Messages in this thread | | | From | Manfred Spraul <> | Subject | [PATCH 0/5 V5] Clarify/standardize memory barriers for lock/unlock | Date | Wed, 31 Aug 2016 15:42:25 +0200 |
| |
Hi,
V5: Major restructuring based on input from Peter and Davidlohr.
As discussed before: If a high-scalability locking scheme is built with multiple spinlocks, then often additional memory barriers are required.
The documentation was not as clear as possible, and memory barriers were missing / superfluous in the implementation.
Patch 1: sem.c: Remove the smp_rmb() after spin_unlock_wait(). Patch 2: Documentation Patch 3: Add spinlock_store_acquire(), update ipc/sem.c Patch 4: Move smp_mb__after_unlock_lock to <linux/spinlock.h> Patch 5: Fix memory ordering for nf_conntrack
The patches are safe for all architectures, the default is smp_mb().
Patch 5 is larger than required, it rewrites the conntrack logic with the code from ipc/sem.c. I think the new code is simpler and more realtime-friendly.
@Peter: A hint if qspinlocks can omit the smp_mb() would be appreciated (everywhere or x86-only). The comments on top of queued_spin_unlock_wait() had convinced me that no additional barrier is required. But the comment also convinced me that on x86, no further barrier would be required for a full smp_mb() - and there you wrote that this is wrong.
@Andrew: The patches are relative to mmots. Could you include them in your tree, with the target of including in linux-next?
-- Manfred
| |