Re: [PATCH v2] locking/rwsem: Take read lock immediate if queue empty with no writer

From: Waiman Long
Date: Wed Jul 18 2018 - 11:15:46 EST


On 07/13/2018 02:30 PM, Waiman Long wrote:
> It was discovered that a constant stream of readers might cause the
> count to go negative most of the time after an initial trigger by a
> writer even if no writer was present afterward. As a result, most of the
> readers would have to go through the slowpath reducing their performance.
>
> To avoid that from happening, an additional check is added to detect
> the special case that the reader in the critical section is the only
> one in the wait queue and no writer is present. When that happens, it
> can just have the lock and return immediately without further action.
> Other incoming readers won't see a waiter is present and be forced
> into the slowpath.
>
> The additional code is in the slowpath and so should not have an impact
> on rwsem performance. However, in the special case listed above, it may
> greatly improve performance.
>
> The issue was found in a customer site where they had an application
> that pounded on the pread64 syscalls heavily on an XFS filesystem. The
> application was run in a recent 4-socket boxes with a lot of CPUs. They
> saw significant spinlock contention in the rwsem_down_read_failed() call.
> With this patch applied, the system CPU usage went from 85% to 57%,
> and the spinlock contention in the pread64 syscalls was gone.
>
> v2: Add customer testing results and remove wording that may cause
> confusion.
>
> Signed-off-by: Waiman Long <longman@xxxxxxxxxx>
> ---
> kernel/locking/rwsem-xadd.c | 13 ++++++++++++-
> 1 file changed, 12 insertions(+), 1 deletion(-)
>
> diff --git a/kernel/locking/rwsem-xadd.c b/kernel/locking/rwsem-xadd.c
> index 3064c50..bf0570e 100644
> --- a/kernel/locking/rwsem-xadd.c
> +++ b/kernel/locking/rwsem-xadd.c
> @@ -233,8 +233,19 @@ static void __rwsem_mark_wake(struct rw_semaphore *sem,
> waiter.type = RWSEM_WAITING_FOR_READ;
>
> raw_spin_lock_irq(&sem->wait_lock);
> - if (list_empty(&sem->wait_list))
> + if (list_empty(&sem->wait_list)) {
> + /*
> + * In the unlikely event that the task is the only one in
> + * the wait queue and a writer isn't present, it can have
> + * the lock and return immediately without going through
> + * the remaining slowpath code.
> + */
> + if (unlikely(atomic_long_read(&sem->count) >= 0)) {
> + raw_spin_unlock_irq(&sem->wait_lock);
> + return sem;
> + }
> adjustment += RWSEM_WAITING_BIAS;
> + }
> list_add_tail(&waiter.list, &sem->wait_list);
>
> /* we're now waiting on the lock, but no longer actively locking */

Ping!

Any comment on this one?

Cheers,
Longman