Re: [PATCH] kprobes, x86/alternatives: use text_mutex to protect smp_alt_modules

From: Masami Hiramatsu
Date: Sat Oct 28 2017 - 04:45:08 EST


On Fri, 27 Oct 2017 21:30:24 +0800
zhouchengming <zhouchengming1@xxxxxxxxxx> wrote:

> On 2017/10/27 20:33, Borislav Petkov wrote:
> > On Fri, Oct 27, 2017 at 07:42:45PM +0800, zhouchengming wrote:
> >> This is a real bug happened on one of our machines, below is the calltrace.
> >> We can see the trigger is at alternatives_text_reserved+0x20/0x80, and
> >> encounter a deleted (poisoned) list_head.
> > Looks like some out-of-tree, old kernel thing. We don't have
> > mlx4_stats_sysfs_create() upstream and looking at the boot timestamps,
> > it could be that register_jprobe() is not ready yet.
>
> Yes, it's an out-of-tree module, loaded when boot kernel. register_kprobe()
> maybe not ready yet, but the bug is not caused by it obviously.
>
> >
> > Looking at the Code, though:
> >
> > 20: 74 59 je 0x7b
> > 22: 66 0f 1f 84 00 00 00 nopw 0x0(%rax,%rax,1)
> > 29: 00 00
> > 2b:* 48 3b 71 20 cmp 0x20(%rcx),%rsi<-- trapping instruction
> > 2f: 72 3a jb 0x6b
> > 31: 48 3b 79 28 cmp 0x28(%rcx),%rdi
> > 35: 77 34 ja 0x6b
> >
> > %rcx is 0xdead0000000000d0 and that is POISON_POINTER_DELTA + 0xd0 so
> > that looks more like smp_alt_modules is not initialized yet but I could
> > could very well be wrong because this is an old kernel. So trigger that
> > with the upstream kernel without out of tree modules.
>
> The smp_alt_modules is defined by LIST_HEAD, so it's initialized at start.
>
> A deleted list_head->next = LIST_POISON1 = 0xdead000000000000 + 0x100, then
> container_of() to get the struct smp_alt_module: -0x30 = 0xdead0000000000d0
>
> Obviously, it's a deleted list_head, and I have explained clearly how it happen in
> the patch comment.

Ah, I see. It looks alternatives_text_reserved() bug at a glance.
But simply adding smp_alt mutex to alternatives_text_reserved() causes
ABBA deadlock in the kprobe's path.
So your solution is to replace the smp_alt with text_mutex, since
alternatives_text_reserved is x86 specific function.

Hmm, let me see... I agree that will be a simple way to solve, but
it also means we have 2 resources protected by text_mutex.

Thank you,


--
Masami Hiramatsu <mhiramat@xxxxxxxxxx>