[PATCH net-next 0/10] xfrm: remove flow cache

[PATCH net-next 0/10] xfrm: remove flow cache

[Florian Westphal]

After RCU-ification of ipsec packet path there are no major scalability
issues anymore without flow cache.

We still incur a performance hit, which comes mostly from the extra xfrm
dst allocation/freeing.
The last patch in the series adds a simple percpu cache to avoid the
extra allocation if a packet matched the same policies as last one.

The main concern with this is that we will see performance drops,
especially with large numbers of policies/SAs.

However, during hallway discussions at nfws 2017 it seemed the issues
with flow caching outweight the removal downsides, and that it
might be best to just 'remove it' and see where the practical issues
(if any) will appear.

It should now be possible to also remove the genid member in the policies
as we don't hold bundles for prolonged time anymore, but I think
this change is controversial (and intrusive) enough as-is, so defer
that to a later point in time.

Changes since last rfc:

- fix build failures due to implicit interrupt.h includes
- rework last patch (pcpu cache):
 * avoid xchg()
 * check policies for walk.dead = 1 instead of more costly bundle_ok().
 * flush pcpu bundles when sa/policies get removed, to allow module
   references to go away (suggested by Ilan Tayari)

Documentation/networking/ip-sysctl.txt         |    7
drivers/net/arcnet/arcdevice.h                 |    2
drivers/net/ethernet/amd/xgbe/xgbe.h           |    1
drivers/net/ethernet/synopsys/dwc-xlgmac-net.c |    1
drivers/net/ieee802154/ca8210.c                |    1
include/net/flow.h                             |   34 -
include/net/netns/xfrm.h                       |   11
include/net/xfrm.h                             |    9
net/core/Makefile                              |    1
net/ipv4/ip_vti.c                              |   31 -
net/ipv4/xfrm4_policy.c                        |   11
net/ipv6/ip6_vti.c                             |   31 -
net/ipv6/xfrm6_policy.c                        |   11
net/key/af_key.c                               |    6
net/xfrm/xfrm_device.c                         |    4
net/xfrm/xfrm_policy.c                         |  417 +++++++-----------
net/xfrm/xfrm_state.c                          |    5
net/xfrm/xfrm_user.c                           |    3
security/selinux/include/xfrm.h                |    4
include/net/flowcache.h                        |   25 -
net/core/flow.c                                |  516 -----------------------
 21 files changed, 173 insertions(+), 957 deletions(-)

[PATCH net-next 01/10] drivers: net: add missing interrupt.h include

[Florian Westphal]

these drivers use tasklets or irq apis, but don't include interrupt.h.
Once flow cache is removed the implicit interrupt.h inclusion goes away
which will break the build.

Signed-off-by: Florian Westphal <fw@strlen.de>
---
 drivers/net/arcnet/arcdevice.h                 | 2 +-
 drivers/net/ethernet/amd/xgbe/xgbe.h           | 1 +
 drivers/net/ethernet/synopsys/dwc-xlgmac-net.c | 1 +
 drivers/net/ieee802154/ca8210.c                | 1 +
 4 files changed, 4 insertions(+), 1 deletion(-)

diff --git a/drivers/net/arcnet/arcdevice.h b/drivers/net/arcnet/arcdevice.h
index cbb4f8566bbe..d09b2b46ab63 100644
--- a/drivers/net/arcnet/arcdevice.h
+++ b/drivers/net/arcnet/arcdevice.h
@@ -20,7 +20,7 @@
 #include <linux/if_arcnet.h>
 
 #ifdef __KERNEL__
-#include  <linux/irqreturn.h>
+#include <linux/interrupt.h>
 
 /*
  * RECON_THRESHOLD is the maximum number of RECON messages to receive
diff --git a/drivers/net/ethernet/amd/xgbe/xgbe.h b/drivers/net/ethernet/amd/xgbe/xgbe.h
index 0938294f640a..e9282c924621 100644
--- a/drivers/net/ethernet/amd/xgbe/xgbe.h
+++ b/drivers/net/ethernet/amd/xgbe/xgbe.h
@@ -129,6 +129,7 @@
 #include <net/dcbnl.h>
 #include <linux/completion.h>
 #include <linux/cpumask.h>
+#include <linux/interrupt.h>
 
 #define XGBE_DRV_NAME		"amd-xgbe"
 #define XGBE_DRV_VERSION	"1.0.3"
diff --git a/drivers/net/ethernet/synopsys/dwc-xlgmac-net.c b/drivers/net/ethernet/synopsys/dwc-xlgmac-net.c
index 3b91257683bc..e1b55b8fb8e0 100644
--- a/drivers/net/ethernet/synopsys/dwc-xlgmac-net.c
+++ b/drivers/net/ethernet/synopsys/dwc-xlgmac-net.c
@@ -17,6 +17,7 @@
 
 #include <linux/netdevice.h>
 #include <linux/tcp.h>
+#include <linux/interrupt.h>
 
 #include "dwc-xlgmac.h"
 #include "dwc-xlgmac-reg.h"
diff --git a/drivers/net/ieee802154/ca8210.c b/drivers/net/ieee802154/ca8210.c
index a626c539fb17..326243fae7e2 100644
--- a/drivers/net/ieee802154/ca8210.c
+++ b/drivers/net/ieee802154/ca8210.c
@@ -66,6 +66,7 @@
 #include <linux/spinlock.h>
 #include <linux/string.h>
 #include <linux/workqueue.h>
+#include <linux/interrupt.h>
 
 #include <net/ieee802154_netdev.h>
 #include <net/mac802154.h>
-- 
2.13.0

[PATCH net-next 02/10] vti: revert flush x-netns xfrm cache when vti interface is removed

[Florian Westphal]

flow cache is removed in next commit.

Signed-off-by: Florian Westphal <fw@strlen.de>
---
 net/ipv4/ip_vti.c  | 31 -------------------------------
 net/ipv6/ip6_vti.c | 31 -------------------------------
 2 files changed, 62 deletions(-)

diff --git a/net/ipv4/ip_vti.c b/net/ipv4/ip_vti.c
index 0192c255e508..5ed63d250950 100644
--- a/net/ipv4/ip_vti.c
+++ b/net/ipv4/ip_vti.c
@@ -584,33 +584,6 @@ static struct rtnl_link_ops vti_link_ops __read_mostly = {
 	.get_link_net	= ip_tunnel_get_link_net,
 };
 
-static bool is_vti_tunnel(const struct net_device *dev)
-{
-	return dev->netdev_ops == &vti_netdev_ops;
-}
-
-static int vti_device_event(struct notifier_block *unused,
-			    unsigned long event, void *ptr)
-{
-	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
-	struct ip_tunnel *tunnel = netdev_priv(dev);
-
-	if (!is_vti_tunnel(dev))
-		return NOTIFY_DONE;
-
-	switch (event) {
-	case NETDEV_DOWN:
-		if (!net_eq(tunnel->net, dev_net(dev)))
-			xfrm_garbage_collect(tunnel->net);
-		break;
-	}
-	return NOTIFY_DONE;
-}
-
-static struct notifier_block vti_notifier_block __read_mostly = {
-	.notifier_call = vti_device_event,
-};
-
 static int __init vti_init(void)
 {
 	const char *msg;
@@ -618,8 +591,6 @@ static int __init vti_init(void)
 
 	pr_info("IPv4 over IPsec tunneling driver\n");
 
-	register_netdevice_notifier(&vti_notifier_block);
-
 	msg = "tunnel device";
 	err = register_pernet_device(&vti_net_ops);
 	if (err < 0)
@@ -652,7 +623,6 @@ static int __init vti_init(void)
 xfrm_proto_esp_failed:
 	unregister_pernet_device(&vti_net_ops);
 pernet_dev_failed:
-	unregister_netdevice_notifier(&vti_notifier_block);
 	pr_err("vti init: failed to register %s\n", msg);
 	return err;
 }
@@ -664,7 +634,6 @@ static void __exit vti_fini(void)
 	xfrm4_protocol_deregister(&vti_ah4_protocol, IPPROTO_AH);
 	xfrm4_protocol_deregister(&vti_esp4_protocol, IPPROTO_ESP);
 	unregister_pernet_device(&vti_net_ops);
-	unregister_netdevice_notifier(&vti_notifier_block);
 }
 
 module_init(vti_init);
diff --git a/net/ipv6/ip6_vti.c b/net/ipv6/ip6_vti.c
index 486c2305f53c..79444a4bfd6d 100644
--- a/net/ipv6/ip6_vti.c
+++ b/net/ipv6/ip6_vti.c
@@ -1145,33 +1145,6 @@ static struct xfrm6_protocol vti_ipcomp6_protocol __read_mostly = {
 	.priority	=	100,
 };
 
-static bool is_vti6_tunnel(const struct net_device *dev)
-{
-	return dev->netdev_ops == &vti6_netdev_ops;
-}
-
-static int vti6_device_event(struct notifier_block *unused,
-			     unsigned long event, void *ptr)
-{
-	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
-	struct ip6_tnl *t = netdev_priv(dev);
-
-	if (!is_vti6_tunnel(dev))
-		return NOTIFY_DONE;
-
-	switch (event) {
-	case NETDEV_DOWN:
-		if (!net_eq(t->net, dev_net(dev)))
-			xfrm_garbage_collect(t->net);
-		break;
-	}
-	return NOTIFY_DONE;
-}
-
-static struct notifier_block vti6_notifier_block __read_mostly = {
-	.notifier_call = vti6_device_event,
-};
-
 /**
  * vti6_tunnel_init - register protocol and reserve needed resources
  *
@@ -1182,8 +1155,6 @@ static int __init vti6_tunnel_init(void)
 	const char *msg;
 	int err;
 
-	register_netdevice_notifier(&vti6_notifier_block);
-
 	msg = "tunnel device";
 	err = register_pernet_device(&vti6_net_ops);
 	if (err < 0)
@@ -1216,7 +1187,6 @@ static int __init vti6_tunnel_init(void)
 xfrm_proto_esp_failed:
 	unregister_pernet_device(&vti6_net_ops);
 pernet_dev_failed:
-	unregister_netdevice_notifier(&vti6_notifier_block);
 	pr_err("vti6 init: failed to register %s\n", msg);
 	return err;
 }
@@ -1231,7 +1201,6 @@ static void __exit vti6_tunnel_cleanup(void)
 	xfrm6_protocol_deregister(&vti_ah6_protocol, IPPROTO_AH);
 	xfrm6_protocol_deregister(&vti_esp6_protocol, IPPROTO_ESP);
 	unregister_pernet_device(&vti6_net_ops);
-	unregister_netdevice_notifier(&vti6_notifier_block);
 }
 
 module_init(vti6_tunnel_init);
-- 
2.13.0

[PATCH net-next 03/10] net: xfrm: revert to lower xfrm dst gc limit

[Florian Westphal]

revert c386578f1cdb4dac230395 ("xfrm: Let the flowcache handle its size by default.").

Once we remove flow cache, we don't have a flow cache limit anymore.
We must not allow (virtually) unlimited allocations of xfrm dst entries.
Revert back to the old xfrm dst gc limits.

Signed-off-by: Florian Westphal <fw@strlen.de>
---
 Documentation/networking/ip-sysctl.txt | 6 ++----
 net/ipv4/xfrm4_policy.c                | 2 +-
 net/ipv6/xfrm6_policy.c                | 2 +-
 3 files changed, 4 insertions(+), 6 deletions(-)

diff --git a/Documentation/networking/ip-sysctl.txt b/Documentation/networking/ip-sysctl.txt
index 974ab47ae53a..f485d553e65c 100644
--- a/Documentation/networking/ip-sysctl.txt
+++ b/Documentation/networking/ip-sysctl.txt
@@ -1291,8 +1291,7 @@ tag - INTEGER
 xfrm4_gc_thresh - INTEGER
 	The threshold at which we will start garbage collecting for IPv4
 	destination cache entries.  At twice this value the system will
-	refuse new allocations. The value must be set below the flowcache
-	limit (4096 * number of online cpus) to take effect.
+	refuse new allocations.
 
 igmp_link_local_mcast_reports - BOOLEAN
 	Enable IGMP reports for link local multicast groups in the
@@ -1778,8 +1777,7 @@ ratelimit - INTEGER
 xfrm6_gc_thresh - INTEGER
 	The threshold at which we will start garbage collecting for IPv6
 	destination cache entries.  At twice this value the system will
-	refuse new allocations. The value must be set below the flowcache
-	limit (4096 * number of online cpus) to take effect.
+	refuse new allocations.
 
 
 IPv6 Update by:
diff --git a/net/ipv4/xfrm4_policy.c b/net/ipv4/xfrm4_policy.c
index 71b4ecc195c7..19455a5fc328 100644
--- a/net/ipv4/xfrm4_policy.c
+++ b/net/ipv4/xfrm4_policy.c
@@ -266,7 +266,7 @@ static struct dst_ops xfrm4_dst_ops_template = {
 	.destroy =		xfrm4_dst_destroy,
 	.ifdown =		xfrm4_dst_ifdown,
 	.local_out =		__ip_local_out,
-	.gc_thresh =		INT_MAX,
+	.gc_thresh =		32768,
 };
 
 static const struct xfrm_policy_afinfo xfrm4_policy_afinfo = {
diff --git a/net/ipv6/xfrm6_policy.c b/net/ipv6/xfrm6_policy.c
index 79651bc71bf0..ae30dc4973e8 100644
--- a/net/ipv6/xfrm6_policy.c
+++ b/net/ipv6/xfrm6_policy.c
@@ -286,7 +286,7 @@ static struct dst_ops xfrm6_dst_ops_template = {
 	.destroy =		xfrm6_dst_destroy,
 	.ifdown =		xfrm6_dst_ifdown,
 	.local_out =		__ip6_local_out,
-	.gc_thresh =		INT_MAX,
+	.gc_thresh =		32768,
 };
 
 static const struct xfrm_policy_afinfo xfrm6_policy_afinfo = {
-- 
2.13.0

[PATCH net-next 04/10] xfrm_policy: bypass flow_cache_lookup

[Florian Westphal]

Instead of consulting flow cache, call the xfrm bundle/policy lookup
functions directly.  This pretends the flow cache had no entry.

This helps to gradually remove flow cache integration,
followup commit will remove the dead code that this change adds.

Signed-off-by: Florian Westphal <fw@strlen.de>
---
 net/xfrm/xfrm_policy.c | 14 +++++---------
 1 file changed, 5 insertions(+), 9 deletions(-)

diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c
index ff61d8557929..1c7126ab752c 100644
--- a/net/xfrm/xfrm_policy.c
+++ b/net/xfrm/xfrm_policy.c
@@ -2052,13 +2052,12 @@ static struct xfrm_dst *xfrm_create_dummy_bundle(struct net *net,
 }
 
 static struct flow_cache_object *
-xfrm_bundle_lookup(struct net *net, const struct flowi *fl, u16 family, u8 dir,
-		   struct flow_cache_object *oldflo, void *ctx)
+xfrm_bundle_lookup(struct net *net, const struct flowi *fl, u16 family, u8 dir, struct xfrm_flo *xflo)
 {
-	struct xfrm_flo *xflo = (struct xfrm_flo *)ctx;
 	struct xfrm_policy *pols[XFRM_POLICY_TYPE_MAX];
 	struct xfrm_dst *xdst, *new_xdst;
 	int num_pols = 0, num_xfrms = 0, i, err, pol_dead;
+	struct flow_cache_object *oldflo = NULL;
 
 	/* Check if the policies from old bundle are usable */
 	xdst = NULL;
@@ -2128,8 +2127,6 @@ xfrm_bundle_lookup(struct net *net, const struct flowi *fl, u16 family, u8 dir,
 		dst_release_immediate(&xdst->u.dst);
 	}
 
-	/* We do need to return one reference for original caller */
-	dst_hold(&new_xdst->u.dst);
 	return &new_xdst->flo;
 
 make_dummy_bundle:
@@ -2242,8 +2239,7 @@ struct dst_entry *xfrm_lookup(struct net *net, struct dst_entry *dst_orig,
 		    !net->xfrm.policy_count[XFRM_POLICY_OUT])
 			goto nopol;
 
-		flo = flow_cache_lookup(net, fl, family, dir,
-					xfrm_bundle_lookup, &xflo);
+		flo = xfrm_bundle_lookup(net, fl, family, dir, &xflo);
 		if (flo == NULL)
 			goto nopol;
 		if (IS_ERR(flo)) {
@@ -2489,8 +2485,8 @@ int __xfrm_policy_check(struct sock *sk, int dir, struct sk_buff *skb,
 	if (!pol) {
 		struct flow_cache_object *flo;
 
-		flo = flow_cache_lookup(net, &fl, family, fl_dir,
-					xfrm_policy_lookup, NULL);
+		flo = xfrm_policy_lookup(net, &fl, family, dir, NULL, NULL);
+
 		if (IS_ERR_OR_NULL(flo))
 			pol = ERR_CAST(flo);
 		else
-- 
2.13.0

[PATCH net-next 05/10] xfrm_policy: remove always true/false branches

[Florian Westphal]

after previous change oldflo and xdst are always NULL.
These branches were already removed by gcc, this doesn't change code.

Signed-off-by: Florian Westphal <fw@strlen.de>
---
 net/xfrm/xfrm_policy.c | 74 ++++++++++----------------------------------------
 1 file changed, 14 insertions(+), 60 deletions(-)

diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c
index 1c7126ab752c..19d457db3a09 100644
--- a/net/xfrm/xfrm_policy.c
+++ b/net/xfrm/xfrm_policy.c
@@ -2056,48 +2056,23 @@ xfrm_bundle_lookup(struct net *net, const struct flowi *fl, u16 family, u8 dir,
 {
 	struct xfrm_policy *pols[XFRM_POLICY_TYPE_MAX];
 	struct xfrm_dst *xdst, *new_xdst;
-	int num_pols = 0, num_xfrms = 0, i, err, pol_dead;
-	struct flow_cache_object *oldflo = NULL;
+	int num_pols = 0, num_xfrms = 0, err;
 
 	/* Check if the policies from old bundle are usable */
 	xdst = NULL;
-	if (oldflo) {
-		xdst = container_of(oldflo, struct xfrm_dst, flo);
-		num_pols = xdst->num_pols;
-		num_xfrms = xdst->num_xfrms;
-		pol_dead = 0;
-		for (i = 0; i < num_pols; i++) {
-			pols[i] = xdst->pols[i];
-			pol_dead |= pols[i]->walk.dead;
-		}
-		if (pol_dead) {
-			/* Mark DST_OBSOLETE_DEAD to fail the next
-			 * xfrm_dst_check()
-			 */
-			xdst->u.dst.obsolete = DST_OBSOLETE_DEAD;
-			dst_release_immediate(&xdst->u.dst);
-			xdst = NULL;
-			num_pols = 0;
-			num_xfrms = 0;
-			oldflo = NULL;
-		}
-	}
-
 	/* Resolve policies to use if we couldn't get them from
 	 * previous cache entry */
-	if (xdst == NULL) {
-		num_pols = 1;
-		pols[0] = __xfrm_policy_lookup(net, fl, family,
-					       flow_to_policy_dir(dir));
-		err = xfrm_expand_policies(fl, family, pols,
+	num_pols = 1;
+	pols[0] = __xfrm_policy_lookup(net, fl, family,
+				       flow_to_policy_dir(dir));
+	err = xfrm_expand_policies(fl, family, pols,
 					   &num_pols, &num_xfrms);
-		if (err < 0)
-			goto inc_error;
-		if (num_pols == 0)
-			return NULL;
-		if (num_xfrms <= 0)
-			goto make_dummy_bundle;
-	}
+	if (err < 0)
+		goto inc_error;
+	if (num_pols == 0)
+		return NULL;
+	if (num_xfrms <= 0)
+		goto make_dummy_bundle;
 
 	new_xdst = xfrm_resolve_and_create_bundle(pols, num_pols, fl, family,
 						  xflo->dst_orig);
@@ -2105,26 +2080,10 @@ xfrm_bundle_lookup(struct net *net, const struct flowi *fl, u16 family, u8 dir,
 		err = PTR_ERR(new_xdst);
 		if (err != -EAGAIN)
 			goto error;
-		if (oldflo == NULL)
-			goto make_dummy_bundle;
-		dst_hold(&xdst->u.dst);
-		return oldflo;
+		goto make_dummy_bundle;
 	} else if (new_xdst == NULL) {
 		num_xfrms = 0;
-		if (oldflo == NULL)
-			goto make_dummy_bundle;
-		xdst->num_xfrms = 0;
-		dst_hold(&xdst->u.dst);
-		return oldflo;
-	}
-
-	/* Kill the previous bundle */
-	if (xdst) {
-		/* The policies were stolen for newly generated bundle */
-		xdst->num_pols = 0;
-		/* Mark DST_OBSOLETE_DEAD to fail the next xfrm_dst_check() */
-		xdst->u.dst.obsolete = DST_OBSOLETE_DEAD;
-		dst_release_immediate(&xdst->u.dst);
+		goto make_dummy_bundle;
 	}
 
 	return &new_xdst->flo;
@@ -2148,12 +2107,7 @@ xfrm_bundle_lookup(struct net *net, const struct flowi *fl, u16 family, u8 dir,
 inc_error:
 	XFRM_INC_STATS(net, LINUX_MIB_XFRMOUTPOLERROR);
 error:
-	if (xdst != NULL) {
-		/* Mark DST_OBSOLETE_DEAD to fail the next xfrm_dst_check() */
-		xdst->u.dst.obsolete = DST_OBSOLETE_DEAD;
-		dst_release_immediate(&xdst->u.dst);
-	} else
-		xfrm_pols_put(pols, num_pols);
+	xfrm_pols_put(pols, num_pols);
 	return ERR_PTR(err);
 }
 
-- 
2.13.0

[PATCH net-next 06/10] xfrm_policy: kill flow to policy dir conversion

[Florian Westphal]

XFRM_POLICY_IN/OUT/FWD are identical to FLOW_DIR_*, so gcc already
removed this function as its just returns the argument.  Again, no
code change.

Signed-off-by: Florian Westphal <fw@strlen.de>
---
 net/xfrm/xfrm_policy.c | 46 ++++------------------------------------------
 1 file changed, 4 insertions(+), 42 deletions(-)

diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c
index 19d457db3a09..9f724a688475 100644
--- a/net/xfrm/xfrm_policy.c
+++ b/net/xfrm/xfrm_policy.c
@@ -1187,24 +1187,6 @@ __xfrm_policy_lookup(struct net *net, const struct flowi *fl, u16 family, u8 dir
 	return xfrm_policy_lookup_bytype(net, XFRM_POLICY_TYPE_MAIN, fl, family, dir);
 }
 
-static int flow_to_policy_dir(int dir)
-{
-	if (XFRM_POLICY_IN == FLOW_DIR_IN &&
-	    XFRM_POLICY_OUT == FLOW_DIR_OUT &&
-	    XFRM_POLICY_FWD == FLOW_DIR_FWD)
-		return dir;
-
-	switch (dir) {
-	default:
-	case FLOW_DIR_IN:
-		return XFRM_POLICY_IN;
-	case FLOW_DIR_OUT:
-		return XFRM_POLICY_OUT;
-	case FLOW_DIR_FWD:
-		return XFRM_POLICY_FWD;
-	}
-}
-
 static struct flow_cache_object *
 xfrm_policy_lookup(struct net *net, const struct flowi *fl, u16 family,
 		   u8 dir, struct flow_cache_object *old_obj, void *ctx)
@@ -1214,7 +1196,7 @@ xfrm_policy_lookup(struct net *net, const struct flowi *fl, u16 family,
 	if (old_obj)
 		xfrm_pol_put(container_of(old_obj, struct xfrm_policy, flo));
 
-	pol = __xfrm_policy_lookup(net, fl, family, flow_to_policy_dir(dir));
+	pol = __xfrm_policy_lookup(net, fl, family, dir);
 	if (IS_ERR_OR_NULL(pol))
 		return ERR_CAST(pol);
 
@@ -1225,23 +1207,6 @@ xfrm_policy_lookup(struct net *net, const struct flowi *fl, u16 family,
 	return &pol->flo;
 }
 
-static inline int policy_to_flow_dir(int dir)
-{
-	if (XFRM_POLICY_IN == FLOW_DIR_IN &&
-	    XFRM_POLICY_OUT == FLOW_DIR_OUT &&
-	    XFRM_POLICY_FWD == FLOW_DIR_FWD)
-		return dir;
-	switch (dir) {
-	default:
-	case XFRM_POLICY_IN:
-		return FLOW_DIR_IN;
-	case XFRM_POLICY_OUT:
-		return FLOW_DIR_OUT;
-	case XFRM_POLICY_FWD:
-		return FLOW_DIR_FWD;
-	}
-}
-
 static struct xfrm_policy *xfrm_sk_policy_lookup(const struct sock *sk, int dir,
 						 const struct flowi *fl, u16 family)
 {
@@ -1261,7 +1226,7 @@ static struct xfrm_policy *xfrm_sk_policy_lookup(const struct sock *sk, int dir,
 			}
 			err = security_xfrm_policy_lookup(pol->security,
 						      fl->flowi_secid,
-						      policy_to_flow_dir(dir));
+						      dir);
 			if (!err) {
 				if (!xfrm_pol_hold_rcu(pol))
 					goto again;
@@ -2063,8 +2028,7 @@ xfrm_bundle_lookup(struct net *net, const struct flowi *fl, u16 family, u8 dir,
 	/* Resolve policies to use if we couldn't get them from
 	 * previous cache entry */
 	num_pols = 1;
-	pols[0] = __xfrm_policy_lookup(net, fl, family,
-				       flow_to_policy_dir(dir));
+	pols[0] = __xfrm_policy_lookup(net, fl, family, dir);
 	err = xfrm_expand_policies(fl, family, pols,
 					   &num_pols, &num_xfrms);
 	if (err < 0)
@@ -2142,7 +2106,7 @@ struct dst_entry *xfrm_lookup(struct net *net, struct dst_entry *dst_orig,
 	struct xfrm_dst *xdst;
 	struct dst_entry *dst, *route;
 	u16 family = dst_orig->ops->family;
-	u8 dir = policy_to_flow_dir(XFRM_POLICY_OUT);
+	u8 dir = XFRM_POLICY_OUT;
 	int i, err, num_pols, num_xfrms = 0, drop_pols = 0;
 
 	dst = NULL;
@@ -2399,12 +2363,10 @@ int __xfrm_policy_check(struct sock *sk, int dir, struct sk_buff *skb,
 	int pi;
 	int reverse;
 	struct flowi fl;
-	u8 fl_dir;
 	int xerr_idx = -1;
 
 	reverse = dir & ~XFRM_POLICY_MASK;
 	dir &= XFRM_POLICY_MASK;
-	fl_dir = policy_to_flow_dir(dir);
 
 	if (__xfrm_decode_session(skb, &fl, family, reverse) < 0) {
 		XFRM_INC_STATS(net, LINUX_MIB_XFRMINHDRERROR);
-- 
2.13.0

[PATCH net-next 07/10] xfrm_policy: remove xfrm_policy_lookup

[Florian Westphal]

This removes the wrapper and renames the __xfrm_policy_lookup variant
to get rid of another place that used flow cache objects.

Signed-off-by: Florian Westphal <fw@strlen.de>
---
 net/xfrm/xfrm_policy.c | 36 ++++--------------------------------
 1 file changed, 4 insertions(+), 32 deletions(-)

diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c
index 9f724a688475..339bb3ac2797 100644
--- a/net/xfrm/xfrm_policy.c
+++ b/net/xfrm/xfrm_policy.c
@@ -1175,7 +1175,7 @@ static struct xfrm_policy *xfrm_policy_lookup_bytype(struct net *net, u8 type,
 }
 
 static struct xfrm_policy *
-__xfrm_policy_lookup(struct net *net, const struct flowi *fl, u16 family, u8 dir)
+xfrm_policy_lookup(struct net *net, const struct flowi *fl, u16 family, u8 dir)
 {
 #ifdef CONFIG_XFRM_SUB_POLICY
 	struct xfrm_policy *pol;
@@ -1187,26 +1187,6 @@ __xfrm_policy_lookup(struct net *net, const struct flowi *fl, u16 family, u8 dir
 	return xfrm_policy_lookup_bytype(net, XFRM_POLICY_TYPE_MAIN, fl, family, dir);
 }
 
-static struct flow_cache_object *
-xfrm_policy_lookup(struct net *net, const struct flowi *fl, u16 family,
-		   u8 dir, struct flow_cache_object *old_obj, void *ctx)
-{
-	struct xfrm_policy *pol;
-
-	if (old_obj)
-		xfrm_pol_put(container_of(old_obj, struct xfrm_policy, flo));
-
-	pol = __xfrm_policy_lookup(net, fl, family, dir);
-	if (IS_ERR_OR_NULL(pol))
-		return ERR_CAST(pol);
-
-	/* Resolver returns two references:
-	 * one for cache and one for caller of flow_cache_lookup() */
-	xfrm_pol_hold(pol);
-
-	return &pol->flo;
-}
-
 static struct xfrm_policy *xfrm_sk_policy_lookup(const struct sock *sk, int dir,
 						 const struct flowi *fl, u16 family)
 {
@@ -2028,7 +2008,7 @@ xfrm_bundle_lookup(struct net *net, const struct flowi *fl, u16 family, u8 dir,
 	/* Resolve policies to use if we couldn't get them from
 	 * previous cache entry */
 	num_pols = 1;
-	pols[0] = __xfrm_policy_lookup(net, fl, family, dir);
+	pols[0] = xfrm_policy_lookup(net, fl, family, dir);
 	err = xfrm_expand_policies(fl, family, pols,
 					   &num_pols, &num_xfrms);
 	if (err < 0)
@@ -2398,16 +2378,8 @@ int __xfrm_policy_check(struct sock *sk, int dir, struct sk_buff *skb,
 		}
 	}
 
-	if (!pol) {
-		struct flow_cache_object *flo;
-
-		flo = xfrm_policy_lookup(net, &fl, family, dir, NULL, NULL);
-
-		if (IS_ERR_OR_NULL(flo))
-			pol = ERR_CAST(flo);
-		else
-			pol = container_of(flo, struct xfrm_policy, flo);
-	}
+	if (!pol)
+		pol = xfrm_policy_lookup(net, &fl, family, dir);
 
 	if (IS_ERR(pol)) {
 		XFRM_INC_STATS(net, LINUX_MIB_XFRMINPOLERROR);
-- 
2.13.0

[PATCH net-next 08/10] xfrm_policy: make xfrm_bundle_lookup return xfrm dst object

[Florian Westphal]

This allows to remove flow cache object embedded in struct xfrm_dst.

Signed-off-by: Florian Westphal <fw@strlen.de>
---
 net/xfrm/xfrm_policy.c | 28 ++++++++++++----------------
 1 file changed, 12 insertions(+), 16 deletions(-)

diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c
index 339bb3ac2797..145d2395f3c0 100644
--- a/net/xfrm/xfrm_policy.c
+++ b/net/xfrm/xfrm_policy.c
@@ -1996,15 +1996,13 @@ static struct xfrm_dst *xfrm_create_dummy_bundle(struct net *net,
 	goto out;
 }
 
-static struct flow_cache_object *
+static struct xfrm_dst *
 xfrm_bundle_lookup(struct net *net, const struct flowi *fl, u16 family, u8 dir, struct xfrm_flo *xflo)
 {
 	struct xfrm_policy *pols[XFRM_POLICY_TYPE_MAX];
-	struct xfrm_dst *xdst, *new_xdst;
 	int num_pols = 0, num_xfrms = 0, err;
+	struct xfrm_dst *xdst;
 
-	/* Check if the policies from old bundle are usable */
-	xdst = NULL;
 	/* Resolve policies to use if we couldn't get them from
 	 * previous cache entry */
 	num_pols = 1;
@@ -2018,19 +2016,19 @@ xfrm_bundle_lookup(struct net *net, const struct flowi *fl, u16 family, u8 dir,
 	if (num_xfrms <= 0)
 		goto make_dummy_bundle;
 
-	new_xdst = xfrm_resolve_and_create_bundle(pols, num_pols, fl, family,
+	xdst = xfrm_resolve_and_create_bundle(pols, num_pols, fl, family,
 						  xflo->dst_orig);
-	if (IS_ERR(new_xdst)) {
-		err = PTR_ERR(new_xdst);
+	if (IS_ERR(xdst)) {
+		err = PTR_ERR(xdst);
 		if (err != -EAGAIN)
 			goto error;
 		goto make_dummy_bundle;
-	} else if (new_xdst == NULL) {
+	} else if (xdst == NULL) {
 		num_xfrms = 0;
 		goto make_dummy_bundle;
 	}
 
-	return &new_xdst->flo;
+	return xdst;
 
 make_dummy_bundle:
 	/* We found policies, but there's no bundles to instantiate:
@@ -2046,7 +2044,7 @@ xfrm_bundle_lookup(struct net *net, const struct flowi *fl, u16 family, u8 dir,
 	memcpy(xdst->pols, pols, sizeof(struct xfrm_policy *) * num_pols);
 
 	dst_hold(&xdst->u.dst);
-	return &xdst->flo;
+	return xdst;
 
 inc_error:
 	XFRM_INC_STATS(net, LINUX_MIB_XFRMOUTPOLERROR);
@@ -2082,7 +2080,6 @@ struct dst_entry *xfrm_lookup(struct net *net, struct dst_entry *dst_orig,
 			      const struct sock *sk, int flags)
 {
 	struct xfrm_policy *pols[XFRM_POLICY_TYPE_MAX];
-	struct flow_cache_object *flo;
 	struct xfrm_dst *xdst;
 	struct dst_entry *dst, *route;
 	u16 family = dst_orig->ops->family;
@@ -2137,14 +2134,13 @@ struct dst_entry *xfrm_lookup(struct net *net, struct dst_entry *dst_orig,
 		    !net->xfrm.policy_count[XFRM_POLICY_OUT])
 			goto nopol;
 
-		flo = xfrm_bundle_lookup(net, fl, family, dir, &xflo);
-		if (flo == NULL)
+		xdst = xfrm_bundle_lookup(net, fl, family, dir, &xflo);
+		if (xdst == NULL)
 			goto nopol;
-		if (IS_ERR(flo)) {
-			err = PTR_ERR(flo);
+		if (IS_ERR(xdst)) {
+			err = PTR_ERR(xdst);
 			goto dropdst;
 		}
-		xdst = container_of(flo, struct xfrm_dst, flo);
 
 		num_pols = xdst->num_pols;
 		num_xfrms = xdst->num_xfrms;
-- 
2.13.0

[PATCH net-next 09/10] xfrm: remove flow cache

[Florian Westphal]

After rcu conversions performance degradation in forward tests isn't that
noticeable anymore.

See next patch for some numbers.

A followup patcg could then also remove genid from the policies
as we do not cache bundles anymore.

Signed-off-by: Florian Westphal <fw@strlen.de>
---
 include/net/flow.h              |  34 ---
 include/net/flowcache.h         |  25 --
 include/net/netns/xfrm.h        |  11 -
 include/net/xfrm.h              |   8 -
 net/core/Makefile               |   1 -
 net/core/flow.c                 | 516 ----------------------------------------
 net/ipv4/xfrm4_policy.c         |   9 -
 net/ipv6/xfrm6_policy.c         |   9 -
 net/key/af_key.c                |   6 -
 net/xfrm/xfrm_device.c          |   2 -
 net/xfrm/xfrm_policy.c          | 108 +--------
 net/xfrm/xfrm_user.c            |   3 -
 security/selinux/include/xfrm.h |   4 +-
 13 files changed, 2 insertions(+), 734 deletions(-)
 delete mode 100644 include/net/flowcache.h
 delete mode 100644 net/core/flow.c

diff --git a/include/net/flow.h b/include/net/flow.h
index bae198b3039e..f3dc61b29bb5 100644
--- a/include/net/flow.h
+++ b/include/net/flow.h
@@ -218,40 +218,6 @@ static inline unsigned int flow_key_size(u16 family)
 	return 0;
 }
 
-#define FLOW_DIR_IN	0
-#define FLOW_DIR_OUT	1
-#define FLOW_DIR_FWD	2
-
-struct net;
-struct sock;
-struct flow_cache_ops;
-
-struct flow_cache_object {
-	const struct flow_cache_ops *ops;
-};
-
-struct flow_cache_ops {
-	struct flow_cache_object *(*get)(struct flow_cache_object *);
-	int (*check)(struct flow_cache_object *);
-	void (*delete)(struct flow_cache_object *);
-};
-
-typedef struct flow_cache_object *(*flow_resolve_t)(
-		struct net *net, const struct flowi *key, u16 family,
-		u8 dir, struct flow_cache_object *oldobj, void *ctx);
-
-struct flow_cache_object *flow_cache_lookup(struct net *net,
-					    const struct flowi *key, u16 family,
-					    u8 dir, flow_resolve_t resolver,
-					    void *ctx);
-int flow_cache_init(struct net *net);
-void flow_cache_fini(struct net *net);
-void flow_cache_hp_init(void);
-
-void flow_cache_flush(struct net *net);
-void flow_cache_flush_deferred(struct net *net);
-extern atomic_t flow_cache_genid;
-
 __u32 __get_hash_from_flowi6(const struct flowi6 *fl6, struct flow_keys *keys);
 
 static inline __u32 get_hash_from_flowi6(const struct flowi6 *fl6)
diff --git a/include/net/flowcache.h b/include/net/flowcache.h
deleted file mode 100644
index 51eb971e8973..000000000000
--- a/include/net/flowcache.h
+++ /dev/null
@@ -1,25 +0,0 @@
-#ifndef _NET_FLOWCACHE_H
-#define _NET_FLOWCACHE_H
-
-#include <linux/interrupt.h>
-#include <linux/types.h>
-#include <linux/timer.h>
-#include <linux/notifier.h>
-
-struct flow_cache_percpu {
-	struct hlist_head		*hash_table;
-	unsigned int			hash_count;
-	u32				hash_rnd;
-	int				hash_rnd_recalc;
-	struct tasklet_struct		flush_tasklet;
-};
-
-struct flow_cache {
-	u32				hash_shift;
-	struct flow_cache_percpu __percpu *percpu;
-	struct hlist_node		node;
-	unsigned int			low_watermark;
-	unsigned int			high_watermark;
-	struct timer_list		rnd_timer;
-};
-#endif	/* _NET_FLOWCACHE_H */
diff --git a/include/net/netns/xfrm.h b/include/net/netns/xfrm.h
index 27bb9633c69d..611521646dd4 100644
--- a/include/net/netns/xfrm.h
+++ b/include/net/netns/xfrm.h
@@ -6,7 +6,6 @@
 #include <linux/workqueue.h>
 #include <linux/xfrm.h>
 #include <net/dst_ops.h>
-#include <net/flowcache.h>
 
 struct ctl_table_header;
 
@@ -73,16 +72,6 @@ struct netns_xfrm {
 	spinlock_t xfrm_state_lock;
 	spinlock_t xfrm_policy_lock;
 	struct mutex xfrm_cfg_mutex;
-
-	/* flow cache part */
-	struct flow_cache	flow_cache_global;
-	atomic_t		flow_cache_genid;
-	struct list_head	flow_cache_gc_list;
-	atomic_t		flow_cache_gc_count;
-	spinlock_t		flow_cache_gc_lock;
-	struct work_struct	flow_cache_gc_work;
-	struct work_struct	flow_cache_flush_work;
-	struct mutex		flow_flush_sem;
 };
 
 #endif
diff --git a/include/net/xfrm.h b/include/net/xfrm.h
index c0916ab18d32..e0feba2ce76a 100644
--- a/include/net/xfrm.h
+++ b/include/net/xfrm.h
@@ -563,7 +563,6 @@ struct xfrm_policy {
 	refcount_t		refcnt;
 	struct timer_list	timer;
 
-	struct flow_cache_object flo;
 	atomic_t		genid;
 	u32			priority;
 	u32			index;
@@ -978,7 +977,6 @@ struct xfrm_dst {
 		struct rt6_info		rt6;
 	} u;
 	struct dst_entry *route;
-	struct flow_cache_object flo;
 	struct xfrm_policy *pols[XFRM_POLICY_TYPE_MAX];
 	int num_pols, num_xfrms;
 	u32 xfrm_genid;
@@ -1226,9 +1224,6 @@ static inline void xfrm_sk_free_policy(struct sock *sk)
 	}
 }
 
-void xfrm_garbage_collect(struct net *net);
-void xfrm_garbage_collect_deferred(struct net *net);
-
 #else
 
 static inline void xfrm_sk_free_policy(struct sock *sk) {}
@@ -1263,9 +1258,6 @@ static inline int xfrm6_policy_check_reverse(struct sock *sk, int dir,
 {
 	return 1;
 }
-static inline void xfrm_garbage_collect(struct net *net)
-{
-}
 #endif
 
 static __inline__
diff --git a/net/core/Makefile b/net/core/Makefile
index 79f9479e9658..d501c4278015 100644
--- a/net/core/Makefile
+++ b/net/core/Makefile
@@ -11,7 +11,6 @@ obj-y		     += dev.o ethtool.o dev_addr_lists.o dst.o netevent.o \
 			neighbour.o rtnetlink.o utils.o link_watch.o filter.o \
 			sock_diag.o dev_ioctl.o tso.o sock_reuseport.o
 
-obj-$(CONFIG_XFRM) += flow.o
 obj-y += net-sysfs.o
 obj-$(CONFIG_PROC_FS) += net-procfs.o
 obj-$(CONFIG_NET_PKTGEN) += pktgen.o
diff --git a/net/core/flow.c b/net/core/flow.c
deleted file mode 100644
index f7f5d1932a27..000000000000
--- a/net/core/flow.c
+++ /dev/null
@@ -1,516 +0,0 @@
-/* flow.c: Generic flow cache.
- *
- * Copyright (C) 2003 Alexey N. Kuznetsov (kuznet@ms2.inr.ac.ru)
- * Copyright (C) 2003 David S. Miller (davem@redhat.com)
- */
-
-#include <linux/kernel.h>
-#include <linux/module.h>
-#include <linux/list.h>
-#include <linux/jhash.h>
-#include <linux/interrupt.h>
-#include <linux/mm.h>
-#include <linux/random.h>
-#include <linux/init.h>
-#include <linux/slab.h>
-#include <linux/smp.h>
-#include <linux/completion.h>
-#include <linux/percpu.h>
-#include <linux/bitops.h>
-#include <linux/notifier.h>
-#include <linux/cpu.h>
-#include <linux/cpumask.h>
-#include <linux/mutex.h>
-#include <net/flow.h>
-#include <linux/atomic.h>
-#include <linux/security.h>
-#include <net/net_namespace.h>
-
-struct flow_cache_entry {
-	union {
-		struct hlist_node	hlist;
-		struct list_head	gc_list;
-	} u;
-	struct net			*net;
-	u16				family;
-	u8				dir;
-	u32				genid;
-	struct flowi			key;
-	struct flow_cache_object	*object;
-};
-
-struct flow_flush_info {
-	struct flow_cache		*cache;
-	atomic_t			cpuleft;
-	struct completion		completion;
-};
-
-static struct kmem_cache *flow_cachep __read_mostly;
-
-#define flow_cache_hash_size(cache)	(1U << (cache)->hash_shift)
-#define FLOW_HASH_RND_PERIOD		(10 * 60 * HZ)
-
-static void flow_cache_new_hashrnd(unsigned long arg)
-{
-	struct flow_cache *fc = (void *) arg;
-	int i;
-
-	for_each_possible_cpu(i)
-		per_cpu_ptr(fc->percpu, i)->hash_rnd_recalc = 1;
-
-	fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
-	add_timer(&fc->rnd_timer);
-}
-
-static int flow_entry_valid(struct flow_cache_entry *fle,
-				struct netns_xfrm *xfrm)
-{
-	if (atomic_read(&xfrm->flow_cache_genid) != fle->genid)
-		return 0;
-	if (fle->object && !fle->object->ops->check(fle->object))
-		return 0;
-	return 1;
-}
-
-static void flow_entry_kill(struct flow_cache_entry *fle,
-				struct netns_xfrm *xfrm)
-{
-	if (fle->object)
-		fle->object->ops->delete(fle->object);
-	kmem_cache_free(flow_cachep, fle);
-}
-
-static void flow_cache_gc_task(struct work_struct *work)
-{
-	struct list_head gc_list;
-	struct flow_cache_entry *fce, *n;
-	struct netns_xfrm *xfrm = container_of(work, struct netns_xfrm,
-						flow_cache_gc_work);
-
-	INIT_LIST_HEAD(&gc_list);
-	spin_lock_bh(&xfrm->flow_cache_gc_lock);
-	list_splice_tail_init(&xfrm->flow_cache_gc_list, &gc_list);
-	spin_unlock_bh(&xfrm->flow_cache_gc_lock);
-
-	list_for_each_entry_safe(fce, n, &gc_list, u.gc_list) {
-		flow_entry_kill(fce, xfrm);
-		atomic_dec(&xfrm->flow_cache_gc_count);
-	}
-}
-
-static void flow_cache_queue_garbage(struct flow_cache_percpu *fcp,
-				     unsigned int deleted,
-				     struct list_head *gc_list,
-				     struct netns_xfrm *xfrm)
-{
-	if (deleted) {
-		atomic_add(deleted, &xfrm->flow_cache_gc_count);
-		fcp->hash_count -= deleted;
-		spin_lock_bh(&xfrm->flow_cache_gc_lock);
-		list_splice_tail(gc_list, &xfrm->flow_cache_gc_list);
-		spin_unlock_bh(&xfrm->flow_cache_gc_lock);
-		schedule_work(&xfrm->flow_cache_gc_work);
-	}
-}
-
-static void __flow_cache_shrink(struct flow_cache *fc,
-				struct flow_cache_percpu *fcp,
-				unsigned int shrink_to)
-{
-	struct flow_cache_entry *fle;
-	struct hlist_node *tmp;
-	LIST_HEAD(gc_list);
-	unsigned int deleted = 0;
-	struct netns_xfrm *xfrm = container_of(fc, struct netns_xfrm,
-						flow_cache_global);
-	unsigned int i;
-
-	for (i = 0; i < flow_cache_hash_size(fc); i++) {
-		unsigned int saved = 0;
-
-		hlist_for_each_entry_safe(fle, tmp,
-					  &fcp->hash_table[i], u.hlist) {
-			if (saved < shrink_to &&
-			    flow_entry_valid(fle, xfrm)) {
-				saved++;
-			} else {
-				deleted++;
-				hlist_del(&fle->u.hlist);
-				list_add_tail(&fle->u.gc_list, &gc_list);
-			}
-		}
-	}
-
-	flow_cache_queue_garbage(fcp, deleted, &gc_list, xfrm);
-}
-
-static void flow_cache_shrink(struct flow_cache *fc,
-			      struct flow_cache_percpu *fcp)
-{
-	unsigned int shrink_to = fc->low_watermark / flow_cache_hash_size(fc);
-
-	__flow_cache_shrink(fc, fcp, shrink_to);
-}
-
-static void flow_new_hash_rnd(struct flow_cache *fc,
-			      struct flow_cache_percpu *fcp)
-{
-	get_random_bytes(&fcp->hash_rnd, sizeof(u32));
-	fcp->hash_rnd_recalc = 0;
-	__flow_cache_shrink(fc, fcp, 0);
-}
-
-static u32 flow_hash_code(struct flow_cache *fc,
-			  struct flow_cache_percpu *fcp,
-			  const struct flowi *key,
-			  unsigned int keysize)
-{
-	const u32 *k = (const u32 *) key;
-	const u32 length = keysize * sizeof(flow_compare_t) / sizeof(u32);
-
-	return jhash2(k, length, fcp->hash_rnd)
-		& (flow_cache_hash_size(fc) - 1);
-}
-
-/* I hear what you're saying, use memcmp.  But memcmp cannot make
- * important assumptions that we can here, such as alignment.
- */
-static int flow_key_compare(const struct flowi *key1, const struct flowi *key2,
-			    unsigned int keysize)
-{
-	const flow_compare_t *k1, *k1_lim, *k2;
-
-	k1 = (const flow_compare_t *) key1;
-	k1_lim = k1 + keysize;
-
-	k2 = (const flow_compare_t *) key2;
-
-	do {
-		if (*k1++ != *k2++)
-			return 1;
-	} while (k1 < k1_lim);
-
-	return 0;
-}
-
-struct flow_cache_object *
-flow_cache_lookup(struct net *net, const struct flowi *key, u16 family, u8 dir,
-		  flow_resolve_t resolver, void *ctx)
-{
-	struct flow_cache *fc = &net->xfrm.flow_cache_global;
-	struct flow_cache_percpu *fcp;
-	struct flow_cache_entry *fle, *tfle;
-	struct flow_cache_object *flo;
-	unsigned int keysize;
-	unsigned int hash;
-
-	local_bh_disable();
-	fcp = this_cpu_ptr(fc->percpu);
-
-	fle = NULL;
-	flo = NULL;
-
-	keysize = flow_key_size(family);
-	if (!keysize)
-		goto nocache;
-
-	/* Packet really early in init?  Making flow_cache_init a
-	 * pre-smp initcall would solve this.  --RR */
-	if (!fcp->hash_table)
-		goto nocache;
-
-	if (fcp->hash_rnd_recalc)
-		flow_new_hash_rnd(fc, fcp);
-
-	hash = flow_hash_code(fc, fcp, key, keysize);
-	hlist_for_each_entry(tfle, &fcp->hash_table[hash], u.hlist) {
-		if (tfle->net == net &&
-		    tfle->family == family &&
-		    tfle->dir == dir &&
-		    flow_key_compare(key, &tfle->key, keysize) == 0) {
-			fle = tfle;
-			break;
-		}
-	}
-
-	if (unlikely(!fle)) {
-		if (fcp->hash_count > fc->high_watermark)
-			flow_cache_shrink(fc, fcp);
-
-		if (atomic_read(&net->xfrm.flow_cache_gc_count) >
-		    2 * num_online_cpus() * fc->high_watermark) {
-			flo = ERR_PTR(-ENOBUFS);
-			goto ret_object;
-		}
-
-		fle = kmem_cache_alloc(flow_cachep, GFP_ATOMIC);
-		if (fle) {
-			fle->net = net;
-			fle->family = family;
-			fle->dir = dir;
-			memcpy(&fle->key, key, keysize * sizeof(flow_compare_t));
-			fle->object = NULL;
-			hlist_add_head(&fle->u.hlist, &fcp->hash_table[hash]);
-			fcp->hash_count++;
-		}
-	} else if (likely(fle->genid == atomic_read(&net->xfrm.flow_cache_genid))) {
-		flo = fle->object;
-		if (!flo)
-			goto ret_object;
-		flo = flo->ops->get(flo);
-		if (flo)
-			goto ret_object;
-	} else if (fle->object) {
-	        flo = fle->object;
-	        flo->ops->delete(flo);
-	        fle->object = NULL;
-	}
-
-nocache:
-	flo = NULL;
-	if (fle) {
-		flo = fle->object;
-		fle->object = NULL;
-	}
-	flo = resolver(net, key, family, dir, flo, ctx);
-	if (fle) {
-		fle->genid = atomic_read(&net->xfrm.flow_cache_genid);
-		if (!IS_ERR(flo))
-			fle->object = flo;
-		else
-			fle->genid--;
-	} else {
-		if (!IS_ERR_OR_NULL(flo))
-			flo->ops->delete(flo);
-	}
-ret_object:
-	local_bh_enable();
-	return flo;
-}
-EXPORT_SYMBOL(flow_cache_lookup);
-
-static void flow_cache_flush_tasklet(unsigned long data)
-{
-	struct flow_flush_info *info = (void *)data;
-	struct flow_cache *fc = info->cache;
-	struct flow_cache_percpu *fcp;
-	struct flow_cache_entry *fle;
-	struct hlist_node *tmp;
-	LIST_HEAD(gc_list);
-	unsigned int deleted = 0;
-	struct netns_xfrm *xfrm = container_of(fc, struct netns_xfrm,
-						flow_cache_global);
-	unsigned int i;
-
-	fcp = this_cpu_ptr(fc->percpu);
-	for (i = 0; i < flow_cache_hash_size(fc); i++) {
-		hlist_for_each_entry_safe(fle, tmp,
-					  &fcp->hash_table[i], u.hlist) {
-			if (flow_entry_valid(fle, xfrm))
-				continue;
-
-			deleted++;
-			hlist_del(&fle->u.hlist);
-			list_add_tail(&fle->u.gc_list, &gc_list);
-		}
-	}
-
-	flow_cache_queue_garbage(fcp, deleted, &gc_list, xfrm);
-
-	if (atomic_dec_and_test(&info->cpuleft))
-		complete(&info->completion);
-}
-
-/*
- * Return whether a cpu needs flushing.  Conservatively, we assume
- * the presence of any entries means the core may require flushing,
- * since the flow_cache_ops.check() function may assume it's running
- * on the same core as the per-cpu cache component.
- */
-static int flow_cache_percpu_empty(struct flow_cache *fc, int cpu)
-{
-	struct flow_cache_percpu *fcp;
-	unsigned int i;
-
-	fcp = per_cpu_ptr(fc->percpu, cpu);
-	for (i = 0; i < flow_cache_hash_size(fc); i++)
-		if (!hlist_empty(&fcp->hash_table[i]))
-			return 0;
-	return 1;
-}
-
-static void flow_cache_flush_per_cpu(void *data)
-{
-	struct flow_flush_info *info = data;
-	struct tasklet_struct *tasklet;
-
-	tasklet = &this_cpu_ptr(info->cache->percpu)->flush_tasklet;
-	tasklet->data = (unsigned long)info;
-	tasklet_schedule(tasklet);
-}
-
-void flow_cache_flush(struct net *net)
-{
-	struct flow_flush_info info;
-	cpumask_var_t mask;
-	int i, self;
-
-	/* Track which cpus need flushing to avoid disturbing all cores. */
-	if (!alloc_cpumask_var(&mask, GFP_KERNEL))
-		return;
-	cpumask_clear(mask);
-
-	/* Don't want cpus going down or up during this. */
-	get_online_cpus();
-	mutex_lock(&net->xfrm.flow_flush_sem);
-	info.cache = &net->xfrm.flow_cache_global;
-	for_each_online_cpu(i)
-		if (!flow_cache_percpu_empty(info.cache, i))
-			cpumask_set_cpu(i, mask);
-	atomic_set(&info.cpuleft, cpumask_weight(mask));
-	if (atomic_read(&info.cpuleft) == 0)
-		goto done;
-
-	init_completion(&info.completion);
-
-	local_bh_disable();
-	self = cpumask_test_and_clear_cpu(smp_processor_id(), mask);
-	on_each_cpu_mask(mask, flow_cache_flush_per_cpu, &info, 0);
-	if (self)
-		flow_cache_flush_tasklet((unsigned long)&info);
-	local_bh_enable();
-
-	wait_for_completion(&info.completion);
-
-done:
-	mutex_unlock(&net->xfrm.flow_flush_sem);
-	put_online_cpus();
-	free_cpumask_var(mask);
-}
-
-static void flow_cache_flush_task(struct work_struct *work)
-{
-	struct netns_xfrm *xfrm = container_of(work, struct netns_xfrm,
-						flow_cache_flush_work);
-	struct net *net = container_of(xfrm, struct net, xfrm);
-
-	flow_cache_flush(net);
-}
-
-void flow_cache_flush_deferred(struct net *net)
-{
-	schedule_work(&net->xfrm.flow_cache_flush_work);
-}
-
-static int flow_cache_cpu_prepare(struct flow_cache *fc, int cpu)
-{
-	struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
-	unsigned int sz = sizeof(struct hlist_head) * flow_cache_hash_size(fc);
-
-	if (!fcp->hash_table) {
-		fcp->hash_table = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu));
-		if (!fcp->hash_table) {
-			pr_err("NET: failed to allocate flow cache sz %u\n", sz);
-			return -ENOMEM;
-		}
-		fcp->hash_rnd_recalc = 1;
-		fcp->hash_count = 0;
-		tasklet_init(&fcp->flush_tasklet, flow_cache_flush_tasklet, 0);
-	}
-	return 0;
-}
-
-static int flow_cache_cpu_up_prep(unsigned int cpu, struct hlist_node *node)
-{
-	struct flow_cache *fc = hlist_entry_safe(node, struct flow_cache, node);
-
-	return flow_cache_cpu_prepare(fc, cpu);
-}
-
-static int flow_cache_cpu_dead(unsigned int cpu, struct hlist_node *node)
-{
-	struct flow_cache *fc = hlist_entry_safe(node, struct flow_cache, node);
-	struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
-
-	__flow_cache_shrink(fc, fcp, 0);
-	return 0;
-}
-
-int flow_cache_init(struct net *net)
-{
-	int i;
-	struct flow_cache *fc = &net->xfrm.flow_cache_global;
-
-	if (!flow_cachep)
-		flow_cachep = kmem_cache_create("flow_cache",
-						sizeof(struct flow_cache_entry),
-						0, SLAB_PANIC, NULL);
-	spin_lock_init(&net->xfrm.flow_cache_gc_lock);
-	INIT_LIST_HEAD(&net->xfrm.flow_cache_gc_list);
-	INIT_WORK(&net->xfrm.flow_cache_gc_work, flow_cache_gc_task);
-	INIT_WORK(&net->xfrm.flow_cache_flush_work, flow_cache_flush_task);
-	mutex_init(&net->xfrm.flow_flush_sem);
-	atomic_set(&net->xfrm.flow_cache_gc_count, 0);
-
-	fc->hash_shift = 10;
-	fc->low_watermark = 2 * flow_cache_hash_size(fc);
-	fc->high_watermark = 4 * flow_cache_hash_size(fc);
-
-	fc->percpu = alloc_percpu(struct flow_cache_percpu);
-	if (!fc->percpu)
-		return -ENOMEM;
-
-	if (cpuhp_state_add_instance(CPUHP_NET_FLOW_PREPARE, &fc->node))
-		goto err;
-
-	setup_timer(&fc->rnd_timer, flow_cache_new_hashrnd,
-		    (unsigned long) fc);
-	fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
-	add_timer(&fc->rnd_timer);
-
-	return 0;
-
-err:
-	for_each_possible_cpu(i) {
-		struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, i);
-		kfree(fcp->hash_table);
-		fcp->hash_table = NULL;
-	}
-
-	free_percpu(fc->percpu);
-	fc->percpu = NULL;
-
-	return -ENOMEM;
-}
-EXPORT_SYMBOL(flow_cache_init);
-
-void flow_cache_fini(struct net *net)
-{
-	int i;
-	struct flow_cache *fc = &net->xfrm.flow_cache_global;
-
-	del_timer_sync(&fc->rnd_timer);
-
-	cpuhp_state_remove_instance_nocalls(CPUHP_NET_FLOW_PREPARE, &fc->node);
-
-	for_each_possible_cpu(i) {
-		struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, i);
-		kfree(fcp->hash_table);
-		fcp->hash_table = NULL;
-	}
-
-	free_percpu(fc->percpu);
-	fc->percpu = NULL;
-}
-EXPORT_SYMBOL(flow_cache_fini);
-
-void __init flow_cache_hp_init(void)
-{
-	int ret;
-
-	ret = cpuhp_setup_state_multi(CPUHP_NET_FLOW_PREPARE,
-				      "net/flow:prepare",
-				      flow_cache_cpu_up_prep,
-				      flow_cache_cpu_dead);
-	WARN_ON(ret < 0);
-}
diff --git a/net/ipv4/xfrm4_policy.c b/net/ipv4/xfrm4_policy.c
index 19455a5fc328..4aefb149fe0a 100644
--- a/net/ipv4/xfrm4_policy.c
+++ b/net/ipv4/xfrm4_policy.c
@@ -213,14 +213,6 @@ _decode_session4(struct sk_buff *skb, struct flowi *fl, int reverse)
 	fl4->flowi4_tos = iph->tos;
 }
 
-static inline int xfrm4_garbage_collect(struct dst_ops *ops)
-{
-	struct net *net = container_of(ops, struct net, xfrm.xfrm4_dst_ops);
-
-	xfrm_garbage_collect_deferred(net);
-	return (dst_entries_get_slow(ops) > ops->gc_thresh * 2);
-}
-
 static void xfrm4_update_pmtu(struct dst_entry *dst, struct sock *sk,
 			      struct sk_buff *skb, u32 mtu)
 {
@@ -259,7 +251,6 @@ static void xfrm4_dst_ifdown(struct dst_entry *dst, struct net_device *dev,
 
 static struct dst_ops xfrm4_dst_ops_template = {
 	.family =		AF_INET,
-	.gc =			xfrm4_garbage_collect,
 	.update_pmtu =		xfrm4_update_pmtu,
 	.redirect =		xfrm4_redirect,
 	.cow_metrics =		dst_cow_metrics_generic,
diff --git a/net/ipv6/xfrm6_policy.c b/net/ipv6/xfrm6_policy.c
index ae30dc4973e8..f44b25a48478 100644
--- a/net/ipv6/xfrm6_policy.c
+++ b/net/ipv6/xfrm6_policy.c
@@ -214,14 +214,6 @@ _decode_session6(struct sk_buff *skb, struct flowi *fl, int reverse)
 	}
 }
 
-static inline int xfrm6_garbage_collect(struct dst_ops *ops)
-{
-	struct net *net = container_of(ops, struct net, xfrm.xfrm6_dst_ops);
-
-	xfrm_garbage_collect_deferred(net);
-	return dst_entries_get_fast(ops) > ops->gc_thresh * 2;
-}
-
 static void xfrm6_update_pmtu(struct dst_entry *dst, struct sock *sk,
 			      struct sk_buff *skb, u32 mtu)
 {
@@ -279,7 +271,6 @@ static void xfrm6_dst_ifdown(struct dst_entry *dst, struct net_device *dev,
 
 static struct dst_ops xfrm6_dst_ops_template = {
 	.family =		AF_INET6,
-	.gc =			xfrm6_garbage_collect,
 	.update_pmtu =		xfrm6_update_pmtu,
 	.redirect =		xfrm6_redirect,
 	.cow_metrics =		dst_cow_metrics_generic,
diff --git a/net/key/af_key.c b/net/key/af_key.c
index ca9d3ae665e7..10d7133e4fe9 100644
--- a/net/key/af_key.c
+++ b/net/key/af_key.c
@@ -2398,8 +2398,6 @@ static int pfkey_spddelete(struct sock *sk, struct sk_buff *skb, const struct sa
 
 out:
 	xfrm_pol_put(xp);
-	if (err == 0)
-		xfrm_garbage_collect(net);
 	return err;
 }
 
@@ -2650,8 +2648,6 @@ static int pfkey_spdget(struct sock *sk, struct sk_buff *skb, const struct sadb_
 
 out:
 	xfrm_pol_put(xp);
-	if (delete && err == 0)
-		xfrm_garbage_collect(net);
 	return err;
 }
 
@@ -2751,8 +2747,6 @@ static int pfkey_spdflush(struct sock *sk, struct sk_buff *skb, const struct sad
 	int err, err2;
 
 	err = xfrm_policy_flush(net, XFRM_POLICY_TYPE_MAIN, true);
-	if (!err)
-		xfrm_garbage_collect(net);
 	err2 = unicast_flush_resp(sk, hdr);
 	if (err || err2) {
 		if (err == -ESRCH) /* empty table - old silent behavior */
diff --git a/net/xfrm/xfrm_device.c b/net/xfrm/xfrm_device.c
index 5f7e8bfa0c2d..1f9a079e08b0 100644
--- a/net/xfrm/xfrm_device.c
+++ b/net/xfrm/xfrm_device.c
@@ -175,8 +175,6 @@ static int xfrm_dev_down(struct net_device *dev)
 	if (dev->features & NETIF_F_HW_ESP)
 		xfrm_dev_state_flush(dev_net(dev), dev, true);
 
-	xfrm_garbage_collect(dev_net(dev));
-
 	return NOTIFY_DONE;
 }
 
diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c
index 145d2395f3c0..0f1db4c18b22 100644
--- a/net/xfrm/xfrm_policy.c
+++ b/net/xfrm/xfrm_policy.c
@@ -246,36 +246,6 @@ static void xfrm_policy_timer(unsigned long data)
 	xfrm_pol_put(xp);
 }
 
-static struct flow_cache_object *xfrm_policy_flo_get(struct flow_cache_object *flo)
-{
-	struct xfrm_policy *pol = container_of(flo, struct xfrm_policy, flo);
-
-	if (unlikely(pol->walk.dead))
-		flo = NULL;
-	else
-		xfrm_pol_hold(pol);
-
-	return flo;
-}
-
-static int xfrm_policy_flo_check(struct flow_cache_object *flo)
-{
-	struct xfrm_policy *pol = container_of(flo, struct xfrm_policy, flo);
-
-	return !pol->walk.dead;
-}
-
-static void xfrm_policy_flo_delete(struct flow_cache_object *flo)
-{
-	xfrm_pol_put(container_of(flo, struct xfrm_policy, flo));
-}
-
-static const struct flow_cache_ops xfrm_policy_fc_ops = {
-	.get = xfrm_policy_flo_get,
-	.check = xfrm_policy_flo_check,
-	.delete = xfrm_policy_flo_delete,
-};
-
 /* Allocate xfrm_policy. Not used here, it is supposed to be used by pfkeyv2
  * SPD calls.
  */
@@ -298,7 +268,6 @@ struct xfrm_policy *xfrm_policy_alloc(struct net *net, gfp_t gfp)
 				(unsigned long)policy);
 		setup_timer(&policy->polq.hold_timer, xfrm_policy_queue_process,
 			    (unsigned long)policy);
-		policy->flo.ops = &xfrm_policy_fc_ops;
 	}
 	return policy;
 }
@@ -798,7 +767,6 @@ int xfrm_policy_insert(int dir, struct xfrm_policy *policy, int excl)
 	else
 		hlist_add_head(&policy->bydst, chain);
 	__xfrm_policy_link(policy, dir);
-	atomic_inc(&net->xfrm.flow_cache_genid);
 
 	/* After previous checking, family can either be AF_INET or AF_INET6 */
 	if (policy->family == AF_INET)
@@ -1490,58 +1458,6 @@ static int xfrm_get_tos(const struct flowi *fl, int family)
 	return tos;
 }
 
-static struct flow_cache_object *xfrm_bundle_flo_get(struct flow_cache_object *flo)
-{
-	struct xfrm_dst *xdst = container_of(flo, struct xfrm_dst, flo);
-	struct dst_entry *dst = &xdst->u.dst;
-
-	if (xdst->route == NULL) {
-		/* Dummy bundle - if it has xfrms we were not
-		 * able to build bundle as template resolution failed.
-		 * It means we need to try again resolving. */
-		if (xdst->num_xfrms > 0)
-			return NULL;
-	} else if (dst->flags & DST_XFRM_QUEUE) {
-		return NULL;
-	} else {
-		/* Real bundle */
-		if (stale_bundle(dst))
-			return NULL;
-	}
-
-	dst_hold(dst);
-	return flo;
-}
-
-static int xfrm_bundle_flo_check(struct flow_cache_object *flo)
-{
-	struct xfrm_dst *xdst = container_of(flo, struct xfrm_dst, flo);
-	struct dst_entry *dst = &xdst->u.dst;
-
-	if (!xdst->route)
-		return 0;
-	if (stale_bundle(dst))
-		return 0;
-
-	return 1;
-}
-
-static void xfrm_bundle_flo_delete(struct flow_cache_object *flo)
-{
-	struct xfrm_dst *xdst = container_of(flo, struct xfrm_dst, flo);
-	struct dst_entry *dst = &xdst->u.dst;
-
-	/* Mark DST_OBSOLETE_DEAD to fail the next xfrm_dst_check() */
-	dst->obsolete = DST_OBSOLETE_DEAD;
-	dst_release_immediate(dst);
-}
-
-static const struct flow_cache_ops xfrm_bundle_fc_ops = {
-	.get = xfrm_bundle_flo_get,
-	.check = xfrm_bundle_flo_check,
-	.delete = xfrm_bundle_flo_delete,
-};
-
 static inline struct xfrm_dst *xfrm_alloc_dst(struct net *net, int family)
 {
 	const struct xfrm_policy_afinfo *afinfo = xfrm_policy_get_afinfo(family);
@@ -1569,7 +1485,6 @@ static inline struct xfrm_dst *xfrm_alloc_dst(struct net *net, int family)
 		struct dst_entry *dst = &xdst->u.dst;
 
 		memset(dst + 1, 0, sizeof(*xdst) - sizeof(*dst));
-		xdst->flo.ops = &xfrm_bundle_fc_ops;
 	} else
 		xdst = ERR_PTR(-ENOBUFS);
 
@@ -2521,11 +2436,9 @@ static struct dst_entry *xfrm_dst_check(struct dst_entry *dst, u32 cookie)
 	 * notice.  That's what we are validating here via the
 	 * stale_bundle() check.
 	 *
-	 * When an xdst is removed from flow cache, DST_OBSOLETE_DEAD will
-	 * be marked on it.
 	 * When a dst is removed from the fib tree, DST_OBSOLETE_DEAD will
 	 * be marked on it.
-	 * Both will force stable_bundle() to fail on any xdst bundle with
+	 * This will force stale_bundle() to fail on any xdst bundle with
 	 * this dst linked in it.
 	 */
 	if (dst->obsolete < 0 && !stale_bundle(dst))
@@ -2565,18 +2478,6 @@ static struct dst_entry *xfrm_negative_advice(struct dst_entry *dst)
 	return dst;
 }
 
-void xfrm_garbage_collect(struct net *net)
-{
-	flow_cache_flush(net);
-}
-EXPORT_SYMBOL(xfrm_garbage_collect);
-
-void xfrm_garbage_collect_deferred(struct net *net)
-{
-	flow_cache_flush_deferred(net);
-}
-EXPORT_SYMBOL(xfrm_garbage_collect_deferred);
-
 static void xfrm_init_pmtu(struct dst_entry *dst)
 {
 	do {
@@ -2914,14 +2815,9 @@ static int __net_init xfrm_net_init(struct net *net)
 	rv = xfrm_sysctl_init(net);
 	if (rv < 0)
 		goto out_sysctl;
-	rv = flow_cache_init(net);
-	if (rv < 0)
-		goto out;
 
 	return 0;
 
-out:
-	xfrm_sysctl_fini(net);
 out_sysctl:
 	xfrm_policy_fini(net);
 out_policy:
@@ -2934,7 +2830,6 @@ static int __net_init xfrm_net_init(struct net *net)
 
 static void __net_exit xfrm_net_exit(struct net *net)
 {
-	flow_cache_fini(net);
 	xfrm_sysctl_fini(net);
 	xfrm_policy_fini(net);
 	xfrm_state_fini(net);
@@ -2948,7 +2843,6 @@ static struct pernet_operations __net_initdata xfrm_net_ops = {
 
 void __init xfrm_init(void)
 {
-	flow_cache_hp_init();
 	register_pernet_subsys(&xfrm_net_ops);
 	seqcount_init(&xfrm_policy_hash_generation);
 	xfrm_input_init();
diff --git a/net/xfrm/xfrm_user.c b/net/xfrm/xfrm_user.c
index 2be4c6af008a..1b539b7dcfab 100644
--- a/net/xfrm/xfrm_user.c
+++ b/net/xfrm/xfrm_user.c
@@ -1815,8 +1815,6 @@ static int xfrm_get_policy(struct sk_buff *skb, struct nlmsghdr *nlh,
 
 out:
 	xfrm_pol_put(xp);
-	if (delete && err == 0)
-		xfrm_garbage_collect(net);
 	return err;
 }
 
@@ -2027,7 +2025,6 @@ static int xfrm_flush_policy(struct sk_buff *skb, struct nlmsghdr *nlh,
 			return 0;
 		return err;
 	}
-	xfrm_garbage_collect(net);
 
 	c.data.type = type;
 	c.event = nlh->nlmsg_type;
diff --git a/security/selinux/include/xfrm.h b/security/selinux/include/xfrm.h
index 1450f85b946d..36a7ce9e11ff 100644
--- a/security/selinux/include/xfrm.h
+++ b/security/selinux/include/xfrm.h
@@ -47,10 +47,8 @@ static inline void selinux_xfrm_notify_policyload(void)
 	struct net *net;
 
 	rtnl_lock();
-	for_each_net(net) {
-		atomic_inc(&net->xfrm.flow_cache_genid);
+	for_each_net(net)
 		rt_genid_bump_all(net);
-	}
 	rtnl_unlock();
 }
 #else
-- 
2.13.0

[PATCH net-next 10/10] xfrm: add xdst pcpu cache

[Florian Westphal]

retain last used xfrm_dst in a pcpu cache.
On next request, reuse this dst if the policies are the same.

The cache will not help with strict RR workloads as there is no hit.

The cache packet-path part is reasonably small, the notifier part is
needed so we do not add long hangs when a device is dismantled but some
pcpu xdst still holds a reference, there are also calls to the flush
operation when userspace deletes SAs so modules can be removed
(there is no hit.

We need to run the dst_release on the correct cpu to avoid races with
packet path.  This is done by adding a work_struct for each cpu and then
doing the actual test/release on each affected cpu via schedule_work_on().

Test results using 4 network namespaces and null encryption:

ns1           ns2          -> ns3           -> ns4
netperf -> xfrm/null enc   -> xfrm/null dec -> netserver

what                    TCP_STREAM      UDP_STREAM      UDP_RR
Flow cache:             14644.61        294.35          327231.64
No flow cache:		14349.81	242.64		202301.72
Pcpu cache:		14629.70	292.21		205595.22

UDP tests used 64byte packets, tests ran for one minute each,
value is average over ten iterations.

'Flow cache' is 'net-next', 'No flow cache' is net-next plus this
series but without this patch.

Signed-off-by: Florian Westphal <fw@strlen.de>
---
 include/net/xfrm.h     |   1 +
 net/xfrm/xfrm_device.c |   2 +
 net/xfrm/xfrm_policy.c | 127 ++++++++++++++++++++++++++++++++++++++++++++++++-
 net/xfrm/xfrm_state.c  |   5 +-
 4 files changed, 132 insertions(+), 3 deletions(-)

diff --git a/include/net/xfrm.h b/include/net/xfrm.h
index e0feba2ce76a..afb4929d7232 100644
--- a/include/net/xfrm.h
+++ b/include/net/xfrm.h
@@ -317,6 +317,7 @@ int xfrm_policy_register_afinfo(const struct xfrm_policy_afinfo *afinfo, int fam
 void xfrm_policy_unregister_afinfo(const struct xfrm_policy_afinfo *afinfo);
 void km_policy_notify(struct xfrm_policy *xp, int dir,
 		      const struct km_event *c);
+void xfrm_policy_cache_flush(void);
 void km_state_notify(struct xfrm_state *x, const struct km_event *c);
 
 struct xfrm_tmpl;
diff --git a/net/xfrm/xfrm_device.c b/net/xfrm/xfrm_device.c
index 1f9a079e08b0..5cd7a244e88d 100644
--- a/net/xfrm/xfrm_device.c
+++ b/net/xfrm/xfrm_device.c
@@ -153,6 +153,7 @@ static int xfrm_dev_register(struct net_device *dev)
 
 static int xfrm_dev_unregister(struct net_device *dev)
 {
+	xfrm_policy_cache_flush();
 	return NOTIFY_DONE;
 }
 
@@ -175,6 +176,7 @@ static int xfrm_dev_down(struct net_device *dev)
 	if (dev->features & NETIF_F_HW_ESP)
 		xfrm_dev_state_flush(dev_net(dev), dev, true);
 
+	xfrm_policy_cache_flush();
 	return NOTIFY_DONE;
 }
 
diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c
index 0f1db4c18b22..06c3bf7ab86b 100644
--- a/net/xfrm/xfrm_policy.c
+++ b/net/xfrm/xfrm_policy.c
@@ -24,6 +24,7 @@
 #include <linux/netfilter.h>
 #include <linux/module.h>
 #include <linux/cache.h>
+#include <linux/cpu.h>
 #include <linux/audit.h>
 #include <net/dst.h>
 #include <net/flow.h>
@@ -44,6 +45,8 @@ struct xfrm_flo {
 	u8 flags;
 };
 
+static DEFINE_PER_CPU(struct xfrm_dst *, xfrm_last_dst);
+static struct work_struct *xfrm_pcpu_work __read_mostly;
 static DEFINE_SPINLOCK(xfrm_policy_afinfo_lock);
 static struct xfrm_policy_afinfo const __rcu *xfrm_policy_afinfo[AF_INET6 + 1]
 						__read_mostly;
@@ -972,6 +975,8 @@ int xfrm_policy_flush(struct net *net, u8 type, bool task_valid)
 	}
 	if (!cnt)
 		err = -ESRCH;
+	else
+		xfrm_policy_cache_flush();
 out:
 	spin_unlock_bh(&net->xfrm.xfrm_policy_lock);
 	return err;
@@ -1700,6 +1705,102 @@ static int xfrm_expand_policies(const struct flowi *fl, u16 family,
 
 }
 
+static void xfrm_last_dst_update(struct xfrm_dst *xdst, struct xfrm_dst *old)
+{
+	this_cpu_write(xfrm_last_dst, xdst);
+	if (old)
+		dst_release(&old->u.dst);
+}
+
+static void __xfrm_pcpu_work_fn(void)
+{
+	struct xfrm_dst *old;
+
+	old = this_cpu_read(xfrm_last_dst);
+	if (old && !xfrm_bundle_ok(old))
+		xfrm_last_dst_update(NULL, old);
+}
+
+static void xfrm_pcpu_work_fn(struct work_struct *work)
+{
+	local_bh_disable();
+	rcu_read_lock();
+	__xfrm_pcpu_work_fn();
+	rcu_read_unlock();
+	local_bh_enable();
+}
+
+void xfrm_policy_cache_flush(void)
+{
+	struct xfrm_dst *old;
+	bool found = 0;
+	int cpu;
+
+	local_bh_disable();
+	rcu_read_lock();
+	for_each_possible_cpu(cpu) {
+		old = per_cpu(xfrm_last_dst, cpu);
+		if (old && !xfrm_bundle_ok(old)) {
+			if (smp_processor_id() == cpu) {
+				__xfrm_pcpu_work_fn();
+				continue;
+			}
+			found = true;
+			break;
+		}
+	}
+
+	rcu_read_unlock();
+	local_bh_enable();
+
+	if (!found)
+		return;
+
+	get_online_cpus();
+
+	for_each_possible_cpu(cpu) {
+		bool bundle_release;
+
+		rcu_read_lock();
+		old = per_cpu(xfrm_last_dst, cpu);
+		bundle_release = old && !xfrm_bundle_ok(old);
+		rcu_read_unlock();
+
+		if (!bundle_release)
+			continue;
+
+		if (cpu_online(cpu)) {
+			schedule_work_on(cpu, &xfrm_pcpu_work[cpu]);
+			continue;
+		}
+
+		rcu_read_lock();
+		old = per_cpu(xfrm_last_dst, cpu);
+		if (old && !xfrm_bundle_ok(old)) {
+			per_cpu(xfrm_last_dst, cpu) = NULL;
+			dst_release(&old->u.dst);
+		}
+		rcu_read_unlock();
+	}
+
+	put_online_cpus();
+}
+
+static bool xfrm_pol_dead(struct xfrm_dst *xdst)
+{
+	unsigned int num_pols = xdst->num_pols;
+	unsigned int pol_dead = 0, i;
+
+	for (i = 0; i < num_pols; i++)
+		pol_dead |= xdst->pols[i]->walk.dead;
+
+	/* Mark DST_OBSOLETE_DEAD to fail the next xfrm_dst_check() */
+	if (pol_dead)
+		xdst->u.dst.obsolete = DST_OBSOLETE_DEAD;
+
+	return pol_dead;
+}
+
 static struct xfrm_dst *
 xfrm_resolve_and_create_bundle(struct xfrm_policy **pols, int num_pols,
 			       const struct flowi *fl, u16 family,
@@ -1707,10 +1808,22 @@ xfrm_resolve_and_create_bundle(struct xfrm_policy **pols, int num_pols,
 {
 	struct net *net = xp_net(pols[0]);
 	struct xfrm_state *xfrm[XFRM_MAX_DEPTH];
+	struct xfrm_dst *xdst, *old;
 	struct dst_entry *dst;
-	struct xfrm_dst *xdst;
 	int err;
 
+	xdst = this_cpu_read(xfrm_last_dst);
+	if (xdst &&
+	    xdst->u.dst.dev == dst_orig->dev &&
+	    xdst->num_pols == num_pols &&
+	    !xfrm_pol_dead(xdst) &&
+	    memcmp(xdst->pols, pols,
+		   sizeof(struct xfrm_policy *) * num_pols) == 0) {
+		dst_hold(&xdst->u.dst);
+		return xdst;
+	}
+
+	old = xdst;
 	/* Try to instantiate a bundle */
 	err = xfrm_tmpl_resolve(pols, num_pols, fl, xfrm, family);
 	if (err <= 0) {
@@ -1731,6 +1844,9 @@ xfrm_resolve_and_create_bundle(struct xfrm_policy **pols, int num_pols,
 	memcpy(xdst->pols, pols, sizeof(struct xfrm_policy *) * num_pols);
 	xdst->policy_genid = atomic_read(&pols[0]->genid);
 
+	atomic_set(&xdst->u.dst.__refcnt, 2);
+	xfrm_last_dst_update(xdst, old);
+
 	return xdst;
 }
 
@@ -2843,6 +2959,15 @@ static struct pernet_operations __net_initdata xfrm_net_ops = {
 
 void __init xfrm_init(void)
 {
+	int i;
+
+	xfrm_pcpu_work = kmalloc_array(NR_CPUS, sizeof(*xfrm_pcpu_work),
+				       GFP_KERNEL);
+	BUG_ON(!xfrm_pcpu_work);
+
+	for (i = 0; i < NR_CPUS; i++)
+		INIT_WORK(&xfrm_pcpu_work[i], xfrm_pcpu_work_fn);
+
 	register_pernet_subsys(&xfrm_net_ops);
 	seqcount_init(&xfrm_policy_hash_generation);
 	xfrm_input_init();
diff --git a/net/xfrm/xfrm_state.c b/net/xfrm/xfrm_state.c
index 6c0956d10db6..82cbbce69b79 100644
--- a/net/xfrm/xfrm_state.c
+++ b/net/xfrm/xfrm_state.c
@@ -724,9 +724,10 @@ int xfrm_state_flush(struct net *net, u8 proto, bool task_valid)
 			}
 		}
 	}
-	if (cnt)
+	if (cnt) {
 		err = 0;
-
+		xfrm_policy_cache_flush();
+	}
 out:
 	spin_unlock_bh(&net->xfrm.xfrm_state_lock);
 	return err;
-- 
2.13.0

Re: [PATCH net-next 0/10] xfrm: remove flow cache

[David Miller]

From: Florian Westphal <fw@strlen.de>
Date: Mon, 17 Jul 2017 13:57:17 +0200

> After RCU-ification of ipsec packet path there are no major scalability
> issues anymore without flow cache.
> 
> We still incur a performance hit, which comes mostly from the extra xfrm
> dst allocation/freeing.
> The last patch in the series adds a simple percpu cache to avoid the
> extra allocation if a packet matched the same policies as last one.
> 
> The main concern with this is that we will see performance drops,
> especially with large numbers of policies/SAs.
> 
> However, during hallway discussions at nfws 2017 it seemed the issues
> with flow caching outweight the removal downsides, and that it
> might be best to just 'remove it' and see where the practical issues
> (if any) will appear.
> 
> It should now be possible to also remove the genid member in the policies
> as we don't hold bundles for prolonged time anymore, but I think
> this change is controversial (and intrusive) enough as-is, so defer
> that to a later point in time.
> 
> Changes since last rfc:
> 
> - fix build failures due to implicit interrupt.h includes
> - rework last patch (pcpu cache):
>  * avoid xchg()
>  * check policies for walk.dead = 1 instead of more costly bundle_ok().
>  * flush pcpu bundles when sa/policies get removed, to allow module
>    references to go away (suggested by Ilan Tayari)

Steffen, I know you have some level of trepidation about this because
there is obviously some performance cost immediately for removing this
DoS problem.

Like the routing cache removal, most of the low hanging fruit will be
fixed shortly, and over time the bulk of the loss will be reparied one
way or another.

And, to me more importantly, killing it now gives a real incentive to
do the work for fixing that stuff now rather than later.

Therefore I have applied this series to net-next, thanks everyone.

Re: [PATCH net-next 0/10] xfrm: remove flow cache

[Kristian Evensen]

Hello,

On Tue, Jul 18, 2017 at 8:15 PM, David Miller <davem@davemloft.net> wrote:
> Steffen, I know you have some level of trepidation about this because
> there is obviously some performance cost immediately for removing this
> DoS problem.

In a project I am involved in, we are running ipsec (Strongswan) on
different mt7621-based routers. Each router is configured as an
initiator and has around ~30 tunnels to different responders (running
on misc. devices). Before the flow cache was removed (kernel 4.9), we
got a combined throughput of around 70Mbit/s for all tunnels on one
router. However, we recently switched to kernel 4.14 (4.14.48), and
the total throughput is somewhere around 57Mbit/s (best-case). I.e., a
drop of around 20%. Reverting the flow cache removal restores, as
expected, performance levels to that of kernel 4.9.

Carrying around a fairly large revert patch is not something we want,
we are more interested in trying to fix at least some of the
performance problems. However, we are not very experienced when it
comes to profiling the kernel code or the xfrm-code itself. Are there
any known areas we should take a special look at, or should we just
read-up on different profiling tools and get started?

Also, the revert went very smooth, which always makes me a bit
nervous. Are there any parts of the flow cache removal that should or
would require a bit of special care when reverted?

Thanks in advance for any help.

BR,
Kristian

Re: [PATCH net-next 0/10] xfrm: remove flow cache

[Florian Westphal]

Kristian Evensen <kristian.evensen@gmail.com> wrote:
> Hello,
> 
> On Tue, Jul 18, 2017 at 8:15 PM, David Miller <davem@davemloft.net> wrote:
> > Steffen, I know you have some level of trepidation about this because
> > there is obviously some performance cost immediately for removing this
> > DoS problem.
> 
> In a project I am involved in, we are running ipsec (Strongswan) on
> different mt7621-based routers. Each router is configured as an
> initiator and has around ~30 tunnels to different responders (running
> on misc. devices). Before the flow cache was removed (kernel 4.9), we
> got a combined throughput of around 70Mbit/s for all tunnels on one
> router. However, we recently switched to kernel 4.14 (4.14.48), and
> the total throughput is somewhere around 57Mbit/s (best-case). I.e., a
> drop of around 20%. Reverting the flow cache removal restores, as
> expected, performance levels to that of kernel 4.9.

Can you test attached patch?

I'd like to see how much the pcpu cache helps or if it actually hurts
in your setup.

Subject: [TEST PATCH 4.14.y] xfrm: remove pcpu policy cache

We need to re-evaluate if this still buys anything after indirect calls
got more expensive (retpolines).
When pcpu xdst exists, it has to be validated first (which needs
indirect calls).  So even if hit rate is good, it might be cheaper to
allocate a new xdst entry.

Furthermore, the current xdst cache needs to run with BH off, which
is also not needed when its removed.

Compile tested only.

Signed-off-by: Florian Westphal <fw@strlen.de>
---
 include/net/xfrm.h     |   1 -
 net/xfrm/xfrm_device.c |  10 ----
 net/xfrm/xfrm_policy.c | 138 +------------------------------------------------
 net/xfrm/xfrm_state.c  |   5 +-
 4 files changed, 3 insertions(+), 151 deletions(-)

diff --git a/include/net/xfrm.h b/include/net/xfrm.h
index db99efb2d1d0..bdf185ae93db 100644
--- a/include/net/xfrm.h
+++ b/include/net/xfrm.h
@@ -323,7 +323,6 @@ int xfrm_policy_register_afinfo(const struct xfrm_policy_afinfo *afinfo, int fam
 void xfrm_policy_unregister_afinfo(const struct xfrm_policy_afinfo *afinfo);
 void km_policy_notify(struct xfrm_policy *xp, int dir,
 		      const struct km_event *c);
-void xfrm_policy_cache_flush(void);
 void km_state_notify(struct xfrm_state *x, const struct km_event *c);
 
 struct xfrm_tmpl;
diff --git a/net/xfrm/xfrm_device.c b/net/xfrm/xfrm_device.c
index 30e5746085b8..4e458fd9236a 100644
--- a/net/xfrm/xfrm_device.c
+++ b/net/xfrm/xfrm_device.c
@@ -153,12 +153,6 @@ static int xfrm_dev_register(struct net_device *dev)
 	return NOTIFY_DONE;
 }
 
-static int xfrm_dev_unregister(struct net_device *dev)
-{
-	xfrm_policy_cache_flush();
-	return NOTIFY_DONE;
-}
-
 static int xfrm_dev_feat_change(struct net_device *dev)
 {
 	if ((dev->features & NETIF_F_HW_ESP) && !dev->xfrmdev_ops)
@@ -178,7 +172,6 @@ static int xfrm_dev_down(struct net_device *dev)
 	if (dev->features & NETIF_F_HW_ESP)
 		xfrm_dev_state_flush(dev_net(dev), dev, true);
 
-	xfrm_policy_cache_flush();
 	return NOTIFY_DONE;
 }
 
@@ -190,9 +183,6 @@ static int xfrm_dev_event(struct notifier_block *this, unsigned long event, void
 	case NETDEV_REGISTER:
 		return xfrm_dev_register(dev);
 
-	case NETDEV_UNREGISTER:
-		return xfrm_dev_unregister(dev);
-
 	case NETDEV_FEAT_CHANGE:
 		return xfrm_dev_feat_change(dev);
 
diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c
index 9c57d6a5816c..cdfb60a9820b 100644
--- a/net/xfrm/xfrm_policy.c
+++ b/net/xfrm/xfrm_policy.c
@@ -45,8 +45,6 @@ struct xfrm_flo {
 	u8 flags;
 };
 
-static DEFINE_PER_CPU(struct xfrm_dst *, xfrm_last_dst);
-static struct work_struct *xfrm_pcpu_work __read_mostly;
 static DEFINE_SPINLOCK(xfrm_policy_afinfo_lock);
 static struct xfrm_policy_afinfo const __rcu *xfrm_policy_afinfo[AF_INET6 + 1]
 						__read_mostly;
@@ -1715,108 +1713,6 @@ static int xfrm_expand_policies(const struct flowi *fl, u16 family,
 
 }
 
-static void xfrm_last_dst_update(struct xfrm_dst *xdst, struct xfrm_dst *old)
-{
-	this_cpu_write(xfrm_last_dst, xdst);
-	if (old)
-		dst_release(&old->u.dst);
-}
-
-static void __xfrm_pcpu_work_fn(void)
-{
-	struct xfrm_dst *old;
-
-	old = this_cpu_read(xfrm_last_dst);
-	if (old && !xfrm_bundle_ok(old))
-		xfrm_last_dst_update(NULL, old);
-}
-
-static void xfrm_pcpu_work_fn(struct work_struct *work)
-{
-	local_bh_disable();
-	rcu_read_lock();
-	__xfrm_pcpu_work_fn();
-	rcu_read_unlock();
-	local_bh_enable();
-}
-
-void xfrm_policy_cache_flush(void)
-{
-	struct xfrm_dst *old;
-	bool found = 0;
-	int cpu;
-
-	might_sleep();
-
-	local_bh_disable();
-	rcu_read_lock();
-	for_each_possible_cpu(cpu) {
-		old = per_cpu(xfrm_last_dst, cpu);
-		if (old && !xfrm_bundle_ok(old)) {
-			if (smp_processor_id() == cpu) {
-				__xfrm_pcpu_work_fn();
-				continue;
-			}
-			found = true;
-			break;
-		}
-	}
-
-	rcu_read_unlock();
-	local_bh_enable();
-
-	if (!found)
-		return;
-
-	get_online_cpus();
-
-	for_each_possible_cpu(cpu) {
-		bool bundle_release;
-
-		rcu_read_lock();
-		old = per_cpu(xfrm_last_dst, cpu);
-		bundle_release = old && !xfrm_bundle_ok(old);
-		rcu_read_unlock();
-
-		if (!bundle_release)
-			continue;
-
-		if (cpu_online(cpu)) {
-			schedule_work_on(cpu, &xfrm_pcpu_work[cpu]);
-			continue;
-		}
-
-		rcu_read_lock();
-		old = per_cpu(xfrm_last_dst, cpu);
-		if (old && !xfrm_bundle_ok(old)) {
-			per_cpu(xfrm_last_dst, cpu) = NULL;
-			dst_release(&old->u.dst);
-		}
-		rcu_read_unlock();
-	}
-
-	put_online_cpus();
-}
-
-static bool xfrm_xdst_can_reuse(struct xfrm_dst *xdst,
-				struct xfrm_state * const xfrm[],
-				int num)
-{
-	const struct dst_entry *dst = &xdst->u.dst;
-	int i;
-
-	if (xdst->num_xfrms != num)
-		return false;
-
-	for (i = 0; i < num; i++) {
-		if (!dst || dst->xfrm != xfrm[i])
-			return false;
-		dst = dst->child;
-	}
-
-	return xfrm_bundle_ok(xdst);
-}
-
 static struct xfrm_dst *
 xfrm_resolve_and_create_bundle(struct xfrm_policy **pols, int num_pols,
 			       const struct flowi *fl, u16 family,
@@ -1824,7 +1720,7 @@ xfrm_resolve_and_create_bundle(struct xfrm_policy **pols, int num_pols,
 {
 	struct net *net = xp_net(pols[0]);
 	struct xfrm_state *xfrm[XFRM_MAX_DEPTH];
-	struct xfrm_dst *xdst, *old;
+	struct xfrm_dst *xdst;
 	struct dst_entry *dst;
 	int err;
 
@@ -1836,21 +1732,6 @@ xfrm_resolve_and_create_bundle(struct xfrm_policy **pols, int num_pols,
 		return ERR_PTR(err);
 	}
 
-	xdst = this_cpu_read(xfrm_last_dst);
-	if (xdst &&
-	    xdst->u.dst.dev == dst_orig->dev &&
-	    xdst->num_pols == num_pols &&
-	    memcmp(xdst->pols, pols,
-		   sizeof(struct xfrm_policy *) * num_pols) == 0 &&
-	    xfrm_xdst_can_reuse(xdst, xfrm, err)) {
-		dst_hold(&xdst->u.dst);
-		while (err > 0)
-			xfrm_state_put(xfrm[--err]);
-		return xdst;
-	}
-
-	old = xdst;
-
 	dst = xfrm_bundle_create(pols[0], xfrm, err, fl, dst_orig);
 	if (IS_ERR(dst)) {
 		XFRM_INC_STATS(net, LINUX_MIB_XFRMOUTBUNDLEGENERROR);
@@ -1863,9 +1744,6 @@ xfrm_resolve_and_create_bundle(struct xfrm_policy **pols, int num_pols,
 	memcpy(xdst->pols, pols, sizeof(struct xfrm_policy *) * num_pols);
 	xdst->policy_genid = atomic_read(&pols[0]->genid);
 
-	atomic_set(&xdst->u.dst.__refcnt, 2);
-	xfrm_last_dst_update(xdst, old);
-
 	return xdst;
 }
 
@@ -2066,11 +1944,8 @@ xfrm_bundle_lookup(struct net *net, const struct flowi *fl, u16 family, u8 dir,
 	if (num_xfrms <= 0)
 		goto make_dummy_bundle;
 
-	local_bh_disable();
 	xdst = xfrm_resolve_and_create_bundle(pols, num_pols, fl, family,
 					      xflo->dst_orig);
-	local_bh_enable();
-
 	if (IS_ERR(xdst)) {
 		err = PTR_ERR(xdst);
 		if (err != -EAGAIN)
@@ -2157,11 +2032,9 @@ struct dst_entry *xfrm_lookup(struct net *net, struct dst_entry *dst_orig,
 				goto no_transform;
 			}
 
-			local_bh_disable();
 			xdst = xfrm_resolve_and_create_bundle(
 					pols, num_pols, fl,
 					family, dst_orig);
-			local_bh_enable();
 
 			if (IS_ERR(xdst)) {
 				xfrm_pols_put(pols, num_pols);
@@ -2982,15 +2855,6 @@ static struct pernet_operations __net_initdata xfrm_net_ops = {
 
 void __init xfrm_init(void)
 {
-	int i;
-
-	xfrm_pcpu_work = kmalloc_array(NR_CPUS, sizeof(*xfrm_pcpu_work),
-				       GFP_KERNEL);
-	BUG_ON(!xfrm_pcpu_work);
-
-	for (i = 0; i < NR_CPUS; i++)
-		INIT_WORK(&xfrm_pcpu_work[i], xfrm_pcpu_work_fn);
-
 	register_pernet_subsys(&xfrm_net_ops);
 	seqcount_init(&xfrm_policy_hash_generation);
 	xfrm_input_init();
diff --git a/net/xfrm/xfrm_state.c b/net/xfrm/xfrm_state.c
index 6c4ec69e11a0..bff47fd1519b 100644
--- a/net/xfrm/xfrm_state.c
+++ b/net/xfrm/xfrm_state.c
@@ -735,10 +735,9 @@ int xfrm_state_flush(struct net *net, u8 proto, bool task_valid)
 	}
 out:
 	spin_unlock_bh(&net->xfrm.xfrm_state_lock);
-	if (cnt) {
+	if (cnt)
 		err = 0;
-		xfrm_policy_cache_flush();
-	}
+
 	return err;
 }
 EXPORT_SYMBOL(xfrm_state_flush);
-- 
2.16.4

> Carrying around a fairly large revert patch is not something we want,
> we are more interested in trying to fix at least some of the
> performance problems. However, we are not very experienced when it
> comes to profiling the kernel code or the xfrm-code itself. Are there
> any known areas we should take a special look at, or should we just
> read-up on different profiling tools and get started?
> 
> Also, the revert went very smooth, which always makes me a bit
> nervous. Are there any parts of the flow cache removal that should or
> would require a bit of special care when reverted?
> 
> Thanks in advance for any help.
> 
> BR,
> Kristian

-- 
Florian Westphal <fw@strlen.de>
4096R/AD5FF600  2015-09-13
Key fingerprint = 80A9 20C5 B203 E069 F586  AE9F 7091 A8D9 AD5F F600
Phone: +49 151 11132303

Re: [PATCH net-next 0/10] xfrm: remove flow cache

[Kristian Evensen]

Hi,

On Wed, Jun 13, 2018 at 2:40 PM, Florian Westphal <fw@strlen.de> wrote:
> Can you test attached patch?
>
> I'd like to see how much the pcpu cache helps or if it actually hurts
> in your setup.
>
> Subject: [TEST PATCH 4.14.y] xfrm: remove pcpu policy cache
>
> We need to re-evaluate if this still buys anything after indirect calls
> got more expensive (retpolines).
> When pcpu xdst exists, it has to be validated first (which needs
> indirect calls).  So even if hit rate is good, it might be cheaper to
> allocate a new xdst entry.
>
> Furthermore, the current xdst cache needs to run with BH off, which
> is also not needed when its removed.
>
> Compile tested only.

Thanks! I will prepare a firmware for one of my devices tonight, start
testing tomorrow and report back when I have some results.

BR,
Kristian

Re: [PATCH net-next 0/10] xfrm: remove flow cache

[Kristian Evensen]

Hi,

On Wed, Jun 13, 2018 at 3:43 PM, Kristian Evensen
<kristian.evensen@gmail.com> wrote:
> Thanks! I will prepare a firmware for one of my devices tonight, start
> testing tomorrow and report back when I have some results.

We have been testing this patch over the weekend and it has a
significant impact on performance. In our tests, the throughput
reduction has been reduced from around -20% to -5%. We also see that
the overall throughput is independent of the number of tunnels, while
before the throughput was reduced as the number of tunnels increased.
We are still running our tests and will report back if we observe
anything new, but so far, so good.

BR,
Kristian