p9fs: locking improvements for p9fs_stat_vnode_dotl()

[kib]

Summary

If the vnode is share-locked:
- Use vn_delayed_setsize() to avoid calling vnode_pager_setsize() with
  the vnode only shared locked.
- Interlock the vnode to get exclusive mode for updating the node
  fields.
  
Reciprocally, interlock the vnode in p9fs_getattr_dotl() to observe the
consistent values on read.

PR:     293492

p9fs: use atomics for updating node->flags

This should prevent seeing inconsistent flags values when updating it
under the shared vnode lock.

[markj]

sys/fs/p9fs/p9fs_vnops.c

1006

Surely this and the assignments above should be exclusively locked? Maybe add an XXX comment for it.

[kib]

sys/fs/p9fs/p9fs_vnops.c

1006

I used the vnode interlock there, instead of adding the XXX comment.

[markj]

sys/fs/p9fs/p9fs_vnops.c

975

I would not bother with the micro-optimization of avoiding the interlock. Oh, it's not an optimization, it's required to avoid a LOR with the VM object lock in vnode_pager_setsize(). I think a comment should explain this.

[kib]

Add comment.

[markj]

sys/fs/p9fs/p9fs_vnops.c

1002

p9fs_getattr_dotl() might still observe a torn write to these fields, since only the shared vnode lock will be held there, and the interlock is not acquired.

1006

Actually I think the exclusive vnode lock is required for this flags field, in the current scheme.

[kib]

sys/fs/p9fs/p9fs_vnops.c

1002

Yes, and it is enough to interlock on read too.

1006

IMO we can convert the flags updates to atomic.

[kib]

Use atomics for flags updates.
Interlock in getattr.