nfs: ensure that sillyrenames run to completion (try #2)
Get series with b4
b4 am [email protected]Versions newest first
- v1 Latest8 patches 10 repliesIn review
Patches in v1 No applies recorded yet
| # | Subject and files | Size | Replies | Applied |
|---|---|---|---|---|
| 1/8 | nfs: eliminate nfs3_renameargsfs/ fs/ include/ | +14 | 0 replies | |
| 2/8 | nfs: convert nfs_renameargs to use qstr structsfs/ fs/ fs/ +2 more | +20 | 0 replies | |
| 3/8 | nfs: eliminate nfs4_rename_argfs/ fs/ include/ | +8 | 4 replies | |
| 4/8 | nfs: standardize the rename response containerfs/ fs/ fs/ +2 more | +22 | 2 replies | |
| 5/8 | nfs: move nfs_sillyrename to unlink.cfs/ fs/ include/ | +85 | 0 replies | |
| 6/8 | nfs: add a rename_setup nfs_rpc_op for async renamesfs/ fs/ fs/ +1 more | +52 | 2 replies | |
| 7/8 | nfs: add rename_done nfs_rpc_opfs/ fs/ fs/ +1 more | +47 | 0 replies | |
| 8/8 | nfs: make sillyrename an async operationfs/ | +183 | 2 replies |
Reviews trailers per patch
No Reviewed-by, Acked-by or Tested-by trailers for v1 yet.
Cover letter [PATCH 0/8] nfs: ensure that sillyrenames run to completion (try #2)
This patchset is pretty much identical to the one that I sent last week. It has one minor fix -- I fixed it to use nfs_free_fattr instead of kfree in nfs_async_rename_release. I also cleaned up the comments a bit.
We had a bug report recently where someone was complaining about silly-renamed files being left around:
https://bugzilla.redhat.com/show_bug.cgi?id=511901
...they attached a reproducer to the bug which involves a pthreads-based program killing a child thread that's unlinking and closing a file.
The unlink triggers a sillyrename (since the file is still open). The parent kills the child thread and if timed just right, the child thread will be killed after the RENAME call is issued but before the reply is processed. The file will end up renamed on the server, but the client won't be aware of it and won't unlink it during dentry_iput.
This patchset is intended to prevent this from happening by having nfs_sillyrename use an asynchronous rename call and simply wait for the call to complete in TASK_KILLABLE sleep before proceeding. If the task is killed while the rename RPC is on the wire, it'll still run to completion even though the thread that initiated it was killed off.
To facilitate this, the set first changes all of the existing rename calls to use standardized argument and response containers. The nfs_sillyrename call is then moved to unlink.c (where the rest of the sillyrename code lives), and then is converted to use an asynchronous RPC call to handle the rename.
I've tested this by running the reproducer in the above bug report in a loop for several hours and never ended up with a leftover .nfs* file.
I'd like to see this in 2.6.37 if possible.
Jeff Layton (8):
nfs: eliminate nfs3_renameargs
nfs: convert nfs_renameargs to use qstr structs
nfs: eliminate nfs4_rename_arg
nfs: standardize the rename response container
nfs: move nfs_sillyrename to unlink.c
nfs: add a rename_setup nfs_rpc_op for async renames
nfs: add rename_done nfs_rpc_op
nfs: make sillyrename an async operation
fs/nfs/dir.c | 70 ------------- fs/nfs/nfs2xdr.c | 8 +- fs/nfs/nfs3proc.c | 62 +++++++++--- fs/nfs/nfs3xdr.c | 16 ++-- fs/nfs/nfs4proc.c | 43 ++++++++- fs/nfs/nfs4xdr.c | 4 +- fs/nfs/proc.c | 34 +++++- fs/nfs/unlink.c | 253 ++++++++++++++++++++++++++++++++++++++++++++++- include/linux/nfs_fs.h | 2 +- include/linux/nfs_xdr.h | 64 +++++-------- 10 files changed, 406 insertions(+), 150 deletions(-)