Bug #72355: client: statfs ignores `path` parameter to fill quota information - CephFS - Ceph

Actions

Copy link

Bug #72355

closed

client: statfs ignores `path` parameter to fill quota information

Added by Venky Shankar 8 months ago. Updated 6 months ago.

Status:

Resolved

Priority:

High

Assignee:

Christopher Hoffman

Category:

Administration/Usability

Target version:

Ceph - v21.0.0

% Done:

Source:

other

Backport:

tentacle,squid

Regression:

Severity:

3 - minor

Reviewed:

Affected Versions:

ceph-qa-suite:

Component(FS):

Client, libcephfs

Labels (FS):

NFS-cluster, quotas

Pull request ID:

64864

Tags (freeform):

backport_processed

Merge Commit:

7b5fe620593d347c0d025daa20ed68b00ba4af35

Fixed In:

v20.3.0-2464-g7b5fe62059

Released In:

Upkeep Timestamp:

2025-09-23T01:37:21+00:00

Description

Client::statfs() ignores the passed in path parameter and uses the mount root to fill in quote free/available space. This causes NFS clients using cmount_path=/ to show full volume size rather than quota size. This should be fixed by finding the appropriate quota root the file is in and using it to infer the free/used blocks. It might also help to have a additional inode based interface for the same.

While we are here, NFS-Ganesha also needs to know "boundaries" are for statfs (ffilz description). Currently, @stx_dev is filled with the snapped, as per Client::fill_statx():

stx->stx_dev = in->snapid;

So, it seems like NFS-Ganesha requires unique "id" in stx_dev field for files under a cephfs subvolume. (needs to be confirmed with @Frank Filz).

Related issues 2 (0 open — 2 closed)

Actions

Copy link

Updated by Frank Filz 8 months ago

Yes, we need statfs to take an inode and return subvolume quota for that inode. The path option is less useful to us (but if only path is available, we could make it work, inode would be much easier).

We need stx_dev to be unique per sub-volume (and every inode in the sub-volume needs to have the same stx_dev). This is so Ganesha and clients can identify the sub-volume boundaries.

Actions

Copy link

Updated by Christopher Hoffman 8 months ago

Status changed from New to In Progress

Actions

Copy link

Updated by Christopher Hoffman 8 months ago

Pull request ID set to 64864

Actions

Copy link

Updated by Christopher Hoffman 8 months ago

I pushed an initial patch for this effort.

What's missing is:
1. Discuss how to identify if an inode is a subvolume or how to get it's parent subvolume inode. The mds currently tracks this. Does the client? I don't see a clear way.
2. We can not swap out 1:1 of stx_dev/st_dev to contain subvolume inode. It is used in fuse_ll.cc and will break some things if swapped out as usage is through fuse.

For number 2, @Greg Farnum mentioned perhaps a config toggle that does the requested change by @Frank Filz that is only used by NFS/someone who knows what they are doing. Another option is perhaps using f_fsid from statfs. It's currently hard coded to -1.

Actions

Copy link

Updated by Frank Filz 8 months ago

I'd be happy for Ganesha to use f_fsid.

Actions

Copy link

Updated by Christopher Hoffman 8 months ago

Updates:
1) Spoke with Venky and discussed a reliable way for the client to obtain subvolume inode.
2) During statfs, f_fsid now is populated with subvolume inode.

Actions

Copy link