Re: [patch 14/22] pollfs: pollable futex

[Davi Arnaut <davi@haxent.com.br>]

Ulrich Drepper wrote:
> On 5/2/07, Davi Arnaut <davi@haxent.com.br> wrote:
>> It's quite easy to implement this scheme by write()ing the futexes all
>> at once but that would break the one futex per fd association. For
>> atomicity: if one of the futexes can't be queued, we would rollback
>> (unqueue) the others.
>>
>> Sounds sane?
>
> I don't know how you use "unqueue" in this context.  If a queued futex
> is one which is /locked/ by te call, then yes, this is the semantics
> needed.  Atomically locking a number of futexes means that if one of
> the set cannot be locked all operations done to lock the others have
> to be undone.  It's an all-or-nothing situation.
The waits are queued, thus then can be "unqueued". It's quite simple to
extend futex_wait_queue() to support this, but again you are thinking of
locks while what I want is fast events.
> Locking is not as easy as you might think, though.  For non-PI futexes
> there is deliberately no protocol in place describing what "locked"
> means.  The locking operation has to be customizable.  This is what
> the FUTEX_OP_* stuff is about.
Events are simple. A event is either signaled or not. A futex value 0 means
not signaled, 1+ signaled.
> And you wrote that currently each futex needs its own file descriptor.
> So this would have to be changed, too.
If it's really worth, I have no problem with it.

--
Davi Arnaut