On 3/9/23 12:44, Hanna Czenczek wrote:
+ * + * Note that TRIM operations call blk_aio_pdiscard() multiple + * times (and finally increment s->blk's in-flight counter while + * ide_trim_bh_cb() is scheduled), so we have to loop blk_drain() + * until the whole operation is done. */ if (s->bus->dma->aiocb) { trace_ide_cancel_dma_sync_remaining(); - blk_drain(s->blk); - assert(s->bus->dma->aiocb == NULL); + while (s->bus->dma->aiocb) { + blk_drain(s->blk); + }
I think having to do this is problematic, because the blk_drain should leave no pending operation.
Here it seems okay because you do it in a controlled situation, but the main thread can also issue blk_drain(), or worse bdrv_drained_begin(), and there would be pending I/O operations when it returns.
Unfortunately I don't have a solution (I'm not considering the idea of using disable_request_queuing and having even more atomics magic in block/block-backend.c), but I'll read the thread.
Paolo