"Christophe Leroy (CS GROUP)" <[email protected]> writes:

> Le 17/11/2026 à 15:25, Ritesh Harjani (IBM) a écrit :
>> With PROVE_LOCKING, boot warns once and lockdep turns itself off:
>>    DEBUG_LOCKS_WARN_ON(lockdep_hardirq_context())
>>    WARNING: kernel/locking/lockdep.c:4448 at 
>> lockdep_hardirqs_on_prepare+0x22c/0x2d4, CPU#0: swapper/0/0
>>    CPU: 0 UID: 0 PID: 0 Comm: swapper/0 Not tainted 
>> 7.4.0-rc2-00027-g074831f06e18-dirty #123 PREEMPT
>>    NIP:  c0000000002b2f2c LR: c0000000002b2f28 CTR: c000000002426f90
>>    REGS: c000000004f075c0 TRAP: 0700   Not tainted  
>> (7.4.0-rc2-00027-g074831f06e18-dirty)
>>    MSR:  8000000002021033 <SF,VEC,ME,IR,DR,RI,LE>  CR: 48022221  XER: 
>> 20040000
>>    NIP [c0000000002b2f2c] lockdep_hardirqs_on_prepare+0x22c/0x2d4
>>    LR [c0000000002b2f28] lockdep_hardirqs_on_prepare+0x228/0x2d4
>>    Call Trace:
>>      lockdep_hardirqs_on_prepare+0x228/0x2d4 (unreliable)
>>      trace_hardirqs_on+0x98/0x11c
>>      irqentry_exit+0x64c/0x1064
>>      timer_interrupt+0x274/0x6bc
>>      __replay_soft_interrupts+0x158/0x380
>>      arch_local_irq_restore+0x1cc/0x244
>>      console_flush_one_record+0x438/0x654
>>      console_unlock+0xb8/0x1e0
>>      vprintk_emit+0x300/0x56c
>>      vprintk+0x20/0x34
>>      _printk+0x40/0x54
>>      lockdep_init+0xdc/0x104
>>      start_kernel+0xa24/0xc68
>>      start_here_common+0x1c/0x20
>> 
>> Since commit bee25f97ad24 ("powerpc: Enable GENERIC_ENTRY feature") a
>> replayed handler always returns via irqentry_exit(), which traces hardirqs
>> on while the replay is still inside irq_enter():
>>    arch_local_irq_restore()
>>      irq_enter()                        hardirq_context = 1
>>      __replay_soft_interrupts()
>>        timer_interrupt(&regs)
>>          irq_enter_rcu()                hardirq_context = 2
>>          irq_exit_rcu()                 hardirq_context = 1
>>          irqentry_exit()
>>            trace_hardirqs_on()          hardirq_context != 0 -> WARN
>>        next_interrupt()
>>          lockdep_hardirq_exit()         too late
>> 
>> Leave the hardirq context before each replayed handler instead of
>> after it:
>>    arch_local_irq_restore()
>>      irq_enter()                        hardirq_context = 1
>>      __replay_soft_interrupts()
>>        lockdep_hardirq_exit()           hardirq_context = 0 // fix
>>        timer_interrupt(&regs)
>>          irq_enter_rcu()                hardirq_context = 1
>>          irq_exit_rcu()                 hardirq_context = 0
>>          irqentry_exit()
>>            trace_hardirqs_on()          ok
>>        next_interrupt()
>>          trace_hardirqs_off()
>>          lockdep_hardirq_enter()        hardirq_context = 1
>>      irq_exit()                         hardirq_context = 0
>> 
>> Fixes: bee25f97ad24 ("powerpc: Enable GENERIC_ENTRY feature")
>> Cc: [email protected]
>> Signed-off-by: Ritesh Harjani (IBM) <[email protected]>
>> ---
>>   arch/powerpc/kernel/irq_64.c | 11 +++++------
>>   1 file changed, 5 insertions(+), 6 deletions(-)
>> 
>> diff --git a/arch/powerpc/kernel/irq_64.c b/arch/powerpc/kernel/irq_64.c
>> index d5c48d1b0a31..b189e796eb34 100644
>> --- a/arch/powerpc/kernel/irq_64.c
>> +++ b/arch/powerpc/kernel/irq_64.c
>> @@ -75,12 +75,6 @@ static inline void next_interrupt(struct pt_regs *regs)
>>              WARN_ON(irq_soft_mask_return() != IRQS_ALL_DISABLED);
>>      }
>> 
>> -    /*
>> -     * We are responding to the next interrupt, so interrupt-off
>> -     * latencies should be reset here.
>> -     */
>> -    lockdep_hardirq_exit();
>> -    trace_hardirqs_on();
>>      trace_hardirqs_off();
>>      lockdep_hardirq_enter();
>>   }
>> @@ -137,18 +131,21 @@ static __no_kcsan void __replay_soft_interrupts(void)
>>      if (IS_ENABLED(CONFIG_PPC_BOOK3S) &&
>>          irq_happened_test_and_clear(PACA_IRQ_HMI)) {
>>              regs.trap = INTERRUPT_HMI;
>> +            lockdep_hardirq_exit();
>
> Can we include that in irq_happened_test_and_clear() instead of 
> duplicating it 5 times ?
>

Thanks for the suggestion. We can actually simplify this if we move this
into the caller itself i.e. __replay_soft_interrupts(). I think we could
change it to this diff then:

diff --git a/arch/powerpc/kernel/irq_64.c b/arch/powerpc/kernel/irq_64.c
index d5c48d1b0a31..e2f1e5b7a3a7 100644
--- a/arch/powerpc/kernel/irq_64.c
+++ b/arch/powerpc/kernel/irq_64.c
@@ -76,13 +76,11 @@ static inline void next_interrupt(struct pt_regs *regs)
        }

        /*
-        * We are responding to the next interrupt, so interrupt-off
-        * latencies should be reset here.
+        * We are responding to the next interrupt. The irqentry_exit() of the
+        * replayed handler has traced hardirqs on, which also reset the
+        * interrupt-off latencies. Trace them off again.
         */
-       lockdep_hardirq_exit();
-       trace_hardirqs_on();
        trace_hardirqs_off();
-       lockdep_hardirq_enter();
 }

 static inline bool irq_happened_test_and_clear(u8 irq)
@@ -116,6 +114,15 @@ static __no_kcsan void __replay_soft_interrupts(void)
         */
        local_paca->irq_happened |= PACA_IRQ_REPLAYING;

+       /*
+        * The irq_enter() of our callers is there to hold off softirqs. Each
+        * replayed handler enters and leaves the hardirq context itself and
+        * returns through irqentry_exit(), which again traces hardirqs on.
+        * Since lockdep only allows that outside of hardirq context, so leave
+        * it for the duration of the replay.
+        */
+       lockdep_hardirq_exit();
+
        ppc_save_regs(&regs);
        regs.softe = IRQS_ENABLED;
        regs.msr |= MSR_EE;
@@ -168,6 +175,8 @@ static __no_kcsan void __replay_soft_interrupts(void)
                next_interrupt(&regs);
        }

+       lockdep_hardirq_enter();
+
        local_paca->irq_happened &= ~PACA_IRQ_REPLAYING;
 }

The above simplifies it a lot, and the new state machine with
hardirq_context and hardirqs_enabled which lockdep cares about, will
look like this:
                                           hardirq_context  hardirqs_enabled
arch_local_irq_restore()
  trace_hardirqs_off()                     0             1 -> 0
  irq_enter()                           0 -> 1           0          (softirqs 
held off)
  __replay_soft_interrupts()
    lockdep_hardirq_exit()              1 -> 0           0          patch: the 
loop is not a handler
    timer_interrupt(&regs)
      irq_enter_rcu()                   0 -> 1           0          handler 
starts
      irq_exit_rcu()                    1 -> 0           0          handler done
      irqentry_exit()
        trace_hardirqs_on()                0             0 -> 1     legal: 
context is 0
    next_interrupt()
      trace_hardirqs_off()                 0             1 -> 0     still 
replaying: back to off
    ...                                 same for each pending interrupt
    lockdep_hardirq_enter()             0 -> 1           0          patch: 
restore for the caller
  irq_exit()                            1 -> 0           0          softirqs 
may run now
  trace_hardirqs_on()                      0             0 -> 1     IRQs really 
enabled


Let me know if this looks good, then I can re-spin the next revision
with the above changes.

-ritesh


Reply via email to