[PATCH v2 06/23] migration/multifd: Separate SYNC request with normal jobs

peterx@redhat.com posted 23 patches 9 months, 4 weeks ago
Maintainers: Peter Xu <peterx@redhat.com>, Fabiano Rosas <farosas@suse.de>
[PATCH v2 06/23] migration/multifd: Separate SYNC request with normal jobs
Posted by peterx@redhat.com 9 months, 4 weeks ago
From: Peter Xu <peterx@redhat.com>

Multifd provide a threaded model for processing jobs.  On sender side,
there can be two kinds of job: (1) a list of pages to send, or (2) a sync
request.

The sync request is a very special kind of job.  It never contains a page
array, but only a multifd packet telling the dest side to synchronize with
sent pages.

Before this patch, both requests use the pending_job field, no matter what
the request is, it will boost pending_job, while multifd sender thread will
decrement it after it finishes one job.

However this should be racy, because SYNC is special in that it needs to
set p->flags with MULTIFD_FLAG_SYNC, showing that this is a sync request.
Consider a sequence of operations where:

  - migration thread enqueue a job to send some pages, pending_job++ (0->1)

  - [...before the selected multifd sender thread wakes up...]

  - migration thread enqueue another job to sync, pending_job++ (1->2),
    setup p->flags=MULTIFD_FLAG_SYNC

  - multifd sender thread wakes up, found pending_job==2
    - send the 1st packet with MULTIFD_FLAG_SYNC and list of pages
    - send the 2nd packet with flags==0 and no pages

This is not expected, because MULTIFD_FLAG_SYNC should hopefully be done
after all the pages are received.  Meanwhile, the 2nd packet will be
completely useless, which contains zero information.

I didn't verify above, but I think this issue is still benign in that at
least on the recv side we always receive pages before handling
MULTIFD_FLAG_SYNC.  However that's not always guaranteed and just tricky.

One other reason I want to separate it is using p->flags to communicate
between the two threads is also not clearly defined, it's very hard to read
and understand why accessing p->flags is always safe; see the current impl
of multifd_send_thread() where we tried to cache only p->flags.  It doesn't
need to be that complicated.

This patch introduces pending_sync, a separate flag just to show that the
requester needs a sync.  Alongside, we remove the tricky caching of
p->flags now because after this patch p->flags should only be used by
multifd sender thread now, which will be crystal clear.  So it is always
thread safe to access p->flags.

With that, we can also safely convert the pending_job into a boolean,
because we don't support >1 pending jobs anyway.

Always use atomic ops to access both flags to make sure no cache effect.
When at it, drop the initial setting of "pending_job = 0" because it's
always allocated using g_new0().

Signed-off-by: Peter Xu <peterx@redhat.com>
---
 migration/multifd.h | 13 +++++++++++--
 migration/multifd.c | 39 +++++++++++++++++++++++++--------------
 2 files changed, 36 insertions(+), 16 deletions(-)

diff --git a/migration/multifd.h b/migration/multifd.h
index 3920bdbcf1..08f26ef3fe 100644
--- a/migration/multifd.h
+++ b/migration/multifd.h
@@ -99,8 +99,17 @@ typedef struct {
     uint32_t flags;
     /* global number of generated multifd packets */
     uint64_t packet_num;
-    /* thread has work to do */
-    int pending_job;
+    /*
+     * The sender thread has work to do if either of below boolean is set.
+     *
+     * @pending_job:  a job is pending
+     * @pending_sync: a sync request is pending
+     *
+     * For both of these fields, they're only set by the requesters, and
+     * cleared by the multifd sender threads.
+     */
+    bool pending_job;
+    bool pending_sync;
     /* array of pages to sent.
      * The owner of 'pages' depends of 'pending_job' value:
      * pending_job == 0 -> migration_thread can use it.
diff --git a/migration/multifd.c b/migration/multifd.c
index 8bb1fd95cf..ea25bbe6bd 100644
--- a/migration/multifd.c
+++ b/migration/multifd.c
@@ -442,8 +442,8 @@ static int multifd_send_pages(void)
         }
         p = &multifd_send_state->params[i];
         qemu_mutex_lock(&p->mutex);
-        if (!p->pending_job) {
-            p->pending_job++;
+        if (qatomic_read(&p->pending_job) == false) {
+            qatomic_set(&p->pending_job, true);
             next_channel = (i + 1) % migrate_multifd_channels();
             break;
         }
@@ -631,8 +631,12 @@ int multifd_send_sync_main(void)
 
         qemu_mutex_lock(&p->mutex);
         p->packet_num = multifd_send_state->packet_num++;
-        p->flags |= MULTIFD_FLAG_SYNC;
-        p->pending_job++;
+        /*
+         * We should be the only user so far, so not possible to be set by
+         * others concurrently.
+         */
+        assert(qatomic_read(&p->pending_sync) == false);
+        qatomic_set(&p->pending_sync, true);
         qemu_mutex_unlock(&p->mutex);
         qemu_sem_post(&p->sem);
     }
@@ -685,10 +689,9 @@ static void *multifd_send_thread(void *opaque)
         }
         qemu_mutex_lock(&p->mutex);
 
-        if (p->pending_job) {
+        if (qatomic_read(&p->pending_job)) {
             uint64_t packet_num = p->packet_num;
             MultiFDPages_t *pages = p->pages;
-            uint32_t flags;
 
             if (use_zero_copy_send) {
                 p->iovs_num = 0;
@@ -704,13 +707,11 @@ static void *multifd_send_thread(void *opaque)
                 }
             }
             multifd_send_fill_packet(p);
-            flags = p->flags;
-            p->flags = 0;
             p->num_packets++;
             p->total_normal_pages += pages->num;
             qemu_mutex_unlock(&p->mutex);
 
-            trace_multifd_send(p->id, packet_num, pages->num, flags,
+            trace_multifd_send(p->id, packet_num, pages->num, p->flags,
                                p->next_packet_size);
 
             if (use_zero_copy_send) {
@@ -738,12 +739,23 @@ static void *multifd_send_thread(void *opaque)
             multifd_pages_reset(p->pages);
             p->next_packet_size = 0;
             qemu_mutex_lock(&p->mutex);
-            p->pending_job--;
+            qatomic_set(&p->pending_job, false);
             qemu_mutex_unlock(&p->mutex);
-
-            if (flags & MULTIFD_FLAG_SYNC) {
-                qemu_sem_post(&p->sem_sync);
+        } else if (qatomic_read(&p->pending_sync)) {
+            p->flags = MULTIFD_FLAG_SYNC;
+            multifd_send_fill_packet(p);
+            ret = qio_channel_write_all(p->c, (void *)p->packet,
+                                        p->packet_len, &local_err);
+            if (ret != 0) {
+                qemu_mutex_unlock(&p->mutex);
+                break;
             }
+            /* p->next_packet_size will always be zero for a SYNC packet */
+            stat64_add(&mig_stats.multifd_bytes, p->packet_len);
+            p->flags = 0;
+            qatomic_set(&p->pending_sync, false);
+            qemu_mutex_unlock(&p->mutex);
+            qemu_sem_post(&p->sem_sync);
         } else {
             qemu_mutex_unlock(&p->mutex);
             /* sometimes there are spurious wakeups */
@@ -907,7 +919,6 @@ int multifd_save_setup(Error **errp)
         qemu_mutex_init(&p->mutex);
         qemu_sem_init(&p->sem, 0);
         qemu_sem_init(&p->sem_sync, 0);
-        p->pending_job = 0;
         p->id = i;
         p->pages = multifd_pages_init(page_count);
         p->packet_len = sizeof(MultiFDPacket_t)
-- 
2.43.0
Re: [PATCH v2 06/23] migration/multifd: Separate SYNC request with normal jobs
Posted by Fabiano Rosas 9 months, 3 weeks ago
peterx@redhat.com writes:

> From: Peter Xu <peterx@redhat.com>
>
> Multifd provide a threaded model for processing jobs.  On sender side,
> there can be two kinds of job: (1) a list of pages to send, or (2) a sync
> request.
>
> The sync request is a very special kind of job.  It never contains a page
> array, but only a multifd packet telling the dest side to synchronize with
> sent pages.
>
> Before this patch, both requests use the pending_job field, no matter what
> the request is, it will boost pending_job, while multifd sender thread will
> decrement it after it finishes one job.
>
> However this should be racy, because SYNC is special in that it needs to
> set p->flags with MULTIFD_FLAG_SYNC, showing that this is a sync request.
> Consider a sequence of operations where:
>
>   - migration thread enqueue a job to send some pages, pending_job++ (0->1)
>
>   - [...before the selected multifd sender thread wakes up...]
>
>   - migration thread enqueue another job to sync, pending_job++ (1->2),
>     setup p->flags=MULTIFD_FLAG_SYNC
>
>   - multifd sender thread wakes up, found pending_job==2
>     - send the 1st packet with MULTIFD_FLAG_SYNC and list of pages
>     - send the 2nd packet with flags==0 and no pages
>
> This is not expected, because MULTIFD_FLAG_SYNC should hopefully be done
> after all the pages are received.  Meanwhile, the 2nd packet will be
> completely useless, which contains zero information.
>
> I didn't verify above, but I think this issue is still benign in that at
> least on the recv side we always receive pages before handling
> MULTIFD_FLAG_SYNC.  However that's not always guaranteed and just tricky.
>
> One other reason I want to separate it is using p->flags to communicate
> between the two threads is also not clearly defined, it's very hard to read
> and understand why accessing p->flags is always safe; see the current impl
> of multifd_send_thread() where we tried to cache only p->flags.  It doesn't
> need to be that complicated.
>
> This patch introduces pending_sync, a separate flag just to show that the
> requester needs a sync.  Alongside, we remove the tricky caching of
> p->flags now because after this patch p->flags should only be used by
> multifd sender thread now, which will be crystal clear.  So it is always
> thread safe to access p->flags.
>
> With that, we can also safely convert the pending_job into a boolean,
> because we don't support >1 pending jobs anyway.
>
> Always use atomic ops to access both flags to make sure no cache effect.
> When at it, drop the initial setting of "pending_job = 0" because it's
> always allocated using g_new0().
>
> Signed-off-by: Peter Xu <peterx@redhat.com>

Reviewed-by: Fabiano Rosas <farosas@suse.de>