diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig index 8943c13502c6..29073532f94c 100644 --- a/arch/x86/Kconfig +++ b/arch/x86/Kconfig @@ -248,6 +248,15 @@ config SPARSE_IRQ If you don't know what to do here, say Y. +config NUMA_MIGRATE_IRQ_DESC + bool "Move irq desc when changing irq smp_affinity" + depends on SPARSE_IRQ && SMP + default n + help + This enables moving irq_desc to cpu/node that irq will use handled. + + If you don't know what to do here, say N. + config X86_FIND_SMP_CONFIG def_bool y depends on X86_MPPARSE || X86_VOYAGER diff --git a/arch/x86/kernel/io_apic.c b/arch/x86/kernel/io_apic.c index a1a2e070f31a..bfe1245b1a3e 100644 --- a/arch/x86/kernel/io_apic.c +++ b/arch/x86/kernel/io_apic.c @@ -141,6 +141,9 @@ struct irq_cfg { unsigned move_cleanup_count; u8 vector; u8 move_in_progress : 1; +#ifdef CONFIG_NUMA_MIGRATE_IRQ_DESC + u8 move_desc_pending : 1; +#endif }; /* irq_cfg is indexed by the sum of all RTEs in all I/O APICs. */ @@ -223,6 +226,121 @@ void arch_init_chip_data(struct irq_desc *desc, int cpu) } } +#ifdef CONFIG_NUMA_MIGRATE_IRQ_DESC + +static void +init_copy_irq_2_pin(struct irq_cfg *old_cfg, struct irq_cfg *cfg, int cpu) +{ + struct irq_pin_list *old_entry, *head, *tail, *entry; + + cfg->irq_2_pin = NULL; + old_entry = old_cfg->irq_2_pin; + if (!old_entry) + return; + + entry = get_one_free_irq_2_pin(cpu); + if (!entry) + return; + + entry->apic = old_entry->apic; + entry->pin = old_entry->pin; + head = entry; + tail = entry; + old_entry = old_entry->next; + while (old_entry) { + entry = get_one_free_irq_2_pin(cpu); + if (!entry) { + entry = head; + while (entry) { + head = entry->next; + kfree(entry); + entry = head; + } + /* still use the old one */ + return; + } + entry->apic = old_entry->apic; + entry->pin = old_entry->pin; + tail->next = entry; + tail = entry; + old_entry = old_entry->next; + } + + tail->next = NULL; + cfg->irq_2_pin = head; +} + +static void free_irq_2_pin(struct irq_cfg *old_cfg, struct irq_cfg *cfg) +{ + struct irq_pin_list *entry, *next; + + if (old_cfg->irq_2_pin == cfg->irq_2_pin) + return; + + entry = old_cfg->irq_2_pin; + + while (entry) { + next = entry->next; + kfree(entry); + entry = next; + } + old_cfg->irq_2_pin = NULL; +} + +void arch_init_copy_chip_data(struct irq_desc *old_desc, + struct irq_desc *desc, int cpu) +{ + struct irq_cfg *cfg; + struct irq_cfg *old_cfg; + + cfg = get_one_free_irq_cfg(cpu); + + if (!cfg) + return; + + desc->chip_data = cfg; + + old_cfg = old_desc->chip_data; + + memcpy(cfg, old_cfg, sizeof(struct irq_cfg)); + + init_copy_irq_2_pin(old_cfg, cfg, cpu); +} + +static void free_irq_cfg(struct irq_cfg *old_cfg) +{ + kfree(old_cfg); +} + +void arch_free_chip_data(struct irq_desc *old_desc, struct irq_desc *desc) +{ + struct irq_cfg *old_cfg, *cfg; + + old_cfg = old_desc->chip_data; + cfg = desc->chip_data; + + if (old_cfg == cfg) + return; + + if (old_cfg) { + free_irq_2_pin(old_cfg, cfg); + free_irq_cfg(old_cfg); + old_desc->chip_data = NULL; + } +} + +static void set_extra_move_desc(struct irq_desc *desc, cpumask_t mask) +{ + struct irq_cfg *cfg = desc->chip_data; + + if (!cfg->move_in_progress) { + /* it means that domain is not changed */ + if (!cpus_intersects(desc->affinity, mask)) + cfg->move_desc_pending = 1; + } +} +#endif + #else static struct irq_cfg *irq_cfg(unsigned int irq) { @@ -231,9 +349,11 @@ static struct irq_cfg *irq_cfg(unsigned int irq) #endif +#ifndef CONFIG_NUMA_MIGRATE_IRQ_DESC static inline void set_extra_move_desc(struct irq_desc *desc, cpumask_t mask) { } +#endif struct io_apic { unsigned int index; @@ -2346,14 +2466,34 @@ static void irq_complete_move(struct irq_desc **descp) struct irq_cfg *cfg = desc->chip_data; unsigned vector, me; - if (likely(!cfg->move_in_progress)) + if (likely(!cfg->move_in_progress)) { +#ifdef CONFIG_NUMA_MIGRATE_IRQ_DESC + if (likely(!cfg->move_desc_pending)) + return; + + /* domain is not change, but affinity is changed */ + me = smp_processor_id(); + if (cpu_isset(me, desc->affinity)) { + *descp = desc = move_irq_desc(desc, me); + /* get the new one */ + cfg = desc->chip_data; + cfg->move_desc_pending = 0; + } +#endif return; + } vector = ~get_irq_regs()->orig_ax; me = smp_processor_id(); if ((vector == cfg->vector) && cpu_isset(me, cfg->domain)) { cpumask_t cleanup_mask; +#ifdef CONFIG_NUMA_MIGRATE_IRQ_DESC + *descp = desc = move_irq_desc(desc, me); + /* get the new one */ + cfg = desc->chip_data; +#endif + cpus_and(cleanup_mask, cfg->old_domain, cpu_online_map); cfg->move_cleanup_count = cpus_weight(cleanup_mask); send_IPI_mask(cleanup_mask, IRQ_MOVE_CLEANUP_VECTOR); diff --git a/include/linux/irq.h b/include/linux/irq.h index b5749db3e5a1..36a015746788 100644 --- a/include/linux/irq.h +++ b/include/linux/irq.h @@ -227,6 +227,16 @@ extern struct irq_desc *move_irq_desc(struct irq_desc *old_desc, int cpu); #endif +static inline struct irq_desc * +irq_remap_to_desc(unsigned int irq, struct irq_desc *desc) +{ +#ifdef CONFIG_NUMA_MIGRATE_IRQ_DESC + return irq_to_desc(irq); +#else + return desc; +#endif +} + /* * Migration helpers for obsolete names, they will go away: */ diff --git a/kernel/irq/Makefile b/kernel/irq/Makefile index 681c52dbfe22..4dd5b1edac98 100644 --- a/kernel/irq/Makefile +++ b/kernel/irq/Makefile @@ -3,3 +3,4 @@ obj-y := handle.o manage.o spurious.o resend.o chip.o devres.o obj-$(CONFIG_GENERIC_IRQ_PROBE) += autoprobe.o obj-$(CONFIG_PROC_FS) += proc.o obj-$(CONFIG_GENERIC_PENDING_IRQ) += migration.o +obj-$(CONFIG_NUMA_MIGRATE_IRQ_DESC) += numa_migrate.o diff --git a/kernel/irq/chip.c b/kernel/irq/chip.c index 8e4fce4a1b1f..de210f4b7a92 100644 --- a/kernel/irq/chip.c +++ b/kernel/irq/chip.c @@ -353,6 +353,7 @@ handle_level_irq(unsigned int irq, struct irq_desc *desc) spin_lock(&desc->lock); mask_ack_irq(desc, irq); + desc = irq_remap_to_desc(irq, desc); if (unlikely(desc->status & IRQ_INPROGRESS)) goto out_unlock; @@ -430,6 +431,7 @@ handle_fasteoi_irq(unsigned int irq, struct irq_desc *desc) desc->status &= ~IRQ_INPROGRESS; out: desc->chip->eoi(irq); + desc = irq_remap_to_desc(irq, desc); spin_unlock(&desc->lock); } @@ -466,12 +468,14 @@ handle_edge_irq(unsigned int irq, struct irq_desc *desc) !desc->action)) { desc->status |= (IRQ_PENDING | IRQ_MASKED); mask_ack_irq(desc, irq); + desc = irq_remap_to_desc(irq, desc); goto out_unlock; } kstat_incr_irqs_this_cpu(irq, desc); /* Start handling the irq */ desc->chip->ack(irq); + desc = irq_remap_to_desc(irq, desc); /* Mark the IRQ currently in progress.*/ desc->status |= IRQ_INPROGRESS; @@ -532,8 +536,10 @@ handle_percpu_irq(unsigned int irq, struct irq_desc *desc) if (!noirqdebug) note_interrupt(irq, desc, action_ret); - if (desc->chip->eoi) + if (desc->chip->eoi) { desc->chip->eoi(irq); + desc = irq_remap_to_desc(irq, desc); + } } void @@ -568,8 +574,10 @@ __set_irq_handler(unsigned int irq, irq_flow_handler_t handle, int is_chained, /* Uninstall? */ if (handle == handle_bad_irq) { - if (desc->chip != &no_irq_chip) + if (desc->chip != &no_irq_chip) { mask_ack_irq(desc, irq); + desc = irq_remap_to_desc(irq, desc); + } desc->status |= IRQ_DISABLED; desc->depth = 1; } diff --git a/kernel/irq/handle.c b/kernel/irq/handle.c index 8aa09547f5ef..f1a23069c20a 100644 --- a/kernel/irq/handle.c +++ b/kernel/irq/handle.c @@ -23,7 +23,7 @@ /* * lockdep: we want to handle all irq_desc locks as a single lock-class: */ -static struct lock_class_key irq_desc_lock_class; +struct lock_class_key irq_desc_lock_class; /** * handle_bad_irq - handle spurious and unhandled irqs @@ -73,7 +73,7 @@ static struct irq_desc irq_desc_init = { #endif }; -static void init_kstat_irqs(struct irq_desc *desc, int cpu, int nr) +void init_kstat_irqs(struct irq_desc *desc, int cpu, int nr) { unsigned long bytes; char *ptr; @@ -113,7 +113,7 @@ static void init_one_irq_desc(int irq, struct irq_desc *desc, int cpu) /* * Protect the sparse_irqs: */ -static DEFINE_SPINLOCK(sparse_irq_lock); +DEFINE_SPINLOCK(sparse_irq_lock); struct irq_desc *irq_desc_ptrs[NR_IRQS] __read_mostly; @@ -337,8 +337,11 @@ unsigned int __do_IRQ(unsigned int irq) /* * No locking required for CPU-local interrupts: */ - if (desc->chip->ack) + if (desc->chip->ack) { desc->chip->ack(irq); + /* get new one */ + desc = irq_remap_to_desc(irq, desc); + } if (likely(!(desc->status & IRQ_DISABLED))) { action_ret = handle_IRQ_event(irq, desc->action); if (!noirqdebug) @@ -349,8 +352,10 @@ unsigned int __do_IRQ(unsigned int irq) } spin_lock(&desc->lock); - if (desc->chip->ack) + if (desc->chip->ack) { desc->chip->ack(irq); + desc = irq_remap_to_desc(irq, desc); + } /* * REPLAY is when Linux resends an IRQ that was dropped earlier * WAITING is used by probe to mark irqs that are being tested diff --git a/kernel/irq/internals.h b/kernel/irq/internals.h index 64c1c7253dae..e6d0a43cc125 100644 --- a/kernel/irq/internals.h +++ b/kernel/irq/internals.h @@ -13,6 +13,11 @@ extern void compat_irq_chip_set_default_handler(struct irq_desc *desc); extern int __irq_set_trigger(struct irq_desc *desc, unsigned int irq, unsigned long flags); +extern struct lock_class_key irq_desc_lock_class; +extern void init_kstat_irqs(struct irq_desc *desc, int cpu, int nr); +extern spinlock_t sparse_irq_lock; +extern struct irq_desc *irq_desc_ptrs[NR_IRQS]; + #ifdef CONFIG_PROC_FS extern void register_irq_proc(unsigned int irq, struct irq_desc *desc); extern void register_handler_proc(unsigned int irq, struct irqaction *action); diff --git a/kernel/irq/numa_migrate.c b/kernel/irq/numa_migrate.c new file mode 100644 index 000000000000..0178e2296990 --- /dev/null +++ b/kernel/irq/numa_migrate.c @@ -0,0 +1,127 @@ +/* + * linux/kernel/irq/handle.c + * + * Copyright (C) 1992, 1998-2006 Linus Torvalds, Ingo Molnar + * Copyright (C) 2005-2006, Thomas Gleixner, Russell King + * + * This file contains the core interrupt handling code. + * + * Detailed information is available in Documentation/DocBook/genericirq + * + */ + +#include +#include +#include +#include +#include + +#include "internals.h" + +static void init_copy_kstat_irqs(struct irq_desc *old_desc, + struct irq_desc *desc, + int cpu, int nr) +{ + unsigned long bytes; + + init_kstat_irqs(desc, cpu, nr); + + if (desc->kstat_irqs != old_desc->kstat_irqs) { + /* Compute how many bytes we need per irq and allocate them */ + bytes = nr * sizeof(unsigned int); + + memcpy(desc->kstat_irqs, old_desc->kstat_irqs, bytes); + } +} + +static void free_kstat_irqs(struct irq_desc *old_desc, struct irq_desc *desc) +{ + if (old_desc->kstat_irqs == desc->kstat_irqs) + return; + + kfree(old_desc->kstat_irqs); + old_desc->kstat_irqs = NULL; +} + +static void init_copy_one_irq_desc(int irq, struct irq_desc *old_desc, + struct irq_desc *desc, int cpu) +{ + memcpy(desc, old_desc, sizeof(struct irq_desc)); + desc->cpu = cpu; + lockdep_set_class(&desc->lock, &irq_desc_lock_class); + init_copy_kstat_irqs(old_desc, desc, cpu, nr_cpu_ids); + arch_init_copy_chip_data(old_desc, desc, cpu); +} + +static void free_one_irq_desc(struct irq_desc *old_desc, struct irq_desc *desc) +{ + free_kstat_irqs(old_desc, desc); + arch_free_chip_data(old_desc, desc); +} + +static struct irq_desc *__real_move_irq_desc(struct irq_desc *old_desc, + int cpu) +{ + struct irq_desc *desc; + unsigned int irq; + unsigned long flags; + int node; + + irq = old_desc->irq; + + spin_lock_irqsave(&sparse_irq_lock, flags); + + /* We have to check it to avoid races with another CPU */ + desc = irq_desc_ptrs[irq]; + + if (desc && old_desc != desc) + goto out_unlock; + + node = cpu_to_node(cpu); + desc = kzalloc_node(sizeof(*desc), GFP_ATOMIC, node); + printk(KERN_DEBUG " move irq_desc for %d to cpu %d node %d\n", + irq, cpu, node); + if (!desc) { + printk(KERN_ERR "can not get new irq_desc for moving\n"); + /* still use old one */ + desc = old_desc; + goto out_unlock; + } + init_copy_one_irq_desc(irq, old_desc, desc, cpu); + + irq_desc_ptrs[irq] = desc; + + /* free the old one */ + free_one_irq_desc(old_desc, desc); + kfree(old_desc); + +out_unlock: + spin_unlock_irqrestore(&sparse_irq_lock, flags); + + return desc; +} + +struct irq_desc *move_irq_desc(struct irq_desc *desc, int cpu) +{ + int old_cpu; + int node, old_node; + + /* those all static, do move them */ + if (desc->irq < NR_IRQS_LEGACY) + return desc; + + old_cpu = desc->cpu; + printk(KERN_DEBUG + "try to move irq_desc from cpu %d to %d\n", old_cpu, cpu); + if (old_cpu != cpu) { + node = cpu_to_node(cpu); + old_node = cpu_to_node(old_cpu); + if (old_node != node) + desc = __real_move_irq_desc(desc, cpu); + else + desc->cpu = cpu; + } + + return desc; +} +