> From: Kevin Wolf [mailto:kw...@redhat.com] > > > Coroutines aren't randomly assigned to threads, but threads actively > > > enter coroutines. To my knowledge this happens only when starting a > > > request (either vcpu or I/O thread; consistent per device) or by a > > > callback when some event happens (only I/O thread). I can't see any > > > non-determinism here. > > > > Behavior of coroutines looks strange for me. > > Consider the code below (co_readv function of the replay driver). > > In record mode it somehow changes the thread it assigned to. > > Code in point A is executed in CPU thread and code in point B - in some > > other thread. > > May this happen because this coroutine yields somewhere and its execution > > is restored > > by aio_poll, which is called from iothread? > > In this case event finishing callback cannot be executed deterministically > > (always in CPU thread or always in IO thread). > > > > static int coroutine_fn blkreplay_co_readv(BlockDriverState *bs, > > int64_t sector_num, int nb_sectors, QEMUIOVector *qiov) > > { > > BDRVBlkreplayState *s = bs->opaque; > > uint32_t reqid = request_id++; > > Request *req; > > // A > > bdrv_co_readv(bs->file->bs, sector_num, nb_sectors, qiov); > > > > if (replay_mode == REPLAY_MODE_RECORD) { > > replay_save_block_event(reqid); > > } else { > > assert(replay_mode == REPLAY_MODE_PLAY); > > if (reqid == current_request) { > > current_finished = true; > > } else { > > req = block_request_insert(reqid, bs, qemu_coroutine_self()); > > qemu_coroutine_yield(); > > block_request_remove(req); > > } > > } > > // B > > return 0; > > } > > Yes, I guess this can happen. As I described above, the coroutine can be > entered from a vcpu thread initially. After yielding for the first time, > it is resumed from the I/O thread. So if there are paths where the > coroutine never yields, the coroutine completes in the original vcpu > thread. (It's not the common case that bdrv_co_readv() doesn't yield, > but it happens e.g. with unallocated sectors in qcow2.) > > If this is a problem for you, you need to force the coroutine into the > I/O thread. You can do that by scheduling a BH, then yield, and then let > the BH reenter the coroutine.
Thanks, this approach seems to work. I got rid of replay_run_block_event, because BH basically does the same job. There is one problem with flush event - callbacks for flush are called for all layers and I couldn't synchronize them correctly yet. I'll probably have to add new callback to block driver, which handles flush request for the whole stack of the drivers. Pavel Dovgalyuk
