lib/win_minmax.c

   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 }