I don't see how we could use them either. They are a solution
for a different problem.
Mark.
On Thu, 25 Sep 2003, Emmanuel Allaud wrote:
> /I did not have news from my request on linux-kernel, but I came across
> this interview of Rusty Russel where he talks about futexes (URL :
> http://kerneltrap.org/node/view/892) :
> /
>
> /
> ---------
> /
>
> /Rusty Russell/: OK, a futex ("Fast Userspace Mutex") isn't a mutex at
> all. It was in a (much) earlier implemetation, but now it's just a
> waitqueue primitive accessible to userspace. FUTEX_WAKE says "wake
> anyone waiting on this address", and FUTEX_WAIT says "when I read ADDR,
> it contained VALUE: if it still contains that value when you look at it,
> sleep until someone calls FUTEX_WAKE on this address".
>
> It's used as follows. Two processes (or threads) share some memory, and
> agree to use a part of it as a lock. 1 means unlocked, 0 means locked, <
> 0 means locked and someone is waiting. To grab the lock, you do an
> (architecture-specific) atomic decrement operation: if it hits 0, then
> it was 1, and you've got the lock. Otherwise you wait for a while. When
> you want to release the lock, you set it to 1 again. This is really fast
> if there's no contention on the lock: not a system call in sight.
>
> Before futexes, the "wait a while" was usually implemented "try three
> times and then call yield()". The problem with yield is that it has no
> clear semantics: the kernel has *no* idea what you are waiting for, so
> it has to guess when to wake you up. In a word, "yield()" is always a hack.
>
> With futexes, you say "this address was -1 when I finished with it,
> sleep on it": the kernel atomically checks the value and either puts you
> to sleep or returns EAGAIN, and you try again. When you release a lock,
> if it's 0, you can simply increment it back to 1, otherwise, you set it
> to 1 and call FUTEX_WAKE, because you know someone is waiting. You can
> read some example code here
> <http://www.kernel.org/pub/linux/kernel/people/rusty/futex-2.2.tar.gz>,
> and of course in the current glibc sources.
>
> -------------------
>
> The thing here is that I still don't know how we would use futexes...
> Bye
> Manu
>
>
>
_______________________________________________
Devel mailing list
[EMAIL PROTECTED]
http://XFree86.Org/mailman/listinfo/devel
