[pbs-devel] [PATCH v5 proxmox-backup 0/8] fix #5982: check atime update is honored

[Christian Ebner <c.ebner@proxmox.com>]

These patches add a check to phase 1 of garbage collection and datastore
creation in order to detect when the filesystem backing the chunk store
does not honor atime updates. This avoids possible data loss for situations
where garbage collection could otherwise delete chunks still referenced by
a backup snaphost's index file.

The check is performed on a fixed size 4 MiB unencrypted and compressed
chunk of all-zeros, inserted if not present yet.
The Linux kernel timestamp granularity is taken into account by sleeping
for 1 second to avoid discarded atime update attempts by utimensat calls.
The test is enabled by default, but an opt-out option can be set via the
datastore tuning parameters for backwards compatibility.

Further, add a datastore tuning parameter to reduce the wait period for
chunk removal in phase 2 of garbage collection.

Most notable changes sice version 4 (thanks Fabian for feedback):
- Decouple gc-atime-safety-check from gc-atime-cutoff
- Improve logging and error handling
- fixed typo in docs (thanks Maximilliano for noticing).

Most notable changes sice version 3 (thanks Fabian for feedback):
- Drop check for relatime like behaviour, as this is not supported and
  does not show up in any of the tests performed on btrfs, cephfs, ext4,
  NFS3, NFS4, ntfs, SMB3_11, xfs or ZFS.
- Additionally check chunk inode to detect possible but very unlikely
  file changes, perform check once again in that case.
- Move atime cutoff selection and min_atime calculation to the same
  location, as they are logically related.

Most notable changes sice version 2 (thanks Fabian and Thomas for
comments and suggestions):
- Take into account Linux timestamp granularity, do not set timestamp
  to the past, as that introduces other error paths such as lack of
  permissions or fs limitations.
- Check relatime behavior, if atime behaviour is not honored. Fallback
  to original cutoff in that case.
- Adapt tuning parameter names.

Most notable changes sice version 1 (thanks Fabian and Thomas for
comments and suggestions):
- Optimize check by using the all zero chunk
- Enable the check by default and fail GC job if not honored, but allow
  to opt-out
- Add GC wait period tuning option

Link to the issue in the bugtracker:
https://bugzilla.proxmox.com/show_bug.cgi?id=5982

proxmox:

Christian Ebner (2):
  pbs api types: add garbage collection atime safety check flag
  pbs api types: add option to set GC chunk cleanup atime cutoff

 pbs-api-types/src/datastore.rs | 25 +++++++++++++++++++++++++
 1 file changed, 25 insertions(+)

proxmox-backup:

Christian Ebner (6):
  fix #5982: garbage collection: check atime updates are honored
  ui: expose GC atime safety check flag in datastore tuning options
  docs: mention GC atime update check for tuning options
  datastore: use custom GC atime cutoff if set
  ui: expose GC atime cutoff in datastore tuning option
  docs: mention gc-atime-cutoff as datastore tuning option

 docs/storage.rst                 |  16 ++++-
 pbs-datastore/src/chunk_store.rs | 111 ++++++++++++++++++++++++++-----
 pbs-datastore/src/datastore.rs   |  39 ++++++++++-
 src/api2/config/datastore.rs     |   1 +
 www/Utils.js                     |   9 +++
 www/datastore/OptionView.js      |  17 +++++
 6 files changed, 174 insertions(+), 19 deletions(-)

-- 
2.39.5



_______________________________________________
pbs-devel mailing list
pbs-devel@lists.proxmox.com
https://lists.proxmox.com/cgi-bin/mailman/listinfo/pbs-devel