Ah! After a bit of test programming I understand what's going on now.
If I'm right then there are a few incorrect assumptions in Dovecot's
logic. (Sorry! :-)
On 17 Jul 2007, at 18:12, Timo Sirainen wrote:
auto_direct /mailstore/messages/p autofs
direct,ignore,dev=4740014 1184648400 crypt2.york.ac.uk:/vol/vol9/p /mailstore/messages/p nfs
proto=tcp,xattr,dev=4700386 1184668792Although there are two entries they have different device
numbers. The mountpoint_get() function attempts to guard against
this by checking the device number obtained matches that of the
path being located.What do you mean different device numbers? Both have the same mount
path, so how can they be different?
Looking again at the above two example lines from our /etc/mnttab I
see that the first entry includes "ignore" as an option. As I
recollected that Dovecot's code in mountpoint_get() checks for
filesystems marked "ignore" and should, umm, ignore them(!) when
searching for the correct entry in the mount tab ... yet it wasn't
working.
So I've just written a minimal program to display all the entries of /
etc/mnttab and their fields using getmntent() and now think Dovecot's
checks are incorrect, at least on Solaris...
PROBLEM 1 :: Skipping swap and ignored filestores is incorrect
The mountpoint_get() function in lib/mountpoint.c uses getmntent() to
fetch each entry from /etc/mnttab in turn. It then checks the
ent.mnt_fstype field to see if it is MNTTYPE_SWAP ("swap") or
MNTTYPE_IGNORE ("ignore") and, if it is, the entry is skipped.
However I this appears to be incorrect for Solaris 10 (and probably
earlier versions too). For example:
- swap does NOT return "swap" as the value of ent.mnt_fstype
- ignored filestores do NOT return "ignore" as the value of
ent.mnt_fstype
Instead you get:
swap returns "swap" has the value of ent.mnt_special and "tmpfs"
for ent.mnt_fstypeignored filesystems return "ignore" as one of the options within
the ent.mnt_mntopts field. (NB: this might be one of several comma- separated options and could appear anywhere in the list.)
So I now believe that:
My fix of also skipping filestores whose ent.mnt_fstype is
"autofs" is not fully correct -- the problem should really be dealt with by
properly checking for the "ignore" option instead;The "ignore" option should be being checked within
ent.mnt_mntopts instead of ent.mnt_fstype; this would deal with the problem of the
template autofs entry being picked up and used as it has the "ignore" option.Swap should be detected by checking either ent.mnt_special being
"swap" or ent.mnt_fstype being "tmpfs" -- I'm not sure which is correct,
or if both tests are needed.
PROBLEM 2 :: Matching an entry in /etc/mnttab using getmntent() is
incorrect
As for the oddity with device numbers I mentioned previously, I now
understand this and think it shows up another incorrect assumption...
When using the automounter you get a permanent entry in /etc/mnttab
that I will refer to as a 'template' entry. In its options field it
includes a device number of the form "dev=XXXXX" where XXXXX is a
hexadecimal number (at least, it is hex on Solaris 10!).
When the automounter REALLY mounts the filestore as it is needed it
ADDS a second entry for the mounted filesystem. This has its own
(different) "dev=XXXXX" number. Don't ask me why this is: presumably
there are Good Reasons.
When locating the entry for a path in /etc/mnttab Dovecot tries to do
this by comparing the device number for the entry with the device
number of the path being located. It does this by using stat() on
the mount point given in the entry.
Whilst this works for manually-listed mounts it does not work when
the auto-mounter is involved. This is how Dovecot is being fooled...
We start off with just the automounter's 'template' entry in /etc/ mnttab:
auto_direct /mailstore/messages/p autofs
direct,ignore,dev=4740014 1184648400Dovecot starts by performing a stat on the path to be located. This causes the automounter to secretly mount the real filestore --
adding the second entry to /etc/mnttab with its different device number:crypt2.york.ac.uk:/vol/vol9/p /mailstore/messages/p nfs
proto=tcp,xattr,dev=4700386 1184668792The number stat() returns is THIS device number -- 4700386 --
and NOT that from the 'template' (4740014). This gets stored away in the
"st" structure ready for later comparisons.Dovecot now uses getmntent() to fetch each entry from /etc/mnttab
in turn. It calls stat() on the mount point in the entry to gets its
device number into the "st2" structure.In the case of using the automounter the first entry reached is the 'template' entry, with mount point /mailstore/messages/p. Using
stat() on this returns its device number ... and the number you get
back is that of the real mounted filesystem.Dovecot then compares the deivce numbers in the "st" and "st2"
structures, finds they match and so assumes it has found the matching entry in /etc/mnttabUnfortunately it hasn't: it has only got the template, which
doesn't have the remote host and fspath of the server from which the
filestore is mounted. This entry even has a different device number
(4740014) which would have shown up the problem ... but unfortunately Dovecot isn't checking the device number from the entry itself.
I believe the correct solution is to change step (3) so that it does
NOT stat() the mount point of each entry: you can't rely on getting
the right device number for that entry when the automounter is used.
Instead the proper solution is to parse the mount options in
ent.mnt_mntopts and extract the device number from the "dev=" option,
then compare THAT with the device number of the path being located
(from step (1)).
Of course this is fraught with problems because the mount options is
a comma-separated text string and the "dev=" option could appear
anywhere within it. Also, the value of "dev=" is in hexadecimal (at
least it is on Solaris 10), making comparison with the decimal
integer from the start() in step (1) non-trivial.
Given this complexity it may be sufficient to instead properly check
and skip filestores marked "ignore". The automounter's 'template'
entries are so marked; thus skipping them would allow Dovecot to pass
over them in the list and locate the proper entry for the real,
mounted, filestore instead.
Unfortunately this check still involves parsing the ent.mnt_mntopts
string, as the "ignore" option could appear as any one of:
ignore
ignore,
,ignore
,ignore,So some clever coding is still needed or a regular expression such as
^(.*,)?ignore(,.*)?$
Alternatively there's the dreaded strtok() function -- dreaded
because it actually changes the source string by over-writing
separator characters with '\0' as it tokenises it. Lovely! :-(
Cheers, Mike B-)
-- The Computing Service, University of York, Heslington, York Yo10 5DD, UK Tel:+44-1904-433811 FAX:+44-1904-433740
- Unsolicited commercial e-mail is NOT welcome at this e-mail address. *