cvs commit: src/sys/kern kern_descrip.c

[Mike Silbersack]

On Thu, 19 Jun 2003, David Schultz wrote:

> Please allow me to re-make the point I believe I made when the
> patch was submitted.  (Of course, you're welcome to ignore it.)
> The reserve should be a tunable constant, not a fixed percentage.
> The idea, as I understand, is to allow the administrator to log in
> and correct the problem.  On a server with 50000 struct files, the
> reserve doesn't need to be 2000; a reserve on the order of 100
> open files should be adequate for just about any system.  In other
> words, the reserve doesn't need to scale with the size of the
> system.

Your point is good, but not relevant to file descriptors.  Why?  File
descriptors are dynamically allocated.  Even if we set maxfiles to
1000000, thereby making the reserve 5 million, we're still not actually
wasting anything.

Also, the idea of a huge huge reserve isn't necessarily a bad idea; it
seems somewhat good that with maxfiles = 50000 that apache / other root
daemons can probably run entirely within the reserve, thereby not being
susceptible to user interference no matter how many fds are allocated.

If we were talking about sockets / procs / anything that used a fixed
allocation set at boot time, then I'd agree with you.

Oh, btw; while running my fd hog program as root, cron decided to core.
Someone should take a look at it and figure out where it doesn't check the
return value of something which allocates a fd (or socket?)

Mike "Silby" Silbersack