Re: [PATCH v4 5/5] qcow2: repair a dirty image when it becomes writable

[Markus Armbruster <armbru@redhat.com>]

"Denis V. Lunev" <den@virtuozzo.com> writes:

> On 8/25/26 11:37, Markus Armbruster wrote:
>> "Denis V. Lunev" <den@openvz.org> writes:
>>
>>> From: Denis V. Lunev <den@openvz.org>
>>>
>>> A dirty image must be repaired before anything allocates a cluster in
>>> it. qcow2_do_open() does that, but only for a node that is writable
>>> from the start. A node opened read-only skips it, and nothing revisits
>>> the question once that node becomes writable, which block-commit does
>>> routinely: commit_active_start() and commit_start() reopen the base
>>> read-write for the duration of the job.
>>>
>>> With lazy refcounts the on-disk refcount block then still accounts for
>>> the metadata clusters only, so the allocator restarts at the front of
>>> the image and hands out clusters that L2 entries point at. Two guest
>>> offsets end up sharing one host cluster. Nothing fails, the corrupt bit
>>> stays clear, and a clean close clears the dirty bit, so no later open
>>> repairs the image either. The bit also stays set for the whole writable
>>> session, so a node which is merely writable says nothing.
>>>
>>> Refusing the reopen instead is simpler and keeps it atomic, but it
>>> leaves nowhere to go: the base belongs to a chain the VM has open, so
>>> the qemu-img check -r such an error would ask for cannot take the write
>>> lock it needs. The repair does the trick in most cases anyway.
>>>
>>> Do the repair in qcow2_reopen_commit_post(), the earliest point where
>>> the node is writable. An inactive node is skipped: bdrv_activate() calls
>>> qcow2_do_open() again through qcow2_co_invalidate_cache().
>>>
>>> commit_post cannot reject the reopen, so a failed repair takes the
>>> driver away from the node instead, which is what stops writes from
>>> aliasing live clusters. qcow2_signal_corruption() does that as well,
>>> but it also sends BLOCK_IMAGE_CORRUPTED and sets the corrupt bit,
>>> which qcow2_do_open() honours by refusing every later read-write open.
>>> The image is dirty and unrepaired, not corrupt, and qemu-img check -r
>>> still fixes it, so neither belongs here. Return the error and skip
>>> the bitmaps.
>>>
>>> Signed-off-by: Denis V. Lunev <den@openvz.org>
>>> Reviewed-by: Andrey Drobyshev <andrey.drobyshev@virtuozzo.com>
>>> CC: Kevin Wolf <kwolf@redhat.com>
>>> CC: Hanna Reitz <hreitz@redhat.com>
>>> CC: Eric Blake <eblake@redhat.com>
>>> CC: Markus Armbruster <armbru@redhat.com>
>>> CC: Andrey Drobyshev <andrey.drobyshev@virtuozzo.com>
>>> Cc: qemu-stable@nongnu.org
>> [...]
>>
>>> diff --git a/qapi/block-core.json b/qapi/block-core.json
>>> index 199efc1e00..940249a5e5 100644
>>> --- a/qapi/block-core.json
>>> +++ b/qapi/block-core.json
>>> @@ -1852,6 +1852,9 @@
>>   ##
>>   # @change-backing-file:
>>   #
>>   # Change the backing file in the image file metadata.  This does not
>>   # cause QEMU to reopen the image file to reparse the backing filename
>>   # (it may, however, perform a reopen to change permissions from r/o ->
>>>  # r/w -> r/o, if needed).  The new backing file string is written into
>>>  # the image file metadata, and the QEMU internal strings are updated.
>>>  #
>>> +# A dirty qcow2 image is repaired during that reopen, which blocks
>>> +# other requests and can fail the command.
>>
>> Pardon my ignorance: what makes a qcow2 image dirty?
> usual obvious reasons are SIGKILL to qemu process (f.e. from OOM)
> or node crash.

So, you have to do some cleaning work before you can use it again, just
like a dirty filesystem.  Correct?

Back to change-backing-file.  It operates on an open image.  Cleaning
happens when that image is read-only and dirty.  Possible because you
can open dirty images read-only, and that doesn't clean them.  Correct?

>> Can you give me an idea of what other requests could be blocked?
> Before the patch reopen was smooth - we have just opened the
> file again. After this patch in a very unlikely corner case
> potentially lengthy procedure has been added - full image
> consistency check. Guest IO is stuck until the check will be
> completed.
>
> The case is unfortunately real for production.
>
>> Double-checking: "that reopen" is the one to change permissions,
>> i.e. the parenthesis above.  Correct?
> yes
>
>>> +#
>>>  # @image-node-name: The name of the block driver state node of the
>>>  #     image to modify.  The "device" argument is used to verify
>>>  #     "image-node-name" is in the chain described by "device".
>>> @@ -1891,6 +1894,9 @@
>>   ##
>>   # @block-commit:
>>   #
>>   # Live commit of data from overlay image nodes into backing nodes -
>>   # i.e., writes data between 'top' and 'base' into 'base'.
>>   #
>>   # If top == base, that is an error.  If top has no overlays on top of
>>   # it, or if it is in use by a writer, the job will not be completed by
>>   # itself.  The user needs to complete the job with the `job-complete`
>>   # command after getting the ready event.  (Since 2.0)
>>   #
>>   # If the base image is smaller than top, then the base image will be
>>   # resized to be the same size as top.  If top is smaller than the base
>>   # image, the base will not be truncated.  If you want the base image
>>>  # size to match the size of the smaller top, you can safely truncate
>>>  # it yourself once the commit operation successfully completes.
>>>  #
>>> +# The base is opened read-write.  A dirty qcow2 base is repaired

Reopend, I presume?

>>> +# first, which blocks other requests and can fail the command.
>>> +#
>>>  # @job-id: identifier for the newly-created block job.  If omitted,
>>>  #     the device name will be used.  (Since 2.7)
>>>  #

Like change-backing-file, block-commit operates on open images.  It
copies down into a base image.  If the base image is read-only, it is
reopened, and cleaning happens when it's dirty.  Correct?

Can it happen in any other way?

>>> @@ -2902,6 +2908,9 @@
>>   ##
>>   # @block-stream:
>>   #
>>   # Copy data from a backing file into a block device.
>>
>> [...]

Whereas block-commit copies down into a base image, block-stream copies
up from a base image.  If the image copied to is read-only, it is
reopened, and cleaning happens when it's dirty.  Correct?

Can it happen in any other way?

>>>  # On successful completion the image file is updated to drop the
>>>  # backing file and the `BLOCK_JOB_COMPLETED` event is emitted.
>>>  #
>>> +# The top image is opened read-write.  A dirty qcow2 image is repaired
>>> +# first, which blocks other requests and can fail the command.
>>> +#
>>>  # In case @device is a filter node, `block-stream` modifies the first
>>>  # non-filter overlay node below it to point to the new backing node
>>>  # instead of modifying @device itself.
>>> @@ -4989,6 +4998,16 @@
>>   ##
>>   # @blockdev-reopen:
>>   #
>>   # Reopens one or more block devices using the given set of options.
>>   # Any option not specified will be reset to its default value
>>   # regardless of its previous status.  If an option cannot be changed
>>   # or a particular driver does not support reopening then the command
>>   # will return an error.  All devices in the list are reopened in one
>>>  # transaction, so if one of them fails then the whole transaction is
>>>  # cancelled.

blockdev-reopen also operates on open images.  Cleaning happens when
reopening a dirty read-only image read/write.  Correct?

Can it happen in any other way?

>>>  #
>>> +# An error is also returned when a device cannot be used once it has
>>> +# been reopened.  Such a reopen is not undone, so an error does not
>>> +# always mean that nothing has changed.
>>
>> Could that be a problem?
> Error reply as itself is not a problem. The situation as a whole
> is a real pain in the ass.
>
> The problem is that such images exists in production and there
> is not way to handle this without downtime.
>
> Under this patch we are trying to fix things hard and this is
> correct thing to do - on RW image we are doing exactly the
> same thing. If automation is unable to fix - the guest denies
> starting.
>
> Here we report an error and render guest as unusable. This
> case is expected to be extremely rare but technically possible.
>
> This is a problem.

I'll come back to this as soon as I understand when exactly cleaning may
happen.

>>> +#                                        A node the driver gave up on
>>> +# serves nothing at all: it keeps its image open, makes
>>> +# `query-named-block-nodes` fail for as long as it is in the graph,
>>> +# and `blockdev-del` removes it only once nothing refers to it.
>>> +#
>>> +# Reopening a dirty qcow2 image read-write repairs it first, which
>>> +# blocks other requests and can fail the command.
>>> +#
>>>  # The command receives a list of block devices to reopen.  For each
>>>  # one of them, the top-level @node-name option (from
>>>  # `BlockdevOptions`) must be specified and is used to select the block
>>
>> These documentation updates suggest the patch affects commands
>> change-backing-file, block-commit, block-stream, and blockdev-reopen.
>> Is that correct?
>>
>> Would it make sense to list them in the commit message?
> That is simple thing. Will add :-)
>
> Thank you,
>     Den

Thanks!