* [PULL 1/4] nbd/server: push pending frames after sending reply
2023-03-28 12:35 [PULL 0/4] Block layer patches Kevin Wolf
@ 2023-03-28 12:35 ` Kevin Wolf
2023-03-28 12:35 ` [PULL 2/4] block/export: only acquire AioContext once for vhost_user_server_stop() Kevin Wolf
` (3 subsequent siblings)
4 siblings, 0 replies; 6+ messages in thread
From: Kevin Wolf @ 2023-03-28 12:35 UTC (permalink / raw)
To: qemu-block; +Cc: kwolf, peter.maydell, qemu-devel
From: Florian Westphal <fw@strlen.de>
qemu-nbd doesn't set TCP_NODELAY on the tcp socket.
Kernel waits for more data and avoids transmission of small packets.
Without TLS this is barely noticeable, but with TLS this really shows.
Booting a VM via qemu-nbd on localhost (with tls) takes more than
2 minutes on my system. tcpdump shows frequent wait periods, where no
packets get sent for a 40ms period.
Add explicit (un)corking when processing (and responding to) requests.
"TCP_CORK, &zero" after earlier "CORK, &one" will flush pending data.
VM Boot time:
main: no tls: 23s, with tls: 2m45s
patched: no tls: 14s, with tls: 15s
VM Boot time, qemu-nbd via network (same lan):
main: no tls: 18s, with tls: 1m50s
patched: no tls: 17s, with tls: 18s
Future optimization: if we could detect if there is another pending
request we could defer the uncork operation because more data would be
appended.
Signed-off-by: Florian Westphal <fw@strlen.de>
Message-Id: <20230324104720.2498-1-fw@strlen.de>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
---
nbd/server.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/nbd/server.c b/nbd/server.c
index a4750e4188..848836d414 100644
--- a/nbd/server.c
+++ b/nbd/server.c
@@ -2667,6 +2667,8 @@ static coroutine_fn void nbd_trip(void *opaque)
goto disconnect;
}
+ qio_channel_set_cork(client->ioc, true);
+
if (ret < 0) {
/* It wasn't -EIO, so, according to nbd_co_receive_request()
* semantics, we should return the error to the client. */
@@ -2692,6 +2694,7 @@ static coroutine_fn void nbd_trip(void *opaque)
goto disconnect;
}
+ qio_channel_set_cork(client->ioc, false);
done:
nbd_request_put(req);
nbd_client_put(client);
--
2.39.2
^ permalink raw reply related [flat|nested] 6+ messages in thread* [PULL 3/4] aio-posix: fix race between epoll upgrade and aio_set_fd_handler()
2023-03-28 12:35 [PULL 0/4] Block layer patches Kevin Wolf
2023-03-28 12:35 ` [PULL 1/4] nbd/server: push pending frames after sending reply Kevin Wolf
2023-03-28 12:35 ` [PULL 2/4] block/export: only acquire AioContext once for vhost_user_server_stop() Kevin Wolf
@ 2023-03-28 12:35 ` Kevin Wolf
2023-03-28 12:35 ` [PULL 4/4] block/export: Fix graph locking in blk_get_geometry() call Kevin Wolf
2023-03-28 19:42 ` [PULL 0/4] Block layer patches Peter Maydell
4 siblings, 0 replies; 6+ messages in thread
From: Kevin Wolf @ 2023-03-28 12:35 UTC (permalink / raw)
To: qemu-block; +Cc: kwolf, peter.maydell, qemu-devel
From: Stefan Hajnoczi <stefanha@redhat.com>
If another thread calls aio_set_fd_handler() while the IOThread event
loop is upgrading from ppoll(2) to epoll(7) then we might miss new
AioHandlers. The epollfd will not monitor the new AioHandler's fd,
resulting in hangs.
Take the AioHandler list lock while upgrading to epoll. This prevents
AioHandlers from changing while epoll is being set up. If we cannot lock
because we're in a nested event loop, then don't upgrade to epoll (it
will happen next time we're not in a nested call).
The downside to taking the lock is that the aio_set_fd_handler() thread
has to wait until the epoll upgrade is finished, which involves many
epoll_ctl(2) system calls. However, this scenario is rare and I couldn't
think of another solution that is still simple.
Reported-by: Qing Wang <qinwang@redhat.com>
Buglink: https://bugzilla.redhat.com/show_bug.cgi?id=2090998
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Fam Zheng <fam@euphon.net>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-Id: <20230323144859.1338495-1-stefanha@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
---
util/fdmon-epoll.c | 25 ++++++++++++++++++-------
1 file changed, 18 insertions(+), 7 deletions(-)
diff --git a/util/fdmon-epoll.c b/util/fdmon-epoll.c
index e11a8a022e..1683aa1105 100644
--- a/util/fdmon-epoll.c
+++ b/util/fdmon-epoll.c
@@ -127,6 +127,8 @@ static bool fdmon_epoll_try_enable(AioContext *ctx)
bool fdmon_epoll_try_upgrade(AioContext *ctx, unsigned npfd)
{
+ bool ok;
+
if (ctx->epollfd < 0) {
return false;
}
@@ -136,14 +138,23 @@ bool fdmon_epoll_try_upgrade(AioContext *ctx, unsigned npfd)
return false;
}
- if (npfd >= EPOLL_ENABLE_THRESHOLD) {
- if (fdmon_epoll_try_enable(ctx)) {
- return true;
- } else {
- fdmon_epoll_disable(ctx);
- }
+ if (npfd < EPOLL_ENABLE_THRESHOLD) {
+ return false;
+ }
+
+ /* The list must not change while we add fds to epoll */
+ if (!qemu_lockcnt_dec_if_lock(&ctx->list_lock)) {
+ return false;
+ }
+
+ ok = fdmon_epoll_try_enable(ctx);
+
+ qemu_lockcnt_inc_and_unlock(&ctx->list_lock);
+
+ if (!ok) {
+ fdmon_epoll_disable(ctx);
}
- return false;
+ return ok;
}
void fdmon_epoll_setup(AioContext *ctx)
--
2.39.2
^ permalink raw reply related [flat|nested] 6+ messages in thread* [PULL 4/4] block/export: Fix graph locking in blk_get_geometry() call
2023-03-28 12:35 [PULL 0/4] Block layer patches Kevin Wolf
` (2 preceding siblings ...)
2023-03-28 12:35 ` [PULL 3/4] aio-posix: fix race between epoll upgrade and aio_set_fd_handler() Kevin Wolf
@ 2023-03-28 12:35 ` Kevin Wolf
2023-03-28 19:42 ` [PULL 0/4] Block layer patches Peter Maydell
4 siblings, 0 replies; 6+ messages in thread
From: Kevin Wolf @ 2023-03-28 12:35 UTC (permalink / raw)
To: qemu-block; +Cc: kwolf, peter.maydell, qemu-devel
blk_get_geometry() eventually calls bdrv_nb_sectors(), which is a
co_wrapper_mixed_bdrv_rdlock. This means that when it is called from
coroutine context, it already assume to have the graph locked.
However, virtio_blk_sect_range_ok() in block/export/virtio-blk-handler.c
(used by vhost-user-blk and VDUSE exports) runs in a coroutine, but
doesn't take the graph lock - blk_*() functions are generally expected
to do that internally. This causes an assertion failure when accessing
an export for the first time if it runs in an iothread.
This is an example of the crash:
$ ./storage-daemon/qemu-storage-daemon --object iothread,id=th0 --blockdev file,filename=/home/kwolf/images/hd.img,node-name=disk --export vhost-user-blk,addr.type=unix,addr.path=/tmp/vhost.sock,node-name=disk,id=exp0,iothread=th0
qemu-storage-daemon: ../block/graph-lock.c:268: void assert_bdrv_graph_readable(void): Assertion `qemu_in_main_thread() || reader_count()' failed.
(gdb) bt
#0 0x00007ffff6eafe5c in __pthread_kill_implementation () from /lib64/libc.so.6
#1 0x00007ffff6e5fa76 in raise () from /lib64/libc.so.6
#2 0x00007ffff6e497fc in abort () from /lib64/libc.so.6
#3 0x00007ffff6e4971b in __assert_fail_base.cold () from /lib64/libc.so.6
#4 0x00007ffff6e58656 in __assert_fail () from /lib64/libc.so.6
#5 0x00005555556337a3 in assert_bdrv_graph_readable () at ../block/graph-lock.c:268
#6 0x00005555555fd5a2 in bdrv_co_nb_sectors (bs=0x5555564c5ef0) at ../block.c:5847
#7 0x00005555555ee949 in bdrv_nb_sectors (bs=0x5555564c5ef0) at block/block-gen.c:256
#8 0x00005555555fd6b9 in bdrv_get_geometry (bs=0x5555564c5ef0, nb_sectors_ptr=0x7fffef7fedd0) at ../block.c:5884
#9 0x000055555562ad6d in blk_get_geometry (blk=0x5555564cb200, nb_sectors_ptr=0x7fffef7fedd0) at ../block/block-backend.c:1624
#10 0x00005555555ddb74 in virtio_blk_sect_range_ok (blk=0x5555564cb200, block_size=512, sector=0, size=512) at ../block/export/virtio-blk-handler.c:44
#11 0x00005555555dd80d in virtio_blk_process_req (handler=0x5555564cbb98, in_iov=0x7fffe8003830, out_iov=0x7fffe8003860, in_num=1, out_num=0) at ../block/export/virtio-blk-handler.c:189
#12 0x00005555555dd546 in vu_blk_virtio_process_req (opaque=0x7fffe8003800) at ../block/export/vhost-user-blk-server.c:66
#13 0x00005555557bf4a1 in coroutine_trampoline (i0=-402635264, i1=32767) at ../util/coroutine-ucontext.c:177
#14 0x00007ffff6e75c20 in ?? () from /lib64/libc.so.6
#15 0x00007fffefffa870 in ?? ()
#16 0x0000000000000000 in ?? ()
Fix this by creating a new blk_co_get_geometry() that takes the lock,
and changing blk_get_geometry() to be a co_wrapper_mixed around it.
To make the resulting code cleaner, virtio-blk-handler.c can directly
call the coroutine version now (though that wouldn't be necessary for
fixing the bug, taking the lock in blk_co_get_geometry() is what fixes
it).
Fixes: 8ab8140a04cf771d63e9754d6ba6c1e676bfe507
Reported-by: Lukáš Doktor <ldoktor@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Message-Id: <20230327113959.60071-1-kwolf@redhat.com>
Reviewed-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
---
include/block/block-io.h | 4 +++-
include/sysemu/block-backend-io.h | 5 ++++-
block.c | 5 +++--
block/block-backend.c | 7 +++++--
block/export/virtio-blk-handler.c | 7 ++++---
5 files changed, 19 insertions(+), 9 deletions(-)
diff --git a/include/block/block-io.h b/include/block/block-io.h
index 5da99d4d60..dbc034b728 100644
--- a/include/block/block-io.h
+++ b/include/block/block-io.h
@@ -89,7 +89,9 @@ int64_t co_wrapper bdrv_get_allocated_file_size(BlockDriverState *bs);
BlockMeasureInfo *bdrv_measure(BlockDriver *drv, QemuOpts *opts,
BlockDriverState *in_bs, Error **errp);
-void bdrv_get_geometry(BlockDriverState *bs, uint64_t *nb_sectors_ptr);
+
+void coroutine_fn GRAPH_RDLOCK
+bdrv_co_get_geometry(BlockDriverState *bs, uint64_t *nb_sectors_ptr);
int coroutine_fn GRAPH_RDLOCK
bdrv_co_delete_file(BlockDriverState *bs, Error **errp);
diff --git a/include/sysemu/block-backend-io.h b/include/sysemu/block-backend-io.h
index 40ab178719..c672b77247 100644
--- a/include/sysemu/block-backend-io.h
+++ b/include/sysemu/block-backend-io.h
@@ -70,7 +70,10 @@ void co_wrapper blk_eject(BlockBackend *blk, bool eject_flag);
int64_t coroutine_fn blk_co_getlength(BlockBackend *blk);
int64_t co_wrapper_mixed blk_getlength(BlockBackend *blk);
-void blk_get_geometry(BlockBackend *blk, uint64_t *nb_sectors_ptr);
+void coroutine_fn blk_co_get_geometry(BlockBackend *blk,
+ uint64_t *nb_sectors_ptr);
+void co_wrapper_mixed blk_get_geometry(BlockBackend *blk,
+ uint64_t *nb_sectors_ptr);
int64_t coroutine_fn blk_co_nb_sectors(BlockBackend *blk);
int64_t co_wrapper_mixed blk_nb_sectors(BlockBackend *blk);
diff --git a/block.c b/block.c
index 0dd604d0f6..e0c6c648b1 100644
--- a/block.c
+++ b/block.c
@@ -5879,9 +5879,10 @@ int64_t coroutine_fn bdrv_co_getlength(BlockDriverState *bs)
}
/* return 0 as number of sectors if no device present or error */
-void bdrv_get_geometry(BlockDriverState *bs, uint64_t *nb_sectors_ptr)
+void coroutine_fn bdrv_co_get_geometry(BlockDriverState *bs,
+ uint64_t *nb_sectors_ptr)
{
- int64_t nb_sectors = bdrv_nb_sectors(bs);
+ int64_t nb_sectors = bdrv_co_nb_sectors(bs);
IO_CODE();
*nb_sectors_ptr = nb_sectors < 0 ? 0 : nb_sectors;
diff --git a/block/block-backend.c b/block/block-backend.c
index 278b04ce69..2ee39229e4 100644
--- a/block/block-backend.c
+++ b/block/block-backend.c
@@ -1615,13 +1615,16 @@ int64_t coroutine_fn blk_co_getlength(BlockBackend *blk)
return bdrv_co_getlength(blk_bs(blk));
}
-void blk_get_geometry(BlockBackend *blk, uint64_t *nb_sectors_ptr)
+void coroutine_fn blk_co_get_geometry(BlockBackend *blk,
+ uint64_t *nb_sectors_ptr)
{
IO_CODE();
+ GRAPH_RDLOCK_GUARD();
+
if (!blk_bs(blk)) {
*nb_sectors_ptr = 0;
} else {
- bdrv_get_geometry(blk_bs(blk), nb_sectors_ptr);
+ bdrv_co_get_geometry(blk_bs(blk), nb_sectors_ptr);
}
}
diff --git a/block/export/virtio-blk-handler.c b/block/export/virtio-blk-handler.c
index 313666e8ab..bc1cec6757 100644
--- a/block/export/virtio-blk-handler.c
+++ b/block/export/virtio-blk-handler.c
@@ -22,8 +22,9 @@ struct virtio_blk_inhdr {
unsigned char status;
};
-static bool virtio_blk_sect_range_ok(BlockBackend *blk, uint32_t block_size,
- uint64_t sector, size_t size)
+static bool coroutine_fn
+virtio_blk_sect_range_ok(BlockBackend *blk, uint32_t block_size,
+ uint64_t sector, size_t size)
{
uint64_t nb_sectors;
uint64_t total_sectors;
@@ -41,7 +42,7 @@ static bool virtio_blk_sect_range_ok(BlockBackend *blk, uint32_t block_size,
if ((sector << VIRTIO_BLK_SECTOR_BITS) % block_size) {
return false;
}
- blk_get_geometry(blk, &total_sectors);
+ blk_co_get_geometry(blk, &total_sectors);
if (sector > total_sectors || nb_sectors > total_sectors - sector) {
return false;
}
--
2.39.2
^ permalink raw reply related [flat|nested] 6+ messages in thread