[Dovecot] Setting location for indexes
We have NFS-mounted user directories which cause significant problems with stale file handles:
dovecot: IMAP(uid): fstat() failed with file /home/uid/mail/.imap/INBOX/dovecot.index.log: Stale NFS file handle
(a) why can't it just close the file and re-open ?
(b) is there a way to specify with the 'default_mail_env' (or in some other way) the location of **only** the indexes (and let dovecot autodetect all else?).
I have already tried 'mmap_disable = yes', but that does not solve the problem, and if I set 'default_mail_env = INDEX=/var/spool/mail/INDEX/%u', I get errors like:
IMAP(uid): Failed to create storage with data: INDEX=/var/spool/mail/INDEX/uid
I don't want to specify mbox:~/mail/ or mbox:~/Mail/, because some users have one and some the other.
On Wed, Apr 27, 2005 at 09:47:25AM +0200, dovecot.7.signal11@spamgourmet.com wrote:
We have NFS-mounted user directories which cause significant problems with stale file handles:
dovecot: IMAP(uid): fstat() failed with file /home/uid/mail/.imap/INBOX/dovecot.index.log: Stale NFS file handle
(a) why can't it just close the file and re-open ?
NFS has issues with files that are open on one client, and get deleted on another. UNIX systems require that files don't disappear while they are in use, but NFS is stateless and it simply can't know when a file is in use. The result is the 'stale NFS handle' kludge, which also prevents a simple close-and-open. Don't use index files on NFS disks.
(b) is there a way to specify with the 'default_mail_env' (or in some other way) the location of **only** the indexes (and let dovecot autodetect all else?).
Haven't figured that one out myself either :)
-- Thomas Wouters thomas@xs4all.net
Hi! I'm a .signature virus! copy me into your .signature file to help me spread!
Thomas Wouters wrote:
(b) is there a way to specify with the 'default_mail_env' (or in some other way) the location of **only** the indexes (and let dovecot autodetect all else?).
Haven't figured that one out myself either :)
can't you just append :INDEX=/some/path to your default_mail_env ? just like you can have :INBOX=/where/ever So, if you had Maildirs in /home/nfs/%u/ and indexes in /home/%u , with mail delivered to /var/mail/%u you could use:
default_mail_env = maildir:/home/nfs/%u:INBOX=/var/mail/%u:INDEX=/home/%u
or is it INDEXES? It's been a while since I've looked at any of this.. .Dovecot has very much been a "set and forget" for me... the way I like it :)
-- Curtis
dovecot: IMAP(uid): fstat() failed with file /home/uid/mail/.imap/INBOX/dovecot.index.log: Stale NFS file handle
(a) why can't it just close the file and re-open ?
NFS has issues with files that are open on one client, and get deleted on another. UNIX systems require that files don't disappear while they are in use, but NFS is stateless and it simply can't know when a file is in use. The result is the 'stale NFS handle' kludge, which also prevents a simple close-and-open. Don't use index files on NFS disks.
First, the file did not disappear - the NFS is flaky under high load, that's all. Closing and re-opening the file would probably fix the issue, at least after the traffic spike.
Second, these are files that are intended for dovecots internal use only. If the NFS filehandle goes stale, it's anyway highly unlikely that someone else has deleted them, and dovecot should know that - IMHO it should at least try to recover.
Third, I'm perfectly willing to put the index files on a local disk, but I need dovecot to autodetect the rest of the mail environment, unless I want to define it completely for each user individually, and apparently dovecot does not let me do that(?).
From the (few) answers I got so far, it seems that it is in fact impossible to specify 'default_mail_env = INDEX=/var/spool/mail/INDEX/%u', so would anybody know where in the sources the root directory for the indexes is set, such that I can patch the code?
Have you tried 'default_mail_env = :INDEX=/var/spool/mail/INDEX/%u'? I've not tried it myself, mind you :)
Looking through the code in 1.0-stable, in src/lib-storage/index/mbox/mbox-storage.c and src/lib-storage/index/maildir/maildir-storage.c it looks for the first ":" then takes root_dir as the leftmost part and looks for INDEX= and INBOX= in the rightmost part. I'm not sure what happens then when root_dir is empty (mbox_create and maildir_create return NULL).
You didn't say which version you're using? I think 1.0-test might be different again (Timo mentioned re-organising the whole mailbox "driver" thing).
Best Wishes, Chris
dovecot.7.signal11@spamgourmet.com wrote:
dovecot: IMAP(uid): fstat() failed with file /home/uid/mail/.imap/INBOX/dovecot.index.log: Stale NFS file handle
(a) why can't it just close the file and re-open ?
NFS has issues with files that are open on one client, and get deleted on another. UNIX systems require that files don't disappear while they are in use, but NFS is stateless and it simply can't know when a file is in use. The result is the 'stale NFS handle' kludge, which also prevents a simple close-and-open. Don't use index files on NFS disks.
First, the file did not disappear - the NFS is flaky under high load, that's all. Closing and re-opening the file would probably fix the issue, at least after the traffic spike.
Second, these are files that are intended for dovecots internal use only. If the NFS filehandle goes stale, it's anyway highly unlikely that someone else has deleted them, and dovecot should know that - IMHO it should at least try to recover.
Third, I'm perfectly willing to put the index files on a local disk, but I need dovecot to autodetect the rest of the mail environment, unless I want to define it completely for each user individually, and apparently dovecot does not let me do that(?).
From the (few) answers I got so far, it seems that it is in fact impossible to specify 'default_mail_env = INDEX=/var/spool/mail/INDEX/%u', so would anybody know where in the sources the root directory for the indexes is set, such that I can patch the code?
-- --+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+- Christopher Wakelin, c.d.wakelin@reading.ac.uk IT Services Centre, The University of Reading, Tel: +44 (0)118 378 8439 Whiteknights, Reading, RG6 2AF, UK Fax: +44 (0)118 975 3094
On Wed, 2005-04-27 at 10:56 +0200, Thomas Wouters wrote:
On Wed, Apr 27, 2005 at 09:47:25AM +0200, dovecot.7.signal11@spamgourmet.com wrote:
We have NFS-mounted user directories which cause significant problems with stale file handles:
dovecot: IMAP(uid): fstat() failed with file /home/uid/mail/.imap/INBOX/dovecot.index.log: Stale NFS file handle
(a) why can't it just close the file and re-open ?
NFS has issues with files that are open on one client, and get deleted on another. UNIX systems require that files don't disappear while they are in use, but NFS is stateless and it simply can't know when a file is in use. The result is the 'stale NFS handle' kludge, which also prevents a simple close-and-open. Don't use index files on NFS disks.
This is actually a TODO item. Dovecot should be able to handle all ESTALE errors, but I just haven't implemented them. With dovecot.index.log handling code the ESTALE is an expected error with NFS because Dovecot rotates the log file continuously overwriting the existing one.
Actually the new code in CVS nowadays rotates .log file to .log.2 file, which should make this problem much more rare.
"First, the file did not disappear - the NFS is flaky under high load, that's all. Closing and re-opening the file would probably fix the issue, at least after the traffic spike."
Can this really happen with high NFS load, or is it just that Dovecot had more time to rotate its log files? I'm not sure if I can reliably fix the ESTALE problems if they happen randomly with the file actually being lost.
(b) is there a way to specify with the 'default_mail_env' (or in some other way) the location of **only** the indexes (and let dovecot autodetect all else?).
Haven't figured that one out myself either :)
Guess not. New configuration file code has split them to different settings though, that'll fix it. Now if I only could get all that code to work so I could commit it.. :)
dovecot.7.signal11@spamgourmet.com wrote:
We have NFS-mounted user directories which cause significant problems with stale file handles:
dovecot: IMAP(uid): fstat() failed with file /home/uid/mail/.imap/INBOX/dovecot.index.log: Stale NFS file handle
(a) why can't it just close the file and re-open ?
The Stale NFS file handle error is a fatal error on the system side.
Closing and re-opening isn't a guaranteed solution.
Actually, the real question, in my opinion, would be what is causing the system to get a stale NFS file handle? IF you are going to store and access mail from NFS filesystems, the filesystem needs to be extremely stable and accessible..
-Rich
participants (6)
-
Chris Wakelin
-
Curtis Maloney
-
dovecot.7.signal11＠spamgourmet.com
-
Rich West
-
Thomas Wouters
-
Timo Sirainen