Re: [PATCH][V3] arm64: perf: Get the wrong PC value in REGS_ABI_32 mode

[Mark Rutland]

On Tue, May 26, 2020 at 08:54:19PM +0100, Will Deacon wrote:
> On Tue, May 26, 2020 at 11:26:11AM +0100, Mark Rutland wrote:
> > On Mon, May 11, 2020 at 10:52:07AM +0800, Jiping Ma wrote:
> > > Modified the patch subject and the change description.
> > > 
> > > PC value is get from regs[15] in REGS_ABI_32 mode, but correct PC
> > > is regs->pc(regs[PERF_REG_ARM64_PC]) in arm64 kernel, which caused
> > > that perf can not parser the backtrace of app with dwarf mode in the 
> > > 32bit system and 64bit kernel.
> > > 
> > > Signed-off-by: Jiping Ma <jiping.ma2@xxxxxxxxxxxxx>
> > 
> > Thanks for this.
> > 
> > 
> > > ---
> > >  arch/arm64/kernel/perf_regs.c | 4 ++++
> > >  1 file changed, 4 insertions(+)
> > > 
> > > diff --git a/arch/arm64/kernel/perf_regs.c b/arch/arm64/kernel/perf_regs.c
> > > index 0bbac61..0ef2880 100644
> > > --- a/arch/arm64/kernel/perf_regs.c
> > > +++ b/arch/arm64/kernel/perf_regs.c
> > > @@ -32,6 +32,10 @@ u64 perf_reg_value(struct pt_regs *regs, int idx)
> > >  	if ((u32)idx == PERF_REG_ARM64_PC)
> > >  		return regs->pc;
> > >  
> > > +	if (perf_reg_abi(current) == PERF_SAMPLE_REGS_ABI_32
> > > +		&& idx == 15)
> > > +		return regs->pc;
> > 
> > I think there are some more issues here, and we may need a more
> > substantial rework. For a compat thread, we always expose
> > PERF_SAMPLE_REGS_ABI_32 via per_reg_abi(), but for some reason
> > perf_reg_value() also munges the compat SP/LR into their ARM64
> > equivalents, which don't exist in the 32-bit sample ABI. We also don't
> > zero the regs that don't exist in 32-bit (including the aliasing PC).
> 
> I think this was for the case where you have a 64-bit perf profiling a
> 32-bit task, and it was passing the registers off to libunwind. Won't that
> break if we follow your suggestion?

Oh yuck; have we messed up the ABI here, or have I misunderstood?

Is arm64's PERF_SAMPLE_REGS_ABI_32 supposed to be the same as the 32-bit
arm's PERF_SAMPLE_REGS_ABI_32?

If yes, and the differences are being relied upon by 64-bit consumers,
that's a nasty ABI issue we've introduced for compat tasks, and I don't
think this patch alone is quite right.

If no, then I don't see that any change is necessary, as we already
expose the information, and it's a userspace bug to expect the PC in a
place where the kernel has never exposed it.

Thanks,
Mark.