xref: /linux/lib/dynamic_queue_limits.c (revision 6025b913)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Dynamic byte queue limits.  See include/linux/dynamic_queue_limits.h
4  *
5  * Copyright (c) 2011, Tom Herbert <therbert@google.com>
6  */
7 #include <linux/types.h>
8 #include <linux/kernel.h>
9 #include <linux/jiffies.h>
10 #include <linux/dynamic_queue_limits.h>
11 #include <linux/compiler.h>
12 #include <linux/export.h>
13 #include <trace/events/napi.h>
14 
15 #define POSDIFF(A, B) ((int)((A) - (B)) > 0 ? (A) - (B) : 0)
16 #define AFTER_EQ(A, B) ((int)((A) - (B)) >= 0)
17 
dql_check_stall(struct dql * dql)18 static void dql_check_stall(struct dql *dql)
19 {
20 	unsigned short stall_thrs;
21 	unsigned long now;
22 
23 	stall_thrs = READ_ONCE(dql->stall_thrs);
24 	if (!stall_thrs)
25 		return;
26 
27 	now = jiffies;
28 	/* Check for a potential stall */
29 	if (time_after_eq(now, dql->last_reap + stall_thrs)) {
30 		unsigned long hist_head, t, start, end;
31 
32 		/* We are trying to detect a period of at least @stall_thrs
33 		 * jiffies without any Tx completions, but during first half
34 		 * of which some Tx was posted.
35 		 */
36 dqs_again:
37 		hist_head = READ_ONCE(dql->history_head);
38 		/* pairs with smp_wmb() in dql_queued() */
39 		smp_rmb();
40 
41 		/* Get the previous entry in the ring buffer, which is the
42 		 * oldest sample.
43 		 */
44 		start = (hist_head - DQL_HIST_LEN + 1) * BITS_PER_LONG;
45 
46 		/* Advance start to continue from the last reap time */
47 		if (time_before(start, dql->last_reap + 1))
48 			start = dql->last_reap + 1;
49 
50 		/* Newest sample we should have already seen a completion for */
51 		end = hist_head * BITS_PER_LONG + (BITS_PER_LONG - 1);
52 
53 		/* Shrink the search space to [start, (now - start_thrs/2)] if
54 		 * `end` is beyond the stall zone
55 		 */
56 		if (time_before(now, end + stall_thrs / 2))
57 			end = now - stall_thrs / 2;
58 
59 		/* Search for the queued time in [t, end] */
60 		for (t = start; time_before_eq(t, end); t++)
61 			if (test_bit(t % (DQL_HIST_LEN * BITS_PER_LONG),
62 				     dql->history))
63 				break;
64 
65 		/* Variable t contains the time of the queue */
66 		if (!time_before_eq(t, end))
67 			goto no_stall;
68 
69 		/* The ring buffer was modified in the meantime, retry */
70 		if (hist_head != READ_ONCE(dql->history_head))
71 			goto dqs_again;
72 
73 		dql->stall_cnt++;
74 		dql->stall_max = max_t(unsigned short, dql->stall_max, now - t);
75 
76 		trace_dql_stall_detected(dql->stall_thrs, now - t,
77 					 dql->last_reap, dql->history_head,
78 					 now, dql->history);
79 	}
80 no_stall:
81 	dql->last_reap = now;
82 }
83 
84 /* Records completed count and recalculates the queue limit */
dql_completed(struct dql * dql,unsigned int count)85 void dql_completed(struct dql *dql, unsigned int count)
86 {
87 	unsigned int inprogress, prev_inprogress, limit;
88 	unsigned int ovlimit, completed, num_queued;
89 	bool all_prev_completed;
90 
91 	num_queued = READ_ONCE(dql->num_queued);
92 
93 	/* Can't complete more than what's in queue */
94 	BUG_ON(count > num_queued - dql->num_completed);
95 
96 	completed = dql->num_completed + count;
97 	limit = dql->limit;
98 	ovlimit = POSDIFF(num_queued - dql->num_completed, limit);
99 	inprogress = num_queued - completed;
100 	prev_inprogress = dql->prev_num_queued - dql->num_completed;
101 	all_prev_completed = AFTER_EQ(completed, dql->prev_num_queued);
102 
103 	if ((ovlimit && !inprogress) ||
104 	    (dql->prev_ovlimit && all_prev_completed)) {
105 		/*
106 		 * Queue considered starved if:
107 		 *   - The queue was over-limit in the last interval,
108 		 *     and there is no more data in the queue.
109 		 *  OR
110 		 *   - The queue was over-limit in the previous interval and
111 		 *     when enqueuing it was possible that all queued data
112 		 *     had been consumed.  This covers the case when queue
113 		 *     may have becomes starved between completion processing
114 		 *     running and next time enqueue was scheduled.
115 		 *
116 		 *     When queue is starved increase the limit by the amount
117 		 *     of bytes both sent and completed in the last interval,
118 		 *     plus any previous over-limit.
119 		 */
120 		limit += POSDIFF(completed, dql->prev_num_queued) +
121 		     dql->prev_ovlimit;
122 		dql->slack_start_time = jiffies;
123 		dql->lowest_slack = UINT_MAX;
124 	} else if (inprogress && prev_inprogress && !all_prev_completed) {
125 		/*
126 		 * Queue was not starved, check if the limit can be decreased.
127 		 * A decrease is only considered if the queue has been busy in
128 		 * the whole interval (the check above).
129 		 *
130 		 * If there is slack, the amount of excess data queued above
131 		 * the amount needed to prevent starvation, the queue limit
132 		 * can be decreased.  To avoid hysteresis we consider the
133 		 * minimum amount of slack found over several iterations of the
134 		 * completion routine.
135 		 */
136 		unsigned int slack, slack_last_objs;
137 
138 		/*
139 		 * Slack is the maximum of
140 		 *   - The queue limit plus previous over-limit minus twice
141 		 *     the number of objects completed.  Note that two times
142 		 *     number of completed bytes is a basis for an upper bound
143 		 *     of the limit.
144 		 *   - Portion of objects in the last queuing operation that
145 		 *     was not part of non-zero previous over-limit.  That is
146 		 *     "round down" by non-overlimit portion of the last
147 		 *     queueing operation.
148 		 */
149 		slack = POSDIFF(limit + dql->prev_ovlimit,
150 		    2 * (completed - dql->num_completed));
151 		slack_last_objs = dql->prev_ovlimit ?
152 		    POSDIFF(dql->prev_last_obj_cnt, dql->prev_ovlimit) : 0;
153 
154 		slack = max(slack, slack_last_objs);
155 
156 		if (slack < dql->lowest_slack)
157 			dql->lowest_slack = slack;
158 
159 		if (time_after(jiffies,
160 			       dql->slack_start_time + dql->slack_hold_time)) {
161 			limit = POSDIFF(limit, dql->lowest_slack);
162 			dql->slack_start_time = jiffies;
163 			dql->lowest_slack = UINT_MAX;
164 		}
165 	}
166 
167 	/* Enforce bounds on limit */
168 	limit = clamp(limit, dql->min_limit, dql->max_limit);
169 
170 	if (limit != dql->limit) {
171 		dql->limit = limit;
172 		ovlimit = 0;
173 	}
174 
175 	dql->adj_limit = limit + completed;
176 	dql->prev_ovlimit = ovlimit;
177 	dql->prev_last_obj_cnt = dql->last_obj_cnt;
178 	dql->num_completed = completed;
179 	dql->prev_num_queued = num_queued;
180 
181 	dql_check_stall(dql);
182 }
183 EXPORT_SYMBOL(dql_completed);
184 
dql_reset(struct dql * dql)185 void dql_reset(struct dql *dql)
186 {
187 	/* Reset all dynamic values */
188 	dql->limit = 0;
189 	dql->num_queued = 0;
190 	dql->num_completed = 0;
191 	dql->last_obj_cnt = 0;
192 	dql->prev_num_queued = 0;
193 	dql->prev_last_obj_cnt = 0;
194 	dql->prev_ovlimit = 0;
195 	dql->lowest_slack = UINT_MAX;
196 	dql->slack_start_time = jiffies;
197 
198 	dql->last_reap = jiffies;
199 	dql->history_head = jiffies / BITS_PER_LONG;
200 	memset(dql->history, 0, sizeof(dql->history));
201 }
202 EXPORT_SYMBOL(dql_reset);
203 
dql_init(struct dql * dql,unsigned int hold_time)204 void dql_init(struct dql *dql, unsigned int hold_time)
205 {
206 	dql->max_limit = DQL_MAX_LIMIT;
207 	dql->min_limit = 0;
208 	dql->slack_hold_time = hold_time;
209 	dql->stall_thrs = 0;
210 	dql_reset(dql);
211 }
212 EXPORT_SYMBOL(dql_init);
213