Skip to contentSkip to conversation

"r/o bind mounts" patch in 2.6.26-rc6 changes (breaks?) error behaviour

DISCUSSION Dave Hansen 3 participants 4 messages Jun 20, 2008 – Jun 26, 2008 1 message not in the archive VFS

lore
Share & tools
lore
#1 lore raw
In reply to a message not in the archive
Quoted 11 lines · Jens Rottmann

I noticed the following change in the error behaviour when attempting to delete nonexisting files/dirs from R/O-mounted directories:

Linux 2.6.25.7:
rm /RO_mounted_path/nonexisting_file
==> "No such file or directory"

Linux 2.6.26-rc6-git6:
rm /RO_mounted_path/nonexisting_file
==> "Read-only file system"

We looked at this way back in the beginning when I started to work on the patches. It was decided that the error codes weren't horribly important. We did have a case a bit ago where it really mattered, and Al fixed that up.

I know this changed behavior a bit, but userspace can surely work around it with read-only operations (stat) when it is really important to get the error to the user just right.

Review static long do_unlinkat(int dfd, const c · line 2381

Quoted 11 lines
(quoted diff) +3 -1
@@ -2371,8 +2373,10 @@ static long do_unlinkat(int dfd, const c
                if (inode)
                        atomic_inc(&inode->i_count);
                error = mnt_want_write(nd.path.mnt);
-               if (error)
+               if (error) {
+                       error = vfs_unlink(nd.path.dentry->d_inode, dentry);
                        goto exit2;
+               }
                error = vfs_unlink(nd.path.dentry->d_inode, dentry);
                mnt_drop_write(nd.path.mnt);

On a r/w fs where this is being performed on a r/o bind mount, this will let you unlink files. I think that's bad. :)

-- Dave

#2 lore raw
Quoted 13 lines · Dave Hansen
Quoted 11 lines · Jens Rottmann

I noticed the following change in the error behaviour when attempting to delete nonexisting files/dirs from R/O-mounted directories:

Linux 2.6.25.7:
rm /RO_mounted_path/nonexisting_file
==> "No such file or directory"

Linux 2.6.26-rc6-git6:
rm /RO_mounted_path/nonexisting_file
==> "Read-only file system"

Quoted 8 lines

We looked at this way back in the beginning when I started to work on the patches. It was decided that the error codes weren't horribly important. We did have a case a bit ago where it really mattered, and Al fixed that up.

I know this changed behavior a bit, but userspace can surely work around it with read-only operations (stat) when it is really important to get the error to the user just right.

Also, there are some nice benefits to the new way of doing it. To figure out if there is a file in that location also requires a path walk. Now, that isn't the most expensive thing in the world, but it is a heck of a lot more expensive than just doing a read-only look at a variable in the mount (in mnt_want_write()).

Don't you *want* warnings when an 'rm -f' runs into a r/o filesystem? ;)

-- Dave

#3 lore raw
Quoted 1 line · Dave Hansen

... requires a path walk ...

... which is spared only in the error path, the normal path still does the path walk anyway.

Quoted 1 line

Don't you *want* warnings when an 'rm -f' runs into a r/o filesystem? ;)

Well, not really. The "rm -f"s are in my startup scripts, supposed to delete stale lock files and lots of stuff that's usually not there unless something went wrong the last time. But most of the time, I leave the whole system mounted R/O, except some tmpfs, dead useful for kernel/hardware hacking/testing. But now I get masses of error messages on bootup - guess I have to litter the scripts with "if [ -e ... ]; then"s or just live with it. :-|

Anyway, thanks for solving this mystery for me. :-)

Jens

#4 lore raw
In reply to Dave Hansen
Quoted 13 lines · Dave Hansen
Quoted 11 lines · Jens Rottmann

I noticed the following change in the error behaviour when attempting to delete nonexisting files/dirs from R/O-mounted directories:

Linux 2.6.25.7:
rm /RO_mounted_path/nonexisting_file
==> "No such file or directory"

Linux 2.6.26-rc6-git6:
rm /RO_mounted_path/nonexisting_file
==> "Read-only file system"

Quoted 8 lines

We looked at this way back in the beginning when I started to work on the patches. It was decided that the error codes weren't horribly important. We did have a case a bit ago where it really mattered, and Al fixed that up.

I know this changed behavior a bit, but userspace can surely work around it with read-only operations (stat) when it is really important to get the error to the user just right.

I'd say userspace should not have to work around this.

Interface change, in stable series, w/o deprecation period, new interface is worse than the old one...

Pavel

Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html