Skip to content

nfs_instantiate() might succeed leaving dentry negative unhashed

Benjamin Coddington linux-nfs 1 version In review

Get series with b4
b4 am [email protected]

Versions newest first

  1. v1 Latest3 patches 1 replyIn review

Patches in v1 No applies recorded yet

#Subject and filesSizeRepliesApplied
1/3NFS: Refactor nfs_instantiate() for dentry referencing callersfs/nfs/dir.c include/linux/nfs_fs.h+30 -140 replies
2/3NFSv3: use nfs_add_or_obtain() to create and reference inodesfs/nfs/nfs3proc.c+36 -91 reply
3/3NFS: remove unused check for negative dentryfs/nfs/dir.c+1 -50 replies

Reviews trailers per patch

No Reviewed-by, Acked-by or Tested-by trailers for v1 yet.

Read the v1 thread

Cover letter [PATCH 0/3] nfs_instantiate() might succeed leaving dentry negative unhashed

After adding commit b0c6108ecf64 ("nfs_instantiate(): prevent multiple aliases for directory inode") my NFS client crashes while doing lustre race tests simultaneously on a local filesystem and the same filesystem exported via knfsd:

    BUG: unable to handle kernel NULL pointer dereference at 0000000000000028
     Call Trace:
      ? iput+0x76/0x200
      ? d_splice_alias+0x307/0x3c0
      ? dput.part.31+0x96/0x110
      ? nfs_instantiate+0x45/0x160 [nfs]
      nfs3_proc_setacls+0xa/0x20 [nfsv3]
      nfs3_proc_create+0x1cc/0x230 [nfsv3]
      nfs_create+0x83/0x160 [nfs]
      path_openat+0x11aa/0x14d0
      do_filp_open+0x93/0x100
      ? __check_object_size+0xa3/0x181
      do_sys_open+0x184/0x220
      do_syscall_64+0x5b/0x1b0
      entry_SYSCALL_64_after_hwframe+0x65/0xca

158 static int __nfs3_proc_setacls(struct inode *inode, struct posix_acl *acl,
159 struct posix_acl *dfacl)
160 {
161 struct nfs_server *server = NFS_SERVER(inode);

The 0x28 offset is i_sb in struct inode, we passed a NULL inode to nfs3_proc_setacls().

After taking this apart, I find the dentry in R12 has a NULL inode after nfs_instantiate(), which makes sense if we move it to the alias just after nfs_fhget() (See the referenced commit above). Indeed, on the list of children is the identical positive dentry that is left behind after d_splice_alias(). Moving it would usualy be fine for callers, except for NFSv3 because we want the inode pointer to ride the dentry back up the stack so we can set ACLs on it and/or set attributes in the case of EXCLUSIVE.

The first patch splits up nfs_instantiate() so that we can have two call paths, the original - whose callers don't care about what dentry ends up hashed, and a new one that returns the dentry or the alias.

The second patch modifies NFSv3 to use the latter path for callers that need to reference the dentry.

The third patch removes a test for positive dentry that seems to be impossible - I can't find a path for it, and it has never been hit in all my testing.

Benjamin Coddington (3):
NFS: Refactor nfs_instantiate() for dentry referencing callers
NFSv3: use nfs_add_or_obtain() to create and reference inodes
NFS: remove unused check for negative dentry

 fs/nfs/dir.c           | 41 +++++++++++++++++++++++---------------
 fs/nfs/nfs3proc.c      | 45 +++++++++++++++++++++++++++++++++---------
 include/linux/nfs_fs.h |  3 +++
 3 files changed, 64 insertions(+), 25 deletions(-)
Signature · 1 line
2.20.1