Currently, the membarrier system call uses a single global mutex (`membarrier_ipi_mutex`) to serialize expedited commands. This causes significant contention on large systems when multiple threads invoke membarrier concurrently, even if they target different CPUs.
This contention becomes critical when combined with CFS bandwidth throttling/unthrottling, during which interrupts can be disabled for relatively long periods on target CPUs. If membarrier is waiting for a response from such a CPU, it holds the global mutex, blocking all other membarrier calls on the system. This cascade effect can lead to hard lockups when thousands of threads stall waiting for the mutex. Optimize `MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ` when a specific CPU is targeted by introducing per-CPU mutexes. Broadcast commands and commands without a specific CPU target continue to use the global mutex. This prevents the cascade lockup scenario. As measured by the stress test introduced in the subsequent patch, on an AMD Turin machine with 384 CPUs (2 NUMA nodes with SMT=2), this optimization yields 200x more throughput. Signed-off-by: Aniket Gattani <[email protected]> --- kernel/sched/membarrier.c | 48 +++++++++++++++++++++++++++------------ 1 file changed, 34 insertions(+), 14 deletions(-) diff --git a/kernel/sched/membarrier.c b/kernel/sched/membarrier.c index 623445603725..dc916e6541d2 100644 --- a/kernel/sched/membarrier.c +++ b/kernel/sched/membarrier.c @@ -165,7 +165,26 @@ | MEMBARRIER_CMD_GET_REGISTRATIONS) static DEFINE_MUTEX(membarrier_ipi_mutex); -#define SERIALIZE_IPI() guard(mutex)(&membarrier_ipi_mutex) +static DEFINE_PER_CPU(struct mutex, membarrier_cpu_mutexes); + +static inline struct mutex *membarrier_get_mutex(int cpu) +{ + if (cpu >= 0) + return &per_cpu(membarrier_cpu_mutexes, cpu); + return &membarrier_ipi_mutex; +} + +#define SERIALIZE_IPI(cpu_id) guard(mutex)(membarrier_get_mutex(cpu_id)) + +static int __init membarrier_init(void) +{ + int i; + + for_each_possible_cpu(i) + mutex_init(&per_cpu(membarrier_cpu_mutexes, i)); + return 0; +} +core_initcall(membarrier_init); static void ipi_mb(void *info) { @@ -264,7 +283,7 @@ static int membarrier_global_expedited(void) if (!zalloc_cpumask_var(&tmpmask, GFP_KERNEL)) return -ENOMEM; - SERIALIZE_IPI(); + SERIALIZE_IPI(-1); cpus_read_lock(); rcu_read_lock(); for_each_online_cpu(cpu) { @@ -358,14 +377,19 @@ static int membarrier_private_expedited(int flags, int cpu_id) if (cpu_id < 0 && !zalloc_cpumask_var(&tmpmask, GFP_KERNEL)) return -ENOMEM; - SERIALIZE_IPI(); + if (cpu_id >= 0 && (cpu_id >= nr_cpu_ids || !cpu_possible(cpu_id))) + return 0; + + SERIALIZE_IPI(cpu_id); + cpus_read_lock(); if (cpu_id >= 0) { struct task_struct *p; - if (cpu_id >= nr_cpu_ids || !cpu_online(cpu_id)) + if (!cpu_online(cpu_id)) goto out; + rcu_read_lock(); p = rcu_dereference(cpu_rq(cpu_id)->curr); if (!p || p->mm != mm) { @@ -373,6 +397,11 @@ static int membarrier_private_expedited(int flags, int cpu_id) goto out; } rcu_read_unlock(); + /* + * smp_call_function_single() will call ipi_func() if cpu_id + * is the calling CPU. + */ + smp_call_function_single(cpu_id, ipi_func, NULL, 1); } else { int cpu; @@ -385,15 +414,6 @@ static int membarrier_private_expedited(int flags, int cpu_id) __cpumask_set_cpu(cpu, tmpmask); } rcu_read_unlock(); - } - - if (cpu_id >= 0) { - /* - * smp_call_function_single() will call ipi_func() if cpu_id - * is the calling CPU. - */ - smp_call_function_single(cpu_id, ipi_func, NULL, 1); - } else { /* * For regular membarrier, we can save a few cycles by * skipping the current cpu -- we're about to do smp_mb() @@ -472,7 +492,7 @@ static int sync_runqueues_membarrier_state(struct mm_struct *mm) * between threads which are users of @mm has its membarrier state * updated. */ - SERIALIZE_IPI(); + SERIALIZE_IPI(-1); cpus_read_lock(); rcu_read_lock(); for_each_online_cpu(cpu) { -- 2.54.0.rc0.605.g598a273b03-goog
