Messages in this thread | | | Date | Sun, 29 Jul 2012 01:13:34 -0400 (EDT) | From | Mikulas Patocka <> | Subject | Re: [dm-devel] [PATCH 2/3] Introduce percpu rw semaphores |
| |
On Sat, 28 Jul 2012, Eric Dumazet wrote:
> On Sat, 2012-07-28 at 12:41 -0400, Mikulas Patocka wrote: > > Introduce percpu rw semaphores > > > > When many CPUs are locking a rw semaphore for read concurrently, cache > > line bouncing occurs. When a CPU acquires rw semaphore for read, the > > CPU writes to the cache line holding the semaphore. Consequently, the > > cache line is being moved between CPUs and this slows down semaphore > > acquisition. > > > > This patch introduces new percpu rw semaphores. They are functionally > > identical to existing rw semaphores, but locking the percpu rw semaphore > > for read is faster and locking for write is slower. > > > > The percpu rw semaphore is implemented as a percpu array of rw > > semaphores, each semaphore for one CPU. When some thread needs to lock > > the semaphore for read, only semaphore on the current CPU is locked for > > read. When some thread needs to lock the semaphore for write, semaphores > > for all CPUs are locked for write. This avoids cache line bouncing. > > > > Note that the thread that is locking percpu rw semaphore may be > > rescheduled, it doesn't cause bug, but cache line bouncing occurs in > > this case. > > > > Signed-off-by: Mikulas Patocka <mpatocka@redhat.com> > > I am curious to see how this performs with 4096 cpus ?
Each cpu should have its own rw semaphore in its cache, so I don't see a problem there.
When you change block size, all 4096 rw semaphores are locked for write, but changing block size is not a performance sensitive operation.
> Really you shouldnt use rwlock in a path if this might hurt performance. > > RCU is probably a better answer.
RCU is meaningless here. RCU allows lockless dereference of a pointer. Here the problem is not pointer dereference, the problem is that integer bd_block_size may change.
> (bdev->bd_block_size should be read exactly once )
Rewrite all direct and non-direct io code so that it reads block size just once ...
Mikulas
| |