| Message ID | 20241021042218.746659-5-yuzhao@google.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | mm/arm64: re-enable HVO | expand |
On Mon, 21 Oct 2024 05:22:16 +0100, Yu Zhao <yuzhao@google.com> wrote: > > Broadcast pseudo-NMI IPIs to pause remote CPUs for a short period of > time, and then reliably resume them when the local CPU exits critical > sections that preclude the execution of remote CPUs. > > A typical example of such critical sections is BBM on kernel PTEs. > HugeTLB Vmemmap Optimization (HVO) on arm64 was disabled by > commit 060a2c92d1b6 ("arm64: mm: hugetlb: Disable > HUGETLB_PAGE_OPTIMIZE_VMEMMAP") due to the folllowing reason: > > This is deemed UNPREDICTABLE by the Arm architecture without a > break-before-make sequence (make the PTE invalid, TLBI, write the > new valid PTE). However, such sequence is not possible since the > vmemmap may be concurrently accessed by the kernel. > > Supporting BBM on kernel PTEs is one of the approaches that can make > HVO theoretically safe on arm64. Is the safety only theoretical? I would have expected that we'd use an approach that is absolutely rock-solid. > > Note that it is still possible for the paused CPUs to perform > speculative translations. Such translations would cause spurious > kernel PFs, which should be properly handled by > is_spurious_el1_translation_fault(). Speculative translation faults are never reported, that'd be a CPU bug. *Spurious* translation faults can be reported if the CPU doesn't implement FEAT_ETS2, for example, and that has to do with the ordering of memory access wrt page-table walking for the purpose of translations. > > Signed-off-by: Yu Zhao <yuzhao@google.com> > --- > arch/arm64/include/asm/smp.h | 3 ++ > arch/arm64/kernel/smp.c | 92 +++++++++++++++++++++++++++++++++--- > 2 files changed, 88 insertions(+), 7 deletions(-) > > diff --git a/arch/arm64/include/asm/smp.h b/arch/arm64/include/asm/smp.h > index 2510eec026f7..cffb0cfed961 100644 > --- a/arch/arm64/include/asm/smp.h > +++ b/arch/arm64/include/asm/smp.h > @@ -133,6 +133,9 @@ bool cpus_are_stuck_in_kernel(void); > extern void crash_smp_send_stop(void); > extern bool smp_crash_stop_failed(void); > > +void pause_remote_cpus(void); > +void resume_remote_cpus(void); > + > #endif /* ifndef __ASSEMBLY__ */ > > #endif /* ifndef __ASM_SMP_H */ > diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c > index 3b3f6b56e733..68829c6de1b1 100644 > --- a/arch/arm64/kernel/smp.c > +++ b/arch/arm64/kernel/smp.c > @@ -85,7 +85,12 @@ static int ipi_irq_base __ro_after_init; > static int nr_ipi __ro_after_init = NR_IPI; > static struct irq_desc *ipi_desc[MAX_IPI] __ro_after_init; > > -static bool crash_stop; > +enum { > + SEND_STOP = BIT(0), > + CRASH_STOP = BIT(1), > +}; > + > +static unsigned long stop_in_progress; > > static void ipi_setup(int cpu); > > @@ -917,6 +922,79 @@ static void __noreturn ipi_cpu_crash_stop(unsigned int cpu, struct pt_regs *regs > #endif > } > > +static DEFINE_SPINLOCK(cpu_pause_lock); PREEMPT_RT will turn this into a sleeping lock. Is it safe to sleep as you are dealing with kernel mappings? > +static cpumask_t paused_cpus; > +static cpumask_t resumed_cpus; > + > +static void pause_local_cpu(void) > +{ > + int cpu = smp_processor_id(); > + > + cpumask_clear_cpu(cpu, &resumed_cpus); > + /* > + * Paired with pause_remote_cpus() to confirm that this CPU not only > + * will be paused but also can be reliably resumed. > + */ > + smp_wmb(); > + cpumask_set_cpu(cpu, &paused_cpus); > + /* paused_cpus must be set before waiting on resumed_cpus. */ > + barrier(); I'm not sure what this is trying to enforce. Yes, the compiler won't reorder the set and the test. But your comment seems to indicate that also need to make sure the CPU preserves that ordering, and short of a DMB, the test below could be reordered. > + while (!cpumask_test_cpu(cpu, &resumed_cpus)) > + cpu_relax(); > + /* A typical example for sleep and wake-up functions. */ I'm not sure this is "typical",... > + smp_mb(); > + cpumask_clear_cpu(cpu, &paused_cpus); > +} > + > +void pause_remote_cpus(void) > +{ > + cpumask_t cpus_to_pause; > + > + lockdep_assert_cpus_held(); > + lockdep_assert_preemption_disabled(); > + > + cpumask_copy(&cpus_to_pause, cpu_online_mask); > + cpumask_clear_cpu(smp_processor_id(), &cpus_to_pause); This bitmap is manipulated outside of your cpu_pause_lock. What guarantees you can't have two CPUs stepping on each other here? > + > + spin_lock(&cpu_pause_lock); > + > + WARN_ON_ONCE(!cpumask_empty(&paused_cpus)); > + > + smp_cross_call(&cpus_to_pause, IPI_CPU_STOP_NMI); > + > + while (!cpumask_equal(&cpus_to_pause, &paused_cpus)) > + cpu_relax(); This can be a lot of things to compare, specially that you are explicitly mentioning large systems. Why can't this be implemented as a counter instead? Overall, this looks like stop_machine() in disguise. Why can't this use the existing infrastructure? Thanks, M.
diff --git a/arch/arm64/include/asm/smp.h b/arch/arm64/include/asm/smp.h index 2510eec026f7..cffb0cfed961 100644 --- a/arch/arm64/include/asm/smp.h +++ b/arch/arm64/include/asm/smp.h @@ -133,6 +133,9 @@ bool cpus_are_stuck_in_kernel(void); extern void crash_smp_send_stop(void); extern bool smp_crash_stop_failed(void); +void pause_remote_cpus(void); +void resume_remote_cpus(void); + #endif /* ifndef __ASSEMBLY__ */ #endif /* ifndef __ASM_SMP_H */ diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c index 3b3f6b56e733..68829c6de1b1 100644 --- a/arch/arm64/kernel/smp.c +++ b/arch/arm64/kernel/smp.c @@ -85,7 +85,12 @@ static int ipi_irq_base __ro_after_init; static int nr_ipi __ro_after_init = NR_IPI; static struct irq_desc *ipi_desc[MAX_IPI] __ro_after_init; -static bool crash_stop; +enum { + SEND_STOP = BIT(0), + CRASH_STOP = BIT(1), +}; + +static unsigned long stop_in_progress; static void ipi_setup(int cpu); @@ -917,6 +922,79 @@ static void __noreturn ipi_cpu_crash_stop(unsigned int cpu, struct pt_regs *regs #endif } +static DEFINE_SPINLOCK(cpu_pause_lock); +static cpumask_t paused_cpus; +static cpumask_t resumed_cpus; + +static void pause_local_cpu(void) +{ + int cpu = smp_processor_id(); + + cpumask_clear_cpu(cpu, &resumed_cpus); + /* + * Paired with pause_remote_cpus() to confirm that this CPU not only + * will be paused but also can be reliably resumed. + */ + smp_wmb(); + cpumask_set_cpu(cpu, &paused_cpus); + /* paused_cpus must be set before waiting on resumed_cpus. */ + barrier(); + while (!cpumask_test_cpu(cpu, &resumed_cpus)) + cpu_relax(); + /* A typical example for sleep and wake-up functions. */ + smp_mb(); + cpumask_clear_cpu(cpu, &paused_cpus); +} + +void pause_remote_cpus(void) +{ + cpumask_t cpus_to_pause; + + lockdep_assert_cpus_held(); + lockdep_assert_preemption_disabled(); + + cpumask_copy(&cpus_to_pause, cpu_online_mask); + cpumask_clear_cpu(smp_processor_id(), &cpus_to_pause); + + spin_lock(&cpu_pause_lock); + + WARN_ON_ONCE(!cpumask_empty(&paused_cpus)); + + smp_cross_call(&cpus_to_pause, IPI_CPU_STOP_NMI); + + while (!cpumask_equal(&cpus_to_pause, &paused_cpus)) + cpu_relax(); + /* + * Paired with pause_local_cpu() to confirm that all CPUs not only will + * be paused but also can be reliably resumed. + */ + smp_rmb(); + WARN_ON_ONCE(cpumask_intersects(&cpus_to_pause, &resumed_cpus)); + + spin_unlock(&cpu_pause_lock); +} + +void resume_remote_cpus(void) +{ + cpumask_t cpus_to_resume; + + lockdep_assert_cpus_held(); + lockdep_assert_preemption_disabled(); + + cpumask_copy(&cpus_to_resume, cpu_online_mask); + cpumask_clear_cpu(smp_processor_id(), &cpus_to_resume); + + spin_lock(&cpu_pause_lock); + + cpumask_setall(&resumed_cpus); + /* A typical example for sleep and wake-up functions. */ + smp_mb(); + while (cpumask_intersects(&cpus_to_resume, &paused_cpus)) + cpu_relax(); + + spin_unlock(&cpu_pause_lock); +} + static void arm64_backtrace_ipi(cpumask_t *mask) { __ipi_send_mask(ipi_desc[IPI_CPU_BACKTRACE], mask); @@ -970,7 +1048,9 @@ static void do_handle_IPI(int ipinr) case IPI_CPU_STOP: case IPI_CPU_STOP_NMI: - if (IS_ENABLED(CONFIG_KEXEC_CORE) && crash_stop) { + if (!test_bit(SEND_STOP, &stop_in_progress)) { + pause_local_cpu(); + } else if (test_bit(CRASH_STOP, &stop_in_progress)) { ipi_cpu_crash_stop(cpu, get_irq_regs()); unreachable(); } else { @@ -1142,7 +1222,6 @@ static inline unsigned int num_other_online_cpus(void) void smp_send_stop(void) { - static unsigned long stop_in_progress; cpumask_t mask; unsigned long timeout; @@ -1154,7 +1233,7 @@ void smp_send_stop(void) goto skip_ipi; /* Only proceed if this is the first CPU to reach this code */ - if (test_and_set_bit(0, &stop_in_progress)) + if (test_and_set_bit(SEND_STOP, &stop_in_progress)) return; /* @@ -1230,12 +1309,11 @@ void crash_smp_send_stop(void) * This function can be called twice in panic path, but obviously * we execute this only once. * - * We use this same boolean to tell whether the IPI we send was a + * We use the CRASH_STOP bit to tell whether the IPI we send was a * stop or a "crash stop". */ - if (crash_stop) + if (test_and_set_bit(CRASH_STOP, &stop_in_progress)) return; - crash_stop = 1; smp_send_stop();
Broadcast pseudo-NMI IPIs to pause remote CPUs for a short period of time, and then reliably resume them when the local CPU exits critical sections that preclude the execution of remote CPUs. A typical example of such critical sections is BBM on kernel PTEs. HugeTLB Vmemmap Optimization (HVO) on arm64 was disabled by commit 060a2c92d1b6 ("arm64: mm: hugetlb: Disable HUGETLB_PAGE_OPTIMIZE_VMEMMAP") due to the folllowing reason: This is deemed UNPREDICTABLE by the Arm architecture without a break-before-make sequence (make the PTE invalid, TLBI, write the new valid PTE). However, such sequence is not possible since the vmemmap may be concurrently accessed by the kernel. Supporting BBM on kernel PTEs is one of the approaches that can make HVO theoretically safe on arm64. Note that it is still possible for the paused CPUs to perform speculative translations. Such translations would cause spurious kernel PFs, which should be properly handled by is_spurious_el1_translation_fault(). Signed-off-by: Yu Zhao <yuzhao@google.com> --- arch/arm64/include/asm/smp.h | 3 ++ arch/arm64/kernel/smp.c | 92 +++++++++++++++++++++++++++++++++--- 2 files changed, 88 insertions(+), 7 deletions(-)