[Qemu-devel] [PATCH 0/3] block: three optimizations

[Qemu-devel] [PATCH 0/3] block: three optimizations

[Paolo Bonzini]

These are three unrelated micro-optimizations of the block layer.
The first is only visible on non-dataplane operation, the others
are generic.

Paolo Bonzini (3):
  block: mark AioContext as recursive
  block: do not allocate an iovec per read of a growable/zero_after_eof BDS
  block: replace g_new0 with g_new for bottom half allocation.

 async.c | 11 +++++++----
 block.c | 12 +++++-------
 2 files changed, 12 insertions(+), 11 deletions(-)

-- 
2.1.0

[Qemu-devel] [PATCH 1/3] block: mark AioContext as recursive

[Paolo Bonzini]

AioContext can be accessed recursively, in fact that's what we do with
aio_poll.  Marking the GSource as recursive avoids that GLib blocks it
and unblocks it around every call to aio_dispatch, which is a pretty
expensive operation.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 async.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/async.c b/async.c
index 3939b79..572f239 100644
--- a/async.c
+++ b/async.c
@@ -300,6 +300,7 @@ AioContext *aio_context_new(Error **errp)
         error_setg_errno(errp, -ret, "Failed to initialize event notifier");
         return NULL;
     }
+    g_source_set_can_recurse(&ctx->source, true);
     aio_set_event_notifier(ctx, &ctx->notifier,
                            (EventNotifierHandler *)
                            event_notifier_test_and_clear);
-- 
2.1.0

[Qemu-devel] [PATCH 2/3] block: do not allocate an iovec per read of a growable/zero_after_eof BDS

[Paolo Bonzini]

Most reads do not go past the end of the file, and they can use the
input QEMUIOVector instead of creating one.  This removes the
qemu_iovec_* functions from the profile.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 block.c | 12 +++++-------
 1 file changed, 5 insertions(+), 7 deletions(-)

diff --git a/block.c b/block.c
index 4165d42..58f8042 100644
--- a/block.c
+++ b/block.c
@@ -3034,18 +3034,16 @@ static int coroutine_fn bdrv_aligned_preadv(BlockDriverState *bs,
 
         max_nb_sectors = ROUND_UP(MAX(0, total_sectors - sector_num),
                                   align >> BDRV_SECTOR_BITS);
-        if (max_nb_sectors > 0) {
+        if (nb_sectors < max_nb_sectors) {
+            ret = drv->bdrv_co_readv(bs, sector_num, nb_sectors, qiov);
+        } else if (max_nb_sectors > 0) {
             QEMUIOVector local_qiov;
-            size_t local_sectors;
-
-            max_nb_sectors = MIN(max_nb_sectors, SIZE_MAX / BDRV_SECTOR_BITS);
-            local_sectors = MIN(max_nb_sectors, nb_sectors);
 
             qemu_iovec_init(&local_qiov, qiov->niov);
             qemu_iovec_concat(&local_qiov, qiov, 0,
-                              local_sectors * BDRV_SECTOR_SIZE);
+                              max_nb_sectors * BDRV_SECTOR_SIZE);
 
-            ret = drv->bdrv_co_readv(bs, sector_num, local_sectors,
+            ret = drv->bdrv_co_readv(bs, sector_num, max_nb_sectors,
                                      &local_qiov);
 
             qemu_iovec_destroy(&local_qiov);
-- 
2.1.0

[Qemu-devel] [PATCH 3/3] block: replace g_new0 with g_new for bottom half allocation.

[Paolo Bonzini]

This saves about 15% of the clock cycles spent on allocation.  Using the
slice allocator does not add a visible improvement; allocation is faster
than malloc, while freeing seems to be slower.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 async.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/async.c b/async.c
index 572f239..2be88cc 100644
--- a/async.c
+++ b/async.c
@@ -44,10 +44,12 @@ struct QEMUBH {
 QEMUBH *aio_bh_new(AioContext *ctx, QEMUBHFunc *cb, void *opaque)
 {
     QEMUBH *bh;
-    bh = g_new0(QEMUBH, 1);
-    bh->ctx = ctx;
-    bh->cb = cb;
-    bh->opaque = opaque;
+    bh = g_new(QEMUBH, 1);
+    *bh = (QEMUBH){
+        .ctx = ctx,
+        .cb = cb,
+        .opaque = opaque,
+    };
     qemu_mutex_lock(&ctx->bh_lock);
     bh->next = ctx->first_bh;
     /* Make sure that the members are ready before putting bh into list */
-- 
2.1.0

Re: [Qemu-devel] [PATCH 0/3] block: three optimizations

[Kevin Wolf]

Am 17.12.2014 um 16:09 hat Paolo Bonzini geschrieben:
> These are three unrelated micro-optimizations of the block layer.
> The first is only visible on non-dataplane operation, the others
> are generic.

Thanks, applied all to the block branch.

Kevin