SUMMARY: large phantom files
Beck, Joseph
jbeck at seic.com
Wed Apr 4 09:47:49 EDT 2007
Thanks to all for their quick replies...
Lsof and find /proc -links 0 - will show you files thar were deleted but
still have a process using them
---
Run "lsof +L1" and look for large entries in the SIZE column.
Restart/kill the corresponding process.
The +L1 option lists files with a link count of zero.
---
lsof is your best friend in these cases. lsof can be had from
www.sunfreeware.com as a package for many releases of Solaris.
Something like
lsof | grep /tmp | grep -v /tmp/ | grep VREG will list unnamed files
open in /tmp and the processes that have them open. Any or all of those
are possible candidates.
Cheers,
Ric Anderson (ric at opus1.com)
---
It's been a while since I've used this strategy, but I've had success in
the past by running:
fsck -n
on the file system in question. Any "phantom" file should show up as
disconnected inodes. If you see any *large* disconnected files, you can
use "pfiles" on all suspect processes to find which has the process
open.
Steve Ehrhardt
---
I may have spoken too quickly. I checked on this approach just after I
sent you my last email, and found that it would not work.
As I said, I haven't used the "fsck" approach recentl. Iit appears that
"fsck -n" chokes on the question:
/dev/dsk/c0t0d0s5 IS CURRENTLY MOUNTED READ/WRITE.
CONTINUE? no
You could still use this basic approach, but would either have to run it
manually - answering "no" to all the questions , or else try using ing
like "expect" to selectively answer "yes" to just this one question.
Sorry for not checking before answering.
Steve Ehrhardt
---
Did you use fuser(1M)?
---
Never fully investigated Disk I/O by process myself but the following
whitepaper may be of interest:
http://brendangregg.com/Perf/paper_diskubyp1.pdf
Regards,
Glenn
---
I have had some success in the past with running "fsck -N" against the
suspect filesystem. If fsck reports an unallocated inode I would then
use lsof to find a pid associated with that inode.
Regards,
Vic
---
you have to kill the process that created the files that are filling up
your disk; what's happening is that a process still has an inode open,
so the space never becomes available. eg: if /var/adm/messages fills up
your disk, you have to HUP syslogd to get the space back.
run lsof to track it down...
cheers
bil
-----Original Message-----
From: Beck, Joseph [mailto:jbeck at seic.com]
Sent: Thursday, March 29, 2007 5:30 PM
To: sunmanagers at sunmanagers.org
Subject: large phantom files
We've had 2 incidents this week where file systems were 100% or quite
close.
In both scenarios du was reporting little to no utilization. So, I
figured it was the old issue where I just needed to identify the
application that still had a file descriptor open on some really large
file and shut the app down to clear the space. Unfortunately, that
didn't prove to be the case.
The first incident was /tmp on an apache server. After much hunting
around, the conclusion was that the application (a weblogic agent) was
writing something to /tmp and someone deleted the file. I suppose the
possibilities are limitless. In the end, after stopping & starting
several of the 30+ apache instances, we ended up rebooting.
Fortunately it was a dev environment, but I still took enough heat.
2nd incident was a less critical orca server, /var was showing 98% full
but only had 25MB worth of files out there.
Orca server was sol10 but at end of day so we went with the easy fix &
rebooted.
I'm looking for other strategies, tools, whatever to be able to locate
the offending process in this scenario. On the web server, I exhausted a
lot of time with this looking at /proc & pfiles command & lsof, etc.
This was solaris 9
Also curious if dtrace could've helped in any way.
Joe Beck Ciber Inc. - a consultant to SEI One Freedom Valley Drive/ 100
Cider Mill Road| Oaks, PA 19456 | p: 610.676.2258 | jbeck at seic.com
_______________________________________________
sunmanagers mailing list
sunmanagers at sunmanagers.org
http://www.sunmanagers.org/mailman/listinfo/sunmanagers
More information about the summaries
mailing list