On 11/13/2014 08:54 AM, Vladimir Sementsov-Ogievskiy wrote:
Hi
I'd just like to start working on persistent dirty bitmap. My thoughts
about it are the following:
- qemu -drive file=file,dirty_bitmap=file
so, bitmap will be loaded with drive open and saved with drive
close.
- save only meaningful (the last) level of the bitmap, restore all
levels on bitmap loading
- bool parameter "persistent" for bdrv_create_dirty_bitmap and
BdrvDirtyBitmap
- internal dirty_bitmaps, saved in qcow2 file
Best regards,
Vladimir
I am thinking:
(1) Command Lines
If you enable dirty bitmaps and give it a file that doesn't exist, it
should error out on you.
If you enable dirty bitmaps and give it a file that's blank, it
understands that it is to create a persistent bitmap file in this
location and it should enable persistence.
If a bitmap file is given and it has valid magic, this should imply
persistence.
I am hesitant to have it auto-create files that don't already exist in
case the files become large in size and a misconfiguration leads to
repeated creation of these files that get orphaned in random folders.
Perhaps we can add a create=auto flag or similar to allow this
behavior if wanted.
(2) File Format
Some standard file magic, which includes:
- Some magic byte(s)
- Dirty flag. Needed to tell if we can trust this data or not.
- The size of the bitmap
- The granularity of the bitmap
- The offset to the first sector of bitmap data (Maybe? It can't hurt
if we give ourselves a sector's worth to write metadata within.)
- Data starting at... PAGESIZE?
(3) Data Integrity
The dirty flag could work something like:
- If, on first open, the file has the dirty flag set, we need to
discard the bitmap data because we can no longer trust it.
- If the bitmap file is clean, proceed as normal, but take a lock
against any of the bitmap functions to prevent them from marking any
bits dirty.
- On first write to a clean persistent bitmap, delay the write until
we can mark the bitmap as dirty first. This incurs a write penalty
when we try to use the bitmap at first...
- Unlock the bitmap functions and allow them to mark blocks as needed.
- At some point, based on a sync policy, re-commit the dirty
information to the file and mark the file as clean once more and
re-take the persistence lock.
(4) Synchronization Policy
- Sync after so many bits become dirty in the bitmap, either as an
absolute threshold or a density percentage?
- Sync periodically on a fixed timer?
- Sync periodically opportunistically when I/O utilization becomes
relatively low? (With some sort of starvation prevention timer?)
- Sync only at shutdown?
In discussing with Stefan, I think we rather liked the idea of a timer
that tries to re-commit the block data during lulls in the I/O.
(5) Partial Persistence
We did not discuss only saving higher levels of the bitmap. What's the
primary benefit you're seeking?
(6) Inclusion as qcow2 Metadata
And lastly, we did discuss the inclusion of the bitmap as qcow2
metadata, but decided it wasn't our principle target for the format to
allow generality to other file formats. We didn't really discuss the
idea of having it as an option or an extension, but I don't (off the
top of my head) have any reasonings against it, but I will likely not
work on it myself.
You didn't CC qemu-devel on this (so I won't!), but perhaps we should
re-send out our ideas to the wider list for feedback before we proceed
any further. Maybe we can split the work if we agree upon a design.
Thanks!
--js
P.S.: I'm still cleaning up Fam's first patchset based on Max's and
your feedback. Hope to have it out by the end of this week.
On 11.11.2014 18:59, John Snow wrote:
On 11/10/2014 03:15 AM, Vladimir Sementsov-Ogievskiy wrote:
Hi Fam, hi Jorn.
Jagane's project - http://wiki.qemu.org/Features/Livebackup
In two words:
Normal delta - like in qemu, while backuping, we save all new
writes to
separate virtual disk - delta. When backup is done, we can merge delta
back to original image.
Reverse delta - while backuping, we don't stop writing to original
image
(and qemu works with it, not with delta), but before every write we
copy
the corresponding block (if backup needs it) to separate virtual disk
(reverse delta). So, for backuping, if the current block is not
rewritten, we take it from original file, otherwise we take it from
reverse delta. The benefit is that, theoretically we'll have two times
less overhead I/Os then with "normal delta backup + merge", because of
part of (about 0.5) blocks which are rewritten while backuping will be
backuped before rewrite, and they will not go to delta then. Also,
there
are should be methods to improve this coefficient, by choosing rules
about what blocks to backup first.
Are someone work now on persistent (while vm is turned off)
bitmaps? If
not, are there any recommendations?
Best regards,
Vladimir
Hi Vladimir:
I recently inherited Fam's backup series and I intended to start
reviewing the comments this week. After I sift through that situation,
the next step for me is persistent dirty bitmap support. I had been
working on AHCI and IDE an awful lot lately, so it will take me a
little bit to switch over.
I don't have any existing notions of how I intend to implement it, so
if you'd like to discuss any concerns you have about that subsystem,
we can talk :)
Thanks,
--John
On 08.11.2014 10:19, Fam Zheng wrote:
On Fri, 11/07 15:23, Vladimir Sementsov-Ogievskiy wrote:
Hi!
Glad to see you working on backup series again. Some time ago I've
already
fixed v4 to be mergeable, but unfortunately didn't publish.
Actually I need (and I'm going to develop) incremental backup with
reverse
delta (like abandoned project
http://wiki.qemu.org/Features/Livebackup by
Jagane). And I've already fixed Jagane's project to be working with
current
Qemu. But the code is bad. Can you give me a peace of advice about
what
should I notice? I have a very superficial understanding of qemu
block
devices, block jobs, etc.
Hi Vladimir,
Thank you for reviewing my series.
I haven't seen Jagane's series, so could you pleased explain what
"reverse
delta" is?
John Snow (CC'ed) will take my work on incremental backup so I think
you can
coordinate. I'll probably count on John to work on your comments too,
in the
future I'll follow the progress and review.
Thanks,
Fam