libceph: support for balanced and localized reads

OSD-side issues with reads from replica have been resolved in
Octopus. Reading from replica should be safe wrt. unstable or
uncommitted state now, so add support for balanced and localized
reads.

There are two cases when a read from replica can't be served:

- OSD may silently drop the request, expecting the client to
notice that the acting set has changed and resend via the usual
means (handled with t->used_replica)

- OSD may return EAGAIN, expecting the client to resend to the
primary, ignoring replica read flags (see handle_reply())

Signed-off-by: Ilya Dryomov <[e-mail]>
Reviewed-by: Jeff Layton <[e-mail]>

include/linux/ceph/osd_client.h

···
 	bool recovery_deletes;
 
 	unsigned int flags;                /* CEPH_OSD_FLAG_* */
+	bool used_replica;
 	bool paused;
 
 	u32 epoch;

include/linux/ceph/osdmap.h

···
 int ceph_compare_crush_locs(struct rb_root *locs1, struct rb_root *locs2);
 void ceph_clear_crush_locs(struct rb_root *locs);
 
+int ceph_get_crush_locality(struct ceph_osdmap *osdmap, int id,
+			    struct rb_root *locs);
+
 extern struct ceph_pg_pool_info *ceph_pg_pool_by_id(struct ceph_osdmap *map,
 						    u64 id);
 extern const char *ceph_pg_pool_name_by_id(struct ceph_osdmap *map, u64 id);

net/ceph/debugfs.c

···
 		u32 state = map->osd_state[i];
 		char sb[64];
 
-		seq_printf(s, "osd%d\t%s\t%3d%%\t(%s)\t%3d%%\n",
+		seq_printf(s, "osd%d\t%s\t%3d%%\t(%s)\t%3d%%\t%2d\n",
 			   i, ceph_pr_addr(addr),
 			   ((map->osd_weight[i]*100) >> 16),
 			   ceph_osdmap_state_str(sb, sizeof(sb), state),
-			   ((ceph_get_primary_affinity(map, i)*100) >> 16));
+			   ((ceph_get_primary_affinity(map, i)*100) >> 16),
+			   ceph_get_crush_locality(map, i,
+					   &client->options->crush_locs));
 	}
 	for (n = rb_first(&map->pg_temp); n; n = rb_next(n)) {
 		struct ceph_pg_mapping *pg =

net/ceph/osd_client.c

···
 	       (osdc->osdmap->epoch < osdc->epoch_barrier);
 }
 
+static int pick_random_replica(const struct ceph_osds *acting)
+{
+	int i = prandom_u32() % acting->size;
+
+	dout("%s picked osd%d, primary osd%d\n", __func__,
+	     acting->osds[i], acting->primary);
+	return i;
+}
+
+/*
+ * Picks the closest replica based on client's location given by
+ * crush_location option.  Prefers the primary if the locality is
+ * the same.
+ */
+static int pick_closest_replica(struct ceph_osd_client *osdc,
+				const struct ceph_osds *acting)
+{
+	struct ceph_options *opt = osdc->client->options;
+	int best_i, best_locality;
+	int i = 0, locality;
+
+	do {
+		locality = ceph_get_crush_locality(osdc->osdmap,
+						   acting->osds[i],
+						   &opt->crush_locs);
+		if (i == 0 ||
+		    (locality >= 0 && best_locality < 0) ||
+		    (locality >= 0 && best_locality >= 0 &&
+		     locality < best_locality)) {
+			best_i = i;
+			best_locality = locality;
+		}
+	} while (++i < acting->size);
+
+	dout("%s picked osd%d with locality %d, primary osd%d\n", __func__,
+	     acting->osds[best_i], best_locality, acting->primary);
+	return best_i;
+}
+
 enum calc_target_result {
 	CALC_TARGET_NO_ACTION = 0,
 	CALC_TARGET_NEED_RESEND,
···
 	struct ceph_pg_pool_info *pi;
 	struct ceph_pg pgid, last_pgid;
 	struct ceph_osds up, acting;
+	bool is_read = t->flags & CEPH_OSD_FLAG_READ;
+	bool is_write = t->flags & CEPH_OSD_FLAG_WRITE;
 	bool force_resend = false;
 	bool unpaused = false;
 	bool legacy_change = false;
···
 	ceph_oid_copy(&t->target_oid, &t->base_oid);
 	ceph_oloc_copy(&t->target_oloc, &t->base_oloc);
 	if ((t->flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) {
-		if (t->flags & CEPH_OSD_FLAG_READ && pi->read_tier >= 0)
+		if (is_read && pi->read_tier >= 0)
 			t->target_oloc.pool = pi->read_tier;
-		if (t->flags & CEPH_OSD_FLAG_WRITE && pi->write_tier >= 0)
+		if (is_write && pi->write_tier >= 0)
 			t->target_oloc.pool = pi->write_tier;
 
 		pi = ceph_pg_pool_by_id(osdc->osdmap, t->target_oloc.pool);
···
 		unpaused = true;
 	}
 	legacy_change = ceph_pg_compare(&t->pgid, &pgid) ||
-			ceph_osds_changed(&t->acting, &acting, any_change);
+			ceph_osds_changed(&t->acting, &acting,
+					  t->used_replica || any_change);
 	if (t->pg_num)
 		split = ceph_pg_is_split(&last_pgid, t->pg_num, pi->pg_num);
 
···
 		t->sort_bitwise = sort_bitwise;
 		t->recovery_deletes = recovery_deletes;
 
-		t->osd = acting.primary;
+		if ((t->flags & (CEPH_OSD_FLAG_BALANCE_READS |
+				 CEPH_OSD_FLAG_LOCALIZE_READS)) &&
+		    !is_write && pi->type == CEPH_POOL_TYPE_REP &&
+		    acting.size > 1) {
+			int pos;
+
+			WARN_ON(!is_read || acting.osds[0] != acting.primary);
+			if (t->flags & CEPH_OSD_FLAG_BALANCE_READS) {
+				pos = pick_random_replica(&acting);
+			} else {
+				pos = pick_closest_replica(osdc, &acting);
+			}
+			t->osd = acting.osds[pos];
+			t->used_replica = pos > 0;
+		} else {
+			t->osd = acting.primary;
+			t->used_replica = false;
+		}
 	}
 
 	if (unpaused || legacy_change || force_resend || split)
···
 		req->r_flags |= CEPH_OSD_FLAG_REDIRECTED |
 				CEPH_OSD_FLAG_IGNORE_OVERLAY |
 				CEPH_OSD_FLAG_IGNORE_CACHE;
+		req->r_tid = 0;
+		__submit_request(req, false);
+		goto out_unlock_osdc;
+	}
+
+	if (m.result == -EAGAIN) {
+		dout("req %p tid %llu EAGAIN\n", req, req->r_tid);
+		unlink_request(osd, req);
+		mutex_unlock(&osd->lock);
+
+		/*
+		 * The object is missing on the replica or not (yet)
+		 * readable.  Clear pgid to force a resend to the primary
+		 * via legacy_change.
+		 */
+		req->r_t.pgid.pool = 0;
+		req->r_t.pgid.seed = 0;
+		WARN_ON(!req->r_t.used_replica);
+		req->r_flags &= ~(CEPH_OSD_FLAG_BALANCE_READS |
+				  CEPH_OSD_FLAG_LOCALIZE_READS);
 		req->r_tid = 0;
 		__submit_request(req, false);
 		goto out_unlock_osdc;

net/ceph/osdmap.c

···
 		free_crush_loc(loc);
 	}
 }
+
+/*
+ * [a-zA-Z0-9-_.]+
+ */
+static bool is_valid_crush_name(const char *name)
+{
+	do {
+		if (!('a' <= *name && *name <= 'z') &&
+		    !('A' <= *name && *name <= 'Z') &&
+		    !('0' <= *name && *name <= '9') &&
+		    *name != '-' && *name != '_' && *name != '.')
+			return false;
+	} while (*++name != '\0');
+
+	return true;
+}
+
+/*
+ * Gets the parent of an item.  Returns its id (<0 because the
+ * parent is always a bucket), type id (>0 for the same reason,
+ * via @parent_type_id) and location (via @parent_loc).  If no
+ * parent, returns 0.
+ *
+ * Does a linear search, as there are no parent pointers of any
+ * kind.  Note that the result is ambigous for items that occur
+ * multiple times in the map.
+ */
+static int get_immediate_parent(struct crush_map *c, int id,
+				u16 *parent_type_id,
+				struct crush_loc *parent_loc)
+{
+	struct crush_bucket *b;
+	struct crush_name_node *type_cn, *cn;
+	int i, j;
+
+	for (i = 0; i < c->max_buckets; i++) {
+		b = c->buckets[i];
+		if (!b)
+			continue;
+
+		/* ignore per-class shadow hierarchy */
+		cn = lookup_crush_name(&c->names, b->id);
+		if (!cn || !is_valid_crush_name(cn->cn_name))
+			continue;
+
+		for (j = 0; j < b->size; j++) {
+			if (b->items[j] != id)
+				continue;
+
+			*parent_type_id = b->type;
+			type_cn = lookup_crush_name(&c->type_names, b->type);
+			parent_loc->cl_type_name = type_cn->cn_name;
+			parent_loc->cl_name = cn->cn_name;
+			return b->id;
+		}
+	}
+
+	return 0;  /* no parent */
+}
+
+/*
+ * Calculates the locality/distance from an item to a client
+ * location expressed in terms of CRUSH hierarchy as a set of
+ * (bucket type name, bucket name) pairs.  Specifically, looks
+ * for the lowest-valued bucket type for which the location of
+ * @id matches one of the locations in @locs, so for standard
+ * bucket types (host = 1, rack = 3, datacenter = 8, zone = 9)
+ * a matching host is closer than a matching rack and a matching
+ * data center is closer than a matching zone.
+ *
+ * Specifying multiple locations (a "multipath" location) such
+ * as "rack=foo1 rack=foo2 datacenter=bar" is allowed -- @locs
+ * is a multimap.  The locality will be:
+ *
+ * - 3 for OSDs in racks foo1 and foo2
+ * - 8 for OSDs in data center bar
+ * - -1 for all other OSDs
+ *
+ * The lowest possible bucket type is 1, so the best locality
+ * for an OSD is 1 (i.e. a matching host).  Locality 0 would be
+ * the OSD itself.
+ */
+int ceph_get_crush_locality(struct ceph_osdmap *osdmap, int id,
+			    struct rb_root *locs)
+{
+	struct crush_loc loc;
+	u16 type_id;
+
+	/*
+	 * Instead of repeated get_immediate_parent() calls,
+	 * the location of @id could be obtained with a single
+	 * depth-first traversal.
+	 */
+	for (;;) {
+		id = get_immediate_parent(osdmap->crush, id, &type_id, &loc);
+		if (id >= 0)
+			return -1;  /* not local */
+
+		if (lookup_crush_loc(locs, &loc))
+			return type_id;
+	}
+}