Cache issue side time querying
Subsystems (5)
Jens Axboe
maintainerv4 cover Cache issue side time querying
Hi,
When I run my peak testing to see if we've regressed, my test script always does:
echo 0 > /sys/block/$DEV/queue/iostats
echo 2 > /sys/block/$DEV/queue/nomerges
for each device being used. It's unfortunate that we need to disable iostats, but without doing that, I lose about 12% performance. The main reason for that is the time querying we need to do, when iostats are enabled. As it turns out, lots of other block code is quite trigger happy with querying time as well. We do have some nice batching in place which helps ammortize that, but it's not perfect.
This trivial patchset simply caches the current time in struct blk_plug, on the premise that any issue side time querying can get adequate granularity through that. Nobody really needs nsec granularity on the timestamp.
Results in patch 3, but tldr is a more than 9% improvement (108M -> 118M IOPS) for my test case, which doesn't even enable most of the costly block layer items that you'd typically find in a distro and which would further increase the number of issue side time calls. This brings iostats enabled _almost_ to the level of turning it off.
Can also be found in my block-issue-ts branch:
https://git.kernel.dk/cgit/linux/log/?h=block-issue-ts
block/bfq-cgroup.c | 14 +++--- block/bfq-iosched.c | 28 +++++------ block/blk-cgroup.c | 2 +- block/blk-core.c | 33 +++++++------ block/blk-flush.c | 2 +- block/blk-iocost.c | 8 ++-- block/blk-iolatency.c | 6 +-- block/blk-mq.c | 18 ++++---- block/blk-throttle.c | 6 +-- block/blk-wbt.c | 5 +- drivers/md/raid1-10.c | 2 +- include/linux/blk_types.h | 42 ----------------- include/linux/blkdev.h | 97 ++++++++++++++++++++++++++++++++++++--- include/linux/sched.h | 2 +- kernel/sched/core.c | 4 +- 15 files changed, 160 insertions(+), 109 deletions(-)
Changes since v3:
- Include a ktime_get() variant, and use that to convert the remaining ktime user (BFQ)
- Remove RFC label, I think this is ready to go
- Rebase on 6.8-rc1
Signature · 1 line
Jens Axboe
Jens Axboe
maintainerPATCH v4 1/6 block: move cgroup time handling code into blkdev.h
In preparation for moving time keeping into blkdev.h, move the cgroup related code for timestamps in here too. This will help avoid a circular dependency.
Leave struct bio_issue in blk_types.h as it's a proper time definition.
--- include/linux/blk_types.h | 42 -------------------------------------- include/linux/blkdev.h | 43 +++++++++++++++++++++++++++++++++++++++ 2 files changed, 43 insertions(+), 42 deletions(-)
include/linux/blk_types.h +0 -42
@@ -206,52 +206,10 @@ static inline bool blk_path_error(blk_status_t error)return true; } -/* -* From most significant bit: -* 1 bit: reserved for other usage, see below -* 12 bits: original size of bio -* 51 bits: issue time of bio -*/ -#define BIO_ISSUE_RES_BITS 1 -#define BIO_ISSUE_SIZE_BITS 12 -#define BIO_ISSUE_RES_SHIFT (64 - BIO_ISSUE_RES_BITS) -#define BIO_ISSUE_SIZE_SHIFT (BIO_ISSUE_RES_SHIFT - BIO_ISSUE_SIZE_BITS) -#define BIO_ISSUE_TIME_MASK ((1ULL << BIO_ISSUE_SIZE_SHIFT) - 1) -#define BIO_ISSUE_SIZE_MASK \ - (((1ULL << BIO_ISSUE_SIZE_BITS) - 1) << BIO_ISSUE_SIZE_SHIFT) -#define BIO_ISSUE_RES_MASK (~((1ULL << BIO_ISSUE_RES_SHIFT) - 1)) - -/* Reserved bit for blk-throtl */ -#define BIO_ISSUE_THROTL_SKIP_LATENCY (1ULL << 63) -struct bio_issue {u64 value; }; -static inline u64 __bio_issue_time (u64 time) -{ -return time &BIO_ISSUE_TIME_MASK ; -} - -static inline u64 bio_issue_time (struct bio_issue *issue) -{ -return __bio_issue_time (issue->value); -} - -static inline sector_t bio_issue_size (struct bio_issue *issue) -{ -return ((issue->value &BIO_ISSUE_SIZE_MASK ) >>BIO_ISSUE_SIZE_SHIFT ); -} - -static inline void bio_issue_init (struct bio_issue *issue, -sector_t size) -{ - size &= (1ULL <<BIO_ISSUE_SIZE_BITS ) -1 ; - issue->value = ((issue->value &BIO_ISSUE_RES_MASK ) | - (ktime_get_ns () &BIO_ISSUE_TIME_MASK ) | - ((u64 )size <<BIO_ISSUE_SIZE_SHIFT )); -} -typedef __u32 __bitwise blk_opf_t ;typedef unsigned int blk_qc_t ;
include/linux/blkdev.h +43 -0
49 lines not on this page · raw message
Signature · 1 line
2.43.0
Jens Axboe
maintainerPATCH v4 2/6 block: add blk_time_get_ns() and blk_time_get() helpers
In reply to Jens AxboeConvert any user of ktime_get_ns() to use blk_time_get_ns(), and ktime_get() to blk_time_get(), so we have a unified API for querying the current time in nanoseconds or as ktime.
No functional changes intended, this patch just wraps ktime_get_ns() and ktime_get() with a block helper.
--- block/bfq-cgroup.c | 14 +++++++------- block/bfq-iosched.c | 28 ++++++++++++++-------------- block/blk-cgroup.c | 2 +- block/blk-flush.c | 2 +- block/blk-iocost.c | 8 ++++---- block/blk-iolatency.c | 6 +++--- block/blk-mq.c | 16 ++++++++-------- block/blk-throttle.c | 6 +++--- block/blk-wbt.c | 5 ++--- include/linux/blkdev.h | 13 ++++++++++++- 10 files changed, 55 insertions(+), 45 deletions(-)
block/bfq-cgroup.c +7 -7
@@ -127,7 +127,7 @@ static void bfqg_stats_update_group_wait_time(struct bfqg_stats *stats)if (!bfqg_stats_waiting (stats))return ; - now =ktime_get_ns (); + now =blk_time_get_ns ();if (now > stats->start_group_wait_time)bfq_stat_add (&stats->group_wait_time, now - stats->start_group_wait_time);
@@ -144,7 +144,7 @@ static void bfqg_stats_set_start_group_wait_time(struct bfq_group *bfqg,return ;if (bfqg == curr_bfqg)return ; - stats->start_group_wait_time =ktime_get_ns (); + stats->start_group_wait_time =blk_time_get_ns ();bfqg_stats_mark_waiting (stats); }
@@ -156,7 +156,7 @@ static void bfqg_stats_end_empty_time(struct bfqg_stats *stats)if (!bfqg_stats_empty (stats))return ; - now =ktime_get_ns (); + now =blk_time_get_ns ();if (now > stats->start_empty_time)bfq_stat_add (&stats->empty_time, now - stats->start_empty_time);
@@ -183,7 +183,7 @@ void bfqg_stats_set_start_empty_time(struct bfq_group *bfqg)if (bfqg_stats_empty (stats))return ; - stats->start_empty_time =ktime_get_ns (); + stats->start_empty_time =blk_time_get_ns ();bfqg_stats_mark_empty (stats); }
@@ -192,7 +192,7 @@ void bfqg_stats_update_idle_time(struct bfq_group *bfqg)struct bfqg_stats *stats = &bfqg->stats;if (bfqg_stats_idling (stats)) { -u64 now =ktime_get_ns (); +u64 now =blk_time_get_ns ();if (now > stats->start_idle_time) bfq_stat_add(&stats->idle_time,
@@ -205,7 +205,7 @@ void bfqg_stats_set_start_idle_time(struct bfq_group *bfqg) {struct bfqg_stats *stats = &bfqg->stats; - stats->start_idle_time =ktime_get_ns (); + stats->start_idle_time =blk_time_get_ns ();bfqg_stats_mark_idling (stats); }
@@ -242,7 +242,7 @@ void bfqg_stats_update_completion(struct bfq_group *bfqg, u64 start_time_ns, u64 io_start_time_ns, blk_opf_t opf) {struct bfqg_stats *stats = &bfqg->stats; -u64 now =ktime_get_ns (); +u64 now =blk_time_get_ns ();if (now > io_start_time_ns) blkg_rwstat_add(&stats->service_time, opf,
block/bfq-iosched.c +14 -14
@@ -1005,7 +1005,7 @@ static struct request *bfq_check_fifo(struct bfq_queue *bfqq, rq =rq_entry_fifo (bfqq->fifo.next); -if (rq == last ||ktime_get_ns () < rq->fifo_time) +if (rq == last ||blk_time_get_ns () < rq->fifo_time)return NULL ;bfq_log_bfqq (bfqq->bfqd, bfqq,"check_fifo: returned %p" , rq);
@@ -1829,7 +1829,7 @@ static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,* bfq_bfqq_update_budg_for_activation for * details on the usage of the next variable. */ - arrived_in_time =ktime_get_ns () <= + arrived_in_time =blk_time_get_ns () <= bfqq->ttime.last_end_request + bfqd->bfq_slice_idle *3 ;unsigned int act_idx =bfq_actuator_index (bfqd, rq->bio);
@@ -2208,7 +2208,7 @@ static void bfq_add_request(struct request *rq)struct request *next_rq, *prev;unsigned int old_wr_coeff = bfqq->wr_coeff;bool interactive = false; -u64 now_ns =ktime_get_ns (); +u64 now_ns =blk_time_get_ns ();bfq_log_bfqq (bfqd, bfqq,"add_request %d" ,rq_is_sync (rq)); bfqq->queued[rq_is_sync (rq)]++;
@@ -2262,7 +2262,7 @@ static void bfq_add_request(struct request *rq) bfqd->rqs_injected && bfqd->tot_rq_in_driver > 0)) && time_is_before_eq_jiffies(bfqq->decrease_time_jif + msecs_to_jiffies(10))) { - bfqd->last_empty_occupied_ns =ktime_get_ns (); + bfqd->last_empty_occupied_ns =blk_time_get_ns ();/* * Start the state machine for measuring the * total service time of rq: setting
@@ -3294,7 +3294,7 @@ static void bfq_set_budget_timeout(struct bfq_data *bfqd,else timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight; - bfqd->last_budget_start =ktime_get (); + bfqd->last_budget_start =blk_time_get (); bfqq->budget_timeout = jiffies + bfqd->bfq_timeout * timeout_coeff;
@@ -3394,7 +3394,7 @@ static void bfq_arm_slice_timer(struct bfq_data *bfqd)else if (bfqq->wr_coeff >1 ) sl =max_t (u32, sl,20ULL *NSEC_PER_MSEC ); - bfqd->last_idling_start =ktime_get (); + bfqd->last_idling_start =blk_time_get (); bfqd->last_idling_start_jiffies = jiffies; hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
@@ -3433,7 +3433,7 @@ static void bfq_reset_rate_computation(struct bfq_data *bfqd, struct request *rq) {if (rq !=NULL ) {/* new rq dispatch now, reset accordingly */ - bfqd->last_dispatch = bfqd->first_dispatch =ktime_get_ns (); + bfqd->last_dispatch = bfqd->first_dispatch =blk_time_get_ns (); bfqd->peak_rate_samples =1 ; bfqd->sequential_samples =0 ; bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
@@ -3590,7 +3590,7 @@ static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)*/ static void bfq_update_peak_rate (struct bfq_data *bfqd,struct request *rq) { -u64 now_ns =ktime_get_ns (); +u64 now_ns =blk_time_get_ns ();if (bfqd->peak_rate_samples ==0 ) {/* first dispatch */ bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
@@ -4162,7 +4162,7 @@ static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,if (compensate) delta_ktime = bfqd->last_idling_start;else - delta_ktime =ktime_get (); + delta_ktime =blk_time_get (); delta_ktime =ktime_sub (delta_ktime, bfqd->last_budget_start); delta_usecs =ktime_to_us (delta_ktime);
@@ -5591,7 +5591,7 @@ static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq, struct bfq_io_cq *bic, pid_t pid, int is_sync, unsigned int act_idx) { - u64 now_ns = ktime_get_ns(); + u64 now_ns = blk_time_get_ns(); bfqq->actuator_idx = act_idx;RB_CLEAR_NODE (&bfqq->entity.rb_node);
@@ -5903,7 +5903,7 @@ static void bfq_update_io_thinktime(struct bfq_data *bfqd,*/ if (bfqq->dispatched ||bfq_bfqq_busy (bfqq))return ; - elapsed =ktime_get_ns () - bfqq->ttime.last_end_request; + elapsed =blk_time_get_ns () - bfqq->ttime.last_end_request; elapsed =min_t (u64, elapsed,2ULL * bfqd->bfq_slice_idle); ttime->ttime_samples = (7 *ttime->ttime_samples +256 ) /8 ;
@@ -6194,7 +6194,7 @@ static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)bfq_add_request (rq); idle_timer_disabled = waiting && !bfq_bfqq_wait_request (bfqq); - rq->fifo_time =ktime_get_ns () + bfqd->bfq_fifo_expire[rq_is_sync (rq)]; + rq->fifo_time =blk_time_get_ns () + bfqd->bfq_fifo_expire[rq_is_sync (rq)];list_add_tail (&rq->queuelist, &bfqq->fifo);bfq_rq_enqueued (bfqd, bfqq, rq);
@@ -6370,7 +6370,7 @@ static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)bfq_weights_tree_remove (bfqq); } - now_ns =ktime_get_ns (); + now_ns =blk_time_get_ns (); bfqq->ttime.last_end_request = now_ns;
@@ -6585,7 +6585,7 @@ static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)static void bfq_update_inject_limit (struct bfq_data *bfqd,struct bfq_queue *bfqq) { -u64 tot_time_ns =ktime_get_ns () - bfqd->last_empty_occupied_ns; +u64 tot_time_ns =blk_time_get_ns () - bfqd->last_empty_occupied_ns;unsigned int old_limit = bfqq->inject_limit;if (bfqq->last_serv_time_ns >0 && bfqd->rqs_injected) {
block/blk-cgroup.c +1 -1
@@ -1846,7 +1846,7 @@ static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay) {unsigned long pflags;bool clamp; -u64 now =ktime_to_ns (ktime_get ()); +u64 now =blk_time_get_ns ();u64 exp;u64 delay_nsec =0 ;int tok;
block/blk-flush.c +1 -1
@@ -143,7 +143,7 @@ static void blk_account_io_flush(struct request *rq)part_stat_lock ();part_stat_inc (part, ios[STAT_FLUSH ]);part_stat_add (part, nsecs[STAT_FLUSH ], -ktime_get_ns () - rq->start_time_ns); +blk_time_get_ns () - rq->start_time_ns);part_stat_unlock (); }
block/blk-iocost.c +4 -4
@@ -829,7 +829,7 @@ static int ioc_autop_idx(struct ioc *ioc, struct gendisk *disk)/* step up/down based on the vrate */ vrate_pct =div64_u64 (ioc->vtime_base_rate *100 ,VTIME_PER_USEC ); - now_ns =ktime_get_ns (); + now_ns =blk_time_get_ns ();if (p->too_fast_vrate_pct && p->too_fast_vrate_pct <= vrate_pct) {if (!ioc->autop_too_fast_at)
@@ -1044,7 +1044,7 @@ static void ioc_now(struct ioc *ioc, struct ioc_now *now)unsigned seq;u64 vrate; - now->now_ns =ktime_get (); + now->now_ns =blk_time_get_ns (); now->now =ktime_to_us (now->now_ns); vrate =atomic64_read (&ioc->vtime_rate);
@@ -2810,7 +2810,7 @@ static void ioc_rqos_done(struct rq_qos *rqos, struct request *rq)return ; } - on_q_ns =ktime_get_ns () - rq->alloc_time_ns; + on_q_ns =blk_time_get_ns () - rq->alloc_time_ns; rq_wait_ns = rq->start_time_ns - rq->alloc_time_ns; size_nsec =div64_u64 (calc_size_vtime_cost (rq, ioc),VTIME_PER_NSEC );
@@ -2893,7 +2893,7 @@ static int blk_iocost_init(struct gendisk *disk) ioc->vtime_base_rate =VTIME_PER_USEC ;atomic64_set (&ioc->vtime_rate,VTIME_PER_USEC );seqcount_spinlock_init (&ioc->period_seqcount, &ioc->lock); - ioc->period_at =ktime_to_us (ktime_get ()); + ioc->period_at =ktime_to_us (blk_time_get ());atomic64_set (&ioc->cur_period,0 );atomic_set (&ioc->hweight_gen,0 );
block/blk-iolatency.c +3 -3
@@ -609,7 +609,7 @@ static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)if (!iolat->blkiolat->enabled)return ; - now =ktime_to_ns (ktime_get ()); + now =blk_time_get_ns ();while (blkg && blkg->parent) { iolat =blkg_to_lat (blkg);if (!iolat) {
@@ -661,7 +661,7 @@ static void blkiolatency_timer_fn(struct timer_list *t)struct blk_iolatency *blkiolat =from_timer (blkiolat, t, timer);struct blkcg_gq *blkg;struct cgroup_subsys_state *pos_css; -u64 now =ktime_to_ns (ktime_get ()); +u64 now =blk_time_get_ns ();rcu_read_lock (); blkg_for_each_descendant_pre(blkg, pos_css,
@@ -985,7 +985,7 @@ static void iolatency_pd_init(struct blkg_policy_data *pd)struct blkcg_gq *blkg =lat_to_blkg (iolat);struct rq_qos *rqos =iolat_rq_qos (blkg->q);struct blk_iolatency *blkiolat =BLKIOLATENCY (rqos); -u64 now =ktime_to_ns (ktime_get ()); +u64 now =blk_time_get_ns ();int cpu;if (blk_queue_nonrot (blkg->q))
block/blk-mq.c +8 -8
@@ -323,7 +323,7 @@ void blk_rq_init(struct request_queue *q, struct request *rq)RB_CLEAR_NODE (&rq->rb_node); rq->tag =BLK_MQ_NO_TAG ; rq->internal_tag =BLK_MQ_NO_TAG ; - rq->start_time_ns =ktime_get_ns (); + rq->start_time_ns =blk_time_get_ns (); rq->part =NULL ;blk_crypto_rq_set_defaults (rq); }
@@ -333,7 +333,7 @@ EXPORT_SYMBOL(blk_rq_init);static inline void blk_mq_rq_time_init (struct request *rq,u64 alloc_time_ns) {if (blk_mq_need_time_stamp (rq)) - rq->start_time_ns =ktime_get_ns (); + rq->start_time_ns =blk_time_get_ns ();else rq->start_time_ns =0 ;
@@ -444,7 +444,7 @@ static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)/* alloc_time includes depth and tag waits */ if (blk_queue_rq_alloc_time (q)) - alloc_time_ns =ktime_get_ns (); + alloc_time_ns =blk_time_get_ns ();if (data->cmd_flags &REQ_NOWAIT ) data->flags |=BLK_MQ_REQ_NOWAIT ;
@@ -629,7 +629,7 @@ struct request *blk_mq_alloc_request_hctx(struct request_queue *q,/* alloc_time includes depth and tag waits */ if (blk_queue_rq_alloc_time (q)) - alloc_time_ns =ktime_get_ns (); + alloc_time_ns =blk_time_get_ns ();/* * If the tag allocator sleeps we could get an allocation for a
@@ -1042,7 +1042,7 @@ static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)inline void __blk_mq_end_request (struct request *rq,blk_status_t error) {if (blk_mq_need_time_stamp (rq)) -__blk_mq_end_request_acct (rq,ktime_get_ns ()); +__blk_mq_end_request_acct (rq,blk_time_get_ns ());blk_mq_finish_request (rq);
@@ -1085,7 +1085,7 @@ void blk_mq_end_request_batch(struct io_comp_batch *iob)u64 now =0 ;if (iob->need_ts) - now =ktime_get_ns (); + now =blk_time_get_ns ();while ((rq =rq_list_pop (&iob->req_list)) !=NULL ) {prefetch (rq->bio);
@@ -1255,7 +1255,7 @@ void blk_mq_start_request(struct request *rq)if (test_bit (QUEUE_FLAG_STATS , &q->queue_flags) && !blk_rq_is_passthrough (rq)) { - rq->io_start_time_ns =ktime_get_ns (); + rq->io_start_time_ns =blk_time_get_ns (); rq->stats_sectors =blk_rq_sectors (rq); rq->rq_flags |=RQF_STATS ;rq_qos_issue (q, rq);
@@ -3107,7 +3107,7 @@ blk_status_t blk_insert_cloned_request(struct request *rq)blk_mq_run_dispatch_ops (q, ret =blk_mq_request_issue_directly (rq, true));if (ret) -blk_account_io_done (rq,ktime_get_ns ()); +blk_account_io_done (rq,blk_time_get_ns ());return ret; }EXPORT_SYMBOL_GPL (blk_insert_cloned_request);
block/blk-throttle.c +3 -3
@@ -1815,7 +1815,7 @@ static bool throtl_tg_is_idle(struct throtl_grp *tg) time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold); ret = tg->latency_target ==DFL_LATENCY_TARGET || tg->idletime_threshold ==DFL_IDLE_THRESHOLD || - (ktime_get_ns () >>10 ) - tg->last_finish_time > time || + (blk_time_get_ns () >>10 ) - tg->last_finish_time > time || tg->avg_idletime > tg->idletime_threshold || (tg->latency_target && tg->bio_cnt && tg->bad_bio_cnt *5 < tg->bio_cnt);
@@ -2060,7 +2060,7 @@ static void blk_throtl_update_idletime(struct throtl_grp *tg)if (last_finish_time ==0 )return ; - now =ktime_get_ns () >>10 ; + now =blk_time_get_ns () >>10 ;if (now <= last_finish_time || last_finish_time == tg->checked_last_finish_time)return ;
@@ -2327,7 +2327,7 @@ void blk_throtl_bio_endio(struct bio *bio)if (!tg->td->limit_valid[LIMIT_LOW ])return ; - finish_time_ns =ktime_get_ns (); + finish_time_ns =blk_time_get_ns (); tg->last_finish_time = finish_time_ns >>10 ; start_time =bio_issue_time (&bio->bi_issue) >>10 ;
block/blk-wbt.c +2 -3
@@ -274,13 +274,12 @@ static inline bool stat_sample_valid(struct blk_rq_stat *stat)static u64 rwb_sync_issue_lat (struct rq_wb *rwb) { -u64 now, issue =READ_ONCE (rwb->sync_issue); +u64 issue =READ_ONCE (rwb->sync_issue);if (!issue || !rwb->sync_cookie)return 0 ; - now =ktime_to_ns (ktime_get ()); -return now - issue; +return blk_time_get_ns () - issue; }static inline unsigned int wbt_inflight (struct rq_wb *rwb)
include/linux/blkdev.h +12 -1
@@ -24,6 +24,7 @@#include <linux/sbitmap.h> #include <linux/uuid.h> #include <linux/xarray.h> +#include <linux/timekeeping.h> struct module ;struct request_queue ;
@@ -975,6 +976,16 @@ static inline void blk_flush_plug(struct blk_plug *plug, bool async)int blkdev_issue_flush (struct block_device *bdev);long nr_blockdev_pages (void ); +static inline u64 blk_time_get_ns (void ) +{ +return ktime_get_ns (); +} + +static inline ktime_t blk_time_get (void ) +{ +return ns_to_ktime (blk_time_get_ns ()); +} +/* * From most significant bit: * 1 bit: reserved for other usage, see below
@@ -1013,7 +1024,7 @@ static inline void bio_issue_init(struct bio_issue *issue, { size &= (1ULL <<BIO_ISSUE_SIZE_BITS ) -1 ; issue->value = ((issue->value &BIO_ISSUE_RES_MASK ) | - (ktime_get_ns () &BIO_ISSUE_TIME_MASK ) | + (blk_time_get_ns () &BIO_ISSUE_TIME_MASK ) | ((u64 )size <<BIO_ISSUE_SIZE_SHIFT )); }
Signature · 1 line
2.43.0
Re: [PATCH 2/6] block: add blk_time_get_ns() and blk_time_get() helpers
Quoted 6 lines · Jens Axboe
Convert any user of
ktime_get_ns()to useblk_time_get_ns(), andktime_get()toblk_time_get(), so we have a unified API for querying the current time in nanoseconds or as ktime.No functional changes intended, this patch just wraps
ktime_get_ns()andktime_get()with a block helper.
Given that the helper only is (and only should be) used in the block layer code, any reason not to keep it private in block/ ?
Jens Axboe
maintainerRe: [PATCH 2/6] block: add blk_time_get_ns() and blk_time_get() helpers
Quoted 10 lines · Christoph Hellwig
Quoted 6 lines · Jens Axboe
Convert any user of
ktime_get_ns()to useblk_time_get_ns(), andktime_get()toblk_time_get(), so we have a unified API for querying the current time in nanoseconds or as ktime.No functional changes intended, this patch just wraps
ktime_get_ns()andktime_get()with a block helper.Given that the helper only is (and only should be) used in the block layer code, any reason not to keep it private in
block/?
I did originally put it in there, but you then run into dependency issues with needing the struct definitions too, eg for struct blk_plug. Hence not sure it makes sense, even if logically it'd fit better there.
Signature · 1 line
Jens Axboe
Jens Axboe
maintainerPATCH v4 3/6 block: cache current nsec time in struct blk_plug
In reply to Jens AxboeQuerying the current time is the most costly thing we do in the block layer per IO, and depending on kernel config settings, we may do it many times per IO.
None of the callers actually need nsec granularity. Take advantage of that by caching the current time in the plug, with the assumption here being that any time checking will be temporally close enough that the slight loss of precision doesn't matter.
If the block plug gets flushed, eg on preempt or schedule out, then we invalidate the cached clock.
On a basic peak IOPS test case with iostats enabled, this changes the performance from:
IOPS=108.41M, BW=52.93GiB/s, IOS/call=31/31
IOPS=108.43M, BW=52.94GiB/s, IOS/call=32/32
IOPS=108.29M, BW=52.88GiB/s, IOS/call=31/32
IOPS=108.35M, BW=52.91GiB/s, IOS/call=32/32
IOPS=108.42M, BW=52.94GiB/s, IOS/call=31/31
IOPS=108.40M, BW=52.93GiB/s, IOS/call=32/32
IOPS=108.31M, BW=52.89GiB/s, IOS/call=32/31
to
IOPS=118.79M, BW=58.00GiB/s, IOS/call=31/32
IOPS=118.62M, BW=57.92GiB/s, IOS/call=31/31
IOPS=118.80M, BW=58.01GiB/s, IOS/call=32/31
IOPS=118.78M, BW=58.00GiB/s, IOS/call=32/32
IOPS=118.69M, BW=57.95GiB/s, IOS/call=32/31
IOPS=118.62M, BW=57.92GiB/s, IOS/call=32/31
IOPS=118.63M, BW=57.92GiB/s, IOS/call=31/32
which is more than a 9% improvement in performance. Looking at perf diff, we can see a huge reduction in time overhead:
10.55% -9.88% [kernel.vmlinux] [k] read_tsc
1.31% -1.22% [kernel.vmlinux] [k] ktime_getNote that since this relies on blk_plug for the caching, it's only applicable to the issue side. But this is where most of the time calls happen anyway. On the completion side, cached time stamping is done with struct io_comp patch, as long as the driver supports it.
It's also worth noting that the above testing doesn't enable any of the higher cost CPU items on the block layer side, like wbt, cgroups, iocost, etc, which all would add additional time querying and hence overhead. IOW, results would likely look even better in comparison with those enabled, as distros would do.
--- block/blk-core.c | 1 + include/linux/blkdev.h | 15 ++++++++++++++- 2 files changed, 15 insertions(+), 1 deletion(-)
block/blk-core.c +1 -0
@@ -1073,6 +1073,7 @@ void blk_start_plug_nr_ios(struct blk_plug *plug, unsigned short nr_ios)if (tsk->plug)return ; + plug->cur_ktime =0 ; plug->mq_list =NULL ; plug->cached_rq =NULL ; plug->nr_ios = min_t(unsigned short, nr_ios, BLK_MAX_REQUEST_COUNT);
include/linux/blkdev.h +14 -1
@@ -943,6 +943,7 @@ struct blk_plug {/* if ios_left is > 1, we can batch tag/rq allocations */ struct request *cached_rq; +u64 cur_ktime;unsigned short nr_ios;unsigned short rq_count;
@@ -978,7 +979,19 @@ long nr_blockdev_pages(void);static inline u64 blk_time_get_ns (void ) { -return ktime_get_ns (); +struct blk_plug *plug = current->plug; + +if (!plug) +return ktime_get_ns (); + +/* +* 0 could very well be a valid time, but rather than flag "this is +* a valid timestamp" separately, just accept that we'll do an extra +* ktime_get_ns() if we just happen to get 0 as the current time. +*/ +if (!plug->cur_ktime) + plug->cur_ktime =ktime_get_ns (); +return plug->cur_ktime; }static inline ktime_t blk_time_get (void )
Signature · 1 line
2.43.0
Jens Axboe
maintainerPATCH v4 4/6 block: update cached timestamp post schedule/preemption
In reply to Jens AxboeMark the task as having a cached timestamp when set assign it, so we can efficiently check if it needs updating post being scheduled back in. This covers both the actual schedule out case, which would've flushed the plug, and the preemption case which doesn't touch the plugged requests (for many reasons, one of them being then we'd need to have preemption disabled around plug state manipulation).
--- block/blk-core.c | 2 ++ include/linux/blkdev.h | 20 +++++++++++++++++++- include/linux/sched.h | 2 +- kernel/sched/core.c | 4 +++- 4 files changed, 25 insertions(+), 3 deletions(-)
block/blk-core.c +2 -0
@@ -1173,6 +1173,8 @@ void __blk_flush_plug(struct blk_plug *plug, bool from_schedule)*/ if (unlikely (!rq_list_empty (plug->cached_rq)))blk_mq_free_plug_rqs (plug); + + current->flags &= ~PF_BLOCK_TS ; }/**
include/linux/blkdev.h +19 -1
@@ -977,6 +977,18 @@ static inline void blk_flush_plug(struct blk_plug *plug, bool async)int blkdev_issue_flush (struct block_device *bdev);long nr_blockdev_pages (void ); +/* +* tsk == current here +*/ +static inline void blk_plug_invalidate_ts (struct task_struct *tsk) +{ +struct blk_plug *plug = tsk->plug; + +if (plug) + plug->cur_ktime =0 ; + current->flags &= ~PF_BLOCK_TS ; +} +static inline u64 blk_time_get_ns (void ) {struct blk_plug *plug = current->plug;
@@ -989,8 +1001,10 @@ static inline u64 blk_time_get_ns(void)* a valid timestamp" separately, just accept that we'll do an extra * ktime_get_ns() if we just happen to get 0 as the current time. */ -if (!plug->cur_ktime) +if (!plug->cur_ktime) { plug->cur_ktime =ktime_get_ns (); + current->flags |=PF_BLOCK_TS ; + }return plug->cur_ktime; }
@@ -1062,6 +1076,10 @@ static inline void blk_flush_plug(struct blk_plug *plug, bool async) { } +static inline void blk_plug_invalidate_ts (struct task_struct *tsk) +{ +} +static inline int blkdev_issue_flush (struct block_device *bdev) {return 0 ;
include/linux/sched.h +1 -1
@@ -1642,7 +1642,7 @@ extern struct pid *cad_pid;#define PF_NO_SETAFFINITY 0x04000000/* Userland is not allowed to meddle with cpus_mask */ #define PF_MCE_EARLY 0x08000000/* Early kill for mce process policy */ #define PF_MEMALLOC_PIN 0x10000000/* Allocation context constrained to zones which allow long term pinning. */ -#define PF__HOLE__20000000 0x20000000 +#define PF_BLOCK_TS 0x20000000/* plug has ts that needs updating */ #define PF__HOLE__40000000 0x40000000#define PF_SUSPEND_TASK 0x80000000/* This thread called freeze_processes() and should not be frozen */
kernel/sched/core.c +3 -1
@@ -6787,7 +6787,9 @@ static inline void sched_submit_work(struct task_struct *tsk)static void sched_update_worker (struct task_struct *tsk) { -if (tsk->flags & (PF_WQ_WORKER |PF_IO_WORKER )) { +if (tsk->flags & (PF_WQ_WORKER |PF_IO_WORKER |PF_BLOCK_TS )) { +if (tsk->flags &PF_BLOCK_TS ) +blk_plug_invalidate_ts (tsk);if (tsk->flags &PF_WQ_WORKER )wq_worker_running (tsk);else
Signature · 1 line
2.43.0
Re: [PATCH 4/6] block: update cached timestamp post schedule/preemption
kernel/sched/core.c static inline void sched_submit_work(struct task_struct *tsk) · line 6795
Quoted 9 lines · Jens Axboe
static void
sched_update_worker(struct task_struct *tsk){kernel/sched/core.c +3
-1-if (tsk->flags & (PF_WQ_WORKER |PF_IO_WORKER )) { +if (tsk->flags & (PF_WQ_WORKER |PF_IO_WORKER |PF_BLOCK_TS )) { +if (tsk->flags &PF_BLOCK_TS ) +blk_plug_invalidate_ts (tsk);if (tsk->flags &PF_WQ_WORKER )wq_worker_running (tsk);else
I don't even know what is in the 'else' case, but it doesn't look right now that you added another flag to enter into this block. Before this patch, it looks like we could assume PF_IO_WORKER was set in the 'else' condition, but that's not necessarily the case with this change.
Jens Axboe
maintainerRe: [PATCH 4/6] block: update cached timestamp post schedule/preemption
Quoted 15 lines · Keith Busch
kernel/sched/core.c static inline void sched_submit_work(struct task_struct *tsk) · line 6795
Quoted 9 lines · Jens Axboe
static void
sched_update_worker(struct task_struct *tsk){kernel/sched/core.c +3
-1-if (tsk->flags & (PF_WQ_WORKER |PF_IO_WORKER )) { +if (tsk->flags & (PF_WQ_WORKER |PF_IO_WORKER |PF_BLOCK_TS )) { +if (tsk->flags &PF_BLOCK_TS ) +blk_plug_invalidate_ts (tsk);if (tsk->flags &PF_WQ_WORKER )wq_worker_running (tsk);else I don't even know what is in the 'else' case, but it doesn't look right now that you added another flag to enter into this block. Before this patch, it looks like we could assume
PF_IO_WORKERwas set in the 'else' condition, but that's not necessarily the case with this change.
Will add an else if for PF_IO_WORKER for that case, to be on the safe side.
Signature · 1 line
Jens Axboe
Jens Axboe
maintainerPATCH v4 5/6 block: shrink plug->{nr_ios, rq_count} to unsigned char
In reply to Jens AxboeWe never use more than 64 max in here, we can change them from unsigned short to just a byte. Add a BUILD_BUG_ON() check, in case the max plug count changes in the future.
--- block/blk-core.c | 4 ++-- block/blk-mq.c | 2 ++ include/linux/blkdev.h | 8 ++++---- 3 files changed, 8 insertions(+), 6 deletions(-)
block/blk-core.c +2 -2
@@ -1063,7 +1063,7 @@ int kblockd_mod_delayed_work_on(int cpu, struct delayed_work *dwork, }EXPORT_SYMBOL (kblockd_mod_delayed_work_on); -void blk_start_plug_nr_ios (struct blk_plug *plug,unsigned short nr_ios) +void blk_start_plug_nr_ios (struct blk_plug *plug,unsigned char nr_ios) {struct task_struct *tsk = current;
@@ -1076,7 +1076,7 @@ void blk_start_plug_nr_ios(struct blk_plug *plug, unsigned short nr_ios) plug->cur_ktime =0 ; plug->mq_list =NULL ; plug->cached_rq =NULL ; - plug->nr_ios = min_t(unsigned short, nr_ios, BLK_MAX_REQUEST_COUNT); + plug->nr_ios = min_t(unsigned char, nr_ios, BLK_MAX_REQUEST_COUNT); plug->rq_count =0 ; plug->multiple_queues = false; plug->has_elevator = false;
block/blk-mq.c +2 -0
@@ -1283,6 +1283,8 @@ EXPORT_SYMBOL(blk_mq_start_request);*/ static inline unsigned short blk_plug_max_rq_count (struct blk_plug *plug) { +BUILD_BUG_ON (2 *BLK_MAX_REQUEST_COUNT >U8_MAX ); +if (plug->multiple_queues)return BLK_MAX_REQUEST_COUNT *2 ;return BLK_MAX_REQUEST_COUNT ;
include/linux/blkdev.h +4 -4
@@ -944,9 +944,9 @@ struct blk_plug {/* if ios_left is > 1, we can batch tag/rq allocations */ struct request *cached_rq;u64 cur_ktime; -unsigned short nr_ios; +unsigned char nr_ios; -unsigned short rq_count; +unsigned char rq_count;bool multiple_queues;bool has_elevator;
@@ -964,7 +964,7 @@ struct blk_plug_cb {extern struct blk_plug_cb *blk_check_plugged (blk_plug_cb_fn unplug,void *data,int size);extern void blk_start_plug (struct blk_plug *); -extern void blk_start_plug_nr_ios (struct blk_plug *,unsigned short ); +extern void blk_start_plug_nr_ios (struct blk_plug *,unsigned char );extern void blk_finish_plug (struct blk_plug *);void __blk_flush_plug (struct blk_plug *plug,bool from_schedule);
@@ -1060,7 +1060,7 @@ struct blk_plug { };static inline void blk_start_plug_nr_ios (struct blk_plug *plug, -unsigned short nr_ios) +unsigned char nr_ios) { }
Signature · 1 line
2.43.0
Re: [PATCH 5/6] block: shrink plug->{nr_ios, rq_count} to unsigned char
Quoted 3 lines · Jens Axboe
We never use more than 64 max in here, we can change them from unsigned short to just a byte. Add a
BUILD_BUG_ON()check, in case the max plug count changes in the future.
Do we care about the size of this once per task structure? byte-level access tends to be quite a bit more expensive on various architectures.
Jens Axboe
maintainerRe: [PATCH 5/6] block: shrink plug->{nr_ios, rq_count} to unsigned char
Quoted 7 lines · Christoph Hellwig
Quoted 3 lines · Jens Axboe
We never use more than 64 max in here, we can change them from unsigned short to just a byte. Add a
BUILD_BUG_ON()check, in case the max plug count changes in the future.Do we care about the size of this once per task structure? byte-level access tends to be quite a bit more expensive on various architectures.
I don't think we care that much, and honestly I'm fine dropping the last two patches. They don't matter that much to me, and we can always revisit if we do care more about shrinking the blk_plug later on.
Signature · 1 line
Jens Axboe
Jens Axboe
maintainerPATCH v4 6/6 block: convert struct blk_plug callback list to hlists
In reply to Jens AxboeWe currently use a doubly linked list, which means the head takes up 16 bytes. As any iteration goes over the full list by first splicing it to an on-stack copy, we never need to remove members from the middle of the list.
Convert it to an hlist instead, saving 8 bytes in the blk_plug structure. This also helps save 40 bytes of text in the core block code, tested on arm64.
This does mean that flush callbacks will be run in reverse. While this should not pose a problem, we can always change the list splicing to just iteration-and-add instead, preservering ordering. These lists are generally just a single entry (or a few entries), either way this should be fine.
--- block/blk-core.c | 26 ++++++++++++++------------ drivers/md/raid1-10.c | 2 +- include/linux/blkdev.h | 4 ++-- 3 files changed, 17 insertions(+), 15 deletions(-)
block/blk-core.c +14 -12
@@ -1080,7 +1080,7 @@ void blk_start_plug_nr_ios(struct blk_plug *plug, unsigned char nr_ios) plug->rq_count =0 ; plug->multiple_queues = false; plug->has_elevator = false; -INIT_LIST_HEAD (&plug->cb_list); +INIT_HLIST_HEAD (&plug->cb_list);/* * Store ordering should not be needed here, since a potential
@@ -1120,16 +1120,18 @@ EXPORT_SYMBOL(blk_start_plug);static void flush_plug_callbacks (struct blk_plug *plug,bool from_schedule) { -LIST_HEAD (callbacks); +HLIST_HEAD (callbacks); -while (!list_empty (&plug->cb_list)) { -list_splice_init (&plug->cb_list, &callbacks); +while (!hlist_empty (&plug->cb_list)) { +struct hlist_node *entry, *tmp; -while (!list_empty (&callbacks)) { -struct blk_plug_cb *cb =list_first_entry (&callbacks, - struct blk_plug_cb, - list); -list_del (&cb->list); +hlist_move_list (&plug->cb_list, &callbacks); + +hlist_for_each_safe (entry, tmp, &callbacks) { +struct blk_plug_cb *cb; + + cb = hlist_entry(entry, struct blk_plug_cb, list); +hlist_del (&cb->list); cb->callback (cb, from_schedule); } }
@@ -1144,7 +1146,7 @@ struct blk_plug_cb *blk_check_plugged(blk_plug_cb_fn unplug, void *data,if (!plug)return NULL ; -list_for_each_entry (cb, &plug->cb_list, list) +hlist_for_each_entry (cb, &plug->cb_list, list)if (cb->callback == unplug && cb->data == data)return cb;
@@ -1154,7 +1156,7 @@ struct blk_plug_cb *blk_check_plugged(blk_plug_cb_fn unplug, void *data,if (cb) { cb->data = data; cb->callback = unplug; -list_add (&cb->list, &plug->cb_list); +hlist_add_head (&cb->list, &plug->cb_list); }return cb; }
@@ -1162,7 +1164,7 @@ EXPORT_SYMBOL(blk_check_plugged);void __blk_flush_plug (struct blk_plug *plug,bool from_schedule) { -if (!list_empty (&plug->cb_list)) +if (!hlist_empty (&plug->cb_list))flush_plug_callbacks (plug, from_schedule);blk_mq_flush_plug_list (plug, from_schedule);/*
drivers/md/raid1-10.c +1 -1
@@ -152,7 +152,7 @@ static inline bool raid1_add_bio_to_plug(struct mddev *mddev, struct bio *bio, plug = container_of(cb, struct raid1_plug_cb, cb);bio_list_add (&plug->pending, bio);if (++plug->count /MAX_PLUG_BIO >= copies) { -list_del (&cb->list); +hlist_del (&cb->list); cb->callback (cb, false); }
include/linux/blkdev.h +2 -2
@@ -951,13 +951,13 @@ struct blk_plug {bool multiple_queues;bool has_elevator; -struct list_head cb_list;/* md requires an unplug callback */ +struct hlist_head cb_list;/* md requires an unplug callback */ };struct blk_plug_cb ;typedef void (*blk_plug_cb_fn )(struct blk_plug_cb *,bool );struct blk_plug_cb { -struct list_head list; +struct hlist_node list;blk_plug_cb_fn callback;void *data; };
Signature · 1 line
2.43.0
Re: [PATCH 6/6] block: convert struct blk_plug callback list to hlists
Quoted 9 lines · Jens Axboe
We currently use a doubly linked list, which means the head takes up 16 bytes. As any iteration goes over the full list by first splicing it to an on-stack copy, we never need to remove members from the middle of the list.
Convert it to an hlist instead, saving 8 bytes in the
blk_plugstructure. This also helps save 40 bytes of text in the core block code, tested on arm64.
Quoted 5 lines
This does mean that flush callbacks will be run in reverse. While this should not pose a problem, we can always change the list splicing to just iteration-and-add instead, preservering ordering. These lists are generally just a single entry (or a few entries), either way this should be fine.
That would complete mess up I/O order for anyone using flush callbacks, I don't think that's great.