* [PATCH v5 0/4] Add packet recirculation
@ 2013-04-08 6:43 Simon Horman
2013-04-08 6:43 ` [PATCH 3/4] Allow recirculation without facets Simon Horman
[not found] ` <1365403431-18102-1-git-send-email-horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org>
0 siblings, 2 replies; 14+ messages in thread
From: Simon Horman @ 2013-04-08 6:43 UTC (permalink / raw)
To: dev-yBygre7rU0TnMu66kgdUjQ, netdev-u79uwXL29TY76Z2rM5mHXA
Cc: Isaku Yamahata, Ravi K
Recirculation is a technique to allow a frame to re-enter
frame processing. This is intended to be used after actions
have been applied to the frame with modify the frame in
some way that makes it possible for richer processing to occur.
An example is and indeed targeted use case is MPLS. If an MPLS frame has an
mpls_pop action applied with the IPv4 ethernet type then it becomes
possible to decode the IPv4 portion of the frame. This may be used to
construct a facet that modifies the IPv4 portion of the frame. This is not
possible prior to the mpls_pop action as the contents of the frame after
the MPLS stack is not known to be IPv4.
Status:
I have dropped the RFC prefix from this series as I now believe
it is feature-complete. Any and all review is greatly appreciated.
Design:
* New recirculation action.
ovs-vswitchd adds a recirculation action to the end of a list of
datapath actions for a flow when the actions are truncated because
insufficient flow match information is available to add the next
OpenFlow action. The recirculation action is preceded by an action
to set the skb_mark to an id which can be used to scope a facet lookup
of a recirculated packet.
e.g. pop_mpls(0x0800),dec_ttl becomes pop_mpls(0x800),set(skb_mark(id)),recirculate
* Datapath behaviour
Then the datapath encounters a recirculate action it:
+ Recalculates the flow key based on the packet
which will typically have been modified by previous actions
+ As the recirculate action is preceded by a set(skb_mark(id)) action,
the new match key will now include skb_mark=id.
+ Performs a lookup using the new match key
+ Processes the packet if a facet matches the key or;
+ Makes an upcall if necessary
* No facet behaviour
+ Loop:
1) translate actions
2) If there is a recirculate action, execute packet
and go back to 1) for remaining actions.
Base/Pre-requisites:
This patch depends on "[PATCH v2.24] datapath: Add basic MPLS support to kernel".
There are currently no other patches in the recirculation series.
Availability:
For reference this patch is available in git at:
git://github.com/horms/openvswitch.git devel/mpls-recirculate.v5
Change Log:
v5
* Correct declaration of facet_find_by_id to match definition:
ovs_be32 -> uint32_t.
* Enhancements to recirculation id code:
- Allow efficient lookup of facets by their recirculation id
- Add RECIRCULATION_ID_DUMMY which may be used in cases
where no facet it used. It is an arbitrary valid id.
- Also add recirculated element to action_xlate_ctx()
to use to detect if a recirculation action was added during
translation. The previous scheme of checking if recirculation_id
was not RECIRCULATION_ID_NONE is broken for cases where
the context is initialised with a recirculation_id other than
RECIRCULATION_ID_NONE. E.g. when RECIRCULATION_ID_DUMMY is used.
- Avoid id collision
rfc4:
* Allow recirculation without facets in ovs-vswitchd
- Handle flow miss without facet
- Packet out
* Minor enhancement to recirculation id management: Add RECIRCULATE_ID_NONE
to use instead of using 0 directly.
* Correct calculation of facet->recirculation_ofpacts and
facet->recirculation_ofpacts_len in subfacet_make_actions()
in the case of more than one level of recirculation.
rfc3
* Use IS_ERR_OR_NULL()
* Handle facet consistency checking by constructing a chain of facets
from the given facet, to its recirculation parent and then its parent
until the topmost facet. If there is no recirculation the chain will
be of length one. If there is one recirculation action then the chain
will be of length two. And so on.
The topmost facet in the chain can is used to lookup the rule to be
verified. The chain is then walked from top to bottom, translating
actions up to the end or the first recirculation action that is
encountered, whichever comes first. As the code walks down the chain
it updates the actions that are executed to start of the actions to
be executed to be just after the end of the actions executed in the
previous facet in the chain. This is similar to the way that facets
are created when a recirculation action is encountered.
rfc2
* As suggested by Jesse Gross
- Update for changes to ovs_dp_process_received_packet()
to no longer check if OVS_CB(skb)->flow is pre-initialised.
- Do not add spurious printk debugging to ovs_execute_actions()
- Do not add spurious debugging messages to commit_set_nw_action()
- Correct typo in comment above commit_odp_actions().
- Do not execute recirculation in ovs-vswitchd, rather allow
the datapath to make an upcall when a recirculation action
is encountered on execute.
+ This implicitly breaks support for recirculation without facets,
so for now force all misses of MPLS frames to be handled with
a facet; and treat handling of recirculation for packet_out as
a todo item.
- Use skb_mark for recirculation_id in match. This avoids
both expanding the match and including a recirculation_id parameter
with the recirculation action: set_skb_mark should be used before
the recirculation action.
- Tidy up ownership of skb in ovs_execute_actions
rfc1
* Initial post
Patch List and Diffstat:
Simon Horman (4):
Add packet recirculation
Move execute_set_action to lib/odp-util.c
Allow recirculation without facets
Avoid recirculation id collision
datapath/actions.c | 9 +-
datapath/datapath.c | 98 +++++---
datapath/datapath.h | 2 +-
include/linux/openvswitch.h | 4 +
lib/dpif-netdev.c | 153 +++++--------
lib/flow.h | 3 +
lib/odp-util.c | 91 +++++++-
lib/odp-util.h | 4 +
ofproto/ofproto-dpif.c | 521 +++++++++++++++++++++++++++++++++++++------
9 files changed, 681 insertions(+), 204 deletions(-)
--
1.7.10.4
^ permalink raw reply [flat|nested] 14+ messages in thread
* [PATCH 1/4] Add packet recirculation
[not found] ` <1365403431-18102-1-git-send-email-horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org>
@ 2013-04-08 6:43 ` Simon Horman
2013-04-09 1:46 ` Jesse Gross
2013-04-08 6:43 ` [PATCH 2/4] Move execute_set_action to lib/odp-util.c Simon Horman
2013-04-08 6:43 ` [PATCH 4/4] Avoid recirculation id collision Simon Horman
2 siblings, 1 reply; 14+ messages in thread
From: Simon Horman @ 2013-04-08 6:43 UTC (permalink / raw)
To: dev-yBygre7rU0TnMu66kgdUjQ, netdev-u79uwXL29TY76Z2rM5mHXA
Cc: Isaku Yamahata, Ravi K
Recirculation is a technique to allow a frame to re-enter
frame processing. This is intended to be used after actions
have been applied to the frame with modify the frame in
some way that makes it possible for richer processing to occur.
An example is and indeed targeted use case is MPLS. If an MPLS frame has an
mpls_pop action applied with the IPv4 ethernet type then it becomes
possible to decode the IPv4 portion of the frame. This may be used to
construct a facet that modifies the IPv4 portion of the frame. This is not
possible prior to the mpls_pop action as the contents of the frame after
the MPLS stack is not known to be IPv4.
Design:
* New recirculation action.
ovs-vswitchd adds a recirculation action to the end of a list of
datapath actions for a flow when the actions are truncated because
insufficient flow match information is available to add the next
OpenFlow action. The recirculation action is preceded by an action
to set the skb_mark to an id which can be used to scope a facet lookup
of a recirculated packet.
e.g. pop_mpls(0x0800),dec_ttl becomes pop_mpls(0x800),set(skb_mark(id)),recirculate
* Datapath behaviour
Then the datapath encounters a recirculate action it:
+ Recalculates the flow key based on the packet
which will typically have been modified by previous actions
+ As the recirculate action is preceded by a set(skb_mark(id)) action,
the new match key will now include skb_mark=id.
+ Performs a lookup using the new match key
+ Processes the packet if a facet matches the key or;
+ Makes an upcall if necessary
* No facet behaviour
+ Loop:
1) translate actions
2) If there is a recirculate action, execute packet
and go back to 1) for remaining actions.
Limitations of this patch:
* Facets are required, support for recirculation without facets.
A proposed implementation is provided in a subsequent patch.
* Recirculation ids may conflicts.
A proposed resolution is provided in a subsequent patch.
Signed-off-by: Simon Horman <horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org>
---
This patch depends on "[PATCH v2.24] datapath: Add basic MPLS support to kernel".
Change Log:
v5
* Correct declaration of facet_find_by_id to match definition:
ovs_be32 -> uint32_t.
* Enhancements to recirculation id code:
- Allow efficient lookup of facets by their recirculation id
- Add RECIRCULATION_ID_DUMMY which may be used in cases
where no facet it used. It is an arbitrary valid id.
- Also add recirculated element to action_xlate_ctx()
to use to detect if a recirculation action was added during
translation. The previous scheme of checking if recirculation_id
was not RECIRCULATION_ID_NONE is broken for cases where
the context is initialised with a recirculation_id other than
RECIRCULATION_ID_NONE. E.g. when RECIRCULATION_ID_DUMMY is used.
rfc4
* Minor enhancement to recirculation id management: Add RECIRCULATE_ID_NONE
to use instead of using 0 directly.
* Correct calculation of facet->recirculation_ofpacts and
facet->recirculation_ofpacts_len in subfacet_make_actions()
in the case of more than one level of recirculation.
rfc3
* Use IS_ERR_OR_NULL()
* Handle facet consistency checking by constructing a chain of facets
from the given facet, to its recirculation parent and then its parent
until the topmost facet. If there is no recirculation the chain will
be of length one. If there is one recirculation action then the chain
will be of length two. And so on.
The topmost facet in the chain can is used to lookup the rule to be
verified. The chain is then walked from top to bottom, translating
actions up to the end or the first recirculation action that is
encountered, whichever comes first. As the code walks down the chain
it updates the actions that are executed to start of the actions to
be executed to be just after the end of the actions executed in the
previous facet in the chain. This is similar to the way that facets
are created when a recirculation action is encountered.
rfc2
* As suggested by Jesse Gross
- Update for changes to ovs_dp_process_received_packet()
to no longer check if OVS_CB(skb)->flow is pre-initialised.
- Do not add spurious printk debugging to ovs_execute_actions()
- Do not add spurious debugging messages to commit_set_nw_action()
- Correct typo in comment above commit_odp_actions().
- Do not execute recirculation in ovs-vswitchd, rather allow
the datapath to make an upcall when a recirculation action
is encountered on execute.
+ This implicitly breaks support for recirculation without facets,
so for now force all misses of MPLS frames to be handled with
a facet; and treat handling of recirculation for packet_out as
a todo item.
- Use skb_mark for recirculation_id in match. This avoids
both expanding the match and including a recirculation_id parameter
with the recirculation action: set_skb_mark should be used before
the recirculation action.
- Tidy up ownership of skb in ovs_execute_actions
rfc1
* Initial post
---
datapath/actions.c | 9 +-
datapath/datapath.c | 98 +++++++----
datapath/datapath.h | 2 +-
include/linux/openvswitch.h | 4 +
lib/dpif-netdev.c | 89 +++++++---
lib/flow.h | 3 +
lib/odp-util.c | 15 +-
lib/odp-util.h | 1 +
ofproto/ofproto-dpif.c | 382 ++++++++++++++++++++++++++++++++++++-------
9 files changed, 476 insertions(+), 127 deletions(-)
diff --git a/datapath/actions.c b/datapath/actions.c
index e9634fe..7b0f022 100644
--- a/datapath/actions.c
+++ b/datapath/actions.c
@@ -617,6 +617,9 @@ static int do_execute_actions(struct datapath *dp, struct sk_buff *skb,
case OVS_ACTION_ATTR_SAMPLE:
err = sample(dp, skb, a);
break;
+
+ case OVS_ACTION_ATTR_RECIRCULATE:
+ return 1;
}
if (unlikely(err)) {
@@ -657,7 +660,7 @@ static int loop_suppress(struct datapath *dp, struct sw_flow_actions *actions)
}
/* Execute a list of actions against 'skb'. */
-int ovs_execute_actions(struct datapath *dp, struct sk_buff *skb)
+struct sk_buff *ovs_execute_actions(struct datapath *dp, struct sk_buff *skb)
{
struct sw_flow_actions *acts = rcu_dereference(OVS_CB(skb)->flow->sf_acts);
struct loop_counter *loop;
@@ -676,6 +679,8 @@ int ovs_execute_actions(struct datapath *dp, struct sk_buff *skb)
OVS_CB(skb)->tun_key = NULL;
error = do_execute_actions(dp, skb, acts->actions,
acts->actions_len, false);
+ if (likely(error <= 0))
+ skb = NULL;
/* Check whether sub-actions looped too much. */
if (unlikely(loop->looping))
@@ -686,5 +691,5 @@ out_loop:
if (!--loop->count)
loop->looping = false;
- return error;
+ return (error < 0) ? ERR_PTR(error) : skb;
}
diff --git a/datapath/datapath.c b/datapath/datapath.c
index e8be795..ab39dd7 100644
--- a/datapath/datapath.c
+++ b/datapath/datapath.c
@@ -202,52 +202,63 @@ void ovs_dp_detach_port(struct vport *p)
ovs_vport_del(p);
}
+#define MAX_RECIRCULATION_DEPTH 4 /* Completely arbitrary */
+
/* Must be called with rcu_read_lock. */
void ovs_dp_process_received_packet(struct vport *p, struct sk_buff *skb)
{
struct datapath *dp = p->dp;
- struct sw_flow *flow;
struct dp_stats_percpu *stats;
- struct sw_flow_key key;
- u64 *stats_counter;
- int error;
- int key_len;
+ int limit = MAX_RECIRCULATION_DEPTH;
stats = this_cpu_ptr(dp->stats_percpu);
- /* Extract flow from 'skb' into 'key'. */
- error = ovs_flow_extract(skb, p->port_no, &key, &key_len);
- if (unlikely(error)) {
- kfree_skb(skb);
- return;
- }
+ while (1) {
+ u64 *stats_counter;
+ struct sw_flow *flow;
+ struct sw_flow_key key;
+ int error, key_len;
- /* Look up flow. */
- flow = ovs_flow_tbl_lookup(rcu_dereference(dp->table), &key, key_len);
- if (unlikely(!flow)) {
- struct dp_upcall_info upcall;
-
- upcall.cmd = OVS_PACKET_CMD_MISS;
- upcall.key = &key;
- upcall.userdata = NULL;
- upcall.portid = p->upcall_portid;
- ovs_dp_upcall(dp, skb, &upcall);
- consume_skb(skb);
- stats_counter = &stats->n_missed;
- goto out;
- }
+ /* Extract flow from 'skb' into 'key'. */
+ error = ovs_flow_extract(skb, p->port_no, &key, &key_len);
+ if (unlikely(error)) {
+ kfree_skb(skb);
+ return;
+ }
- OVS_CB(skb)->flow = flow;
+ /* Look up flow. */
+ flow = ovs_flow_tbl_lookup(rcu_dereference(dp->table),
+ &key, key_len);
+ if (unlikely(!flow)) {
+ struct dp_upcall_info upcall;
+
+ upcall.cmd = OVS_PACKET_CMD_MISS;
+ upcall.key = &key;
+ upcall.userdata = NULL;
+ upcall.portid = p->upcall_portid;
+ ovs_dp_upcall(dp, skb, &upcall);
+ consume_skb(skb);
+ stats_counter = &stats->n_missed;
+ skb = NULL;
+ } else {
+ OVS_CB(skb)->flow = flow;
+ stats_counter = &stats->n_hit;
+ ovs_flow_used(flow, skb);
+ skb = ovs_execute_actions(dp, skb);
+ }
- stats_counter = &stats->n_hit;
- ovs_flow_used(OVS_CB(skb)->flow, skb);
- ovs_execute_actions(dp, skb);
+ /* Update datapath statistics. */
+ u64_stats_update_begin(&stats->sync);
+ (*stats_counter)++;
+ u64_stats_update_end(&stats->sync);
-out:
- /* Update datapath statistics. */
- u64_stats_update_begin(&stats->sync);
- (*stats_counter)++;
- u64_stats_update_end(&stats->sync);
+ if (IS_ERR_OR_NULL(skb)) {
+ break;
+ } else if (unlikely(!limit--)) {
+ kfree_skb(skb);
+ return;
+ }
+ }
}
static struct genl_family dp_packet_genl_family = {
@@ -818,6 +829,7 @@ static int validate_and_copy_actions__(const struct nlattr *attr,
[OVS_ACTION_ATTR_POP_MPLS] = sizeof(__be16),
[OVS_ACTION_ATTR_PUSH_VLAN] = sizeof(struct ovs_action_push_vlan),
[OVS_ACTION_ATTR_POP_VLAN] = 0,
+ [OVS_ACTION_ATTR_RECIRCULATE] = 0,
[OVS_ACTION_ATTR_SET] = (u32)-1,
[OVS_ACTION_ATTR_SAMPLE] = (u32)-1
};
@@ -901,6 +913,9 @@ static int validate_and_copy_actions__(const struct nlattr *attr,
skip_copy = true;
break;
+ case OVS_ACTION_ATTR_RECIRCULATE:
+ break;
+
default:
return -EINVAL;
}
@@ -1005,12 +1020,23 @@ static int ovs_packet_cmd_execute(struct sk_buff *skb, struct genl_info *info)
goto err_unlock;
local_bh_disable();
- err = ovs_execute_actions(dp, packet);
+ packet = ovs_execute_actions(dp, packet);
+ if (!IS_ERR_OR_NULL(packet)) {
+ struct vport *vport;
+ vport = ovs_lookup_vport(dp, flow->key.phy.in_port);
+ if (!vport) {
+ err = -ENODEV;
+ goto err_unlock;
+ }
+ /* Recirculate */
+ ovs_dp_process_received_packet(vport, packet);
+ packet = NULL;
+ }
local_bh_enable();
rcu_read_unlock();
ovs_flow_free(flow);
- return err;
+ return PTR_ERR(packet);
err_unlock:
rcu_read_unlock();
diff --git a/datapath/datapath.h b/datapath/datapath.h
index 7665742..8da5e8a 100644
--- a/datapath/datapath.h
+++ b/datapath/datapath.h
@@ -188,7 +188,7 @@ const char *ovs_dp_name(const struct datapath *dp);
struct sk_buff *ovs_vport_cmd_build_info(struct vport *, u32 portid, u32 seq,
u8 cmd);
-int ovs_execute_actions(struct datapath *dp, struct sk_buff *skb);
+struct sk_buff *ovs_execute_actions(struct datapath *dp, struct sk_buff *skb);
unsigned char *skb_cb_mpls_stack(const struct sk_buff *skb);
#endif /* datapath.h */
diff --git a/include/linux/openvswitch.h b/include/linux/openvswitch.h
index e890fd8..0fff7cc 100644
--- a/include/linux/openvswitch.h
+++ b/include/linux/openvswitch.h
@@ -516,6 +516,9 @@ struct ovs_action_push_vlan {
* indicate the new packet contents This could potentially still be
* %ETH_P_MPLS_* if the resulting MPLS label stack is not empty. If there
* is no MPLS label stack, as determined by ethertype, no action is taken.
+ * @OVS_ACTION_ATTR_RECIRCULATE: Restart processing of packet.
+ * The packet must have been modified by a previous action in such a way
+ * that it does not match its original flow again.
*
* Only a single header can be set with a single %OVS_ACTION_ATTR_SET. Not all
* fields within a header are modifiable, e.g. the IPv4 protocol and fragment
@@ -532,6 +535,7 @@ enum ovs_action_attr {
OVS_ACTION_ATTR_SAMPLE, /* Nested OVS_SAMPLE_ATTR_*. */
OVS_ACTION_ATTR_PUSH_MPLS, /* struct ovs_action_push_mpls. */
OVS_ACTION_ATTR_POP_MPLS, /* __be16 ethertype. */
+ OVS_ACTION_ATTR_RECIRCULATE, /* No argument */
__OVS_ACTION_ATTR_MAX
};
diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
index e4a2f75..31255f6 100644
--- a/lib/dpif-netdev.c
+++ b/lib/dpif-netdev.c
@@ -152,10 +152,14 @@ static int dpif_netdev_open(const struct dpif_class *, const char *name,
static int dp_netdev_output_userspace(struct dp_netdev *, const struct ofpbuf *,
int queue_no, const struct flow *,
const struct nlattr *userdata);
-static void dp_netdev_execute_actions(struct dp_netdev *,
+static bool dp_netdev_execute_actions(struct dp_netdev *,
struct ofpbuf *, struct flow *,
const struct nlattr *actions,
- size_t actions_len);
+ size_t actions_len,
+ uint32_t *skb_mark);
+static void dp_netdev_port_input(struct dp_netdev *dp,
+ struct dp_netdev_port *port,
+ struct ofpbuf *packet);
static struct dpif_netdev *
dpif_netdev_cast(const struct dpif *dpif)
@@ -940,8 +944,22 @@ dpif_netdev_execute(struct dpif *dpif, const struct dpif_execute *execute)
error = dpif_netdev_flow_from_nlattrs(execute->key, execute->key_len,
&key);
if (!error) {
- dp_netdev_execute_actions(dp, ©, &key,
- execute->actions, execute->actions_len);
+ bool recirculate;
+ uint32_t skb_mark = 0;
+
+ recirculate = dp_netdev_execute_actions(dp, ©, &key,
+ execute->actions,
+ execute->actions_len,
+ &skb_mark);
+ if (recirculate) {
+ struct dp_netdev_port *port;
+ port = (key.in_port < MAX_PORTS) ? dp->ports[key.in_port] : NULL;
+ if (port) {
+ dp_netdev_port_input(dp, port, ©);
+ return 0;
+ }
+ error = ENOENT;
+ }
}
ofpbuf_uninit(©);
@@ -1028,23 +1046,32 @@ static void
dp_netdev_port_input(struct dp_netdev *dp, struct dp_netdev_port *port,
struct ofpbuf *packet)
{
- struct dp_netdev_flow *flow;
- struct flow key;
+ bool recirculate;
+ uint32_t skb_mark = 0;
+ int limit = MAX_RECIRCULATION_DEPTH;
- if (packet->size < ETH_HEADER_LEN) {
- return;
- }
- flow_extract(packet, 0, 0, NULL, port->port_no, &key);
- flow = dp_netdev_lookup_flow(dp, &key);
- if (flow) {
- dp_netdev_flow_used(flow, packet);
- dp_netdev_execute_actions(dp, packet, &key,
- flow->actions, flow->actions_len);
- dp->n_hit++;
- } else {
- dp->n_missed++;
- dp_netdev_output_userspace(dp, packet, DPIF_UC_MISS, &key, NULL);
- }
+ do {
+ struct dp_netdev_flow *flow;
+ struct flow key;
+
+ if (packet->size < ETH_HEADER_LEN) {
+ return;
+ }
+ flow_extract(packet, 0, skb_mark, NULL, port->port_no, &key);
+ flow = dp_netdev_lookup_flow(dp, &key);
+ if (flow) {
+ dp_netdev_flow_used(flow, packet);
+ recirculate = dp_netdev_execute_actions(dp, packet, &key,
+ flow->actions,
+ flow->actions_len,
+ &skb_mark);
+ dp->n_hit++;
+ } else {
+ dp->n_missed++;
+ dp_netdev_output_userspace(dp, packet, DPIF_UC_MISS, &key, NULL);
+ recirculate = false;
+ }
+ } while (recirculate && limit--);
}
static void
@@ -1163,6 +1190,7 @@ dp_netdev_sample(struct dp_netdev *dp,
const struct nlattr *subactions = NULL;
const struct nlattr *a;
size_t left;
+ uint32_t skb_mark;
NL_NESTED_FOR_EACH_UNSAFE (a, left, action) {
int type = nl_attr_type(a);
@@ -1186,7 +1214,7 @@ dp_netdev_sample(struct dp_netdev *dp,
}
dp_netdev_execute_actions(dp, packet, key, nl_attr_get(subactions),
- nl_attr_get_size(subactions));
+ nl_attr_get_size(subactions), &skb_mark);
}
static void
@@ -1201,7 +1229,8 @@ dp_netdev_action_userspace(struct dp_netdev *dp,
}
static void
-execute_set_action(struct ofpbuf *packet, const struct nlattr *a)
+execute_set_action(struct ofpbuf *packet, const struct nlattr *a,
+ uint32_t *skb_mark)
{
enum ovs_key_attr type = nl_attr_type(a);
const struct ovs_key_ipv4 *ipv4_key;
@@ -1211,11 +1240,14 @@ execute_set_action(struct ofpbuf *packet, const struct nlattr *a)
switch (type) {
case OVS_KEY_ATTR_PRIORITY:
- case OVS_KEY_ATTR_SKB_MARK:
case OVS_KEY_ATTR_TUNNEL:
/* not implemented */
break;
+ case OVS_KEY_ATTR_SKB_MARK:
+ *skb_mark = nl_attr_get_u32(a);
+ break;
+
case OVS_KEY_ATTR_ETHERNET:
dp_netdev_set_dl(packet,
nl_attr_get_unspec(a, sizeof(struct ovs_key_ethernet)));
@@ -1263,11 +1295,11 @@ execute_set_action(struct ofpbuf *packet, const struct nlattr *a)
}
}
-static void
+static bool
dp_netdev_execute_actions(struct dp_netdev *dp,
struct ofpbuf *packet, struct flow *key,
const struct nlattr *actions,
- size_t actions_len)
+ size_t actions_len, uint32_t *skb_mark)
{
const struct nlattr *a;
unsigned int left;
@@ -1305,18 +1337,23 @@ dp_netdev_execute_actions(struct dp_netdev *dp,
break;
case OVS_ACTION_ATTR_SET:
- execute_set_action(packet, nl_attr_get(a));
+ execute_set_action(packet, nl_attr_get(a), skb_mark);
break;
case OVS_ACTION_ATTR_SAMPLE:
dp_netdev_sample(dp, packet, key, a);
break;
+ case OVS_ACTION_ATTR_RECIRCULATE:
+ return true;
+
case OVS_ACTION_ATTR_UNSPEC:
case __OVS_ACTION_ATTR_MAX:
NOT_REACHED();
}
}
+
+ return false;
}
const struct dpif_class dpif_netdev_class = {
diff --git a/lib/flow.h b/lib/flow.h
index 6e169d6..66f89e3 100644
--- a/lib/flow.h
+++ b/lib/flow.h
@@ -296,4 +296,7 @@ uint32_t minimask_hash(const struct minimask *, uint32_t basis);
bool minimask_has_extra(const struct minimask *, const struct minimask *);
bool minimask_is_catchall(const struct minimask *);
+#define MAX_RECIRCULATION_DEPTH 4 /* Completely arbitrary value to
+ * guard against infinite loops */
+
#endif /* flow.h */
diff --git a/lib/odp-util.c b/lib/odp-util.c
index 3206dc9..e18e109 100644
--- a/lib/odp-util.c
+++ b/lib/odp-util.c
@@ -75,6 +75,7 @@ odp_action_len(uint16_t type)
case OVS_ACTION_ATTR_POP_VLAN: return 0;
case OVS_ACTION_ATTR_PUSH_MPLS: return sizeof(struct ovs_action_push_mpls);
case OVS_ACTION_ATTR_POP_MPLS: return sizeof(ovs_be16);
+ case OVS_ACTION_ATTR_RECIRCULATE: return 0;
case OVS_ACTION_ATTR_SET: return -2;
case OVS_ACTION_ATTR_SAMPLE: return -2;
@@ -376,6 +377,10 @@ format_odp_action(struct ds *ds, const struct nlattr *a)
ds_put_format(ds, "pop_mpls(eth_type=0x%"PRIx16")", ntohs(ethertype));
break;
}
+ case OVS_ACTION_ATTR_RECIRCULATE: {
+ ds_put_format(ds, "recirculate");
+ break;
+ }
case OVS_ACTION_ATTR_SAMPLE:
format_odp_sample_action(ds, a);
break;
@@ -2172,6 +2177,12 @@ commit_odp_tunnel_action(const struct flow *flow, struct flow *base,
}
}
+void
+commit_odp_recirculate_action(struct ofpbuf *odp_actions)
+{
+ nl_msg_put_flag(odp_actions, OVS_ACTION_ATTR_RECIRCULATE);
+}
+
static void
commit_set_ether_addr_action(const struct flow *flow, struct flow *base,
struct ofpbuf *odp_actions)
@@ -2385,14 +2396,14 @@ commit_set_skb_mark_action(const struct flow *flow, struct flow *base,
return;
}
base->skb_mark = flow->skb_mark;
-
odp_put_skb_mark_action(base->skb_mark, odp_actions);
}
/* If any of the flow key data that ODP actions can modify are different in
* 'base' and 'flow', appends ODP actions to 'odp_actions' that change the flow
* key from 'base' into 'flow', and then changes 'base' the same way. Does not
* commit set_tunnel actions. Users should call commit_odp_tunnel_action()
- * in addition to this function if needed. */
+ * and commit_odp_recirculate_action() in addition to those functions are
+ * needed. */
void
commit_odp_actions(const struct flow *flow, struct flow *base,
struct ofpbuf *odp_actions)
diff --git a/lib/odp-util.h b/lib/odp-util.h
index ad0fb30..da62aa5 100644
--- a/lib/odp-util.h
+++ b/lib/odp-util.h
@@ -115,6 +115,7 @@ const char *odp_key_fitness_to_string(enum odp_key_fitness);
void commit_odp_tunnel_action(const struct flow *, struct flow *base,
struct ofpbuf *odp_actions);
+void commit_odp_recirculate_action(struct ofpbuf *odp_actions);
void commit_odp_actions(const struct flow *, struct flow *base,
struct ofpbuf *odp_actions);
\f
diff --git a/ofproto/ofproto-dpif.c b/ofproto/ofproto-dpif.c
index 47830c1..5129da1 100644
--- a/ofproto/ofproto-dpif.c
+++ b/ofproto/ofproto-dpif.c
@@ -119,7 +119,8 @@ static struct rule_dpif *rule_dpif_miss_rule(struct ofproto_dpif *ofproto,
static void rule_credit_stats(struct rule_dpif *,
const struct dpif_flow_stats *);
-static void flow_push_stats(struct facet *, const struct dpif_flow_stats *);
+static void flow_push_stats(struct facet *, const struct dpif_flow_stats *,
+ const struct ofpact *, size_t ofpacts_len);
static tag_type rule_calculate_tag(const struct flow *,
const struct minimask *, uint32_t basis);
static void rule_invalidate(const struct rule_dpif *);
@@ -276,6 +277,17 @@ struct action_xlate_ctx {
uint16_t nf_output_iface; /* Output interface index for NetFlow. */
mirror_mask_t mirrors; /* Bitmap of associated mirrors. */
+ size_t ofpacts_len; /* The number of bytes of the ofpacts
+ * argument to xlate_actions() processed
+ * by it. This is used to calculate an
+ * offset into ofpacts for calls to
+ * xlate_actions on recirculated packets */
+
+ uint32_t recirculation_id; /* skb_mark to use to identify
+ * recirculation. */
+ bool recircualted; /* True if the context does not add a
+ * recirculate action. False otherwise. */
+
/* xlate_actions() initializes and uses these members, but the client has no
* reason to look at them. */
@@ -312,7 +324,8 @@ static void action_xlate_ctx_init(struct action_xlate_ctx *,
struct ofproto_dpif *, const struct flow *,
const struct initial_vals *initial_vals,
struct rule_dpif *,
- uint8_t tcp_flags, const struct ofpbuf *);
+ uint8_t tcp_flags, const struct ofpbuf *,
+ uint32_t recirculation_id);
static void xlate_actions(struct action_xlate_ctx *,
const struct ofpact *ofpacts, size_t ofpacts_len,
struct ofpbuf *odp_actions);
@@ -494,13 +507,40 @@ struct facet {
struct subfacet one_subfacet;
long long int learn_rl; /* Rate limiter for facet_learn(). */
+
+ const struct ofpact *ofpacts; /* ofpacts for this facet.
+ * Will differ from rule->up.ofpacts
+ * if facet is for a recirculated packet. */
+ size_t ofpacts_len; /* ofpacts_len for this facet
+ * Will differ from * rule->up.ofpacts_len
+ * if facet is for a recirculated packet. */
+
+ uint32_t recirculation_id; /* Recirculation id.
+ * Non-sero for a facet
+ * that recirculates packets;
+ * used as the value of flow.skb_mark
+ * in the facet of recirculated packets.
+ * Zero otherwise. */
+ struct hmap_node recirculation_id_hmap_node;
+ /* In owning ofproto's 'recirculation_id'
+ * hmap. */
+ const struct ofpact *recirculation_ofpacts;
+ /* ofpacts for facets of packets
+ * recirculated by this facet */
+ size_t recirculation_ofpacts_len;
+ /* ofpacts_len for facets of packets
+ * recirculated by this facet */
+
+ bool recirculated; /* Facet of a recirculated packet? */
};
-static struct facet *facet_create(struct rule_dpif *,
- const struct flow *, uint32_t hash);
+static struct facet *facet_create(struct rule_dpif *, const struct flow *,
+ const struct ofpact *, size_t ofpacts_len,
+ bool recirculated, uint32_t hash);
static void facet_remove(struct facet *);
static void facet_free(struct facet *);
+static struct facet *facet_find_by_id(struct ofproto_dpif *, uint32_t id);
static struct facet *facet_find(struct ofproto_dpif *,
const struct flow *, uint32_t hash);
static struct facet *facet_lookup_valid(struct ofproto_dpif *,
@@ -703,6 +743,7 @@ struct ofproto_dpif {
/* Facets. */
struct hmap facets;
+ struct hmap recirculation_ids;
struct hmap subfacets;
struct governor *governor;
long long int consistency_rl;
@@ -1358,6 +1399,7 @@ construct(struct ofproto *ofproto_)
ofproto->has_bonded_bundles = false;
hmap_init(&ofproto->facets);
+ hmap_init(&ofproto->recirculation_ids);
hmap_init(&ofproto->subfacets);
ofproto->governor = NULL;
ofproto->consistency_rl = LLONG_MIN;
@@ -3408,6 +3450,31 @@ port_is_lacp_current(const struct ofport *ofport_)
: -1);
}
\f
+/* Recirculation Id */
+#define RECIRCULATION_ID_NONE 0
+#define RECIRCULATION_ID_DUMMY 2
+#define RECIRCULATION_ID_MIN RECIRCULATION_ID_DUMMY
+
+static uint32_t recirculation_id_hash(uint32_t id)
+{
+ return hash_words(&id, 1, 0);
+}
+
+/* XXX: This does not prevent id collision */
+static uint32_t get_recirculation_id(void)
+{
+ static uint32_t id = RECIRCULATION_ID_MIN;
+
+ if (id < RECIRCULATION_ID_MIN)
+ id = RECIRCULATION_ID_MIN;
+ /* Skip IPSEC_MARK bit it is reserved */
+ if (id & IPSEC_MARK) {
+ id++;
+ ovs_assert(!(id & IPSEC_MARK));
+ }
+ return id++;
+}
+\f
/* Upcall handling. */
/* Flow miss batching.
@@ -3565,6 +3632,15 @@ static bool
flow_miss_should_make_facet(struct ofproto_dpif *ofproto,
struct flow_miss *miss, uint32_t hash)
{
+ /* A facet is currently required to handle recirculation.
+ * There currently isn't a good way to detect if recirculation will
+ * occur or not. So in the mean time assume that it can't occur
+ * for non-MPLS packets and it may occur for MPLS packets
+ */
+ if (eth_type_mpls(miss->flow.dl_type)) {
+ return true;
+ }
+
if (!ofproto->governor) {
size_t n_subfacets;
@@ -3584,8 +3660,8 @@ flow_miss_should_make_facet(struct ofproto_dpif *ofproto,
* or creating any datapath flow. May add an "execute" operation to 'ops' and
* increment '*n_ops'. */
static void
-handle_flow_miss_without_facet(struct flow_miss *miss,
- struct rule_dpif *rule,
+handle_flow_miss_without_facet(struct flow_miss *miss, struct rule_dpif *rule,
+ const struct ofpact *ofpacts, size_t ofpacts_len,
struct flow_miss_op *ops, size_t *n_ops)
{
struct ofproto_dpif *ofproto = ofproto_dpif_cast(rule->up.ofproto);
@@ -3606,10 +3682,10 @@ handle_flow_miss_without_facet(struct flow_miss *miss,
rule_credit_stats(rule, &stats);
action_xlate_ctx_init(&ctx, ofproto, &miss->flow,
- &miss->initial_vals, rule, 0, packet);
+ &miss->initial_vals, rule, 0, packet,
+ RECIRCULATION_ID_DUMMY);
ctx.resubmit_stats = &stats;
- xlate_actions(&ctx, rule->up.ofpacts, rule->up.ofpacts_len,
- &odp_actions);
+ xlate_actions(&ctx, ofpacts, ofpacts_len, &odp_actions);
if (odp_actions.size) {
struct dpif_execute *execute = &op->dpif_op.u.execute;
@@ -3723,14 +3799,30 @@ handle_flow_miss(struct flow_miss *miss, struct flow_miss_op *ops,
facet = facet_lookup_valid(ofproto, &miss->flow, hash);
if (!facet) {
- struct rule_dpif *rule = rule_dpif_lookup(ofproto, &miss->flow);
+ struct rule_dpif *rule;
+ const struct ofpact *ofpacts;
+ size_t ofpacts_len;
+ struct facet *parent_facet;
+
+ parent_facet = facet_find_by_id(ofproto, miss->flow.skb_mark);
+ if (parent_facet) {
+ rule = parent_facet->rule;
+ ofpacts = parent_facet->recirculation_ofpacts;
+ ofpacts_len = parent_facet->recirculation_ofpacts_len;
+ } else {
+ rule = rule_dpif_lookup(ofproto, &miss->flow);
+ ofpacts = rule->up.ofpacts;
+ ofpacts_len = rule->up.ofpacts_len;
+ }
if (!flow_miss_should_make_facet(ofproto, miss, hash)) {
- handle_flow_miss_without_facet(miss, rule, ops, n_ops);
+ handle_flow_miss_without_facet(miss, rule, ofpacts,
+ ofpacts_len, ops, n_ops);
return;
}
- facet = facet_create(rule, &miss->flow, hash);
+ facet = facet_create(rule, &miss->flow, ofpacts, ofpacts_len,
+ parent_facet != NULL, hash);
now = facet->used;
} else {
now = time_msec();
@@ -4494,7 +4586,9 @@ rule_expire(struct rule_dpif *rule)
* The facet will initially have no subfacets. The caller should create (at
* least) one subfacet with subfacet_create(). */
static struct facet *
-facet_create(struct rule_dpif *rule, const struct flow *flow, uint32_t hash)
+facet_create(struct rule_dpif *rule, const struct flow *flow,
+ const struct ofpact *ofpacts, size_t ofpacts_len,
+ bool recirculated, uint32_t hash)
{
struct ofproto_dpif *ofproto = ofproto_dpif_cast(rule->up.ofproto);
struct facet *facet;
@@ -4502,9 +4596,13 @@ facet_create(struct rule_dpif *rule, const struct flow *flow, uint32_t hash)
facet = xzalloc(sizeof *facet);
facet->used = time_msec();
hmap_insert(&ofproto->facets, &facet->hmap_node, hash);
+ hmap_node_nullify(&facet->recirculation_id_hmap_node);
list_push_back(&rule->facets, &facet->list_node);
facet->rule = rule;
facet->flow = *flow;
+ facet->ofpacts = ofpacts;
+ facet->ofpacts_len = ofpacts_len;
+ facet->recirculated = recirculated;
list_init(&facet->subfacets);
netflow_flow_init(&facet->nf_flow);
netflow_flow_update_time(ofproto->netflow, &facet->nf_flow, facet->used);
@@ -4574,6 +4672,10 @@ facet_remove(struct facet *facet)
}
hmap_remove(&ofproto->facets, &facet->hmap_node);
list_remove(&facet->list_node);
+ if (!hmap_node_is_null(&facet->recirculation_id_hmap_node)) {
+ hmap_remove(&ofproto->recirculation_ids,
+ &facet->recirculation_id_hmap_node);
+ }
facet_free(facet);
}
@@ -4603,10 +4705,10 @@ facet_learn(struct facet *facet)
action_xlate_ctx_init(&ctx, ofproto, &facet->flow,
&subfacet->initial_vals,
- facet->rule, facet->tcp_flags, NULL);
+ facet->rule, facet->tcp_flags, NULL,
+ facet->recirculation_id);
ctx.may_learn = true;
- xlate_actions_for_side_effects(&ctx, facet->rule->up.ofpacts,
- facet->rule->up.ofpacts_len);
+ xlate_actions_for_side_effects(&ctx, facet->ofpacts, facet->ofpacts_len);
}
static void
@@ -4742,6 +4844,36 @@ facet_find(struct ofproto_dpif *ofproto,
return NULL;
}
+/* Searches 'ofproto''s table of facets with recircualtion ids
+ * for a facet whose recicualtion_id is 'id'.
+ * Returns it if found, otherwise a null pointer.
+ *
+ * The returned facet might need revalidation; use facet_lookup_valid()
+ * instead if that is important. */
+static struct facet *
+facet_find_by_id(struct ofproto_dpif *ofproto, uint32_t id)
+{
+ uint32_t hash = recirculation_id_hash(id);
+ struct facet *facet;
+
+ /* some values are never used */
+ if (id == RECIRCULATION_ID_NONE || (id & IPSEC_MARK)) {
+ return NULL;
+ }
+
+ /* This is a ridiculous way to look things up, most likely the id
+ * should be cooked somehow to allow a more efficient lookup.
+ */
+ HMAP_FOR_EACH_WITH_HASH (facet, recirculation_id_hmap_node,
+ hash, &ofproto->recirculation_ids) {
+ if (facet->recirculation_id == id) {
+ return facet;
+ }
+ }
+
+ return NULL;
+}
+
/* Searches 'ofproto''s table of facets for one exactly equal to 'flow'.
* Returns it if found, otherwise a null pointer.
*
@@ -4814,8 +4946,10 @@ subfacet_should_install(struct subfacet *subfacet, enum slow_path_reason slow,
subfacet->actions_len))));
}
-static bool
-facet_check_consistency(struct facet *facet)
+static size_t
+facet_check_actions_consistency(struct facet *facet, struct rule_dpif *rule,
+ const struct ofpact *ofpacts,
+ size_t ofpacts_len)
{
static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 15);
@@ -4824,33 +4958,10 @@ facet_check_consistency(struct facet *facet)
uint64_t odp_actions_stub[1024 / 8];
struct ofpbuf odp_actions;
- struct rule_dpif *rule;
struct subfacet *subfacet;
bool may_log = false;
- bool ok;
-
- /* Check the rule for consistency. */
- rule = rule_dpif_lookup(ofproto, &facet->flow);
- ok = rule == facet->rule;
- if (!ok) {
- may_log = !VLOG_DROP_WARN(&rl);
- if (may_log) {
- struct ds s;
-
- ds_init(&s);
- flow_format(&s, &facet->flow);
- ds_put_format(&s, ": facet associated with wrong rule (was "
- "table=%"PRIu8",", facet->rule->up.table_id);
- cls_rule_format(&facet->rule->up.cr, &s);
- ds_put_format(&s, ") (should have been table=%"PRIu8",",
- rule->up.table_id);
- cls_rule_format(&rule->up.cr, &s);
- ds_put_char(&s, ')');
-
- VLOG_WARN("%s", ds_cstr(&s));
- ds_destroy(&s);
- }
- }
+ bool ok = true;
+ size_t ofpacts_consumed_len = 0;
/* Check the datapath actions for consistency. */
ofpbuf_use_stub(&odp_actions, odp_actions_stub, sizeof odp_actions_stub);
@@ -4860,9 +4971,10 @@ facet_check_consistency(struct facet *facet)
struct ds s;
action_xlate_ctx_init(&ctx, ofproto, &facet->flow,
- &subfacet->initial_vals, rule, 0, NULL);
- xlate_actions(&ctx, rule->up.ofpacts, rule->up.ofpacts_len,
- &odp_actions);
+ &subfacet->initial_vals, rule, 0, NULL,
+ facet->recirculation_id);
+ xlate_actions(&ctx, ofpacts, ofpacts_len, &odp_actions);
+ ofpacts_consumed_len = ctx.ofpacts_len;
if (subfacet->path == SF_NOT_INSTALLED) {
/* This only happens if the datapath reported an error when we
@@ -4921,6 +5033,84 @@ facet_check_consistency(struct facet *facet)
}
ofpbuf_uninit(&odp_actions);
+ return ok ? ofpacts_consumed_len : 0;
+}
+
+static bool
+facet_check_consistency(struct facet *facet)
+{
+ static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 15);
+
+ struct ofproto_dpif *ofproto = ofproto_dpif_cast(facet->rule->up.ofproto);
+
+ const struct ofpact *ofpacts;
+ size_t ofpacts_len;
+
+ struct rule_dpif *rule;
+ struct facet *chain[MAX_RECIRCULATION_DEPTH + 1];
+ int top;
+ bool may_log = false;
+ bool ok;
+
+ top = 0;
+ chain[0] = facet;
+
+ while (chain[top]->recirculated && top <= MAX_RECIRCULATION_DEPTH) {
+ chain[top + 1] = facet_find_by_id(ofproto, chain[top]->flow.skb_mark);
+ if (!chain[top + 1]) {
+ may_log = !VLOG_DROP_WARN(&rl);
+ if (may_log) {
+ struct ds s;
+
+ ds_init(&s);
+ flow_format(&s, &chain[top]->flow);
+ ds_put_format(&s, ": parent facet of facet for "
+ "recirculated packets could not be found");
+
+ VLOG_WARN("%s", ds_cstr(&s));
+ ds_destroy(&s);
+ }
+ break;
+ }
+ top++;
+ }
+
+ rule = rule_dpif_lookup(ofproto, &chain[top]->flow);
+ ok = rule == chain[top]->rule;
+ if (!ok) {
+ may_log = !VLOG_DROP_WARN(&rl);
+ if (may_log) {
+ struct ds s;
+
+ ds_init(&s);
+ flow_format(&s, &chain[top]->flow);
+ ds_put_format(&s, ": facet associated with wrong rule (was "
+ "table=%"PRIu8",", chain[top]->rule->up.table_id);
+ cls_rule_format(&chain[top]->rule->up.cr, &s);
+ ds_put_format(&s, ") (should have been table=%"PRIu8",",
+ rule->up.table_id);
+ cls_rule_format(&rule->up.cr, &s);
+ ds_put_char(&s, ')');
+
+ VLOG_WARN("%s", ds_cstr(&s));
+ ds_destroy(&s);
+ }
+ }
+
+ ofpacts = rule->up.ofpacts;
+ ofpacts_len = rule->up.ofpacts_len;
+ do {
+ size_t consumed;
+ consumed = facet_check_actions_consistency(chain[top], rule,
+ ofpacts, ofpacts_len);
+ if (!consumed) {
+ ok = false;
+ break;
+ }
+ ofpacts = ofpact_end(ofpacts, consumed);
+ ofpacts_len -= consumed;
+ } while(top--);
+
return ok;
}
@@ -4991,7 +5181,8 @@ facet_revalidate(struct facet *facet)
enum slow_path_reason slow;
action_xlate_ctx_init(&ctx, ofproto, &facet->flow,
- &subfacet->initial_vals, new_rule, 0, NULL);
+ &subfacet->initial_vals, new_rule, 0, NULL,
+ facet->recirculation_id);
xlate_actions(&ctx, new_rule->up.ofpacts, new_rule->up.ofpacts_len,
&odp_actions);
@@ -5089,11 +5280,13 @@ facet_push_stats(struct facet *facet)
stats.tcp_flags = 0;
if (stats.n_packets || stats.n_bytes || facet->used > facet->prev_used) {
+
facet->prev_packet_count = facet->packet_count;
facet->prev_byte_count = facet->byte_count;
facet->prev_used = facet->used;
- flow_push_stats(facet, &stats);
+ flow_push_stats(facet, &stats,
+ facet->ofpacts, facet->ofpacts_len);
update_mirror_stats(ofproto_dpif_cast(facet->rule->up.ofproto),
facet->mirrors, stats.n_packets, stats.n_bytes);
@@ -5133,7 +5326,8 @@ rule_credit_stats(struct rule_dpif *rule, const struct dpif_flow_stats *stats)
/* Pushes flow statistics to the rules which 'facet->flow' resubmits
* into given 'facet->rule''s actions and mirrors. */
static void
-flow_push_stats(struct facet *facet, const struct dpif_flow_stats *stats)
+flow_push_stats(struct facet *facet, const struct dpif_flow_stats *stats,
+ const struct ofpact *ofpacts, size_t ofpacts_len)
{
struct rule_dpif *rule = facet->rule;
struct ofproto_dpif *ofproto = ofproto_dpif_cast(rule->up.ofproto);
@@ -5143,10 +5337,11 @@ flow_push_stats(struct facet *facet, const struct dpif_flow_stats *stats)
ofproto_rule_update_used(&rule->up, stats->used);
action_xlate_ctx_init(&ctx, ofproto, &facet->flow,
- &subfacet->initial_vals, rule, 0, NULL);
+ &subfacet->initial_vals, rule, 0, NULL,
+ facet->recirculation_id);
ctx.resubmit_stats = stats;
- xlate_actions_for_side_effects(&ctx, rule->up.ofpacts,
- rule->up.ofpacts_len);
+
+ xlate_actions_for_side_effects(&ctx, ofpacts, ofpacts_len);
}
\f
/* Subfacets. */
@@ -5306,8 +5501,19 @@ subfacet_make_actions(struct subfacet *subfacet, const struct ofpbuf *packet,
struct action_xlate_ctx ctx;
action_xlate_ctx_init(&ctx, ofproto, &facet->flow,
- &subfacet->initial_vals, rule, 0, packet);
- xlate_actions(&ctx, rule->up.ofpacts, rule->up.ofpacts_len, odp_actions);
+ &subfacet->initial_vals, rule, 0, packet,
+ facet->recirculation_id);
+ xlate_actions(&ctx, facet->ofpacts, facet->ofpacts_len, odp_actions);
+ if (ctx.recircualted) {
+ facet->recirculation_id = ctx.recirculation_id;
+ facet->recirculation_ofpacts = ofpact_end(facet->ofpacts,
+ ctx.ofpacts_len);
+ facet->recirculation_ofpacts_len =
+ facet->ofpacts_len - ctx.ofpacts_len;
+ hmap_insert(&ofproto->recirculation_ids,
+ &facet->recirculation_id_hmap_node,
+ recirculation_id_hash(facet->recirculation_id));
+ }
facet->tags = ctx.tags;
facet->has_learn = ctx.has_learn;
facet->has_normal = ctx.has_normal;
@@ -5638,7 +5844,8 @@ rule_dpif_execute(struct rule_dpif *rule, const struct flow *flow,
initial_vals.tunnel_ip_tos = flow->tunnel.ip_tos;
ofpbuf_use_stub(&odp_actions, odp_actions_stub, sizeof odp_actions_stub);
action_xlate_ctx_init(&ctx, ofproto, flow, &initial_vals,
- rule, stats.tcp_flags, packet);
+ rule, stats.tcp_flags, packet,
+ RECIRCULATION_ID_DUMMY);
ctx.resubmit_stats = &stats;
xlate_actions(&ctx, rule->up.ofpacts, rule->up.ofpacts_len, &odp_actions);
@@ -6320,6 +6527,16 @@ execute_dec_mpls_ttl_action(struct action_xlate_ctx *ctx)
}
static void
+execute_recircualte_action(struct action_xlate_ctx *ctx)
+{
+ if (ctx->recirculation_id == RECIRCULATION_ID_NONE) {
+ ctx->recirculation_id = get_recirculation_id();
+ }
+ ctx->recircualted = true;
+ ctx->flow.skb_mark = ctx->recirculation_id;
+}
+
+static void
xlate_output_action(struct action_xlate_ctx *ctx,
uint16_t port, uint16_t max_len, bool may_packet_in)
{
@@ -6560,6 +6777,7 @@ do_xlate_actions(const struct ofpact *ofpacts, size_t ofpacts_len,
struct action_xlate_ctx *ctx)
{
bool was_evictable = true;
+ bool may_recirculate = false;
const struct ofpact *a;
if (ctx->rule) {
@@ -6628,18 +6846,30 @@ do_xlate_actions(const struct ofpact *ofpacts, size_t ofpacts_len,
case OFPACT_SET_IPV4_SRC:
if (ctx->flow.dl_type == htons(ETH_TYPE_IP)) {
+ if (may_recirculate) {
+ execute_recircualte_action(ctx);
+ goto out;
+ }
ctx->flow.nw_src = ofpact_get_SET_IPV4_SRC(a)->ipv4;
}
break;
case OFPACT_SET_IPV4_DST:
if (ctx->flow.dl_type == htons(ETH_TYPE_IP)) {
+ if (may_recirculate) {
+ execute_recircualte_action(ctx);
+ goto out;
+ }
ctx->flow.nw_dst = ofpact_get_SET_IPV4_DST(a)->ipv4;
}
break;
case OFPACT_SET_IPV4_DSCP:
/* OpenFlow 1.0 only supports IPv4. */
+ if (may_recirculate) {
+ execute_recircualte_action(ctx);
+ goto out;
+ }
if (ctx->flow.dl_type == htons(ETH_TYPE_IP)) {
ctx->flow.nw_tos &= ~IP_DSCP_MASK;
ctx->flow.nw_tos |= ofpact_get_SET_IPV4_DSCP(a)->dscp;
@@ -6648,12 +6878,20 @@ do_xlate_actions(const struct ofpact *ofpacts, size_t ofpacts_len,
case OFPACT_SET_L4_SRC_PORT:
if (is_ip_any(&ctx->flow)) {
+ if (may_recirculate) {
+ execute_recircualte_action(ctx);
+ goto out;
+ }
ctx->flow.tp_src = htons(ofpact_get_SET_L4_SRC_PORT(a)->port);
}
break;
case OFPACT_SET_L4_DST_PORT:
if (is_ip_any(&ctx->flow)) {
+ if (may_recirculate) {
+ execute_recircualte_action(ctx);
+ goto out;
+ }
ctx->flow.tp_dst = htons(ofpact_get_SET_L4_DST_PORT(a)->port);
}
break;
@@ -6694,10 +6932,15 @@ do_xlate_actions(const struct ofpact *ofpacts, size_t ofpacts_len,
case OFPACT_PUSH_MPLS:
execute_mpls_push_action(ctx, ofpact_get_PUSH_MPLS(a)->ethertype);
+ may_recirculate = false;
break;
case OFPACT_POP_MPLS:
execute_mpls_pop_action(ctx, ofpact_get_POP_MPLS(a)->ethertype);
+ if (ctx->flow.dl_type == htons(ETH_TYPE_IP) ||
+ ctx->flow.dl_type == htons(ETH_TYPE_IPV6)) {
+ may_recirculate = true;
+ }
break;
case OFPACT_SET_MPLS_TTL:
@@ -6713,7 +6956,10 @@ do_xlate_actions(const struct ofpact *ofpacts, size_t ofpacts_len,
break;
case OFPACT_DEC_TTL:
- if (compose_dec_ttl(ctx, ofpact_get_DEC_TTL(a))) {
+ if (may_recirculate) {
+ execute_recircualte_action(ctx);
+ goto out;
+ } else if (compose_dec_ttl(ctx, ofpact_get_DEC_TTL(a))) {
goto out;
}
break;
@@ -6800,6 +7046,7 @@ do_xlate_actions(const struct ofpact *ofpacts, size_t ofpacts_len,
}
out:
+ ctx->ofpacts_len = (char *)(a) - (char *)ofpacts;
if (ctx->rule) {
ctx->rule->up.evictable = was_evictable;
}
@@ -6810,7 +7057,8 @@ action_xlate_ctx_init(struct action_xlate_ctx *ctx,
struct ofproto_dpif *ofproto, const struct flow *flow,
const struct initial_vals *initial_vals,
struct rule_dpif *rule,
- uint8_t tcp_flags, const struct ofpbuf *packet)
+ uint8_t tcp_flags, const struct ofpbuf *packet,
+ uint32_t recirculation_id)
{
ovs_be64 initial_tun_id = flow->tunnel.tun_id;
@@ -6833,7 +7081,13 @@ action_xlate_ctx_init(struct action_xlate_ctx *ctx,
* registers.
* - Tunnel 'base_flow' is completely cleared since that is what the
* kernel does. If we wish to maintain the original values an action
- * needs to be generated. */
+ * needs to be generated.
+ * - The recirculation_id element of flow and base flow are set to
+ * recirculate_id, which is the id that will be used by a recirculation
+ * action of one is added. It is stored in flow and base_flow for
+ * convenience as the recirculation_id element of flow and base flow
+ * are otherwise unused by action_xlate_ctx_init().
+ */
ctx->ofproto = ofproto;
ctx->flow = *flow;
@@ -6849,6 +7103,7 @@ action_xlate_ctx_init(struct action_xlate_ctx *ctx,
ctx->resubmit_hook = NULL;
ctx->report_hook = NULL;
ctx->resubmit_stats = NULL;
+ ctx->recirculation_id = recirculation_id;
}
/* Translates the 'ofpacts_len' bytes of "struct ofpacts" starting at 'ofpacts'
@@ -6885,6 +7140,7 @@ xlate_actions(struct action_xlate_ctx *ctx,
ctx->orig_skb_priority = ctx->flow.skb_priority;
ctx->table_id = 0;
ctx->exit = false;
+ ctx->recircualted = false;
ofpbuf_use_stub(&ctx->stack, ctx->init_stack, sizeof ctx->init_stack);
@@ -6933,6 +7189,11 @@ xlate_actions(struct action_xlate_ctx *ctx,
if (tunnel_ecn_ok(ctx) && (!in_port || may_receive(in_port, ctx))) {
do_xlate_actions(ofpacts, ofpacts_len, ctx);
+ if (ctx->recircualted) {
+ commit_odp_actions(&ctx->flow, &ctx->base_flow,
+ ctx->odp_actions);
+ commit_odp_recirculate_action(odp_actions);
+ }
/* We've let OFPP_NORMAL and the learning action look at the
* packet, so drop it now if forwarding is disabled. */
@@ -7692,7 +7953,8 @@ packet_out(struct ofproto *ofproto_, struct ofpbuf *packet,
initial_vals.vlan_tci = flow->vlan_tci;
initial_vals.tunnel_ip_tos = 0;
action_xlate_ctx_init(&ctx, ofproto, flow, &initial_vals, NULL,
- packet_get_tcp_flags(packet, flow), packet);
+ packet_get_tcp_flags(packet, flow), packet,
+ RECIRCULATION_ID_DUMMY);
ctx.resubmit_stats = &stats;
ofpbuf_use_stub(&odp_actions,
@@ -8077,7 +8339,7 @@ ofproto_trace(struct ofproto_dpif *ofproto, const struct flow *flow,
ofpbuf_use_stub(&odp_actions,
odp_actions_stub, sizeof odp_actions_stub);
action_xlate_ctx_init(&trace.ctx, ofproto, flow, initial_vals,
- rule, tcp_flags, packet);
+ rule, tcp_flags, packet, RECIRCULATION_ID_DUMMY);
trace.ctx.resubmit_hook = trace_resubmit;
trace.ctx.report_hook = trace_report;
xlate_actions(&trace.ctx, rule->up.ofpacts, rule->up.ofpacts_len,
--
1.7.10.4
^ permalink raw reply related [flat|nested] 14+ messages in thread
* [PATCH 2/4] Move execute_set_action to lib/odp-util.c
[not found] ` <1365403431-18102-1-git-send-email-horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org>
2013-04-08 6:43 ` [PATCH 1/4] Add packet recirculation Simon Horman
@ 2013-04-08 6:43 ` Simon Horman
[not found] ` <1365403431-18102-3-git-send-email-horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org>
2013-04-08 6:43 ` [PATCH 4/4] Avoid recirculation id collision Simon Horman
2 siblings, 1 reply; 14+ messages in thread
From: Simon Horman @ 2013-04-08 6:43 UTC (permalink / raw)
To: dev-yBygre7rU0TnMu66kgdUjQ, netdev-u79uwXL29TY76Z2rM5mHXA
Cc: Isaku Yamahata, Ravi K
Move execute_set_action from lib/dpif-netedev.c to lib/odp-util.c
This is in preparation for using execute_set_action()
in lib/odp-util.c to handle recirculation/
Signed-off-by: Simon Horman <horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org>
---
packet.c might be a better place for execute_set_action()
but I'm unsure if accessing struct ovs_key_ethernet would
lead to a layering violation.
This patch depends on the patch "Add packet recirculation"
v5
* No change
rfc4
* make use of skb_mark
rfc2 - rfc3
* omitted
rfc1
* Initial post
Conflicts:
lib/dpif-netdev.c
---
lib/dpif-netdev.c | 76 -----------------------------------------------------
lib/odp-util.c | 76 +++++++++++++++++++++++++++++++++++++++++++++++++++++
lib/odp-util.h | 3 +++
3 files changed, 79 insertions(+), 76 deletions(-)
diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
index 31255f6..e698e1e 100644
--- a/lib/dpif-netdev.c
+++ b/lib/dpif-netdev.c
@@ -1115,15 +1115,6 @@ dpif_netdev_wait(struct dpif *dpif)
}
static void
-dp_netdev_set_dl(struct ofpbuf *packet, const struct ovs_key_ethernet *eth_key)
-{
- struct eth_header *eh = packet->l2;
-
- memcpy(eh->eth_src, eth_key->eth_src, sizeof eh->eth_src);
- memcpy(eh->eth_dst, eth_key->eth_dst, sizeof eh->eth_dst);
-}
-
-static void
dp_netdev_output_port(struct dp_netdev *dp, struct ofpbuf *packet,
uint32_t out_port)
{
@@ -1228,73 +1219,6 @@ dp_netdev_action_userspace(struct dp_netdev *dp,
dp_netdev_output_userspace(dp, packet, DPIF_UC_ACTION, key, userdata);
}
-static void
-execute_set_action(struct ofpbuf *packet, const struct nlattr *a,
- uint32_t *skb_mark)
-{
- enum ovs_key_attr type = nl_attr_type(a);
- const struct ovs_key_ipv4 *ipv4_key;
- const struct ovs_key_ipv6 *ipv6_key;
- const struct ovs_key_tcp *tcp_key;
- const struct ovs_key_udp *udp_key;
-
- switch (type) {
- case OVS_KEY_ATTR_PRIORITY:
- case OVS_KEY_ATTR_TUNNEL:
- /* not implemented */
- break;
-
- case OVS_KEY_ATTR_SKB_MARK:
- *skb_mark = nl_attr_get_u32(a);
- break;
-
- case OVS_KEY_ATTR_ETHERNET:
- dp_netdev_set_dl(packet,
- nl_attr_get_unspec(a, sizeof(struct ovs_key_ethernet)));
- break;
-
- case OVS_KEY_ATTR_IPV4:
- ipv4_key = nl_attr_get_unspec(a, sizeof(struct ovs_key_ipv4));
- packet_set_ipv4(packet, ipv4_key->ipv4_src, ipv4_key->ipv4_dst,
- ipv4_key->ipv4_tos, ipv4_key->ipv4_ttl);
- break;
-
- case OVS_KEY_ATTR_IPV6:
- ipv6_key = nl_attr_get_unspec(a, sizeof(struct ovs_key_ipv6));
- packet_set_ipv6(packet, ipv6_key->ipv6_proto, ipv6_key->ipv6_src,
- ipv6_key->ipv6_dst, ipv6_key->ipv6_tclass,
- ipv6_key->ipv6_label, ipv6_key->ipv6_hlimit);
- break;
-
- case OVS_KEY_ATTR_TCP:
- tcp_key = nl_attr_get_unspec(a, sizeof(struct ovs_key_tcp));
- packet_set_tcp_port(packet, tcp_key->tcp_src, tcp_key->tcp_dst);
- break;
-
- case OVS_KEY_ATTR_UDP:
- udp_key = nl_attr_get_unspec(a, sizeof(struct ovs_key_udp));
- packet_set_udp_port(packet, udp_key->udp_src, udp_key->udp_dst);
- break;
-
- case OVS_KEY_ATTR_MPLS:
- set_mpls_lse(packet, nl_attr_get_be32(a));
- break;
-
- case OVS_KEY_ATTR_UNSPEC:
- case OVS_KEY_ATTR_ENCAP:
- case OVS_KEY_ATTR_ETHERTYPE:
- case OVS_KEY_ATTR_IN_PORT:
- case OVS_KEY_ATTR_VLAN:
- case OVS_KEY_ATTR_ICMP:
- case OVS_KEY_ATTR_ICMPV6:
- case OVS_KEY_ATTR_ARP:
- case OVS_KEY_ATTR_ND:
- case __OVS_KEY_ATTR_MAX:
- default:
- NOT_REACHED();
- }
-}
-
static bool
dp_netdev_execute_actions(struct dp_netdev *dp,
struct ofpbuf *packet, struct flow *key,
diff --git a/lib/odp-util.c b/lib/odp-util.c
index e18e109..ad5873c 100644
--- a/lib/odp-util.c
+++ b/lib/odp-util.c
@@ -2420,3 +2420,79 @@ commit_odp_actions(const struct flow *flow, struct flow *base,
commit_set_priority_action(flow, base, odp_actions);
commit_set_skb_mark_action(flow, base, odp_actions);
}
+
+static void
+dp_netdev_set_dl(struct ofpbuf *packet, const struct ovs_key_ethernet *eth_key)
+{
+ struct eth_header *eh = packet->l2;
+
+ memcpy(eh->eth_src, eth_key->eth_src, sizeof eh->eth_src);
+ memcpy(eh->eth_dst, eth_key->eth_dst, sizeof eh->eth_dst);
+}
+
+void
+execute_set_action(struct ofpbuf *packet, const struct nlattr *a,
+ uint32_t *skb_mark)
+{
+ enum ovs_key_attr type = nl_attr_type(a);
+ const struct ovs_key_ipv4 *ipv4_key;
+ const struct ovs_key_ipv6 *ipv6_key;
+ const struct ovs_key_tcp *tcp_key;
+ const struct ovs_key_udp *udp_key;
+
+ switch (type) {
+ case OVS_KEY_ATTR_PRIORITY:
+ case OVS_KEY_ATTR_TUNNEL:
+ /* not implemented */
+ break;
+
+ case OVS_KEY_ATTR_SKB_MARK:
+ *skb_mark = nl_attr_get_u32(a);
+ break;
+
+ case OVS_KEY_ATTR_ETHERNET:
+ dp_netdev_set_dl(packet,
+ nl_attr_get_unspec(a, sizeof(struct ovs_key_ethernet)));
+ break;
+
+ case OVS_KEY_ATTR_IPV4:
+ ipv4_key = nl_attr_get_unspec(a, sizeof(struct ovs_key_ipv4));
+ packet_set_ipv4(packet, ipv4_key->ipv4_src, ipv4_key->ipv4_dst,
+ ipv4_key->ipv4_tos, ipv4_key->ipv4_ttl);
+ break;
+
+ case OVS_KEY_ATTR_IPV6:
+ ipv6_key = nl_attr_get_unspec(a, sizeof(struct ovs_key_ipv6));
+ packet_set_ipv6(packet, ipv6_key->ipv6_proto, ipv6_key->ipv6_src,
+ ipv6_key->ipv6_dst, ipv6_key->ipv6_tclass,
+ ipv6_key->ipv6_label, ipv6_key->ipv6_hlimit);
+ break;
+
+ case OVS_KEY_ATTR_TCP:
+ tcp_key = nl_attr_get_unspec(a, sizeof(struct ovs_key_tcp));
+ packet_set_tcp_port(packet, tcp_key->tcp_src, tcp_key->tcp_dst);
+ break;
+
+ case OVS_KEY_ATTR_UDP:
+ udp_key = nl_attr_get_unspec(a, sizeof(struct ovs_key_udp));
+ packet_set_udp_port(packet, udp_key->udp_src, udp_key->udp_dst);
+ break;
+
+ case OVS_KEY_ATTR_MPLS:
+ set_mpls_lse(packet, nl_attr_get_be32(a));
+ break;
+
+ case OVS_KEY_ATTR_UNSPEC:
+ case OVS_KEY_ATTR_ENCAP:
+ case OVS_KEY_ATTR_ETHERTYPE:
+ case OVS_KEY_ATTR_IN_PORT:
+ case OVS_KEY_ATTR_VLAN:
+ case OVS_KEY_ATTR_ICMP:
+ case OVS_KEY_ATTR_ICMPV6:
+ case OVS_KEY_ATTR_ARP:
+ case OVS_KEY_ATTR_ND:
+ case __OVS_KEY_ATTR_MAX:
+ default:
+ NOT_REACHED();
+ }
+}
diff --git a/lib/odp-util.h b/lib/odp-util.h
index da62aa5..637d6a5 100644
--- a/lib/odp-util.h
+++ b/lib/odp-util.h
@@ -159,6 +159,9 @@ void odp_put_tunnel_action(const struct flow_tnl *tunnel,
void odp_put_skb_mark_action(const uint32_t skb_mark,
struct ofpbuf *odp_actions);
+void execute_set_action(struct ofpbuf *packet, const struct nlattr *a,
+ uint32_t *skb_mark);
+
/* Reasons why a subfacet might not be fast-pathable. */
enum slow_path_reason {
/* These reasons are mutually exclusive. */
--
1.7.10.4
^ permalink raw reply related [flat|nested] 14+ messages in thread
* [PATCH 3/4] Allow recirculation without facets
2013-04-08 6:43 [PATCH v5 0/4] Add packet recirculation Simon Horman
@ 2013-04-08 6:43 ` Simon Horman
[not found] ` <1365403431-18102-1-git-send-email-horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org>
1 sibling, 0 replies; 14+ messages in thread
From: Simon Horman @ 2013-04-08 6:43 UTC (permalink / raw)
To: dev, netdev; +Cc: Ravi K, Isaku Yamahata, Jesse Gross, Ben Pfaff
This covers the following cases:
* Handle flow miss without facet
- Previously the use of facets was forced if there was
any chance of a recirculation action. That is, for
all flows misses of MPLS packets.
* Packet Out
Signed-off-by: Simon Horman <horms@verge.net.au>
---
This patch depends on the patch "Move execute_set_action to lib/odp-util.c"
v5
* Use RECIRCULATION_ID_DUMMY
rfc4
* Initial post
---
ofproto/ofproto-dpif.c | 132 ++++++++++++++++++++++++++++++++++++++++--------
1 file changed, 111 insertions(+), 21 deletions(-)
diff --git a/ofproto/ofproto-dpif.c b/ofproto/ofproto-dpif.c
index 5129da1..67121f2 100644
--- a/ofproto/ofproto-dpif.c
+++ b/ofproto/ofproto-dpif.c
@@ -3571,6 +3571,64 @@ flow_miss_find(struct hmap *todo, const struct ofproto_dpif *ofproto,
return NULL;
}
+static void
+execute_actions_for_recircualtion(struct ofpbuf *packet,
+ const struct nlattr *actions,
+ size_t actions_len, uint32_t *skb_mark)
+{
+ const struct nlattr *a;
+ unsigned int left;
+
+ NL_ATTR_FOR_EACH_UNSAFE (a, left, actions, actions_len) {
+ int type = nl_attr_type(a);
+
+ switch ((enum ovs_action_attr) type) {
+
+ case OVS_ACTION_ATTR_PUSH_VLAN: {
+ const struct ovs_action_push_vlan *vlan = nl_attr_get(a);
+ eth_push_vlan(packet, vlan->vlan_tci);
+ break;
+ }
+
+ case OVS_ACTION_ATTR_POP_VLAN:
+ eth_pop_vlan(packet);
+ break;
+
+ case OVS_ACTION_ATTR_PUSH_MPLS: {
+ const struct ovs_action_push_mpls *mpls = nl_attr_get(a);
+ push_mpls(packet, mpls->mpls_ethertype, mpls->mpls_lse);
+ break;
+ }
+
+ case OVS_ACTION_ATTR_POP_MPLS:
+ pop_mpls(packet, nl_attr_get_be16(a));
+ break;
+
+ case OVS_ACTION_ATTR_SET:
+ execute_set_action(packet, nl_attr_get(a), skb_mark);
+ break;
+
+ case OVS_ACTION_ATTR_RECIRCULATE:
+ if (packet->l2) {
+ ofpbuf_push_uninit(packet, (char *)packet->l2 -
+ (char *)packet->data);
+ }
+ return;
+
+ case OVS_ACTION_ATTR_OUTPUT:
+ case OVS_ACTION_ATTR_USERSPACE:
+ case OVS_ACTION_ATTR_SAMPLE:
+ case OVS_ACTION_ATTR_UNSPEC:
+ case __OVS_ACTION_ATTR_MAX:
+ NOT_REACHED();
+ }
+ }
+
+ /* There should always be a OVS_ACTION_ATTR_RECIRCULATE present
+ * in actions if this function is called */
+ NOT_REACHED();
+}
+
/* Partially Initializes 'op' as an "execute" operation for 'miss' and
* 'packet'. The caller must initialize op->actions and op->actions_len. If
* 'miss' is associated with a subfacet the caller must also initialize the
@@ -3632,15 +3690,6 @@ static bool
flow_miss_should_make_facet(struct ofproto_dpif *ofproto,
struct flow_miss *miss, uint32_t hash)
{
- /* A facet is currently required to handle recirculation.
- * There currently isn't a good way to detect if recirculation will
- * occur or not. So in the mean time assume that it can't occur
- * for non-MPLS packets and it may occur for MPLS packets
- */
- if (eth_type_mpls(miss->flow.dl_type)) {
- return true;
- }
-
if (!ofproto->governor) {
size_t n_subfacets;
@@ -3656,6 +3705,50 @@ flow_miss_should_make_facet(struct ofproto_dpif *ofproto,
list_size(&miss->packets));
}
+static const struct flow *
+xlate_with_recirculate(struct ofproto_dpif *ofproto, struct rule_dpif *rule,
+ const struct flow *flow, struct flow *flow_storage,
+ const struct initial_vals *initial_vals,
+ const struct ofpact *ofpacts, size_t ofpacts_len,
+ struct ofpbuf *odp_actions,
+ struct dpif_flow_stats *stats, struct ofpbuf *packet)
+{
+ struct initial_vals initial_vals_ = *initial_vals;
+
+ while (1) {
+ struct action_xlate_ctx ctx;
+ uint32_t skb_mark = flow->skb_mark;
+
+ ofpbuf_clear(odp_actions);
+ action_xlate_ctx_init(&ctx, ofproto, flow, &initial_vals_,
+ rule, stats->tcp_flags, packet,
+ RECIRCULATION_ID_DUMMY);
+ ctx.resubmit_stats = stats;
+ xlate_actions(&ctx, ofpacts, ofpacts_len, odp_actions);
+
+ if (!ctx.recircualted) {
+ break;
+ }
+
+ /* Update the packet */
+ execute_actions_for_recircualtion(packet, odp_actions->data,
+ odp_actions->size, &skb_mark);
+ ofpbuf_clear(odp_actions);
+
+ /* Replace the flow */
+ flow_extract(packet, flow->skb_priority, skb_mark,
+ NULL, flow->in_port, flow_storage);
+ flow = flow_storage;
+ initial_vals_.vlan_tci = flow->vlan_tci;
+ initial_vals_.tunnel_ip_tos = flow->tunnel.ip_tos;
+
+ ofpacts = ofpact_end(ofpacts, ctx.ofpacts_len);
+ ofpacts_len -= ctx.ofpacts_len;
+ }
+
+ return flow;
+}
+
/* Handles 'miss', which matches 'rule', without creating a facet or subfacet
* or creating any datapath flow. May add an "execute" operation to 'ops' and
* increment '*n_ops'. */
@@ -3666,8 +3759,8 @@ handle_flow_miss_without_facet(struct flow_miss *miss, struct rule_dpif *rule,
{
struct ofproto_dpif *ofproto = ofproto_dpif_cast(rule->up.ofproto);
long long int now = time_msec();
- struct action_xlate_ctx ctx;
struct ofpbuf *packet;
+ struct flow flow_storage;
LIST_FOR_EACH (packet, list_node, &miss->packets) {
struct flow_miss_op *op = &ops[*n_ops];
@@ -3681,11 +3774,9 @@ handle_flow_miss_without_facet(struct flow_miss *miss, struct rule_dpif *rule,
dpif_flow_stats_extract(&miss->flow, packet, now, &stats);
rule_credit_stats(rule, &stats);
- action_xlate_ctx_init(&ctx, ofproto, &miss->flow,
- &miss->initial_vals, rule, 0, packet,
- RECIRCULATION_ID_DUMMY);
- ctx.resubmit_stats = &stats;
- xlate_actions(&ctx, ofpacts, ofpacts_len, &odp_actions);
+ xlate_with_recirculate(ofproto, rule, &miss->flow, &flow_storage,
+ &miss->initial_vals, ofpacts, ofpacts_len,
+ &odp_actions, &stats, packet);
if (odp_actions.size) {
struct dpif_execute *execute = &op->dpif_op.u.execute;
@@ -7937,10 +8028,10 @@ packet_out(struct ofproto *ofproto_, struct ofpbuf *packet,
struct initial_vals initial_vals;
struct odputil_keybuf keybuf;
struct dpif_flow_stats stats;
+ struct flow flow_storage;
struct ofpbuf key;
- struct action_xlate_ctx ctx;
uint64_t odp_actions_stub[1024 / 8];
struct ofpbuf odp_actions;
@@ -7952,14 +8043,13 @@ packet_out(struct ofproto *ofproto_, struct ofpbuf *packet,
initial_vals.vlan_tci = flow->vlan_tci;
initial_vals.tunnel_ip_tos = 0;
- action_xlate_ctx_init(&ctx, ofproto, flow, &initial_vals, NULL,
- packet_get_tcp_flags(packet, flow), packet,
- RECIRCULATION_ID_DUMMY);
- ctx.resubmit_stats = &stats;
ofpbuf_use_stub(&odp_actions,
odp_actions_stub, sizeof odp_actions_stub);
- xlate_actions(&ctx, ofpacts, ofpacts_len, &odp_actions);
+ flow = xlate_with_recirculate(ofproto, NULL, flow, &flow_storage,
+ &initial_vals, ofpacts, ofpacts_len,
+ &odp_actions, &stats, packet);
+
dpif_execute(ofproto->backer->dpif, key.data, key.size,
odp_actions.data, odp_actions.size, packet);
ofpbuf_uninit(&odp_actions);
--
1.7.10.4
^ permalink raw reply related [flat|nested] 14+ messages in thread
* [PATCH 4/4] Avoid recirculation id collision
[not found] ` <1365403431-18102-1-git-send-email-horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org>
2013-04-08 6:43 ` [PATCH 1/4] Add packet recirculation Simon Horman
2013-04-08 6:43 ` [PATCH 2/4] Move execute_set_action to lib/odp-util.c Simon Horman
@ 2013-04-08 6:43 ` Simon Horman
2 siblings, 0 replies; 14+ messages in thread
From: Simon Horman @ 2013-04-08 6:43 UTC (permalink / raw)
To: dev-yBygre7rU0TnMu66kgdUjQ, netdev-u79uwXL29TY76Z2rM5mHXA
Cc: Isaku Yamahata, Ravi K
Avoid recirculation id collision by checking that an id is
not already associated with a facet.
Consecutive recirculation ids are used and thus it is possible for
there to be situations where a very large number of ids have to
be checked before finding one that is not already associated with a facet.
To mitigate the performance impact of such situations a limit on
the number of checks is in place and if no unused recirculation id
can be found then the miss is handled without facets as this can
be done using a dummy recirculation id.
Signed-off-by: Simon Horman <horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org>
This patch depends on the patch "Allow recirculation without facets"
---
v5
* First post
---
ofproto/ofproto-dpif.c | 55 +++++++++++++++++++++++++++++++++++++++---------
1 file changed, 45 insertions(+), 10 deletions(-)
diff --git a/ofproto/ofproto-dpif.c b/ofproto/ofproto-dpif.c
index 67121f2..e9ab58c 100644
--- a/ofproto/ofproto-dpif.c
+++ b/ofproto/ofproto-dpif.c
@@ -3455,24 +3455,51 @@ port_is_lacp_current(const struct ofport *ofport_)
#define RECIRCULATION_ID_DUMMY 2
#define RECIRCULATION_ID_MIN RECIRCULATION_ID_DUMMY
+#define RECIRCULATION_ID_MAX_LOOP 1024 /* Arbitrary value to prevent
+ * endless loop */
+
static uint32_t recirculation_id_hash(uint32_t id)
{
return hash_words(&id, 1, 0);
}
-/* XXX: This does not prevent id collision */
-static uint32_t get_recirculation_id(void)
+static uint32_t recirculation_id = RECIRCULATION_ID_MIN;
+static uint32_t validated_recirculation_id = RECIRCULATION_ID_NONE;
+
+static uint32_t peek_recirculation_id(struct ofproto_dpif *ofproto)
{
- static uint32_t id = RECIRCULATION_ID_MIN;
+ static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 15);
+
+ int loop = RECIRCULATION_ID_MAX_LOOP;
- if (id < RECIRCULATION_ID_MIN)
- id = RECIRCULATION_ID_MIN;
- /* Skip IPSEC_MARK bit it is reserved */
- if (id & IPSEC_MARK) {
- id++;
- ovs_assert(!(id & IPSEC_MARK));
+ if (validated_recirculation_id == recirculation_id) {
+ return recirculation_id;
+ }
+
+ while (loop--) {
+ if (recirculation_id < RECIRCULATION_ID_MIN)
+ recirculation_id = RECIRCULATION_ID_MIN;
+ /* Skip IPSEC_MARK bit it is reserved */
+ if (recirculation_id & IPSEC_MARK) {
+ recirculation_id++;
+ ovs_assert(!(recirculation_id & IPSEC_MARK));
+ }
+ if (!facet_find_by_id(ofproto, recirculation_id)) {
+ validated_recirculation_id = recirculation_id;
+ return recirculation_id;
+ }
+ recirculation_id++;
}
- return id++;
+
+ VLOG_WARN_RL(&rl, "Failed to allocate recirulation id after %d attempts\n",
+ RECIRCULATION_ID_MAX_LOOP);
+ return RECIRCULATION_ID_NONE;
+}
+
+static uint32_t get_recirculation_id(void)
+{
+ ovs_assert(recirculation_id == validated_recirculation_id);
+ return recirculation_id++;
}
\f
/* Upcall handling. */
@@ -3690,6 +3717,14 @@ static bool
flow_miss_should_make_facet(struct ofproto_dpif *ofproto,
struct flow_miss *miss, uint32_t hash)
{
+ /* If the packet is MPLS then recirculation may be used and
+ * this will not be possible with facets if there are no recirculation
+ * ids available */
+ if (eth_type_mpls(miss->flow.dl_type) &&
+ peek_recirculation_id(ofproto) == RECIRCULATION_ID_NONE) {
+ return false;
+ }
+
if (!ofproto->governor) {
size_t n_subfacets;
--
1.7.10.4
^ permalink raw reply related [flat|nested] 14+ messages in thread
* Re: [PATCH 2/4] Move execute_set_action to lib/odp-util.c
[not found] ` <1365403431-18102-3-git-send-email-horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org>
@ 2013-04-08 20:29 ` Jesse Gross
2013-04-09 3:11 ` Simon Horman
0 siblings, 1 reply; 14+ messages in thread
From: Jesse Gross @ 2013-04-08 20:29 UTC (permalink / raw)
To: Simon Horman
Cc: dev-yBygre7rU0TnMu66kgdUjQ@public.gmane.org, netdev, Ravi K,
Isaku Yamahata
On Sun, Apr 7, 2013 at 11:43 PM, Simon Horman <horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org> wrote:
> Move execute_set_action from lib/dpif-netedev.c to lib/odp-util.c
>
> This is in preparation for using execute_set_action()
> in lib/odp-util.c to handle recirculation/
>
> Signed-off-by: Simon Horman <horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org>
>
> ---
>
> packet.c might be a better place for execute_set_action()
> but I'm unsure if accessing struct ovs_key_ethernet would
> lead to a layering violation.
I'd be tempted to just put this in it's own file. As you say, it
doesn't really fit in either of the two existing ones.
> diff --git a/lib/odp-util.c b/lib/odp-util.c
> index e18e109..ad5873c 100644
> --- a/lib/odp-util.c
> +++ b/lib/odp-util.c
> @@ -2420,3 +2420,79 @@ commit_odp_actions(const struct flow *flow, struct flow *base,
> commit_set_priority_action(flow, base, odp_actions);
> commit_set_skb_mark_action(flow, base, odp_actions);
> }
> +
> +static void
> +dp_netdev_set_dl(struct ofpbuf *packet, const struct ovs_key_ethernet *eth_key)
I think this function should be given a more generic name and possibly
moved to packet.c.
> +void
> +execute_set_action(struct ofpbuf *packet, const struct nlattr *a,
> + uint32_t *skb_mark)
> +{
> + enum ovs_key_attr type = nl_attr_type(a);
> + const struct ovs_key_ipv4 *ipv4_key;
> + const struct ovs_key_ipv6 *ipv6_key;
> + const struct ovs_key_tcp *tcp_key;
> + const struct ovs_key_udp *udp_key;
> +
> + switch (type) {
> + case OVS_KEY_ATTR_PRIORITY:
> + case OVS_KEY_ATTR_TUNNEL:
> + /* not implemented */
> + break;
Don't we need to carry this information along as well similar to skb->mark?
Also, is there a reason to not have the code for push/pop actions here as well?
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH 1/4] Add packet recirculation
2013-04-08 6:43 ` [PATCH 1/4] Add packet recirculation Simon Horman
@ 2013-04-09 1:46 ` Jesse Gross
2013-04-09 7:50 ` Simon Horman
0 siblings, 1 reply; 14+ messages in thread
From: Jesse Gross @ 2013-04-09 1:46 UTC (permalink / raw)
To: Simon Horman
Cc: dev@openvswitch.org, netdev, Ravi K, Isaku Yamahata, Ben Pfaff
On Sun, Apr 7, 2013 at 11:43 PM, Simon Horman <horms@verge.net.au> wrote:
> diff --git a/datapath/actions.c b/datapath/actions.c
> index e9634fe..7b0f022 100644
> --- a/datapath/actions.c
> +++ b/datapath/actions.c
> @@ -617,6 +617,9 @@ static int do_execute_actions(struct datapath *dp, struct sk_buff *skb,
> case OVS_ACTION_ATTR_SAMPLE:
> err = sample(dp, skb, a);
> break;
> +
> + case OVS_ACTION_ATTR_RECIRCULATE:
> + return 1;
I think that if we've had a previous output action with the port
stored in prev_port then this will cause the packet to not actually be
output.
> diff --git a/datapath/datapath.c b/datapath/datapath.c
> index e8be795..ab39dd7 100644
> --- a/datapath/datapath.c
> +++ b/datapath/datapath.c
> void ovs_dp_process_received_packet(struct vport *p, struct sk_buff *skb)
[...]
> + if (IS_ERR_OR_NULL(skb)) {
> + break;
> + } else if (unlikely(!limit--)) {
Should this be a predecrement?
> + kfree_skb(skb);
Should we log some kind of rate limited warning here?
> + return;
In the first case we use break to exit the loop and here we use
return. Both should have the same effect so it might be nice to make
them the same.
> @@ -901,6 +913,9 @@ static int validate_and_copy_actions__(const struct nlattr *attr,
> skip_copy = true;
> break;
>
> + case OVS_ACTION_ATTR_RECIRCULATE:
> + break;
I think we might want to jump out the loop here to better model how
the actions are actually executed.
> diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
> index e4a2f75..31255f6 100644
> --- a/lib/dpif-netdev.c
> +++ b/lib/dpif-netdev.c
> dp_netdev_port_input(struct dp_netdev *dp, struct dp_netdev_port *port,
> struct ofpbuf *packet)
[...]
> + } else {
> + dp->n_missed++;
> + dp_netdev_output_userspace(dp, packet, DPIF_UC_MISS, &key, NULL);
> + recirculate = false;
> + }
> + } while (recirculate && limit--);
I have the same question about predecrement here.
> @@ -1163,6 +1190,7 @@ dp_netdev_sample(struct dp_netdev *dp,
> const struct nlattr *subactions = NULL;
> const struct nlattr *a;
> size_t left;
> + uint32_t skb_mark;
I don't think it's right to have a new (and uninitialized) copy of
skb_mark here. We should have the same one all the way through, like
we do in the kernel.
> diff --git a/ofproto/ofproto-dpif.c b/ofproto/ofproto-dpif.c
> index 47830c1..5129da1 100644
> --- a/ofproto/ofproto-dpif.c
> +++ b/ofproto/ofproto-dpif.c
I'm still working on more detailed comments for this. However, I'm
concerned about whether the behavior for revalidation and stats is
correct.
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH 2/4] Move execute_set_action to lib/odp-util.c
2013-04-08 20:29 ` Jesse Gross
@ 2013-04-09 3:11 ` Simon Horman
2013-04-09 15:17 ` Jesse Gross
0 siblings, 1 reply; 14+ messages in thread
From: Simon Horman @ 2013-04-09 3:11 UTC (permalink / raw)
To: Jesse Gross
Cc: dev@openvswitch.org, netdev, Ravi K, Isaku Yamahata, Ben Pfaff
On Mon, Apr 08, 2013 at 01:29:52PM -0700, Jesse Gross wrote:
> On Sun, Apr 7, 2013 at 11:43 PM, Simon Horman <horms@verge.net.au> wrote:
> > Move execute_set_action from lib/dpif-netedev.c to lib/odp-util.c
> >
> > This is in preparation for using execute_set_action()
> > in lib/odp-util.c to handle recirculation/
> >
> > Signed-off-by: Simon Horman <horms@verge.net.au>
> >
> > ---
> >
> > packet.c might be a better place for execute_set_action()
> > but I'm unsure if accessing struct ovs_key_ethernet would
> > lead to a layering violation.
>
> I'd be tempted to just put this in it's own file. As you say, it
> doesn't really fit in either of the two existing ones.
perhaps execute-action.c ?
>
> > diff --git a/lib/odp-util.c b/lib/odp-util.c
> > index e18e109..ad5873c 100644
> > --- a/lib/odp-util.c
> > +++ b/lib/odp-util.c
> > @@ -2420,3 +2420,79 @@ commit_odp_actions(const struct flow *flow, struct flow *base,
> > commit_set_priority_action(flow, base, odp_actions);
> > commit_set_skb_mark_action(flow, base, odp_actions);
> > }
> > +
> > +static void
> > +dp_netdev_set_dl(struct ofpbuf *packet, const struct ovs_key_ethernet *eth_key)
>
> I think this function should be given a more generic name and possibly
> moved to packet.c.
Sure, how about eth_set_src_and_dst()
> > +void
> > +execute_set_action(struct ofpbuf *packet, const struct nlattr *a,
> > + uint32_t *skb_mark)
> > +{
> > + enum ovs_key_attr type = nl_attr_type(a);
> > + const struct ovs_key_ipv4 *ipv4_key;
> > + const struct ovs_key_ipv6 *ipv6_key;
> > + const struct ovs_key_tcp *tcp_key;
> > + const struct ovs_key_udp *udp_key;
> > +
> > + switch (type) {
> > + case OVS_KEY_ATTR_PRIORITY:
> > + case OVS_KEY_ATTR_TUNNEL:
> > + /* not implemented */
> > + break;
>
> Don't we need to carry this information along as well similar to skb->mark?
Most likely, sorry for missing that.
> Also, is there a reason to not have the code for push/pop actions here as well?
Good point.
With that in mind perhaps execute_set_or_mpls_action() would
be a good name for the function?
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH 1/4] Add packet recirculation
2013-04-09 1:46 ` Jesse Gross
@ 2013-04-09 7:50 ` Simon Horman
2013-04-09 15:44 ` Jesse Gross
0 siblings, 1 reply; 14+ messages in thread
From: Simon Horman @ 2013-04-09 7:50 UTC (permalink / raw)
To: Jesse Gross
Cc: dev@openvswitch.org, netdev, Ravi K, Isaku Yamahata, Ben Pfaff
On Mon, Apr 08, 2013 at 06:46:29PM -0700, Jesse Gross wrote:
> On Sun, Apr 7, 2013 at 11:43 PM, Simon Horman <horms@verge.net.au> wrote:
> > diff --git a/datapath/actions.c b/datapath/actions.c
> > index e9634fe..7b0f022 100644
> > --- a/datapath/actions.c
> > +++ b/datapath/actions.c
> > @@ -617,6 +617,9 @@ static int do_execute_actions(struct datapath *dp, struct sk_buff *skb,
> > case OVS_ACTION_ATTR_SAMPLE:
> > err = sample(dp, skb, a);
> > break;
> > +
> > + case OVS_ACTION_ATTR_RECIRCULATE:
> > + return 1;
>
> I think that if we've had a previous output action with the port
> stored in prev_port then this will cause the packet to not actually be
> output.
I'm not so sure.
I see something like this occurring:
1. Iteration of for loop for output action
switch (nla_type(a)) {
case OVS_ACTION_ATTR_OUTPUT:
prev_port = nla_get_u32(a);
break;
...
}
2. Iteration of of for loop for next action, lets say its is recirculate
i. Output packet
if (prev_port != -1) {
do_output(dp, skb_clone(skb, GFP_ATOMIC), prev_port);
prev_port = -1;
}
ii. Return due to recirculate
switch (nla_type(a)) {
...
case OVS_ACTION_ATTR_RECIRCULATE:
return 1;
}
Am I missing something?
> > diff --git a/datapath/datapath.c b/datapath/datapath.c
> > index e8be795..ab39dd7 100644
> > --- a/datapath/datapath.c
> > +++ b/datapath/datapath.c
> > void ovs_dp_process_received_packet(struct vport *p, struct sk_buff *skb)
> [...]
> > + if (IS_ERR_OR_NULL(skb)) {
> > + break;
> > + } else if (unlikely(!limit--)) {
>
> Should this be a predecrement?
I will make it so.
> > + kfree_skb(skb);
>
> Should we log some kind of rate limited warning here?
Sure.
> > + return;
>
> In the first case we use break to exit the loop and here we use
> return. Both should have the same effect so it might be nice to make
> them the same.
>
> > @@ -901,6 +913,9 @@ static int validate_and_copy_actions__(const struct nlattr *attr,
> > skip_copy = true;
> > break;
> >
> > + case OVS_ACTION_ATTR_RECIRCULATE:
> > + break;
>
> I think we might want to jump out the loop here to better model how
> the actions are actually executed.
Sure, perhaps something like this?
diff --git a/datapath/datapath.c b/datapath/datapath.c
index ab39dd7..721a52c 100644
--- a/datapath/datapath.c
+++ b/datapath/datapath.c
@@ -914,7 +914,7 @@ static int validate_and_copy_actions__(const struct nlattr *attr,
break;
case OVS_ACTION_ATTR_RECIRCULATE:
- break;
+ goto out;
default:
return -EINVAL;
@@ -926,6 +926,7 @@ static int validate_and_copy_actions__(const struct nlattr *attr,
}
}
+out:
if (rem > 0)
return -EINVAL;
> > diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
> > index e4a2f75..31255f6 100644
> > --- a/lib/dpif-netdev.c
> > +++ b/lib/dpif-netdev.c
> > dp_netdev_port_input(struct dp_netdev *dp, struct dp_netdev_port *port,
> > struct ofpbuf *packet)
> [...]
> > + } else {
> > + dp->n_missed++;
> > + dp_netdev_output_userspace(dp, packet, DPIF_UC_MISS, &key, NULL);
> > + recirculate = false;
> > + }
> > + } while (recirculate && limit--);
>
> I have the same question about predecrement here.
I will change this one too.
> > @@ -1163,6 +1190,7 @@ dp_netdev_sample(struct dp_netdev *dp,
> > const struct nlattr *subactions = NULL;
> > const struct nlattr *a;
> > size_t left;
> > + uint32_t skb_mark;
>
> I don't think it's right to have a new (and uninitialized) copy of
> skb_mark here. We should have the same one all the way through, like
> we do in the kernel.
Sure. I will pass it as an argument to dp_netdev_sample()
> > diff --git a/ofproto/ofproto-dpif.c b/ofproto/ofproto-dpif.c
> > index 47830c1..5129da1 100644
> > --- a/ofproto/ofproto-dpif.c
> > +++ b/ofproto/ofproto-dpif.c
>
> I'm still working on more detailed comments for this. However, I'm
> concerned about whether the behavior for revalidation and stats is
> correct.
I am a little concerned about that too.
Perhaps Ben could look over it?
^ permalink raw reply related [flat|nested] 14+ messages in thread
* Re: [PATCH 2/4] Move execute_set_action to lib/odp-util.c
2013-04-09 3:11 ` Simon Horman
@ 2013-04-09 15:17 ` Jesse Gross
0 siblings, 0 replies; 14+ messages in thread
From: Jesse Gross @ 2013-04-09 15:17 UTC (permalink / raw)
To: Simon Horman
Cc: dev@openvswitch.org, netdev, Ravi K, Isaku Yamahata, Ben Pfaff
On Mon, Apr 8, 2013 at 8:11 PM, Simon Horman <horms@verge.net.au> wrote:
> On Mon, Apr 08, 2013 at 01:29:52PM -0700, Jesse Gross wrote:
>> On Sun, Apr 7, 2013 at 11:43 PM, Simon Horman <horms@verge.net.au> wrote:
>> > Move execute_set_action from lib/dpif-netedev.c to lib/odp-util.c
>> >
>> > This is in preparation for using execute_set_action()
>> > in lib/odp-util.c to handle recirculation/
>> >
>> > Signed-off-by: Simon Horman <horms@verge.net.au>
>> >
>> > ---
>> >
>> > packet.c might be a better place for execute_set_action()
>> > but I'm unsure if accessing struct ovs_key_ethernet would
>> > lead to a layering violation.
>>
>> I'd be tempted to just put this in it's own file. As you say, it
>> doesn't really fit in either of the two existing ones.
>
> perhaps execute-action.c ?
Sure.
>>
>> > diff --git a/lib/odp-util.c b/lib/odp-util.c
>> > index e18e109..ad5873c 100644
>> > --- a/lib/odp-util.c
>> > +++ b/lib/odp-util.c
>> > @@ -2420,3 +2420,79 @@ commit_odp_actions(const struct flow *flow, struct flow *base,
>> > commit_set_priority_action(flow, base, odp_actions);
>> > commit_set_skb_mark_action(flow, base, odp_actions);
>> > }
>> > +
>> > +static void
>> > +dp_netdev_set_dl(struct ofpbuf *packet, const struct ovs_key_ethernet *eth_key)
>>
>> I think this function should be given a more generic name and possibly
>> moved to packet.c.
>
> Sure, how about eth_set_src_and_dst()
That sounds fine.
>> > +void
>> > +execute_set_action(struct ofpbuf *packet, const struct nlattr *a,
>> > + uint32_t *skb_mark)
>> > +{
>> > + enum ovs_key_attr type = nl_attr_type(a);
>> > + const struct ovs_key_ipv4 *ipv4_key;
>> > + const struct ovs_key_ipv6 *ipv6_key;
>> > + const struct ovs_key_tcp *tcp_key;
>> > + const struct ovs_key_udp *udp_key;
>> > +
>> > + switch (type) {
>> > + case OVS_KEY_ATTR_PRIORITY:
>> > + case OVS_KEY_ATTR_TUNNEL:
>> > + /* not implemented */
>> > + break;
>>
>> Don't we need to carry this information along as well similar to skb->mark?
>
> Most likely, sorry for missing that.
>
>> Also, is there a reason to not have the code for push/pop actions here as well?
>
> Good point.
>
> With that in mind perhaps execute_set_or_mpls_action() would
> be a good name for the function?
I'm not sure that this is specific to MPLS. Won't we basically just
have the execute loop from dpif-netdev.c here?
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH 1/4] Add packet recirculation
2013-04-09 7:50 ` Simon Horman
@ 2013-04-09 15:44 ` Jesse Gross
[not found] ` <CAEP_g=-baZGttBowXKGdgKafxp9fkUrQ=44y562ZzfnOL-XaQg-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>
0 siblings, 1 reply; 14+ messages in thread
From: Jesse Gross @ 2013-04-09 15:44 UTC (permalink / raw)
To: Simon Horman
Cc: dev@openvswitch.org, netdev, Ravi K, Isaku Yamahata, Ben Pfaff
On Tue, Apr 9, 2013 at 12:50 AM, Simon Horman <horms@verge.net.au> wrote:
> On Mon, Apr 08, 2013 at 06:46:29PM -0700, Jesse Gross wrote:
>> On Sun, Apr 7, 2013 at 11:43 PM, Simon Horman <horms@verge.net.au> wrote:
>> > diff --git a/datapath/actions.c b/datapath/actions.c
>> > index e9634fe..7b0f022 100644
>> > --- a/datapath/actions.c
>> > +++ b/datapath/actions.c
>> > @@ -617,6 +617,9 @@ static int do_execute_actions(struct datapath *dp, struct sk_buff *skb,
>> > case OVS_ACTION_ATTR_SAMPLE:
>> > err = sample(dp, skb, a);
>> > break;
>> > +
>> > + case OVS_ACTION_ATTR_RECIRCULATE:
>> > + return 1;
>>
>> I think that if we've had a previous output action with the port
>> stored in prev_port then this will cause the packet to not actually be
>> output.
>
> I'm not so sure.
>
> I see something like this occurring:
>
> 1. Iteration of for loop for output action
>
> switch (nla_type(a)) {
> case OVS_ACTION_ATTR_OUTPUT:
> prev_port = nla_get_u32(a);
> break;
> ...
> }
>
> 2. Iteration of of for loop for next action, lets say its is recirculate
>
> i. Output packet
>
> if (prev_port != -1) {
> do_output(dp, skb_clone(skb, GFP_ATOMIC), prev_port);
> prev_port = -1;
> }
>
> ii. Return due to recirculate
> switch (nla_type(a)) {
> ...
> case OVS_ACTION_ATTR_RECIRCULATE:
> return 1;
> }
>
>
> Am I missing something?
Sorry, you're right.
>> > @@ -901,6 +913,9 @@ static int validate_and_copy_actions__(const struct nlattr *attr,
>> > skip_copy = true;
>> > break;
>> >
>> > + case OVS_ACTION_ATTR_RECIRCULATE:
>> > + break;
>>
>> I think we might want to jump out the loop here to better model how
>> the actions are actually executed.
>
> Sure, perhaps something like this?
>
> diff --git a/datapath/datapath.c b/datapath/datapath.c
> index ab39dd7..721a52c 100644
> --- a/datapath/datapath.c
> +++ b/datapath/datapath.c
> @@ -914,7 +914,7 @@ static int validate_and_copy_actions__(const struct nlattr *attr,
> break;
>
> case OVS_ACTION_ATTR_RECIRCULATE:
> - break;
> + goto out;
>
> default:
> return -EINVAL;
> @@ -926,6 +926,7 @@ static int validate_and_copy_actions__(const struct nlattr *attr,
> }
> }
>
> +out:
> if (rem > 0)
> return -EINVAL;
Since this function is now both validating and copying I think this
will result in the recirculate action not being copied.
>> > diff --git a/ofproto/ofproto-dpif.c b/ofproto/ofproto-dpif.c
>> > index 47830c1..5129da1 100644
>> > --- a/ofproto/ofproto-dpif.c
>> > +++ b/ofproto/ofproto-dpif.c
>>
>> I'm still working on more detailed comments for this. However, I'm
>> concerned about whether the behavior for revalidation and stats is
>> correct.
>
> I am a little concerned about that too.
> Perhaps Ben could look over it?
To rephrase, there are problems in both of those areas. Validation in
particular I don't think handles resubmitted facets and I believe that
stats on rules will be the sum of all resubmitted passes.
Both of these will likely significantly affect the data structures, so
please look into this before we go further. In general, I'd also like
to see patches that are standalone without needing follow on patches
to fix known problems (for example, the recirculation ID patches or
MPLS GSO) unless there is a good reason.
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH 1/4] Add packet recirculation
[not found] ` <CAEP_g=-baZGttBowXKGdgKafxp9fkUrQ=44y562ZzfnOL-XaQg-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>
@ 2013-04-10 9:16 ` Simon Horman
2013-04-10 16:21 ` Jesse Gross
0 siblings, 1 reply; 14+ messages in thread
From: Simon Horman @ 2013-04-10 9:16 UTC (permalink / raw)
To: Jesse Gross
Cc: dev-yBygre7rU0TnMu66kgdUjQ@public.gmane.org, netdev, Ravi K,
Isaku Yamahata
On Tue, Apr 09, 2013 at 08:44:02AM -0700, Jesse Gross wrote:
> On Tue, Apr 9, 2013 at 12:50 AM, Simon Horman <horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org> wrote:
> > On Mon, Apr 08, 2013 at 06:46:29PM -0700, Jesse Gross wrote:
> >> On Sun, Apr 7, 2013 at 11:43 PM, Simon Horman <horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org> wrote:
> >> > diff --git a/datapath/actions.c b/datapath/actions.c
> >> > index e9634fe..7b0f022 100644
> >> > --- a/datapath/actions.c
> >> > +++ b/datapath/actions.c
> >> > @@ -617,6 +617,9 @@ static int do_execute_actions(struct datapath *dp, struct sk_buff *skb,
> >> > case OVS_ACTION_ATTR_SAMPLE:
> >> > err = sample(dp, skb, a);
> >> > break;
> >> > +
> >> > + case OVS_ACTION_ATTR_RECIRCULATE:
> >> > + return 1;
> >>
> >> I think that if we've had a previous output action with the port
> >> stored in prev_port then this will cause the packet to not actually be
> >> output.
> >
> > I'm not so sure.
> >
> > I see something like this occurring:
> >
> > 1. Iteration of for loop for output action
> >
> > switch (nla_type(a)) {
> > case OVS_ACTION_ATTR_OUTPUT:
> > prev_port = nla_get_u32(a);
> > break;
> > ...
> > }
> >
> > 2. Iteration of of for loop for next action, lets say its is recirculate
> >
> > i. Output packet
> >
> > if (prev_port != -1) {
> > do_output(dp, skb_clone(skb, GFP_ATOMIC), prev_port);
> > prev_port = -1;
> > }
> >
> > ii. Return due to recirculate
> > switch (nla_type(a)) {
> > ...
> > case OVS_ACTION_ATTR_RECIRCULATE:
> > return 1;
> > }
> >
> >
> > Am I missing something?
>
> Sorry, you're right.
>
> >> > @@ -901,6 +913,9 @@ static int validate_and_copy_actions__(const struct nlattr *attr,
> >> > skip_copy = true;
> >> > break;
> >> >
> >> > + case OVS_ACTION_ATTR_RECIRCULATE:
> >> > + break;
> >>
> >> I think we might want to jump out the loop here to better model how
> >> the actions are actually executed.
> >
> > Sure, perhaps something like this?
> >
> > diff --git a/datapath/datapath.c b/datapath/datapath.c
> > index ab39dd7..721a52c 100644
> > --- a/datapath/datapath.c
> > +++ b/datapath/datapath.c
> > @@ -914,7 +914,7 @@ static int validate_and_copy_actions__(const struct nlattr *attr,
> > break;
> >
> > case OVS_ACTION_ATTR_RECIRCULATE:
> > - break;
> > + goto out;
> >
> > default:
> > return -EINVAL;
> > @@ -926,6 +926,7 @@ static int validate_and_copy_actions__(const struct nlattr *attr,
> > }
> > }
> >
> > +out:
> > if (rem > 0)
> > return -EINVAL;
>
> Since this function is now both validating and copying I think this
> will result in the recirculate action not being copied.
Thanks, I'll look into that.
> >> > diff --git a/ofproto/ofproto-dpif.c b/ofproto/ofproto-dpif.c
> >> > index 47830c1..5129da1 100644
> >> > --- a/ofproto/ofproto-dpif.c
> >> > +++ b/ofproto/ofproto-dpif.c
> >>
> >> I'm still working on more detailed comments for this. However, I'm
> >> concerned about whether the behavior for revalidation and stats is
> >> correct.
> >
> > I am a little concerned about that too.
> > Perhaps Ben could look over it?
>
> To rephrase, there are problems in both of those areas. Validation in
> particular I don't think handles resubmitted facets and I believe that
> stats on rules will be the sum of all resubmitted passes.
Some questions:
By resubmitted do you mean recirculated?
What is the stats behaviour that you would like?
With regards to validation, I assume the area of concern
is around facet_revalidate(). I will look into that.
> Both of these will likely significantly affect the data structures, so
> please look into this before we go further.
Sure. I was not planning to push (much) further until this series
is reviewed properly.
> In general, I'd also like
> to see patches that are standalone without needing follow on patches
> to fix known problems (for example, the recirculation ID patches or
> MPLS GSO) unless there is a good reason.
Thanks, I understand. I'll try and structure my patches accordingly.
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH 1/4] Add packet recirculation
2013-04-10 9:16 ` Simon Horman
@ 2013-04-10 16:21 ` Jesse Gross
[not found] ` <CAEP_g=8vE9Ykdnro+Puw34ORRnPEMn1LMBxv+f7w_Dp-uty_iw-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>
0 siblings, 1 reply; 14+ messages in thread
From: Jesse Gross @ 2013-04-10 16:21 UTC (permalink / raw)
To: Simon Horman
Cc: dev@openvswitch.org, netdev, Ravi K, Isaku Yamahata, Ben Pfaff
On Wed, Apr 10, 2013 at 2:16 AM, Simon Horman <horms@verge.net.au> wrote:
> On Tue, Apr 09, 2013 at 08:44:02AM -0700, Jesse Gross wrote:
>> On Tue, Apr 9, 2013 at 12:50 AM, Simon Horman <horms@verge.net.au> wrote:
>> > On Mon, Apr 08, 2013 at 06:46:29PM -0700, Jesse Gross wrote:
>> >> On Sun, Apr 7, 2013 at 11:43 PM, Simon Horman <horms@verge.net.au> wrote:
>> >> > diff --git a/ofproto/ofproto-dpif.c b/ofproto/ofproto-dpif.c
>> >> > index 47830c1..5129da1 100644
>> >> > --- a/ofproto/ofproto-dpif.c
>> >> > +++ b/ofproto/ofproto-dpif.c
>> >>
>> >> I'm still working on more detailed comments for this. However, I'm
>> >> concerned about whether the behavior for revalidation and stats is
>> >> correct.
>> >
>> > I am a little concerned about that too.
>> > Perhaps Ben could look over it?
>>
>> To rephrase, there are problems in both of those areas. Validation in
>> particular I don't think handles resubmitted facets and I believe that
>> stats on rules will be the sum of all resubmitted passes.
>
> Some questions:
> By resubmitted do you mean recirculated?
Yes.
> What is the stats behaviour that you would like?
A given rule should have byte and packet counts equal to the number of
times it is matched (i.e. the first time) even if we have to decompose
it into multiple passes internally.
> With regards to validation, I assume the area of concern
> is around facet_revalidate(). I will look into that.
Yes.
>> Both of these will likely significantly affect the data structures, so
>> please look into this before we go further.
>
> Sure. I was not planning to push (much) further until this series
> is reviewed properly.
I'm planning on waiting on further reviews of this file until you've
had a chance to look into validation and stats since I think that may
change some of the data structures.
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH 1/4] Add packet recirculation
[not found] ` <CAEP_g=8vE9Ykdnro+Puw34ORRnPEMn1LMBxv+f7w_Dp-uty_iw-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>
@ 2013-04-11 0:14 ` Simon Horman
0 siblings, 0 replies; 14+ messages in thread
From: Simon Horman @ 2013-04-11 0:14 UTC (permalink / raw)
To: Jesse Gross
Cc: dev-yBygre7rU0TnMu66kgdUjQ@public.gmane.org, netdev, Ravi K,
Isaku Yamahata
On Wed, Apr 10, 2013 at 09:21:23AM -0700, Jesse Gross wrote:
> On Wed, Apr 10, 2013 at 2:16 AM, Simon Horman <horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org> wrote:
> > On Tue, Apr 09, 2013 at 08:44:02AM -0700, Jesse Gross wrote:
> >> On Tue, Apr 9, 2013 at 12:50 AM, Simon Horman <horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org> wrote:
> >> > On Mon, Apr 08, 2013 at 06:46:29PM -0700, Jesse Gross wrote:
> >> >> On Sun, Apr 7, 2013 at 11:43 PM, Simon Horman <horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org> wrote:
> >> >> > diff --git a/ofproto/ofproto-dpif.c b/ofproto/ofproto-dpif.c
> >> >> > index 47830c1..5129da1 100644
> >> >> > --- a/ofproto/ofproto-dpif.c
> >> >> > +++ b/ofproto/ofproto-dpif.c
> >> >>
> >> >> I'm still working on more detailed comments for this. However, I'm
> >> >> concerned about whether the behavior for revalidation and stats is
> >> >> correct.
> >> >
> >> > I am a little concerned about that too.
> >> > Perhaps Ben could look over it?
> >>
> >> To rephrase, there are problems in both of those areas. Validation in
> >> particular I don't think handles resubmitted facets and I believe that
> >> stats on rules will be the sum of all resubmitted passes.
> >
> > Some questions:
> > By resubmitted do you mean recirculated?
>
> Yes.
>
> > What is the stats behaviour that you would like?
>
> A given rule should have byte and packet counts equal to the number of
> times it is matched (i.e. the first time) even if we have to decompose
> it into multiple passes internally.
>
> > With regards to validation, I assume the area of concern
> > is around facet_revalidate(). I will look into that.
>
> Yes.
>
> >> Both of these will likely significantly affect the data structures, so
> >> please look into this before we go further.
> >
> > Sure. I was not planning to push (much) further until this series
> > is reviewed properly.
>
> I'm planning on waiting on further reviews of this file until you've
> had a chance to look into validation and stats since I think that may
> change some of the data structures.
Sure, I assumed as much.
I'll try and prepare and post a version with those issues, and the other
ones you raised elsewhere in your review, in the not to distant future.
^ permalink raw reply [flat|nested] 14+ messages in thread
end of thread, other threads:[~2013-04-11 0:14 UTC | newest]
Thread overview: 14+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2013-04-08 6:43 [PATCH v5 0/4] Add packet recirculation Simon Horman
2013-04-08 6:43 ` [PATCH 3/4] Allow recirculation without facets Simon Horman
[not found] ` <1365403431-18102-1-git-send-email-horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org>
2013-04-08 6:43 ` [PATCH 1/4] Add packet recirculation Simon Horman
2013-04-09 1:46 ` Jesse Gross
2013-04-09 7:50 ` Simon Horman
2013-04-09 15:44 ` Jesse Gross
[not found] ` <CAEP_g=-baZGttBowXKGdgKafxp9fkUrQ=44y562ZzfnOL-XaQg-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>
2013-04-10 9:16 ` Simon Horman
2013-04-10 16:21 ` Jesse Gross
[not found] ` <CAEP_g=8vE9Ykdnro+Puw34ORRnPEMn1LMBxv+f7w_Dp-uty_iw-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>
2013-04-11 0:14 ` Simon Horman
2013-04-08 6:43 ` [PATCH 2/4] Move execute_set_action to lib/odp-util.c Simon Horman
[not found] ` <1365403431-18102-3-git-send-email-horms-/R6kz+dDXgpPR4JQBCEnsQ@public.gmane.org>
2013-04-08 20:29 ` Jesse Gross
2013-04-09 3:11 ` Simon Horman
2013-04-09 15:17 ` Jesse Gross
2013-04-08 6:43 ` [PATCH 4/4] Avoid recirculation id collision Simon Horman
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).