Date: Wed, 13 Sep 2017 03:42:19 +0100
From: Will Deacon <will.deacon@arm.com>
To: James Morse <james.morse@arm.com>
Cc: Pratyush Anand <panand@redhat.com>,
        linux-arm-kernel@lists.infradead.org,
        Catalin Marinas <catalin.marinas@arm.com>,
        linux-kernel@vger.kernel.org
Subject: Re: [PATCH v2] arm64: fix unwind_frame() for filtered out fn for
 function graph tracing
Message-ID: <20170913024218.GB12411@arm.com>
References: <3c175a7a1f7c2e08098f6d5e84dc247ce94846d2.1504244801.git.panand@redhat.com>
 <59B7AED4.1070106@arm.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <59B7AED4.1070106@arm.com>
User-Agent: Mutt/1.5.23 (2014-03-12)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1885
Lines: 42

On Tue, Sep 12, 2017 at 10:54:28AM +0100, James Morse wrote:
> Hi Pratyush,
> 
> On 01/09/17 06:48, Pratyush Anand wrote:
> > do_task_stat() calls get_wchan(), which further does unbind_frame().
> > unbind_frame() restores frame->pc to original value in case function
> > graph tracer has modified a return address (LR) in a stack frame to hook
> > a function return. However, if function graph tracer has hit a filtered
> > function, then we can't unwind it as ftrace_push_return_trace() has
> > biased the index(frame->graph) with a 'huge negative'
> > offset(-FTRACE_NOTRACE_DEPTH).
> > 
> > Moreover, arm64 stack walker defines index(frame->graph) as unsigned
> > int, which can not compare a -ve number.
> > 
> > Similar problem we can have with calling of walk_stackframe() from
> > save_stack_trace_tsk() or dump_backtrace().
> > 
> > This patch fixes unwind_frame() to test the index for -ve value and
> > restore index accordingly before we can restore frame->pc.
> 
> I've just spotted arm64's profile_pc, which does this:
> From arch/arm64/kernel/time.c:profile_pc():
> > #ifdef CONFIG_FUNCTION_GRAPH_TRACER
> > 	frame.graph = -1; /* no task info */
> > #endif
> 
> Is this another elaborate way of hitting this problem?
> 
> I guess the options are skip any return-address restore in the unwinder if
> frame.graph is -1. (and profile_pc may have a bug here). Or, put
> current->curr_ret_stack in there.
> 
> profile_pc() always passes tsk=NULL, so the unwinder assumes its current...
> kernel/profile.c pulls the pt_regs from a per-cpu irq_regs variable, that is
> updated by handle_IPI ... so it looks like this should always be current...

Hmmm... is profile_pc the *only* case where frame->graph isn't equal to
tsk->curr_ret_stack in unwind_frame? If so, maybe unwind_frame should just
use that, and we could kill the graph member of struct stackframe completely?

Will