Hey everyone,
This is an attempty to allow concurrent mounters and iterators to wait
on superblock state changes without having to hold s_umount. This is
made necessary by recent attempts to open block devices after superblock
creation and fixing deadlocks due to blkdev_put() trying to acquire
s_umount while s_umount is already held.
This is on top of Jan's and Christoph's work in vfs.super. Obviously not
for v6.6. I hope I got it right but this is intricate.
It reliably survives xfstests for btrfs, ext4, and xfs while
concurrently having 7 processes running ustat() hammering on
super_blocks and a while true loop that tries to mount a filsystem with
an invalid superblock hammering on sget{_fc}() concurrently as well.
Even with inserting an arbitrary 5s delay after generic_shutdown_super()
and blkdev_put() things work fine.
Thanks and don't hit me over the head with things.
Christian
Signed-off-by: Christian Brauner <brauner@xxxxxxxxxx>
---
Changes in v2:
- Rename various functions according to Jan's and Willy's suggestions.
- Remove smp_wmb() as smp_store_release() is enough.
- Remove hlist_unhashed() checks now that we wait on SB_DYING.
- Link to v1: https://lore.kernel.org/r/20230817-vfs-super-fixes-v3-v1-0-06ddeca7059b@xxxxxxxxxx
---
Christian Brauner (4):
super: use locking helpers
super: make locking naming consistent
super: wait for nascent superblocks
super: wait until we passed kill super
fs/fs-writeback.c | 4 +-
fs/internal.h | 2 +-
fs/super.c | 404 ++++++++++++++++++++++++++++++++++++++++-------------
include/linux/fs.h | 2 +
4 files changed, 313 insertions(+), 99 deletions(-)
---
base-commit: f3aeab61fb15edef1e81828da8dbf0814541e49b
change-id: 20230816-vfs-super-fixes-v3-f2cff6192a50
