| Message ID | 20170131112308.54189-1-vsementsov@virtuozzo.com |
|---|---|
| State | New |
| Headers | show |
v2 missed in topic, sorry for that. First version was "[PATCH] block: bdrv_invalidate_cache_all: invalidate children first"
On 31.01.2017 12:23, Vladimir Sementsov-Ogievskiy wrote: > Current implementation invalidates firstly parent bds and then its > children. This leads to the following bug: > > after incoming migration, in bdrv_invalidate_cache_all: > 1. invalidate parent bds - reopen it with BDRV_O_INACTIVE cleared > 2. child is not yet invalidated > 3. parent check that its BDRV_O_INACTIVE is cleared > 4. parent writes to child > 5. assert in bdrv_co_pwritev, as BDRV_O_INACTIVE is set for child > > This patch fixes it by just changing invalidate sequence: invalidate > children first. > > Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> > --- > > v2: I've missed that bdrv_invalidate_cache is already recursive, so we > can change sequence here. Also v1 doesn't cover the case when > bdrv_invalidate_cache is called not from bdrv_invalidate_cache_all. > > block.c | 11 +++++------ > 1 file changed, 5 insertions(+), 6 deletions(-) Reviewed-by: Max Reitz <mreitz@redhat.com> I'll wait a bit until I apply this patch (please remind me next week if I forget it...), though, because migration-related things are not exactly my forte, so maybe someone else has something to say. Max
On Tue, Jan 31, 2017 at 02:23:08PM +0300, Vladimir Sementsov-Ogievskiy wrote: > Current implementation invalidates firstly parent bds and then its > children. This leads to the following bug: > > after incoming migration, in bdrv_invalidate_cache_all: > 1. invalidate parent bds - reopen it with BDRV_O_INACTIVE cleared > 2. child is not yet invalidated > 3. parent check that its BDRV_O_INACTIVE is cleared > 4. parent writes to child > 5. assert in bdrv_co_pwritev, as BDRV_O_INACTIVE is set for child > > This patch fixes it by just changing invalidate sequence: invalidate > children first. > > Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> > --- > > v2: I've missed that bdrv_invalidate_cache is already recursive, so we > can change sequence here. Also v1 doesn't cover the case when > bdrv_invalidate_cache is called not from bdrv_invalidate_cache_all. > > block.c | 11 +++++------ > 1 file changed, 5 insertions(+), 6 deletions(-) Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
On 31.01.2017 12:23, Vladimir Sementsov-Ogievskiy wrote: > Current implementation invalidates firstly parent bds and then its > children. This leads to the following bug: > > after incoming migration, in bdrv_invalidate_cache_all: > 1. invalidate parent bds - reopen it with BDRV_O_INACTIVE cleared > 2. child is not yet invalidated > 3. parent check that its BDRV_O_INACTIVE is cleared > 4. parent writes to child > 5. assert in bdrv_co_pwritev, as BDRV_O_INACTIVE is set for child > > This patch fixes it by just changing invalidate sequence: invalidate > children first. > > Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> > --- > > v2: I've missed that bdrv_invalidate_cache is already recursive, so we > can change sequence here. Also v1 doesn't cover the case when > bdrv_invalidate_cache is called not from bdrv_invalidate_cache_all. Thanks, applied to my block tree: https://github.com/XanClic/qemu/commits/block Max
diff --git a/block.c b/block.c index a0346c80c6..dce1dc02af 100644 --- a/block.c +++ b/block.c @@ -3235,19 +3235,18 @@ void bdrv_invalidate_cache(BlockDriverState *bs, Error **errp) if (!(bs->open_flags & BDRV_O_INACTIVE)) { return; } - bs->open_flags &= ~BDRV_O_INACTIVE; - if (bs->drv->bdrv_invalidate_cache) { - bs->drv->bdrv_invalidate_cache(bs, &local_err); + QLIST_FOREACH(child, &bs->children, next) { + bdrv_invalidate_cache(child->bs, &local_err); if (local_err) { - bs->open_flags |= BDRV_O_INACTIVE; error_propagate(errp, local_err); return; } } - QLIST_FOREACH(child, &bs->children, next) { - bdrv_invalidate_cache(child->bs, &local_err); + bs->open_flags &= ~BDRV_O_INACTIVE; + if (bs->drv->bdrv_invalidate_cache) { + bs->drv->bdrv_invalidate_cache(bs, &local_err); if (local_err) { bs->open_flags |= BDRV_O_INACTIVE; error_propagate(errp, local_err);
Current implementation invalidates firstly parent bds and then its children. This leads to the following bug: after incoming migration, in bdrv_invalidate_cache_all: 1. invalidate parent bds - reopen it with BDRV_O_INACTIVE cleared 2. child is not yet invalidated 3. parent check that its BDRV_O_INACTIVE is cleared 4. parent writes to child 5. assert in bdrv_co_pwritev, as BDRV_O_INACTIVE is set for child This patch fixes it by just changing invalidate sequence: invalidate children first. Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> --- v2: I've missed that bdrv_invalidate_cache is already recursive, so we can change sequence here. Also v1 doesn't cover the case when bdrv_invalidate_cache is called not from bdrv_invalidate_cache_all. block.c | 11 +++++------ 1 file changed, 5 insertions(+), 6 deletions(-)