Re: 2.6.28, rlimits, performance and debian etch

[David Daney <ddaney@caviumnetworks.com>]

Florian Weimer wrote:
> * Peter Palfrader:
> 
>> Turns out that script is forking a lot and something in it or python or
>> whereever closes all the file descriptors it doesn't want to pass on.
>> That is, it starts at zero and goes up to ulimit -n/RLIMIT_NOFILE and
>> closes them all with a few exceptions.
>>
>> Turns out that takes a long time when your limit -n is now 2^20 (1048576).
> 
> Interesting.
> 
> Can we make /proc more-or-less mandatory, so that the file descriptor
> list can be retrieved explicitly?

One problem is that for values of RLIMIT_NOFILE less than something like 
4096, it is much faster to call sys_close() on all possible values than 
iterate through a handful of open files from /proc/self/fd using 
opendir(3)/readdir(3).

Obviously for some large values of RLIMIT_NOFILE, this is no longer true.

People who have written code based on measuring the difference end up 
getting screwed when RLIMIT_NOFILE unexpectedly increases.

The real solution is to convert your user space programs to use the new 
syscalls that allow for race-free setting of close-on-exec.  Then you no 
longer need to mess around with iterating over these things.

David Daney