[PVFS2-users] errors accessing files

[Phil Carns]

Hi Elliot,

Thanks for the helpful log output!

There are actually two levels of bugs here.  The first is how the file 
got into that state in the first place.  The second is why pvfs2-ls and 
pvfs2-fsck blow up when given bad information from the server.  I don't 
know about the second issue (though I don't think it would be hard to 
track down given the test case below), but I suspect I know how the file 
got damaged.

You can see from the logs that the dfile count and distribution have 
been zeroed out for the file in question, while the rest of the 
attributes appear to be intact.  I think the problem is in some trove 
subsystem code that copies attributes when they are modified.  You can 
actually trigger this pretty easily, at least with the current pvfs2 
code in cvs:

apps/admin> ./pvfs2-cp /etc/hosts /mnt/pvfs2/testfile
apps/admin> ./pvfs2-chmod 700 /mnt/pvfs2/testfile
apps/admin> ./pvfs2-ls -alh /mnt/pvfs2/testfile
Segmentation fault

I think the fix is to remove these lines from the 
PVFS_object_attr_overwrite_setable macro in pvfs2-storage.h:

dest->u.meta.dfile_count = 0;                              \
dest->u.meta.dist_size = 0;                                \

Those fields are not supposed to be zeroed out in this code path.

I am not 100% that this is what happened to your file, but it is pretty 
likely.  I apologize for not posting this fix earlier, but I had it 
lumped in with a larger set of attribute changes that I have been 
playing with but don't have patches ready for yet.

As a side note, the way that the kernel driver happens to handle 
attributes in the current pvfs2 cvs tree makes it very difficult 
(impossible?) to trigger this bug through the kernel interface, or else 
someone probably would have spotted it much sooner.

-Phil

Elliot Peele wrote:
> I have a file in my pvfs2 file system that gives me Permission Denied
> errors if I'm accessing it as root or a normal user. From taking to the
> user that caused this error, he was reading from the file using netCDF
> (may have been in rw mode).
> 
> Our pvfs2 file system consists of three IO nodes, one meta data node,
> and 22 unix IO clients.
> 
> Trying to run any command on the file gives a permission denied error.
> 
> All of the pvfs2-* commands that I have tried now segfault when trying
> to access the file. This also includes pvfs2-fsck. Output and logs from
> pvfs2-ls can be found at:
> 
> 	http://rudi.ncren.net/~elliot/pvfs2log.txt
> 
> Any help would be appreciated.
> 
> Elliot
> 
> _______________________________________________
> PVFS2-users mailing list
> PVFS2-users at beowulf-underground.org
> http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users