Details
-
Improvement
-
Resolution: Fixed
-
Minor
-
None
-
9223372036854775807
Description
It would be useful to store the JobID that created a file in an xattr on the MDT inode. This would allow tracking the source of files after the fact, and can help users with the provenance of their file data (e.g. which run the files came from, which can then be used to determine runtime parameters used, whether the run was successful or not, etc.).
It isn't clear whether this would be needed on the OST inodes as well, but I think initially the MDT inode would be enough, as this would be the only xattr visible to users. The OST xattr would only be accessible for low-level scanning tools.
Attachments
Issue Links
- duplicates
-
-
- Resolved
-
- has to be done before
-
-
- Closed
-
- is related to
-
LU-18798 user.job xattr is storing quotes around job name
-
- Resolved
-
-
-
- Open
-
-
-
- Resolved
-
- is related to
-
-
- Open
-
-
-
- Open
-
-
-
- Open
-
- mentioned in
-
Page Loading...
Activity
"Oleg Drokin <green@whamcloud.com>" merged in patch https://review.whamcloud.com/c/fs/lustre-release/+/50982/
Subject: LU-13031 jobstats: store jobid in xattr when files are created
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 23a2db28dcf1422a6a6da575e907fd257106d402
For the implementation, any such restrictions/checking would be done in the sysfs write handler when the xattr is first set, and for actual xattr creation this would just copy the job_xattr string during actual file creation with no need for additional sanity checking, since it should always be consistent.
For the "don't want user.job to be cp.12345" issue, I think that should be handled properly. Initially, the xattr would be set to cp.PID or SLURM_JOBID or whatever by the MDS internally at file creation time. However, if cp, rsync, tar, dcp, dsync ... is also copying xattrs, the initial value would be overwritten later in the copy process, which is what we would want I think.
I'm inclined toward the user.job xattr, but it is never a good idea to unilaterally make policy decisions in the kernel that cannot be changed.
As such, it might make sense to have a tunable parameter like "mdt.*.job_xattr=user.job" (or mdd.*.job_xattr is possibly better, though e.g. "mdt.*.job_cleanup_interval" already exists), and then this could be changed in the future if there is some conflict (e.g. some site already uses the "user.job" xattr for some other purpose).
I don't think the job_xattr should allow totally arbitrary values (e.g. overwriting trusted.lov or trusted.lma or security.* would be bad). One option is to only allow a limited selection of valid xattr namespaces, and possibly names:
- NONE to turn this feature off
- user, or trusted or system (if admin wants to restrict the ability of regular users to change this value?), with implicit ".job" added
- user.* (or trusted.* or system.*) to also allow specifying the xattr name
If we allow the xattr name portion to be specified (which I'm not sure about, but putting it out for completeness), it should have some reasonable limits:
- <= 7 characters long to avoid wasting valuable xattr space in the inode
- only ASCII alpha-numeric characters, plus "." for the name separator, to avoid spaces, control characters, symbols (=, ,, etc.) that might cause problems
- should not conflict with other known xattrs, which is tricky if we allow the name to be arbitrary. Possibly if in trusted (and system?) it should only allow trusted.job to avoid future conflicts?
- maybe restrict it to contain "job" (or maybe "pbs", "slurm", ...) to reduce the chance of namespace clashes in user or system? However, I'm reluctant to restrict names in user since this shouldn't have any fatal side effects (e.g. data corruption like in trusted or system), and the admin is supposed to know what they are doing...
Thoughts?
Thank you for that; I'm persuaded we should do the user xattr, the issues around interop and user tools look decisive. If we're forced to change tack in the future, we can deal with that then. (I think I was the only one actively objecting, Andreas indicated OK at LUG, so you should be good to go.)
Requirements (or at least, Wants):
- Can preserve xattr with tools like `cp` and `rsync` (it's not useful if you copy your data and the jobid gets overwritten with stuff like "cp.1234")
- Can preserve xattr when moving files to other filesystems, e.g., to an archive, since the jobid is likely useful to keep around long term
User namespace
Pros:
- Enables things like `cp --preserve=xattr` and `rsync --xattrs`
- Allows interoperability with other filesystems (xattr can be preserved if a file is copied to an FS that doesn't know about the "lustre" namespace)
Cons:
- violates expectation that the kernel will not write user data, which could be problematic if this becomes enforced at some point
Lustre namespace:
Pros:
- Allows flexibility in terms of being written from both user and kernel space; we can decide our own rules
Cons:
- Lustre xattr is "virtual" and must be backed by an xattr in a traditional namespace (example, lustre.lov and trusted.lov) since the kernel (AFAICT) will reject storing a lustre.* xattr directly, which increases complexity.
- Prevents interoperability-- standard tools like `rsync` won't be able to preserve this xattr because tools that copy files to other filesystems will have to be specially coded to translate lustre.job into user.job
Trusted namespace: I think this would only be considered as backing the lustre namespace and wouldn't be used on its own.
System namespace:
Pros:
-Arguably the most correct choice. According to man 7 xattr: "Extended system attributes are used by the kernel to store system objects... Read and write access permissions to system attributes depend on the policy ... implemented by filesystems in the kernel." So this fits our need of being able to write it from both user and kernel processes.
Cons:
-Is not interoperable with other filesystems because other filesystems likely will not allow user processes to write a "system.job" xattr.
My opinion is that the user namespace is the best choice because every other choice fails to meet the goals above, but I will admit that not being very experienced with the kernel side of things, I can't say how severe the problem with the user namespace is.
Thanks, Thomas.
Could you lay out the implications of using user vs lustre namespaces as you see them? What do we gain, what we do lose, and/or what requirements can't we meet in certain settings?
@Patrick Farrell
Here's a link to my question on the mailing list archive http://lists.lustre.org/pipermail/lustre-discuss-lustre.org/2023-May/018629.html If you want to respond with your comments either in the mailing list or in the ticket here to continue the discussion.
"Thomas Bertschinger <bertschinger@lanl.gov>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/50982
Subject: LU-13031 jobstats: store jobid in xattr when files are created
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 26c63b60e0dcd145a8d032a9f4780c921435e184
Does interop refer to some testing that mixes different client and server versions? Is the failure occurring when the server version is older than the client and the MDS doesn't have this new change?
If it just needs a version check like:
(( $MDS1_VERSION >= $(version_code 2.15.??) )) || skip ...
that should be simple enough, I can submit a fix for that.
The version I get when I check out this commit and run `lctl get_param version` is 2.15.57. Is that the correct code to put in the version check? Or do I need one higher than that (since I guess you could check out a commit newer than the 2.15.57 tag but older than this change and its version would pass...). If I need one higher than 2.15.57, do I use 2.15.58, 2.16.0, or something else?