Re: Fwd: Memory corruption in multithreaded user space program while calling fork

[Suren Baghdasaryan]

On Tue, Jul 4, 2023 at 2:28 PM Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> On Tue, 4 Jul 2023 13:22:54 -0700 Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
>
> > On Tue, Jul 4, 2023 at 9:18 AM Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
> > >
> > > On Tue, 4 Jul 2023 09:00:19 +0100 Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> > >
> > > > > > > > Thanks! I'll investigate this later today. After discussing with
> > > > > > > > Andrew, we would like to disable CONFIG_PER_VMA_LOCK by default until
> > > > > > > > the issue is fixed. I'll post a patch shortly.
> > > > > > >
> > > > > > > Posted at: https://lore.kernel.org/all/20230703182150.2193578-1-surenb@xxxxxxxxxx/
> > > > > >
> > > > > > As that change fixes something in 6.4, why not cc: stable on it as well?
> > > > >
> > > > > Sorry, I thought since per-VMA locks were introduced in 6.4 and this
> > > > > patch is fixing 6.4 I didn't need to send it to stable for older
> > > > > versions. Did I miss something?
> > > >
> > > > 6.4.y is a stable kernel tree right now, so yes, it needs to be included
> > > > there :)
> > >
> > > I'm in wait-a-few-days-mode on this.  To see if we have a backportable
> > > fix rather than disabling the feature in -stable.
> >
> > Ok, I think we have a fix posted at [2]  and it's cleanly applies to
> > 6.4.y stable branch as well. However fork() performance might slightly
> > regress, therefore disabling per-VMA locks by default for now seems to
> > be preferable even with this fix (see discussion at
> > https://lore.kernel.org/all/54cd9ffb-8f4b-003f-c2d6-3b6b0d2cb7d9@xxxxxxxxxx/).
> > IOW, both [1] and [2] should be applied to 6.4.y stable. Both apply
> > cleanly and I CC'ed stable on [2]. Greg, should I send [1] separately
> > to stable@vger?
> >
> > [1] https://lore.kernel.org/all/20230703182150.2193578-1-surenb@xxxxxxxxxx/
>
> This one isn't sufficient for .configs which already have
> PER_VMA_LOCK=y.  Using `depends on BROKEN' would be better.
>
> > [2] https://lore.kernel.org/all/20230704200656.2526715-1-surenb@xxxxxxxxxx/
> >
>
> We're still awaiting tester input on this?

Yeah, and it seems to be negative... Anyway, I'll post a dependency on BROKEN.

>
> I think a clean new fully-changelogged two-patch series would be the
> best way to handle this.  Please ensure that the [0/2] intro clearly
> explains what we're proposing here, and why.
>
> Also, "might slightly regress" is a bit weak.  These things are
> measurable, no?  Because a better solution would be to fix 6.4.x and
> mainline and leave it at that.
>