On 5/24/13 1:05 AM, "Stefan Hajnoczi" <stefa...@gmail.com> wrote:
>On Thu, May 23, 2013 at 09:58:31PM +0000, Mark Trumpold wrote:
>> I have a working configuration using the signal approach suggested by
>>Stefan.
>>
>> 'qemu-nbd.c' is patched as follows:
>>
>> do {
>> main_loop_wait(false);
>> + if (sighup_reported) {
>> + sighup_reported = false;
>> + bdrv_drain_all();
>> + bdrv_flush_all();
>> }
>> } while (!sigterm_reported && (persistent || !nbd_started || nb_fds
>>> 0));
>>
>> The driving script was patched as follows:
>>
>> mount -o remount,ro /dev/nbd0
>> blockdev --flushbufs /dev/nbd0
>> + kill -HUP <qemu-nbd process id>
>>
>> I needed to retain 'blockdev --flushbufs' for things to work. Seems
>>the 'bdrv_flush_all' is flushing what is being missed by the blockdev
>>flush. I did not go back an retest with 'fsync' or other approaches I
>>had tried before.
>
>Okay, that makes sense:
>
>'blockdev --flushbufs' is writing dirty pages to the NBD device.
>
>bdrv_drain_all() + bdrv_flush_all() ensures that image file writes reach
>the physical disk.
>
>One thing to be careful of is whether these operations are asynchronous.
>The signal is asynchronous, you have no way of knowing when qemu-nbd is
>finished flushing to the physical disk.
Right, of course. I missed the obvious.
>
>I didn't check blockdev(8) but it could be the same there.
>
>So watch out, otherwise your script is timing-dependent and may not
>actually have finished flushing when you take the snapshot.
>
>Stefan
>
The race condition would not be acceptable. You had mentioned another
approach using the socket interface:
>2. Instantiate a block/nbd.c client that connects to the running
> qemu-nbd server (make sure format=raw). Then call bdrv_flush() on
> the NBD client. You must use the qemu-nbd --shared=2 option.
>
In my case I only have a 'qemu-nbd' process per loop device. Would the
'qemu-nbd' process act as the socket server, and I would then write a
simple socket client to instruct him to do the flush? And, would the
client block until the flush is complete?
Thank you,
Mark T.
