Re: [PATCH v2] libfs: Attempt exact-match comparison first during casefold lookup

[Gabriel Krisman Bertazi <krisman@suse.de>]

Linus Torvalds <torvalds@linux-foundation.org> writes:

> On Thu, 18 Jan 2024 at 07:42, Gabriel Krisman Bertazi <krisman@suse.de> wrote:
>>
>> But I don't see how this could be a problem.  the str pointer itself
>> can't change since it's already a copy of the dentry->d_name pointer; if
>> the string is external, it is guaranteed to exist until the end of the
>> lookup; Finally, If it's inlined, the string can change concurrently
>> which, in the worst case scenario, gives us a false result. But then,
>> VFS will retry, like we do for the case-insensitive comparison right
>> below.
>>
>> ..Right? :)
>
> Wrong, for two subtle reasons.
>
> The issue is not that the string can go away. The issue is that the
> string and the length have been loaded independently - and may not
> match.
>
> So when you do that
>
>         if (len == name->len && !memcmp(str, name->name, len))
>
> the 'name->len' you test for equality with 'len' may not match the
> length of the string allocated in 'name->name', because they are two
> different loads of two different values, and we do not hold the lock
> that makes them consistent.
>
> See the big comment (and the READ_ONCE()" in dentry_cmp(), and notice
> how dentry_cmp() intentionally doesn't use 'name->len' at all.
>

I see what you mean.  I really appreciate the explanation, by the way.
Thanks for that.

I'll follow up with a v3.

-- 
Gabriel Krisman Bertazi