Re: git: 40a42785dbba - main - fcntl(F_SETFL): only allow one thread to perform F_SETFL

From: Mateusz Guzik <mjguzik_at_gmail.com>
Date: Tue, 23 Sep 2025 09:50:49 UTC
On Mon, Sep 22, 2025 at 11:05 PM Konstantin Belousov
<kostikbel@gmail.com> wrote:
>
> > On 9/22/25 13:40, Mateusz Guzik wrote:
> > > Per my previous e-mail the offset lock is already contested.
> I am curious how they are?
> If several threads do read(2)/write(2) kind of io on the same fd,
> they get what they deserve.
>

I don't know who, I had seen them with poudriere:
https://people.freebsd.org/~mjg/poudriere-offcpu-32kpkg.svg?s=vofflock

Note the % would be higher if it was not for other problems.

> > >
> > > Using a pool over a lock embedded into the struct would hinder performance.
> > >
> > > I explained why I don't consider embedding sx into struct file to be a problem.
>
> Are you saying that struct file falls into 128-byte zone, and has enough
> space for sx?

It is 80 bytes now. adding 32 still leaves 16 bytes free.