[PATCH] powerpc: Avoid nmi_enter/nmi_exit in real mode interrupt.

[Christophe Leroy]

Le 25/02/2022 à 11:24, Mahesh Salgaonkar a écrit :
> nmi_enter()/nmi_exit() touches per cpu variables which can lead to kernel
> crash when invoked during real mode interrupt handling (e.g. early HMI/MCE
> interrupt handler) if percpu allocation comes from vmalloc area.
> 
> Early HMI/MCE handlers are called through DEFINE_INTERRUPT_HANDLER_NMI()
> wrapper which invokes nmi_enter/nmi_exit calls. We don't see any issue when
> percpu allocation is from the embedded first chunk. However with
> CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK enabled there are chances where percpu
> allocation can come from the vmalloc area.
> 
> With kernel command line "percpu_alloc=page" we can force percpu allocation
> to come from vmalloc area and can see kernel crash in machine_check_early:
> 
> [    1.215714] NIP [c000000000e49eb4] rcu_nmi_enter+0x24/0x110
> [    1.215717] LR [c0000000000461a0] machine_check_early+0xf0/0x2c0
> [    1.215719] --- interrupt: 200
> [    1.215720] [c000000fffd73180] [0000000000000000] 0x0 (unreliable)
> [    1.215722] [c000000fffd731b0] [0000000000000000] 0x0
> [    1.215724] [c000000fffd73210] [c000000000008364] machine_check_early_common+0x134/0x1f8
> 
> Fix this by avoiding use of nmi_enter()/nmi_exit() in real mode if percpu
> first chunk is not embedded.
> 
> Signed-off-by: Mahesh Salgaonkar <mahesh at linux.ibm.com>
> ---
>   arch/powerpc/include/asm/interrupt.h |   15 +++++++++++++++
>   arch/powerpc/include/asm/percpu.h    |    2 ++
>   arch/powerpc/kernel/setup_64.c       |    3 +++
>   3 files changed, 20 insertions(+)
> 
> diff --git a/arch/powerpc/include/asm/interrupt.h b/arch/powerpc/include/asm/interrupt.h
> index fc28f46d2f9dc..8c5b3a7ca2ab6 100644
> --- a/arch/powerpc/include/asm/interrupt.h
> +++ b/arch/powerpc/include/asm/interrupt.h
> @@ -327,6 +327,16 @@ static inline void interrupt_nmi_enter_prepare(struct pt_regs *regs, struct inte
>   	}
>   #endif
>   
> +	/*
> +	 * Do not use nmi_enter() in real mode if percpu first chunk is
> +	 * not embedded. With CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK enabled
> +	 * there are chances where percpu allocation can come from vmalloc
> +	 * area.
> +	 */
> +	if (IS_ENABLED(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK) &&
> +	    !(mfmsr() & MSR_DR) && !__percpu_embed_first_chunk)
> +		return;
> +

It would probably be better to refactor with the following test and 
avoid calling mfmsr() twice.

>   	/*
>   	 * Do not use nmi_enter() for pseries hash guest taking a real-mode
>   	 * NMI because not everything it touches is within the RMA limit.
> @@ -339,6 +349,10 @@ static inline void interrupt_nmi_enter_prepare(struct pt_regs *regs, struct inte
>   
>   static inline void interrupt_nmi_exit_prepare(struct pt_regs *regs, struct interrupt_nmi_state *state)
>   {
> +	if (IS_ENABLED(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK) &&
> +	    !(mfmsr() & MSR_DR) && !__percpu_embed_first_chunk)
> +		goto skip_nmi_exit;
> +
>   	if (!IS_ENABLED(CONFIG_PPC_BOOK3S_64) ||
>   			!firmware_has_feature(FW_FEATURE_LPAR) ||
>   			radix_enabled() || (mfmsr() & MSR_DR))

Same here, if you can avoid calling mfmsr() two times that would be better.


> @@ -349,6 +363,7 @@ static inline void interrupt_nmi_exit_prepare(struct pt_regs *regs, struct inter
>   	 * new work to do (must use irq_work for that).
>   	 */
>   
> +skip_nmi_exit:
>   #ifdef CONFIG_PPC64
>   #ifdef CONFIG_PPC_BOOK3S
>   	if (arch_irq_disabled_regs(regs)) {
> diff --git a/arch/powerpc/include/asm/percpu.h b/arch/powerpc/include/asm/percpu.h
> index 8e5b7d0b851c6..45c08f5b0b4e9 100644
> --- a/arch/powerpc/include/asm/percpu.h
> +++ b/arch/powerpc/include/asm/percpu.h
> @@ -12,6 +12,8 @@
>   
>   #define __my_cpu_offset local_paca->data_offset
>   
> +extern bool __percpu_embed_first_chunk;
> +
>   #endif /* CONFIG_SMP */
>   #endif /* __powerpc64__ */
>   
> diff --git a/arch/powerpc/kernel/setup_64.c b/arch/powerpc/kernel/setup_64.c
> index be8577ac93971..39dbf2fb23d61 100644
> --- a/arch/powerpc/kernel/setup_64.c
> +++ b/arch/powerpc/kernel/setup_64.c
> @@ -786,6 +786,7 @@ static __init int pcpu_cpu_to_node(int cpu)
>   
>   unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
>   EXPORT_SYMBOL(__per_cpu_offset);
> +bool __percpu_embed_first_chunk;
>   
>   void __init setup_per_cpu_areas(void)
>   {
> @@ -821,6 +822,8 @@ void __init setup_per_cpu_areas(void)
>   			pr_warn("PERCPU: %s allocator failed (%d), "
>   				"falling back to page size\n",
>   				pcpu_fc_names[pcpu_chosen_fc], rc);
> +		else
> +			__percpu_embed_first_chunk = true;

If I understand correctly that happens once at system startup ?

Then could you use jump_labels in order to avoid redoing the test at 
each interrupt entry ?


>   	}
>   
>   	if (rc < 0)
> 
>