Re: [PATCH RFC 1/2] qrwlock: A queue read/write lock implementation

[Ingo Molnar <mingo@kernel.org>]

* Waiman Long <waiman.long@hp.com> wrote:

> On 07/18/2013 03:42 AM, Ingo Molnar wrote:
> >* Waiman Long<waiman.long@hp.com>  wrote:
> >
> >>>>+ *    stealing the lock if come at the right moment, the granting of the
> >>>>+ *    lock is mostly in FIFO order.
> >>>>+ * 2. It is faster in high contention situation.
> >>>Again, why is it faster?
> >>The current rwlock implementation suffers from a thundering herd
> >>problem. When many readers are waiting for the lock hold by a writer,
> >>they will all jump in more or less at the same time when the writer
> >>releases the lock. That is not the case with qrwlock. It has been shown
> >>in many cases that avoiding this thundering herd problem can lead to
> >>better performance.
> >Btw., it's possible to further optimize this "writer releases the lock to
> >multiple readers spinning" thundering herd scenario in the classic
> >read_lock() case, without changing the queueing model.
> >
> >Right now read_lock() fast path is a single atomic instruction. When a
> >writer releases the lock then it makes it available to all readers and
> >each reader will execute a LOCK DEC instruction which will succeed.
> >
> >This is the relevant code in arch/x86/lib/rwlock.S [edited for
> >readability]:
> >
> >__read_lock_failed():
> >
> >0:      LOCK_PREFIX
> >         READ_LOCK_SIZE(inc) (%__lock_ptr)
> >
> >1:      rep; nop
> >         READ_LOCK_SIZE(cmp) $1, (%__lock_ptr)
> >         js      1b
> >
> >         LOCK_PREFIX READ_LOCK_SIZE(dec) (%__lock_ptr)
> >         js      0b
> >
> >         ret
> >
> >This is where we could optimize: instead of signalling to each reader that
> >it's fine to decrease the count and letting dozens of readers do that on
> >the same cache-line, which ping-pongs around the numa cross-connect
> >touching every other CPU as they execute the LOCK DEC instruction, we
> >could let the _writer_ modify the count on unlock in essence, to the exact
> >value that readers expect.
> >
> >Since read_lock() can never abort this should be relatively
> >straightforward: the INC above could be left out, and the writer side
> >needs to detect that there are no other writers waiting and can set the
> >count to 'reader locked' value - which the readers will detect without
> >modifying the cache line:
> >
> >__read_lock_failed():
> >
> >0:      rep; nop
> >         READ_LOCK_SIZE(cmp) $1, (%__lock_ptr)
> >         js      0b
> >
> >         ret
> >
> >(Unless I'm missing something that is.)
> >
> >That way the current write_unlock() followed by a 'thundering herd' of
> >__read_lock_failed() atomic accesses is transformed into an efficient
> >read-only broadcast of information with only a single update to the
> >cacheline: the writer-updated cacheline propagates in parallel to every
> >CPU and is cached there.
> >
> >On typical hardware this will be broadcast to all CPUs as part of regular
> >MESI invalidation bus traffic.
> >
> >reader unlock will still have to modify the cacheline, so rwlocks will
> >still have a fundamental scalability limit even in the read-only usecase.
> 
> I think that will work. The only drawback that I can see is the fairness 
> argument. The current read/write lock implementation is unfair to the 
> writer. That change will make it even more unfair to the writer and 
> there is no easy way to detect a waiting writer unless we change the 
> structure to add such a field. As a result, a steady stream of readers 
> will have a higher chance of blocking out a writer indefinitely.

The effect will have to be measured - but I don't think it's particularly 
hard to tune the fairness balance between readers and writers: the change 
I suggested would only affect the case when a writer already holding the 
lock unlocks it.

But when a writer already holds the lock it can decide to pass that lock 
to another writer-spinning instead of unlocking to all readers. This too 
should be relatively straightforward to implement because neither 
read_lock() nor write_lock() can abort and race.

Instead of doing:

static inline void arch_write_unlock(arch_rwlock_t *rw)
{
        asm volatile(LOCK_PREFIX WRITE_LOCK_ADD(%1) "%0"
                     : "+m" (rw->write) : "i" (RW_LOCK_BIAS) : "memory");
}

the current owner could check whether there are other writers waiting and 
could drop into a slowpath that passes ownership to one of the writers via 
toggling bit 30 or so. This reduces the max number of writers by a factor 
of 2.

But I'd implement this only if it proves to be a problem in practice.

I'd strongly suggest to first address the thundering herd problem of 
write_unlock() and see how it affects scalability - before totally 
replacing it all with a new, fundamentally heavier locking primitive!

Thanks,

	Ingo