+static void
+metaslab_aux_histograms_clear(metaslab_t *msp)
+{
+ /*
+ * Auxiliary histograms are only cleared when resetting them,
+ * which can only happen while the metaslab is loaded.
+ */
+ ASSERT(msp->ms_loaded);
+
+ bzero(msp->ms_synchist, sizeof (msp->ms_synchist));
+ for (int t = 0; t < TXG_DEFER_SIZE; t++)
+ bzero(msp->ms_deferhist[t], sizeof (msp->ms_deferhist[t]));
+}
+
+static void
+metaslab_aux_histogram_add(uint64_t *histogram, uint64_t shift,
+ range_tree_t *rt)
+{
+ /*
+ * This is modeled after space_map_histogram_add(), so refer to that
+ * function for implementation details. We want this to work like
+ * the space map histogram, and not the range tree histogram, as we
+ * are essentially constructing a delta that will be later subtracted
+ * from the space map histogram.
+ */
+ int idx = 0;
+ for (int i = shift; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
+ ASSERT3U(i, >=, idx + shift);
+ histogram[idx] += rt->rt_histogram[i] << (i - idx - shift);
+
+ if (idx < SPACE_MAP_HISTOGRAM_SIZE - 1) {
+ ASSERT3U(idx + shift, ==, i);
+ idx++;
+ ASSERT3U(idx, <, SPACE_MAP_HISTOGRAM_SIZE);
+ }
+ }
+}
+
+/*
+ * Called at every sync pass that the metaslab gets synced.
+ *
+ * The reason is that we want our auxiliary histograms to be updated
+ * wherever the metaslab's space map histogram is updated. This way
+ * we stay consistent on which parts of the metaslab space map's
+ * histogram are currently not available for allocations (e.g because
+ * they are in the defer, freed, and freeing trees).
+ */
+static void
+metaslab_aux_histograms_update(metaslab_t *msp)
+{
+ space_map_t *sm = msp->ms_sm;
+ ASSERT(sm != NULL);
+
+ /*
+ * This is similar to the metaslab's space map histogram updates
+ * that take place in metaslab_sync(). The only difference is that
+ * we only care about segments that haven't made it into the
+ * ms_allocatable tree yet.
+ */
+ if (msp->ms_loaded) {
+ metaslab_aux_histograms_clear(msp);
+
+ metaslab_aux_histogram_add(msp->ms_synchist,
+ sm->sm_shift, msp->ms_freed);
+
+ for (int t = 0; t < TXG_DEFER_SIZE; t++) {
+ metaslab_aux_histogram_add(msp->ms_deferhist[t],
+ sm->sm_shift, msp->ms_defer[t]);
+ }
+ }
+
+ metaslab_aux_histogram_add(msp->ms_synchist,
+ sm->sm_shift, msp->ms_freeing);
+}
+
+/*
+ * Called every time we are done syncing (writing to) the metaslab,
+ * i.e. at the end of each sync pass.
+ * [see the comment in metaslab_impl.h for ms_synchist, ms_deferhist]
+ */
+static void
+metaslab_aux_histograms_update_done(metaslab_t *msp, boolean_t defer_allowed)
+{
+ spa_t *spa = msp->ms_group->mg_vd->vdev_spa;
+ space_map_t *sm = msp->ms_sm;
+
+ if (sm == NULL) {
+ /*
+ * We came here from metaslab_init() when creating/opening a
+ * pool, looking at a metaslab that hasn't had any allocations
+ * yet.
+ */
+ return;
+ }
+
+ /*
+ * This is similar to the actions that we take for the ms_freed
+ * and ms_defer trees in metaslab_sync_done().
+ */
+ uint64_t hist_index = spa_syncing_txg(spa) % TXG_DEFER_SIZE;
+ if (defer_allowed) {
+ bcopy(msp->ms_synchist, msp->ms_deferhist[hist_index],
+ sizeof (msp->ms_synchist));
+ } else {
+ bzero(msp->ms_deferhist[hist_index],
+ sizeof (msp->ms_deferhist[hist_index]));
+ }
+ bzero(msp->ms_synchist, sizeof (msp->ms_synchist));
+}
+
+/*
+ * Ensure that the metaslab's weight and fragmentation are consistent
+ * with the contents of the histogram (either the range tree's histogram
+ * or the space map's depending whether the metaslab is loaded).
+ */
+static void
+metaslab_verify_weight_and_frag(metaslab_t *msp)
+{
+ ASSERT(MUTEX_HELD(&msp->ms_lock));
+
+ if ((zfs_flags & ZFS_DEBUG_METASLAB_VERIFY) == 0)
+ return;
+
+ /* see comment in metaslab_verify_unflushed_changes() */
+ if (msp->ms_group == NULL)
+ return;
+
+ /*
+ * Devices being removed always return a weight of 0 and leave
+ * fragmentation and ms_max_size as is - there is nothing for
+ * us to verify here.
+ */
+ vdev_t *vd = msp->ms_group->mg_vd;
+ if (vd->vdev_removing)
+ return;
+
+ /*
+ * If the metaslab is dirty it probably means that we've done
+ * some allocations or frees that have changed our histograms
+ * and thus the weight.
+ */
+ for (int t = 0; t < TXG_SIZE; t++) {
+ if (txg_list_member(&vd->vdev_ms_list, msp, t))
+ return;
+ }
+
+ /*
+ * This verification checks that our in-memory state is consistent
+ * with what's on disk. If the pool is read-only then there aren't
+ * any changes and we just have the initially-loaded state.
+ */
+ if (!spa_writeable(msp->ms_group->mg_vd->vdev_spa))
+ return;
+
+ /* some extra verification for in-core tree if you can */
+ if (msp->ms_loaded) {
+ range_tree_stat_verify(msp->ms_allocatable);
+ VERIFY(space_map_histogram_verify(msp->ms_sm,
+ msp->ms_allocatable));
+ }
+
+ uint64_t weight = msp->ms_weight;
+ uint64_t was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
+ boolean_t space_based = WEIGHT_IS_SPACEBASED(msp->ms_weight);
+ uint64_t frag = msp->ms_fragmentation;
+ uint64_t max_segsize = msp->ms_max_size;
+
+ msp->ms_weight = 0;
+ msp->ms_fragmentation = 0;
+ msp->ms_max_size = 0;
+
+ /*
+ * This function is used for verification purposes. Regardless of
+ * whether metaslab_weight() thinks this metaslab should be active or
+ * not, we want to ensure that the actual weight (and therefore the
+ * value of ms_weight) would be the same if it was to be recalculated
+ * at this point.
+ */
+ msp->ms_weight = metaslab_weight(msp) | was_active;
+
+ VERIFY3U(max_segsize, ==, msp->ms_max_size);
+
+ /*
+ * If the weight type changed then there is no point in doing
+ * verification. Revert fields to their original values.
+ */
+ if ((space_based && !WEIGHT_IS_SPACEBASED(msp->ms_weight)) ||
+ (!space_based && WEIGHT_IS_SPACEBASED(msp->ms_weight))) {
+ msp->ms_fragmentation = frag;
+ msp->ms_weight = weight;
+ return;
+ }
+
+ VERIFY3U(msp->ms_fragmentation, ==, frag);
+ VERIFY3U(msp->ms_weight, ==, weight);
+}
+