Re: [PATCH 3/3] arm64: debug: Remove rcu_read_lock from debug exception

[James Morse]

Hi,

On 7/18/19 3:31 PM, Masami Hiramatsu wrote:
> On Thu, 18 Jul 2019 10:20:23 +0100
> Mark Rutland <mark.rutland@xxxxxxx> wrote:
>
> > On Wed, Jul 17, 2019 at 11:22:15PM -0700, Paul E. McKenney wrote:
> > > On Thu, Jul 18, 2019 at 02:43:58PM +0900, Masami Hiramatsu wrote:
> > > > Remove rcu_read_lock()/rcu_read_unlock() from debug exception
> > > > handlers since the software breakpoint can be hit on idle task.
> >
> > Why precisely do we need to elide these? Are we seeing warnings today?
>
> Yes, unfortunately, or fortunately. Naresh reported that warns when
> ftracetest ran. I confirmed that happens if I probe on default_idle_call too.
>
> /sys/kernel/debug/tracing # echo p default_idle_call >> kprobe_events
> /sys/kernel/debug/tracing # echo 1 > events/kprobes/enable
> /sys/kernel/debug/tracing # [  135.122237]
> [  135.125035] =============================
> [  135.125310] WARNING: suspicious RCU usage

> [  135.132224] Call trace:
> [  135.132491]  dump_backtrace+0x0/0x140
> [  135.132806]  show_stack+0x24/0x30
> [  135.133133]  dump_stack+0xc4/0x10c
> [  135.133726]  lockdep_rcu_suspicious+0xf8/0x108
> [  135.134171]  call_break_hook+0x170/0x178
> [  135.134486]  brk_handler+0x28/0x68
> [  135.134792]  do_debug_exception+0x90/0x150
> [  135.135051]  el1_dbg+0x18/0x8c
> [  135.135260]  default_idle_call+0x0/0x44
> [  135.135516]  cpu_startup_entry+0x2c/0x30
> [  135.135815]  rest_init+0x1b0/0x280
> [  135.136044]  arch_call_rest_init+0x14/0x1c
> [  135.136305]  start_kernel+0x4d4/0x500

> > > The exception entry and exit use irq_enter() and irq_exit(), in this
> > > case, correct?  Otherwise RCU will be ignoring this CPU.
> >
> > This is missing today, which sounds like the underlying bug.
>
> Agreed. I'm not so familier with how debug exception is handled on arm64,
> would it be a kind of NMI or IRQ?

Debug exceptions can interrupt both SError (think: machine check) and 
pseudo-NMI, which both in turn interrupt interrupt-masked code. So they 
are a kind of NMI. But, be careful not to call 'nmi_enter()' twice, see 
do_serror() for how we work around this...


> Anyway, it seems that normal irqs are also not calling irq_enter/exit
> except for arch/arm64/kernel/smp.c
drivers/irqchip/irq-gic.c:gic_handle_irq() either calls 
handle_domain_irq() or handle_IPI(). The enter/exit calls live in those 
functions.


Thanks,

James