too many open file descriptors messages since bind 9.4.2-P1 (port dns94)

[Bakul Shah]

On Tue, 15 Jul 2008 14:18:41 PDT JINMEI Tatuya / =?ISO-2022-JP?B?GyRCP0BMQEMjOkgbKEI=?= <Jinmei_Tatuya at isc.org>  wrote:
> At Tue, 15 Jul 2008 23:09:30 +0200,
> Kris Kennaway <kris at FreeBSD.org> wrote:
> 
> > > If that's regularly happening, I'm afraid recent P1 versions don't
> > > handle that well, and recommend you try 9.4.3b2 ore 9.5.1b1.
> > 
> > Or increase the number of file descriptors as a workaround, per my email :)
> 
> Does FreeBSD allow an application to increase FD_SETSIZE (at its
> compilation time)?  I thought FD_SETSIZE defaults to 1024.  Any
> 9.x.y-P1 versions can only open FD_SETSIZE sockets, regardless of the
> # FDs limit.

$ cvs log /sys/kern/kern_generic.c
...
revision 1.19
date: 1996/08/20 07:17:48;  author: smpatel;  state: Exp;  lines: +43 -15
Remove the kernel FD_SETSIZE limit for select().
...

Unless things have changed, you can completely ignore
FD_SETSIZE (& struct fd_set) and decide at runtime how many
fds you want in a select read/write set (subject to the
openfiles limit).

Hmm... things have reverted back....

cvs blame -r1.71 /sys/kern/kern_generic.c # the earliest reversal I can find
...
1.71         (peter    07-Feb-01):       * This is kinda bogus.  We have fd limi
ts, but that doesn't
1.71         (peter    07-Feb-01):       * map too well to the size of the pfd[] array.  Make sure
1.71         (peter    07-Feb-01):       * we let the process use at least FD_SETSIZE entries.
1.71         (peter    07-Feb-01):       * The specs say we only have to support OPEN_MAX entries (64).
1.71         (peter    07-Feb-01):       */
1.71         (peter    07-Feb-01):      lim = min((int)p->p_rlimit[RLIMIT_NOFILE].rlim_cur, maxfilesperproc);
1.71         (peter    07-Feb-01):      lim = min(lim, FD_SETSIZE);
1.71         (peter    07-Feb-01):      if (nfds > lim)
1.71         (peter    07-Feb-01):              return (EINVAL);

Sigh....  This is a mistake.  I don't see why a user is not
allowed to select on all the fds he can open.  The
corresponding log indicates perhaps the author didn't know
select used to work for # of fds > FD_SETSIZE.

revision 1.71
date: 2001/02/07 23:28:01;  author: peter;  state: Exp;  lines: +16 -8
The code I picked up from NetBSD in '97 had a nasty bug.  It limited
the index of the pollfd array to the number of fd's currently open, not
the maximum number of fd's.  ie: if you had 0,1,2 open, you could not
use pollfd slots higher than 20.  The specs say we only have to support
OPEN_MAX [64] entries but we allow way more than that.

> Besides, I guess that the P1 versions severely suffer from heavy
> overhead of select(2) when it regularly opens more than 1000 sockets.
> Even if 'too many open file' messages are gone, many users won't
> accept the increased load due to the overhead.  Beta versions use
> kqueue, eliminating the fundamental overhead as well as the (too low)
> limitation of # of descriptors.

Or more portably you can use poll(2).