xref: /linux/lib/win_minmax.c (revision f3a8b6645dc2e60d11f20c1c23afd964ff4e55ae)
1 /**
2  * lib/minmax.c: windowed min/max tracker
3  *
4  * Kathleen Nichols' algorithm for tracking the minimum (or maximum)
5  * value of a data stream over some fixed time interval.  (E.g.,
6  * the minimum RTT over the past five minutes.) It uses constant
7  * space and constant time per update yet almost always delivers
8  * the same minimum as an implementation that has to keep all the
9  * data in the window.
10  *
11  * The algorithm keeps track of the best, 2nd best & 3rd best min
12  * values, maintaining an invariant that the measurement time of
13  * the n'th best >= n-1'th best. It also makes sure that the three
14  * values are widely separated in the time window since that bounds
15  * the worse case error when that data is monotonically increasing
16  * over the window.
17  *
18  * Upon getting a new min, we can forget everything earlier because
19  * it has no value - the new min is <= everything else in the window
20  * by definition and it's the most recent. So we restart fresh on
21  * every new min and overwrites 2nd & 3rd choices. The same property
22  * holds for 2nd & 3rd best.
23  */
24 #include <linux/module.h>
25 #include <linux/win_minmax.h>
26 
27 /* As time advances, update the 1st, 2nd, and 3rd choices. */
28 static u32 minmax_subwin_update(struct minmax *m, u32 win,
29 				const struct minmax_sample *val)
30 {
31 	u32 dt = val->t - m->s[0].t;
32 
33 	if (unlikely(dt > win)) {
34 		/*
35 		 * Passed entire window without a new val so make 2nd
36 		 * choice the new val & 3rd choice the new 2nd choice.
37 		 * we may have to iterate this since our 2nd choice
38 		 * may also be outside the window (we checked on entry
39 		 * that the third choice was in the window).
40 		 */
41 		m->s[0] = m->s[1];
42 		m->s[1] = m->s[2];
43 		m->s[2] = *val;
44 		if (unlikely(val->t - m->s[0].t > win)) {
45 			m->s[0] = m->s[1];
46 			m->s[1] = m->s[2];
47 			m->s[2] = *val;
48 		}
49 	} else if (unlikely(m->s[1].t == m->s[0].t) && dt > win/4) {
50 		/*
51 		 * We've passed a quarter of the window without a new val
52 		 * so take a 2nd choice from the 2nd quarter of the window.
53 		 */
54 		m->s[2] = m->s[1] = *val;
55 	} else if (unlikely(m->s[2].t == m->s[1].t) && dt > win/2) {
56 		/*
57 		 * We've passed half the window without finding a new val
58 		 * so take a 3rd choice from the last half of the window
59 		 */
60 		m->s[2] = *val;
61 	}
62 	return m->s[0].v;
63 }
64 
65 /* Check if new measurement updates the 1st, 2nd or 3rd choice max. */
66 u32 minmax_running_max(struct minmax *m, u32 win, u32 t, u32 meas)
67 {
68 	struct minmax_sample val = { .t = t, .v = meas };
69 
70 	if (unlikely(val.v >= m->s[0].v) ||	  /* found new max? */
71 	    unlikely(val.t - m->s[2].t > win))	  /* nothing left in window? */
72 		return minmax_reset(m, t, meas);  /* forget earlier samples */
73 
74 	if (unlikely(val.v >= m->s[1].v))
75 		m->s[2] = m->s[1] = val;
76 	else if (unlikely(val.v >= m->s[2].v))
77 		m->s[2] = val;
78 
79 	return minmax_subwin_update(m, win, &val);
80 }
81 EXPORT_SYMBOL(minmax_running_max);
82 
83 /* Check if new measurement updates the 1st, 2nd or 3rd choice min. */
84 u32 minmax_running_min(struct minmax *m, u32 win, u32 t, u32 meas)
85 {
86 	struct minmax_sample val = { .t = t, .v = meas };
87 
88 	if (unlikely(val.v <= m->s[0].v) ||	  /* found new min? */
89 	    unlikely(val.t - m->s[2].t > win))	  /* nothing left in window? */
90 		return minmax_reset(m, t, meas);  /* forget earlier samples */
91 
92 	if (unlikely(val.v <= m->s[1].v))
93 		m->s[2] = m->s[1] = val;
94 	else if (unlikely(val.v <= m->s[2].v))
95 		m->s[2] = val;
96 
97 	return minmax_subwin_update(m, win, &val);
98 }
99