1 // SPDX-License-Identifier: GPL-2.0
6  * - Monitor latencies in a defined window of time.
7  * - If the minimum latency in the above window exceeds some target, increment
8  *   scaling step and scale down queue depth by a factor of 2x. The monitoring
9  *   window is then shrunk to 100 / sqrt(scaling step + 1).
10  * - For any window where we don't have solid data on what the latencies
12  * - If latencies look good, decrement scaling step.
13  * - If we're only doing writes, allow the scaling step to go negative. This
15  *   scaling step of 0 if reads show up or the heavy writers finish. Unlike
17  *   scaling step retains the default step==0 window size.
25 #include <linux/backing-dev.h>
28 #include "blk-stat.h"
29 #include "blk-wbt.h"
30 #include "blk-rq-qos.h"
75 	 * information to make a firm scale up/down decision.
89 	unsigned long last_issue;		/* last non-throttled issue */
90 	unsigned long last_comp;		/* last non-throttled comp */
104 	rq->wbt_flags = 0;  in wbt_clear_state()
109 	return rq->wbt_flags;  in wbt_flags()
114 	return rq->wbt_flags & WBT_TRACKED;  in wbt_is_tracked()
119 	return rq->wbt_flags & WBT_READ;  in wbt_is_read()
124 	 * Default setting, we'll scale up (to 75% of QD max) or down (min 1)
141 	 * information to scale up or down, scale up.
148 	return rwb && rwb->enable_state != WBT_STATE_OFF_DEFAULT &&  in rwb_enabled()
149 		      rwb->enable_state != WBT_STATE_OFF_MANUAL;  in rwb_enabled()
168 	struct bdi_writeback *wb = &rwb->rqos.disk->bdi->wb;  in wb_recent_wait()
170 	return time_before(jiffies, wb->dirty_sleep + HZ);  in wb_recent_wait()
177 		return &rwb->rq_wait[WBT_RWQ_KSWAPD];  in get_rq_wait()
179 		return &rwb->rq_wait[WBT_RWQ_DISCARD];  in get_rq_wait()
181 	return &rwb->rq_wait[WBT_RWQ_BG];  in get_rq_wait()
189 		struct rq_wait *rqw = &rwb->rq_wait[i];  in rwb_wake_all()
191 		if (wq_has_sleeper(&rqw->wait))  in rwb_wake_all()
192 			wake_up_all(&rqw->wait);  in rwb_wake_all()
201 	inflight = atomic_dec_return(&rqw->inflight);  in wbt_rqw_done()
205 	 * the device does write back caching, drop further down before we  in wbt_rqw_done()
209 		limit = rwb->wb_background;  in wbt_rqw_done()
210 	else if (rwb->wc && !wb_recent_wait(rwb))  in wbt_rqw_done()
213 		limit = rwb->wb_normal;  in wbt_rqw_done()
221 	if (wq_has_sleeper(&rqw->wait)) {  in wbt_rqw_done()
222 		int diff = limit - inflight;  in wbt_rqw_done()
224 		if (!inflight || diff >= rwb->wb_background / 2)  in wbt_rqw_done()
225 			wake_up_all(&rqw->wait);  in wbt_rqw_done()
250 		if (rwb->sync_cookie == rq) {  in wbt_done()
251 			rwb->sync_issue = 0;  in wbt_done()
252 			rwb->sync_cookie = NULL;  in wbt_done()
256 			wb_timestamp(rwb, &rwb->last_comp);  in wbt_done()
258 		WARN_ON_ONCE(rq == rwb->sync_cookie);  in wbt_done()
278 	u64 now, issue = READ_ONCE(rwb->sync_issue);  in rwb_sync_issue_lat()
280 	if (!issue || !rwb->sync_cookie)  in rwb_sync_issue_lat()
284 	return now - issue;  in rwb_sync_issue_lat()
292 		ret += atomic_read(&rwb->rq_wait[i].inflight);  in wbt_inflight()
306 	struct backing_dev_info *bdi = rwb->rqos.disk->bdi;  in latency_exceeded()
307 	struct rq_depth *rqd = &rwb->rq_depth;  in latency_exceeded()
320 	if (thislat > rwb->cur_win_nsec ||  in latency_exceeded()
321 	    (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) {  in latency_exceeded()
343 	 * If the 'min' latency exceeds our target, step down.  in latency_exceeded()
345 	if (stat[READ].min > rwb->min_lat_nsec) {  in latency_exceeded()
351 	if (rqd->scale_step)  in latency_exceeded()
359 	struct backing_dev_info *bdi = rwb->rqos.disk->bdi;  in rwb_trace_step()
360 	struct rq_depth *rqd = &rwb->rq_depth;  in rwb_trace_step()
362 	trace_wbt_step(bdi, msg, rqd->scale_step, rwb->cur_win_nsec,  in rwb_trace_step()
363 			rwb->wb_background, rwb->wb_normal, rqd->max_depth);  in rwb_trace_step()
368 	if (rwb->min_lat_nsec == 0) {  in calc_wb_limits()
369 		rwb->wb_normal = rwb->wb_background = 0;  in calc_wb_limits()
370 	} else if (rwb->rq_depth.max_depth <= 2) {  in calc_wb_limits()
371 		rwb->wb_normal = rwb->rq_depth.max_depth;  in calc_wb_limits()
372 		rwb->wb_background = 1;  in calc_wb_limits()
374 		rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2;  in calc_wb_limits()
375 		rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4;  in calc_wb_limits()
381 	if (!rq_depth_scale_up(&rwb->rq_depth))  in scale_up()
384 	rwb->unknown_cnt = 0;  in scale_up()
391 	if (!rq_depth_scale_down(&rwb->rq_depth, hard_throttle))  in scale_down()
394 	rwb->unknown_cnt = 0;  in scale_down()
395 	rwb_trace_step(rwb, tracepoint_string("scale down"));  in scale_down()
400 	struct rq_depth *rqd = &rwb->rq_depth;  in rwb_arm_timer()
402 	if (rqd->scale_step > 0) {  in rwb_arm_timer()
409 		rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,  in rwb_arm_timer()
410 					int_sqrt((rqd->scale_step + 1) << 8));  in rwb_arm_timer()
413 		 * For step < 0, we don't want to increase/decrease the  in rwb_arm_timer()
416 		rwb->cur_win_nsec = rwb->win_nsec;  in rwb_arm_timer()
419 	blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec);  in rwb_arm_timer()
424 	struct rq_wb *rwb = cb->data;  in wb_timer_fn()
425 	struct rq_depth *rqd = &rwb->rq_depth;  in wb_timer_fn()
429 	if (!rwb->rqos.disk)  in wb_timer_fn()
432 	status = latency_exceeded(rwb, cb->stat);  in wb_timer_fn()
434 	trace_wbt_timer(rwb->rqos.disk->bdi, status, rqd->scale_step, inflight);  in wb_timer_fn()
437 	 * If we exceeded the latency target, step down. If we did not,  in wb_timer_fn()
438 	 * step one level up. If we don't know enough to say either exceeded  in wb_timer_fn()
450 		 * We started a the center step, but don't have a valid  in wb_timer_fn()
452 		 * Allow step to go negative, to increase write perf.  in wb_timer_fn()
457 		if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP)  in wb_timer_fn()
462 		 * case, slowly return to center state (step == 0).  in wb_timer_fn()
464 		if (rqd->scale_step > 0)  in wb_timer_fn()
466 		else if (rqd->scale_step < 0)  in wb_timer_fn()
474 	 * Re-arm timer, if we have IO in flight  in wb_timer_fn()
476 	if (rqd->scale_step || inflight)  in wb_timer_fn()
482 	struct rq_depth *rqd = &rwb->rq_depth;  in wbt_update_limits()
484 	rqd->scale_step = 0;  in wbt_update_limits()
485 	rqd->scaled_max = false;  in wbt_update_limits()
505 	return RQWB(rqos)->min_lat_nsec;  in wbt_get_min_lat()
514 	RQWB(rqos)->min_lat_nsec = val;  in wbt_set_min_lat()
516 		RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL;  in wbt_set_min_lat()
518 		RQWB(rqos)->enable_state = WBT_STATE_OFF_MANUAL;  in wbt_set_min_lat()
528 	return time_before(now, rwb->last_issue + HZ / 10) ||  in close_io()
529 		time_before(now, rwb->last_comp + HZ / 10);  in close_io()
539 		return rwb->wb_background;  in get_limit()
550 		limit = rwb->rq_depth.max_depth;  in get_limit()
556 		limit = rwb->wb_background;  in get_limit()
558 		limit = rwb->wb_normal;  in get_limit()
572 	return rq_wait_inc_below(rqw, get_limit(data->rwb, data->opf));  in wbt_inflight_cb()
578 	wbt_rqw_done(data->rwb, rqw, data->wb_acct);  in wbt_cleanup_cb()
605 		if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) ==  in wbt_should_throttle()
645  * If we do sleep, we'll release and re-grab it.
655 			wb_timestamp(rwb, &rwb->last_issue);  in wbt_wait()
659 	__wbt_wait(rwb, flags, bio->bi_opf);  in wbt_wait()
661 	if (!blk_stat_is_active(rwb->cb))  in wbt_wait()
668 	rq->wbt_flags |= bio_to_wbt_flags(rwb, bio);  in wbt_track()
685 	if (wbt_is_read(rq) && !rwb->sync_issue) {  in wbt_issue()
686 		rwb->sync_cookie = rq;  in wbt_issue()
687 		rwb->sync_issue = rq->io_start_time_ns;  in wbt_issue()
696 	if (rq == rwb->sync_cookie) {  in wbt_requeue()
697 		rwb->sync_issue = 0;  in wbt_requeue()
698 		rwb->sync_cookie = NULL;  in wbt_requeue()
706 		RQWB(rqos)->wc = write_cache_on;  in wbt_set_write_cache()
714 	struct request_queue *q = disk->queue;  in wbt_enable_default()
718 	if (q->elevator &&  in wbt_enable_default()
719 	    test_bit(ELEVATOR_FLAG_DISABLE_WBT, &q->elevator->flags))  in wbt_enable_default()
725 		if (enable && RQWB(rqos)->enable_state == WBT_STATE_OFF_DEFAULT)  in wbt_enable_default()
726 			RQWB(rqos)->enable_state = WBT_STATE_ON_DEFAULT;  in wbt_enable_default()
730 	/* Queue not registered? Maybe shutting down... */  in wbt_enable_default()
742 	 * We default to 2msec for non-rotational storage, and 75msec  in wbt_default_latency_nsec()
761 	return -1;  in wbt_data_dir()
766 	RQWB(rqos)->rq_depth.queue_depth = blk_queue_depth(rqos->disk->queue);  in wbt_queue_depth_changed()
774 	blk_stat_remove_callback(rqos->disk->queue, rwb->cb);  in wbt_exit()
775 	blk_stat_free_callback(rwb->cb);  in wbt_exit()
784 	struct rq_qos *rqos = wbt_rq_qos(disk->queue);  in wbt_disable_default()
789 	if (rwb->enable_state == WBT_STATE_ON_DEFAULT) {  in wbt_disable_default()
790 		blk_stat_deactivate(rwb->cb);  in wbt_disable_default()
791 		rwb->enable_state = WBT_STATE_OFF_DEFAULT;  in wbt_disable_default()
802 	seq_printf(m, "%llu\n", rwb->cur_win_nsec);  in wbt_curr_win_nsec_show()
811 	seq_printf(m, "%d\n", rwb->enable_state);  in wbt_enabled_show()
819 	seq_printf(m, "%u\n", rqos->id);  in wbt_id_show()
831 			   atomic_read(&rwb->rq_wait[i].inflight));  in wbt_inflight_show()
840 	seq_printf(m, "%lu\n", rwb->min_lat_nsec);  in wbt_min_lat_nsec_show()
849 	seq_printf(m, "%u\n", rwb->unknown_cnt);  in wbt_unknown_cnt_show()
858 	seq_printf(m, "%u\n", rwb->wb_normal);  in wbt_normal_show()
867 	seq_printf(m, "%u\n", rwb->wb_background);  in wbt_background_show()
900 	struct request_queue *q = disk->queue;  in wbt_init()
907 		return -ENOMEM;  in wbt_init()
909 	rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb);  in wbt_init()
910 	if (!rwb->cb) {  in wbt_init()
912 		return -ENOMEM;  in wbt_init()
916 		rq_wait_init(&rwb->rq_wait[i]);  in wbt_init()
918 	rwb->last_comp = rwb->last_issue = jiffies;  in wbt_init()
919 	rwb->win_nsec = RWB_WINDOW_NSEC;  in wbt_init()
920 	rwb->enable_state = WBT_STATE_ON_DEFAULT;  in wbt_init()
921 	rwb->wc = test_bit(QUEUE_FLAG_WC, &q->queue_flags);  in wbt_init()
922 	rwb->rq_depth.default_depth = RWB_DEF_DEPTH;  in wbt_init()
923 	rwb->min_lat_nsec = wbt_default_latency_nsec(q);  in wbt_init()
924 	rwb->rq_depth.queue_depth = blk_queue_depth(q);  in wbt_init()
930 	mutex_lock(&q->rq_qos_mutex);  in wbt_init()
931 	ret = rq_qos_add(&rwb->rqos, disk, RQ_QOS_WBT, &wbt_rqos_ops);  in wbt_init()
932 	mutex_unlock(&q->rq_qos_mutex);  in wbt_init()
936 	blk_stat_add_callback(q, rwb->cb);  in wbt_init()
941 	blk_stat_free_callback(rwb->cb);  in wbt_init()