brettkettering
commented
12 years ago Due to more messing around today, we have discovered that modifying the permissions in any way on a N to N mount point through the fuse mount does not propagate to the backends.
This is a security issue now in that users will be unable to deterministically modify the permissions of their PLFS files.
agtorre
In creating the lustre PLFS mount points on LANL's smog machine, we found that chown doesn't properly chown backend directories when the mount point is N to N. Here is the order of what we did:
None of the /lustre/lscratch/.plfs_store_nn-*/dshrader directories where owned by dshrader; they were all owned by root.
The admin helping me with the install then created his own directory inside /plfs/lscratch_nn/ and chowned it to himself. All but one of the backend directories were owned by root; the one exception was owned by him.
The admin then created /plfs/scratch3_nn/test (/plfs/scratch3_nn is a N to N PLFS mount point connected to PLFS backends on a panasas filesystem) and chowned it to himself; all but one of the test directories on the backends were owned by root.
The admin then created /plfs/scratch3_n1/test (/plfs/scratch3_n1 is a N to 1 PLFS mount point connected to PLFS backends on a panasas filesystem) and chowned it to himself. All of the directories on the backends were owned by him.
The admin then did the same experiment on cielito by creating a test directory in a N to N mount point with backends on panasas and chowning it to himself. All but one of the directories on the backend where owned by root. So, we have consistent behaviour across machines and backed filesystem types.