[PATCH][CIFS] Use FS_RENAME_DOES_D_MOVE to minimize races in rename

[PATCH][CIFS] Use FS_RENAME_DOES_D_MOVE to minimize races in rename

[Steve French]

[-- Attachment #1: Type: text/plain, Size: 110 bytes --]

Should be safer to do the dentry move immediately after the rename
rather than later.





-- 
Thanks,

Steve

[-- Attachment #2: 0001-cifs-do-d_move-in-rename.patch --]
[-- Type: text/x-patch, Size: 1415 bytes --]

From c1bc75178780665716ad14b8c40a8d8e58d0f3e3 Mon Sep 17 00:00:00 2001
From: Steve French <stfrench@microsoft.com>
Date: Tue, 25 Feb 2020 18:08:54 -0600
Subject: [PATCH] cifs: do d_move in rename

It is safer to do the d_move closer to the rename to
avoid races.

Signed-off-by: Steve French <stfrench@microsoft.com>
---
 fs/cifs/cifsfs.c | 4 ++--
 fs/cifs/inode.c  | 2 ++
 2 files changed, 4 insertions(+), 2 deletions(-)

diff --git a/fs/cifs/cifsfs.c b/fs/cifs/cifsfs.c
index fa77fe5258b0..94e3ed4850b5 100644
--- a/fs/cifs/cifsfs.c
+++ b/fs/cifs/cifsfs.c
@@ -1018,7 +1018,7 @@ struct file_system_type cifs_fs_type = {
 	.name = "cifs",
 	.mount = cifs_do_mount,
 	.kill_sb = cifs_kill_sb,
-	/*  .fs_flags */
+	.fs_flags = FS_RENAME_DOES_D_MOVE,
 };
 MODULE_ALIAS_FS("cifs");
 
@@ -1027,7 +1027,7 @@ static struct file_system_type smb3_fs_type = {
 	.name = "smb3",
 	.mount = smb3_do_mount,
 	.kill_sb = cifs_kill_sb,
-	/*  .fs_flags */
+	.fs_flags = FS_RENAME_DOES_D_MOVE,
 };
 MODULE_ALIAS_FS("smb3");
 MODULE_ALIAS("smb3");
diff --git a/fs/cifs/inode.c b/fs/cifs/inode.c
index 6543465595f6..7bb7ff110dc6 100644
--- a/fs/cifs/inode.c
+++ b/fs/cifs/inode.c
@@ -1835,6 +1835,8 @@ cifs_do_rename(const unsigned int xid, struct dentry *from_dentry,
 		CIFSSMBClose(xid, tcon, fid.netfid);
 	}
 do_rename_exit:
+	if (rc == 0)
+		d_move(from_dentry, to_dentry);
 	cifs_put_tlink(tlink);
 	return rc;
 }
-- 
2.20.1

Re: [PATCH][CIFS] Use FS_RENAME_DOES_D_MOVE to minimize races in rename

[Amir Goldstein]

On Wed, Feb 26, 2020 at 8:37 AM Steve French <smfrench@gmail.com> wrote:
>

What sort of rename races? Is that a real/reproducible problem?

> Should be safer to do the dentry move immediately after the rename
> rather than later.
>

Looking at what "later" means, you moved d_move() before
cifs_put_tlink(tlink);
and
shrink_dcache_parent(new_dentry);
detach_mounts(new_dentry);

I suppose cifs_put_tlink() is not the issue.

It makes me wonder about shrink_dcache_parent()/detach_mounts() -
they happen for some fs before d_move() and for some fs after d_move()

I think it kind of makes sense to have them moved after dentry is
unhashed anyway? (that is after d_move()).

Thanks,
Amir.