DAOS-19526 dfs: Enable Hardlink support for dfuse - #19065

This commit implements the following changes:
1. Introduced a callback handler for hardlink (link) in dfuse.
2. Modified the dfuse unlink handler to treat the unlink of a hardlink
   differently from the removal of the last link.
3. Introduced two new dfs functions for unlink and rename that tell the
   caller whether the target file was really removed/deleted. These apis
   are called by dfuse rename and unlink handlers.
4. Added support for tracking multiple dentries per inode entry. The
   entries are updated as they are encountered during link, unlink,
   rename and lookup/readdir operations, and every name is invalidated
   when the inode is dropped.
5. NLT functional dfuse tests and DFS unit tests for hardlinks.

Allow-unstable-test: true

Signed-off-by: Sherin T George <sherin-t.george@hpe.com>

…9526

Signed-off-by: Sherin T George <sherin-t.george@hpe.com>

…link

Rework for the dfuse hardlink support.

- dfs: in remove_hardlink(), set the 'deleted' out-param on the branch that
  drops the final link (link_cnt reaches 0), so the value is set on every
  path and stays correct across a transaction restart instead of retaining
  a stale result from an earlier attempt; previously it was only cleared to
  false when a link survived.

- dfuse/unlink: evict the surviving inode's metadata cache in
  dfuse_hardlink_removed() so a peer-name unlink/rename no longer leaves
  stale st_nlink/ctime for the kernel's follow-up GETATTR to read.

- dfuse/fuseops: reject cross-container hardlinks in df_ll_link() with
  EXDEV before calling into DFS, instead of relying on an oid lookup that
  fails with a misleading ENOENT (oids are container-scoped).

- dfuse/core: in dfuse_ie_dentry_replace(), de-duplicate when the rename
  destination name is already tracked (primary or secondary) so the
  secondary list never gains a duplicate entry.  Defer freeing a removed
  dentry in both dfuse_ie_dentry_replace() and dfuse_ie_dentry_set_single()
  until after the spinlock is released, matching the other dentry helpers.

- tests: add test_hardlink_stale_notify_delete (out-of-band removal then a
  local final unlink to exercise notify_delete on stale same-parent
  dentries), add test_hardlink_stale_name, add test_hardlink_cross_container
  (pool-level mount with two containers, os.link across them expects EXDEV),
  and assert the refreshed link count after a peer-name unlink in
  test_hardlink_caching.

Features: dfs dfuse pil4dfs
Allow-unstable-test: true

Signed-off-by: Sherin T George <sherin-t.george@hpe.com>

Addressed a regression seen in CI/CD where open_stat() in DFS was not
sanitizing the mode passed as argument. The fix strips the internal bits
(DFS_EXTERNAL_MODE) from caller-supplied modes at the DFS boundary so an
untrusted mode can never forge internal state:
- open_stat(): mask the mode before it is stored; legitimate hardlinks are
  still flagged from fetched entry state after creation.
- dfs_mkdir(): mask the mode before building the directory entry.
- dfs_osetattr(): return the sanitized external mode in the out stat instead
  of echoing the caller's raw mode.

Features: dfs dfuse pil4dfs
Allow-unstable-test: true

Signed-off-by: Sherin T George <sherin-t.george@hpe.com>

Addressed a regression seen in CI/CD where simul tests expected hardlink
related tests to fail. Now hardlink related tests are removed from the
faillist.

Features: dfs dfuse pil4dfs
Allow-unstable-test: true

Signed-off-by: Sherin T George <sherin-t.george@hpe.com>

Rework addressing review feedback on the hardlink support.

- dfs_link() built the returned stat directly from the stored inode
  entry, whose mtime lags the array's modification epoch.  Route it
  through update_stbuf_times() like every other stat path so the
  reconciled mtime is returned.  Without this, dfuse's attribute cache
  is poisoned with the stale value and an immediate stat after link()
  reports the modification time moving backwards.

- dfuse_cb_link() never evicted the destination directory's cache, so a
  retained readdir stream kept dfuse's shared readdir handle resident
  and an independently opened stream was served a stale listing that
  omitted the new name.  Evict it, matching create/unlink/rename.

- dfuse_ie_inode_delete() issued fuse_lowlevel_notify_delete()
  synchronously on the FUSE worker pool.  That call blocks acquiring the
  parent's kernel i_rwsem, which may be held by another in-flight
  operation waiting on the same pool, stalling the worker.  Queue the
  deletes on the invalidation thread instead, extending struct
  dfuse_inval_item with an 'ino' and a 'delete_entry' flag.  Delaying is
  safe because the kernel matches the child nodeid before deleting, so a
  name re-created in the meantime is left untouched.

Added NLT tests test_hardlink_dir_cache and test_hardlink_mtime, which
fail on the unfixed code and pass with these changes.

Features: dfs dfuse pil4dfs
Allow-unstable-test: true

Signed-off-by: Sherin T George <sherin-t.george@hpe.com>