Re: KCSAN: data-race in __alloc_file / __alloc_file

[Linus Torvalds]

On Fri, Nov 8, 2019 at 5:28 AM Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote:
>
> Linus, what do you think of the following fix ?

I think it's incredibly ugly.

I realize that avoiding the cacheline dirtying might be worth it, but
I'd like to see some indication that it actually matters and helps
from a performance angle. We've already dirtied memory fairly close,
even if it might not share a cacheline (that structure is randomized,
we've touched - or will touch - 'cred->usage') too.

Honestly, I don't think get_cred() is even in a hotpath. Most cred use
just use the current cred that doesn't need the 'get'. So the
optimization looks somewhat questionable - for all we know it just
makes things worse.

I also don't like using a "WRITE_ONCE()" without a reason for it. In
this case, the only "reason" is that KCSAN special-cases that thing.
I'd much rather have some other way to mark it.

So it just looks hacky to me.

I like that people are looking at KCSAN, but I get a very strong
feeling that right now the workarounds for KCSAN false-positives are
incredibly ugly, and not always appropriate.

There is absolutely zero need for a WRITE_ONCE() in this case. The
code would work fine if the compiler did the zero write fifty times,
and re-ordered it wildly. We have a flag that starts out set, and we
clear it.  There's really no "write-once" about it.

               Linus