From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from eggs.gnu.org ([2001:4830:134:3::10]:55951) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1aVZsr-0005or-I9 for qemu-devel@nongnu.org; Tue, 16 Feb 2016 02:17:18 -0500 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1aVZso-0006wy-A8 for qemu-devel@nongnu.org; Tue, 16 Feb 2016 02:17:17 -0500 Received: from mx1.redhat.com ([209.132.183.28]:43983) by eggs.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1aVZso-0006wl-2V for qemu-devel@nongnu.org; Tue, 16 Feb 2016 02:17:14 -0500 Received: from int-mx11.intmail.prod.int.phx2.redhat.com (int-mx11.intmail.prod.int.phx2.redhat.com [10.5.11.24]) by mx1.redhat.com (Postfix) with ESMTPS id A3DA0804E7 for ; Tue, 16 Feb 2016 07:17:13 +0000 (UTC) Date: Tue, 16 Feb 2016 15:17:11 +0800 From: Fam Zheng Message-ID: <20160216071711.GI18664@ad.usersys.redhat.com> References: <1455470231-5223-1-git-send-email-pbonzini@redhat.com> <1455470231-5223-2-git-send-email-pbonzini@redhat.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <1455470231-5223-2-git-send-email-pbonzini@redhat.com> Subject: Re: [Qemu-devel] [PATCH 1/8] block-migration: acquire AioContext as necessary List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: Paolo Bonzini Cc: qemu-devel@nongnu.org, stefanha@redhat.com, mst@redhat.com On Sun, 02/14 18:17, Paolo Bonzini wrote: > This is needed because dataplane will run during block migration as well. > > The block device migration code is quite liberal in taking the iothread > mutex. For simplicity, keep it the same way, even though one could > actually choose between the BQL (for regular BlockDriverStates) and > the AioContext (for dataplane BlockDriverStates). When the block layer > is made fully thread safe, aio_context_acquire shall go away altogether. > > Signed-off-by: Paolo Bonzini > --- > migration/block.c | 61 ++++++++++++++++++++++++++++++++++++++++++++----------- > 1 file changed, 49 insertions(+), 12 deletions(-) > > diff --git a/migration/block.c b/migration/block.c > index a444058..6dd2327 100644 > --- a/migration/block.c > +++ b/migration/block.c > @@ -60,9 +60,15 @@ typedef struct BlkMigDevState { > int64_t cur_sector; > int64_t cur_dirty; > > - /* Protected by block migration lock. */ > + /* Data in the aio_bitmap is protected by block migration lock. > + * Allocation and free happen during setup and cleanup respectively. > + */ > unsigned long *aio_bitmap; > + > + /* Protected by block migration lock. */ > int64_t completed_sectors; > + > + /* Protected by iothread lock / AioContext. */ > BdrvDirtyBitmap *dirty_bitmap; > Error *blocker; > } BlkMigDevState; > @@ -100,7 +106,7 @@ typedef struct BlkMigState { > int prev_progress; > int bulk_completed; > > - /* Lock must be taken _inside_ the iothread lock. */ > + /* Lock must be taken _inside_ the iothread lock and any AioContexts. */ > QemuMutex lock; > } BlkMigState; > > @@ -264,11 +270,13 @@ static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds) > > if (bmds->shared_base) { > qemu_mutex_lock_iothread(); > + aio_context_acquire(bdrv_get_aio_context(bs)); > while (cur_sector < total_sectors && > !bdrv_is_allocated(bs, cur_sector, MAX_IS_ALLOCATED_SEARCH, > &nr_sectors)) { > cur_sector += nr_sectors; > } > + aio_context_release(bdrv_get_aio_context(bs)); > qemu_mutex_unlock_iothread(); > } > > @@ -302,11 +310,21 @@ static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds) > block_mig_state.submitted++; > blk_mig_unlock(); > > + /* We do not know if bs is under the main thread (and thus does > + * not acquire the AioContext when doing AIO) or rather under > + * dataplane. Thus acquire both the iothread mutex and the > + * AioContext. > + * > + * This is ugly and will disappear when we make bdrv_* thread-safe, > + * without the need to acquire the AioContext. > + */ > qemu_mutex_lock_iothread(); > + aio_context_acquire(bdrv_get_aio_context(bmds->bs)); > blk->aiocb = bdrv_aio_readv(bs, cur_sector, &blk->qiov, > nr_sectors, blk_mig_read_cb, blk); > > bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, cur_sector, nr_sectors); > + aio_context_release(bdrv_get_aio_context(bmds->bs)); > qemu_mutex_unlock_iothread(); > > bmds->cur_sector = cur_sector + nr_sectors; > @@ -321,8 +339,9 @@ static int set_dirty_tracking(void) > int ret; > > QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) { > + /* Creating/dropping dirty bitmaps only requires the big QEMU lock. */ Why? I don't think it is safe today. The BDS state is mutated and it can race with bdrv_set_dirty() etc. (Also the refresh_total_sectors in bdrv_nb_sectors can even do read/write, no?) > bmds->dirty_bitmap = bdrv_create_dirty_bitmap(bmds->bs, BLOCK_SIZE, > NULL, NULL); > if (!bmds->dirty_bitmap) { > ret = -errno; > goto fail; > @@ -332,11 +352,14 @@ static int set_dirty_tracking(void) > return ret; > } > > +/* Called with iothread lock taken. */ > + > static void unset_dirty_tracking(void) > { > BlkMigDevState *bmds; > > QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) { > + /* Creating/dropping dirty bitmaps only requires the big QEMU lock. */ Ditto. > bdrv_release_dirty_bitmap(bmds->bs, bmds->dirty_bitmap); > } > } > @@ -597,21 +627,28 @@ static void block_migration_cleanup(void *opaque) > { > BlkMigDevState *bmds; > BlkMigBlock *blk; > + AioContext *ctx; > > bdrv_drain_all(); > > unset_dirty_tracking(); > > - blk_mig_lock(); Why is it okay to skip the blk_mig_lock() for block_mig_state.bmds_list? > while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) { > QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry); > bdrv_op_unblock_all(bmds->bs, bmds->blocker); > error_free(bmds->blocker); > + > + /* Save ctx, because bmds->bs can disappear during bdrv_unref. */ > + ctx = bdrv_get_aio_context(bmds->bs); > + aio_context_acquire(ctx); > bdrv_unref(bmds->bs); > + aio_context_release(ctx); > + > g_free(bmds->aio_bitmap); > g_free(bmds); > } > > + blk_mig_lock(); > while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) { > QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry); > g_free(blk->buf);