Hiya,
On 15/04/2021 17:45, Will Deacon wrote:
On Thu, Apr 15, 2021 at 04:26:46PM +0000, Ali Saidi wrote:
On Thu, 15 Apr 2021 16:02:29 +0100, Will Deacon wrote:
On Thu, Apr 15, 2021 at 02:25:52PM +0000, Ali Saidi wrote:
While this code is executed with the wait_lock held, a reader can acquire the lock without holding wait_lock. The writer side loops checking the value with the atomic_cond_read_acquire(), but only truly acquires the lock when the compare-and-exchange is completed successfully which isn’t ordered. The other atomic operations from this point are release-ordered and thus reads after the lock acquisition can be completed before the lock is truly acquired which violates the guarantees the lock should be making.
I think it would be worth spelling this out with an example. The issue appears to be a concurrent reader in interrupt context taking and releasing the lock in the window where the writer has returned from the atomic_cond_read_acquire() but has not yet performed the cmpxchg(). Loads can be speculated during this time, but the A-B-A of the lock word from _QW_WAITING to (_QW_WAITING | _QR_BIAS) and back to _QW_WAITING allows the atomic_cmpxchg_relaxed() to succeed. Is that right?
You're right. What we're seeing is an A-B-A problem that can allow atomic_cond_read_acquire() to succeed and before the cmpxchg succeeds a reader performs an A-B-A on the lock which allows the core to observe a read that follows the cmpxchg ahead of the cmpxchg succeeding.
We've seen a problem in epoll where the reader does a xchg while holding the read lock, but the writer can see a value change out from under it.
Writer | Reader 2
ep_scan_ready_list() | |- write_lock_irq() | |- queued_write_lock_slowpath() | |- atomic_cond_read_acquire() | | read_lock_irqsave(&ep->lock, flags); | chain_epi_lockless() | epi->next = xchg(&ep->ovflist, epi); | read_unlock_irqrestore(&ep->lock, flags); | atomic_cmpxchg_relaxed() | READ_ONCE(ep->ovflist);
Please stick this in the commit message, preferably annotated a bit like Peter's example to show the READ_ONCE() being speculated.
I can confirm that this patch fixes a problem observed in ep_scan_ready_list(.) whereby ovflist appeared to change when the write lock was held.
So please feel free to add: Tested-by: Steve Capper steve.capper@arm.com
Also, I have spent a decent chunk of time looking at the above issue and went through qrwlock, so FWIW, please feel free to add: Reviewed-by: Steve Capper steve.capper@arm.com
Cheers,