An idea of remove MUTEX_WAKE_ALL

[Scott Long]

Daniel Eischen wrote:

> On Tue, 3 Jan 2006, John Baldwin wrote:
> 
> 
>>On Sunday 01 January 2006 02:21 am, prime wrote:
>>
>>>Hi hackers,
>>>   I have an idea about remove the kernel option MUTEX_WAKE_ALL.
>>>   When we unlock the mutex(in _mtx_unlock_sleep),we can directly
>>>give the lock to the first thread waiting on the turnstile.And a
>>>thread gets the mutex after he returned from turnstile_wait so he
>>>can simply jump out the _obtain_lock loop in _mtx_lock_sleep.
>>>This makes a mutex always be owned by a thread when there are threads
>>>waiting on the turnstile,so priority inheritance can work now.
>>>   This idea need only a few changes in kern/kern_mutex.c .But when
>>>NO_ADAPTIVE_MUTEXS not set,it makes threads that spinning on other CPU
>>>to get the mutex have to spin for a long time,and this makes the short
>>>term mutex more expensive(maybe should use spin mutex instead).
>>>
>>>What do think about the idea? Thanks.
>>
>>Sun actually found that the performance was better when you did MUTEX_WAKE_ALL
>>because once you woke up N threads, if they don't all resume at once then
>>they will acquire the lock in sequence and the lock acquires and releaes will
>>all be simple ones rather than all being the complicated contested case.
>>There are more details in _Solaris Internals_.
> 
> 
> Yes, but doesn't this partly rely on having the threads spin(*)
> for a bit if the current lock owner is running on another CPU?
> Do we currently do that?
> 
> (*) No, I am not referring to spin mutexes.
> 

Adaptive mutexes are enabled by default and have been for at least a
year.

Scott