[pve-devel] [PATCH container/storage v3] add fsfreeze/thaw for rbd snapshots

[Stoiko Ivanov <s.ivanov@proxmox.com>]

this patchset addresses #2991 and #2528.

v2->v3:
* incoroprated Wolfgang's feedback (huge thanks!!):
** /proc/$pid/root contains as magic-link [0] to the container's rootfs -
   so use these for the FIFREEZE/FITHAW ioctls instead of fork+nsenter
** thus moved the fsfreeze_mountpoint to PVE::LXC::Config (it's only needed
   there)

[0] https://www.kernel.org/doc/html/latest/filesystems/path-lookup.html

original cover-letter for the v2:
v1->v2:
mostly incorporated Thomas' feedback (huge thanks!!):
* moved fsfreeze from pve-common to pve-container (it's only used here, and
  it avoids one versioned dependency bump).
* for this needed to drop O_CLOEXEC (only defined in PVE::Tools) flag from
  sysopen (fsfreeze(8) does not use it either...)
* increased APIVER and APIAGE in PVE::Storage
* don't use sync_container_namespace for unfreezing (no need to parse
  /proc/mounts and pass it to the child process) - moved that part to a new
  sub
* found a tiny typo in a comment (patch 1/2 for pve-storage)

original cover-letter for v1:

As discussed in #2991 (and off-list with Wolfgang B. and Dominik) - it does
not address the fundamental problem of the snapshot being created outside of
the open krbd block-device, by an independend 'rbd' call (which is most likely
the reason for the inconsistency).

However according to the reporter in #2991 it does help in their case to
actually get backups of their containers.

I put the ioctl call inside sync_container_namespace since it:
* should happen shortly after the syncfs call
* needs to happen inside the container's mount namespace (else we'd need to
  mount the filesystem in order to freeze/thaw it - see the proposed patch
  in #2528)

and I wanted to avoid to fork+nsenter for each volume twice (in
__snapshot_create_vol_snapshs_hook)

Would be grateful for feedback if this approach is ok (reading containerconfig
+ storage config in __snapshot_freeze) or if some other way would be nicer.

Tested on my testsetup with a ceph-backed container (and 2 additional
mountpoints (one ceph, one on LVM thin).

pve-storage:
Stoiko Ivanov (2):
  fix typo in comment
  add check for fsfreeze before snapshot

 PVE/Storage.pm           | 18 +++++++++++++++---
 PVE/Storage/Plugin.pm    |  4 ++++
 PVE/Storage/RBDPlugin.pm |  5 +++++
 3 files changed, 24 insertions(+), 3 deletions(-)
i
pve-container:
Stoiko Ivanov (2):
  add fsfreeze helper:
  snapshot creation: fsfreeze mountpoints, if needed

 src/PVE/LXC/Config.pm     | 47 +++++++++++++++++++++++++++++++++++++++
 src/test/snapshot-test.pm | 12 +++++++++-
 2 files changed, 58 insertions(+), 1 deletion(-)

-- 
2.20.1