The following QMP command leads to a crash when iothreads are used:
{ 'execute': 'device_del', 'arguments': {'id': 'data'} }
The backtrace involves the queue restart coroutine where
tgm->throttle_state is a NULL pointer because
throttle_group_unregister_tgm() has already been called:
(gdb) bt full
#0 0x00005585a7a3b378 in qemu_mutex_lock_impl (mutex=0xffffffffffffffd0, file=0x5585a7bb3d54 "block/throttle-groups.c", line=412) at util/qemu-thread-posix.c:64
err = <optimized out>
__PRETTY_FUNCTION__ = "qemu_mutex_lock_impl"
__func__ = "qemu_mutex_lock_impl"
#1 0x00005585a79be074 in throttle_group_restart_queue_entry (opaque=0x5585a9de4eb0) at block/throttle-groups.c:412
_f = <optimized out>
data = 0x5585a9de4eb0
tgm = 0x5585a9079440
ts = 0x0
tg = 0xffffffffffffff98
is_write = false
empty_queue = 255
This coroutine should not execute in the iothread after the throttle
group member has been unregistered!
The root cause is that the device_del code path schedules the restart
coroutine in the iothread while holding the AioContext lock. Therefore
the iothread cannot execute the coroutine until after device_del
releases the lock - by this time it's too late.
This patch adds a reference count to ThrottleGroupMember so we can
synchronously wait for restart coroutines to complete. Once they are
done it is safe to unregister the ThrottleGroupMember.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Alberto Garcia <berto@igalia.com>
Message-id: 20190114133257.30299-2-stefanha@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Commit 6fccbb475b fixed a bug caused by
QEMU attempting to remove a throttle group member with no pending
requests but an active timer set. This was the result of a previous
bdrv_drained_begin() call processing the throttled requests but
leaving the timer untouched.
Although the commit does solve the problem, the situation shouldn't
happen in the first place. If we try to drain a throttle group member
which has a timer set, we should cancel the timer instead of ignoring
it.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
In the throttling code after an I/O request has been completed the
next one is selected from a different member using a round-robin
algorithm. This ensures that all members get a chance to finish their
pending I/O requests.
However, if a group member has its I/O limits disabled (because it's
being drained) then we should always give it priority in order to have
all its pending requests finished as soon as possible.
If we don't do this we could have a member in the process of being
drained waiting for the throttled requests of other members, for which
the I/O limits still apply.
This can have additional consequences: if we're running in qtest mode
(with QEMU_CLOCK_VIRTUAL) then timers can only fire if we advance the
clock manually, so attempting to drain a block device can hang QEMU in
the BDRV_POLL_WHILE() loop at the end of bdrv_do_drained_begin().
Signed-off-by: Alberto Garcia <berto@igalia.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Throttle groups consist of members sharing one throttling state
(including bps/iops limits). Round-robin scheduling is used to ensure
fairness. If a group member already has a timer pending then other
groups members do not schedule their own timers. The next group member
will have its turn when the existing timer expires.
A hang may occur when a group member leaves while it had a timer
scheduled. Although the code carefully removes the group member from
the round-robin list, it does not schedule the next member. Therefore
remaining members continue to wait for the removed member's timer to
expire.
This patch schedules the next request if a timer is pending.
Unfortunately the actual bug is a race condition that I've been unable
to capture in a test case.
Sometimes drive2 hangs when drive1 is removed from the throttling group:
$ qemu ... -drive if=none,id=drive1,cache=none,format=qcow2,file=data1.qcow2,iops=100,group=foo \
-device virtio-blk-pci,id=virtio-blk-pci0,drive=drive1 \
-drive if=none,id=drive2,cache=none,format=qcow2,file=data2.qcow2,iops=10,group=foo \
-device virtio-blk-pci,id=virtio-blk-pci1,drive=drive2
(guest-console1)# fio -filename /dev/vda 4k-seq-read.job
(guest-console2)# fio -filename /dev/vdb 4k-seq-read.job
(qmp) {"execute": "block_set_io_throttle", "arguments": {"device": "drive1","bps": 0,"bps_rd": 0,"bps_wr": 0,"iops": 0,"iops_rd": 0,"iops_wr": 0}}
Reported-by: Nini Gu <ngu@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 20180704145410.794-1-stefanha@redhat.com
RHBZ: https://bugzilla.redhat.com/show_bug.cgi?id=1535914
Cc: Alberto Garcia <berto@igalia.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
In my "build everything" tree, a change to the types in
qapi-schema.json triggers a recompile of about 4800 out of 5100
objects.
The previous commit split up qmp-commands.h, qmp-event.h, qmp-visit.h,
qapi-types.h. Each of these headers still includes all its shards.
Reduce compile time by including just the shards we actually need.
To illustrate the benefits: adding a type to qapi/migration.json now
recompiles some 2300 instead of 4800 objects. The next commit will
improve it further.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Message-Id: <20180211093607.27351-24-armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Marc-André Lureau <marcandre.lureau@redhat.com>
[eblake: rebase to master]
Signed-off-by: Eric Blake <eblake@redhat.com>
tg->any_timer_armed[] must be cleared when detaching pending timers from
the AioContext. Failure to do so leads to hung I/O because it looks
like there are still timers pending when in fact they have been removed.
Other ThrottleGroupMembers might have requests pending too so it's
necessary to schedule the next TGM so it can set a timer.
This patch fixes hung I/O when QEMU is launched with drives that are in
the same throttling group:
(guest)$ dd if=/dev/zero of=/dev/vdb oflag=direct bs=512 &
(guest)$ dd if=/dev/zero of=/dev/vdc oflag=direct bs=512 &
(qemu) stop
(qemu) cont
...I/O is stuck...
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: 20171116112150.27607-1-stefanha@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
I/O requests hang after stop/cont commands at least since QEMU 2.10.0
with -drive iops=100:
(guest)$ dd if=/dev/zero of=/dev/vdb oflag=direct count=1000
(qemu) stop
(qemu) cont
...I/O is stuck...
This happens because blk_set_aio_context() detaches the ThrottleState
while requests may still be in flight:
if (tgm->throttle_state) {
throttle_group_detach_aio_context(tgm);
throttle_group_attach_aio_context(tgm, new_context);
}
This patch encloses the detach/attach calls in a drained region so no
I/O request is left hanging. Also add assertions so we don't make the
same mistake again in the future.
Reported-by: Yongxue Hong <yhong@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Alberto Garcia <berto@igalia.com>
Message-id: 20171110151934.16883-1-stefanha@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
RestartData is the opaque data of the throttle_group_restart_queue_entry
coroutine. By being stack allocated, it isn't available anymore if
aio_co_enter schedules the coroutine with a bottom half and runs after
throttle_group_restart_queue returns.
Cc: qemu-stable@nongnu.org
Signed-off-by: Manos Pitsidianakis <el13635@mail.ntua.gr>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Alberto Garcia <berto@igalia.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
block/throttle.c uses existing I/O throttle infrastructure inside a
block filter driver. I/O operations are intercepted in the filter's
read/write coroutines, and referred to block/throttle-groups.c
The driver can be used with the syntax
-drive driver=throttle,file.filename=foo.qcow2,throttle-group=bar
which registers the throttle filter node with the ThrottleGroup 'bar'. The
given group must be created beforehand with object-add or -object.
Reviewed-by: Alberto Garcia <berto@igalia.com>
Signed-off-by: Manos Pitsidianakis <el13635@mail.ntua.gr>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
ThrottleGroup is converted to an object. This will allow the future
throttle block filter drive easy creation and configuration of throttle
groups in QMP and cli.
A new QAPI struct, ThrottleLimits, is introduced to provide a shared
struct for all throttle configuration needs in QMP.
ThrottleGroups can be created via CLI as
-object throttle-group,id=foo,x-iops-total=100,x-..
where x-* are individual limit properties. Since we can't add non-scalar
properties in -object this interface must be used instead. However,
setting these properties must be disabled after initialization because
certain combinations of limits are forbidden and thus configuration
changes should be done in one transaction. The individual properties
will go away when support for non-scalar values in CLI is implemented
and thus are marked as experimental.
ThrottleGroup also has a `limits` property that uses the ThrottleLimits
struct. It can be used to create ThrottleGroups or set the
configuration in existing groups as follows:
{ "execute": "object-add",
"arguments": {
"qom-type": "throttle-group",
"id": "foo",
"props" : {
"limits": {
"iops-total": 100
}
}
}
}
{ "execute" : "qom-set",
"arguments" : {
"path" : "foo",
"property" : "limits",
"value" : {
"iops-total" : 99
}
}
}
This also means a group's configuration can be fetched with qom-get.
Signed-off-by: Manos Pitsidianakis <el13635@mail.ntua.gr>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Alberto Garcia <berto@igalia.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Move the CoMutex and CoQueue inits inside throttle_group_register_tgm()
which is called whenever a ThrottleGroupMember is initialized. There's
no need for them to be separate.
Reviewed-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Manos Pitsidianakis <el13635@mail.ntua.gr>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
timer_cb() needs to know about the current Aio context of the throttle
request that is woken up. In order to make ThrottleGroupMember backend
agnostic, this information is stored in an aio_context field instead of
accessing it from BlockBackend.
Reviewed-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Manos Pitsidianakis <el13635@mail.ntua.gr>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This commit eliminates the 1:1 relationship between BlockBackend and
throttle group state. Users will be able to create multiple throttle
nodes, each with its own throttle group state, in the future. The
throttle group state cannot be per-BlockBackend anymore, it must be
per-throttle node. This is done by gathering ThrottleGroup membership
details from BlockBackendPublic into ThrottleGroupMember and refactoring
existing code to use the structure.
Reviewed-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Manos Pitsidianakis <el13635@mail.ntua.gr>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
throttle_config() cancels the timers of the calling BlockBackend. This
doesn't make sense because other BlockBackends in the group remain
untouched. There's no need to cancel the timers in the one specific
BlockBackend so let's not do that. Throttled requests will run as
scheduled and future requests will follow the new configuration. This
also allows a throttle group's configuration to be changed even when it
has no members.
Signed-off-by: Manos Pitsidianakis <el13635@mail.ntua.gr>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Clock type in throttling is currently inferred by the ThrottleTimer's
clock type even though it is a per-ThrottleGroup property; it doesn't
make sense to have different clock types in the same group. Moving this
to a field in ThrottleGroup can simplify some of the throttle functions.
Signed-off-by: Manos Pitsidianakis <el13635@mail.ntua.gr>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
There used to be throttle_timers_{detach,attach}_aio_context() calls
in bdrv_set_aio_context(), but since 7ca7f0f6db
they are now in blk_set_aio_context().
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Another possibility is to use tg->lock, which we're holding anyway in
both schedule_next_request and throttle_group_co_io_limits_intercept.
This would require open-coding the CoQueue however, so I've chosen this
alternative.
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20170605123908.18777-10-pbonzini@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
Prepare for removing this function; always restart throttled requests
from coroutine context. This will matter when restarting throttled
requests will have to acquire a CoMutex.
Reviewed-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20170605123908.18777-9-pbonzini@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
Starting all waiting coroutines from bdrv_drain_all is unnecessary;
throttle_group_co_io_limits_intercept calls schedule_next_request as
soon as the coroutine restarts, which in turn will restart the next
request if possible.
If we only start the first request and let the coroutines dance from
there the code is simpler and there is more reuse between
throttle_group_config, throttle_group_restart_blk and timer_cb. The
next patch will benefit from this.
We also stop accessing from throttle_group_restart_blk the
blkp->throttled_reqs CoQueues even when there was no
attached throttling group. This worked but is not pretty.
The only thing that can interrupt the dance is the QEMU_CLOCK_VIRTUAL
timer when switching from one block device to the next, because the
timer is set to "now + 1" but QEMU_CLOCK_VIRTUAL might not be running.
Set that timer to point in the present ("now") rather than the future
and things work.
Reviewed-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20170605123908.18777-8-pbonzini@redhat.com>
Signed-off-by: Fam Zheng <famz@redhat.com>
All that CoQueue needs in order to become thread-safe is help
from an external mutex. Add this to the API.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Message-id: 20170213181244.16297-6-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Reviewed-by: Daniel P. Berrange <berrange@redhat.com>
Message-id: 20170213135235.12274-13-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
In 27ccdd5259 the throttling fields were
moved from BlockDriverState to BlockBackend. However in a few cases
the code started using throttling fields from the active BlockBackend
instead of the round-robin token, making the algorithm behave
incorrectly.
This can cause starvation if there's a throttling group with several
drives but only one of them has I/O.
Cc: qemu-stable@nongnu.org
Reported-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This patch changes where the throttling state is stored (used to be the
BlockDriverState, now it is the BlockBackend), but it doesn't actually
make it a BB level feature yet. For example, throttling is still
disabled when the BDS is detached from the BB.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
As a first step towards moving I/O throttling to the BlockBackend level,
this patch changes all pointers in struct ThrottleGroup from referencing
a BlockDriverState to referencing a BlockBackend.
This change is valid because we made sure that throttling can only be
enabled on BDSes which have a BB attached.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Alberto Garcia <berto@igalia.com>
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
Extract the handling of throttling from bdrv_flush_io_queue. These
new functions will soon become BdrvChildRole callbacks, as they can
be generalized to "beginning of drain" and "end of drain".
Reviewed-by: Alberto Garcia <berto@igalia.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
We want to remove throttled_reqs from block/io.c. This is the easy
part---hide the handling of throttled_reqs during disable/enable of
throttling within throttle-groups.c.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Clean up includes so that osdep.h is included first and headers
which it implies are not included manually.
This commit was created with scripts/clean-includes.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
throttle_group_unregister_bs() removes a BlockDriverState from its
throttling group and destroys the timers. This means that there must
be no pending throttled requests at that point (because it would be
impossible to complete them), so the caller has to drain them first.
At the moment throttle_group_unregister_bs() is only called from
bdrv_io_limits_disable(), which already takes care of draining the
requests, so there's nothing to worry about, but this patch makes
this invariant explicit in the documentation and adds the relevant
assertions.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The group throttling code was always meant to handle its locking
internally. However, bdrv_swap() was touching the ThrottleGroup
structure directly and therefore needed an API for that.
Now that bdrv_swap() no longer exists there's no need for the
throttle_group_lock() API anymore.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Throttle groups are not necessarily referenced by BDSs alone; a later
patch will essentially allow BBs to reference them, too. Make the
ref/unref functions public so that reference can be properly accounted
for.
Their interface is slightly adjusted in that they return and take a
ThrottleState pointer, respectively, instead of a ThrottleGroup pointer.
Functionally, they are equivalent, but since ThrottleGroup is not meant
to be used outside of block/throttle-groups.c, ThrottleState is easier
to handle.
Signed-off-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Calling throttle_group_config() cancels all timers from a particular
BlockDriverState, so any_timer_armed[] should be updated accordingly.
However, with the current code it may happen that a timer is armed in
a different BlockDriverState from the same group, so any_timer_armed[]
would be set to false in a situation where there is still a timer
armed.
The consequence is that we might end up with two timers armed. This
should not have any noticeable impact however, since all accesses to
the ThrottleGroup are protected by a lock, and the situation would
become normal again shortly thereafter as soon as all timers have been
fired.
The correct way to solve this is to check that we're actually
cancelling a timer before updating any_timer_armed[].
Signed-off-by: Alberto Garcia <berto@igalia.com>
Message-id: 1434382875-3998-1-git-send-email-berto@igalia.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
bdrv_swap() touches the fields of a BlockDriverState that are
protected by the ThrottleGroup lock. Although those fields end up in
their original place, they are temporarily swapped in the process,
so there's a chance that an operation on a member of the same group
happening on a different thread can try to use them.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: d92dc40d7c4f1fc5cda5cbbf4ffb7a4670b79d17.1433779731.git.berto@igalia.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
The throttle group support use a cooperative round robin scheduling
algorithm.
The principles of the algorithm are simple:
- Each BDS of the group is used as a token in a circular way.
- The active BDS computes if a wait must be done and arms the right
timer.
- If a wait must be done the token timer will be armed so the token
will become the next active BDS.
Signed-off-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-id: f0082a86f3ac01c46170f7eafe2101a92e8fde39.1433779731.git.berto@igalia.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Stefan Hajnoczi