v2: add small fix by Stefano, Hanna's series fixed
-----BEGIN PGP SIGNATURE-----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=TVTu
-----END PGP SIGNATURE-----
Merge remote-tracking branch 'remotes/vsementsov/tags/pull-jobs-2021-10-07-v2' into staging
mirror: Handle errors after READY cancel
v2: add small fix by Stefano, Hanna's series fixed
# gpg: Signature made Thu 07 Oct 2021 08:25:07 AM PDT
# gpg: using RSA key 8B9C26CDB2FD147C880E86A1561F24C1F19F79FB
# gpg: Good signature from "Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>" [unknown]
# gpg: WARNING: This key is not certified with a trusted signature!
# gpg: There is no indication that the signature belongs to the owner.
# Primary key fingerprint: 8B9C 26CD B2FD 147C 880E 86A1 561F 24C1 F19F 79FB
* remotes/vsementsov/tags/pull-jobs-2021-10-07-v2:
iotests: Add mirror-ready-cancel-error test
mirror: Do not clear .cancelled
mirror: Stop active mirroring after force-cancel
mirror: Check job_is_cancelled() earlier
mirror: Use job_is_cancelled()
job: Add job_cancel_requested()
job: Do not soft-cancel after a job is done
jobs: Give Job.force_cancel more meaning
job: @force parameter for job_cancel_sync()
job: Force-cancel jobs in a failed transaction
mirror: Drop s->synced
mirror: Keep s->synced on error
job: Context changes in job_completed_txn_abort()
block/aio_task: assert `max_busy_tasks` is greater than 0
block/backup: avoid integer overflow of `max-workers`
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Clearing .cancelled before leaving the main loop when the job has been
soft-cancelled is no longer necessary since job_is_cancelled() only
returns true for jobs that have been force-cancelled.
Therefore, this only makes a differences in places that call
job_cancel_requested(). In block/mirror.c, this is done only before
.cancelled was cleared.
In job.c, there are two callers:
- job_completed_txn_abort() asserts that .cancelled is true, so keeping
it true will not affect this place.
- job_complete() refuses to let a job complete that has .cancelled set.
It is correct to refuse to let the user invoke job-complete on mirror
jobs that have already been soft-cancelled.
With this change, there are no places that reset .cancelled to false and
so we can be sure that .force_cancel can only be true if .cancelled is
true as well. Assert this in job_is_cancelled().
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20211006151940.214590-13-hreitz@redhat.com>
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Once the mirror job is force-cancelled (job_is_cancelled() is true), we
should not generate new I/O requests. This applies to active mirroring,
too, so stop it once the job is cancelled.
(We must still forward all I/O requests to the source, though, of
course, but those are not really I/O requests generated by the job, so
this is fine.)
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20211006151940.214590-12-hreitz@redhat.com>
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
We must check whether the job is force-cancelled early in our main loop,
most importantly before any `continue` statement. For example, we used
to have `continue`s before our current checking location that are
triggered by `mirror_flush()` failing. So, if `mirror_flush()` kept
failing, force-cancelling the job would not terminate it.
Jobs can be cancelled while they yield, and once they are
(force-cancelled), they should not generate new I/O requests.
Therefore, we should put the check after the last yield before
mirror_iteration() is invoked.
Buglink: https://gitlab.com/qemu-project/qemu/-/issues/462
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20211006151940.214590-11-hreitz@redhat.com>
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
mirror_drained_poll() returns true whenever the job is cancelled,
because "we [can] be sure that it won't issue more requests". However,
this is only true for force-cancelled jobs, so use job_is_cancelled().
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20211006151940.214590-10-hreitz@redhat.com>
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Most callers of job_is_cancelled() actually want to know whether the job
is on its way to immediate termination. For example, we refuse to pause
jobs that are cancelled; but this only makes sense for jobs that are
really actually cancelled.
A mirror job that is cancelled during READY with force=false should
absolutely be allowed to pause. This "cancellation" (which is actually
a kind of completion) may take an indefinite amount of time, and so
should behave like any job during normal operation. For example, with
on-target-error=stop, the job should stop on write errors. (In
contrast, force-cancelled jobs should not get write errors, as they
should just terminate and not do further I/O.)
Therefore, redefine job_is_cancelled() to only return true for jobs that
are force-cancelled (which as of HEAD^ means any job that interprets the
cancellation request as a request for immediate termination), and add
job_cancel_requested() as the general variant, which returns true for
any jobs which have been requested to be cancelled, whether it be
immediately or after an arbitrarily long completion phase.
Finally, here is a justification for how different job_is_cancelled()
invocations are treated by this patch:
- block/mirror.c (mirror_run()):
- The first invocation is a while loop that should loop until the job
has been cancelled or scheduled for completion. What kind of cancel
does not matter, only the fact that the job is supposed to end.
- The second invocation wants to know whether the job has been
soft-cancelled. Calling job_cancel_requested() is a bit too broad,
but if the job were force-cancelled, we should leave the main loop
as soon as possible anyway, so this should not matter here.
- The last two invocations already check force_cancel, so they should
continue to use job_is_cancelled().
- block/backup.c, block/commit.c, block/stream.c, anything in tests/:
These jobs know only force-cancel, so there is no difference between
job_is_cancelled() and job_cancel_requested(). We can continue using
job_is_cancelled().
- job.c:
- job_pause_point(), job_yield(), job_sleep_ns(): Only force-cancelled
jobs should be prevented from being paused. Continue using job_is_cancelled().
- job_update_rc(), job_finalize_single(), job_finish_sync(): These
functions are all called after the job has left its main loop. The
mirror job (the only job that can be soft-cancelled) will clear
.cancelled before leaving the main loop if it has been
soft-cancelled. Therefore, these functions will observe .cancelled
to be true only if the job has been force-cancelled. We can
continue to use job_is_cancelled().
(Furthermore, conceptually, a soft-cancelled mirror job should not
report to have been cancelled. It should report completion (see
also the block-job-cancel QAPI documentation). Therefore, it makes
sense for these functions not to distinguish between a
soft-cancelled mirror job and a job that has completed as normal.)
- job_completed_txn_abort(): All jobs other than @job have been
force-cancelled. job_is_cancelled() must be true for them.
Regarding @job itself: job_completed_txn_abort() is mostly called
when the job's return value is not 0. A soft-cancelled mirror has a
return value of 0, and so will not end up here then.
However, job_cancel() invokes job_completed_txn_abort() if the job
has been deferred to the main loop, which is mostly the case for
completed jobs (which skip the assertion), but not for sure.
To be safe, use job_cancel_requested() in this assertion.
- job_complete(): This is function eventually invoked by the user
(through qmp_block_job_complete() or qmp_job_complete(), or
job_complete_sync(), which comes from qemu-img). The intention here
is to prevent a user from invoking job-complete after the job has
been cancelled. This should also apply to soft cancelling: After a
mirror job has been soft-cancelled, the user should not be able to
decide otherwise and have it complete as normal (i.e. pivoting to
the target).
- job_cancel(): Both functions are equivalent (see comment there), but
we want to use job_is_cancelled(), because this shows that we call
job_completed_txn_abort() only for force-cancelled jobs. (As
explained for job_update_rc(), soft-cancelled jobs should be treated
as if they have completed as normal.)
Buglink: https://gitlab.com/qemu-project/qemu/-/issues/462
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20211006151940.214590-9-hreitz@redhat.com>
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
We largely have two cancel modes for jobs:
First, there is actual cancelling. The job is terminated as soon as
possible, without trying to reach a consistent result.
Second, we have mirror in the READY state. Technically, the job is not
really cancelled, but it just is a different completion mode. The job
can still run for an indefinite amount of time while it tries to reach a
consistent result.
We want to be able to clearly distinguish which cancel mode a job is in
(when it has been cancelled). We can use Job.force_cancel for this, but
right now it only reflects cancel requests from the user with
force=true, but clearly, jobs that do not even distinguish between
force=false and force=true are effectively always force-cancelled.
So this patch has Job.force_cancel signify whether the job will
terminate as soon as possible (force_cancel=true) or whether it will
effectively remain running despite being "cancelled"
(force_cancel=false).
To this end, we let jobs that provide JobDriver.cancel() tell the
generic job code whether they will terminate as soon as possible or not,
and for jobs that do not provide that method we assume they will.
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-Id: <20211006151940.214590-7-hreitz@redhat.com>
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Callers should be able to specify whether they want job_cancel_sync() to
force-cancel the job or not.
In fact, almost all invocations do not care about consistency of the
result and just want the job to terminate as soon as possible, so they
should pass force=true. The replication block driver is the exception,
specifically the active commit job it runs.
As for job_cancel_sync_all(), all callers want it to force-cancel all
jobs, because that is the point of it: To cancel all remaining jobs as
quickly as possible (generally on process termination). So make it
invoke job_cancel_sync() with force=true.
This changes some iotest outputs, because quitting qemu while a mirror
job is active will now lead to it being cancelled instead of completed,
which is what we want. (Cancelling a READY mirror job with force=false
may take an indefinite amount of time, which we do not want when
quitting. If users want consistent results, they must have all jobs be
done before they quit qemu.)
Buglink: https://gitlab.com/qemu-project/qemu/-/issues/462
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20211006151940.214590-6-hreitz@redhat.com>
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
As of HEAD^, there is no meaning to s->synced other than whether the job
is READY or not. job_is_ready() gives us that information, too.
Suggested-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-Id: <20211006151940.214590-4-hreitz@redhat.com>
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
An error does not take us out of the READY phase, which is what
s->synced signifies. It does of course mean that source and target are
no longer in sync, but that is what s->actively_sync is for -- s->synced
never meant that source and target are in sync, only that they were at
some point (and at that point we transitioned into the READY phase).
The tangible problem is that we transition to READY once we are in sync
and s->synced is false. By resetting s->synced here, we will transition
from READY to READY once the error is resolved (if the job keeps
running), and that transition is not allowed.
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-Id: <20211006151940.214590-3-hreitz@redhat.com>
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Linux limits the size of iovecs to 1024 (UIO_MAXIOV in the kernel
sources, IOV_MAX in POSIX). Because of this, on some host adapters
requests with many iovecs are rejected with -EINVAL by the
io_submit() or readv()/writev() system calls.
In fact, the same limit applies to SG_IO as well. To fix both the
EINVAL and the possible performance issues from using fewer iovecs
than allowed by Linux (some HBAs have max_segments as low as 128),
introduce a separate entry in BlockLimits to hold the max_segments
value from sysfs. This new limit is used only for SG_IO and clamped
to bs->bl.max_iov anyway, just like max_hw_transfer is clamped to
bs->bl.max_transfer.
Reported-by: Halil Pasic <pasic@linux.ibm.com>
Cc: Hanna Reitz <hreitz@redhat.com>
Cc: Kevin Wolf <kwolf@redhat.com>
Cc: qemu-block@nongnu.org
Cc: qemu-stable@nongnu.org
Fixes: 18473467d5 ("file-posix: try BLKSECTGET on block devices too, do not round to power of 2", 2021-06-25)
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20210923130436.1187591-1-pbonzini@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
All code in block/aio_task.c expects `max_busy_tasks` to always
be greater than 0.
Assert this condition during the AioTaskPool creation where
`max_busy_tasks` is set.
Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20211005161157.282396-3-sgarzare@redhat.com>
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
QAPI generates `struct BackupPerf` where `max-workers` value is stored
in an `int64_t` variable.
But block_copy_async(), and the underlying code, uses an `int` parameter.
At the end that variable is used to initialize `max_busy_tasks` in
block/aio_task.c causing the following assertion failure if a value
greater than INT_MAX(2147483647) is used:
../block/aio_task.c:63: aio_task_pool_wait_one: Assertion `pool->busy_tasks > 0' failed.
Let's check that `max-workers` doesn't exceed INT_MAX and print an
error in that case.
Buglink: https://bugzilla.redhat.com/show_bug.cgi?id=2009310
Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20211005161157.282396-2-sgarzare@redhat.com>
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
If we don't have active request, that waiting for this handle to be
received, we should report an error.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20210902103805.25686-6-vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
OK, that's a big rewrite of the logic.
Pre-patch we have an always running coroutine - connection_co. It does
reply receiving and reconnecting. And it leads to a lot of difficult
and unobvious code around drained sections and context switch. We also
abuse bs->in_flight counter which is increased for connection_co and
temporary decreased in points where we want to allow drained section to
begin. One of these place is in another file: in nbd_read_eof() in
nbd/client.c.
We also cancel reconnect and requests waiting for reconnect on drained
begin which is not correct. And this patch fixes that.
Let's finally drop this always running coroutine and go another way:
do both reconnect and receiving in request coroutines.
The detailed list of changes below (in the sequence of diff hunks).
1. receiving coroutines are woken directly from nbd_channel_error, when
we change s->state
2. nbd_co_establish_connection_cancel(): we don't have drain_begin now,
and in nbd_teardown_connection() all requests should already be
finished (and reconnect is done from request). So
nbd_co_establish_connection_cancel() is called from
nbd_cancel_in_flight() (to cancel the request that is doing
nbd_co_establish_connection()) and from reconnect_delay_timer_cb()
(previously we didn't need it, as reconnect delay only should cancel
active requests not the reconnection itself). But now reconnection
itself is done in the separate thread (we now call
nbd_client_connection_enable_retry() in nbd_open()), and we need to
cancel the requests that wait in nbd_co_establish_connection()
now).
2A. We do receive headers in request coroutine. But we also should
dispatch replies for other pending requests. So,
nbd_connection_entry() is turned into nbd_receive_replies(), which
does reply dispatching while it receives other request headers, and
returns when it receives the requested header.
3. All old staff around drained sections and context switch is dropped.
In details:
- we don't need to move connection_co to new aio context, as we
don't have connection_co anymore
- we don't have a fake "request" of connection_co (extra increasing
in_flight), so don't care with it in drain_begin/end
- we don't stop reconnection during drained section anymore. This
means that drain_begin may wait for a long time (up to
reconnect_delay). But that's an improvement and more correct
behavior see below[*]
4. In nbd_teardown_connection() we don't have to wait for
connection_co, as it is dropped. And cleanup for s->ioc and nbd_yank
is moved here from removed connection_co.
5. In nbd_co_do_establish_connection() we now should handle
NBD_CLIENT_CONNECTING_NOWAIT: if new request comes when we are in
NBD_CLIENT_CONNECTING_NOWAIT, it still should call
nbd_co_establish_connection() (who knows, maybe the connection was
already established by another thread in the background). But we
shouldn't wait: if nbd_co_establish_connection() can't return new
channel immediately the request should fail (we are in
NBD_CLIENT_CONNECTING_NOWAIT state).
6. nbd_reconnect_attempt() is simplified: it's now easier to wait for
other requests in the caller, so here we just assert that fact.
Also delay time is now initialized here: we can easily detect first
attempt and start a timer.
7. nbd_co_reconnect_loop() is dropped, we don't need it. Reconnect
retries are fully handle by thread (nbd/client-connection.c), delay
timer we initialize in nbd_reconnect_attempt(), we don't have to
bother with s->drained and friends. nbd_reconnect_attempt() now
called from nbd_co_send_request().
8. nbd_connection_entry is dropped: reconnect is now handled by
nbd_co_send_request(), receiving reply is now handled by
nbd_receive_replies(): all handled from request coroutines.
9. So, welcome new nbd_receive_replies() called from request coroutine,
that receives reply header instead of nbd_connection_entry().
Like with sending requests, only one coroutine may receive in a
moment. So we introduce receive_mutex, which is locked around
nbd_receive_reply(). It also protects some related fields. Still,
full audit of thread-safety in nbd driver is a separate task.
New function waits for a reply with specified handle being received
and works rather simple:
Under mutex:
- if current handle is 0, do receive by hand. If another handle
received - switch to other request coroutine, release mutex and
yield. Otherwise return success
- if current handle == requested handle, we are done
- otherwise, release mutex and yield
10: in nbd_co_send_request() we now do nbd_reconnect_attempt() if
needed. Also waiting in free_sema queue we now wait for one of two
conditions:
- connectED, in_flight < MAX_NBD_REQUESTS (so we can start new one)
- connectING, in_flight == 0, so we can call
nbd_reconnect_attempt()
And this logic is protected by s->send_mutex
Also, on failure we don't have to care of removed s->connection_co
11. nbd_co_do_receive_one_chunk(): now instead of yield() and wait for
s->connection_co we just call new nbd_receive_replies().
12. nbd_co_receive_one_chunk(): place where s->reply.handle becomes 0,
which means that handling of the whole reply is finished. Here we
need to wake one of coroutines sleeping in nbd_receive_replies().
If none are sleeping - do nothing. That's another behavior change: we
don't have endless recv() in the idle time. It may be considered as
a drawback. If so, it may be fixed later.
13. nbd_reply_chunk_iter_receive(): don't care about removed
connection_co, just ping in_flight waiters.
14. Don't create connection_co, enable retry in the connection thread
(we don't have own reconnect loop anymore)
15. We now need to add a nbd_co_establish_connection_cancel() call in
nbd_cancel_in_flight(), to cancel the request that is doing a
connection attempt.
[*], ok, now we don't cancel reconnect on drain begin. That's correct:
reconnect feature leads to possibility of long-running requests (up
to reconnect delay). Still, drain begin is not a reason to kill
long requests. We should wait for them.
This also means, that we can again reproduce a dead-lock, described
in 8c517de24a.
Why we are OK with it:
1. Now this is not absolutely-dead dead-lock: the vm is unfrozen
after reconnect delay. Actually 8c517de24a fixed a bug in
NBD logic, that was not described in 8c517de24a and led to
forever dead-lock. The problem was that nobody woke the free_sema
queue, but drain_begin can't finish until there is a request in
free_sema queue. Now we have a reconnect delay timer that works
well.
2. It's not a problem of the NBD driver, but of the ide code,
because it does drain_begin under the global mutex; the problem
doesn't reproduce when using scsi instead of ide.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20210902103805.25686-5-vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
[eblake: grammar and comment tweaks]
Signed-off-by: Eric Blake <eblake@redhat.com>
Split out nbd_recv_coroutine_wake_one(), as it will be used
separately.
Rename the function and add a possibility to wake only first found
sleeping coroutine.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20210902103805.25686-4-vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
[eblake: grammar tweak]
Signed-off-by: Eric Blake <eblake@redhat.com>
We are going to use it in nbd_channel_error(), so move it up. Note,
that we are going also refactor and rename
nbd_recv_coroutines_wake_all() in future anyway, so keeping it where it
is and making forward declaration doesn't make real sense.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20210902103805.25686-3-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Don't rely on connection being totally broken in case of -EIO. Safer
and more correct is to just shut down the channel anyway, since we
change the state and plan on reconnecting.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20210902103805.25686-2-vsementsov@virtuozzo.com>
[eblake: grammar tweaks]
Signed-off-by: Eric Blake <eblake@redhat.com>
Now that all drivers are updated by the previous commit, we can drop
the last limiter on pdiscard path: INT_MAX in bdrv_co_pdiscard().
Now everything is prepared for implementing incredibly cool and fast
big-discard requests in NBD and qcow2. And any other driver which wants
it of course.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20210903102807.27127-12-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
We are generally moving to int64_t for both offset and bytes parameters
on all io paths.
Main motivation is realization of 64-bit write_zeroes operation for
fast zeroing large disk chunks, up to the whole disk.
We chose signed type, to be consistent with off_t (which is signed) and
with possibility for signed return type (where negative value means
error).
So, convert driver discard handlers bytes parameter to int64_t.
The only caller of all updated function is bdrv_co_pdiscard in
block/io.c. It is already prepared to work with 64bit requests, but
pass at most max(bs->bl.max_pdiscard, INT_MAX) to the driver.
Let's look at all updated functions:
blkdebug: all calculations are still OK, thanks to
bdrv_check_qiov_request().
both rule_check and bdrv_co_pdiscard are 64bit
blklogwrites: pass to blk_loc_writes_co_log which is 64bit
blkreplay, copy-on-read, filter-compress: pass to bdrv_co_pdiscard, OK
copy-before-write: pass to bdrv_co_pdiscard which is 64bit and to
cbw_do_copy_before_write which is 64bit
file-posix: one handler calls raw_account_discard() is 64bit and both
handlers calls raw_do_pdiscard(). Update raw_do_pdiscard, which pass
to RawPosixAIOData::aio_nbytes, which is 64bit (and calls
raw_account_discard())
gluster: somehow, third argument of glfs_discard_async is size_t.
Let's set max_pdiscard accordingly.
iscsi: iscsi_allocmap_set_invalid is 64bit,
!is_byte_request_lun_aligned is 64bit.
list.num is uint32_t. Let's clarify max_pdiscard and
pdiscard_alignment.
mirror_top: pass to bdrv_mirror_top_do_write() which is
64bit
nbd: protocol limitation. max_pdiscard is alredy set strict enough,
keep it as is for now.
nvme: buf.nlb is uint32_t and we do shift. So, add corresponding limits
to nvme_refresh_limits().
preallocate: pass to bdrv_co_pdiscard() which is 64bit.
rbd: pass to qemu_rbd_start_co() which is 64bit.
qcow2: calculations are still OK, thanks to bdrv_check_qiov_request(),
qcow2_cluster_discard() is 64bit.
raw-format: raw_adjust_offset() is 64bit, bdrv_co_pdiscard too.
throttle: pass to bdrv_co_pdiscard() which is 64bit and to
throttle_group_co_io_limits_intercept() which is 64bit as well.
test-block-iothread: bytes argument is unused
Great! Now all drivers are prepared to handle 64bit discard requests,
or else have explicit max_pdiscard limits.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20210903102807.27127-11-vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
We are going to support 64 bit discard requests. Now update the
limit variable. It's absolutely safe. The variable is set in some
drivers, and used in bdrv_co_pdiscard().
Update also max_pdiscard variable in bdrv_co_pdiscard(), so that
bdrv_co_pdiscard() is now prepared for 64bit requests. The remaining
logic including num, offset and bytes variables is already
supporting 64bit requests.
So the only thing that prevents 64 bit requests is limiting
max_pdiscard variable to INT_MAX in bdrv_co_pdiscard().
We'll drop this limitation after updating all block drivers.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20210903102807.27127-10-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Now that all drivers are updated by previous commit, we can drop two
last limiters on write-zeroes path: INT_MAX in
bdrv_co_do_pwrite_zeroes() and bdrv_check_request32() in
bdrv_co_pwritev_part().
Now everything is prepared for implementing incredibly cool and fast
big-write-zeroes in NBD and qcow2. And any other driver which wants it
of course.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20210903102807.27127-9-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
We are generally moving to int64_t for both offset and bytes parameters
on all io paths.
Main motivation is realization of 64-bit write_zeroes operation for
fast zeroing large disk chunks, up to the whole disk.
We chose signed type, to be consistent with off_t (which is signed) and
with possibility for signed return type (where negative value means
error).
So, convert driver write_zeroes handlers bytes parameter to int64_t.
The only caller of all updated function is bdrv_co_do_pwrite_zeroes().
bdrv_co_do_pwrite_zeroes() itself is of course OK with widening of
callee parameter type. Also, bdrv_co_do_pwrite_zeroes()'s
max_write_zeroes is limited to INT_MAX. So, updated functions all are
safe, they will not get "bytes" larger than before.
Still, let's look through all updated functions, and add assertions to
the ones which are actually unprepared to values larger than INT_MAX.
For these drivers also set explicit max_pwrite_zeroes limit.
Let's go:
blkdebug: calculations can't overflow, thanks to
bdrv_check_qiov_request() in generic layer. rule_check() and
bdrv_co_pwrite_zeroes() both have 64bit argument.
blklogwrites: pass to blk_log_writes_co_log() with 64bit argument.
blkreplay, copy-on-read, filter-compress: pass to
bdrv_co_pwrite_zeroes() which is OK
copy-before-write: Calls cbw_do_copy_before_write() and
bdrv_co_pwrite_zeroes, both have 64bit argument.
file-posix: both handler calls raw_do_pwrite_zeroes, which is updated.
In raw_do_pwrite_zeroes() calculations are OK due to
bdrv_check_qiov_request(), bytes go to RawPosixAIOData::aio_nbytes
which is uint64_t.
Check also where that uint64_t gets handed:
handle_aiocb_write_zeroes_block() passes a uint64_t[2] to
ioctl(BLKZEROOUT), handle_aiocb_write_zeroes() calls do_fallocate()
which takes off_t (and we compile to always have 64-bit off_t), as
does handle_aiocb_write_zeroes_unmap. All look safe.
gluster: bytes go to GlusterAIOCB::size which is int64_t and to
glfs_zerofill_async works with off_t.
iscsi: Aha, here we deal with iscsi_writesame16_task() that has
uint32_t num_blocks argument and iscsi_writesame16_task() has
uint16_t argument. Make comments, add assertions and clarify
max_pwrite_zeroes calculation.
iscsi_allocmap_() functions already has int64_t argument
is_byte_request_lun_aligned is simple to update, do it.
mirror_top: pass to bdrv_mirror_top_do_write which has uint64_t
argument
nbd: Aha, here we have protocol limitation, and NBDRequest::len is
uint32_t. max_pwrite_zeroes is cleanly set to 32bit value, so we are
OK for now.
nvme: Again, protocol limitation. And no inherent limit for
write-zeroes at all. But from code that calculates cdw12 it's obvious
that we do have limit and alignment. Let's clarify it. Also,
obviously the code is not prepared to handle bytes=0. Let's handle
this case too.
trace events already 64bit
preallocate: pass to handle_write() and bdrv_co_pwrite_zeroes(), both
64bit.
rbd: pass to qemu_rbd_start_co() which is 64bit.
qcow2: offset + bytes and alignment still works good (thanks to
bdrv_check_qiov_request()), so tail calculation is OK
qcow2_subcluster_zeroize() has 64bit argument, should be OK
trace events updated
qed: qed_co_request wants int nb_sectors. Also in code we have size_t
used for request length which may be 32bit. So, let's just keep
INT_MAX as a limit (aligning it down to pwrite_zeroes_alignment) and
don't care.
raw-format: Is OK. raw_adjust_offset and bdrv_co_pwrite_zeroes are both
64bit.
throttle: Both throttle_group_co_io_limits_intercept() and
bdrv_co_pwrite_zeroes() are 64bit.
vmdk: pass to vmdk_pwritev which is 64bit
quorum: pass to quorum_co_pwritev() which is 64bit
Hooray!
At this point all block drivers are prepared to support 64bit
write-zero requests, or have explicitly set max_pwrite_zeroes.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20210903102807.27127-8-vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
[eblake: use <= rather than < in assertions relying on max_pwrite_zeroes]
Signed-off-by: Eric Blake <eblake@redhat.com>
We are going to support 64 bit write-zeroes requests. Now update the
limit variable. It's absolutely safe. The variable is set in some
drivers, and used in bdrv_co_do_pwrite_zeroes().
Update also max_write_zeroes variable in bdrv_co_do_pwrite_zeroes(), so
that bdrv_co_do_pwrite_zeroes() is now prepared to 64bit requests. The
remaining logic including num, offset and bytes variables is already
supporting 64bit requests.
So the only thing that prevents 64 bit requests is limiting
max_write_zeroes variable to INT_MAX in bdrv_co_do_pwrite_zeroes().
We'll drop this limitation after updating all block drivers.
Ah, we also have bdrv_check_request32() in bdrv_co_pwritev_part(). It
will be modified to do bdrv_check_request() for write-zeroes path.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20210903102807.27127-7-vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
We are generally moving to int64_t for both offset and bytes parameters
on all io paths.
Main motivation is realization of 64-bit write_zeroes operation for
fast zeroing large disk chunks, up to the whole disk.
We chose signed type, to be consistent with off_t (which is signed) and
with possibility for signed return type (where negative value means
error).
So, convert driver copy_range handlers parameters which are already
64bit to signed type.
Now let's consider all callers. Simple
git grep '\->bdrv_co_copy_range'
shows the only caller:
bdrv_co_copy_range_internal(), which does bdrv_check_request32(),
so everything is OK.
Still, the functions may be called directly, not only by drv->...
Let's check:
git grep '\.bdrv_co_copy_range_\(from\|to\)\s*=' | \
awk '{print $4}' | sed 's/,//' | sed 's/&//' | sort | uniq | \
while read func; do git grep "$func(" | \
grep -v "$func(BlockDriverState"; done
shows no more callers. So, we are done.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20210903102807.27127-6-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
We are generally moving to int64_t for both offset and bytes parameters
on all io paths.
Main motivation is realization of 64-bit write_zeroes operation for
fast zeroing large disk chunks, up to the whole disk.
We chose signed type, to be consistent with off_t (which is signed) and
with possibility for signed return type (where negative value means
error).
So, convert driver write handlers parameters which are already 64bit to
signed type.
While being here, convert also flags parameter to be BdrvRequestFlags.
Now let's consider all callers. Simple
git grep '\->bdrv_\(aio\|co\)_pwritev\(_part\)\?'
shows that's there three callers of driver function:
bdrv_driver_pwritev() and bdrv_driver_pwritev_compressed() in
block/io.c, both pass int64_t, checked by bdrv_check_qiov_request() to
be non-negative.
qcow2_save_vmstate() does bdrv_check_qiov_request().
Still, the functions may be called directly, not only by drv->...
Let's check:
git grep '\.bdrv_\(aio\|co\)_pwritev\(_part\)\?\s*=' | \
awk '{print $4}' | sed 's/,//' | sed 's/&//' | sort | uniq | \
while read func; do git grep "$func(" | \
grep -v "$func(BlockDriverState"; done
shows several callers:
qcow2:
qcow2_co_truncate() write at most up to @offset, which is checked in
generic qcow2_co_truncate() by bdrv_check_request().
qcow2_co_pwritev_compressed_task() pass the request (or part of the
request) that already went through normal write path, so it should
be OK
qcow:
qcow_co_pwritev_compressed() pass int64_t, it's updated by this patch
quorum:
quorum_co_pwrite_zeroes() pass int64_t and int - OK
throttle:
throttle_co_pwritev_compressed() pass int64_t, it's updated by this
patch
vmdk:
vmdk_co_pwritev_compressed() pass int64_t, it's updated by this
patch
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20210903102807.27127-5-vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
We are generally moving to int64_t for both offset and bytes parameters
on all io paths.
Main motivation is realization of 64-bit write_zeroes operation for
fast zeroing large disk chunks, up to the whole disk.
We chose signed type, to be consistent with off_t (which is signed) and
with possibility for signed return type (where negative value means
error).
So, convert driver read handlers parameters which are already 64bit to
signed type.
While being here, convert also flags parameter to be BdrvRequestFlags.
Now let's consider all callers. Simple
git grep '\->bdrv_\(aio\|co\)_preadv\(_part\)\?'
shows that's there three callers of driver function:
bdrv_driver_preadv() in block/io.c, passes int64_t, checked by
bdrv_check_qiov_request() to be non-negative.
qcow2_load_vmstate() does bdrv_check_qiov_request().
do_perform_cow_read() has uint64_t argument. And a lot of things in
qcow2 driver are uint64_t, so converting it is big job. But we must
not work with requests that don't satisfy bdrv_check_qiov_request(),
so let's just assert it here.
Still, the functions may be called directly, not only by drv->...
Let's check:
git grep '\.bdrv_\(aio\|co\)_preadv\(_part\)\?\s*=' | \
awk '{print $4}' | sed 's/,//' | sed 's/&//' | sort | uniq | \
while read func; do git grep "$func(" | \
grep -v "$func(BlockDriverState"; done
The only one such caller:
QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, &data, 1);
...
ret = bdrv_replace_test_co_preadv(bs, 0, 1, &qiov, 0);
in tests/unit/test-bdrv-drain.c, and it's OK obviously.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20210903102807.27127-4-vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
[eblake: fix typos]
Signed-off-by: Eric Blake <eblake@redhat.com>
We modify the request by adding an offset to vmstate. Let's check the
modified request. It will help us to safely move .bdrv_co_preadv_part
and .bdrv_co_pwritev_part to int64_t type of offset and bytes.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20210903102807.27127-3-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Only qcow2 driver supports vmstate.
In qcow2 these requests go through .bdrv_co_p{read,write}v_part
handlers.
So, let's do our basic check for the request on vmstate generic
handlers.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-Id: <20210903102807.27127-2-vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Split checking for reserved bits out of aligned offset check.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Tested-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Hanna Reitz <hreitz@redhat.com>
Message-Id: <20210914122454.141075-11-vsementsov@virtuozzo.com>
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
- use g_autofree for l1_table
- better name for size in bytes variable
- reduce code blocks nesting
- whitespaces, braces, newlines
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Hanna Reitz <hreitz@redhat.com>
Message-Id: <20210914122454.141075-9-vsementsov@virtuozzo.com>
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Check subcluster bitmap of the l2 entry for different types of
clusters:
- for compressed it must be zero
- for allocated check consistency of two parts of the bitmap
- for unallocated all subclusters should be unallocated
(or zero-plain)
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Tested-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Message-Id: <20210914122454.141075-7-vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Hanna Reitz <hreitz@redhat.com>
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
We'll reuse the function to fix wrong L2 entry bitmap. Support it now.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Hanna Reitz <hreitz@redhat.com>
Message-Id: <20210914122454.141075-6-vsementsov@virtuozzo.com>
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Split fix_l2_entry_by_zero() out of check_refcounts_l2() to be
reused in further patch.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Hanna Reitz <hreitz@redhat.com>
Message-Id: <20210914122454.141075-5-vsementsov@virtuozzo.com>
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Add helper to parse compressed l2_entry and use it everywhere instead
of open-coding.
Note, that in most places we move to precise coffset/csize instead of
sector-aligned. Still it should work good enough for updating
refcounts.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Hanna Reitz <hreitz@redhat.com>
Message-Id: <20210914122454.141075-4-vsementsov@virtuozzo.com>
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Let's pass the whole L2 entry and not bother with
L2E_COMPRESSED_OFFSET_SIZE_MASK.
It also helps further refactoring that adds generic
qcow2_parse_compressed_l2_entry() helper.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Alberto Garcia <berto@igalia.com>
Reviewed-by: Hanna Reitz <hreitz@redhat.com>
Message-Id: <20210914122454.141075-3-vsementsov@virtuozzo.com>
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
- don't use same name for size in bytes and in entries
- use g_autofree for l2_table
- add whitespace
- fix block comment style
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Hanna Reitz <hreitz@redhat.com>
Message-Id: <20210914122454.141075-2-vsementsov@virtuozzo.com>
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
There is no conflict and no dependency if we have parallel writes to
different subclusters of one cluster when the cluster itself is already
allocated. So, relax extra dependency.
Measure performance:
First, prepare build/qemu-img-old and build/qemu-img-new images.
cd scripts/simplebench
./img_bench_templater.py
Paste the following to stdin of running script:
qemu_img=../../build/qemu-img-{old|new}
$qemu_img create -f qcow2 -o extended_l2=on /ssd/x.qcow2 1G
$qemu_img bench -c 100000 -d 8 [-s 2K|-s 2K -o 512|-s $((1024*2+512))] \
-w -t none -n /ssd/x.qcow2
The result:
All results are in seconds
------------------ --------- ---------
old new
-s 2K 6.7 ± 15% 6.2 ± 12%
-7%
-s 2K -o 512 13 ± 3% 11 ± 5%
-16%
-s $((1024*2+512)) 9.5 ± 4% 8.4
-12%
------------------ --------- ---------
So small writes are more independent now and that helps to keep deeper
io queue which improves performance.
271 iotest output becomes racy for three allocation in one cluster.
Second and third writes may finish in different order. Second and
third requests don't depend on each other any more. Still they both
depend on first request anyway. Filter out second and third write
offsets to cover both possible outputs.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20210824101517.59802-4-vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Hanna Reitz <hreitz@redhat.com>
[hreitz: s/ an / and /]
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
No logic change, just prepare for the following commit. While being
here do also small grammar fix in a comment.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Hanna Reitz <hreitz@redhat.com>
Message-Id: <20210824101517.59802-3-vsementsov@virtuozzo.com>
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
In mirror_iteration() we call mirror_wait_on_conflicts() with
`self` parameter set to NULL.
Starting from commit d44dae1a7c we dereference `self` pointer in
mirror_wait_on_conflicts() without checks if it is not NULL.
Backtrace:
Program terminated with signal SIGSEGV, Segmentation fault.
#0 mirror_wait_on_conflicts (self=0x0, s=<optimized out>, offset=<optimized out>, bytes=<optimized out>)
at ../block/mirror.c:172
172 self->waiting_for_op = op;
[Current thread is 1 (Thread 0x7f0908931ec0 (LWP 380249))]
(gdb) bt
#0 mirror_wait_on_conflicts (self=0x0, s=<optimized out>, offset=<optimized out>, bytes=<optimized out>)
at ../block/mirror.c:172
#1 0x00005610c5d9d631 in mirror_run (job=0x5610c76a2c00, errp=<optimized out>) at ../block/mirror.c:491
#2 0x00005610c5d58726 in job_co_entry (opaque=0x5610c76a2c00) at ../job.c:917
#3 0x00005610c5f046c6 in coroutine_trampoline (i0=<optimized out>, i1=<optimized out>)
at ../util/coroutine-ucontext.c:173
#4 0x00007f0909975820 in ?? () at ../sysdeps/unix/sysv/linux/x86_64/__start_context.S:91
from /usr/lib64/libc.so.6
Buglink: https://bugzilla.redhat.com/show_bug.cgi?id=2001404
Fixes: d44dae1a7c ("block/mirror: fix active mirror dead-lock in mirror_wait_on_conflicts")
Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Message-Id: <20210910124533.288318-1-sgarzare@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
bdrv_co_block_status() does it for us, we do not need to do it here.
The advantage of not capping *pnum is that bdrv_co_block_status() can
cache larger data regions than requested by its caller.
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-Id: <20210812084148.14458-7-hreitz@redhat.com>
bdrv_co_block_status() does it for us, we do not need to do it here.
The advantage of not capping *pnum is that bdrv_co_block_status() can
cache larger data regions than requested by its caller.
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-Id: <20210812084148.14458-6-hreitz@redhat.com>
bdrv_co_block_status() does it for us, we do not need to do it here.
The advantage of not capping *pnum is that bdrv_co_block_status() can
cache larger data regions than requested by its caller.
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Kevin Wolf <kwolf@redhat.com>
Message-Id: <20210812084148.14458-5-hreitz@redhat.com>
As we have attempted before
(https://lists.gnu.org/archive/html/qemu-devel/2019-01/msg06451.html,
"file-posix: Cache lseek result for data regions";
https://lists.nongnu.org/archive/html/qemu-block/2021-02/msg00934.html,
"file-posix: Cache next hole"), this patch seeks to reduce the number of
SEEK_DATA/HOLE operations the file-posix driver has to perform. The
main difference is that this time it is implemented as part of the
general block layer code.
The problem we face is that on some filesystems or in some
circumstances, SEEK_DATA/HOLE is unreasonably slow. Given the
implementation is outside of qemu, there is little we can do about its
performance.
We have already introduced the want_zero parameter to
bdrv_co_block_status() to reduce the number of SEEK_DATA/HOLE calls
unless we really want zero information; but sometimes we do want that
information, because for files that consist largely of zero areas,
special-casing those areas can give large performance boosts. So the
real problem is with files that consist largely of data, so that
inquiring the block status does not gain us much performance, but where
such an inquiry itself takes a lot of time.
To address this, we want to cache data regions. Most of the time, when
bad performance is reported, it is in places where the image is iterated
over from start to end (qemu-img convert or the mirror job), so a simple
yet effective solution is to cache only the current data region.
(Note that only caching data regions but not zero regions means that
returning false information from the cache is not catastrophic: Treating
zeroes as data is fine. While we try to invalidate the cache on zero
writes and discards, such incongruences may still occur when there are
other processes writing to the image.)
We only use the cache for nodes without children (i.e. protocol nodes),
because that is where the problem is: Drivers that rely on block-status
implementations outside of qemu (e.g. SEEK_DATA/HOLE).
Resolves: https://gitlab.com/qemu-project/qemu/-/issues/307
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Message-Id: <20210812084148.14458-3-hreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
[hreitz: Added `local_file == bs` assertion, as suggested by Vladimir]
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
gluster's block-status implementation is basically a copy of that in
block/file-posix.c, there is only one thing missing, and that is
aligning trailing data extents to the request alignment (as added by
commit 9c3db310ff).
Note that 9c3db310ff mentions that "there seems to be no other block
driver that sets request_alignment and [...]", but while block/gluster.c
does indeed not set request_alignment, block/io.c's
bdrv_refresh_limits() will still default to an alignment of 512 because
block/gluster.c does not provide a byte-aligned read function.
Therefore, unaligned tails can conceivably occur, and so we should apply
the change from 9c3db310ff to gluster's block-status implementation.
Reported-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
Message-Id: <20210805143603.59503-1-mreitz@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Hanna Reitz <hreitz@redhat.com>
Paolo Bonzini
Richard Henderson
Hanna Reitz
Stefano Garzarella
Vladimir Sementsov-Ogievskiy
Peter Maydell
Max Reitz
Michael Tokarev