[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Dazuko-devel] Dazuko driver takes lots of cpu time, bottleneck disc
From: |
John Ogness |
Subject: |
Re: [Dazuko-devel] Dazuko driver takes lots of cpu time, bottleneck discovered |
Date: |
Tue, 05 Oct 2004 09:43:00 +0200 |
User-agent: |
Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.6) Gecko/20040113 |
Sami Tikka wrote:
What is the reason for keeping this list? Is it used only for
determining if a file is 'dirty', i.e. has been written to after open()?
The list is necessary because the sys_close() function does not provide any
information except a file descriptor. This list is for looking up files from
file descriptors. As long as ON_CLOSE events should report file names, this
list is necessary.
And with the new improvements in the FreeBSD extension of Dazuko, this list
becomes even more important for ON_OPEN. In FreeBSD the dup() and dup2()
functions are also hooked, which generate additional ON_OPEN events.
If that's the case, then how about abandoning the list altogether and
using brute force: send DAZUKO_CLOSE_MODIFIED event for every close() if
the file was opened with flags O_WRONLY|O_RDWR and do not send the event
if the file was opened O_RDONLY. I believe the Linux 'struct file'
carries the open flags in its f_flags field.
The dirty flag is a secondary use of the list to also mark when open files
have been written to. Since this happens with each sys_write(), it would be
wise for us to optimize this lookup scheme further (although "Swade" already
did some excellent work for optimizing ON_CLOSE_MODIFIED).
I guess there could be situations where a process forgets to call
close, either on purpose or otherwise...
This usually doesn't matter because the system will close the descriptor,
even if the program itself did not. I think leaks are occurring because
certain types of "files" don't require a close from the system and so are
never closed. /dev/null might be an example of this. It definately needs to
be investigated further.
(Lots of references to /var/run/utmp. I believe it is the 'top' utility
that is responsible for those.)
Yes, this is exactly the problem. The question is, why is dazuko not getting
a corresponding close for each of those opens? I have thought about if
dazuko should have an expire time on the list entries so that older items
can just be removed. Of course, understanding the cause of the problem would
make me happier.
John Ogness
--
Dazuko Maintainer