Re: [PATCH 2/2] perf/arm64: Allow per-task kernel breakpoints

[Bhupesh Sharma]

Hi Will,

Thanks for your inputs.

On Thu, Feb 6, 2020 at 4:09 PM Will Deacon <will@xxxxxxxxxx> wrote:
>
> On Tue, Feb 04, 2020 at 03:56:24AM +0530, Bhupesh Sharma wrote:
> > commit 478fcb2cdb23 ("arm64: Debugging support") disallowed per-task
> > kernel breakpoints on arm64 since these would have potentially
> > complicated the stepping code.
> >
> > However, we now have several use-cases (for e.g. perf) which require
> > per-task address execution h/w breakpoint to be exercised/set on arm64:
>
> To be honest, the perf interface to hw_breakpoint is an abomination and
> I think we should remove it entirely for arm64. It's flakey, complicated,
> adds code to context-switch and reduces the capabilities available to
> ptrace.

Sure, I agree.

> > For e.g. we can set address execution h/w breakpoints, using the
> > format prescribed by 'perf-list' command:
> > mem:<addr>[/len][:access]                          [Hardware breakpoint]
> >
> > Without this patch, currently 'perf stat -e' reports that per-task
> > address execution h/w breakpoints are 'not supported' on arm64. See
> > below:
> >
> > $ TEST_FUNC="vfs_read"
> > $ ADDR=0x`cat /proc/kallsyms | grep -P "\\s$TEST_FUNC\$" | cut -f1 -d' '`
> > $ perf stat -e mem:$ADDR:x -x';' -- cat /proc/cpuinfo > /dev/null
> >
> > <not supported>;;mem:0xffff00001031dd68:x;0;100.00;;
> >
> > After this patch, this use-case can be supported:
> >
> > $ TEST_FUNC="vfs_read"
> > $ ADDR=0x`cat /proc/kallsyms | grep -P "\\s$TEST_FUNC\$" | cut -f1 -d' '`
> > $ perf stat -e mem:$ADDR:x -x';' -- cat /proc/cpuinfo > /dev/null
> >
> > 5;;mem:0xfffffe0010361d20:x;912200;100.00;;
> >
> > Cc: Mark Rutland <mark.rutland@xxxxxxx>
> > Cc: Will Deacon <will@xxxxxxxxxx>
> > Cc: Catalin Marinas <catalin.marinas@xxxxxxx>
> > Cc: Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx>
> > Signed-off-by: Bhupesh Sharma <bhsharma@xxxxxxxxxx>
> > ---
> >  arch/arm64/kernel/hw_breakpoint.c | 7 -------
> >  1 file changed, 7 deletions(-)
> >
> > diff --git a/arch/arm64/kernel/hw_breakpoint.c b/arch/arm64/kernel/hw_breakpoint.c
> > index 0b727edf4104..c28f04e02845 100644
> > --- a/arch/arm64/kernel/hw_breakpoint.c
> > +++ b/arch/arm64/kernel/hw_breakpoint.c
> > @@ -562,13 +562,6 @@ int hw_breakpoint_arch_parse(struct perf_event *bp,
> >       hw->address &= ~alignment_mask;
> >       hw->ctrl.len <<= offset;
> >
> > -     /*
> > -      * Disallow per-task kernel breakpoints since these would
> > -      * complicate the stepping code.
> > -      */
> > -     if (hw->ctrl.privilege == AARCH64_BREAKPOINT_EL1 && bp->hw.target)
> > -             return -EINVAL;
> > -
>
> Sorry, but this is broken; the check is there for a reason, not just for
> fun!
>
> Look at how the step handler toggles the bp registers.

Not sure I follow. Can you please give me some pointers. All the perf
tests I have from test-suite run fine with this chunk removed.

Thanks for your help.
Regards,
Bhupesh