Re: [PATCH v5] ARM, ARM64, LONGARCH, XTENSA: Delay HW BP notification to task_work()

From: Peter Zijlstra

Date: Fri Nov 20 2026 - 10:58:22 EST


On Fri, Nov 20, 2026 at 04:03:20PM +0100, Peter Zijlstra wrote:
> On Fri, Nov 20, 2026 at 03:45:53PM +0100, Sebastian Andrzej Siewior wrote:
> > Waiman, Luis, Ada reported that HW breakpoints on ARM64 trigger
> > "sleeping while atomic" warnings on PREEMPT_RT. The hardware event is
> > delivered with disabled interrupts and perf intrastrucure expects
> > disabled interrupts while the overflow callback is invoked.
> >
> > The callback then sends a SIGTRAP signal for which it acquires
> > sighand_struct::siglock, a spinlock_t which becomes a sleeping lock and
> > must not be acquired in atomic context.
> >
> > Delay the event callback until the return to userland.
> > Add perf_hwbp_notify(), a generic perf callback which delayes the
> > actual callback invocation to task_work_add() callback. This callback
> > invokes the architecture defined callback perf_arch_hwbp_send_sig().
> > This requires struct callback_head and the functions require
> > ARCH_NEED_PERF_HW_NOTIFY to be defined.
> >
> > This was reported against ARM64. ARM, LongARCH and Xtensa follow the
> > same pattern are also converted. Xtensa is the only not supporting
> > PREEMPT_RT but now we have all architectures using the same pattern.
> >
> > Fixes: d8fccd9ca5f90 ("arm64: Allow to enable PREEMPT_RT.")
> > Reported-by: Luis Claudio R. Goncalves <lgoncalv@xxxxxxxxxx>
> > Reported-by: Waiman Long <longman@xxxxxxxxxx>
> > Closes: https://lore.kernel.org/all/aho0eqjMESuHxECr@xxxxxxxxxx/
> > Reviewed-by: Ada Couprie Diaz <ada.coupriediaz@xxxxxxx>
> > Tested-by: Ada Couprie Diaz <ada.coupriediaz@xxxxxxx>
> > Reviewed-by: Huacai Chen <chenhuacai@xxxxxxxxxxx>
> > Signed-off-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
> > ---
>
> Mark; you was going to stare at this :-)
>
> I'm thinking the archs are doing it wrong and should be sending the
> signal elsewhere.

Notably, I'm thinking arm64 should be sending the signal from either
do_watchpoint()/do_breakpoint(), after it has done perf_bp_event(),
or from el0_watchpt()/el0_breakpt().

Perf overflow handlers are not the place to send signals from.

That is very much what x86 also does, we send the signal from
exc_debug_user(); which is roughly equivalent to the above mentioned
functions.