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>

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)))
 
 /*
@@ -342,6 +345,14 @@ void zero_fill_bio(struct bio *bio);
 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.

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)
 {

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)
 
 /*