On Tue, 05/19 10:49, Paolo Bonzini wrote: > > > On 19/05/2015 18:48, Fam Zheng wrote: > >> > > >> > This is too late. As a rule, the blocker must be established before > >> > calling bdrv_drain, and removed on the next yield (in this case, before > >> > the assignment to last_pause_ns). > > I don't understand. If the blocker is removed before mirror_run returns, > > wouldn't more device IO already hit source image by the time mirror_exit > > runs? > > If you go to mirror_exit, you won't reach the assignment (so you have to > remove the blocker in mirror_exit too). > > But if you don't go to mirror_exit because cnt != 0, you must remove the > blocker before the next I/O. >
OK, but I'm still not clear how is it too late in this patch? Although the blocker is set after bdrv_drain, we know there is no dirty data because cnt is 0, and we'll be holding a blocker when releasing the AioContext, no new IO is allowed. Fam
