Re: [PATCH 0/2] convert mmap_sem to a scalable rw_mutex

[Esben Nielsen <nielsen.esben@googlemail.com>]

On Sat, 12 May 2007, Peter Zijlstra wrote:

> On Sat, 2007-05-12 at 11:27 +0200, Esben Nielsen wrote:
> > On Fri, 11 May 2007, Peter Zijlstra wrote:
> >
> > > I was toying with a scalable rw_mutex and found that it gives ~10% reduction in
> > > system time on ebizzy runs (without the MADV_FREE patch).
> >
> > You break priority enheritance on user space futexes! :-(
> > The problems is that the futex waiter have to take the mmap_sem. And as
> > your rw_mutex isn't PI enabled you get priority inversions :-(
>
> Do note that rwsems have no PI either.
> PI is not a concern for mainline - yet, I do have ideas here though.
If PI wasn't a concern for mainline, why is PI futexes merged into the 
mainline?

I notice that the rwsems used now isn't priority inversion safe (thus 
destroyingthe perpose of having PI futexes). We thus already have a bug in 
the mainline.

I suggest making a rw_mutex which does read side PI: A reader boosts the 
writer, but a writer can't boost the readers, since there can be a large 
amount of those.

I don't have time to make such a rw_mutex but I have a simple idea for 
one, where the rt_mutex can be reused.

 struct pi_rw_mutex {
      int count; /*  0   -> unoccupied,
                     >0 -> the number of current readers,
                           Second highest bit: there are a waiting writer
                     -1 -> A writer have it. */
      struct rt_mutex mutex;
 }

Use atomic exchange on count.

When locking:

A writer checks if count <= 0. If so it sets the value to -1 and takes
the mutex. When it gets the mutex it rechecks the count and proceeds.
If count > 0 the writer sets the second highest bit and add itself to
the wait-list in the mutex and sleeps. (The mutex will now be in a state 
where owner==NULL but there are waiters. It must be cheched if the 
rt_mutex code can handle this.)

A reader checks if count >= 0. If so it does count++ and proceeds.
If count < 0 it takes the rtmutex. When it gets the mutex it sets the 
count to 1, unlocks the mutex and proceeds.

When unlocking:

The writer sets count to 0 or 0x8000000 (second highest bit) depending 
on how many waiters the mutex have and unlocks the mutex.

The reader checks if count is 0x80000001. If so it sets count to 0 and 
wakes up the first waiter on the mutex (if there are any). Otherwise it 
just do count--.

Esben
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/