All of lore.kernel.org
 help / color / mirror / Atom feed
From: Leo Luan <leoluan@gmail.com>
To: John Snow <jsnow@redhat.com>
Cc: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>,
	qemu-devel@nongnu.org, Qemu-block <qemu-block@nongnu.org>,
	Max Reitz <mreitz@redhat.com>
Subject: Re: Avoid copying unallocated clusters during full backup
Date: Fri, 17 Apr 2020 18:43:04 -0700	[thread overview]
Message-ID: <CAD_3s9+pSepV76-wp_AQTB-EQGwmga1FS+fT-Cb4n0Z48Zhxpg@mail.gmail.com> (raw)
In-Reply-To: <93052f9b-6539-0d4a-c922-fff7618c542d@redhat.com>

[-- Attachment #1: Type: text/plain, Size: 6347 bytes --]

On Fri, Apr 17, 2020 at 5:34 PM John Snow <jsnow@redhat.com> wrote:

>
>
> On 4/17/20 6:57 PM, Leo Luan wrote:
> > On Fri, Apr 17, 2020 at 1:24 PM Eric Blake <eblake@redhat.com
> > <mailto:eblake@redhat.com>> wrote:
> >
> >     On 4/17/20 3:11 PM, John Snow wrote:
> >
> >     >> +
> >     >> +    if (s->sync_mode == MIRROR_SYNC_MODE_FULL &&
> >     >> +       s->bcs->target->bs->drv != NULL &&
> >     >> +       strncmp(s->bcs->target->bs->drv->format_name, "qcow2", 5)
> >     == 0 &&
> >     >> +       s->bcs->source->bs->backing_file[0] == '\0')
> >     >
> >     > This isn't going to suffice upstream; the backup job can't be
> >     performing
> >     > format introspection to determine behavior on the fly.
> >
> >     Agreed.  The idea is right (we NEED to make backup operations smarter
> >     based on knowledge about both source and destination block status),
> but
> >     the implementation is not (a check for strcncmp("qcow2") is not
> ideal).
> >
> >
> > I see/agree that using strncmp("qcow2") is not general enough for the
> > upstream.  Would changing it to bdrv_unallocated_blocks_are_zero()
> suffice?
> >
>
> I don't know, to be really honest with you. Vladimir reworked the backup
> code recently and Virtuozzo et al have shown a very aggressive interest
> in optimizing the backup loop. I haven't really worked on that code
> since their rewrite.
>
> Dropping unallocated regions from the backup manifest is one strategy,
> but I think there will be cases where we won't be able to treat it like
> "TOP", but may still have unallocated regions we don't want to copy (We
> have a backing file which is itself unallocated.)
>
> I'm interested in a more general purpose mechanism for efficient
> copying. I think that instead of the backup job itself doing this in
> backup.c by populating the copy manifest, that it's also appropriate to
> try to copy every last block and have the backup loop implementation
> decide it doesn't actually need to copy that block.
>
> That way, the copy optimizations can be shared by any implementation
> that needs to do efficient copying, and we can avoid special format and
> graph-inspection code in the backup job main interface code.
>
> To be clear, I see these as identical amounts of work:
>
> - backup job runs a loop to inspect every cluster to see if it is
> allocated or not, and modifies its cluster backup manifest accordingly
>

This inspection can detect more than 1GB of unallocated (64KB) clusters per
loop and it's a shallower path.

>
> - backup job loops through the entire block and calls a smart_copy()
> function that might degrade into a no-op if the right conditions are met
> (source is unallocated, explicit zeroes are not needed on the destination)
>

If I am not mistaken, the copy loop does one cluster per iteration using a
twice deeper call path (trying to copy and eventually finding unallocated
clusters).  So with 64KB cluster size, it's 2 * 1G/64K ~= 32 million times
less efficient with the CPU cycles for large sparse virtual disks.

>
> Either way, you're looping and interrogating the disk, but in one case
> the efficiencies go deeper than *just* the backup code.
>

I think the early stop of inefficiency can help minimize the CPU impact of
the backup job on the VM instance.


> I think Vladimir has put a lot of work into making the backup code
> highly optimized, so I would consult with him to find out where the best
> place to put new optimizations are, if any -- he'll know!
>

Yes, hope that he will chime in.

Thanks!

>
> --js
>
>
> >
> >     >
> >     > I think what you're really after is something like
> >     > bdrv_unallocated_blocks_are_zero().
> >
> >     The fact that qemu-img already has a lot of optimizations makes me
> >     wonder what we can salvage from there into reusable code that both
> >     qemu-img and block backup can share, so that we're not reimplementing
> >     block status handling in multiple places.
> >
> >
> > A general fix reusing some existing code would be great.  When will it
> > appear in the upstream?  We are hoping to avoid needing to use a private
> > branch if possible.
> >
> >
> >     > So the basic premise is that if you are copying a qcow2 file and
> the
> >     > unallocated portions as defined by the qcow2 metadata are zero,
> it's
> >     > safe to skip those, so you can treat it like SYNC_MODE_TOP.
> >     >
> >     > I think you *also* have to know if the *source* needs those regions
> >     > explicitly zeroed, and it's not always safe to just skip them at
> the
> >     > manifest level.
> >     >
> >     > I thought there was code that handled this to some extent already,
> >     but I
> >     > don't know. I think Vladimir has worked on it recently and can
> >     probably
> >     > let you know where I am mistaken :)
> >
> >     Yes, I'm hoping Vladimir (or his other buddies at Virtuozzo) can
> chime
> >     in.  Meanwhile, I've working on v2 of some patches that will improve
> >     qemu's ability to tell if a destination qcow2 file already reads as
> all
> >     zeroes, and we already have bdrv_block_status() for telling which
> >     portions of a source image already read as all zeroes (whether or
> >     not it
> >     is due to not being allocated, the goal here is that we should NOT
> have
> >     to copy anything that reads as zero on the source over to the
> >     destination if the destination already starts life as reading all
> zero).
> >
> >
> > Can the eventual/optimal solution allow unallocated clusters to be
> > skipped entirely in the backup loop and make the detection of allocated
> > zeroes an option, not forcing the backup thread to loop through a
> > potentially huge empty virtual disk?
> >
>
> I mean, using the TOP code is doing the same thing, really: it's looking
> at allocation status and marking those blocks as "already copied", more
> or less.
>
> >
> >     And if nothing else, qemu 5.0 just added 'qemu-img convert
> >     --target-is-zero' as a last-ditch means of telling qemu to assume the
> >     destination reads as all zeroes, even if it cannot quickly prove it;
> we
> >     probably want to add a similar knob into the QMP commands for
> >     initiating
> >     block backup, for the same reasons.
> >
> >
> > This seems a good way of assuring the status of the target file.
> >
> > Thanks!
> >
>
>

[-- Attachment #2: Type: text/html, Size: 8568 bytes --]

  reply	other threads:[~2020-04-18  1:44 UTC|newest]

Thread overview: 11+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-04-17 18:33 Avoid copying unallocated clusters during full backup Leo Luan
2020-04-17 20:11 ` John Snow
2020-04-17 20:24   ` Eric Blake
2020-04-17 22:57     ` Leo Luan
2020-04-18  0:34       ` John Snow
2020-04-18  1:43         ` Leo Luan [this message]
2020-04-20 10:56           ` Vladimir Sementsov-Ogievskiy
2020-04-20 14:31             ` Bryan S Rosenburg
2020-04-20 15:04               ` Vladimir Sementsov-Ogievskiy
2020-04-21 14:41                 ` Bryan S Rosenburg
2020-04-17 22:31   ` Leo Luan

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=CAD_3s9+pSepV76-wp_AQTB-EQGwmga1FS+fT-Cb4n0Z48Zhxpg@mail.gmail.com \
    --to=leoluan@gmail.com \
    --cc=jsnow@redhat.com \
    --cc=mreitz@redhat.com \
    --cc=qemu-block@nongnu.org \
    --cc=qemu-devel@nongnu.org \
    --cc=vsementsov@virtuozzo.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.