On Mon, Dec 14, 2009 at 06:21:55AM +0000, David Holland wrote:
> On Sun, Dec 13, 2009 at 08:02:24PM +0000, David Laight wrote:
> > Log Message:
> > Another, better, fix for PR/26567.
> > Only sleep once within each pipe_read/pipe_write call.
> > If there is no data/space available after we wakeup return ERESTART so
> > then the 'fd' number is validated again.
> > A simple broadcast of the cvs is then enough to evict the correct threads
> > when close() is called from an active thread.
>
> Isn't this going to cause a thundering herd problem if there are a lot
> of readers competing for the input? The first one will wake up and
> consume the available data, and then the rest will take a trip all the
> way to userspace and back.
>
> This may or may not be a practical problem for pipes but I'd expect
> the same approach to suck pretty hard for e.g. accept() on sockets. :(
Sockets have other problems - like not always having a syscall interface
for ERESTART to loop on.
I was actually thinking of making the fo_abort() call set a flag on
the socket and only do the ERESTART if that flag is set.
Then you only take the hit on 'file's associated with 'fd' that get
closed with a read/write in progress.
I also need to rename fo_abort() to fo_restart().
David
--
David Laight: da...@l8s.co.uk
