This version includes both the AioContext lock and the block graph
lock, so there might be some deadlocks lurking. It's not possible to
disable the block graph lock like was done in QEMU 8.1, because there
are no changes like the function bdrv_schedule_unref() that require
it. QEMU 9.0 will finally get rid of the AioContext locking.
During live-restore with a VirtIO SCSI drive with iothread there is a
known racy deadlock related to the AioContext lock. Not new [1], but
not sure if more likely now. Should be fixed in QEMU 9.0.
The block graph lock comes with annotations that can be checked by
clang's TSA. This required changes to the block drivers, i.e.
alloc-track, pbs, zeroinit as well as taking the appropriate locks
in pve-backup, savevm-async, vma-reader.
Local variable shadowing is prohibited via a compiler flag now,
required slight adaptation in vma.c.
Major changes only affect alloc-track:
* It is not possible to call a generated co-wrapper like
bdrv_get_info() while holding the block graph lock exclusively [0],
which does happen during initialization of alloc-track when the
backing hd is set and the refresh_limits driver callback is invoked.
The bdrv_get_info() call to get the cluster size is moved to
directly after opening the file child in track_open().
The important thing is that at least the request alignment for the
write target is used, because then the RMW cycle in bdrv_pwritev
will gather enough data from the backing file. Partial cluster
allocations in the target are not a fundamental issue, because the
driver returns its allocation status based on the bitmap, so any
other data that maps to the same cluster will still be copied later
by a stream job (or during writes to that cluster).
* Replacing the node cannot be done in the
track_co_change_backing_file() callback, because it is a coroutine
and cannot hold the block graph lock exclusively. So it is moved to
the stream job itself with the auto-remove option not having an
effect anymore (qemu-server would always set it anyways).
In the future, there could either be a special option for the stream
job, or maybe the upcoming blockdev-replace QMP command can be used.
Replacing the backing child is actually already done in the stream
job, so no need to do it in the track_co_change_backing_file()
callback. It also cannot be called from a coroutine. Looking at the
implementation in the qcow2 driver, it doesn't seem to be intended
to change the backing child itself, just update driver-internal
state.
Other changes:
* alloc-track: Error out early when used without auto-remove. Since
replacing the node now happens in the stream job, where the option
cannot be read from (it's internal to the driver), it will always be
treated as 'on'. Makes sure to have users beside qemu-server notice
the change (should they even exist). The option can be fully dropped
in the future while adding a version guard in qemu-server.
* alloc-track: Avoid seemingly superfluous child permission update.
Doesn't seem necessary nowadays (maybe after commit "alloc-track:
fix deadlock during drop" where the dropping is not rescheduled and
delayed anymore or some upstream change). Replacing the block node
will already update the permissions of the new node (which was the
file child before). Should there really be some issue, instead of
having a drop state, this could also be just based off the fact
whether there is still a backing child.
Dumping the cumulative (shared) permissions for the BDS with a debug
print yields the same values after this patch and with QEMU 8.1,
namely 3 and 5.
* PBS block driver: compile unconditionally. Proxmox VE always needs
it and something in the build process changed to make it not enabled
by default. Probably would need to move the build option to meson
otherwise.
* backup: job unreferencing during cleanup needs to happen outside of
coroutine, so it was moved to before invoking the clean
* mirror: Cherry-pick stable fix to avoid potential deadlock.
* savevm-async: migrate_init now can fail, so propagate potential
error.
* savevm-async: compression counters are not accessible outside
migration/ram-compress now, so drop code that prophylactically set
it to zero.
[0]: https://lore.kernel.org/qemu-devel/220be383-3b0d-4938-b584-69ad214e5d5d@proxmox.com/
[1]: https://lore.kernel.org/qemu-devel/e13b488e-bf13-44f2-acca-e724d14f43fd@proxmox.com/
Signed-off-by: Fiona Ebner <f.ebner@proxmox.com>
With pvebackup_propagate_error(), the first error wins. When one job
in the transaction fails, it is expected that later jobs get the
ECANCELED error. Those are not interesting and by skipping them a more
interesting error, which is likely the actual root cause, can win.
Signed-off-by: Fiona Ebner <f.ebner@proxmox.com>
Most notable fixes from a Proxmox VE perspective are:
* "virtio-net: correctly copy vnet header when flushing TX"
To prevent a stack overflow that could lead to leaking parts of the
QEMU process's memory.
* "hw/pflash: implement update buffer for block writes"
To prevent an edge case for half-completed writes. This potentially
affected EFI disks.
* Fixes to i386 emulation and ARM emulation.
No changes for patches were necessary (all are just automatic context
changes).
Signed-off-by: Fiona Ebner <f.ebner@proxmox.com>
While the patch gives bdrv_graph_wrlock() as an example where the
issue can manifest, something similar can happen even when that is
disabled. Was able to reproduce the issue with
while true; do qm resize 115 scsi0 +4M; sleep 1; done
while running
fio --name=make-mirror-work --size=100M --direct=1 --rw=randwrite \
--bs=4k --ioengine=psync --numjobs=5 --runtime=1200 --time_based
in the VM.
Fix picked up from:
https://lists.nongnu.org/archive/html/qemu-devel/2023-12/msg01102.html
Signed-off-by: Fiona Ebner <f.ebner@proxmox.com>
Bigger notable changes:
* Commit 1a30b0f5d7 ("block: .bdrv_open is non-coroutine and
unlocked") broke the PVE backup patches, in particular setting up
the backup dump block driver, because bdrv_new_open_driver() cannot
be called from a coroutine. To fix it, bdrv_co_open() is used
instead, and while it's a much more involved function, the result
should be essentially the same. The only difference I noticed is
that the BDRV_O_ALLOW_RDWR flag is also set in the resulting bds
(block driver state), but that shouldn't hurt.
Smaller notable changes:
* aio_set_fd_handler() dropped its 'is_external' parameter stating
that all callers now pass false in 60f782b6b7 ("aio: remove
aio_disable_external() API"). The calls in the PVE patches also
passed false, so just drop the parameter too.
* global_state_store() does not have a return value anymore, so the
user in the PVE savevm-async patch was adapted. For context, see
c33f1829f8 ("migration: never fail in global_state_store()").
* Renames affecting the PVE savevm-async patch:
migrate_use_block() -> migrate_block() and ram_counters -> mig_stats
9d4b1e5f22 ("migration: Move migrate_use_block() to options.c")
aff3f6606d ("migration: Rename ram_counters to mig_stats")
Signed-off-by: Fiona Ebner <f.ebner@proxmox.com>
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
Since upstream QEMU 8.0, it's no longer possible to call
bdrv_img_create() from a coroutine anymore, meaning a backup with the
directory format would crash the QEMU instance.
The feature is only exposed via the monitor and was intended to be
experimental. There were no user reports about the breakage and it
only was noticed during the rebase for QEMU 8.1, because other parts
of the backup code needed adaptation and I decided to check the
BACKUP_FORMAT_DIR case too.
It should not stay in a broken state of course, but avoid the
maintenance cost and just make it a removed feature for Proxmox VE 8
retroactively.
Signed-off-by: Fiona Ebner <f.ebner@proxmox.com>
With the drive-backup QMP command, upstream QEMU uses a drained
section for the source drive when creating the backup job. Do the same
here to avoid subtle bugs.
There, the drained section extends until after the job is started, but
this cannot be done here for multi-disk backups (could at most start
the first job). The important thing is that the cbw
(copy-before-write) node is in place and the bcs (block-copy-state)
bitmap is initialized, which both happen during job creation (ensured
by the "block/backup: move bcs bitmap initialization to job creation"
PVE patch).
One such bug is one reported in the community forum [0], where using a
drive with iothread can lead to an overlapping block-copy request and
consequently an assertion failure. The block-copy code relies on the
bcs bitmap to determine if a request for a certain range can be
created. Each time a request is created, it resets the bcs bitmap at
that range to indicate that it's being handled.
The duplicate request can happen as follows:
Thread A attaches the cbw node
Thread B creates a request and resets the bitmap at that range
Thread A clears the bitmap and merges it with the PBS bitmap
The merging can lead to the bitmap being set again at the range of
the previous request, so the block-copy code thinks it's fine to
create a request there.
Thread B creates another requests at an overlapping range before the
other request is finished.
The drained section ensures that nothing else can interfere with the
bcs bitmap between attaching the copy-before-write block node and
initialization of the bitmap.
[0]: https://forum.proxmox.com/threads/133149/
Signed-off-by: Fiona Ebner <f.ebner@proxmox.com>
Reported in the community forum [0]. By default, there can be large
amounts of memory left assigned to the QEMU process after backup.
Likely because of fragmentation, it's necessary to explicitly call
malloc_trim() to tell glibc that it shouldn't keep all that memory
resident for the process.
QEMU itself already does a malloc_trim() in the RCU thread, but that
code path might not be reached (or not for a long time) under usual
operation. The value of 4 MiB for the argument was also copied from
there.
Example with the following configuration:
> agent: 1
> boot: order=scsi0
> cores: 4
> cpu: x86-64-v2-AES
> ide2: none,media=cdrom
> memory: 1024
> name: backup-mem
> net0: virtio=DA:58:18:26:59:9F,bridge=vmbr0,firewall=1
> numa: 0
> ostype: l26
> scsi0: rbd:base-107-disk-0/vm-106-disk-1,size=4302M
> scsihw: virtio-scsi-pci
> smbios1: uuid=b2d4511e-8d01-44f1-afd6-9581b30c24a6
> sockets: 2
> startup: order=2
> virtio0: lvmthin:vm-106-disk-1,iothread=1,size=1G
> virtio1: lvmthin:vm-106-disk-2,iothread=1,size=1G
> virtio2: lvmthin:vm-106-disk-3,iothread=1,size=1G
> vmgenid: 0a1d8751-5e02-449d-977e-c0160e900231
Before the change:
> root@pve8a1 ~ # grep VmRSS /proc/$(cat /var/run/qemu-server/106.pid)/status
> VmRSS: 370948 kB
> root@pve8a1 ~ # vzdump 106 --storage pbs
> (...)
> INFO: Backup job finished successfully
> root@pve8a1 ~ # grep VmRSS /proc/$(cat /var/run/qemu-server/106.pid)/status
> VmRSS: 2114964 kB
After the change:
> root@pve8a1 ~ # grep VmRSS /proc/$(cat /var/run/qemu-server/106.pid)/status
> VmRSS: 398788 kB
> root@pve8a1 ~ # vzdump 106 --storage pbs
> (...)
> INFO: Backup job finished successfully
> root@pve8a1 ~ # grep VmRSS /proc/$(cat /var/run/qemu-server/106.pid)/status
> VmRSS: 424356 kB
[0]: https://forum.proxmox.com/threads/131339/
Co-diagnosed-by: Friedrich Weber <f.weber@proxmox.com>
Co-diagnosed-by: Dominik Csapak <d.csapak@proxmox.com>
Signed-off-by: Fiona Ebner <f.ebner@proxmox.com>
Acked-by: Wolfgang Bumiller <w.bumiller@proxmox.com>
The QAPI change for QEMU 8.0 dropped redundant has_foo parameters, but
in the blockdev_mirror_common() function (which is not part of the
QAPI itself but called from there) the argument pair was has_bitmap
and bitmap_name rather than has_bitmap and bitmap.
Reported-by: Aaron Lauterer <a.lauterer@proxmox.com>
Signed-off-by: Fiona Ebner <f.ebner@proxmox.com>
