[pve-devel] [PATCH qemu 3/4] savevm-async: flush IOThread-drives async before entering blocking part

[Stefan Reiter]

By flushing all drives where its possible to so before entering the
blocking part (where the VM is stopped), we can reduce the time spent in
said part for every disk that has an IOThread (other drives cannot be
flushed async anyway).

Signed-off-by: Stefan Reiter <s.reiter at proxmox.com>
---

It's a bit hard to get benchmark numbers here, since snapshot timings tend to
vary greatly for me. But with some very unscientific testing using patch 4 and
some gut feeling, on a VM running 'stress-ng -d 4' I'd say it shaves off about
2-3 seconds of VM downtime on average. I've never seen it produce worse results
than without it though, even for idle VMs.

 savevm-async.c | 23 +++++++++++++++++++++++
 1 file changed, 23 insertions(+)

diff --git a/savevm-async.c b/savevm-async.c
index 2894c94233..4ce83a0691 100644
--- a/savevm-async.c
+++ b/savevm-async.c
@@ -253,6 +253,8 @@ static void coroutine_fn process_savevm_co(void *opaque)
 {
     int ret;
     int64_t maxlen;
+    BdrvNextIterator it;
+    BlockDriverState *bs = NULL;
 
     ret = qemu_file_get_error(snap_state.file);
     if (ret < 0) {
@@ -288,6 +290,27 @@ static void coroutine_fn process_savevm_co(void *opaque)
         }
     }
 
+    /* If a drive runs in an IOThread we can flush it async, and only
+     * need to sync-flush whatever IO happens between now and
+     * vm_stop_force_state. bdrv_next can only be called from main AioContext,
+     * so move there now and after every flush.
+     */
+    aio_co_reschedule_self(qemu_get_aio_context());
+    for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
+        /* target has BDRV_O_NO_FLUSH, no sense calling bdrv_flush on it */
+        if (bs == blk_bs(snap_state.target)) {
+            continue;
+        }
+
+        AioContext *bs_ctx = bdrv_get_aio_context(bs);
+        if (bs_ctx != qemu_get_aio_context()) {
+            DPRINTF("savevm: async flushing drive %s\n", bs->filename);
+            aio_co_reschedule_self(bs_ctx);
+            bdrv_flush(bs);
+            aio_co_reschedule_self(qemu_get_aio_context());
+        }
+    }
+
     qemu_bh_schedule(snap_state.finalize_bh);
 }
 
-- 
2.20.1