Re: [PATCH] namei: revert old behaviour for filename_lookup with LOOKUP_PARENT flag

[Vineeth Remanan Pillai <vineethp@amazon.com>]

On 10/13/2016 01:26 PM, Al Viro wrote:
> On Thu, Oct 13, 2016 at 01:09:04PM -0700, Christoph Hellwig wrote:
>> On Thu, Oct 13, 2016 at 07:58:51PM +0000, Vineeth Remanan Pillai wrote:
>>> filename_lookup used to return success for non-existing file when called
>>> with LOOKUP_PARENT flag. This behaviour was changed with
>>> commit 8bcb77fabd7c ("namei: split off filename_lookupat()
>>> with LOOKUP_PARENT")
>>>
>>> The above patch split parent lookup functionality to a different function
>>> filename_parentat and changed all calls to filename_lookup(LOOKUP_PARENT)
>>> to the new function filename_parentat. But functions like kern_path which
>>> passed the flags directly to filename_lookup regressed due to this.
>>>
>>> This patch aims to fix the regressed behaviour by calling
>>> filename_parentat from filename_lookup if the flags contain LOOKUP_PARENT.
>> What callers shows te problems?  That's probaby were the fix need to got
>> in, and even if not that's still part of a good bug report.
> Out-of-tree, at a guess...  Incidentally, since filename_lookup() is not
> exported, it's probably kern_path() that gets used wherever it is.
> Depending on the details of what's being attempted,  kern_path_locked()
> might or might not be the right primitive to use, but I would probably
> start with checking if that's what the code in question really wants.
>
> Use:
> 	// kernel_string is an kernel pointer to NUL-terminated array of char
>
> 	struct path path;
> 	struct dentry *dentry;
>
> 	dentry = kern_path_locked(kernel_string, &path);
> 	if (IS_ERR(dentry))
> 		// failed while getting the parent, or not a regular last
> 		// component (/, ., .., <something>/., <something>/..)
> 		sod off // no cleanup needed
>
> 	// now path contains the resolved parent and dentry points to the
> 	// dentry of child, possibly negative; the last component of the
> 	// name can be determined from dentry->d_name.  Parent directory
> 	// is locked, making sure that directory entry won't be changed
> 	// until you are done.
>
> 	if (d_is_really_negative(dentry)) {
> 		// parent exists, but child doesn't
> 	} else {
> 		// child exists
> 	}
>
> 	// clean up: drop dentry, unlock parent, drop dentry/vfsmount of parent
> 	dput(dentry);
> 	inode_unlock(path.dentry->d_inode);
> 	path_put(&path);
>
Yes, the use case is out-of-tree and the code snippet above depicts the 
use .
Since kern_path_locked is also not exported, out-of-tree code used kern_path
for the existence check for directories.

One reference about this issue can be seen here.

http://www.gossamer-threads.com/lists/linux/kernel/2459690?do=post_view_flat#2459690

We also have a customer who complained about this functionality change.

I understand that there has been no API promises been made to this API. 
But since this is an
exported function, the change in function could cause break in 
out-of-tree kernel code. I will
rephrase the commit message to say "change in functionality" instead of 
regression