1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* |
3 | * lib/minmax.c: windowed min/max tracker |
4 | * |
5 | * Kathleen Nichols' algorithm for tracking the minimum (or maximum) |
6 | * value of a data stream over some fixed time interval. (E.g., |
7 | * the minimum RTT over the past five minutes.) It uses constant |
8 | * space and constant time per update yet almost always delivers |
9 | * the same minimum as an implementation that has to keep all the |
10 | * data in the window. |
11 | * |
12 | * The algorithm keeps track of the best, 2nd best & 3rd best min |
13 | * values, maintaining an invariant that the measurement time of |
14 | * the n'th best >= n-1'th best. It also makes sure that the three |
15 | * values are widely separated in the time window since that bounds |
16 | * the worse case error when that data is monotonically increasing |
17 | * over the window. |
18 | * |
19 | * Upon getting a new min, we can forget everything earlier because |
20 | * it has no value - the new min is <= everything else in the window |
21 | * by definition and it's the most recent. So we restart fresh on |
22 | * every new min and overwrites 2nd & 3rd choices. The same property |
23 | * holds for 2nd & 3rd best. |
24 | */ |
25 | #include <linux/module.h> |
26 | #include <linux/win_minmax.h> |
27 | |
28 | /* As time advances, update the 1st, 2nd, and 3rd choices. */ |
29 | static u32 minmax_subwin_update(struct minmax *m, u32 win, |
30 | const struct minmax_sample *val) |
31 | { |
32 | u32 dt = val->t - m->s[0].t; |
33 | |
34 | if (unlikely(dt > win)) { |
35 | /* |
36 | * Passed entire window without a new val so make 2nd |
37 | * choice the new val & 3rd choice the new 2nd choice. |
38 | * we may have to iterate this since our 2nd choice |
39 | * may also be outside the window (we checked on entry |
40 | * that the third choice was in the window). |
41 | */ |
42 | m->s[0] = m->s[1]; |
43 | m->s[1] = m->s[2]; |
44 | m->s[2] = *val; |
45 | if (unlikely(val->t - m->s[0].t > win)) { |
46 | m->s[0] = m->s[1]; |
47 | m->s[1] = m->s[2]; |
48 | m->s[2] = *val; |
49 | } |
50 | } else if (unlikely(m->s[1].t == m->s[0].t) && dt > win/4) { |
51 | /* |
52 | * We've passed a quarter of the window without a new val |
53 | * so take a 2nd choice from the 2nd quarter of the window. |
54 | */ |
55 | m->s[2] = m->s[1] = *val; |
56 | } else if (unlikely(m->s[2].t == m->s[1].t) && dt > win/2) { |
57 | /* |
58 | * We've passed half the window without finding a new val |
59 | * so take a 3rd choice from the last half of the window |
60 | */ |
61 | m->s[2] = *val; |
62 | } |
63 | return m->s[0].v; |
64 | } |
65 | |
66 | /* Check if new measurement updates the 1st, 2nd or 3rd choice max. */ |
67 | u32 minmax_running_max(struct minmax *m, u32 win, u32 t, u32 meas) |
68 | { |
69 | struct minmax_sample val = { .t = t, .v = meas }; |
70 | |
71 | if (unlikely(val.v >= m->s[0].v) || /* found new max? */ |
72 | unlikely(val.t - m->s[2].t > win)) /* nothing left in window? */ |
73 | return minmax_reset(m, t, meas); /* forget earlier samples */ |
74 | |
75 | if (unlikely(val.v >= m->s[1].v)) |
76 | m->s[2] = m->s[1] = val; |
77 | else if (unlikely(val.v >= m->s[2].v)) |
78 | m->s[2] = val; |
79 | |
80 | return minmax_subwin_update(m, win, val: &val); |
81 | } |
82 | EXPORT_SYMBOL(minmax_running_max); |
83 | |
84 | /* Check if new measurement updates the 1st, 2nd or 3rd choice min. */ |
85 | u32 minmax_running_min(struct minmax *m, u32 win, u32 t, u32 meas) |
86 | { |
87 | struct minmax_sample val = { .t = t, .v = meas }; |
88 | |
89 | if (unlikely(val.v <= m->s[0].v) || /* found new min? */ |
90 | unlikely(val.t - m->s[2].t > win)) /* nothing left in window? */ |
91 | return minmax_reset(m, t, meas); /* forget earlier samples */ |
92 | |
93 | if (unlikely(val.v <= m->s[1].v)) |
94 | m->s[2] = m->s[1] = val; |
95 | else if (unlikely(val.v <= m->s[2].v)) |
96 | m->s[2] = val; |
97 | |
98 | return minmax_subwin_update(m, win, val: &val); |
99 | } |
100 | |