[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <Pine.LNX.4.64.0703052230300.4907@alien.or.mcafeemobile.com>
Date: Mon, 5 Mar 2007 22:31:11 -0800 (PST)
From: Davide Libenzi <davidel@...ilserver.org>
To: Eric Dumazet <dada1@...mosbay.com>
cc: "H. Peter Anvin" <hpa@...or.com>,
Linus Torvalds <torvalds@...ux-foundation.org>,
Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
Andrew Morton <akpm@...ux-foundation.org>
Subject: Re: [patch v2] epoll use a single inode ...
On Tue, 6 Mar 2007, Eric Dumazet wrote:
> Davide Libenzi a écrit :
> > On Mon, 5 Mar 2007, H. Peter Anvin wrote:
> >
> > > Davide Libenzi wrote:
> > > > Right now is using:
> > > >
> > > > this.len = sprintf(name, "[%u.%d]", current->pid, fd);
> > > >
> > > > That should be unique and not have the wraparound problem. Ack?
> > > >
> > > NAK, very much NAK.
> > >
> > > File descriptors aren't file structures, they're *pointers* to file
> > > structures.
> > >
> > > It's perfectly possible -- downright common -- for a file descriptor to be
> > > inherited by another process, and then the pid is recycled -- collision.
> >
> > Ugh! Right! 64 bit counter it is ... :)
>
> For epoll, I suspect this is harmless : Programs dont allocate epolls fd every
> milli second, but at startup only.
>
> For pipes/sockets, using a 64 bits would be problematic, because sprintf()
> uses a divide for each digit. And a divide is slow. Ten divides are *very*
> slow.
This would be used for epoll, signalfd and timerfd. And the printf format
is %llu ;)
- Davide
Powered by blists - more mailing lists