On 2019-10-12, Aleksa Sarai [off-list ref] wrote:

On 2019-10-10, Linus Torvalds [off-list ref] wrote:

quoted

On Wed, Oct 9, 2019 at 10:42 PM Aleksa Sarai [off-list ref] wrote:

quoted

--- a/fs/namei.c
+++ b/fs/namei.c

@@ -2277,6 +2277,11 @@ static const char *path_init(struct nameidata *nd, unsigned flags)

        nd->m_seq = read_seqbegin(&mount_lock);

+       /* LOOKUP_IN_ROOT treats absolute paths as being relative-to-dirfd. */
+       if (flags & LOOKUP_IN_ROOT)
+               while (*s == '/')
+                       s++;
+
        /* Figure out the starting path and root (if needed). */
        if (*s == '/') {
                error = nd_jump_root(nd);

Hmm. Wouldn't this make more sense all inside the if (*s =- '/') test?
That way if would be where we check for "should we start at the root",
which seems to make more sense conceptually.

I don't really agree (though I do think that both options are pretty
ugly). Doing it before the block makes it clear that absolute paths are
just treated relative-to-dirfd -- doing it inside the block makes it
look more like "/" is a special-case for nd_jump_root(). And while that

Sorry, I meant "special-case for LOOKUP_IN_ROOT".

is somewhat true, this is just a side-effect of making the code more
clean -- my earlier versions reworked the dirfd handling to always grab
nd->root first if LOOKUP_IS_SCOPED. I switched to this method based on
Al's review.

In fairness, I do agree that the lonely while loop looks ugly.

And with the old way I did it (where we grabbed nd->root first) the
semantics were slightly more clear -- stripping leading "/"s doesn't
really look as "clearly obvious" as grabbing nd->root beforehand and
treating "/"s normally. But the code was also needlessly more complex.

quoted

That test for '/' currently has a "} else if (..)", but that's
pointless since it ends with a "return" anyway. So the "else" logic is
just noise.

This depends on the fact that LOOKUP_BENEATH always triggers -EXDEV for
nd_jump_root() -- if we ever add another "scoped lookup" flag then the
logic will have to be further reworked.

(It should be noted that the new version doesn't always end with a
"return", but you could change it to act that way given the above
assumption.)

quoted

And if you get rid of the unnecessary else, moving the LOOKUP_IN_ROOT
inside the if-statement works fine.

So this could be something like

    --- a/fs/namei.c
    +++ b/fs/namei.c
    @@ -2194,11 +2196,19 @@ static const char *path_init(struct
nameidata *nd, unsigned flags)

        nd->m_seq = read_seqbegin(&mount_lock);
        if (*s == '/') {
    -           set_root(nd);
    -           if (likely(!nd_jump_root(nd)))
    -                   return s;
    -           return ERR_PTR(-ECHILD);
    -   } else if (nd->dfd == AT_FDCWD) {
    +           /* LOOKUP_IN_ROOT treats absolute paths as being
relative-to-dirfd. */
    +           if (!(flags & LOOKUP_IN_ROOT)) {
    +                   set_root(nd);
    +                   if (likely(!nd_jump_root(nd)))
    +                           return s;
    +                   return ERR_PTR(-ECHILD);
    +           }
    +
    +           /* Skip initial '/' for LOOKUP_IN_ROOT */
    +           do { s++; } while (*s == '/');
    +   }
    +
    +   if (nd->dfd == AT_FDCWD) {
                if (flags & LOOKUP_RCU) {
                        struct fs_struct *fs = current->fs;
                        unsigned seq;

instead. The patch ends up slightly bigger (due to the re-indentation)
but now it handles all the "start at root" in the same place. Doesn't
that make sense?

It is correct (though I'd need to clean it up a bit to handle
nd_jump_root() correctly), and if you really would like me to change it
I will -- but I just don't agree that it's cleaner.

-- 
Aleksa Sarai
Senior Software Engineer (Containers)
SUSE Linux GmbH
<https://www.cyphar.com/>