POSIX states that poll() shall fail with EINVAL if nfds > OPEN_MAX.  In
this context, POSIX is referring to sysconf(OPEN_MAX), which is the value
of current->signal->rlim[RLIMIT_NOFILE].rlim_cur in the linux kernel, not
the compile-time constant which happens to also be named OPEN_MAX.  In the
current code, an application may poll up to max_fdset file descriptors,
even if this exceeds RLIMIT_NOFILE.  The current code also breaks
applications which poll more than max_fdset descriptors, which worked circa
2.4.18 when the check was against NR_OPEN, which is 1024*1024.  This patch
enforces the limit precisely as POSIX defines, even if RLIMIT_NOFILE has
been changed at run time with ulimit -n.

To elaborate on the rationale for this, there are three cases:

1) RLIMIT_NOFILE is at the default value of 1024

In this (default) case, the patch changes nothing.  Calls with nfds > 1024
fail with EINVAL both before and after the patch, and calls with nfds <=
1024 pass the check both before and after the patch, since 1024 is the
initial value of max_fdset.

2) RLIMIT_NOFILE has been raised above the default

In this case, poll() becomes more permissive, allowing polling up to
RLIMIT_NOFILE file descriptors even if less than 1024 have been opened.
The patch won't introduce new errors here.  If an application somehow
depends on poll() failing when it polls with duplicate or invalid file
descriptors, it's already broken, since this is already allowed below 1024,
and will also work above 1024 if enough file descriptors have been open at
some point to cause max_fdset to have been increased above nfds.

3) RLIMIT_NOFILE has been lowered below the default

In this case, the system administrator or the user has gone out of their
way to protect the system from inefficient (or malicious) applications
wasting kernel memory.  The current code allows polling up to 1024 file
descriptors even if RLIMIT_NOFILE is much lower, which is not what the user
or administrator intended.  Well-written applications which only poll
valid, unique file descriptors will never notice the difference, because
they'll hit the limit on open() first.  If an application gets broken
because of the patch in this case, then it was already poorly/maliciously
designed, and allowing it to work in the past was a violation of POSIX and
a DoS risk on low-resource systems.

With this patch, poll() will permit exactly what POSIX suggests, no more,
no less, and for any run-time value set with ulimit -n, not just 256 or
1024.  There are existing apps which which poll a large number of file
descriptors, some of which may be invalid, and if those numbers stradle
1024, they currently fail with or without the patch in -mm, though they
worked fine under 2.4.18.

Signed-off-by: Chris Snook <csnook@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>