Re: Access to non-RAM pages

[Linus Torvalds]

On Sun, Sep 2, 2018 at 7:25 PM Benjamin Herrenschmidt
<benh@xxxxxxxxxxxxxxxxxxx> wrote:
> Ah, my bad reading, I was looking at read_word_at_a_time() instead of
> load_unaligned_zeropad(). I'm not familiar enough with the dentry qstr
> stuff, I assume this is safe ?

The dentry qstr should always be 8-byte aligned because it's a kernel
name allocation.

So it's the path component in the actual pathname string that can be
unaligned (ct/tcount in dentry_string_cmp).

The comment actually does talk about it, although the comment also
claims that the cs read would use load_unaligned_zeropad(), which it
no longer does (now it only does the read_word_at_a_time).

And read_word_at_a_time() is purely a KASAN thing. The thing can't
fault, but it *can* read uninitialized data past the end of the
string, making KASAN unhappy.

So that's actually a different issue, where KASAN does byte-level
validity testing, and doing word-at-a-time accesses obviously violates
that for strings.

               Linus