On Wed, Apr 01, 2015 at 04:49:39PM +0800, Bin Wu wrote: > > On 2015/4/1 16:19, Fam Zheng wrote: > > On Wed, 04/01 12:42, Bin Wu wrote: > >> From: Bin Wu <wu.wu...@huawei.com> > > > > What's the issue are you fixing? I think the coroutine already is running in > > the AioContext of bs. > > > > Fam > > > In the current implementation of bdrv_drain, it should be placed in a critical > section as suggested in the comments above the function: "Note that unlike > bdrv_drain_all(), the caller must hold the BlockDriverState AioContext". > > However, the mirror coroutine starting with mirror_run doesn't do this. I just > found qmp_drive_mirror protects the AioCentext, but it is out of the scope of > the mirror coroutine.
There are 3 possibilities: 1. qmp_drive_mirror() under QEMU main loop thread. AioContext is held. 2. IOThread aio_poll(). AioContext is held. 3. QEMU main loop thread when IOThread (dataplane) is not used. Here the AioContext is the global qemu_aio_context. We don't need to acquire it explicitly, we're protected by the global mutex. If #3 was a problem then virtio-blk.c's bdrv_aio_writev() would also be a problem, for example. This patch is unnecessary. Stefan
pgp6U2PGqjyjN.pgp
Description: PGP signature
