On 30 nov. 2009, at 20:25, Jan Kiszka wrote: > Pierre Riteau wrote: >> On 30 nov. 2009, at 19:34, Anthony Liguori wrote: >> >>> Jan Kiszka wrote: >>>> This series is a larger rework of the block migration support qemu >>>> recently gained. Besides lots of code refactorings the major changes >>>> are: >>>> - Faster restore due to larger block sizes (even if the target disk is >>>> unallocated) >>>> - Off-by-one fixes in the block dirty tracking code >>>> - Allow for multiple migrations (after cancellation or if migrating >>>> into a backup image) >>>> - Proper error handling >>>> - Progress reporting fixes: report to monitor instead of stdout, report >>>> sum of multiple disks >>>> - Report disk migration progress via 'info migrate' >>>> - Progress report during restore >>>> >>>> One patch is directly taken from Pierre Riteau queue [1] who happend to >>>> work on the some topic the last days, two more are derived from his >>>> commits. >>>> >>>> These patches make block migration usable for us. Still, there are two >>>> more major improvements on my wish/todo list: >>>> - Respect specified maximum migration downtime (will require tracking >>>> of the number of dirty blocks + some coordination with ram migration) >>>> - Do not transfere unallocated disk space (also for raw images, ie. add >>>> bdrv_is_allocated support for the latter) >>>> >>>> In an off-list chat, Liran additionally brought up the topic that RAM >>>> migration should not start too early so that we avoid re-transmitting >>>> dirty pages over and over again while the disk image is slowly beamed >>>> over. >>>> >>>> I hope we can join our efforts to resolve the open topics quickly, the >>>> critical ones ideally before the merge window closes. >>>> >>> That really needs to happen no later than the end of this week. >>> >>> So Pierre/Liran, what do you think about Jan's series? >>> >>> Regards, >>> >>> Anthony Liguori >> >> >> I'm currently testing these patches. Here are a few issues I noticed, before >> I forget about them. >> >> - "migrate -d -b tcp:dest:port" works, but "migrate -b -d tcp:dest:port" >> doesn't, although "help migrate" doesn't really specify ordering as >> important. But anyway I think Liran is working on a new version of the >> command. > > Saw that too. I think the monitor commands simply do very primitive > option parsing so far. Should be addressed if the final format comes > with this issue as well. > >> - We use bdrv_aio_readv() to read blocks from the disk. This function >> increments rd_bytes and rd_ops, which are reported by "info blockstats". I >> don't think this read operations should appear in VM activity, especially if >> this interface is used by libvirt to report VM stats (and draw graphs in >> virt-manager, etc.). Same for write stats. > > Ack. > >> - We may need to call bdrv_reset_dirty() _before_ sending the data, to be >> sure the block is not rewritten in the meantime (maybe it's an issue only >> with kvm?) > > Can you elaborate? Even in case of multi-threaded qemu, the iomutex > should protect us here.
I only said that because I remember seeing this kind of behavior, but with ram migration on kvm. As I'm not familiar with the I/O emulation in qemu, if you say that it's OK, no problem. By multi-threaded, are you talking about the IO thread feature? >> - I seem to remember that disk images with 0 size are now possible. I'm >> afraid we will hit a divide by zero in this case: "progress = >> completed_sector_sum * 100 / block_mig_state.total_sector_sum;" > > Although I don't see their use, it should be handled gracefully, likely > by skipping such disks. From a patch by Stefan Weil a few weeks ago: > Images with disk size 0 may be used for > VM snapshots, but not to save normal block data. > > It is possible to create such images using > qemu-img, but opening them later fails. > > So even "qemu-img info image.qcow2" is not > possible for an image created with > "qemu-img create -f qcow2 image.qcow2 0". I'm not sure if that concerns us... -- Pierre Riteau -- http://perso.univ-rennes1.fr/pierre.riteau/
