Description of problem: Reading data from files over Samba share causes them to temporarily appear as modified when same files are examined via NFS share. Does not happen if cached 'ls' data is present. Results in spurious rebuilding of output files when 'make' is run concurrently from Samba and NFS shares. Note that in our environment objects for different platforms are kept in separate directories in a manner similar to the way 'gcc' builds work. The fictious source file timestamp modifications are what cause the problem. Version-Release number of selected component (if applicable): Samba+NFS server, x86_64 samba-3.0.10-1.4E.12.2 kernel-smp-2.6.9-55.EL Samba client Windows 2003 R2 SP2 X64, current patches NFS client, i686 kernel-2.6.9-55.EL How reproducible: 1) place a bunch of files in a directory (verified this with 'gzip' source) 2) issue 'touch time_mark' 3) create Samba share of directory 4) create NFSv3 share of directory 5) issue 'find . -type f -print | xargs cat >/dev/null' in directory via Samba client (alternately 'make -n' could be run) 6) issue 'find . -type f -newer time_mark -ls' in directory via NFS client 7) observe that all files are reported as modified!!! 8) repeat (6) after 'acregmax' interval and observe that files are no longer reported as modified Actual results: In practice this has been causing 'make' to rebuild things that it should not. Putting it nicely, it's been driving us nuts for many months. Expected results: Obviously, the files were never modified and should not be reported as such. Additional info: If you run 'ls -l' in the directory via the NFS share, the problem is cured. You may have to unmount and remount the share to bring back the bug. Before we figured this out, tried NFSv4 for awhile. Seemed like it fixed the above issue but it's not certain. NFSv4 + PoPToP started hosing our system so it's gone and I have no further patience or time for more experiments.
A closer looks shows that the 'cat' command is not require. Simply 'find'ing the files on the Windows machine with find . -type f -print >/dev/null produces the behavior. Also determined the problem appears for both 'i686' and 'x86_64' servers, so it's a portable bug. Also it seems that more than a handful of files is required to produce the problem.
Further details: does not happen at top-level of NFSv3 mount, must be in subdirectory if any directories are touched by the 'find' on Samba share, problem is suppressed does not happen when files reside on 'ext2' file system does happen when files reside on 'ext3' file system have 'noreservation' option active too lazy to try it with 'noreservation' turned off tried mounting Samba share remotely with CIFS problem does not happen with 'find' run on CIFS mount files accessed via Windows UNC path, not mapped drive obviously CYGWIN is in use on the Windows system can also produce the problem using CMD shell and 'type \\server\...\file' command CMD 'dir' doesn't work, probably because directory file is examined does happen with even tiny number of files
does happen with WXP SP2 32-bit Samba client can produce problem with MKS, but a 'cat' is also required: find . -type f -print | xargs cat >/dev/null
Created attachment 267991 [details] TAR archive of traces files showing problem Traces showing problem. Debug level 10 Samba trace plus Wireshark capture. # Windows 'type' access of file 04:00:41 SMB access of 'djl0' at packet 21 Last Write: Nov 24, 2007 02:20:44.000000000 # first 'find' invocation 04:00:45 NFS access of 'djl0' at packet 97 mtime: Nov 24, 2007 04:00:45.000000000 04:00:45 NFS access of 'djl0' at packet 107 mtime: Nov 24, 2007 04:00:45.000000000 # second 'find' invocation 04:00:54 NFS access of 'djl0' at packet 176 mtime: Nov 24, 2007 02:20:44.000000000
Created attachment 268101 [details] file lease test program Examined Samba traces and discerned that the cause of the problem: Samba issues fcntl(fd, F_SETLEASE, F_WRLCK); on files that are opened for read access. The Linux kernel sets the modify time as seen via NFS as current when the activity trace is put in place and then restores the modify time once it's released. Reproduced this with the attached test case. The altered modify time appears *only* via NFS. On the local system it's not visible. Definitely a NFS bug.
Checked NFSv4 and issue is not present there. Appears only with NVSv3 and is in the server logic, not in the client side.
The Plot Thickens ================= It turns out the observed behavior is by design; full details at http://bugzilla.kernel.org/show_bug.cgi?id=9454 I don't fully understand all the interactions, so I'm pasting comments from Bruce Fields, a 'nfsd' maintainer. Appended the same information to the Samba bug report. >------- Comment #5 From bfields 2007-11-26 08:57:56 [reply] ------- > >This is by design--see fs/locks.c:lease_get_mtime(). > >The argument is: if somebody has a write lease on the file (are >you exporting this via Samba? That'd be the typical user), then >they're caching writes--they're explicitly telling us that we do >not know whether the file is still the same, because we may have >modified it on a remote client and not told us about it. So >lease_get_mtime() reports the current time as the mtime, >prompting you to actually try opening the file, at which point >the write lease gets broken and any cached writes get flushed >out. > >It's a terrible kludge, I agree, and maybe we should remove it. >But I'd first like to understand what circumstances prompted >somebody to add it originally, and talk to the Samba people >about how they're using these write leases. > >(In NFSv4, by the way, there's a callback to the client to allow >the server to find out the attributes of a file that the client >holding a write lease is caching, which solves this problem. We >haven't implemented that yet; it seems likely to be an enormous >pain. I wonder if Samba would want something similar?) > > >------- Comment #6 From starlight 2007-11-26 10:17:08 [reply] ------- > >Yes, using a Samba mount along with NFS mounts to 'make' build >an application on multiple platforms simultaneously. When >re-building existing trees the result is spurious target >building, a real problem. Took months of frustration to >figure out what was going on. > >Actually reported this as a bug to Samba along with RH, and a >request for info from the Samba team provoked the analysis that >isolated the problem. Closed the Samba bug but it's all still >there: > >https://bugzilla.samba.org/show_bug.cgi?id=5103 > >My impression from looking at the Samba code is that the leases >are put in place so that files can be monitored for changes, >but that's just a guess. Perhaps Samba can use a read lease >instead of a write lease. > > >------- Comment #7 From bfields 2007-11-26 10:55:17 [reply] ------- > >Thanks for the pointer to the samba bug. I understand the >frustration. > >My feeling is that: yes, it's suboptimal (but perhaps not a bug) >for samba to be requesting a write lease when a read lease >(sufficient to alert it to modifications of the file) would be >enough. > >It seems to me that the real bug is the incomplete lease >implementation--if the purpose of a lease is to allow a remote >client to cache writes to the file, then there's no way for us >to give a sensible answer to a stat call, unless we break the >lease first. > >Perhaps we could find out (in the samba case) what the >consequences would be of not updating mtime in this case? I >suppose the worst case would be that a modification to a file >made on a samba client could be indefinitely delayed from being >flushed to the server. > >If we agree that that would be less of a problem than these >spurious bumps in the mtime, then the best solution for now may >just be to rip out lease_get_mtime(). I'll cook up a straw-man >patch.... > > >------- Comment #8 From bfields 2007-11-26 10:56:16 [reply] ------- > >Created an attachment (id=13763) [details] >remove lease_get_mtime > >Not sure if this is what we want to do, but it's at least >something you could test to confirm the source of the problem. > > >------- Comment #9 From starlight 2007-11-26 >11:13:01 [reply] ------- > >Thanks! Happen to have a kernel source tree set up for the >RH images running on the server, so I can try this out in >the next week or so. Seems certain to work. > >In the interim perhaps you could start a dialog with Samba about >the reasoning behind the application of the F_WRLCK? Or if you >wish I could do so using the Samaba bug report cited above. >
Created attachment 270671 [details] kernel-2.6.9-55_remove_lease_get_mtime.patch Adapted kernel.org authored patch for RHEL 4.5 2.6.9-55 kernel (attached). Works as expected and eliminates spurious target rebuilding when a concurrent Samaba and NFS 'make' is run on same source tree. Test case also shows expected behavior. Obviously did not perform any regression testing with other Samba file scenarios. In our build trees the same files and directories are never written by more than one server.
Concise description of patch rationale from original: >From: J. Bruce Fields <> >Date: Mon, 26 Nov 2007 13:48:34 -0500 >Subject: [PATCH] nfsd: stop incrementing mtime on presence of write lease > >The lease_get_mtime() function has the effect of setting the mtime of a >file (as far as any nfs client is concerned) to the current time, >whenever there is a write lease held on the file. > >The presence of a write lease may mean that some client (almost >certainly a samba client) is caching writes to that file. Thus >increasing the mtime has the effect of making the nfs client read data >from that file, thus opening the file for read, thus breaking the write >lease, thus forcing the samba client to flush any cached writes. > >However Samba seems to be requesting write leases even when a read lease >would do. And unfortunately the consequences of the spurious mtime >updates--make unnecessarily rebuilding, for example--may be worse than >the consequences of no mtime update--modifications from a samba client >taking longer to be noticed on the nfs client. > >So perhaps we should rip out lease_get_mtime().
Hmmm...at first glance this sounds tricky and I'm not clear whether it's something we'd want to take for RHEL4. I need to do a more thorough review of the problem and patch though, so let me do that before I NAK anything...
This definitely needs to go upstream first and will need some soak time before it can go into any RHEL product. Given that RHEL4 is approaching maintenance phase, it's probably unlikely that this will ever make it into there. RHEL5 might be a possibility if there's resolution upstream. Most likely, samba is getting leases in order to emulate opportunistic locks (aka oplocks) on the share. One possible workaround might be to disable kernel oplocks in smb.conf: kernel oplocks = no ...or maybe by judicious use of the "veto oplock files" parameter. Of course, that means that oplocks/leases won't be broken by actvity on the files outside of samba, but that's probably similar to the behavior you'll see by removing lease_get_mtime anyway. An even better option might be to remove NFS from the picture altogether and just use CIFS with unix extensions on the Linux clients. In any case, I don't see this happening for RHEL4 so I'm going to NAK this for now. Feel free to open a similar case for RHEL5 if you like, though until this patch goes upstream it's not an option there either...
Development Management has reviewed and declined this request. You may appeal this decision by reopening this request.
Tried "kernel oplocks = no" in the Samba config and it does work around the issue.