On 02/13/2019 02:45 AM, Ingo Molnar wrote:

* Waiman Long [off-list ref] wrote:

quoted

I looked at the assembly code in arch/x86/include/asm/rwsem.h. For both
trylocks (read & write), the count is read first before attempting to
lock it. We did the same for all trylock functions in other locks.
Depending on how the trylock is used and how contended the lock is, it
may help or hurt performance. Changing down_read_trylock to do an
unconditional cmpxchg will change the performance profile of existing
code. So I would prefer keeping the current code.

I do notice now that the generic down_write_trylock() code is doing an
unconditional compxchg. So I wonder if we should change it to read the
lock first like other trylocks or just leave it as it is.

No, I think we should instead move the other trylocks to the 
try-for-ownership model as well, like Linus suggested.

That's the general assumption we make in locking primitives, that we 
optimize for the common, expected case - which would be that the trylock 
succeeds, and I don't see why trylock primitives should be different.

In fact I can see more ways for read-for-sharing to perform suboptimally 
on larger systems.

I don't mind changing to the try-for-ownership model for rwsem and
mutex. I do have some concern to do that for spinlock. Some of the lock
slowpath code do optimistic trylock. Making them unconditional cmpxchg
will impact lock contention performance.

I will update this rwsem patch to make the change while I am working on
it. For other locks, I will suggest we go slow and carefully evaluate
the performance implication before we make the changes.

Cheers,
Longman



_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel