Skip site navigation (1)Skip section navigation (2)
Date:      Mon, 22 Sep 2025 19:40:42 +0200
From:      Mateusz Guzik <mjguzik@gmail.com>
To:        John Baldwin <jhb@freebsd.org>
Cc:        Konstantin Belousov <kib@freebsd.org>, src-committers@freebsd.org,  dev-commits-src-all@freebsd.org, dev-commits-src-main@freebsd.org
Subject:   Re: git: 40a42785dbba - main - fcntl(F_SETFL): only allow one thread to perform F_SETFL
Message-ID:  <CAGudoHEMeDzTUBrFo1Gzzq-NhGOkOS0U_-JcfZS_ukKRZYt9XQ@mail.gmail.com>
In-Reply-To: <ac66d750-3daa-47c6-8b97-51bb69f9a88a@FreeBSD.org>
References:  <202509191419.58JEJsvj031867@gitrepo.freebsd.org> <92831372-745d-4612-b38f-aeb235dd8cca@FreeBSD.org> <CAGudoHGZs3iOLmbRBwhanNHtDRmd5BE%2Buorq8onCAbCkFw39iw@mail.gmail.com> <ac66d750-3daa-47c6-8b97-51bb69f9a88a@FreeBSD.org>

next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, Sep 22, 2025 at 7:39=E2=80=AFPM John Baldwin <jhb@freebsd.org> wrot=
e:
>
> On 9/22/25 04:54, Mateusz Guzik wrote:
> > On Mon, Sep 22, 2025 at 10:41=E2=80=AFAM John Baldwin <jhb@freebsd.org>=
 wrote:
> >>
> >> On 9/19/25 10:19, Konstantin Belousov wrote:
> >>> The branch main has been updated by kib:
> >>>
> >>> URL: https://cgit.FreeBSD.org/src/commit/?id=3D40a42785dbba93cc519617=
8fc49d340c1a89cabe
> >>>
> >>> commit 40a42785dbba93cc5196178fc49d340c1a89cabe
> >>> Author:     Konstantin Belousov <kib@FreeBSD.org>
> >>> AuthorDate: 2025-09-11 10:05:04 +0000
> >>> Commit:     Konstantin Belousov <kib@FreeBSD.org>
> >>> CommitDate: 2025-09-19 14:19:13 +0000
> >>>
> >>>       fcntl(F_SETFL): only allow one thread to perform F_SETFL
> >>>
> >>>       Use f_vflags file locking for this.
> >>>       Allowing more than one thread handling F_SETFL might cause de-s=
ync
> >>>       between real driver state and flags.
> >>>
> >>>       Reviewed by:    markj
> >>>       Tested by:      pho
> >>>       Sponsored by:   The FreeBSD Foundation
> >>>       MFC after:      2 weeks
> >>>       Differential revision:  https://reviews.freebsd.org/D52487
> >>
> >> Thanks for fixing this.  I still slightly worry that "home-grown" lock=
s
> >> aren't visible to WITNESS and it's checking.
> >>
> >
> > Another problem with these is that they don't do adaptive spinning.
> >
> > In particular for file offset, it *is* putting threads off cpu in real
> > workloads when it plausibly could be avoided.
> >
> > I think the real thing to do here is to drop the hand-rolled machinery
> > and use an sx lock.
> >
> > Currently struct file is 80 bytes which is a very nasty size from
> > caching standpoint.
> >
> > Locks are 32 bytes in size, which is another problem, but ultimately
> > one can be added here without growing the struct past 128 bytes.
> >
> > The only issue here is that files are marked as NOFREE, so this memory
> > can *never* be reclaimed.
> >
> > One could be tempted to use smr here, but the cost of smr_enter is
> > prohibitive. There is a lazy variant which does not do atomics, which
> > perhaps could work, but that 0 users in the tree and was probably
> > never tested.
> >
> > With 32-bit archs going away I don't think it's a big deal though.
> >
> > For interested, on Linux the struct is 256 bytes.
>
> I had suggested in an earlier review adding an sx-pool similar to our
> existing mtxpool and using that.  That would avoid bloating the structure
> with a dedicated lock.
>

Per my previous e-mail the offset lock is already contested.

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 prob=
lem.



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CAGudoHEMeDzTUBrFo1Gzzq-NhGOkOS0U_-JcfZS_ukKRZYt9XQ>