block: add support for IO CPU affinity

This patch adds support for controlling the IO completion CPU of
either all requests on a queue, or on a per-request basis. We export
a sysfs variable (rq_affinity) which, if set, migrates completions
of requests to the CPU that originally submitted it. A bio helper
(bio_set_completion_cpu()) is also added, so that queuers can ask
for completion on that specific CPU.

In testing, this has been show to cut the system time by as much
as 20-40% on synthetic workloads where CPU affinity is desired.

This requires a little help from the architecture, so it'll only
work as designed for archs that are using the new generic smp
helper infrastructure.

Signed-off-by: Jens Axboe <jens.axboe@oracle.com>

9 files changed, 182 insertions, 60 deletions

diff --git a/block/blk-core.c b/block/blk-core.c
index 9c6f818d0c3..5484838f46e 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -110,7 +110,7 @@ void blk_rq_init(struct request_queue *q, struct request *rq)
 	memset(rq, 0, sizeof(*rq));
 
 	INIT_LIST_HEAD(&rq->queuelist);
-	INIT_LIST_HEAD(&rq->donelist);
+	rq->cpu = -1;
 	rq->q = q;
 	rq->sector = rq->hard_sector = (sector_t) -1;
 	INIT_HLIST_NODE(&rq->hash);
@@ -322,6 +322,21 @@ void blk_unplug(struct request_queue *q)
 }
 EXPORT_SYMBOL(blk_unplug);
 
+static void blk_invoke_request_fn(struct request_queue *q)
+{
+	/*
+	 * one level of recursion is ok and is much faster than kicking
+	 * the unplug handling
+	 */
+	if (!queue_flag_test_and_set(QUEUE_FLAG_REENTER, q)) {
+		q->request_fn(q);
+		queue_flag_clear(QUEUE_FLAG_REENTER, q);
+	} else {
+		queue_flag_set(QUEUE_FLAG_PLUGGED, q);
+		kblockd_schedule_work(q, &q->unplug_work);
+	}
+}
+
 /**
  * blk_start_queue - restart a previously stopped queue
  * @q:    The &struct request_queue in question
@@ -336,18 +351,7 @@ void blk_start_queue(struct request_queue *q)
 	WARN_ON(!irqs_disabled());
 
 	queue_flag_clear(QUEUE_FLAG_STOPPED, q);
-
-	/*
-	 * one level of recursion is ok and is much faster than kicking
-	 * the unplug handling
-	 */
-	if (!queue_flag_test_and_set(QUEUE_FLAG_REENTER, q)) {
-		q->request_fn(q);
-		queue_flag_clear(QUEUE_FLAG_REENTER, q);
-	} else {
-		blk_plug_device(q);
-		kblockd_schedule_work(q, &q->unplug_work);
-	}
+	blk_invoke_request_fn(q);
 }
 EXPORT_SYMBOL(blk_start_queue);
 
@@ -405,15 +409,8 @@ void __blk_run_queue(struct request_queue *q)
 	 * Only recurse once to avoid overrunning the stack, let the unplug
 	 * handling reinvoke the handler shortly if we already got there.
 	 */
-	if (!elv_queue_empty(q)) {
-		if (!queue_flag_test_and_set(QUEUE_FLAG_REENTER, q)) {
-			q->request_fn(q);
-			queue_flag_clear(QUEUE_FLAG_REENTER, q);
-		} else {
-			blk_plug_device(q);
-			kblockd_schedule_work(q, &q->unplug_work);
-		}
-	}
+	if (!elv_queue_empty(q))
+		blk_invoke_request_fn(q);
 }
 EXPORT_SYMBOL(__blk_run_queue);
 
@@ -1056,6 +1053,7 @@ EXPORT_SYMBOL(blk_put_request);
 
 void init_request_from_bio(struct request *req, struct bio *bio)
 {
+	req->cpu = bio->bi_comp_cpu;
 	req->cmd_type = REQ_TYPE_FS;
 
 	/*
@@ -1198,13 +1196,15 @@ get_rq:
 	init_request_from_bio(req, bio);
 
 	spin_lock_irq(q->queue_lock);
+	if (test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags) ||
+	    bio_flagged(bio, BIO_CPU_AFFINE))
+		req->cpu = blk_cpu_to_group(smp_processor_id());
 	if (elv_queue_empty(q))
 		blk_plug_device(q);
 	add_request(q, req);
 out:
 	if (sync)
 		__generic_unplug_device(q);
-
 	spin_unlock_irq(q->queue_lock);
 	return 0;
 
diff --git a/block/blk-settings.c b/block/blk-settings.c
index d70692badcd..a60e959a12c 100644
--- a/block/blk-settings.c
+++ b/block/blk-settings.c
@@ -443,7 +443,7 @@ void blk_queue_update_dma_alignment(struct request_queue *q, int mask)
 }
 EXPORT_SYMBOL(blk_queue_update_dma_alignment);
 
-static int __init blk_settings_init(void)
+int __init blk_settings_init(void)
 {
 	blk_max_low_pfn = max_low_pfn - 1;
 	blk_max_pfn = max_pfn - 1;
diff --git a/block/blk-softirq.c b/block/blk-softirq.c
index 9e1c43bff66..3a1af551191 100644
--- a/block/blk-softirq.c
+++ b/block/blk-softirq.c
@@ -13,6 +13,70 @@
 
 static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
 
+/*
+ * Softirq action handler - move entries to local list and loop over them
+ * while passing them to the queue registered handler.
+ */
+static void blk_done_softirq(struct softirq_action *h)
+{
+	struct list_head *cpu_list, local_list;
+
+	local_irq_disable();
+	cpu_list = &__get_cpu_var(blk_cpu_done);
+	list_replace_init(cpu_list, &local_list);
+	local_irq_enable();
+
+	while (!list_empty(&local_list)) {
+		struct request *rq;
+
+		rq = list_entry(local_list.next, struct request, csd.list);
+		list_del_init(&rq->csd.list);
+		rq->q->softirq_done_fn(rq);
+	}
+}
+
+#if defined(CONFIG_SMP) && defined(CONFIG_USE_GENERIC_SMP_HELPERS)
+static void trigger_softirq(void *data)
+{
+	struct request *rq = data;
+	unsigned long flags;
+	struct list_head *list;
+
+	local_irq_save(flags);
+	list = &__get_cpu_var(blk_cpu_done);
+	list_add_tail(&rq->csd.list, list);
+
+	if (list->next == &rq->csd.list)
+		raise_softirq_irqoff(BLOCK_SOFTIRQ);
+
+	local_irq_restore(flags);
+}
+
+/*
+ * Setup and invoke a run of 'trigger_softirq' on the given cpu.
+ */
+static int raise_blk_irq(int cpu, struct request *rq)
+{
+	if (cpu_online(cpu)) {
+		struct call_single_data *data = &rq->csd;
+
+		data->func = trigger_softirq;
+		data->info = rq;
+		data->flags = 0;
+
+		__smp_call_function_single(cpu, data);
+		return 0;
+	}
+
+	return 1;
+}
+#else /* CONFIG_SMP && CONFIG_USE_GENERIC_SMP_HELPERS */
+static int raise_blk_irq(int cpu, struct request *rq)
+{
+	return 1;
+}
+#endif
+
 static int __cpuinit blk_cpu_notify(struct notifier_block *self,
 				    unsigned long action, void *hcpu)
 {
@@ -33,33 +97,10 @@ static int __cpuinit blk_cpu_notify(struct notifier_block *self,
 	return NOTIFY_OK;
 }
 
-
-static struct notifier_block blk_cpu_notifier __cpuinitdata = {
+static struct notifier_block __cpuinitdata blk_cpu_notifier = {
 	.notifier_call	= blk_cpu_notify,
 };
 
-/*
- * splice the completion data to a local structure and hand off to
- * process_completion_queue() to complete the requests
- */
-static void blk_done_softirq(struct softirq_action *h)
-{
-	struct list_head *cpu_list, local_list;
-
-	local_irq_disable();
-	cpu_list = &__get_cpu_var(blk_cpu_done);
-	list_replace_init(cpu_list, &local_list);
-	local_irq_enable();
-
-	while (!list_empty(&local_list)) {
-		struct request *rq;
-
-		rq = list_entry(local_list.next, struct request, donelist);
-		list_del_init(&rq->donelist);
-		rq->q->softirq_done_fn(rq);
-	}
-}
-
 /**
  * blk_complete_request - end I/O on a request
  * @req:      the request being processed
@@ -71,25 +112,48 @@ static void blk_done_softirq(struct softirq_action *h)
  *     through a softirq handler. The user must have registered a completion
  *     callback through blk_queue_softirq_done().
  **/
-
 void blk_complete_request(struct request *req)
 {
-	struct list_head *cpu_list;
+	struct request_queue *q = req->q;
 	unsigned long flags;
+	int ccpu, cpu, group_cpu;
 
-	BUG_ON(!req->q->softirq_done_fn);
+	BUG_ON(!q->softirq_done_fn);
 
 	local_irq_save(flags);
+	cpu = smp_processor_id();
+	group_cpu = blk_cpu_to_group(cpu);
 
-	cpu_list = &__get_cpu_var(blk_cpu_done);
-	list_add_tail(&req->donelist, cpu_list);
-	raise_softirq_irqoff(BLOCK_SOFTIRQ);
+	/*
+	 * Select completion CPU
+	 */
+	if (test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags) && req->cpu != -1)
+		ccpu = req->cpu;
+	else
+		ccpu = cpu;
+
+	if (ccpu == cpu || ccpu == group_cpu) {
+		struct list_head *list;
+do_local:
+		list = &__get_cpu_var(blk_cpu_done);
+		list_add_tail(&req->csd.list, list);
+
+		/*
+		 * if the list only contains our just added request,
+		 * signal a raise of the softirq. If there are already
+		 * entries there, someone already raised the irq but it
+		 * hasn't run yet.
+		 */
+		if (list->next == &req->csd.list)
+			raise_softirq_irqoff(BLOCK_SOFTIRQ);
+	} else if (raise_blk_irq(ccpu, req))
+		goto do_local;
 
 	local_irq_restore(flags);
 }
 EXPORT_SYMBOL(blk_complete_request);
 
-int __init blk_softirq_init(void)
+__init int blk_softirq_init(void)
 {
 	int i;
 
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index b9a6ed16664..21e275d7eed 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -156,6 +156,30 @@ static ssize_t queue_nomerges_store(struct request_queue *q, const char *page,
 	return ret;
 }
 
+static ssize_t queue_rq_affinity_show(struct request_queue *q, char *page)
+{
+	unsigned int set = test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags);
+
+	return queue_var_show(set != 0, page);
+}
+
+static ssize_t
+queue_rq_affinity_store(struct request_queue *q, const char *page, size_t count)
+{
+	ssize_t ret = -EINVAL;
+#if defined(CONFIG_USE_GENERIC_SMP_HELPERS)
+	unsigned long val;
+
+	ret = queue_var_store(&val, page, count);
+	spin_lock_irq(q->queue_lock);
+	if (val)
+		queue_flag_set(QUEUE_FLAG_SAME_COMP, q);
+	else
+		queue_flag_clear(QUEUE_FLAG_SAME_COMP,  q);
+	spin_unlock_irq(q->queue_lock);
+#endif
+	return ret;
+}
 
 static struct queue_sysfs_entry queue_requests_entry = {
 	.attr = {.name = "nr_requests", .mode = S_IRUGO | S_IWUSR },
@@ -197,6 +221,12 @@ static struct queue_sysfs_entry queue_nomerges_entry = {
 	.store = queue_nomerges_store,
 };
 
+static struct queue_sysfs_entry queue_rq_affinity_entry = {
+	.attr = {.name = "rq_affinity", .mode = S_IRUGO | S_IWUSR },
+	.show = queue_rq_affinity_show,
+	.store = queue_rq_affinity_store,
+};
+
 static struct attribute *default_attrs[] = {
 	&queue_requests_entry.attr,
 	&queue_ra_entry.attr,
@@ -205,6 +235,7 @@ static struct attribute *default_attrs[] = {
 	&queue_iosched_entry.attr,
 	&queue_hw_sector_size_entry.attr,
 	&queue_nomerges_entry.attr,
+	&queue_rq_affinity_entry.attr,
 	NULL,
 };
 
diff --git a/block/blk.h b/block/blk.h
index c79f30e1df5..de74254cb91 100644
--- a/block/blk.h
+++ b/block/blk.h
@@ -59,4 +59,16 @@ static inline int queue_congestion_off_threshold(struct request_queue *q)
 
 #endif /* BLK_DEV_INTEGRITY */
 
+static inline int blk_cpu_to_group(int cpu)
+{
+#ifdef CONFIG_SCHED_MC
+	cpumask_t mask = cpu_coregroup_map(cpu);
+	return first_cpu(mask);
+#elif defined(CONFIG_SCHED_SMT)
+	return first_cpu(per_cpu(cpu_sibling_map, cpu));
+#else
+	return cpu;
+#endif
+}
+
 #endif
diff --git a/fs/bio.c b/fs/bio.c
index bee4deca774..6a637b5c24b 100644
--- a/fs/bio.c
+++ b/fs/bio.c
@@ -111,6 +111,7 @@ void bio_init(struct bio *bio)
 {
 	memset(bio, 0, sizeof(*bio));
 	bio->bi_flags = 1 << BIO_UPTODATE;
+	bio->bi_comp_cpu = -1;
 	atomic_set(&bio->bi_cnt, 1);
 }
 
diff --git a/include/linux/bio.h b/include/linux/bio.h
index 2c0c09034fd..13aba20edb2 100644
--- a/include/linux/bio.h
+++ b/include/linux/bio.h
@@ -81,6 +81,8 @@ struct bio {
 
 	unsigned int		bi_max_vecs;	/* max bvl_vecs we can hold */
 
+	unsigned int		bi_comp_cpu;	/* completion CPU */
+
 	struct bio_vec		*bi_io_vec;	/* the actual vec list */
 
 	bio_end_io_t		*bi_end_io;
@@ -105,6 +107,7 @@ struct bio {
 #define BIO_BOUNCED	5	/* bio is a bounce bio */
 #define BIO_USER_MAPPED 6	/* contains user pages */
 #define BIO_EOPNOTSUPP	7	/* not supported */
+#define BIO_CPU_AFFINE	8	/* complete bio on same CPU as submitted */
 #define bio_flagged(bio, flag)	((bio)->bi_flags & (1 << (flag)))
 
 /*
@@ -343,6 +346,14 @@ extern struct bio_vec *bvec_alloc_bs(gfp_t, int, unsigned long *, struct bio_set
 extern unsigned int bvec_nr_vecs(unsigned short idx);
 
 /*
+ * Allow queuer to specify a completion CPU for this bio
+ */
+static inline void bio_set_completion_cpu(struct bio *bio, unsigned int cpu)
+{
+	bio->bi_comp_cpu = cpu;
+}
+
+/*
  * bio_set is used to allow other portions of the IO system to
  * allocate their own private memory pools for bio and iovec structures.
  * These memory pools in turn all allocate from the bio_slab
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 10aa46c8f17..93204bf7b29 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -17,6 +17,7 @@
 #include <linux/module.h>
 #include <linux/stringify.h>
 #include <linux/bsg.h>
+#include <linux/smp.h>
 
 #include <asm/scatterlist.h>
 
@@ -139,7 +140,8 @@ enum rq_flag_bits {
  */
 struct request {
 	struct list_head queuelist;
-	struct list_head donelist;
+	struct call_single_data csd;
+	int cpu;
 
 	struct request_queue *q;
 
@@ -420,6 +422,7 @@ struct request_queue
 #define QUEUE_FLAG_ELVSWITCH	8	/* don't use elevator, just do FIFO */
 #define QUEUE_FLAG_BIDI		9	/* queue supports bidi requests */
 #define QUEUE_FLAG_NOMERGES    10	/* disable merge attempts */
+#define QUEUE_FLAG_SAME_COMP   11	/* force complete on same CPU */
 
 static inline int queue_is_locked(struct request_queue *q)
 {
diff --git a/include/linux/elevator.h b/include/linux/elevator.h
index 639624b55fb..bb791c311a5 100644
--- a/include/linux/elevator.h
+++ b/include/linux/elevator.h
@@ -173,15 +173,15 @@ enum {
 #define rb_entry_rq(node)	rb_entry((node), struct request, rb_node)
 
 /*
- * Hack to reuse the donelist list_head as the fifo time holder while
+ * Hack to reuse the csd.list list_head as the fifo time holder while
  * the request is in the io scheduler. Saves an unsigned long in rq.
  */
-#define rq_fifo_time(rq)	((unsigned long) (rq)->donelist.next)
-#define rq_set_fifo_time(rq,exp)	((rq)->donelist.next = (void *) (exp))
+#define rq_fifo_time(rq)	((unsigned long) (rq)->csd.list.next)
+#define rq_set_fifo_time(rq,exp)	((rq)->csd.list.next = (void *) (exp))
 #define rq_entry_fifo(ptr)	list_entry((ptr), struct request, queuelist)
 #define rq_fifo_clear(rq)	do {		\
 	list_del_init(&(rq)->queuelist);	\
-	INIT_LIST_HEAD(&(rq)->donelist);	\
+	INIT_LIST_HEAD(&(rq)->csd.list);	\
 	} while (0)
 
 /*