Re: [PATCH] Optimize libc_lock_lock for MIPS XLP.

[Maxim Kuvyrkov <maxim at codesourcery dot com>]

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?

>  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