core: run many bricks within one glusterfsd process

This patch adds support for multiple brick translator stacks running
in a single brick server process.  This reduces our per-brick memory usage by
approximately 3x, and our appetite for TCP ports even more.  It also creates
potential to avoid process/thread thrashing, and to improve QoS by scheduling
more carefully across the bricks, but realizing that potential will require
further work.

Multiplexing is controlled by the "cluster.brick-multiplex" global option.  By
default it's off, and bricks are started in separate processes as before.  If
multiplexing is enabled, then *compatible* bricks (mostly those with the same
transport options) will be started in the same process.

Change-Id: I45059454e51d6f4cbb29a4953359c09a408695cb
BUG: 1385758
Signed-off-by: Jeff Darcy <jdarcy@redhat.com>
Reviewed-on: https://review.gluster.org/14763
Smoke: Gluster Build System <jenkins@build.gluster.org>
NetBSD-regression: NetBSD Build System <jenkins@build.gluster.org>
CentOS-regression: Gluster Build System <jenkins@build.gluster.org>
Reviewed-by: Vijay Bellur <vbellur@redhat.com>

9 files changed, 261 insertions, 35 deletions

diff --git a/libglusterfs/src/client_t.c b/libglusterfs/src/client_t.c
index b3eb4e4df8c..c20c4089ec3 100644
--- a/libglusterfs/src/client_t.c
+++ b/libglusterfs/src/client_t.c
@@ -331,11 +331,25 @@ gf_client_ref (client_t *client)
 
 
 static void
+gf_client_destroy_recursive (xlator_t *xl, client_t *client)
+{
+        xlator_list_t   *trav;
+
+        if (xl->cbks->client_destroy) {
+                xl->cbks->client_destroy (xl, client);
+        }
+
+        for (trav = xl->children; trav; trav = trav->next) {
+                gf_client_destroy_recursive (trav->xlator, client);
+        }
+}
+
+
+static void
 client_destroy (client_t *client)
 {
         clienttable_t     *clienttable = NULL;
         glusterfs_graph_t *gtrav       = NULL;
-        xlator_t          *xtrav       = NULL;
 
         if (client == NULL){
                 gf_msg_callingfn ("xlator", GF_LOG_ERROR, EINVAL,
@@ -358,12 +372,7 @@ client_destroy (client_t *client)
         UNLOCK (&clienttable->lock);
 
         list_for_each_entry (gtrav, &client->this->ctx->graphs, list) {
-                xtrav = gtrav->top;
-                while (xtrav != NULL) {
-                        if (xtrav->cbks->client_destroy != NULL)
-                                xtrav->cbks->client_destroy (xtrav, client);
-                        xtrav = xtrav->next;
-                }
+                gf_client_destroy_recursive (gtrav->top, client);
         }
         GF_FREE (client->auth.data);
         GF_FREE (client->auth.username);
@@ -375,22 +384,32 @@ out:
         return;
 }
 
+static int
+gf_client_disconnect_recursive (xlator_t *xl, client_t *client)
+{
+        int             ret     = 0;
+        xlator_list_t   *trav;
+
+        if (xl->cbks->client_disconnect) {
+                ret = xl->cbks->client_disconnect (xl, client);
+        }
+
+        for (trav = xl->children; trav; trav = trav->next) {
+                ret |= gf_client_disconnect_recursive (trav->xlator, client);
+        }
+
+        return ret;
+}
+
 
 int
 gf_client_disconnect (client_t *client)
 {
         int                ret   = 0;
         glusterfs_graph_t *gtrav = NULL;
-        xlator_t          *xtrav = NULL;
 
         list_for_each_entry (gtrav, &client->this->ctx->graphs, list) {
-                xtrav = gtrav->top;
-                while (xtrav != NULL) {
-                        if (xtrav->cbks->client_disconnect != NULL)
-                                if (xtrav->cbks->client_disconnect (xtrav, client) != 0)
-                                        ret = -1;
-                        xtrav = xtrav->next;
-                }
+                ret |= gf_client_disconnect_recursive (gtrav->top, client);
         }
 
         return ret;
diff --git a/libglusterfs/src/common-utils.c b/libglusterfs/src/common-utils.c
index 0486409a849..e180dd3eec0 100644
--- a/libglusterfs/src/common-utils.c
+++ b/libglusterfs/src/common-utils.c
@@ -3646,15 +3646,17 @@ gf_is_service_running (char *pidfile, int *pid)
         int             fno = 0;
 
         file = fopen (pidfile, "r+");
-        if (!file)
+        if (!file) {
                 goto out;
+        }
 
         fno = fileno (file);
         ret = lockf (fno, F_TEST, 0);
         if (ret == -1)
                 running = _gf_true;
-        if (!pid)
+        if (!pid) {
                 goto out;
+        }
 
         ret = fscanf (file, "%d", pid);
         if (ret <= 0) {
@@ -3663,6 +3665,15 @@ gf_is_service_running (char *pidfile, int *pid)
                 *pid = -1;
         }
 
+        if (!*pid) {
+                /*
+                 * PID 0 means we've started the process, but it hasn't gotten
+                 * far enough to put in a real PID yet.  More details are in
+                 * glusterd_brick_start.
+                 */
+                running = _gf_true;
+        }
+
 out:
         if (file)
                 fclose (file);
diff --git a/libglusterfs/src/event-epoll.c b/libglusterfs/src/event-epoll.c
index 3fd580d9d1a..e2b40602e7a 100644
--- a/libglusterfs/src/event-epoll.c
+++ b/libglusterfs/src/event-epoll.c
@@ -263,6 +263,7 @@ event_pool_new_epoll (int count, int eventthreadcount)
         event_pool->count = count;
 
         event_pool->eventthreadcount = eventthreadcount;
+        event_pool->auto_thread_count = 0;
 
         pthread_mutex_init (&event_pool->mutex, NULL);
 
@@ -363,7 +364,7 @@ event_register_epoll (struct event_pool *event_pool, int fd,
 		   time as well.
 		*/
 
-		slot->events = EPOLLPRI | EPOLLONESHOT;
+		slot->events = EPOLLPRI | EPOLLHUP | EPOLLERR | EPOLLONESHOT;
 		slot->handler = handler;
 		slot->data = data;
 
diff --git a/libglusterfs/src/event.h b/libglusterfs/src/event.h
index b01ef24bb8e..1348f5d05c0 100644
--- a/libglusterfs/src/event.h
+++ b/libglusterfs/src/event.h
@@ -28,7 +28,7 @@ typedef int (*event_handler_t) (int fd, int idx, void *data,
 
 #define EVENT_EPOLL_TABLES 1024
 #define EVENT_EPOLL_SLOTS 1024
-#define EVENT_MAX_THREADS  32
+#define EVENT_MAX_THREADS  1024
 
 struct event_pool {
 	struct event_ops *ops;
@@ -57,6 +57,20 @@ struct event_pool {
                                                      * and live status */
         int destroy;
         int activethreadcount;
+
+        /*
+         * Number of threads created by auto-scaling, *in addition to* the
+         * configured number of threads.  This is only applicable on the
+         * server, where we try to keep the number of threads around the number
+         * of bricks.  In that case, the configured number is just "extra"
+         * threads to handle requests in excess of one per brick (including
+         * requests on the GlusterD connection).  For clients or GlusterD, this
+         * number will always be zero, so the "extra" is all we have.
+         *
+         * TBD: consider auto-scaling for clients as well
+         */
+        int auto_thread_count;
+
 };
 
 struct event_ops {
diff --git a/libglusterfs/src/glusterfs.h b/libglusterfs/src/glusterfs.h
index 0d073154934..4f1f27b5857 100644
--- a/libglusterfs/src/glusterfs.h
+++ b/libglusterfs/src/glusterfs.h
@@ -557,16 +557,19 @@ typedef struct lock_migration_info {
  */
 #define SECURE_ACCESS_FILE     GLUSTERD_DEFAULT_WORKDIR "/secure-access"
 
-int glusterfs_graph_prepare (glusterfs_graph_t *graph, glusterfs_ctx_t *ctx);
+int glusterfs_graph_prepare (glusterfs_graph_t *graph, glusterfs_ctx_t *ctx,
+                             char *volume_name);
 int glusterfs_graph_destroy_residual (glusterfs_graph_t *graph);
 int glusterfs_graph_deactivate (glusterfs_graph_t *graph);
 int glusterfs_graph_destroy (glusterfs_graph_t *graph);
 int glusterfs_get_leaf_count (glusterfs_graph_t *graph);
 int glusterfs_graph_activate (glusterfs_graph_t *graph, glusterfs_ctx_t *ctx);
 glusterfs_graph_t *glusterfs_graph_construct (FILE *fp);
+int glusterfs_graph_init (glusterfs_graph_t *graph);
 glusterfs_graph_t *glusterfs_graph_new (void);
 int glusterfs_graph_reconfigure (glusterfs_graph_t *oldgraph,
                                   glusterfs_graph_t *newgraph);
+int glusterfs_graph_attach (glusterfs_graph_t *orig_graph, char *path);
 
 void
 gf_free_mig_locks (lock_migration_info_t *locks);
diff --git a/libglusterfs/src/graph.c b/libglusterfs/src/graph.c
index 04bb92c7c75..b090f8a3554 100644
--- a/libglusterfs/src/graph.c
+++ b/libglusterfs/src/graph.c
@@ -407,13 +407,11 @@ fill_uuid (char *uuid, int size)
 
 
 int
-glusterfs_graph_settop (glusterfs_graph_t *graph, glusterfs_ctx_t *ctx)
+glusterfs_graph_settop (glusterfs_graph_t *graph, glusterfs_ctx_t *ctx,
+                        char *volume_name)
 {
-        const char *volume_name = NULL;
         xlator_t   *trav = NULL;
 
-        volume_name = ctx->cmd_args.volume_name;
-
         if (!volume_name) {
                 graph->top = graph->first;
                 return 0;
@@ -454,7 +452,8 @@ glusterfs_graph_parent_up (glusterfs_graph_t *graph)
 
 
 int
-glusterfs_graph_prepare (glusterfs_graph_t *graph, glusterfs_ctx_t *ctx)
+glusterfs_graph_prepare (glusterfs_graph_t *graph, glusterfs_ctx_t *ctx,
+                         char *volume_name)
 {
         xlator_t    *trav = NULL;
         int          ret = 0;
@@ -462,12 +461,20 @@ glusterfs_graph_prepare (glusterfs_graph_t *graph, glusterfs_ctx_t *ctx)
         /* XXX: CHECKSUM */
 
         /* XXX: attach to -n volname */
-        ret = glusterfs_graph_settop (graph, ctx);
+        ret = glusterfs_graph_settop (graph, ctx, volume_name);
         if (ret) {
+                char *slash = rindex (volume_name, '/');
+                if (slash) {
+                        ret = glusterfs_graph_settop (graph, ctx, slash + 1);
+                        if (!ret) {
+                                goto ok;
+                        }
+                }
                 gf_msg ("graph", GF_LOG_ERROR, 0, LG_MSG_GRAPH_ERROR,
                         "glusterfs graph settop failed");
                 return -1;
         }
+ok:
 
         /* XXX: WORM VOLUME */
         ret = glusterfs_graph_worm (graph, ctx);
@@ -749,7 +756,7 @@ xlator_equal_rec (xlator_t *xl1, xlator_t *xl2)
         }
 
 	/* type could have changed even if xlator names match,
-	   e.g cluster/distrubte and cluster/nufa share the same
+	   e.g cluster/distribute and cluster/nufa share the same
 	   xlator name
 	*/
         if (strcmp (xl1->type, xl2->type)) {
@@ -764,13 +771,27 @@ out :
 gf_boolean_t
 is_graph_topology_equal (glusterfs_graph_t *graph1, glusterfs_graph_t *graph2)
 {
-        xlator_t    *trav1    = NULL;
-        xlator_t    *trav2    = NULL;
-        gf_boolean_t ret      = _gf_true;
+        xlator_t      *trav1    = NULL;
+        xlator_t      *trav2    = NULL;
+        gf_boolean_t   ret      = _gf_true;
+        xlator_list_t *ltrav;
 
         trav1 = graph1->first;
         trav2 = graph2->first;
 
+        if (strcmp (trav2->type, "protocol/server") == 0) {
+                trav2 = trav2->children->xlator;
+                for (ltrav = trav1->children; ltrav; ltrav = ltrav->next) {
+                        trav1 = ltrav->xlator;
+                        if (strcmp (trav1->name, trav2->name) == 0) {
+                                break;
+                        }
+                }
+                if (!ltrav) {
+                        return _gf_false;
+                }
+        }
+
         ret = xlator_equal_rec (trav1, trav2);
 
         if (ret) {
@@ -869,7 +890,8 @@ glusterfs_volfile_reconfigure (int oldvollen, FILE *newvolfile_fp,
                 goto out;
         }
 
-	glusterfs_graph_prepare (newvolfile_graph, ctx);
+	glusterfs_graph_prepare (newvolfile_graph, ctx,
+                                 ctx->cmd_args.volume_name);
 
         if (!is_graph_topology_equal (oldvolfile_graph,
                                       newvolfile_graph)) {
@@ -917,8 +939,9 @@ int
 glusterfs_graph_reconfigure (glusterfs_graph_t *oldgraph,
                              glusterfs_graph_t *newgraph)
 {
-        xlator_t   *old_xl   = NULL;
-        xlator_t   *new_xl   = NULL;
+        xlator_t        *old_xl   = NULL;
+        xlator_t        *new_xl   = NULL;
+        xlator_list_t   *trav;
 
         GF_ASSERT (oldgraph);
         GF_ASSERT (newgraph);
@@ -933,7 +956,25 @@ glusterfs_graph_reconfigure (glusterfs_graph_t *oldgraph,
                 new_xl = new_xl->children->xlator;
         }
 
-        return xlator_tree_reconfigure (old_xl, new_xl);
+        if (strcmp (old_xl->type, "protocol/server") != 0) {
+                return xlator_tree_reconfigure (old_xl, new_xl);
+        }
+
+        /* Some options still need to be handled by the server translator. */
+        if (old_xl->reconfigure) {
+                old_xl->reconfigure (old_xl, new_xl->options);
+        }
+
+        (void) copy_opts_to_child (new_xl, FIRST_CHILD (new_xl), "*auth*");
+        new_xl = FIRST_CHILD (new_xl);
+
+        for (trav = old_xl->children; trav; trav = trav->next) {
+                if (strcmp (trav->xlator->name, new_xl->name) == 0) {
+                        return xlator_tree_reconfigure (trav->xlator, new_xl);
+                }
+        }
+
+        return -1;
 }
 
 int
@@ -987,3 +1028,61 @@ glusterfs_graph_destroy (glusterfs_graph_t *graph)
 out:
         return ret;
 }
+
+
+int
+glusterfs_graph_attach (glusterfs_graph_t *orig_graph, char *path)
+{
+        xlator_t                *this   = THIS;
+        FILE                    *fp;
+        glusterfs_graph_t       *graph;
+        xlator_t                *xl;
+        char                    *volfile_id;
+
+        fp = fopen (path, "r");
+        if (!fp) {
+                gf_log (THIS->name, GF_LOG_WARNING,
+                        "oops, %s disappeared on us", path);
+                return -EIO;
+        }
+
+        graph = glusterfs_graph_construct (fp);
+        fclose(fp);
+        if (!graph) {
+                gf_log (this->name, GF_LOG_WARNING,
+                        "could not create graph from %s", path);
+                return -EIO;
+        }
+
+        /*
+         * If there's a server translator on top, we want whatever's below
+         * that.
+         */
+        xl = graph->first;
+        if (strcmp(xl->type, "protocol/server") == 0) {
+                (void) copy_opts_to_child (xl, FIRST_CHILD (xl), "*auth*");
+                xl = FIRST_CHILD(xl);
+        }
+        graph->first = xl;
+
+
+        volfile_id = strstr (path, "/snaps/");
+        if (!volfile_id) {
+                volfile_id = rindex (path, '/');
+                if (volfile_id) {
+                        ++volfile_id;
+                }
+        }
+        if (volfile_id) {
+                xl->volfile_id = gf_strdup (volfile_id);
+                /* There's a stray ".vol" at the end. */
+                xl->volfile_id[strlen(xl->volfile_id)-4] = '\0';
+        }
+
+        /* TBD: memory leaks everywhere */
+        glusterfs_graph_prepare (graph, this->ctx, xl->name);
+        glusterfs_graph_init (graph);
+        glusterfs_xlator_link (orig_graph->top, graph->top);
+
+        return 0;
+}
diff --git a/libglusterfs/src/locking.c b/libglusterfs/src/locking.c
index d3b9754ef76..f27b0d05b35 100644
--- a/libglusterfs/src/locking.c
+++ b/libglusterfs/src/locking.c
@@ -22,7 +22,7 @@ int use_spinlocks = 0;
 static void __attribute__((constructor))
 gf_lock_setup (void)
 {
-        use_spinlocks = (sysconf(_SC_NPROCESSORS_ONLN) > 1);
+        //use_spinlocks = (sysconf(_SC_NPROCESSORS_ONLN) > 1);
 }
 
 #endif
diff --git a/libglusterfs/src/xlator.c b/libglusterfs/src/xlator.c
index 2edebc0aec2..4702ea3eb77 100644
--- a/libglusterfs/src/xlator.c
+++ b/libglusterfs/src/xlator.c
@@ -406,6 +406,59 @@ out:
         return search;
 }
 
+
+/*
+ * With brick multiplexing, we sort of have multiple graphs, so
+ * xlator_search_by_name might not find what we want.  Also, the translator
+ * we're looking for might not be a direct child if something else was put in
+ * between (as already happened with decompounder before that was fixed) and
+ * it's hard to debug why our translator wasn't found.  Using a recursive tree
+ * search instead of a linear search works around both problems.
+ */
+static xlator_t *
+get_xlator_by_name_or_type (xlator_t *this, char *target, int is_name)
+{
+        xlator_list_t   *trav;
+        xlator_t        *child_xl;
+        char            *value;
+
+        for (trav = this->children; trav; trav = trav->next) {
+                value = is_name ? trav->xlator->name : trav->xlator->type;
+                if (strcmp(value, target) == 0) {
+                        return trav->xlator;
+                }
+                child_xl = get_xlator_by_name_or_type (trav->xlator, target,
+                                                       is_name);
+                if (child_xl) {
+                        /*
+                         * If the xlator we're looking for is somewhere down
+                         * the stack, get_xlator_by_name expects to get a
+                         * pointer to the top of its subtree (child of "this")
+                         * while get_xlator_by_type expects a pointer to what
+                         * we actually found.  Handle both cases here.
+                         *
+                         * TBD: rename the functions and fix callers to better
+                         * reflect the difference in semantics.
+                         */
+                        return is_name ? trav->xlator : child_xl;
+                }
+        }
+
+        return NULL;
+}
+
+xlator_t *
+get_xlator_by_name (xlator_t *this, char *target)
+{
+        return get_xlator_by_name_or_type (this, target, 1);
+}
+
+xlator_t *
+get_xlator_by_type (xlator_t *this, char *target)
+{
+        return get_xlator_by_name_or_type (this, target, 0);
+}
+
 static int
 __xlator_init(xlator_t *xl)
 {
@@ -1104,3 +1157,22 @@ xlator_subvolume_count (xlator_t *this)
                 i++;
         return i;
 }
+
+static int
+_copy_opt_to_child (dict_t *options, char *key, data_t *value, void *data)
+{
+        xlator_t        *child = data;
+
+        gf_log (__func__, GF_LOG_DEBUG,
+                "copying %s to child %s", key, child->name);
+        dict_set (child->options, key, value);
+
+        return 0;
+}
+
+int
+copy_opts_to_child (xlator_t *src, xlator_t *dst, char *glob)
+{
+        return dict_foreach_fnmatch (src->options, glob,
+                                     _copy_opt_to_child, dst);
+}
diff --git a/libglusterfs/src/xlator.h b/libglusterfs/src/xlator.h
index e28790cc034..1e2698bb61f 100644
--- a/libglusterfs/src/xlator.h
+++ b/libglusterfs/src/xlator.h
@@ -950,6 +950,9 @@ struct _xlator {
         /* for the memory pool of 'frame->local' */
         struct mem_pool    *local_pool;
         gf_boolean_t        is_autoloaded;
+
+        /* Saved volfile ID (used for multiplexing) */
+        char               *volfile_id;
 };
 
 typedef struct {
@@ -1004,6 +1007,8 @@ void xlator_foreach_depth_first (xlator_t *this,
 				 void *data);
 
 xlator_t *xlator_search_by_name (xlator_t *any, const char *name);
+xlator_t *get_xlator_by_name (xlator_t *this, char *target);
+xlator_t *get_xlator_by_type (xlator_t *this, char *target);
 
 void
 xlator_set_inode_lru_limit (xlator_t *this, void *data);
@@ -1050,5 +1055,7 @@ xlator_subvolume_count (xlator_t *this);
 
 void xlator_init_lock (void);
 void xlator_init_unlock (void);
+int
+copy_opts_to_child (xlator_t *src, xlator_t *dst, char *glob);
 
 #endif /* _XLATOR_H */