[dm-devel] [PATCH 3/3] dm clone: Flush destination device before committing metadata
Mike Snitzer
snitzer at redhat.com
Thu Dec 5 20:07:47 UTC 2019
On Thu, Dec 05 2019 at 2:46pm -0500,
Mike Snitzer <snitzer at redhat.com> wrote:
> On Wed, Dec 04 2019 at 9:06P -0500,
> Nikos Tsironis <ntsironis at arrikto.com> wrote:
>
> > dm-clone maintains an on-disk bitmap which records which regions are
> > valid in the destination device, i.e., which regions have already been
> > hydrated, or have been written to directly, via user I/O.
> >
> > Setting a bit in the on-disk bitmap meas the corresponding region is
> > valid in the destination device and we redirect all I/O regarding it to
> > the destination device.
> >
> > Suppose the destination device has a volatile write-back cache and the
> > following sequence of events occur:
> >
> > 1. A region gets hydrated, either through the background hydration or
> > because it was written to directly, via user I/O.
> >
> > 2. The commit timeout expires and we commit the metadata, marking that
> > region as valid in the destination device.
> >
> > 3. The system crashes and the destination device's cache has not been
> > flushed, meaning the region's data are lost.
> >
> > The next time we read that region we read it from the destination
> > device, since the metadata have been successfully committed, but the
> > data are lost due to the crash, so we read garbage instead of the old
> > data.
> >
> > This has several implications:
> >
> > 1. In case of background hydration or of writes with size smaller than
> > the region size (which means we first copy the whole region and then
> > issue the smaller write), we corrupt data that the user never
> > touched.
> >
> > 2. In case of writes with size equal to the device's logical block size,
> > we fail to provide atomic sector writes. When the system recovers the
> > user will read garbage from the sector instead of the old data or the
> > new data.
> >
> > 3. In case of writes without the FUA flag set, after the system
> > recovers, the written sectors will contain garbage instead of a
> > random mix of sectors containing either old data or new data, thus we
> > fail again to provide atomic sector writes.
> >
> > 4. Even when the user flushes the dm-clone device, because we first
> > commit the metadata and then pass down the flush, the same risk for
> > corruption exists (if the system crashes after the metadata have been
> > committed but before the flush is passed down).
> >
> > The only case which is unaffected is that of writes with size equal to
> > the region size and with the FUA flag set. But, because FUA writes
> > trigger metadata commits, this case can trigger the corruption
> > indirectly.
> >
> > To solve this and avoid the potential data corruption we flush the
> > destination device **before** committing the metadata.
> >
> > This ensures that any freshly hydrated regions, for which we commit the
> > metadata, are properly written to non-volatile storage and won't be lost
> > in case of a crash.
> >
> > Fixes: 7431b7835f55 ("dm: add clone target")
> > Cc: stable at vger.kernel.org # v5.4+
> > Signed-off-by: Nikos Tsironis <ntsironis at arrikto.com>
> > ---
> > drivers/md/dm-clone-target.c | 46 ++++++++++++++++++++++++++++++++++++++------
> > 1 file changed, 40 insertions(+), 6 deletions(-)
> >
> > diff --git a/drivers/md/dm-clone-target.c b/drivers/md/dm-clone-target.c
> > index 613c913c296c..d1e1b5b56b1b 100644
> > --- a/drivers/md/dm-clone-target.c
> > +++ b/drivers/md/dm-clone-target.c
> > @@ -86,6 +86,12 @@ struct clone {
> >
> > struct dm_clone_metadata *cmd;
> >
> > + /*
> > + * bio used to flush the destination device, before committing the
> > + * metadata.
> > + */
> > + struct bio flush_bio;
> > +
> > /* Region hydration hash table */
> > struct hash_table_bucket *ht;
> >
> > @@ -1108,10 +1114,13 @@ static bool need_commit_due_to_time(struct clone *clone)
> > /*
> > * A non-zero return indicates read-only or fail mode.
> > */
> > -static int commit_metadata(struct clone *clone)
> > +static int commit_metadata(struct clone *clone, bool *dest_dev_flushed)
> > {
> > int r = 0;
> >
> > + if (dest_dev_flushed)
> > + *dest_dev_flushed = false;
> > +
> > mutex_lock(&clone->commit_lock);
> >
> > if (!dm_clone_changed_this_transaction(clone->cmd))
> > @@ -1128,6 +1137,19 @@ static int commit_metadata(struct clone *clone)
> > goto out;
> > }
> >
> > + bio_reset(&clone->flush_bio);
> > + bio_set_dev(&clone->flush_bio, clone->dest_dev->bdev);
> > + clone->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
> > +
> > + r = submit_bio_wait(&clone->flush_bio);
> > + if (unlikely(r)) {
> > + __metadata_operation_failed(clone, "flush destination device", r);
> > + goto out;
> > + }
> > +
> > + if (dest_dev_flushed)
> > + *dest_dev_flushed = true;
> > +
> > r = dm_clone_metadata_commit(clone->cmd);
> > if (unlikely(r)) {
> > __metadata_operation_failed(clone, "dm_clone_metadata_commit", r);
> > @@ -1199,6 +1221,7 @@ static void process_deferred_bios(struct clone *clone)
> > static void process_deferred_flush_bios(struct clone *clone)
> > {
> > struct bio *bio;
> > + bool dest_dev_flushed;
> > struct bio_list bios = BIO_EMPTY_LIST;
> > struct bio_list bio_completions = BIO_EMPTY_LIST;
> >
> > @@ -1218,7 +1241,7 @@ static void process_deferred_flush_bios(struct clone *clone)
> > !(dm_clone_changed_this_transaction(clone->cmd) && need_commit_due_to_time(clone)))
> > return;
> >
> > - if (commit_metadata(clone)) {
> > + if (commit_metadata(clone, &dest_dev_flushed)) {
> > bio_list_merge(&bios, &bio_completions);
> >
> > while ((bio = bio_list_pop(&bios)))
> > @@ -1232,8 +1255,17 @@ static void process_deferred_flush_bios(struct clone *clone)
> > while ((bio = bio_list_pop(&bio_completions)))
> > bio_endio(bio);
> >
> > - while ((bio = bio_list_pop(&bios)))
> > - generic_make_request(bio);
> > + while ((bio = bio_list_pop(&bios))) {
> > + if ((bio->bi_opf & REQ_PREFLUSH) && dest_dev_flushed) {
> > + /* We just flushed the destination device as part of
> > + * the metadata commit, so there is no reason to send
> > + * another flush.
> > + */
> > + bio_endio(bio);
> > + } else {
> > + generic_make_request(bio);
> > + }
> > + }
> > }
> >
> > static void do_worker(struct work_struct *work)
> > @@ -1405,7 +1437,7 @@ static void clone_status(struct dm_target *ti, status_type_t type,
> >
> > /* Commit to ensure statistics aren't out-of-date */
> > if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
> > - (void) commit_metadata(clone);
> > + (void) commit_metadata(clone, NULL);
> >
> > r = dm_clone_get_free_metadata_block_count(clone->cmd, &nr_free_metadata_blocks);
> >
> > @@ -1839,6 +1871,7 @@ static int clone_ctr(struct dm_target *ti, unsigned int argc, char **argv)
> > bio_list_init(&clone->deferred_flush_completions);
> > clone->hydration_offset = 0;
> > atomic_set(&clone->hydrations_in_flight, 0);
> > + bio_init(&clone->flush_bio, NULL, 0);
> >
> > clone->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
> > if (!clone->wq) {
> > @@ -1912,6 +1945,7 @@ static void clone_dtr(struct dm_target *ti)
> > struct clone *clone = ti->private;
> >
> > mutex_destroy(&clone->commit_lock);
> > + bio_uninit(&clone->flush_bio);
> >
> > for (i = 0; i < clone->nr_ctr_args; i++)
> > kfree(clone->ctr_args[i]);
> > @@ -1966,7 +2000,7 @@ static void clone_postsuspend(struct dm_target *ti)
> > wait_event(clone->hydration_stopped, !atomic_read(&clone->hydrations_in_flight));
> > flush_workqueue(clone->wq);
> >
> > - (void) commit_metadata(clone);
> > + (void) commit_metadata(clone, NULL);
> > }
> >
> > static void clone_resume(struct dm_target *ti)
> > --
> > 2.11.0
> >
>
>
> Like the dm-thin patch I replied to, would rather avoid open-coding
> blkdev_issue_flush (also I check !bio_has_data), here is incremental:
Sorry for the noise relative to !bio_has_data check.. we don't need it.
DM core will split flush from data (see dec_pending()'s REQ_PREFLUSH
check).
I'm dropping the extra !bio_has_data() checks from the incrementals I
did; will review again and push out to linux-next.. still have time to
change if you fundamentally disagree with using blkdev_issue_flush()
Thanks,
Mike
More information about the dm-devel
mailing list