Re: [PATCH] cifs: When "refer file directly", make new inode cache if "uniqueid is different"

[Steve French <smfrench-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>]

On Tue, Apr 7, 2015 at 5:45 AM, Jeff Layton <jlayton-eUNUBHrolfbYtjvyW6yDsg@public.gmane.org> wrote:
> On Wed, 24 Dec 2014 11:27:38 +0900
> Nakajima Akira <nakajima.akira-o7dWnD6vFTHqq2nvvmkE/A@public.gmane.org> wrote:
>
>> When refer file "directly" (e.g. ls -li <filename>),
>>  if file is same name, old inode cache is used.
>> This causes that client shows wrong(old) inode number.
>> So this patch is that if uniqueid is different, return error.
>>
>> ## But this patch is applicable to when Server is UNIX.
>> ## When Server is Windows, we need another new patch.
>>
>>
>> Reproducible sample :
>> 1. create file 'a' at cifs client.
>> 2. rm 'a' and touch 'b a' at server.
>> 3. ls -li 'a' at client, then client shows wrong(old) inode number.
>>
>> Bug link:
>> https://bugzilla.kernel.org/show_bug.cgi?id=90021
>>
>>
>>
>> Signed-off-by: Nakajima Akira <nakajima.akira-o7dWnD6vFTHqq2nvvmkE/A@public.gmane.org>
>> diff -uprN -X linux-3.18-vanilla/Documentation/dontdiff linux-3.18-vanilla/fs/cifs/inode.c linux-3.18/fs/cifs/inode.c
>> --- linux-3.18-vanilla/fs/cifs/inode.c        2014-12-08 07:21:05.000000000 +0900
>> +++ linux-3.18/fs/cifs/inode.c        2014-12-19 11:07:59.127000000 +0900
>> @@ -402,9 +402,18 @@ int cifs_get_inode_info_unix(struct inod
>>                       rc = -ENOMEM;
>>       } else {
>>               /* we already have inode, update it */
>> +
>> +             /* if uniqueid is different, return error */
>> +             if (unlikely(cifs_sb->mnt_cifs_flags & CIFS_MOUNT_SERVER_INUM &&
>> +                 CIFS_I(*pinode)->uniqueid != fattr.cf_uniqueid)) {
>> +                     rc = -ENOENT;
>> +                     goto cgiiu_exit;
>> +             }
>> +
>>               cifs_fattr_to_inode(*pinode, &fattr);
>>       }
>>
>> +cgiiu_exit:
>>       return rc;
>>  }
>>
>
> Returning ENOENT here seems like the wrong error to me. That path does
> exist, it just no longer refers to the same file as before.
>
> Maybe ESTALE would be better as it would allow the VFS layer
> to revalidate the dcache and invalidate the old dentry?
>
> --
> Jeff Layton <jlayton-eUNUBHrolfbYtjvyW6yDsg@public.gmane.org>

Similar to what Jeff mentioned, isn't the nfs_relavidate_inode path
roughly equivalent to what we want here (where nfs.ko returns ESTALE
on various cases where we detect an inode that doesn't match what we
expect)?

-- 
Thanks,

Steve