptfs_copy_create_frags(struct sk_buff **skbp,
iptfs_output_prepare_skb(nskb, to_copy);
offset += copy_len;
to_copy -= copy_len;
+ blkoff = to_copy;
}
skb_abort_seq_read(&skbseq);
/* return last fragment that will be unsent (or NULL) */
*skbp = nskb;
+ if (nskb)
+ trace_iptfs_first_final_fragment(nskb, mtu, blkoff, NULL);
/* trim the original skb to MTU */
if (!err)
@@ -2133,6 +2190,8 @@ static int iptfs_first_skb(struct sk_buff **skbp, struct xfrm_iptfs_data *xtfs,
/* We've split these up before queuing */
BUG_ON(skb_is_gso(skb));
+ trace_iptfs_first_dequeue(skb, mtu, 0, ip_hdr(skb));
+
/* Simple case -- it fits. `mtu` accounted for all the overhead
* including the basic IPTFS header.
*/
@@ -2233,6 +2292,7 @@ static void iptfs_output_queued(struct xfrm_state *x, struct sk_buff_head *list)
*/
XFRM_INC_STATS(xs_net(x), LINUX_MIB_XFRMOUTERROR);
+ trace_iptfs_first_toobig(skb, mtu, 0, ip_hdr(skb));
kfree_skb_reason(skb, SKB_DROP_REASON_PKT_TOO_BIG);
continue;
}
@@ -2280,6 +2340,7 @@ static void iptfs_output_queued(struct xfrm_state *x, struct sk_buff_head *list)
* case.
*/
while ((skb2 = skb_peek(list))) {
+ trace_iptfs_ingress_nth_peek(skb2, remaining);
if (skb2->len > remaining)
break;
@@ -2315,6 +2376,8 @@ static void iptfs_output_queued(struct xfrm_state *x, struct sk_buff_head *list)
skb->len += skb2->len;
remaining -= skb2->len;
+ trace_iptfs_ingress_nth_add(skb2, share_ok);
+
if (share_ok) {
iptfs_consume_frags(skb, skb2);
} else {
@@ -2338,6 +2401,7 @@ static enum hrtimer_restart iptfs_delay_timer(struct hrtimer *me)
struct sk_buff_head list;
struct xfrm_iptfs_data *xtfs;
struct xfrm_state *x;
+ time64_t settime;
xtfs = container_of(me, typeof(*xtfs), iptfs_timer);
x = xtfs->x;
@@ -2354,6 +2418,7 @@ static enum hrtimer_restart iptfs_delay_timer(struct hrtimer *me)
__skb_queue_head_init(&list);
skb_queue_splice_init(&xtfs->queue, &list);
xtfs->queue_size = 0;
+ settime = xtfs->iptfs_settime;
spin_unlock(&x->lock);
/* After the above unlock, packets can begin queuing again, and the
@@ -2362,6 +2427,9 @@ static enum hrtimer_restart iptfs_delay_timer(struct hrtimer *me)
* already).
*/
+ trace_iptfs_timer_expire(
+ xtfs, (unsigned long long)(ktime_get_raw_fast_ns() - settime));
+
iptfs_output_queued(x, &list);
return HRTIMER_NORESTART;
--
2.46.0
^ permalink raw reply related [flat|nested] 40+ messages in thread