Use percpu inflight counters to avoid cache line bouncing and improve
performance.
Signed-off-by: Mikulas Patocka <mpatocka at redhat.com>
---
drivers/md/dm-core.h | 5 +++++
drivers/md/dm.c | 50 ++++++++++++++++++++++++++++++++++----------------
2 files changed, 39 insertions(+), 16 deletions(-)
Index: linux-dm/drivers/md/dm-core.h
===================================================================
--- linux-dm.orig/drivers/md/dm-core.h 2018-11-15 22:06:37.000000000 +0100
+++ linux-dm/drivers/md/dm-core.h 2018-11-15 22:06:37.000000000 +0100
@@ -24,6 +24,10 @@ struct dm_kobject_holder {
struct completion completion;
};
+struct dm_percpu {
+ unsigned inflight[2];
+};
+
/*
* DM core internal structure that used directly by dm.c and dm-rq.c
* DM targets must _not_ deference a mapped_device to directly access its members!
@@ -63,6 +67,7 @@ struct mapped_device {
/*
* A list of ios that arrived while we were suspended.
*/
+ struct dm_percpu __percpu *counters;
struct work_struct work;
wait_queue_head_t wait;
spinlock_t deferred_lock;
Index: linux-dm/drivers/md/dm.c
===================================================================
--- linux-dm.orig/drivers/md/dm.c 2018-11-15 22:06:37.000000000 +0100
+++ linux-dm/drivers/md/dm.c 2018-11-15 22:09:31.000000000 +0100
@@ -648,19 +648,32 @@ static void free_tio(struct dm_target_io
int md_in_flight(struct mapped_device *md)
{
- return atomic_read(&dm_disk(md)->part0.in_flight[READ]) +
- atomic_read(&dm_disk(md)->part0.in_flight[WRITE]);
+ int cpu;
+ unsigned sum = 0;
+ for_each_possible_cpu(cpu) {
+ struct dm_percpu *p = per_cpu_ptr(md->counters, cpu);
+ sum += p->inflight[READ] + p->inflight[WRITE];
+ }
+ return (int)sum;
}
static void start_io_acct(struct dm_io *io)
{
struct mapped_device *md = io->md;
struct bio *bio = io->orig_bio;
+ struct hd_struct *part;
+ int sgrp, cpu;
io->start_time = jiffies;
- generic_start_io_acct(md->queue, bio_op(bio), bio_sectors(bio),
- &dm_disk(md)->part0);
+ part = &dm_disk(md)->part0;
+ sgrp = op_stat_group(bio_op(bio));
+ cpu = part_stat_lock();
+ __part_stat_add(cpu, part, ios[sgrp], 1);
+ __part_stat_add(cpu, part, sectors[sgrp], bio_sectors(bio));
+ part_stat_unlock();
+
+ this_cpu_inc(md->counters->inflight[bio_data_dir(bio)]);
if (unlikely(dm_stats_used(&md->stats)))
dm_stats_account_io(&md->stats, bio_data_dir(bio),
@@ -673,25 +686,24 @@ static void end_io_acct(struct dm_io *io
struct mapped_device *md = io->md;
struct bio *bio = io->orig_bio;
unsigned long duration = jiffies - io->start_time;
+ struct hd_struct *part;
+ int sgrp, cpu;
if (unlikely(dm_stats_used(&md->stats)))
dm_stats_account_io(&md->stats, bio_data_dir(bio),
bio->bi_iter.bi_sector, bio_sectors(bio),
true, duration, &io->stats_aux);
- /*
- * make sure that atomic_dec in generic_end_io_acct is not reordered
- * with previous writes
- */
- smp_mb__before_atomic();
- generic_end_io_acct(md->queue, bio_op(bio), &dm_disk(md)->part0,
- io->start_time);
- /*
- * generic_end_io_acct does atomic_dec, this barrier makes sure that
- * atomic_dec is not reordered with waitqueue_active
- */
- smp_mb__after_atomic();
+ part = &dm_disk(md)->part0;
+ sgrp = op_stat_group(bio_op(bio));
+ cpu = part_stat_lock();
+ __part_stat_add(cpu, part, nsecs[sgrp], jiffies_to_nsecs(duration));
+ part_stat_unlock();
+
+ smp_wmb();
+ this_cpu_dec(md->counters->inflight[bio_data_dir(bio)]);
+ smp_mb();
/* nudge anyone waiting on suspend queue */
if (unlikely(waitqueue_active(&md->wait))) {
if (!md_in_flight(md))
@@ -1822,6 +1834,8 @@ static void cleanup_mapped_device(struct
if (md->queue)
blk_cleanup_queue(md->queue);
+ free_percpu(md->counters);
+
cleanup_srcu_struct(&md->io_barrier);
if (md->bdev) {
@@ -1892,6 +1906,10 @@ static struct mapped_device *alloc_dev(i
if (!md->disk)
goto bad;
+ md->counters = alloc_percpu(struct dm_percpu);
+ if (!md->counters)
+ goto bad;
+
init_waitqueue_head(&md->wait);
INIT_WORK(&md->work, dm_wq_work);
init_waitqueue_head(&md->eventq);