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
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
25 #include <linux/backing-dev.h>
28 #include "blk-wbt.h"
29 #include "blk-rq-qos.h"
36 	rq->wbt_flags = 0;  in wbt_clear_state()
41 	return rq->wbt_flags;  in wbt_flags()
46 	return rq->wbt_flags & WBT_TRACKED;  in wbt_is_tracked()
51 	return rq->wbt_flags & WBT_READ;  in wbt_is_read()
80 	return rwb && rwb->enable_state != WBT_STATE_OFF_DEFAULT &&  in rwb_enabled()
81 		      rwb->wb_normal != 0;  in rwb_enabled()
100 	struct bdi_writeback *wb = &rwb->rqos.q->disk->bdi->wb;  in wb_recent_wait()
102 	return time_before(jiffies, wb->dirty_sleep + HZ);  in wb_recent_wait()
109 		return &rwb->rq_wait[WBT_RWQ_KSWAPD];  in get_rq_wait()
111 		return &rwb->rq_wait[WBT_RWQ_DISCARD];  in get_rq_wait()
113 	return &rwb->rq_wait[WBT_RWQ_BG];  in get_rq_wait()
121 		struct rq_wait *rqw = &rwb->rq_wait[i];  in rwb_wake_all()
123 		if (wq_has_sleeper(&rqw->wait))  in rwb_wake_all()
124 			wake_up_all(&rqw->wait);  in rwb_wake_all()
133 	inflight = atomic_dec_return(&rqw->inflight);  in wbt_rqw_done()
150 		limit = rwb->wb_background;  in wbt_rqw_done()
151 	else if (rwb->wc && !wb_recent_wait(rwb))  in wbt_rqw_done()
154 		limit = rwb->wb_normal;  in wbt_rqw_done()
162 	if (wq_has_sleeper(&rqw->wait)) {  in wbt_rqw_done()
163 		int diff = limit - inflight;  in wbt_rqw_done()
165 		if (!inflight || diff >= rwb->wb_background / 2)  in wbt_rqw_done()
166 			wake_up_all(&rqw->wait);  in wbt_rqw_done()
191 		if (rwb->sync_cookie == rq) {  in wbt_done()
192 			rwb->sync_issue = 0;  in wbt_done()
193 			rwb->sync_cookie = NULL;  in wbt_done()
197 			wb_timestamp(rwb, &rwb->last_comp);  in wbt_done()
199 		WARN_ON_ONCE(rq == rwb->sync_cookie);  in wbt_done()
210 	 * that it's writes impacting us, and not just some sole read on  in stat_sample_valid()
219 	u64 now, issue = READ_ONCE(rwb->sync_issue);  in rwb_sync_issue_lat()
221 	if (!issue || !rwb->sync_cookie)  in rwb_sync_issue_lat()
225 	return now - issue;  in rwb_sync_issue_lat()
237 	struct backing_dev_info *bdi = rwb->rqos.q->disk->bdi;  in latency_exceeded()
238 	struct rq_depth *rqd = &rwb->rq_depth;  in latency_exceeded()
244 	 * flag the latency as exceeded. wbt works off completion latencies,  in latency_exceeded()
248 	 * window, then count that sync IO as a violation of the latency.  in latency_exceeded()
251 	if (thislat > rwb->cur_win_nsec ||  in latency_exceeded()
252 	    (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) {  in latency_exceeded()
264 		 * waited or still has writes in flights, consider us doing  in latency_exceeded()
274 	 * If the 'min' latency exceeds our target, step down.  in latency_exceeded()
276 	if (stat[READ].min > rwb->min_lat_nsec) {  in latency_exceeded()
282 	if (rqd->scale_step)  in latency_exceeded()
290 	struct backing_dev_info *bdi = rwb->rqos.q->disk->bdi;  in rwb_trace_step()
291 	struct rq_depth *rqd = &rwb->rq_depth;  in rwb_trace_step()
293 	trace_wbt_step(bdi, msg, rqd->scale_step, rwb->cur_win_nsec,  in rwb_trace_step()
294 			rwb->wb_background, rwb->wb_normal, rqd->max_depth);  in rwb_trace_step()
299 	if (rwb->min_lat_nsec == 0) {  in calc_wb_limits()
300 		rwb->wb_normal = rwb->wb_background = 0;  in calc_wb_limits()
301 	} else if (rwb->rq_depth.max_depth <= 2) {  in calc_wb_limits()
302 		rwb->wb_normal = rwb->rq_depth.max_depth;  in calc_wb_limits()
303 		rwb->wb_background = 1;  in calc_wb_limits()
305 		rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2;  in calc_wb_limits()
306 		rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4;  in calc_wb_limits()
312 	if (!rq_depth_scale_up(&rwb->rq_depth))  in scale_up()
315 	rwb->unknown_cnt = 0;  in scale_up()
322 	if (!rq_depth_scale_down(&rwb->rq_depth, hard_throttle))  in scale_down()
325 	rwb->unknown_cnt = 0;  in scale_down()
331 	struct rq_depth *rqd = &rwb->rq_depth;  in rwb_arm_timer()
333 	if (rqd->scale_step > 0) {  in rwb_arm_timer()
340 		rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,  in rwb_arm_timer()
341 					int_sqrt((rqd->scale_step + 1) << 8));  in rwb_arm_timer()
347 		rwb->cur_win_nsec = rwb->win_nsec;  in rwb_arm_timer()
350 	blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec);  in rwb_arm_timer()
355 	struct rq_wb *rwb = cb->data;  in wb_timer_fn()
356 	struct rq_depth *rqd = &rwb->rq_depth;  in wb_timer_fn()
360 	status = latency_exceeded(rwb, cb->stat);  in wb_timer_fn()
362 	trace_wbt_timer(rwb->rqos.q->disk->bdi, status, rqd->scale_step,  in wb_timer_fn()
366 	 * If we exceeded the latency target, step down. If we did not,  in wb_timer_fn()
386 		if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP)  in wb_timer_fn()
393 		if (rqd->scale_step > 0)  in wb_timer_fn()
395 		else if (rqd->scale_step < 0)  in wb_timer_fn()
403 	 * Re-arm timer, if we have IO in flight  in wb_timer_fn()
405 	if (rqd->scale_step || inflight)  in wb_timer_fn()
411 	struct rq_depth *rqd = &rwb->rq_depth;  in wbt_update_limits()
413 	rqd->scale_step = 0;  in wbt_update_limits()
414 	rqd->scaled_max = false;  in wbt_update_limits()
427 	return RQWB(rqos)->min_lat_nsec;  in wbt_get_min_lat()
435 	RQWB(rqos)->min_lat_nsec = val;  in wbt_set_min_lat()
436 	RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL;  in wbt_set_min_lat()
445 	return time_before(now, rwb->last_issue + HZ / 10) ||  in close_io()
446 		time_before(now, rwb->last_comp + HZ / 10);  in close_io()
457 	 * we'll properly inc a new IO, and dec+wakeup at the end.  in get_limit()
463 		return rwb->wb_background;  in get_limit()
474 		limit = rwb->rq_depth.max_depth;  in get_limit()
478 		 * limit us to half the depth for background writeback.  in get_limit()
480 		limit = rwb->wb_background;  in get_limit()
482 		limit = rwb->wb_normal;  in get_limit()
496 	return rq_wait_inc_below(rqw, get_limit(data->rwb, data->rw));  in wbt_inflight_cb()
502 	wbt_rqw_done(data->rwb, rqw, data->wb_acct);  in wbt_cleanup_cb()
529 		if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) ==  in wbt_should_throttle()
569  * If we do sleep, we'll release and re-grab it.
579 			wb_timestamp(rwb, &rwb->last_issue);  in wbt_wait()
583 	__wbt_wait(rwb, flags, bio->bi_opf);  in wbt_wait()
585 	if (!blk_stat_is_active(rwb->cb))  in wbt_wait()
592 	rq->wbt_flags |= bio_to_wbt_flags(rwb, bio);  in wbt_track()
603 	 * Track sync issue, in case it takes a long time to complete. Allows us  in wbt_issue()
609 	if (wbt_is_read(rq) && !rwb->sync_issue) {  in wbt_issue()
610 		rwb->sync_cookie = rq;  in wbt_issue()
611 		rwb->sync_issue = rq->io_start_time_ns;  in wbt_issue()
620 	if (rq == rwb->sync_cookie) {  in wbt_requeue()
621 		rwb->sync_issue = 0;  in wbt_requeue()
622 		rwb->sync_cookie = NULL;  in wbt_requeue()
630 		RQWB(rqos)->wc = write_cache_on;  in wbt_set_write_cache()
642 		if (RQWB(rqos)->enable_state == WBT_STATE_OFF_DEFAULT)  in wbt_enable_default()
643 			RQWB(rqos)->enable_state = WBT_STATE_ON_DEFAULT;  in wbt_enable_default()
659 	 * We default to 2msec for non-rotational storage, and 75msec  in wbt_default_latency_nsec()
678 	return -1;  in wbt_data_dir()
683 	RQWB(rqos)->rq_depth.queue_depth = blk_queue_depth(rqos->q);  in wbt_queue_depth_changed()
690 	struct request_queue *q = rqos->q;  in wbt_exit()
692 	blk_stat_remove_callback(q, rwb->cb);  in wbt_exit()
693 	blk_stat_free_callback(rwb->cb);  in wbt_exit()
707 	if (rwb->enable_state == WBT_STATE_ON_DEFAULT) {  in wbt_disable_default()
708 		blk_stat_deactivate(rwb->cb);  in wbt_disable_default()
709 		rwb->enable_state = WBT_STATE_OFF_DEFAULT;  in wbt_disable_default()
720 	seq_printf(m, "%llu\n", rwb->cur_win_nsec);  in wbt_curr_win_nsec_show()
729 	seq_printf(m, "%d\n", rwb->enable_state);  in wbt_enabled_show()
737 	seq_printf(m, "%u\n", rqos->id);  in wbt_id_show()
749 			   atomic_read(&rwb->rq_wait[i].inflight));  in wbt_inflight_show()
758 	seq_printf(m, "%lu\n", rwb->min_lat_nsec);  in wbt_min_lat_nsec_show()
767 	seq_printf(m, "%u\n", rwb->unknown_cnt);  in wbt_unknown_cnt_show()
776 	seq_printf(m, "%u\n", rwb->wb_normal);  in wbt_normal_show()
785 	seq_printf(m, "%u\n", rwb->wb_background);  in wbt_background_show()
823 		return -ENOMEM;  in wbt_init()
825 	rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb);  in wbt_init()
826 	if (!rwb->cb) {  in wbt_init()
828 		return -ENOMEM;  in wbt_init()
832 		rq_wait_init(&rwb->rq_wait[i]);  in wbt_init()
834 	rwb->rqos.id = RQ_QOS_WBT;  in wbt_init()
835 	rwb->rqos.ops = &wbt_rqos_ops;  in wbt_init()
836 	rwb->rqos.q = q;  in wbt_init()
837 	rwb->last_comp = rwb->last_issue = jiffies;  in wbt_init()
838 	rwb->win_nsec = RWB_WINDOW_NSEC;  in wbt_init()
839 	rwb->enable_state = WBT_STATE_ON_DEFAULT;  in wbt_init()
840 	rwb->wc = 1;  in wbt_init()
841 	rwb->rq_depth.default_depth = RWB_DEF_DEPTH;  in wbt_init()
846 	rq_qos_add(q, &rwb->rqos);  in wbt_init()
847 	blk_stat_add_callback(q, rwb->cb);  in wbt_init()
849 	rwb->min_lat_nsec = wbt_default_latency_nsec(q);  in wbt_init()
851 	wbt_queue_depth_changed(&rwb->rqos);  in wbt_init()
852 	wbt_set_write_cache(q, test_bit(QUEUE_FLAG_WC, &q->queue_flags));  in wbt_init()