RE: [PATCH 1/4] staging: lustre: obdclass: change spinlock of key to rwlock

From: NeilBrown
Date: Thu May 03 2018 - 19:26:57 EST


On Thu, May 03 2018, David Laight wrote:

> From: James Simmons
>> Sent: 02 May 2018 19:22
>> From: Li Xi <lixi@xxxxxxx>
>>
>> Most of the time, keys are never changed. So rwlock might be
>> better for the concurrency of key read.
>
> OTOH unless there is contention on the spin lock during reads the
> additional cost of a rwlock (probably double that of a spinlock)
> will hurt performance.

That's roughly what I was going to say - rwlocks are rarely a win.
I think the second patch which caused the lock to be taken less often
would have a bigger impact that the switch to rwlocks.

However I suspect a better approach would be to investigate some sort of
lockless solution.
I think the use of the spinlock in lu_context_key_register() could be
replaced with a careful cmp_xchg(). I'm less sure about
lu_context_key_degister(), but it might be possible.

>
> ...
>> - spin_lock(&lu_keys_guard);
>> + read_lock(&lu_keys_guard);
>> atomic_inc(&lu_key_initing_cnt);
>> - spin_unlock(&lu_keys_guard);
>> + read_unlock(&lu_keys_guard);
>
> WTF, seems unlikely that you need to hold any kind of lock
> over an atomic_inc().
>
> If this is just ensuring that no code holds the lock then
> it would need to request the write_lock().
> (and would need a comment)

There is a comment - that patch showed the last 2 lines of it.
This is for synchronization with lu_context_key_quiesce().
That spins(!! calling schedule, but still... not good) until
the lu_key_initing_cnt is zero while it holds the write lock.
Then it is sure that the code protected by this counter isn't
running.
I'm sure this can be improved! I would need to study it carefully to
see how.

Note that I don't object to these patches going in - if they provide a
measurable improvement which seems likely, then in they go. But I
hope the code won't stay like this long term.

Thanks,
NeilBrown

Attachment: signature.asc
Description: PGP signature