[PATCH] nfs: don't match flexfiles mirrors that have different creds

[PATCH] nfs: don't match flexfiles mirrors that have different creds

[Jeff Layton]

When we're comparing mirrors to avoid adding duplicates to the to the
list, we must also consider the creds when comparing them. Otherwise,
mirrors that differ only by the credential will be improperly merged.

This is a problem if the client does a LAYOUTGET for a READ layout
first and then one for a RW, and the server sends a layout with a
non-usable uid for the READ layout. The RW mirror entry will get
dropped on the second layoutget as the client thinks the mirror is
identical to the one from the first. Then, when it goes to do a
WRITE, it ends up using the creds from the READ layout and the write
fails with EACCES.

Another possibility is to allow the RW segment to supersede the READ
one. The problem there is that the RW layout is returned, then the
client will still end up using the creds in the RW layout, which may
not be correct.

Signed-off-by: Jeff Layton <jeff.layton@primarydata.com>
---
 fs/nfs/flexfilelayout/flexfilelayout.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/fs/nfs/flexfilelayout/flexfilelayout.c b/fs/nfs/flexfilelayout/flexfilelayout.c
index 0cb1abd535e3..5d0789a41bc2 100644
--- a/fs/nfs/flexfilelayout/flexfilelayout.c
+++ b/fs/nfs/flexfilelayout/flexfilelayout.c
@@ -172,6 +172,8 @@ ff_layout_add_mirror(struct pnfs_layout_hdr *lo,
 			continue;
 		if (!ff_mirror_match_fh(mirror, pos))
 			continue;
+		if (mirror->uid != pos->uid || mirror->gid != pos->gid)
+			continue;
 		if (atomic_inc_not_zero(&pos->ref)) {
 			spin_unlock(&inode->i_lock);
 			return pos;
-- 
2.5.5

Re: [PATCH] nfs: don't match flexfiles mirrors that have different creds

[Jeff Layton]

On Sat, 2016-04-09 at 10:00 -0400, Jeff Layton wrote:
> When we're comparing mirrors to avoid adding duplicates to the to the
> list, we must also consider the creds when comparing them. Otherwise,
> mirrors that differ only by the credential will be improperly merged.
> 
> This is a problem if the client does a LAYOUTGET for a READ layout
> first and then one for a RW, and the server sends a layout with a
> non-usable uid for the READ layout. The RW mirror entry will get
> dropped on the second layoutget as the client thinks the mirror is
> identical to the one from the first. Then, when it goes to do a
> WRITE, it ends up using the creds from the READ layout and the write
> fails with EACCES.
> 
> Another possibility is to allow the RW segment to supersede the READ
> one. The problem there is that the RW layout is returned, then the
> client will still end up using the creds in the RW layout, which may
> not be correct.
> 
> Signed-off-by: Jeff Layton <jeff.layton@primarydata.com>
> ---
>  fs/nfs/flexfilelayout/flexfilelayout.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/fs/nfs/flexfilelayout/flexfilelayout.c b/fs/nfs/flexfilelayout/flexfilelayout.c
> index 0cb1abd535e3..5d0789a41bc2 100644
> --- a/fs/nfs/flexfilelayout/flexfilelayout.c
> +++ b/fs/nfs/flexfilelayout/flexfilelayout.c
> @@ -172,6 +172,8 @@ ff_layout_add_mirror(struct pnfs_layout_hdr *lo,
>  			continue;
>  		if (!ff_mirror_match_fh(mirror, pos))
>  			continue;
> +		if (mirror->uid != pos->uid || mirror->gid != pos->gid)
> +			continue;
>  		if (atomic_inc_not_zero(&pos->ref)) {
>  			spin_unlock(&inode->i_lock);
>  			return pos;

Self-NAK on this patch. As Trond pointed out to me in a private
conversation, we rely on this merging to handle layout stats correctly.
The patchset I sent with the cover letter below should supersede this
patch.

	[PATCH v2 0/7] nfs/sunrpc: fix flexfiles credential handling

Thanks!
-- 
Jeff Layton <jlayton@poochiereds.net>