lkml.org 
[lkml]   [2013]   [Jul]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH RFC 1/2] qrwlock: A queue read/write lock implementation
On 07/18/2013 02:46 PM, George Spelvin wrote:
>> Thank for the revision, I will make such a change in the next version of
>> my patch.
> I'm relying on you to correct any technical errors in my description.
> I just meant "something more like this", not impose that exact wording.
>
>> As I said in my response to Ingo, that change will make the lock more
>> unfair to the writers. However, I can run some tests to find out the
>> performance impact of such a way on the benchmarks that I used.
> You can do a lot with the basic structure if you're willing to
> go to XADD in the _lock_failed handlers.
>
> Adding writer priority is possible, but at least some call sites have
> to be reader-priority because they're recursive or are in interrupts,
> and readers don't disable interrupts.
>
> The basic solution is for writers to first detect if they're the *first*
> writer. If a write lock fails, look and see if it's> -RW_LOCK_BIAS.
>
> If not, drop it and spin reading until it's>= 0, then try again.
> (Possibly with XADD this time to combine the acquire and add.)
>
> But when we *are* the first writer, subtract an additional -RW_LOCK_BIAS/2.
> This tells pending readers "writer wating, back off!".
>
>
> What readers do when they see that bit set in rwlock->lock-1 depends on
> whether they're fair or unfair:
>
> - Fair readers re-increment the lock and spin waiting for it to become>
> -RW_LOCK_BIAS/2, at which point they try to reacquire.
> - Unfair readers say "ah, that means I can go ahead, thank you!".
>
>
> A waiting writer waits for the lock to equal -RW_LOCK_BIAS/2, meaning
> all readers have left. Then it adds -RW_LOCK_BIAS/2, meaning "write
> lock taken" and goes ahead.
>
> At this point, there is a thundering horde while the held-off readers
> get back in line. But hopefully it overlaps with the writer's lock tenure.
>
> When the writer drops its lock, the waiting readers go ahead, and the
> spinning writers advance in a thundering horde to contend for first place.
> (Again, the latency for this is hopefully covered by the readers'
> lock tenure.)

Thank for the suggestion. What you have proposed will be somewhat
similar to what my new code is doing with readers/writers spinning on
the cacheline without the waiting queue. I need to run some tests to see
if it can really help.

> I count 531 calls to read_lock in the kernel (10 of them are in
> lib/locking-selftest.c, called RL). I don't have any idea how difficult
> it would be to divide them into read_lock_fair and read_lock_unfair.

What I have in mind is to have 2 separate rwlock initializers - one for
fair and one for reader-bias behavior. So the lock owners can decide
what behavior do they want with a one line change.

Regards,
Longman


\
 
 \ /
  Last update: 2013-07-19 18:21    [W:0.314 / U:0.156 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site