In the block layer, synchronous APIs are often implemented by creating a
coroutine that calls the asynchronous coroutine-based implementation and
then waiting for completion with BDRV_POLL_WHILE().
For this to work with iothreads (more specifically, when the synchronous
API is called in a thread that is not the home thread of the block
device, so that the coroutine will run in a different thread), we must
make sure to call aio_wait_kick() at the end of the operation. Many
places are missing this, so that BDRV_POLL_WHILE() keeps hanging even if
the condition has long become false.
Note that bdrv_dec_in_flight() involves an aio_wait_kick() call. This
corresponds to the BDRV_POLL_WHILE() in the drain functions, but it is
generally not enough for most other operations because they haven't set
the return value in the coroutine entry stub yet. To avoid race
conditions there, we need to kick after setting the return value.
The race window is small enough that the problem doesn't usually surface
in the common path. However, it does surface and causes easily
reproducible hangs if the operation can return early before even calling
bdrv_inc/dec_in_flight, which many of them do (trivial error or no-op
success paths).
The bug in bdrv_truncate(), bdrv_check() and bdrv_invalidate_cache() is
slightly different: These functions even neglected to schedule the
coroutine in the home thread of the node. This avoids the hang, but is
obviously wrong, too. Fix those to schedule the coroutine in the right
AioContext in addition to adding aio_wait_kick() calls.
Cc: qemu-stable@nongnu.org
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
When the IO size is larger than 2 pages, we move the the pointer one by
one in the pagelist, this is inefficient.
This is a simple benchmark result:
Before:
$ qemu-io -c 'write 0 1G' nvme://0000:00:04.0/1
wrote 1073741824/1073741824 bytes at offset 0
1 GiB, 1 ops; 0:00:02.41 (424.504 MiB/sec and 0.4146 ops/sec)
$ qemu-io -c 'read 0 1G' nvme://0000:00:04.0/1
read 1073741824/1073741824 bytes at offset 0
1 GiB, 1 ops; 0:00:02.03 (503.055 MiB/sec and 0.4913 ops/sec)
After:
$ qemu-io -c 'write 0 1G' nvme://0000:00:04.0/1
wrote 1073741824/1073741824 bytes at offset 0
1 GiB, 1 ops; 0:00:02.17 (471.517 MiB/sec and 0.4605 ops/sec)
$ qemu-io -c 'read 0 1G' nvme://0000:00:04.0/1
read 1073741824/1073741824 bytes at offset 0
1 GiB, 1 ops; 0:00:01.94 (526.770 MiB/sec and 0.5144 ops/sec)
Signed-off-by: Li Feng <lifeng1519@gmail.com>
Message-Id: <20181101103807.25862-1-lifeng1519@gmail.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
nvme_poll_queues is already protected by q->lock, and
AIO callbacks are invoked outside the AioContext lock.
So remove the acquire/release pair in nvme_handle_event.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20180814062739.19640-1-pbonzini@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
bdrv_io_plug/bdrv_io_unplug take care of keeping a nesting count,
so change s->plugged to just a bool.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20180813144320.12382-2-pbonzini@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
It is wrong to leave this field as 1, as nvme_close() called in the
error handling code in nvme_file_open() will use it and try to free
s->queues again.
Another problem is the cleaning ups are duplicated between the fail*
labels of nvme_init() and nvme_file_open(), which calls nvme_close().
A third problem is nvme_close() misses g_free() and
event_notifier_cleanup().
Fix all of them.
Cc: qemu-stable@nongnu.org
Signed-off-by: Fam Zheng <famz@redhat.com>
Message-Id: <20180712025420.4932-1-famz@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
For convenience and clarity, make it possible to call qobject_ref() at
the time when the reference is associated with a variable, or
argument, by making qobject_ref() return the same pointer as given.
Use that to simplify the callers.
Signed-off-by: Marc-André Lureau <marcandre.lureau@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180419150145.24795-5-marcandre.lureau@redhat.com>
Reviewed-by: Markus Armbruster <armbru@redhat.com>
[Useless change to qobject_ref_impl() dropped, commit message improved
slightly]
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Now that we can safely call QOBJECT() on QObject * as well as its
subtypes, we can have macros qobject_ref() / qobject_unref() that work
everywhere instead of having to use QINCREF() / QDECREF() for QObject
and qobject_incref() / qobject_decref() for its subtypes.
The replacement is mechanical, except I broke a long line, and added a
cast in monitor_qmp_cleanup_req_queue_locked(). Unlike
qobject_decref(), qobject_unref() doesn't accept void *.
Note that the new macros evaluate their argument exactly once, thus no
need to shout them.
Signed-off-by: Marc-André Lureau <marcandre.lureau@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20180419150145.24795-4-marcandre.lureau@redhat.com>
Reviewed-by: Markus Armbruster <armbru@redhat.com>
[Rebased, semantic conflict resolved, commit message improved]
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Commit bdd6a90 has a bug: drivers should never directly set
BDRV_BLOCK_ALLOCATED, but only io.c should do that (as needed).
Instead, drivers should report BDRV_BLOCK_DATA if it knows that
data comes from this BDS.
But let's look at the bigger picture: semantically, the nvme
driver is similar to the nbd, null, and raw drivers (no backing
file, all data comes from this BDS). But while two of those
other drivers have to supply the callback (null because it can
special-case BDRV_BLOCK_ZERO, raw because it can special-case
a different offset), in this case the block layer defaults are
good enough without the callback at all (similar to nbd).
So, fix the bug by deletion ;)
Signed-off-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
1) string not null terminated in sysfs_find_group_file
2) NULL pointer dereference and dead local variable in nvme_init.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
Message-Id: <20180213015240.9352-1-famz@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
qemu-common.h includes qemu/option.h, but most places that include the
former don't actually need the latter. Drop the include, and add it
to the places that actually need it.
While there, drop superfluous includes of both headers, and
separate #include from file comment with a blank line.
This cleanup makes the number of objects depending on qemu/option.h
drop from 4545 (out of 4743) to 284 in my "build everything" tree.
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Message-Id: <20180201111846.21846-20-armbru@redhat.com>
[Semantic conflict with commit bdd6a90a9e in block/nvme.c resolved]
Forward these two calls to the IOVA manager.
Signed-off-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-Id: <20180116060901.17413-6-famz@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
This is a new protocol driver that exclusively opens a host NVMe
controller through VFIO. It achieves better latency than linux-aio by
completely bypassing host kernel vfs/block layer.
$rw-$bs-$iodepth linux-aio nvme://
----------------------------------------
randread-4k-1 10.5k 21.6k
randread-512k-1 745 1591
randwrite-4k-1 30.7k 37.0k
randwrite-512k-1 1945 1980
(unit: IOPS)
The driver also integrates with the polling mechanism of iothread.
This patch is co-authored by Paolo and me.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
Message-Id: <20180116060901.17413-4-famz@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
