samba - Oct 2002 - SMBFS files receiving incorrect timestamps

Hello all.

Our system consists of two linux machines, each running Red Hat 7.1
(kernel 2.4.9-34), using SMB to mount multiple shares hosted by a
Windows 2000 Advance Server.  smbclient from Samba 2.2.5 is used to do
the actual mounting.

Over the weekend, a number of files on these SMBFS shares were created
with incorrect timestamps (modification times).  In some cases, the
timestamps were off by as much as four days!  Today, every file created
by the linux machines, even those created by "touch" called without
any
extra arguments, received timestamps that were one hour ahead.

When I discovered this was happening, I unmounted all SMBFS shares, and
remounted them.  This fixed the problem; all files created now have the
correct timestamp.

The Windows host and the linux clients all have their system clocks set
correctly, and their system clocks were updated automatically during the
switch from daylight savings time.  I suspect the 1-hour discrepancy I
saw today has something to do with the daylight savings switch, but if
that is the case, what caused some files created on Saturday (26
October) to receive timestamps for this coming Wednesday (30 October)?

Most importantly, how do I keep this from happening again?  Please let
me know if more information is needed.  Thank you.

---Kris Kelley

On Mon, 28 Oct 2002, Kris Kelley wrote:
> When I discovered this was happening, I unmounted all SMBFS shares, and
> remounted them.  This fixed the problem; all files created now have the
> correct timestamp.
smbfs sets the mtimes itself. Not really sure why it was done like that,
but some comments claim that NT4 doesn't set mtime properly otherwise. It
would be better to just let the server handle time.

The 1-hour diff is probably the daylight savings change. smbfs is told by
smbmount what the server timezone is, as minutes from GMT. Of course this
changes with the daylight savings ...

> switch from daylight savings time.  I suspect the 1-hour discrepancy I
> saw today has something to do with the daylight savings switch, but if
> that is the case, what caused some files created on Saturday (26
> October) to receive timestamps for this coming Wednesday (30 October)?
No idea. Never heard anything like it before and would guess that the
client time actually was wrong (at that particular time).

/Urban

Hi,

My tar (GNU tar) 1.13.25 on a RedHat 7.3 Linux system has been giving me non
fatal, but discouraging messages for some time now as I try to archive files
across smb mounts to a Windows 2000 Server system. it doesn't happen with
all files, fortunatly, but it does happen with a few files per archive.

tar: <filename>: file changed as we read it

The strange thing was when I would look at the files the mtime (on either
system) was several months old. Today I dug deeper because I am increasing
the number of files that I am backing up across the smb mounts.

What I found was that without fail, tar would get one mtime value when it
started reading the file, but that time must have drifted while reading the
file, hence the message.

mtime tar puts into the archive: (no matter how many times I run tar)
2002-01-30 13:41:57
This is the same value reported by the file properties on the Windows 2k
server.

mtime ls -l --full-time reports
Wed Jan 30 13:41:58 2002

the first time I ran stat (after tarring the file a few times) it reported:
Wed Jan 30 13:41:57 2002
I ran it again, immediatly after that output, and it reported:
Wed Jan 30 13:41:58 2002

stat continues to report this value no matter even after running tar again,
which continues to put the (correct) Wed Jan 30 13:41:57 2002 value into the
archive

When I copied (cp -a) the file from the smb mount to the local ext3 fs, the
mtime was Wed Jan 30 13:41:58 2002 (incorrect)

Now, one second isn't any big deal except that tar sees it as a change and
then I have to track down and see if tar realy got the right version of the
file(s) (maybe the file wasn't months old). I have tried remounting, and
that didn't help.

Other program versions:
kernel 2.4.18-3
samba 2.2.7-1.7.3
stat: 2.5-5
ls (GNU fileutils) 4.1
mount 2.11n-12.7.3

I found this thread in the archives in October 2002, and it seems slightly
related. I gather from the reply that smbfs is actualy managing the mtime
values instead of passing on what it reads from the remote file system? If
so, is there one library call (the one tar uses for the archive headers)
that returns the right mtime, and another (the one tar uses to see the file
has changed and that ls -l uses) that returns the wrong mtime? More likely
by some wierd twist tar always gets smbfs to report the right time at first
but then as the file is accessed (read by tar, ls) smbfs is then reporting
the wrong mtime?

Any suggested work-arounds in configuration settings are appreciated. Maybe
when someone is going over the timestamp code, they might use this
information to fix the glitch (if it is in the smbfs/smb mount)

Jacob Anawalt

On Mon, 28 Oct 2002, Urban Widmark wrote:

smbfs sets the mtimes itself. Not really sure why it was done like that,
but some comments claim that NT4 doesn't set mtime properly otherwise. It
would be better to just let the server handle time.

The 1-hour diff is probably the daylight savings change. smbfs is told by
smbmount what the server timezone is, as minutes from GMT. Of course this
changes with the daylight savings ...

On Mon, 28 Oct 2002, Kris Kelley wrote:

Hello all.

Our system consists of two linux machines, each running Red Hat 7.1
(kernel 2.4.9-34), using SMB to mount multiple shares hosted by a
Windows 2000 Advance Server.  smbclient from Samba 2.2.5 is used to do
the actual mounting.

Over the weekend, a number of files on these SMBFS shares were created
with incorrect timestamps (modification times).  In some cases, the
timestamps were off by as much as four days!  Today, every file created
by the linux machines, even those created by "touch" called without
any
extra arguments, received timestamps that were one hour ahead.

When I discovered this was happening, I unmounted all SMBFS shares, and
remounted them.  This fixed the problem; all files created now have the
correct timestamp.

The Windows host and the linux clients all have their system clocks set
correctly, and their system clocks were updated automatically during the
switch from daylight savings time.  I suspect the 1-hour discrepancy I
saw today has something to do with the daylight savings switch, but if
that is the case, what caused some files created on Saturday (26
October) to receive timestamps for this coming Wednesday (30 October)?

Most importantly, how do I keep this from happening again?  Please let
me know if more information is needed.  Thank you.

---Kris Kelley

On Thu, 5 Dec 2002, Jacob Anawalt wrote:
> tar: <filename>: file changed as we read it
...
> mtime tar puts into the archive: (no matter how many times I run tar)
> 2002-01-30 13:41:57
> This is the same value reported by the file properties on the Windows 2k
> server.
I posted something regarding this just now, but here is a repeat:

Find fs/smbfs/proc.c in your kernel source tree.
Search for smb_proc_close_inode().
See if changing any of the operations it does there helps.
  - remove the code after the "Kludge alert"
  - remove smb_proc_setattr_ext()
  - replace ino->i_mtime in the smb_proc_close() call with a 0


I have been a bit busy recently but I will eventually get around to
looking more closely at this.

/Urban