Re: [PATCH v3 0/8] blockdev: avoid acquiring AioContext lock twice at do_drive_backup and do_blockdev_backup

[Kevin Wolf]

Am 19.11.2019 um 12:35 hat Sergio Lopez geschrieben:
> 
> Kevin Wolf <address@hidden> writes:
> 
> > Am 19.11.2019 um 11:54 hat Sergio Lopez geschrieben:
> >> 
> >> Max Reitz <address@hidden> writes:
> >> 
> >> > On 13.11.19 14:24, Sergio Lopez wrote:
> >> >> 
> >> >> Sergio Lopez <address@hidden> writes:
> >> >> 
> >> >>> address@hidden writes:
> >> >>>
> >> >>>> Patchew URL: https://patchew.org/QEMU/address@hidden/
> >> >>>>
> >> >>>>
> >> >>>>
> >> >>>> Hi,
> >> >>>>
> >> >>>> This series failed the docker-quick@centos7 build test. Please find 
> >> >>>> the testing commands and
> >> >>>> their output below. If you have Docker installed, you can probably 
> >> >>>> reproduce it
> >> >>>> locally.
> >> >>>>
> >> >>>> === TEST SCRIPT BEGIN ===
> >> >>>> #!/bin/bash
> >> >>>> make docker-image-centos7 V=1 NETWORK=1
> >> >>>> time make docker-test-quick@centos7 SHOW_ENV=1 J=14 NETWORK=1
> >> >>>> === TEST SCRIPT END ===
> >> >>>>
> >> >>>>   TEST    iotest-qcow2: 268
> >> >>>> Failures: 141
> >> >>>
> >> >>> Hm... 141 didn't fail in my test machine. I'm going to have a look.
> >> >> 
> >> >> So here's the output:
> >> >> 
> >> >> --- /root/qemu/tests/qemu-iotests/141.out       2019-11-12 
> >> >> 04:43:27.651557587 -0500
> >> >> +++ /root/qemu/build/tests/qemu-iotests/141.out.bad     2019-11-13 
> >> >> 08:12:06.575967337 -0500
> >> >> @@ -10,6 +10,8 @@
> >> >>  Formatting 'TEST_DIR/o.IMGFMT', fmt=IMGFMT size=1048576 
> >> >> backing_file=TEST_DIR/t.IMGFMT backing_fmt=IMGFMT
> >> >>  {"timestamp": {"seconds":  TIMESTAMP, "microseconds":  TIMESTAMP}, 
> >> >> "event": "JOB_STATUS_CHANGE", "data": {"status": "created", "id": 
> >> >> "job0"}}
> >> >>  {"timestamp": {"seconds":  TIMESTAMP, "microseconds":  TIMESTAMP}, 
> >> >> "event": "JOB_STATUS_CHANGE", "data": {"status": "running", "id": 
> >> >> "job0"}}
> >> >> +{"timestamp": {"seconds":  TIMESTAMP, "microseconds":  TIMESTAMP}, 
> >> >> "event": "JOB_STATUS_CHANGE", "data": {"status": "paused", "id": 
> >> >> "job0"}}
> >> >> +{"timestamp": {"seconds":  TIMESTAMP, "microseconds":  TIMESTAMP}, 
> >> >> "event": "JOB_STATUS_CHANGE", "data": {"status": "running", "id": 
> >> >> "job0"}}
> >> >>  {"error": {"class": "GenericError", "desc": "Node 'drv0' is busy: node 
> >> >> is used as backing hd of 'NODE_NAME'"}}
> >> >>  {"return": {}}
> >> >>  {"timestamp": {"seconds":  TIMESTAMP, "microseconds":  TIMESTAMP}, 
> >> >> "event": "JOB_STATUS_CHANGE", "data": {"status": "aborting", "id": 
> >> >> "job0"}}
> >> >> 
> >> >> Those extra lines, the "paused" and "running", are a result of the job
> >> >> being done in a transaction, within a drained section.
> >> >> 
> >> >> We can update 141.out, but now I'm wondering, was it safe creating the
> >> >> job at do_drive_backup() outside of a drained section, as
> >> >> qmp_drive_backup was doing?
> >> >
> >> > I think it is.  Someone needs to drain the source node before attaching
> >> > the job filter (which intercepts writes), and bdrv_backup_top_append()
> >> > does precisely this.
> >> >
> >> > If the source node is in an I/O thread, you could argue that the drain
> >> > starts later than when the user has invoked the backup command, and so
> >> > some writes might slip through.  That’s correct.  But at the same time,
> >> > it’s impossible to drain it the instant the command is received.  So
> >> > some writes might always slip through (and the drain will not stop them
> >> > either, it will just let them happen).
> >> >
> >> > Therefore, I think it’s fine the way it is.
> >> >
> >> >> Do you think there may be any potential drawbacks as a result of always
> >> >> doing it now inside a drained section?
> >> >
> >> > Well, one drawback is clearly visible.  The job goes to paused for no
> >> > reason.
> >> 
> >> This is something that already happens when requesting the drive-backup
> >> through a transaction:
> >> 
> >> {"execute":"transaction","arguments":{"actions":[{"type":"drive-backup","data":{"device":"drv0","target":"o.qcow2","sync":"full","format":"qcow2"}}]}}
> >> 
> >> I don't think it makes sense to have two different behaviors for the
> >> same action. So we either accept the additional pause+resume iteration
> >> for qmp_drive_backup, or we remove the drained section from the
> >> transaction based one.
> >> 
> >> What do you think?
> >
> > Draining all involved nodes is necessary for transactions, because you
> > want a consistent backup across all involved disks. That is, you want it
> > to be a snapshot at the same point in time for all of them - no requests
> > may happen between starting backup on the first and the second disk.
> >
> > For a single device operation, this requirement doesn't exist, because
> > there is nothing else that must happen at the same point in time.
> 
> This poses a problem with the unification strategy you suggested for qmp
> commands and transactions. I guess that, if we really want to preserve
> the original behavior, we can extend DriveBackup to add a flag to
> indicate whether the transaction should create a drained section or not.
> 
> Does this sound reasonable to you?

I think we can accept an unnecessary drain for the single-device case.
It's only minimally worse than not draining early (because, as Max said,
we'll drain the node anyway later).

I'm not sure what the code looks like, but does the job go to paused
even when it's already created inside the drained section? (As opposed
to first creating the job and then draining.) I assume that this is what
you're already doing, just double-checking.

If this is how things work, I'd just adjust the test output and explain
the change in the commit message.

Kevin

signature.asc
Description: PGP signature