Re: [PATCH] hw_breakpoint: fix single-stepping when using bpf_overflow_handler

From: Tomislav Novak <tnovak@meta.com>
To: Catalin Marinas <catalin.marinas@arm.com>
Cc: Alexei Starovoitov <ast@kernel.org>,
	Will Deacon <will@kernel.org>,
	"bpf@vger.kernel.org" <bpf@vger.kernel.org>,
	"linux-perf-users@vger.kernel.org"
	<linux-perf-users@vger.kernel.org>,
	"linux-arm-kernel@lists.infradead.org" 
	<linux-arm-kernel@lists.infradead.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	Mark Rutland <mark.rutland@arm.com>,
	Russell King <linux@armlinux.org.uk>,
	Peter Zijlstra <peterz@infradead.org>,
	Ingo Molnar <mingo@redhat.com>,
	Arnaldo Carvalho de Melo <acme@kernel.org>
Subject: Re: [PATCH] hw_breakpoint: fix single-stepping when using bpf_overflow_handler
Date: Wed, 30 Nov 2022 12:08:05 +0000	[thread overview]
Message-ID: <Y4dHocXjElVzTOms@tnovak-mbp.dhcp.thefacebook.com> (raw)
In-Reply-To: <Y4c1zOZYi3sCxzo9@arm.com>

On Wed, Nov 30, 2022 at 10:51:56AM +0000, Catalin Marinas wrote:
> On Mon, Nov 28, 2022 at 11:59:37AM +0000, Tomislav Novak wrote:
> > On Tue, Nov 15, 2022 at 03:09:37PM +0000, Catalin Marinas wrote:
> > > > On ARM platforms is_default_overflow_handler() is used to determine if
> > > > hw_breakpoint code should single-step over the watchpoint trigger or
> > > > let the custom handler deal with it.
> > > > 
> > > > Attaching a BPF program to a watchpoint replaces the handler with
> > > > bpf_overflow_handler, which isn't recognized as a default handler so we
> > > > never step over the instruction triggering the data abort exception (the
> > > > watchpoint keeps firing):
> > > > 
> > > >   # bpftrace -e 'watchpoint:0x10000000:4:w { printf("hit\n"); }' ./wp_test
> > > >   Attaching 1 probe...
> > > >   hit
> > > >   hit
> > > >   hit
> > > >   [...]
> > > > 
> > > > (wp_test performs a single 4-byte store to address 0x10000000)
> > > > 
> > > > This patch replaces the check with uses_default_overflow_handler(), which
> > > > accounts for the bpf_overflow_handler() case by also testing if the handler
> > > > invokes one of the perf_event_output functions via orig_default_handler.
> > > > 
> > > > Signed-off-by: Tomislav Novak <tnovak@fb.com>
> > > > Tested-by: Samuel Gosselin <sgosselin@fb.com> # arm64
> > > > ---
> > > >  arch/arm/kernel/hw_breakpoint.c   |  8 ++++----
> > > >  arch/arm64/kernel/hw_breakpoint.c |  4 ++--
> > > >  include/linux/perf_event.h        | 22 +++++++++++++++++++---
> > > >  3 files changed, 25 insertions(+), 9 deletions(-)
> > > 
> > > It looks like this slipped through the cracks. I'm fine with the patch
> > > but could you split the arm and arm64 parts in separate patches? Unless
> > > rmk acks it and we can take the patch through the arm64 (or perf) tree.
> > 
> > Thanks for reviewing!
> > 
> > Given the changes in the arch-independent perf_event.h, I think merging it
> > as a single commit may be easiest (assuming rmk acks it).
> > 
> > Alternatively I could move arm changes into a separate patch, keeping arm64
> > and perf_event.h in this one (possibly splitting out the latter into its own
> > commit). One that's merged, the arm patch could be submitted to linux-arm.
> > What would you prefer?
> 
> Actually, arch/arm*/kernel/hw_breakpoint.c come under the ARM PMU
> profiling, so no need to split the patch. It may need an ack from the
> generic perf maintainers for include/linux/perf.h.

Good point! I realized I've neglected to CC perf_event maintainers (sorry!),
doing so now.

-- 
T.