I've spent a lot of time staring at the locking code and speculating about all sorts of weird problems that could happen due to memory ordering or lost wakeups or if the custom locking is safe at all, also regarding the recent changes. Inevitably I found something but also wrote documentation. Please read it and if you see need for more clarifications, I'm happy to add it as I'm now in a state that things become temporarily obvious and trivial. I've tested it in fstests with KCSAN (the new concurrency sanitizer), no problems found but this is not considered sufficient, more tests will follow. David Sterba (5): btrfs: merge blocking_writers branches in btrfs_tree_read_lock btrfs: set blocking_writers directly, no increment or decrement btrfs: access eb::blocking_writers according to ACCESS_ONCE policies btrfs: serialize blocking_writers updates btrfs: document extent buffer locking fs/btrfs/locking.c | 184 +++++++++++++++++++++++++++++++++++---------- 1 file changed, 144 insertions(+), 40 deletions(-) -- 2.23.0
