Re: [PATCH] arch/arm64 : fix error in dump_backtrace

[Daniel Thompson]

On Tue, Nov 06, 2018 at 11:00:19AM +0000, Mark Rutland wrote:
> On Tue, Nov 06, 2018 at 08:57:51AM +0000, Daniel Thompson wrote:
> > On Tue, Nov 06, 2018 at 08:39:01AM +0000, Mark Rutland wrote:
> > > On Tue, Nov 06, 2018 at 03:19:35PM +0800, Zhaoyang Huang wrote:
> > > > From: Zhaoyang Huang <zhaoyang.huang@xxxxxxxxxx>
> > > > 
> > > > In some cases, the instruction of "bl foo1" will be the last one of the
> > > > foo2[1], which will cause the lr be the first instruction of the adjacent
> > > > foo3[2]. Hence, the backtrace will show the weird result as bellow[3].
> > > > The patch will fix it by miner 4 of the lr when dump_backtrace
> > > 
> > > This has come up in the past (and a similar patch has been applied, then
> > > reverted).
> > > 
> > > In general, we don't know that a function call was made via BL, and therefore
> > > cannot know that LR - 4 is the address of the caller. The caller could set up
> > > the LR as it likes, then B or BR to the callee, and depending on how the basic
> > > blocks get laid out in memory, LR - 4 might point at something completely
> > > different.
> > > 
> > > More ideally, the compiler wouldn't end a function with a BL. When does that
> > > happen, and is there some way we could arrange for that to not happen? e.g.
> > > somehow pad a NOP after the BL.
> > 
> > It's a consequence of having __noreturn isn't it? __noreturn frees the
> > compiler from the burden of having to produce a valid return stack... so
> > it doesn't and unwinding becomes hard.
> 
> In that case, the compiler could equally just use B rather than BL,
> which this patch doesn't solve.
> 
> The documentation for the GCC noreturn attribute [1] says:
> 
> | In order to preserve backtraces, GCC will never turn calls to noreturn
> | functions into tail calls. 
> 
> ... so clearly it's not intended to mess up backtracing.

I guess that explains why the compiler chooses BL over B (since B would
be a tail call).


> IIUC we mostly use noreturn to prevent warings about uninitialised
> variables and such after a call to a noreturn function. I think
> optimization is a secondary concern.
> 
> We could ask the GCC folk if they can ensure that a noreturn function
> call leave thes LR pointing into the caller, e.g. by padding with a NOP:
> 
> 	BL	<noreturn function>
> 	NOP
> 
> That seems cheap enough, and would keep backtraces reliable.

It might be worth discussing. One related question though... is there
any other case there the symbol name other than __noreturn where the
symbol can change between LR and LR-1?

If not can't we just switch over to %pB which is designed for this case
(and in case you ask: no... I didn't know about %pB until this morning
when I started trying to implement the heuristic it uses by hand and
then discovering it by accident).

Something like: