7291 Attempting to destroy ZFS filesystem and create ZFS volume with …

[deiter]

Please see https://www.illumos.org/issues/7291 for details.

1. In case of ZFS filesystem with nested child datasets we have /dev/zvol/{r,}dsk directories hierarchy: /dev/zvol/{r,}dsk/pool/filesystem/child1/child2/child3
2. When pool/filesystem/child1 filesystem and its child datasets are destroyed, all /dev/zvol/{r,}dsk directories hierarchy for pool/filesystem/child1 and sdev cache still exists (by design of usr/src/uts/common/fs/dev)
3. After re-creation of pool/filesystem/child1 with another dataset type (volume instead of filesystem) and accessing this new ZFS volume (devzvol_lookup -> devname_lookup_func)
   sdev_cache_update(ddv, &dv, nm, SDEV_CACHE_DELETE) call updates the in-core directory cache ONLY.
4. Therefore all /dev/zvol/{r,}dsk/pool/filesystem/child1 directories hierarchy still exists.
   So we have dv->sdev_nlink == number_of_child_datasets + 2 for /dev/zvol/{r,}dsk/pool/filesystem/child1 and system panic on VERIFY:

http://src.illumos.org/source/xref/illumos-gate/usr/src/uts/common/fs/dev/sdev_subr.c#3133

Proposed patch: sdev_cleandir for stalled directory path.

Reviewed by: Sanjay Nadkarni [email protected]
Reviewed by: Saso Kiselkov [email protected]
Reviewed by: Roman Strashkin [email protected]
Reviewed by: Alek Pinchuk [email protected]

Patch was successfully tested and using internally at Nexenta and already included into the release NexentaStor 5.0.

Thank you!

[ahrens]

@deiter You'll have more luck getting this reviewed and integrated over at the OpenZFS github project -- I don't think that the illumos PR's are monitored. I took a quick look and it's great that you're adding a new test for this!