Re: [PATCH] NFS: Add "device" tag to /proc/self/mountstats

[Bryan Schumaker <bjschuma@netapp.com>]

On 08/29/2011 10:43 AM, Chuck Lever wrote:
> 
> On Aug 29, 2011, at 9:46 AM, Bryan Schumaker wrote:
> 
>> On 08/27/2011 07:52 PM, Chuck Lever wrote:
>>>
>>> On Aug 26, 2011, at 4:52 PM, Bryan Schumaker wrote:
>>>
>>>> nfsiostat was failing to find mounted filesystems on recent kernels due
>>>> to changes in the VFS that resulted in a missing "device" tag in the
>>>> /proc/self/mountstats file.
>>>
>>> What was the commit ID that changed this formal kernel API?  I'd like to see the justification.
>>
>> It was changed by this commit:
>>
>> commit c7f404b40a3665d9f4e9a927cc5c1ee0479ed8f9
>> Author: Al Viro <viro@zeniv.linux.org.uk>
>> Date:   Wed Mar 16 06:59:40 2011 -0400
>>
>>    vfs: new superblock methods to override /proc/*/mount{s,info}
>>
>>    a) ->show_devname(m, mnt) - what to put into devname columns in mounts,
>>    mountinfo and mountstats
>>    b) ->show_path(m, mnt) - what to put into relative path column in mountinfo
>>
>>    Leaving those NULL gives old behaviour.  NFS switched to using those.
>>
>>    Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
> 
> It looks like dropping the "device" tag from /proc/self/mountstats was not intentional, so this is a kernel API regression.  At a guess, 2.6.39 probably works right, but 3.0 is broken.  Whatever solution we come up with might need to be cc:d to stable@kernel.org.  Be sure to verify which released kernels are broken.

I'll double check that 39 works.
> 
> Does changing nfs_show_devname() to add a "device" tag have any effect on the other two /proc files (mounts and mountinfo)?  I suspect you do want to fix show_vfsstat() instead of nfs_show_devname().  This should add "device" only where it is needed, and would work for all file systems.

Yeah, looks like it changes both.  I'm looking at show_vfsstat() now, but nfs_show_devname() will return an error if a call to nfs_path() fails.  This would probably result in an empty "device " line printed to /proc/self/mountstats.  Do you know if that would be a problem for anything?  I'm thinking about doing something like this:

diff --git a/fs/namespace.c b/fs/namespace.c
index 22bfe82..1429114 100644
--- a/fs/namespace.c
+++ b/fs/namespace.c
@@ -1109,6 +1109,7 @@ static int show_vfsstat(struct seq_file *m, void *v)
 
        /* device */
        if (mnt->mnt_sb->s_op->show_devname) {
+               seq_puts(m, "device ");
                err = mnt->mnt_sb->s_op->show_devname(m, mnt);
        } else {
                if (mnt->mnt_devname) {


> 
>>
>>
>>>
>>>> This patch re-adds the "device" tag for NFS
>>>> mount data in /proc/self/mountstats.
>>>
>>> The purpose of mountstats is to be an interface that all file systems can use.  If we can't add "device" back to all mounts listed in /proc/self/mountstats, it might be better to adjust nfsiostats to cope.
>>>
>>>> Signed-off-by: Bryan Schumaker <bjschuma@netapp.com>
>>>> ---
>>>> fs/nfs/super.c |    4 +++-
>>>> 1 files changed, 3 insertions(+), 1 deletions(-)
>>>>
>>>> diff --git a/fs/nfs/super.c b/fs/nfs/super.c
>>>> index b961cea..e6a0317 100644
>>>> --- a/fs/nfs/super.c
>>>> +++ b/fs/nfs/super.c
>>>> @@ -766,8 +766,10 @@ static int nfs_show_devname(struct seq_file *m, struct vfsmount *mnt)
>>>> 	devname = nfs_path(&dummy, mnt->mnt_root, page, PAGE_SIZE);
>>>> 	if (IS_ERR(devname))
>>>> 		err = PTR_ERR(devname);
>>>> -	else
>>>> +	else {
>>>> +		seq_puts(m, "device ");
>>>> 		seq_escape(m, devname, " \t\n\\");
>>>> +	}
>>>> 	free_page((unsigned long)page);
>>>> 	return err;
>>>> }
>>>> -- 
>>>> 1.7.6
>>>
>>
>