rmlock(9) two additions

[Max Laier]

Hi again,

On Monday 16 August 2010 05:15:21 I wrote:
> I'd like to run two additions to rmlock(9) by you:
> 
> 1) See the attached patch.  It adds rm_try_rlock() which avoids taking the
> mutex if the lock is currently in a write episode.  The only overhead to
> the hot path is an additional argument and a return value for _rm_rlock*. 
> If you are worried about that, it can obviously be done in a separate code
> path, but I reckon it not worth the code crunch.  Finally, there is one
> additional branch to check the "trylock" argument, but that's well past
> the hot path.
> 
> 2) No code for this yet - testing the waters first.  I'd like to add the
> ability to replace the mutex for writer synchronization with a general lock
> - especially to be able to use a sx(9) lock here.

the attached patch includes code for this as well and updates the man pages 
accordingly.  Instead of using a general lock, I just added a rm_init_flags() 
flag called "RM_SLEEPABLE" that changes the lock to use an sx(9) lock instead 
of the mutex.  This is enough to support the use case I have in mind.  See 
below for details on that from my original post.  There is a bit of "creative" 
use of witness in this patch, but the result provides proper warnings in all 
cases.

Again:  Are there any objections or concerns against this?  Otherwise I will 
commit it soon.

Thanks,
  Max

> The reason for #2 is the following use case in a debugging facility:
> 
> 	"reader":
> 		if (rm_try_rlock()) {
> 			grab_per_cpu_buffer();
> 			fill_per_cpu_buffer();
> 			rm_runlock();
> 		}
> 
> 	"writer" - better exclusive access thread:
> 		rm_wlock();
> 		collect_buffers_and_copy_out_to_userspace();
> 		rm_wunlock();
> 
> This is much cleaner and possibly cheaper than the various hand rolled
> versions I've come across, that try to get the same synchronization with
> atomic operations.  If we could sleep with the wlock held, we can also
> avoid copying the buffer contents, or swapping buffers.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: rmlock.try_and_sleep.diff
Type: text/x-patch
Size: 12570 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-arch/attachments/20100820/a45c5981/rmlock.try_and_sleep.bin