On 30 Apr 2019, at 14:17, David Marchand wrote:
pmd reloads are currently serialised in each steps calling
reload_affected_pmds.
Any pmd processing packets, waiting on a mutex etc... will make other
pmd threads wait for a delay that can be undeterministic when syscalls
adds up.
Switch to a little busy loop on the control thread using an atomic
count.
Signed-off-by: David Marchand <[email protected]>
---
lib/dpif-netdev.c | 43 ++++++++++++++++++++++++++++++-------------
1 file changed, 30 insertions(+), 13 deletions(-)
diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
index 30774ed..b2b21fd 100644
--- a/lib/dpif-netdev.c
+++ b/lib/dpif-netdev.c
@@ -335,6 +335,9 @@ struct dp_netdev {
/* The time that a packet can wait in output batch for sending.
*/
atomic_uint32_t tx_flush_interval;
+ /* Count of pmds currently reloading */
+ atomic_count reloading_pmds;
+
/* Meters. */
struct ovs_mutex meter_locks[N_METER_LOCKS];
struct dp_meter *meters[MAX_METERS]; /* Meter bands. */
@@ -647,9 +650,6 @@ struct dp_netdev_pmd_thread {
struct ovs_refcount ref_cnt; /* Every reference must be
refcount'ed. */
struct cmap_node node; /* In 'dp->poll_threads'. */
- pthread_cond_t cond; /* For synchronizing pmd thread
reload. */
- struct ovs_mutex cond_mutex; /* Mutex for condition variable.
*/
-
/* Per thread exact-match cache. Note, the instance for cpu core
* NON_PMD_CORE_ID can be accessed by multiple threads, and
thusly
* need to be protected by 'non_pmd_mutex'. Every other instance
@@ -1525,6 +1525,8 @@ create_dp_netdev(const char *name, const struct
dpif_class *class,
atomic_init(&dp->emc_insert_min, DEFAULT_EM_FLOW_INSERT_MIN);
atomic_init(&dp->tx_flush_interval, DEFAULT_TX_FLUSH_INTERVAL);
+ atomic_count_init(&dp->reloading_pmds, 0);
+
cmap_init(&dp->poll_threads);
dp->pmd_rxq_assign_cyc = true;
@@ -1754,11 +1756,8 @@ dp_netdev_reload_pmd__(struct
dp_netdev_pmd_thread *pmd)
return;
}
- ovs_mutex_lock(&pmd->cond_mutex);
seq_change(pmd->reload_seq);
atomic_store_relaxed(&pmd->reload, true);
- ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
- ovs_mutex_unlock(&pmd->cond_mutex);
}
static uint32_t
@@ -4641,9 +4640,27 @@ rxq_scheduling(struct dp_netdev *dp, bool
pinned) OVS_REQUIRES(dp->port_mutex)
}
static void
+wait_reloading_pmds(struct dp_netdev *dp)
+{
+ while (atomic_count_get(&dp->reloading_pmds) != 0) {
+ }
+}
+
+static void
reload_affected_pmds(struct dp_netdev *dp)
{
struct dp_netdev_pmd_thread *pmd;
+ unsigned int pmd_count = 0;
+
+ CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
+ if (pmd->core_id == NON_PMD_CORE_ID) {
+ continue;
+ }
+ if (pmd->need_reload) {
+ pmd_count++;
+ }
+ }
+ atomic_count_set(&dp->reloading_pmds, pmd_count);
CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
if (pmd->need_reload) {
@@ -4652,6 +4669,10 @@ reload_affected_pmds(struct dp_netdev *dp)
The above atomic_count_set() is a relaxed set, so is the
atomic_store_relaxed(&pmd->reload, true) in dp_netdev_reload_pmd__(pmd).
Which could lead to the PMDs decreasing the dp->reloading_pmd before
it’s set. Guess if the correct memory_order is selected for
pmd->reload as suggested by Ilya in patch 1/5 it should work.
pmd->need_reload = false;
}
}
+
+ if (pmd_count != 0) {
+ wait_reloading_pmds(dp);
+ }
}
static void
@@ -5813,11 +5834,9 @@ dpif_netdev_enable_upcall(struct dpif *dpif)
static void
dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
{
- ovs_mutex_lock(&pmd->cond_mutex);
atomic_store_relaxed(&pmd->reload, false);
pmd->last_reload_seq = seq_read(pmd->reload_seq);
- xpthread_cond_signal(&pmd->cond);
- ovs_mutex_unlock(&pmd->cond_mutex);
+ atomic_count_dec(&pmd->dp->reloading_pmds);
}
/* Finds and refs the dp_netdev_pmd_thread on core 'core_id'.
Returns
@@ -5902,8 +5921,6 @@ dp_netdev_configure_pmd(struct
dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
pmd->reload_seq = seq_create();
pmd->last_reload_seq = seq_read(pmd->reload_seq);
atomic_init(&pmd->reload, false);
- xpthread_cond_init(&pmd->cond, NULL);
- ovs_mutex_init(&pmd->cond_mutex);
ovs_mutex_init(&pmd->flow_mutex);
ovs_mutex_init(&pmd->port_mutex);
cmap_init(&pmd->flow_table);
@@ -5946,8 +5963,6 @@ dp_netdev_destroy_pmd(struct
dp_netdev_pmd_thread *pmd)
cmap_destroy(&pmd->flow_table);
ovs_mutex_destroy(&pmd->flow_mutex);
seq_destroy(pmd->reload_seq);
- xpthread_cond_destroy(&pmd->cond);
- ovs_mutex_destroy(&pmd->cond_mutex);
ovs_mutex_destroy(&pmd->port_mutex);
free(pmd);
}
@@ -5967,7 +5982,9 @@ dp_netdev_del_pmd(struct dp_netdev *dp, struct
dp_netdev_pmd_thread *pmd)
ovs_mutex_unlock(&dp->non_pmd_mutex);
} else {
atomic_store_relaxed(&pmd->exit, true);
+ atomic_count_set(&dp->reloading_pmds, 1);
dp_netdev_reload_pmd__(pmd);
+ wait_reloading_pmds(dp);
xpthread_join(pmd->thread, NULL);
}
--
1.8.3.1
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
