Date: Mon, 28 Nov 2016 16:40:21 -0800
From: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
To: Josh Poimboeuf <jpoimboe@redhat.com>
Cc: Vince Weaver <vincent.weaver@maine.edu>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        Peter Zijlstra <peterz@infradead.org>, Ingo Molnar <mingo@redhat.com>,
        Arnaldo Carvalho de Melo <acme@kernel.org>,
        "dvyukov@google.com" <dvyukov@google.com>
Subject: Re: perf: fuzzer BUG: KASAN: stack-out-of-bounds in __unwind_start
Reply-To: paulmck@linux.vnet.ibm.com
References: <alpine.DEB.2.20.1611241229180.25241@macbook-air>
 <20161128215411.fkis7bbimjy4v4j7@treble>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20161128215411.fkis7bbimjy4v4j7@treble>
User-Agent: Mutt/1.5.21 (2010-09-15)
Message-Id: <20161129004021.GL3924@linux.vnet.ibm.com>
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 3650
Lines: 69

On Mon, Nov 28, 2016 at 03:54:11PM -0600, Josh Poimboeuf wrote:
> On Thu, Nov 24, 2016 at 12:33:48PM -0500, Vince Weaver wrote:
> > 
> > This is on a skylake machine, linus git as of yesterday after the various 
> > kasan-related fixes went in.  Not sure if there were any that hadn't hit 
> > upstream yet.
> > 
> > Anyway I can't tell from this one what the actual trigger is.  After this 
> > mess the fuzzer process was locked, udev started complaining, and it 
> > eventually died completely after a few hours of repeated messages like 
> > this.
> > 
> > [38898.373183] INFO: rcu_sched self-detected stall on CPU
> > [38898.378452]  7-...: (5249 ticks this GP) idle=727/140000000000001/0 softirq=3141908/3141908 fqs=2625 
> > [38898.381211] INFO: rcu_sched detected stalls on CPUs/tasks:
> > [38898.381214]  0-...: (1 GPs behind) idle=05f/140000000000001/2 softirq=3285458/3285459 fqs=2625 
> > [38898.381217]  7-...: (5249 ticks this GP) idle=727/140000000000001/0 softirq=3141908/3141908 fqs=2625 
> > [38898.381218]  (detected by 1, t=5252 jiffies, g=3685053, c=3685052, q=32)
> > [38898.381244] ==================================================================
> > [38898.381247] BUG: KASAN: stack-out-of-bounds in __unwind_start+0x1a2/0x1c0 at addr ffff8801e9727c28
> > [38898.381248] Read of size 8 by task swapper/1/0
> > [38898.381250] page:ffffea0007a5c9c0 count:0 mapcount:0 mapping:          (null) index:0x0
> > [38898.381251] flags: 0x2ffff8000000000()
> > [38898.381251] page dumped because: kasan: bad access detected
> > [38898.381328] Memory state around the buggy address:
> > [38898.381330]  ffff8801e9727b00: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> > [38898.381331]  ffff8801e9727b80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> > [38898.381332] >ffff8801e9727c00: 00 00 00 00 f1 f1 f1 f1 00 00 00 00 f3 f3 f3 f3
> > [38898.381333]                                   ^
> > [38898.381334]  ffff8801e9727c80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 f1
> > [38898.381335]  ffff8801e9727d00: f1 f1 f1 00 00 00 f4 f2 f2 f2 f2 00 00 00 00 f3
> > [38898.381335] ==================================================================
> > [38898.510702]   (t=5284 jiffies g=3685053 c=3685052 q=32)
> > 
> > (That's all, the report above repeats but no useful things like a 
> > backtrace are ever printed)
> 
> After looking at the RCU stall detection code, I think the KASAN error
> and missing stack dump aren't very surprising.  RCU calls the scheduler
> dump_cpu_task() function, which seems inherently problematic: it tries
> to dump the stack of a task while it's running on another CPU.
> 
> There are some issues with that:
> 
> 1) There's no way to find the starting frame of a currently running task
>    from another CPU.
>    
>    In fact, I'm wondering how dump_cpu_task() ever worked at all?  It
>    seems like you'd have to get lucky that the sp/bp registers stored by
>    the last call to schedule() happen to point to a currently valid
>    stack frame.
> 
> 2) Even if there were a way to find the starting frame, it's racy
>    because the target task could be overwriting the stack while we're
>    reading it.
> 
> 3) IRQ/exception stack dumps would be missing anyway because the stack
>    dump code only looks at the current CPU's interrupt stacks.
> 
> Maybe dump_cpu_task() should instead run the stack dump directly from
> the target CPU, e.g. with trigger_single_cpu_backtrace() or
> smp_call_function_single()?
> 
> Paul, Peter, Ingo, any thoughts?

We used to do that, but the resulting NMIs were problematic on some
platforms.  Perhaps things have gotten better?

							Thaxn, Paul