diff options
| author | Lars Ellenberg | 2010-10-18 23:04:07 +0200 |
|---|---|---|
| committer | Philipp Reisner | 2010-10-22 15:50:27 +0200 |
| commit | 6719fb036cea56a5ee9d0ac912ed8c7cabb27f49 (patch) | |
| tree | eed4e4d3161cc6f7a4009c5c856e88c6707c88a3 /drivers/block/drbd/drbd_req.c | |
| parent | drbd: fix potential deadlock on detach (diff) | |
| download | kernel-qcow2-linux-6719fb036cea56a5ee9d0ac912ed8c7cabb27f49.tar.gz kernel-qcow2-linux-6719fb036cea56a5ee9d0ac912ed8c7cabb27f49.tar.xz kernel-qcow2-linux-6719fb036cea56a5ee9d0ac912ed8c7cabb27f49.zip | |
drbd: fix potential data divergence after multiple failures
If we get an IO-error during an activity log transaction,
if we failed to write the bitmap of the evicted extent,
we must not write the transaction itself.
If we failed to write the transaction,
we must not even submit the corresponding bio,
as its extent is not yet marked in the activity log.
Otherwise, if this was a disconneted Primary (degraded cluster), which
now lost its disk as well, and we later re-attach the same backend
storage, we possibly "forget" to resync some parts of the disk that
potentially have been changed.
On the receiving side, when receiving from a peer with unhealthy disk,
checking for pdsk == D_DISKLESS is not enough, we need to set out of
sync and do AL transactions for everything pdsk < D_INCONSISTENT on the
receiving side.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Diffstat (limited to 'drivers/block/drbd/drbd_req.c')
| -rw-r--r-- | drivers/block/drbd/drbd_req.c | 19 |
1 files changed, 14 insertions, 5 deletions
diff --git a/drivers/block/drbd/drbd_req.c b/drivers/block/drbd/drbd_req.c index 9e91a2545fc8..d26b213dbf15 100644 --- a/drivers/block/drbd/drbd_req.c +++ b/drivers/block/drbd/drbd_req.c @@ -942,12 +942,21 @@ allocate_barrier: if (local) { req->private_bio->bi_bdev = mdev->ldev->backing_bdev; - if (FAULT_ACTIVE(mdev, rw == WRITE ? DRBD_FAULT_DT_WR - : rw == READ ? DRBD_FAULT_DT_RD - : DRBD_FAULT_DT_RA)) + /* State may have changed since we grabbed our reference on the + * mdev->ldev member. Double check, and short-circuit to endio. + * In case the last activity log transaction failed to get on + * stable storage, and this is a WRITE, we may not even submit + * this bio. */ + if (get_ldev(mdev)) { + if (FAULT_ACTIVE(mdev, rw == WRITE ? DRBD_FAULT_DT_WR + : rw == READ ? DRBD_FAULT_DT_RD + : DRBD_FAULT_DT_RA)) + bio_endio(req->private_bio, -EIO); + else + generic_make_request(req->private_bio); + put_ldev(mdev); + } else bio_endio(req->private_bio, -EIO); - else - generic_make_request(req->private_bio); } /* we need to plug ALWAYS since we possibly need to kick lo_dev. |