Re: [PATCH v2 1/2] arm64: entry: Skip single stepping interrupt handlers

[Daniel Thompson]

On Mon, May 09, 2022 at 01:25:21PM +0530, Sumit Garg wrote:
> Hi Daniel,
> 
> On Fri, 6 May 2022 at 21:36, Daniel Thompson <daniel.thompson@xxxxxxxxxx> wrote:
> >
> > On Wed, Apr 13, 2022 at 12:24:57PM +0530, Sumit Garg wrote:
> > > Current implementation allows single stepping into interrupt handlers
> > > for interrupts that were received during single stepping. But interrupt
> > > handlers aren't something that the user expect to debug. Moreover single
> > > stepping interrupt handlers is risky as it may sometimes leads to
> > > unbalanced locking when we resume from single-step debug.
> >
> > Why does single stepping interrupt handlers cause unbalanced locking
> > (with the current code)?
> >
> 
> I have to dig deeper to completely answer this question but this is
> based on observation from following warning splats seen while single
> stepping interrupt handlers:
> 
> [  300.328300] WARNING: bad unlock balance detected!
> [  300.328608] 5.18.0-rc1-00016-g3e732ebf7316-dirty #6 Not tainted
> [  300.329058] -------------------------------------
> [  300.329298] sh/173 is trying to release lock (dbg_slave_lock) at:
> [  300.329718] [<ffffd57c951c016c>] kgdb_cpu_enter+0x7ac/0x820
> [  300.330029] but there are no more locks to release!
> [  300.330265]
> [  300.330265] other info that might help us debug this:
> [  300.330668] 4 locks held by sh/173:
> [  300.330891]  #0: ffff4f5e454d8438 (sb_writers#3){.+.+}-{0:0}, at:
> vfs_write+0x98/0x204
> [  300.331735]  #1: ffffd57c973bc2f0 (dbg_slave_lock){+.+.}-{2:2}, at:
> kgdb_cpu_enter+0x5b4/0x820
> [  300.332259]  #2: ffffd57c973a9460 (rcu_read_lock){....}-{1:2}, at:
> kgdb_cpu_enter+0xe0/0x820
> [  300.332717]  #3: ffffd57c973bc2a8 (dbg_master_lock){....}-{2:2},
> at: kgdb_cpu_enter+0x1ec/0x820

This splat tells us we entered the debugger from a regular task
(echo g > /proc/sysrq-trigger by the looks of it) and exited the
debugger from interrupt.

As such I'd be inclined to describe this as a false positive: it occurs
because we have not taught lockdep that the locks are effectively
owned by the debug trap handler rather than the task/interrupt we trapped
from.

To be honest, for a very long time been inclined to replace
dbg_slave_lock with an atomic flag instead. dbg_slave_lock isn't a lock
in any meaningful sense since it is never contended for and literally
only exists so that other cores can use raw_spin_is_locked() on it.
However sorting out dbg_master_lock is a bit more complex though which
is why I've never found the time for it (at least not yet).

BTW I suspect this patch only avoids the simplest reproduction of this
splat: I think a temporary breakpoint in the ISR that runs during the
single step would also cause the single step to complete early, yielding
the same splat.


> > > Fix broken single-step implementation via skipping single-step over
> > > interrupt handlers. The methodology is when we receive an interrupt from
> > > EL1, check if we are single stepping (pstate.SS). If yes then we save
> > > MDSCR_EL1.SS and clear the register bit if it was set. Then unmask only
> > > D and leave I set. On return from the interrupt, set D and restore
> > > MDSCR_EL1.SS. Along with this skip reschedule if we were stepping.
> >
> > Does this description really explains the motivation here.
> >
> > Isn't the goal to meet the user's expectation that when they step then
> > the system will stop at PC+4 relative the instruction they are stepping
> > (or PC+I for a branch that gets taken)?
> >
> 
> Yeah this matches my understanding as well.
> 
> > To be clear, I've no objections to leaving interrupt handling enabled
> > when single stepping (AFAIK it is similar to what x86 does) but I think
> > this patch description will be difficult for future adventurers to
> > comprehend.
> >
> 
> Okay, so does the below patch description sound apt?

Not entirely.

The lockdep splat still distracts from any explanation about why the
patch is "the right thing to do".

That the patch partially conceals a (false-positive) lockdep splat is
nice but I don't think it should dominate the description of why
single-step is unusable if a single step ends during interrupt handling.

Surely a far more powerful motivation is that (currently) on systems when
the timer interrupt (or any other fast-at-human-scale periodic interrupt)
is active then it is impossible to step any code with interrupts unlocked
because we will always end up stepping into the timer interrupt instead of
stepping the user code.


> "
>     Current implementation allows single stepping into interrupt handlers
>     for interrupts that were received during single stepping. But interrupt
>     handlers aren't something that the user expect to debug.

As above, I think it is more than user expectation. With a 100Hz timer
humans cannot react to the debugger fast enough to step before the next timer
interrupt is due.


>     Moreover single
>     stepping interrupt handlers is risky as it may sometimes leads to
>     unbalanced locking when we resume from single-step debug:

If you want to keep the comments about lockdep as an aside at the bottom
the it is important to:

s/stepping interrupt handlers/stepping into interrupt handlers/
                                      ^^^^^^

>     [  300.328300] WARNING: bad unlock balance detected!
>     [  300.328608] 5.18.0-rc1-00016-g3e732ebf7316-dirty #6 Not tainted
>     [  300.329058] -------------------------------------
>     [  300.329298] sh/173 is trying to release lock (dbg_slave_lock) at:
>     [  300.329718] [<ffffd57c951c016c>] kgdb_cpu_enter+0x7ac/0x820
>     [  300.330029] but there are no more locks to release!
>     [  300.330265]
>     [  300.330265] other info that might help us debug this:
>     [  300.330668] 4 locks held by sh/173:
>     [  300.330891]  #0: ffff4f5e454d8438 (sb_writers#3){.+.+}-{0:0},
> at: vfs_write+0x98/0x204
>     [  300.331735]  #1: ffffd57c973bc2f0 (dbg_slave_lock){+.+.}-{2:2},
> at: kgdb_cpu_enter+0x5b4/0x820
>     [  300.332259]  #2: ffffd57c973a9460 (rcu_read_lock){....}-{1:2},
> at: kgdb_cpu_enter+0xe0/0x820
>     [  300.332717]  #3: ffffd57c973bc2a8
> (dbg_master_lock){....}-{2:2}, at: kgdb_cpu_enter+0x1ec/0x820
> 
>     The common user's expectation while single stepping is that when they
>     step then the system will stop at PC+4 or PC+I for a branch that gets
>     taken relative to the instruction they are stepping. So, fix broken single
>     step implementation via skipping single stepping interrupt handlers.
> 
>     The methodology is when we receive an interrupt from EL1, check if we
>     are single stepping (pstate.SS). If yes then we save MDSCR_EL1.SS and
>     clear the register bit if it was set. Then unmask only D and leave I set. On
>     return from the interrupt, set D and restore MDSCR_EL1.SS. Along with this
>     skip reschedule if we were stepping.
> "
> 
> -Sumit
> 
> >
> > Daniel.
> >
> >
> > > Suggested-by: Will Deacon <will@xxxxxxxxxx>
> > > Signed-off-by: Sumit Garg <sumit.garg@xxxxxxxxxx>
> > > ---
> > >  arch/arm64/kernel/entry-common.c | 18 +++++++++++++++++-
> > >  1 file changed, 17 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/arch/arm64/kernel/entry-common.c b/arch/arm64/kernel/entry-common.c
> > > index 878c65aa7206..dd2d3af615de 100644
> > > --- a/arch/arm64/kernel/entry-common.c
> > > +++ b/arch/arm64/kernel/entry-common.c
> > > @@ -458,19 +458,35 @@ static __always_inline void __el1_irq(struct pt_regs *regs,
> > >       do_interrupt_handler(regs, handler);
> > >       irq_exit_rcu();
> > >
> > > -     arm64_preempt_schedule_irq();
> > > +     /* Don't reschedule in case we are single stepping */
> > > +     if (!(regs->pstate & DBG_SPSR_SS))
> > > +             arm64_preempt_schedule_irq();
> > >
> > >       exit_to_kernel_mode(regs);
> > >  }
> > > +
> > >  static void noinstr el1_interrupt(struct pt_regs *regs,
> > >                                 void (*handler)(struct pt_regs *))
> > >  {
> > > +     unsigned long reg;
> > > +
> > > +     /* Disable single stepping within interrupt handler */
> > > +     if (regs->pstate & DBG_SPSR_SS) {
> > > +             reg = read_sysreg(mdscr_el1);
> > > +             write_sysreg(reg & ~DBG_MDSCR_SS, mdscr_el1);
> > > +     }
> > > +
> > >       write_sysreg(DAIF_PROCCTX_NOIRQ, daif);
> > >
> > >       if (IS_ENABLED(CONFIG_ARM64_PSEUDO_NMI) && !interrupts_enabled(regs))
> > >               __el1_pnmi(regs, handler);
> > >       else
> > >               __el1_irq(regs, handler);
> > > +
> > > +     if (regs->pstate & DBG_SPSR_SS) {
> > > +             write_sysreg(DAIF_PROCCTX_NOIRQ | PSR_D_BIT, daif);
> > > +             write_sysreg(reg, mdscr_el1);
> > > +     }
> > >  }
> > >
> > >  asmlinkage void noinstr el1h_64_irq_handler(struct pt_regs *regs)
> > > --
> > > 2.25.1