Remove preempt_enable/disable calls around sched_clock()
Subsystems (2)
PATCH 1/2 Remove preempt_enable/disable calls around sched_clock()
calls in the block layer. This was a temporary fix added.
--- include/linux/blkdev.h | 4 ---- 1 files changed, 0 insertions(+), 4 deletions(-)
include/linux/blkdev.h +0 -4
@@ -1218,16 +1218,12 @@ int kblockd_schedule_work(struct request_queue *q, struct work_struct *work);*/ static inline void set_start_time_ns (struct request *req) { -preempt_disable (); req->start_time_ns =sched_clock (); -preempt_enable (); }static inline void set_io_start_time_ns (struct request *req) { -preempt_disable (); req->io_start_time_ns =sched_clock (); -preempt_enable (); }static inline uint64_t rq_start_time_ns (struct request *req)
PATCH 2/2 Use ktime_get() instead of sched_clock() for blkio cgroup stats.
This will take care of the pre-emptive kernel issue and the unbounded TSC drift problem. We will lose resolution though in some cases.
--- block/blk-cgroup.c | 22 +++++++++++----------- include/linux/blkdev.h | 4 ++-- 2 files changed, 13 insertions(+), 13 deletions(-)
block/blk-cgroup.c +11 -11
@@ -135,19 +135,19 @@ static void blkio_set_start_group_wait_time(struct blkio_group *blkg,return ;if (blkg == curr_blkg)return ; - blkg->stats.start_group_wait_time =sched_clock (); + blkg->stats.start_group_wait_time =ktime_to_ns (ktime_get ());blkio_mark_blkg_waiting (&blkg->stats); }/* This should be called with the blkg->stats_lock held. */ static void blkio_update_group_wait_time (struct blkio_group_stats *stats) { -unsigned long long now; +u64 now;if (!blkio_blkg_waiting (stats))return ; - now =sched_clock (); + now =ktime_to_ns (ktime_get ());if (time_after64 (now, stats->start_group_wait_time)) stats->group_wait_time += now - stats->start_group_wait_time;blkio_clear_blkg_waiting (stats);
@@ -156,12 +156,12 @@ static void blkio_update_group_wait_time(struct blkio_group_stats *stats)/* This should be called with the blkg->stats_lock held. */ static void blkio_end_empty_time (struct blkio_group_stats *stats) { -unsigned long long now; +u64 now;if (!blkio_blkg_empty (stats))return ; - now =sched_clock (); + now =ktime_to_ns (ktime_get ());if (time_after64 (now, stats->start_empty_time)) stats->empty_time += now - stats->start_empty_time;blkio_clear_blkg_empty (stats);
@@ -173,7 +173,7 @@ void blkiocg_update_set_idle_time_stats(struct blkio_group *blkg)spin_lock_irqsave (&blkg->stats_lock, flags);BUG_ON (blkio_blkg_idling (&blkg->stats)); - blkg->stats.start_idle_time =sched_clock (); + blkg->stats.start_idle_time =ktime_to_ns (ktime_get ());blkio_mark_blkg_idling (&blkg->stats);spin_unlock_irqrestore (&blkg->stats_lock, flags); }
@@ -182,13 +182,13 @@ EXPORT_SYMBOL_GPL(blkiocg_update_set_idle_time_stats);void blkiocg_update_idle_time_stats (struct blkio_group *blkg) {unsigned long flags; -unsigned long long now; +u64 now;struct blkio_group_stats *stats;spin_lock_irqsave (&blkg->stats_lock, flags); stats = &blkg->stats;if (blkio_blkg_idling (stats)) { - now =sched_clock (); + now =ktime_to_ns (ktime_get ());if (time_after64 (now, stats->start_idle_time)) stats->idle_time += now - stats->start_idle_time;blkio_clear_blkg_idling (stats);
@@ -237,7 +237,7 @@ void blkiocg_set_start_empty_time(struct blkio_group *blkg)return ; } - stats->start_empty_time =sched_clock (); + stats->start_empty_time =ktime_to_ns (ktime_get ());blkio_mark_blkg_empty (stats);spin_unlock_irqrestore (&blkg->stats_lock, flags); }
@@ -314,7 +314,7 @@ void blkiocg_update_completion_stats(struct blkio_group *blkg, {struct blkio_group_stats *stats;unsigned long flags; -unsigned long long now =sched_clock (); +u64 now =ktime_to_ns (ktime_get ());spin_lock_irqsave (&blkg->stats_lock, flags); stats = &blkg->stats;
@@ -464,7 +464,7 @@ blkiocg_reset_stats(struct cgroup *cgroup, struct cftype *cftype, u64 val)int i;#ifdef CONFIG_DEBUG_BLK_CGROUP bool idling, waiting, empty; -unsigned long long now =sched_clock (); +u64 now =ktime_to_ns (ktime_get ());#endif blkcg =cgroup_to_blkio_cgroup (cgroup);
include/linux/blkdev.h +2 -2
@@ -1218,12 +1218,12 @@ int kblockd_schedule_work(struct request_queue *q, struct work_struct *work);*/ static inline void set_start_time_ns (struct request *req) { - req->start_time_ns =sched_clock (); + req->start_time_ns =ktime_to_ns (ktime_get ()); }static inline void set_io_start_time_ns (struct request *req) { - req->io_start_time_ns =sched_clock (); + req->io_start_time_ns =ktime_to_ns (ktime_get ()); }static inline uint64_t rq_start_time_ns (struct request *req)
Re: [PATCH 2/2] Use ktime_get() instead of sched_clock() for blkio cgroup stats.
block/blk-cgroup.c static void blkio_set_start_group_wait_time(struct blkio_group *blkg, · line 138
Quoted 5 lines · Divyesh Shah
This will take care of the pre-emptive kernel issue and the unbounded TSC drift problem. We will lose resolution though in some cases.
block/blk-cgroup.c +1
-1- blkg->stats.start_group_wait_time =sched_clock (); + blkg->stats.start_group_wait_time =ktime_to_ns (ktime_get ());
Ugh!
ktime_get() can have insanely high overhead. Peter has added local_clock(), if then you should use that and apply checks to make sure the result isnt negative.
Ingo
Re: [PATCH 2/2] Use ktime_get() instead of sched_clock() for blkio cgroup stats.
Quoted 6 lines · Ingo Molnar
Quoted 2 lines · Divyesh Shah
This will take care of the pre-emptive kernel issue and the unbounded TSC drift problem. We will lose resolution though in some cases.
Quoted 8 lines
Quoted 2 lines
-
blkg->stats.start_group_wait_time=sched_clock();
+blkg->stats.start_group_wait_time= ktime_to_ns(ktime_get());Ugh!
ktime_get()can have insanely high overhead. Peter has addedlocal_clock(), if then you should use that and apply checks to make sure the result isnt negative.
Ok. The negative checks are already in the code.
Quoted 3 lines
Ingo
Quoted 1 line · Divyesh Shah
calls in the block layer. This was a temporary fix added.
Not good, this commit reintroduces those ugly warnings, if someone happens to bisect straight into this commit.
Ingo
Quoted 7 lines · Ingo Molnar
Quoted 1 line · Divyesh Shah
calls in the block layer. This was a temporary fix added.
Not good, this commit reintroduces those ugly warnings, if someone happens to bisect straight into this commit.
You're right. In the next version, I'll merge both patches into one so this cannot happen.
Quoted 3 lines
Ingo
Quoted 11 lines · Divyesh Shah
Quoted 7 lines · Ingo Molnar
Quoted 1 line · Divyesh Shah
calls in the block layer. This was a temporary fix added.
Not good, this commit reintroduces those ugly warnings, if someone happens to bisect straight into this commit.
You're right. In the next version, I'll merge both patches into one so this cannot happen.
Yeah - or if you really want to make a point of doing the two things in two separate patches you can first fix the API, _then_ remove the preemption enable/disable.
Thanks,
Ingo