--- /dev/null
+#ifndef _HAPROXY_WINDOW_FILTER_H
+#define _HAPROXY_WINDOW_FILTER_H
+
+/* Kathleen Nichols' algorithm to track the maximum values of a data type during
+ * a fixed time interval. This algorithm makes usage of three samples to track
+ * the best, second best and third best values with 1st >= 2nd >= 3rd as
+ * invariant.
+ *
+ * This code is used in Linux kernel in linux/win_minmax.c to track both
+ * minimal and maximum values.
+ *
+ * Here the code has been adapted to track 64 bits values and only their
+ * maximum.
+ *
+ * Note that these windowed filters are used by BBR to filter the maximum
+ * estimated bandwidth with counters as time values. A length has been
+ * added to simulate the fixed time interval with counter which are
+ * monotonically increasing.
+ */
+
+/* Windowed filter sample */
+struct wf_smp {
+ uint64_t v;
+ uint32_t t;
+};
+
+/* Windowed filter */
+struct wf {
+ size_t len;
+ struct wf_smp smp[3];
+};
+
+/* Reset all the <wf> windowed filter samples with <v> and <t> as value and
+ * time value respectively.
+ */
+static inline uint64_t wf_reset(struct wf *wf, uint64_t v, uint32_t t)
+{
+ struct wf_smp smp = { .v = v, .t = t };
+
+ wf->smp[2] = wf->smp[1] = wf->smp[0] = smp;
+
+ return wf->smp[0].v;
+}
+
+/* Initialize <wf> windowed filter to track maximum values, with <len> as
+ * length and <v> and <t> as value and time value respectively.
+ */
+static inline void wf_init(struct wf *wf, size_t len, uint64_t v, uint32_t t)
+{
+ wf->len = len;
+ wf_reset(wf, v, t);
+}
+
+/* Similar to minmax_running_max() Linux kernel function to update the best
+ * estimation of <wf> windowed filted with <v> and <t> as value and time value
+ * respectively
+ */
+static inline uint64_t wf_max_update(struct wf *wf, uint64_t v, uint32_t t)
+{
+ uint64_t delta_t;
+ struct wf_smp smp = { .v = v, .t = t };
+
+ /* Reset all estimates if they have not yet been initialized, if new
+ sample is a new best, or if the newest recorded estimate is too
+ old. */
+ if (unlikely(v > wf->smp[0].v) || unlikely(t - wf->smp[2].t > wf->len))
+ return wf_reset(wf, v, t);
+
+ if (unlikely(v > wf->smp[1].v))
+ wf->smp[2] = wf->smp[1] = smp;
+ else if (unlikely(v > wf->smp[2].v))
+ wf->smp[2] = smp;
+
+ delta_t = t - wf->smp[0].t;
+ /* From here, similar to minmax_subwin_update() from Linux kernel. */
+ if (unlikely(delta_t > wf->len)) {
+ wf->smp[0] = wf->smp[1];
+ wf->smp[1] = wf->smp[2];
+ wf->smp[2].v = v;
+ wf->smp[2].t = t;
+
+ if (unlikely(t - wf->smp[0].t > wf->len)) {
+ wf->smp[0] = wf->smp[1];
+ wf->smp[1] = wf->smp[2];
+ }
+ } else if (unlikely(wf->smp[1].v == wf->smp[0].v) && delta_t > wf->len / 4) {
+ wf->smp[2].v = v;
+ wf->smp[2].t = t;
+ wf->smp[1] = wf->smp[2];
+ } else if (unlikely(wf->smp[2].v == wf->smp[1].v) && delta_t > wf->len / 2) {
+ wf->smp[2].v = v;
+ wf->smp[2].t = t;
+ }
+
+ return wf->smp[0].v;
+}
+
+/* Return <wf> windowed filter best maximum estimation. */
+static inline uint64_t wf_get_max(struct wf *wf)
+{
+ return wf->smp[0].v;
+}
+
+#endif /* _HAPROXY_WINDOW_FILTER_H */