Re: [PATCH] powerpc: fix NMI exit ordering: restore soft masks after irqentry_nmi_exit()

[Shrikanth Hegde <sshegde@linux.ibm.com>]

Hi Mukesh.

On 9/29/26 12:46 PM, Mukesh Kumar Chaurasiya (IBM) wrote:
> When running with CONFIG_PPC_IRQ_SOFT_MASK_DEBUG=y and LOCKDEP=y, an
> Oops is triggered during NMI exit:
> 
>    [  202.258332] Oops: Exception in kernel mode, sig: 4 [#1]
>    NIP [000000000003e178] 0x3e178
>    LR  [c00000000003e174] arch_local_irq_restore+0x2a4/0x328
>    TRAP: 0700  MSR: 0000000000081000 <ME>
> 
> The faulting instruction at NIP is a 'trap' (0fe00000) inside the
> rfi_flush_fallback trampoline at a userspace-looking low address, reached
> after a bad rfid caused by an incorrect SRR0 value.
> 
> Root cause
> ----------
> DEFINE_INTERRUPT_HANDLER_NMI used to call arch_interrupt_nmi_exit_prepare()
> *before* irqentry_nmi_exit(), giving this sequence:
> 
>    arch_interrupt_nmi_enter_prepare()   // saves irq_soft_mask, irq_happened
>    irqentry_nmi_enter()                 // __nmi_enter() -> nmi_count++, in_nmi()=true
>    ____func(regs)                       // NMI handler body
>    arch_interrupt_nmi_exit_prepare()    // RESTORES irq_soft_mask=0 (IRQS_ENABLED)
>                                         // RESTORES irq_happened with pending bits
>    irqentry_nmi_exit()                  // instrumentation block runs here...
>      trace_hardirqs_on_prepare()        //   ...with in_nmi()==true still
>      lockdep_hardirqs_on_prepare()      //   ...and irq_soft_mask already restored to 0
>        -> arch_local_irq_restore(0)     // mask=0 -> enable path entered
>             WARN_ON_ONCE(in_nmi())      // FIRES: __nmi_exit() hasn't run yet
>      __nmi_exit()                       // nmi_count-- (too late)
> 
> arch_interrupt_nmi_exit_prepare() restores irq_soft_mask to whatever
> the interrupted thread had (IRQS_ENABLED / 0) and restores irq_happened
> with any pending deferred-interrupt bits that were present when the NMI
> fired.  This is correct and necessary before rfid back to the interrupted
> context, but it must not happen until irqentry_nmi_exit() has fully
> completed.
> 
> With irq_soft_mask=0 and pending bits in irq_happened, the
> trace_hardirqs_on_prepare() call inside irqentry_nmi_exit()'s
> instrumentation block reaches lockdep_hardirqs_on_prepare() which
> calls __trace_hardirqs_on_caller(), which eventually calls
> arch_local_irq_restore(0) (the IRQS_ENABLED / enable path).
> 
> arch_local_irq_restore() with mask=0 hits the CONFIG_PPC_IRQ_SOFT_MASK_DEBUG
> guards at irq_64.c:217:
> 
>    WARN_ON_ONCE(in_nmi());   // TRUE: __nmi_exit() not yet called
> 
> The WARN_ON_ONCE fires a 'trap' instruction (tw 31,0,0 / 0fe00000) in
> kernel mode with MSR_PR=0, MSR_IR=0 (real-mode).  The program check
> handler (TRAP 0700) takes the sig:4 (SIGILL) path via do_program_check()
> and calls die(), producing the Oops.
> 
> The RESTART_TABLE mechanism in arch_local_irq_restore then causes the
> bad rfid that lands at the userspace-looking address visible in the NIP.
> 
> The bug is invisible with LOCKDEP=n because trace_hardirqs_on_prepare()
> is a no-op when CONFIG_TRACE_IRQFLAGS is not set (TRACE_IRQFLAGS is
> selected by PROVE_LOCKING which selects LOCKDEP).
> 
> Fix
> ---
> Move arch_interrupt_nmi_exit_prepare() to after the irqentry_nmi_exit()
> block, so that:
> 
>    irqentry_nmi_exit()                 // __nmi_exit() -> nmi_count--, in_nmi()=false
>    arch_interrupt_nmi_exit_prepare()   // now safe to restore soft mask state
> 
> The instrumentation inside irqentry_nmi_exit() now executes while
> in_nmi() is still true (nmi_count not yet decremented), but
> irq_soft_mask is still IRQS_ALL_DISABLED (NMI-entry value), so
> arch_local_irq_restore() is never called with mask=0 from that path.
> After __nmi_exit() returns, arch_interrupt_nmi_exit_prepare() restores
> the pre-NMI soft mask and irq_happened, which is the correct state for
> the rfid back to the interrupted context.
> 
> Reported-by: Venkat Rao Bagalkote <venkat88@linux.ibm.com>
> Closes: https://lore.kernel.org/all/d4cccba9-ce5c-424b-864c-2248fb0d99e6@linux.ibm.com/
> Fixes: bee25f97ad24 ("powerpc: Enable GENERIC_ENTRY feature")
> Signed-off-by: Mukesh Kumar Chaurasiya (IBM) <mkchauras@gmail.com>
> ---
>   arch/powerpc/include/asm/interrupt.h | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/arch/powerpc/include/asm/interrupt.h b/arch/powerpc/include/asm/interrupt.h
> index 1b45a49e9bed..56f855c6f07e 100644
> --- a/arch/powerpc/include/asm/interrupt.h
> +++ b/arch/powerpc/include/asm/interrupt.h
> @@ -301,7 +301,6 @@ interrupt_handler long func(struct pt_regs *regs)			\
>   		state = irqentry_nmi_enter(regs);			\
>   	}								\
>   	ret = ____##func (regs);					\
> -	arch_interrupt_nmi_exit_prepare(regs, &nmi_state);		\
>   	if (mfmsr() & MSR_DR) {						\
>   		/* nmi_exit if relocations are on */			\
>   		irqentry_nmi_exit(regs, state);				\
> @@ -317,6 +316,7 @@ interrupt_handler long func(struct pt_regs *regs)			\
>   	} else {							\
>   		irqentry_nmi_exit(regs, state);				\
>   	}								\
> +	arch_interrupt_nmi_exit_prepare(regs, &nmi_state);		\
>   									\
>   	return ret;							\
>   }									\

Thanks for the fix.

Ordering also seems right now visually and it reflects similar to code previous
to generic entry.

arch_interrupt_nmi_enter_prepare
irqentry_nmi_enter
irqentry_nmi_exit
arch_interrupt_nmi_exit_prepare

So,
Reviewed-by: Shrikanth Hegde <sshegde@linux.ibm.com>