[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <alpine.DEB.2.21.2308132148500.8596@angie.orcam.me.uk>
Date: Sun, 13 Aug 2023 22:30:52 +0100 (BST)
From: "Maciej W. Rozycki" <macro@...am.me.uk>
To: Tiezhu Yang <yangtiezhu@...ngson.cn>
cc: Thomas Bogendoerfer <tsbogend@...ha.franken.de>,
linux-mips@...r.kernel.org, linux-kernel@...r.kernel.org,
loongson-kernel@...ts.loongnix.cn
Subject: Re: [PATCH] MIPS: Remove noreturn attribute for die()
On Wed, 9 Aug 2023, Tiezhu Yang wrote:
> > So you want to keep a task alive that has caused a kernel oops in the
> > process context in this case, right? What purpose would it be for and
> > what condition causes `notify_die' to return NOTIFY_STOP? IOW why is
> > there no need to call `make_task_dead' in this case?
>
> I did some research, hope it is useful.
>
> There is a related description in Documentation/input/notifier.rst:
>
> For each kind of event but the last, the callback may return
> NOTIFY_STOP in order to "eat" the event: the notify loop is
> stopped and the keyboard event is dropped.
I saw that, but this is irrelevant. Dropping a keyboard event won't make
the system unstable (though it can make a console user unstable, out of
irritation).
> In commit 748f2edb5271 ("x86 NMI: better support for debuggers"), it said:
>
> If the notify is handled with a NOTIFY_STOP return, the
> system is given a new lease on life.
>
> In commit 004429956b48 ("handle recursive calls to bust_spinlocks()"),
> it said:
>
> However, at least on i386 die() has been capable of returning
> (and on other architectures this should really be that way, too)
> when notify_die() returns NOTIFY_STOP.
>
> In commit 22f5991c85de ("x86-64: honor notify_die() returning NOTIFY_STOP"),
> it said:
>
> This requires making die() return a value, making its callers honor
> this (and be prepared that it may return)
>
> In commit 620de2f5dc69 ("[IA64] honor notify_die() returning NOTIFY_STOP"),
> it said:
>
> This requires making die() and die_if_kernel() return a value,
> and their callers to honor this (and be prepared that it returns).
Thanks, that indeed helps, though indirectly. I think the most relevant,
though still terse explanation comes from commit 20c0d2d44029 ("[PATCH]
i386: pass proper trap numbers to die chain handlers"), which I believe is
the earliest of similar changes. The patch was originally submitted here:
<https://lore.kernel.org/r/43DDF02E.76F0.0078.0@novell.com/> and hardly
any discussion emerged, but I think the key statement is:
"[...] honor the return value from the handler chain invocation in die()
as, through a debugger, the fault may have been fixed."
Now it makes sense to me: even if ignoring the event will make the system
unstable, by allowing access through a debugger it has been compromised
already anyway.
So I think your change will be good if you update the change description
to include the justification quoted above rather than just: "the others do
it too, so it must be good" (though you can of course mention that your
change also makes our port consistent with other ones). I suggest linking
to the original i386 submission too for future reference.
Also I note that you combine three independent changes into one, so
please split it into individual patches as per our requirements.
Maciej
Powered by blists - more mailing lists