apr-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Cliff Woolley <cliffwool...@yahoo.com>
Subject Re: Buckets destroy not cleaning up private structures?
Date Sun, 29 Apr 2001 03:00:23 GMT
On Sat, 28 Apr 2001, Justin Erenkrantz wrote:

> Aren't some of the resource leakages in httpd coming from the fact
> that the buckets aren't cleaning up after themselves?  They do
> reference counting, but my guess is that they should close their
> private structures as well when they are no longer around.

There's a sideband discussion on this.  Ryan's strongly opposed to closing
the files/mmaps when the last bucket goes away, as I believe he pointed
out in another response to this message, because it incurs an extra
syscall during the request proper.  I've argued against that to little/no

> mod_core opens a file and gives it apr_bucket_file_create.  After
> that point, shouldn't the bucket "own" the fd and close it when
> it is done?

That's what Greg and I have said.  But we just can't convince Ryan.  And
admittedly he has a good reason... it's a toss-up.  There are pros and
cons either way.  1.3 did it by letting the pool do the work, so I guess
we should too.

> I'm going to implement this in my local tree and see if it blows
> up things spectacularly.  -- justin

There are issues to be aware of if you're going to try it.  For example,
you can't count on an MMAP surviving if its underlying file gets closed.
Assuming Ryan doesn't object too strongly, I plan on changing the file
buckets very soon to be more like pool buckets, in that when one file
bucket gets MMAPed, all of its siblings get lazily morphed to MMAP-type as
well.  The main issue you need to watch out for is in which cases you can
safely close the file--that's tough to do with the current code but should
be easy with the change I propose.  Try it on MMAP's if you want... that
should be safe, or at least safer than closing the files.  PS: Another
reason not to close the files right now is that it would break the module
that caches FD's.  I've got a proposed fix for that as well.


   Cliff Woolley
   Charlottesville, VA

View raw message