Re: [Qemu-devel] [Qemu-block] [PATCH 0/3] mirror: Fix guest responsiveness during bitmap scan

[Alexandre DERUMIER <aderumier@odiso.com>]

>>The other block jobs (stream, backup, commit) are still using 
>>a 0 ns delay and are therefore broken. 

Also in mirror.c, they are block_job_sleep_ns with delay_ns, where delay_ns can be 0.

I have seen sometime,some server/qmp hangs, on a pretty old server/slow disks.
(no more in bitmap scan with theses last patches, but on the block mirror itself)


       if (!s->synced) {
            block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, delay_ns);
            if (block_job_is_cancelled(&s->common)) {
                break;
            }
        } else if (!should_complete) {
            delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0);


----- Mail original -----
De: "Stefan Hajnoczi" <stefanha@gmail.com>
À: "Fam Zheng" <famz@redhat.com>
Cc: "Kevin Wolf" <kwolf@redhat.com>, "qemu-devel" <qemu-devel@nongnu.org>, qemu-block@nongnu.org
Envoyé: Jeudi 9 Juillet 2015 15:02:08
Objet: Re: [Qemu-devel] [Qemu-block] [PATCH 0/3] mirror: Fix guest responsiveness during bitmap scan

On Thu, Jul 09, 2015 at 11:47:55AM +0800, Fam Zheng wrote: 
> This supersedes: 
> 
> http://patchwork.ozlabs.org/patch/491415/ 
> 
> and [1] which is currently in Jeff's tree. 
> 
> Although [1] fixed the QMP responsiveness, Alexandre DERUMIER reported that 
> guest responsiveness still suffers when we are busy in the initial dirty bitmap 
> scanning loop of mirror job. That is because 1) we issue too many lseeks; 2) we 
> only sleep for 0 ns which turns out quite ineffective in yielding BQL to vcpu 
> threads. Both are fixed. 
> 
> To reproduce: start a guest, attach a 10G raw image, then mirror it. Your 
> guest will immediately start to stutter (with patch [1] testing on a local ext4 
> raw image, and "while echo -n .; do sleep 0.05; done" in guest console). 
> 
> This series adds block_job_relax_cpu as suggested by Stefan Hajnoczi and uses 
> it in mirror job; and lets bdrv_is_allocated_above return a larger p_num as 
> suggested by Paolo Bonzini (although it's done without changing the API). 
> 
> [1]: http://patchwork.ozlabs.org/patch/471656/ "block/mirror: Sleep 
> periodically during bitmap scanning" 
> 
> Fam Zheng (3): 
> blockjob: Introduce block_job_relax_cpu 
> mirror: Use block_job_relax_cpu during bitmap scanning 
> mirror: Speed up bitmap initial scanning 
> 
> block/mirror.c | 17 +++++++---------- 
> include/block/blockjob.h | 16 ++++++++++++++++ 
> 2 files changed, 23 insertions(+), 10 deletions(-) 

This patch only converts the mirror block job to use the new relax 
function. The other block jobs (stream, backup, commit) are still using 
a 0 ns delay and are therefore broken. They should probably be 
converted in the same series.