Re: [PATCH] Optimize libc_lock_lock for MIPS XLP.

[Chris Metcalf <cmetcalf@tilera.com>]

On 6/14/2012 9:20 PM, Maxim Kuvyrkov wrote:
> On 15/06/2012, at 12:39 AM, Chris Metcalf wrote:
>
>> On 6/14/2012 1:03 AM, Maxim Kuvyrkov wrote:
>>> These two patches (libc part and ports part) optimize libc_lock_lock() macro that GLIBC uses for locking internally to take advantage of fetch_and_add instruction that is available as an extension on certain processors, e.g., MIPS-architecture XLP.
>>>
>>> The libc_lock_lock macros implement boolean lock: 0 corresponds to unlocked state and non-zero corresponds to locked state.
>> Just to be clear, if you put this comment somewhere when you commit, you
>> should say locks are tristate, where 0 is unlocked, 1 is locked and
>> uncontended, and >1 is locked and contended.
> Right, it's all coming back now.  I will update the comments to mention this.  [This optimization was written around 6 months ago, and not by me.  This and below points are worth elaborating on, thanks for bringing them up.]
>
> I've CC'ed Tom de Vries, who is the original author of patch.  Tom, please let us know if I'm misrepresenting the optimization or the rationale for its correctness.
>
>>> It is, therefore, possible to use fetch_and_add semantics to acquire lock in libc_lock_lock.  For XLP this translates to a single LDADD instruction.  This optimization allows architectures that can perform fetch_and_add faster than compare_and_exchange, such situation is indicated by defining the new macro "lll_add_lock".
>>>
>>> The unlocking counterpart doesn't require any change as it is already uses plain atomic_exchange operation, which, incidentally, also supported on XLP as a single instruction.
>> This seems like it would work well for a single thread acquiring the lock,
>> but I have some questions about it in the presence of multiple threads
>> trying to acquire the lock.
>>
>> First, the generic __lll_lock_wait() code assumes the contended value is
>> exactly "2".
> Um, not exactly.  __lll_lock_wait() *sets* the contended lock to a value of "2", but it will work as well with >2 values.
>
> void
> __lll_lock_wait (int *futex, int private)
> {
>   if (*futex == 2)
>     lll_futex_wait (futex, 2, private);
>
>   while (atomic_exchange_acq (futex, 2) != 0)
>     lll_futex_wait (futex, 2, private);
> }
>
>>  So if two or more threads both try and fail to acquire the
>> lock, the value will be >2.  This will cause the waiters to busywait,
>> spinning on atomic exchange instructions, rather than calling into
>> futex_wait().
> As I read it, in case of a contended lock __lll_lock_wait will reset the value of the lock to "2" before calling lll_futex_wait().  I agree that there is a timing window in which the other threads will see a value of the lock greater than "2", but the value will not get as high as hundreds or billions as it will be constantly reset to "2" in atomic_exchange in lll_lock_wait().
>
> I do not see how threads will get into a busywait state, though.  Would you please elaborate on that?

You are correct.  I was thinking the that the while loop had a cmpxchg, but
since it's just a straight-up exchange, the flow will be something like:

- Fail to initially call lll_futex_wait() if the lock is contended
- Fall through to while loop
- Spin as long as the lock is contended enough that *futex > 2
- Enter futex_wait

So a little busy under high contention, but probably settles out reasonably
well.

Since Tilera makes chips with 64 cores I tend to worry more about spinning
race cases with a lot of cores contending at once :-)

>>  I think it might be possible to change the generic code to
>> support the more general ">1" semantics of contended locks, but it might be
>> a bit less efficient, so you might end up wanting to provide overrides for
>> these functions on MIPS.  Even on MIPS it might result in a certain amount
>> of spinning since you'd have to hit the race window correctly to feed the
>> right value of the lock to futex_wait.
>>
>> Second, if a lock is held long enough for 4 billion threads to try to
>> acquire it and fail, you will end up with an unlocked lock. :-)  I'm not
>> sure how likely this seems, but it is a potential issue.  You might
>> consider, for example, doing a cmpxchg on the contended-lock path to try to
>> reset the lock value back to 2 again; if it fails, it's not a big deal,
>> since statistically I would expect the occasional thread to succeed, which
>> is all you need.
> Thank you,
>
> --
> Maxim Kuvyrkov
> CodeSourcery / Mentor Graphics

-- 
Chris Metcalf, Tilera Corp.
http://www.tilera.com