Re: [PATCH 1/2] lockref: speculatively spin waiting for the lock to be released

[Mateusz Guzik]

On Thu, Jun 13, 2024 at 8:56 PM Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> On Thu, 13 Jun 2024 at 11:48, Mateusz Guzik <mjguzik@xxxxxxxxx> wrote:
> >
> > perhaps lockdep in your config?
>
> Yes, happily it was just lockdep, and the fact that my regular tree
> doesn't have debug info, so I looked at my allmodconfig tree.
>
> I didn't *think* anything in the dentry struct should care about
> debugging, but clearly that sequence number thing did.
>
> But with that fixed, it's still the case that "we used to know about
> this", but what you actually fixed is the case of larger names than 8
> bytes.
>
> You did mention the name clashing in your commit message, but I think
> that should be the important part in the code comments too: make them
> about "these fields are hot and pretty much read-only", "these fields
> don't matter" and "this field is hot and written, and shouldn't be
> near the read-only ones".
>
> The exact byte counts may change, the basic notion doesn't.
>
> (Of course, putting it at the *end* of the structure then possibly
> causes cache conflicts with the next one - we don't force the dentries
> be cacheline aligned even if we've tried to make them generally work
> that way)
>

Look mate, I'm not going to go back and forth about this bit.

Nobody is getting a turing award for moving a field elsewhere, so as
far as I am concerned you are free to write your own version of the
patch, I don't even need to be mentioned. You are also free to grab
the commit message in whatever capacity (I guess you would at least
bench results?).

As long as lockref (the facility) and d_lockref are out of the way I'm
a happy camper.
-- 
Mateusz Guzik <mjguzik gmail.com>