From: Mahesh Salgaonkar <mahesh@xxxxxxxxxxxxx> commit 0db880fc865ffb522141ced4bfa66c12ab1fbb70 upstream. 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. Reviewed-by: Christophe Leroy <christophe.leroy@xxxxxxxxxx> Tested-by: Shirisha Ganta <shirisha@xxxxxxxxxxxxx> Signed-off-by: Mahesh Salgaonkar <mahesh@xxxxxxxxxxxxx> Signed-off-by: Michael Ellerman <mpe@xxxxxxxxxxxxxx> Link: https://msgid.link/20240410043006.81577-1-mahesh@xxxxxxxxxxxxx [ Conflicts in arch/powerpc/include/asm/interrupt.h because machine_check_early() and machine_check_exception() has been refactored. ] Signed-off-by: Jinjie Ruan <ruanjinjie@xxxxxxxxxx> Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx> --- arch/powerpc/include/asm/percpu.h | 10 ++++++++++ arch/powerpc/kernel/mce.c | 14 +++++++++++--- arch/powerpc/kernel/setup_64.c | 2 ++ arch/powerpc/kernel/traps.c | 8 +++++++- 4 files changed, 30 insertions(+), 4 deletions(-) --- a/arch/powerpc/include/asm/percpu.h +++ b/arch/powerpc/include/asm/percpu.h @@ -15,6 +15,16 @@ #endif /* CONFIG_SMP */ #endif /* __powerpc64__ */ +#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK) && defined(CONFIG_SMP) +#include <linux/jump_label.h> +DECLARE_STATIC_KEY_FALSE(__percpu_first_chunk_is_paged); + +#define percpu_first_chunk_is_paged \ + (static_key_enabled(&__percpu_first_chunk_is_paged.key)) +#else +#define percpu_first_chunk_is_paged false +#endif /* CONFIG_PPC64 && CONFIG_SMP */ + #include <asm-generic/percpu.h> #include <asm/paca.h> --- a/arch/powerpc/kernel/mce.c +++ b/arch/powerpc/kernel/mce.c @@ -594,8 +594,15 @@ long notrace machine_check_early(struct u8 ftrace_enabled = this_cpu_get_ftrace_enabled(); this_cpu_set_ftrace_enabled(0); - /* Do not use nmi_enter/exit for pseries hpte guest */ - if (radix_enabled() || !firmware_has_feature(FW_FEATURE_LPAR)) + /* + * Do not use nmi_enter/exit for pseries hpte guest + * + * Likewise, do not use it 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 ((radix_enabled() || !firmware_has_feature(FW_FEATURE_LPAR)) && + !percpu_first_chunk_is_paged) nmi_enter(); hv_nmi_check_nonrecoverable(regs); @@ -606,7 +613,8 @@ long notrace machine_check_early(struct if (ppc_md.machine_check_early) handled = ppc_md.machine_check_early(regs); - if (radix_enabled() || !firmware_has_feature(FW_FEATURE_LPAR)) + if ((radix_enabled() || !firmware_has_feature(FW_FEATURE_LPAR)) && + !percpu_first_chunk_is_paged) nmi_exit(); this_cpu_set_ftrace_enabled(ftrace_enabled); --- a/arch/powerpc/kernel/setup_64.c +++ b/arch/powerpc/kernel/setup_64.c @@ -824,6 +824,7 @@ static int pcpu_cpu_distance(unsigned in unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; EXPORT_SYMBOL(__per_cpu_offset); +DEFINE_STATIC_KEY_FALSE(__percpu_first_chunk_is_paged); static void __init pcpu_populate_pte(unsigned long addr) { @@ -903,6 +904,7 @@ void __init setup_per_cpu_areas(void) if (rc < 0) panic("cannot initialize percpu area (err=%d)", rc); + static_key_enable(&__percpu_first_chunk_is_paged.key); delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; for_each_possible_cpu(cpu) { __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; --- a/arch/powerpc/kernel/traps.c +++ b/arch/powerpc/kernel/traps.c @@ -835,8 +835,14 @@ void machine_check_exception(struct pt_r * This is silly. The BOOK3S_64 should just call a different function * rather than expecting semantics to magically change. Something * like 'non_nmi_machine_check_exception()', perhaps? + * + * Do not use nmi_enter/exit 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. */ - const bool nmi = !IS_ENABLED(CONFIG_PPC_BOOK3S_64); + const bool nmi = !IS_ENABLED(CONFIG_PPC_BOOK3S_64) && + !percpu_first_chunk_is_paged; if (nmi) nmi_enter(); Patches currently in stable-queue which might be from mahesh@xxxxxxxxxxxxx are queue-5.10/powerpc-avoid-nmi_enter-nmi_exit-in-real-mode-interrupt.patch
