[ovs-dev,v3,9/9] dpif-netdev: do hw flow offload in another thread

Message ID 1506404199-23579-10-git-send-email-yliu@fridaylinux.org
State Changes Requested
Headers show
Series
  • OVS-DPDK flow offload with rte_flow
Related show

Commit Message

Yuanhan Liu Sept. 26, 2017, 5:36 a.m.
Currently, the major trigger for hw flow offload is at upcall handling,
which is actually in the datapath. Moreover, the hw offload installation
and modification is not that lightweight. Meaning, if there are so many
flows being added or modified frequently, it could stall the datapath,
which could result to packet loss.

To diminish that, all those flow operations will be recorded and appended
to a list. A thread is then introduced to process this list (to do the
real flow offloading put/del operations). This could leave the datapath
as lightweight as possible.

Signed-off-by: Yuanhan Liu <yliu@fridaylinux.org>
---
 lib/dpif-netdev.c | 301 ++++++++++++++++++++++++++++++++++++++++++++----------
 1 file changed, 249 insertions(+), 52 deletions(-)

Comments

Chandran, Sugesh Oct. 2, 2017, 4:25 p.m. | #1
Regards
_Sugesh


> -----Original Message-----
> From: Yuanhan Liu [mailto:yliu@fridaylinux.org]
> Sent: Tuesday, September 26, 2017 6:37 AM
> To: dev@openvswitch.org
> Cc: Finn Christensen <fc@napatech.com>; Darrell Ball <dball@vmware.com>;
> Chandran, Sugesh <sugesh.chandran@intel.com>; Simon Horman
> <simon.horman@netronome.com>; Yuanhan Liu <yliu@fridaylinux.org>
> Subject: [PATCH v3 9/9] dpif-netdev: do hw flow offload in another thread
> 
> Currently, the major trigger for hw flow offload is at upcall handling, which is
> actually in the datapath. Moreover, the hw offload installation and modification
> is not that lightweight. Meaning, if there are so many flows being added or
> modified frequently, it could stall the datapath, which could result to packet
> loss.
> 
> To diminish that, all those flow operations will be recorded and appended to a
> list. A thread is then introduced to process this list (to do the real flow offloading
> put/del operations). This could leave the datapath as lightweight as possible.
[Sugesh] This is really very useful to reduce the flow install overhead on PMD threads.

> 
> Signed-off-by: Yuanhan Liu <yliu@fridaylinux.org>
> ---
>  lib/dpif-netdev.c | 301 ++++++++++++++++++++++++++++++++++++++++++++---
> -------
>  1 file changed, 249 insertions(+), 52 deletions(-)
> 
> diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c index 13fd012..ef5c2e9 100644
> --- a/lib/dpif-netdev.c
> +++ b/lib/dpif-netdev.c
> @@ -1949,6 +1949,223 @@ dp_netdev_pmd_find_flow_by_mark(const
> uint32_t mark)
>      return NULL;
>  }

[Snip] 
> 
> +                          offload->actions, offload->actions_len,
> +                          &offload->ufid, &info, NULL);
> +    if (ret) {
> +        if (create) {
> +            dp_netdev_remove_flow_mark_map(info.flow_mark);
> +        }
> +        return ret;
> +    }
> +
> +    ovs_mutex_lock(&offload->pmd->flow_mutex);
> +    if (create) {
> +        flow->has_mark = true;
> +        flow->mark = info.flow_mark;
> +        if (!flow->dead) {
[Sugesh] Just to confirm, so the assumption here is the flow delete from
revalidator will set the flow as dead and its still keep the structure as its referenced
here. There wont be a situation where flow is deleted and this strcture may freed up.

> +            /*
> +             * A flow could have been dead after we regain the lock,
> +             * while the flow has offloaded to the netdev. When that
> +             * happens, there should be an offload item in the offload
> +             * list for the flow removal. To make sure the flow will

[Snip] 
> --
> 2.7.4

Patch

diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
index 13fd012..ef5c2e9 100644
--- a/lib/dpif-netdev.c
+++ b/lib/dpif-netdev.c
@@ -1949,6 +1949,223 @@  dp_netdev_pmd_find_flow_by_mark(const uint32_t mark)
     return NULL;
 }
 
+struct dp_flow_offload_item {
+    struct dp_netdev_pmd_thread *pmd;
+    odp_port_t in_port;
+    struct dp_netdev_flow *flow;
+    ovs_u128 ufid;
+    int op;
+    struct match match;
+    struct nlattr *actions;
+    size_t actions_len;
+    int rxq;
+    uint32_t flow_mark;
+
+    struct ovs_list node;
+};
+
+struct dp_flow_offload {
+    struct ovs_mutex mutex;
+    struct ovs_list list;
+    pthread_cond_t cond;
+};
+
+static struct dp_flow_offload dp_flow_offload = {
+    .mutex = OVS_MUTEX_INITIALIZER,
+    .list  = OVS_LIST_INITIALIZER(&dp_flow_offload.list),
+};
+
+static struct ovsthread_once offload_thread_once
+    = OVSTHREAD_ONCE_INITIALIZER;
+
+enum {
+    DP_NETDEV_FLOW_OFFLOAD_OP_ADD,
+    DP_NETDEV_FLOW_OFFLOAD_OP_MOD,
+    DP_NETDEV_FLOW_OFFLOAD_OP_DEL,
+};
+
+static struct dp_flow_offload_item *
+dp_netdev_alloc_flow_offload(struct dp_netdev_pmd_thread *pmd,
+                                  struct dp_netdev_flow *flow,
+                                  odp_port_t in_port, const ovs_u128 *ufid,
+                                  int op)
+{
+    struct dp_flow_offload_item *offload;
+
+    offload = xzalloc(sizeof(*offload));
+    offload->pmd = pmd;
+    offload->flow = flow;
+    offload->in_port = in_port;
+    offload->ufid = *ufid;
+    offload->op = op;
+
+    ovs_refcount_ref(&pmd->ref_cnt);
+    ovs_refcount_ref(&flow->ref_cnt);
+
+    return offload;
+}
+
+static void
+dp_netdev_free_flow_offload(struct dp_flow_offload_item *offload)
+{
+    ovs_refcount_unref(&offload->pmd->ref_cnt);
+    ovs_refcount_unref(&offload->flow->ref_cnt);
+
+    free(offload->actions);
+    free(offload);
+}
+
+static void
+dp_netdev_append_flow_offload(struct dp_flow_offload_item *offload)
+{
+    ovs_mutex_lock(&dp_flow_offload.mutex);
+    ovs_list_push_back(&dp_flow_offload.list, &offload->node);
+    ovs_mutex_unlock(&dp_flow_offload.mutex);
+
+    pthread_cond_signal(&dp_flow_offload.cond);
+}
+
+static int
+dp_netdev_flow_offload_del(struct dp_flow_offload_item *offload)
+{
+    struct dp_netdev_flow *flow = offload->flow;
+    struct dp_netdev_port *port;
+    int ret = -1;
+
+    ovs_mutex_lock(&offload->pmd->flow_mutex);
+    port = dp_netdev_lookup_port(offload->pmd->dp, offload->in_port);
+    if (flow->has_mark && port) {
+        ret = netdev_flow_del(port->netdev, &offload->ufid, NULL);
+    }
+
+    dp_netdev_remove_flow_mark_map(flow->mark);
+    ovsrcu_quiesce_start();
+    flow->has_mark = false;
+    ovs_mutex_unlock(&offload->pmd->flow_mutex);
+
+    return ret;
+}
+
+static int
+dp_netdev_flow_offload_put(struct dp_flow_offload_item *offload)
+{
+    struct dp_netdev_port *port;
+    struct dp_netdev_flow *flow = offload->flow;
+    bool create = offload->op == DP_NETDEV_FLOW_OFFLOAD_OP_ADD;
+    struct offload_info info;
+    int ret;
+
+    port = dp_netdev_lookup_port(offload->pmd->dp, offload->in_port);
+    if (!port) {
+        return -1;
+    }
+
+    if (create) {
+        if (!dp_netdev_alloc_flow_mark(&info.flow_mark)) {
+            VLOG_ERR("failed to allocate flow mark!\n");
+            return -1;
+        }
+        offload->flow_mark = info.flow_mark;
+    } else {
+        info.flow_mark = offload->flow_mark;
+    }
+    info.rxq = offload->rxq;
+
+    ret = netdev_flow_put(port->netdev, &offload->match,
+                          offload->actions, offload->actions_len,
+                          &offload->ufid, &info, NULL);
+    if (ret) {
+        if (create) {
+            dp_netdev_remove_flow_mark_map(info.flow_mark);
+        }
+        return ret;
+    }
+
+    ovs_mutex_lock(&offload->pmd->flow_mutex);
+    if (create) {
+        flow->has_mark = true;
+        flow->mark = info.flow_mark;
+        if (!flow->dead) {
+            /*
+             * A flow could have been dead after we regain the lock,
+             * while the flow has offloaded to the netdev. When that
+             * happens, there should be an offload item in the offload
+             * list for the flow removal. To make sure the flow will
+             * be deleted successfully later, above 2 fields (has_mark
+             * and mark) have to be set properly.
+             */
+            dp_netdev_install_flow_mark_map(flow->mark, flow);
+            ovsrcu_quiesce_start();
+        }
+    }
+    ovs_mutex_unlock(&offload->pmd->flow_mutex);
+
+    return 0;
+}
+
+static void *
+dp_netdev_flow_offload_main(void *data OVS_UNUSED)
+{
+    struct dp_flow_offload_item *offload;
+    struct ovs_list *list;
+    const char *op;
+    int ret;
+
+    for (;;) {
+        ovs_mutex_lock(&dp_flow_offload.mutex);
+        if (ovs_list_is_empty(&dp_flow_offload.list)) {
+            pthread_cond_wait(&dp_flow_offload.cond,
+                              &dp_flow_offload.mutex.lock);
+        }
+        list = ovs_list_pop_front(&dp_flow_offload.list);
+        offload = CONTAINER_OF(list, struct dp_flow_offload_item, node);
+        ovs_mutex_unlock(&dp_flow_offload.mutex);
+
+        switch (offload->op) {
+        case DP_NETDEV_FLOW_OFFLOAD_OP_ADD:
+            op = "install";
+            ret = dp_netdev_flow_offload_put(offload);
+            break;
+        case DP_NETDEV_FLOW_OFFLOAD_OP_MOD:
+            op = "modify";
+            ret = dp_netdev_flow_offload_put(offload);
+            break;
+        case DP_NETDEV_FLOW_OFFLOAD_OP_DEL:
+            op = "delete";
+            ret = dp_netdev_flow_offload_del(offload);
+            break;
+        default:
+            OVS_NOT_REACHED();
+        }
+
+        VLOG_INFO("%s to %s netdev flow with mark %u\n",
+                  ret == 0 ? "succeed" : "failed",
+                  op, offload->flow_mark);
+        dp_netdev_free_flow_offload(offload);
+    }
+
+    return NULL;
+}
+
+static void
+queue_netdev_flow_del(struct dp_netdev_pmd_thread *pmd,
+                      struct dp_netdev_flow *flow,
+                      odp_port_t in_port, const ovs_u128 *ufid,
+                      int op)
+{
+    struct dp_flow_offload_item *offload;
+
+    if (ovsthread_once_start(&offload_thread_once)) {
+        pthread_cond_init(&dp_flow_offload.cond, NULL);
+        ovs_thread_create("dp_netdev_flow_offload",
+                          dp_netdev_flow_offload_main, NULL);
+        ovsthread_once_done(&offload_thread_once);
+    }
+
+    offload = dp_netdev_alloc_flow_offload(pmd, flow, in_port, ufid, op);
+    offload->flow_mark = flow->mark;
+    dp_netdev_append_flow_offload(offload);
+}
 
 static void
 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
@@ -1963,16 +2180,8 @@  dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
     ovs_assert(cls != NULL);
     dpcls_remove(cls, &flow->cr);
     cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
-    if (flow->has_mark) {
-        struct dp_netdev_port *port;
-
-        port = dp_netdev_lookup_port(pmd->dp, in_port);
-        if (port) {
-            netdev_flow_del(port->netdev, &flow->ufid, NULL);
-        }
-        dp_netdev_remove_flow_mark_map(flow->mark);
-        flow->has_mark = false;
-    }
+    queue_netdev_flow_del(pmd, flow, in_port, &flow->ufid,
+                          DP_NETDEV_FLOW_OFFLOAD_OP_DEL);
     flow->dead = true;
 
     dp_netdev_flow_unref(flow);
@@ -2553,53 +2762,41 @@  out:
 }
 
 static void
-try_netdev_flow_put(struct dp_netdev_pmd_thread *pmd, odp_port_t in_port,
-                    struct dp_netdev_flow *flow, struct match *match,
-                    const ovs_u128 *ufid, const struct nlattr *actions,
-                    size_t actions_len, int rxq)
+queue_netdev_flow_put(struct dp_netdev_pmd_thread *pmd, odp_port_t in_port,
+                      struct dp_netdev_flow *flow, struct match *match,
+                      const ovs_u128 *ufid, const struct nlattr *actions,
+                      size_t actions_len, int rxq)
 {
-    struct offload_info info;
-    struct dp_netdev_port *port;
-    bool modification = flow->has_mark;
-    const char *op = modification ? "modify" : "install";
-    int ret;
+    struct dp_flow_offload_item *offload;
+    int op;
 
-    port = dp_netdev_lookup_port(pmd->dp, in_port);
-    if (!port) {
+    if (!netdev_is_flow_api_enabled()) {
         return;
     }
-    info.rxq = rxq;
-
-    if (modification) {
-        info.flow_mark = flow->mark;
-    } else {
-        if (!netdev_is_flow_api_enabled()) {
-            return;
-        }
-
-        if (!dp_netdev_alloc_flow_mark(&info.flow_mark)) {
-            VLOG_ERR("failed to allocate flow mark!\n");
-            return;
-        }
-    }
 
-    ret = netdev_flow_put(port->netdev, match,
-                          CONST_CAST(struct nlattr *, actions),
-                          actions_len, ufid, &info, NULL);
-    if (ret) {
-        VLOG_ERR("failed to %s netdev flow with mark %u\n",
-                 op, info.flow_mark);
-        return;
+    if (ovsthread_once_start(&offload_thread_once)) {
+        pthread_cond_init(&dp_flow_offload.cond, NULL);
+        ovs_thread_create("dp_netdev_flow_offload",
+                          dp_netdev_flow_offload_main, NULL);
+        ovsthread_once_done(&offload_thread_once);
     }
 
-    if (!modification) {
-        flow->has_mark = true;
-        flow->mark = info.flow_mark;
-        dp_netdev_install_flow_mark_map(info.flow_mark, flow);
+    if (flow->has_mark) {
+        op = DP_NETDEV_FLOW_OFFLOAD_OP_MOD;
+    } else {
+        op = DP_NETDEV_FLOW_OFFLOAD_OP_ADD;
+    }
+    offload = dp_netdev_alloc_flow_offload(pmd, flow, in_port, ufid, op);
+    offload->match = *match;
+    offload->rxq = rxq;
+    offload->actions = xmalloc(actions_len);
+    memcpy(offload->actions, actions, actions_len);
+    offload->actions_len = actions_len;
+    if (flow->has_mark) {
+        offload->flow_mark = flow->mark;
     }
 
-    VLOG_INFO("succeed to %s netdev flow with mark %u\n",
-              op, flow->mark);
+    dp_netdev_append_flow_offload(offload);
 }
 
 static struct dp_netdev_flow *
@@ -2654,8 +2851,8 @@  dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
     cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
                 dp_netdev_flow_hash(&flow->ufid));
 
-    try_netdev_flow_put(pmd, in_port, flow, match, ufid,
-                        actions, actions_len, rxq);
+    queue_netdev_flow_put(pmd, in_port, flow, match, ufid,
+                          actions, actions_len, rxq);
 
     if (OVS_UNLIKELY(!VLOG_DROP_DBG((&upcall_rl)))) {
         struct ds ds = DS_EMPTY_INITIALIZER;
@@ -2745,8 +2942,8 @@  flow_put_on_pmd(struct dp_netdev_pmd_thread *pmd,
             old_actions = dp_netdev_flow_get_actions(netdev_flow);
             ovsrcu_set(&netdev_flow->actions, new_actions);
 
-            try_netdev_flow_put(pmd, in_port, netdev_flow, match, ufid,
-                                put->actions, put->actions_len, -1);
+            queue_netdev_flow_put(pmd, in_port, netdev_flow, match, ufid,
+                                  put->actions, put->actions_len, -1);
 
             if (stats) {
                 get_dpif_flow_stats(netdev_flow, stats);