features/bitrot: cleanup, v1

    Backport of http://review.gluster.org/11147

This is a short series of patches (with other cleanups) aimed at
cleaning up some of the incorrect assumptions taken in reconfigure()
leading to crashes when subvolumes are not fully initialized (as
reported here[1] on gluster-devel@). Furthermore, there is some
amount of code cleanup to handle disconnection and cleanup up data
structure (as part of subsequent patch).

[1] http://www.gluster.org/pipermail/gluster-devel/2015-June/045410.html

Change-Id: I68ac4bccfbac4bf02fcc31615bd7d2d191021132
BUG: 1226830
Signed-off-by: Venky Shankar <vshankar@redhat.com>
Reviewed-on: http://review.gluster.org/11539
Reviewed-by: Raghavendra Bhat <raghavendra@redhat.com>
Tested-by: NetBSD Build System <jenkins@build.gluster.org>
Tested-by: Gluster Build System <jenkins@build.gluster.com>

3 files changed, 271 insertions, 126 deletions

diff --git a/xlators/features/bit-rot/src/bitd/bit-rot.c b/xlators/features/bit-rot/src/bitd/bit-rot.c
index 5f297dbab76..4e86f5071ed 100644
--- a/xlators/features/bit-rot/src/bitd/bit-rot.c
+++ b/xlators/features/bit-rot/src/bitd/bit-rot.c
@@ -30,6 +30,18 @@
 
 #define BR_HASH_CALC_READ_SIZE  (128 * 1024)
 
+typedef int32_t (br_child_handler)(xlator_t *, br_child_t *);
+
+struct br_child_event {
+        xlator_t *this;
+
+        br_child_t *child;
+
+        br_child_handler *call;
+
+        struct list_head list;
+};
+
 static int
 br_find_child_index (xlator_t *this, xlator_t *child)
 {
@@ -54,26 +66,6 @@ out:
         return index;
 }
 
-static void
-br_free_children (xlator_t *this)
-{
-        br_private_t *priv = NULL;
-        int32_t       i    = 0;
-        br_child_t   *child = NULL;
-
-        priv = this->private;
-
-        for (i = 0; i < priv->child_count; i++) {
-                child = &priv->children[i];
-                mem_pool_destroy (child->timer_pool);
-                list_del_init (&priv->children[i].list);
-        }
-
-        GF_FREE (priv->children);
-
-        priv->children = NULL;
-}
-
 br_child_t *
 br_get_child_from_brick_path (xlator_t *this, char *brick_path)
 {
@@ -1090,6 +1082,16 @@ br_oneshot_signer (void *arg)
         return NULL;
 }
 
+static void
+br_set_child_state (br_child_t *child, br_child_state_t state)
+{
+        LOCK (&child->lock);
+        {
+                _br_set_child_state (child, state);
+        }
+        UNLOCK (&child->lock);
+}
+
 /**
  * At this point a thread is spawned to crawl the filesystem (in
  * tortoise pace) to sign objects that were not signed in previous run(s).
@@ -1177,7 +1179,12 @@ br_enact_scrubber (xlator_t *this, br_child_t *child)
                 goto error_return;
         }
 
-        ret = br_fsscan_schedule (this, child, fsscan, fsscrub);
+        /* this needs to be serialized with reconfigure() */
+        pthread_mutex_lock (&priv->lock);
+        {
+                ret = br_fsscan_schedule (this, child, fsscan, fsscrub);
+        }
+        pthread_mutex_unlock (&priv->lock);
         if (ret)
                 goto error_return;
 
@@ -1202,6 +1209,30 @@ br_enact_scrubber (xlator_t *this, br_child_t *child)
         return -1;
 }
 
+static int32_t
+br_child_enaction (xlator_t *this, br_child_t *child, br_stub_init_t *stub)
+{
+        int32_t ret = -1;
+        br_private_t *priv = this->private;
+
+        LOCK (&child->lock);
+        {
+                if (priv->iamscrubber)
+                        ret = br_enact_scrubber (this, child);
+                else
+                        ret = br_enact_signer (this, child, stub);
+
+                if (!ret) {
+                        _br_set_child_state (child, BR_CHILD_STATE_CONNECTED);
+                        gf_log (this->name, GF_LOG_INFO,
+                                "Connected to brick %s..", child->brick_path);
+                }
+        }
+        UNLOCK (&child->lock);
+
+        return ret;
+}
+
 /**
  * This routine fetches various attributes associated with a child which
  * is basically a subvolume. Attributes include brick path and the stub
@@ -1209,7 +1240,7 @@ br_enact_scrubber (xlator_t *this, br_child_t *child)
  * by getxattr() on a virtual key. Depending on the configuration, the
  * process either acts as a signer or a scrubber.
  */
-static inline int32_t
+int32_t
 br_brick_connect (xlator_t *this, br_child_t *child)
 {
         int32_t         ret      = -1;
@@ -1218,14 +1249,13 @@ br_brick_connect (xlator_t *this, br_child_t *child)
         struct iatt     parent   = {0, };
         br_stub_init_t *stub     = NULL;
         dict_t         *xattr    = NULL;
-        br_private_t   *priv     = NULL;
         int             op_errno = 0;
 
         GF_VALIDATE_OR_GOTO ("bit-rot", this, out);
         GF_VALIDATE_OR_GOTO (this->name, child, out);
         GF_VALIDATE_OR_GOTO (this->name, this->private, out);
 
-        priv = this->private;
+        br_set_child_state (child, BR_CHILD_STATE_INITIALIZING);
 
         loc.inode = inode_ref (child->table->root);
         gf_uuid_copy (loc.gfid, loc.inode->gfid);
@@ -1262,20 +1292,15 @@ br_brick_connect (xlator_t *this, br_child_t *child)
         child->tv.tv_sec = ntohl (stub->timebuf[0]);
         child->tv.tv_usec = ntohl (stub->timebuf[1]);
 
-        if (priv->iamscrubber)
-                ret = br_enact_scrubber (this, child);
-        else
-                ret = br_enact_signer (this, child, stub);
-
-        if (!ret)
-                gf_msg (this->name, GF_LOG_INFO, 0, BRB_MSG_CONNECTED_TO_BRICK,
-                        "Connected to brick %s..", child->brick_path);
+        ret = br_child_enaction (this, child, stub);
 
  free_dict:
         dict_unref (xattr);
  wipeloc:
         loc_wipe (&loc);
  out:
+        if (ret)
+                br_set_child_state (child, BR_CHILD_STATE_CONNFAILED);
         return ret;
 }
 
@@ -1290,7 +1315,8 @@ br_handle_events (void *arg)
         int32_t       ret   = 0;
         xlator_t     *this  = NULL;
         br_private_t *priv  = NULL;
-        br_child_t   *child = NULL;
+        br_child_t *child = NULL;
+        struct br_child_event *childev = NULL;
 
         this = arg;
         priv = this->private;
@@ -1309,17 +1335,20 @@ br_handle_events (void *arg)
                         while (list_empty (&priv->bricks))
                                 pthread_cond_wait (&priv->cond, &priv->lock);
 
-                        child = list_first_entry
-                                          (&priv->bricks, br_child_t, list);
-                        list_del_init (&child->list);
+                        childev = list_first_entry
+                                   (&priv->bricks, struct br_child_event, list);
+                        list_del_init (&childev->list);
                 }
                 pthread_mutex_unlock (&priv->lock);
 
-                ret = br_brick_connect (this, child);
+                child = childev->child;
+                ret = childev->call (this, child);
                 if (ret)
                         gf_msg (this->name, GF_LOG_ERROR, 0,
-                                BRB_MSG_SUBVOL_CONNECT_FAILED, "failed to "
-                                "connect to subvolume %s", child->xl->name);
+                                BRB_MSG_SUBVOL_CONNECT_FAILED,
+                                "callback handler for subvolume [%s] failed",
+                                child->xl->name);
+                GF_FREE (childev);
         }
 
         return NULL;
@@ -1344,6 +1373,29 @@ mem_acct_init (xlator_t *this)
         return ret;
 }
 
+static void
+_br_qchild_event (xlator_t *this, br_child_t *child, br_child_handler *call)
+{
+        br_private_t *priv = NULL;
+        struct br_child_event *childev = NULL;
+
+        priv = this->private;
+
+        childev = GF_CALLOC (1, sizeof (*childev), gf_br_mt_br_child_event_t);
+        if (!childev) {
+                gf_log (this->name, GF_LOG_ERROR, "Event unhandled for "
+                        "child.. [Brick: %s]", child->xl->name);
+                return;
+        }
+
+        INIT_LIST_HEAD (&childev->list);
+        childev->this  = this;
+        childev->child = child;
+        childev->call  = call;
+
+        list_add_tail (&childev->list, &priv->bricks);
+}
+
 int
 notify (xlator_t *this, int32_t event, void *data, ...)
 {
@@ -1373,14 +1425,14 @@ notify (xlator_t *this, int32_t event, void *data, ...)
                         child = &priv->children[idx];
                         if (child->child_up == 1)
                                 goto unblock;
+                        priv->up_children++;
 
                         child->child_up = 1;
                         child->xl = subvol;
-                        child->table = inode_table_new (4096, subvol);
+                        if (!child->table)
+                                child->table = inode_table_new (4096, subvol);
 
-                        priv->up_children++;
-
-                        list_add_tail (&child->list, &priv->bricks);
+                        _br_qchild_event (this, child, br_brick_connect);
                         pthread_cond_signal (&priv->cond);
                 }
         unblock:
@@ -1410,6 +1462,7 @@ notify (xlator_t *this, int32_t event, void *data, ...)
                 if (priv->up_children == 0)
                         default_notify (this, event, data);
                 break;
+
         default:
                 default_notify (this, event, data);
         }
@@ -1563,59 +1616,94 @@ br_signer_init (xlator_t *this, br_private_t *priv)
 
 }
 
-int32_t
-init (xlator_t *this)
+static void
+br_free_children (xlator_t *this, br_private_t *priv, int count)
 {
-        int            i    = 0;
-        int32_t        ret  = -1;
-        br_private_t  *priv = NULL;
-        xlator_list_t *trav = NULL;
+        br_child_t *child = NULL;
 
-        if (!this->children) {
-                gf_msg (this->name, GF_LOG_ERROR, 0, BRB_MSG_NO_CHILD,
-                        "FATAL: no children");
-                goto out;
+        for (--count; count >= 0; count--) {
+                child = &priv->children[count];
+                mem_pool_destroy (child->timer_pool);
+                LOCK_DESTROY (&child->lock);
         }
 
-        priv = GF_CALLOC (1, sizeof (*priv), gf_br_mt_br_private_t);
-        if (!priv) {
-                gf_msg (this->name, GF_LOG_ERROR, ENOMEM, BRB_MSG_NO_MEMORY,
-                        "failed to allocate memory (->priv)");
-                goto out;
-        }
+        GF_FREE (priv->children);
+        priv->children = NULL;
+}
 
-        GF_OPTION_INIT ("scrubber", priv->iamscrubber, bool, out);
+static int
+br_init_children (xlator_t *this, br_private_t *priv)
+{
+        int i = 0;
+        br_child_t *child = NULL;
+        xlator_list_t *trav = NULL;
 
         priv->child_count = xlator_subvolume_count (this);
         priv->children = GF_CALLOC (priv->child_count, sizeof (*priv->children),
                                     gf_br_mt_br_child_t);
         if (!priv->children)
-                goto free_priv;
+                goto err;
 
         trav = this->children;
         while (trav) {
-                priv->children[i].this = this;
-                priv->children[i].xl = trav->xlator;
-
-                priv->children[i].timer_pool =
-                                  mem_pool_new (struct gf_tw_timer_list,  4096);
-                if (!priv->children[i].timer_pool) {
-                        gf_msg (this->name, GF_LOG_ERROR, ENOMEM,
-                                BRB_MSG_NO_MEMORY, "failed to allocate mem-pool"
-                                " for timer");
+                child = &priv->children[i];
+
+                LOCK_INIT (&child->lock);
+                br_set_child_state (child, BR_CHILD_STATE_DISCONNECTED);
+
+                child->this = this;
+                child->xl = trav->xlator;
+
+                child->timer_pool = mem_pool_new
+                                    (struct gf_tw_timer_list,  4096);
+                if (!child->timer_pool) {
+                        gf_log (this->name, GF_LOG_ERROR,
+                                "failed to allocate mem-pool for timer");
                         errno = ENOMEM;
-                        goto free_children;
+                        goto freechild;
                 }
 
+                INIT_LIST_HEAD (&child->list);
+
                 i++;
                 trav = trav->next;
         }
 
+        return 0;
+
+ freechild:
+        br_free_children (this, priv, i);
+ err:
+        return -1;
+}
+
+int32_t
+init (xlator_t *this)
+{
+        int32_t       ret  = -1;
+	br_private_t *priv = NULL;
+
+	if (!this->children) {
+		gf_log (this->name, GF_LOG_ERROR, "FATAL: no children");
+		goto out;
+	}
+
+        priv = GF_CALLOC (1, sizeof (*priv), gf_br_mt_br_private_t);
+        if (!priv) {
+                gf_log (this->name, GF_LOG_ERROR,
+                        "failed to allocate memory (->priv)");
+                goto out;
+        }
+
+        GF_OPTION_INIT ("scrubber", priv->iamscrubber, bool, out);
+
+        ret = br_init_children (this, priv);
+        if (ret)
+                goto free_priv;
+
         pthread_mutex_init (&priv->lock, NULL);
         pthread_cond_init (&priv->cond, NULL);
 
-        for (i = 0; i < priv->child_count; i++)
-                INIT_LIST_HEAD (&priv->children[i].list);
         INIT_LIST_HEAD (&priv->bricks);
         INIT_LIST_HEAD (&priv->signing);
 
@@ -1624,7 +1712,7 @@ init (xlator_t *this)
                 gf_msg (this->name, GF_LOG_ERROR, 0,
                         BRB_MSG_TIMER_WHEEL_UNAVAILABLE,
                         "global timer wheel unavailable");
-                goto cleanup_mutex;
+                goto cleanup;
         }
 
 	this->private = priv;
@@ -1640,7 +1728,7 @@ init (xlator_t *this)
         }
 
         if (ret)
-                goto cleanup_mutex;
+                goto cleanup;
 
         ret = gf_thread_create (&priv->thread, NULL, br_handle_events, this);
         if (ret != 0) {
@@ -1656,16 +1744,12 @@ init (xlator_t *this)
                 return 0;
         }
 
- cleanup_mutex:
+ cleanup:
         (void) pthread_cond_destroy (&priv->cond);
         (void) pthread_mutex_destroy (&priv->lock);
- free_children:
-        for (i = 0; i < priv->child_count; i++) {
-                if (priv->children[i].timer_pool)
-                        mem_pool_destroy (priv->children[i].timer_pool);
-        }
 
-        GF_FREE (priv->children);
+        br_free_children (this, priv, priv->child_count);
+
  free_priv:
         GF_FREE (priv);
  out:
@@ -1683,7 +1767,7 @@ fini (xlator_t *this)
 
         if (!priv->iamscrubber)
                 br_fini_signer (this, priv);
-        br_free_children (this);
+        br_free_children (this, priv, priv->child_count);
 
         this->private = NULL;
 	GF_FREE (priv);
@@ -1691,64 +1775,96 @@ fini (xlator_t *this)
 	return;
 }
 
-int
-reconfigure (xlator_t *this, dict_t *options)
+static void
+br_reconfigure_child (xlator_t *this,
+                      br_child_t *child, struct br_scrubber *fsscrub)
 {
-        int i = 0;
-        int32_t ret = -1;
-        br_child_t *child = NULL;
-        br_private_t *priv = NULL;
-        struct br_scanfs *fsscan = NULL;
+        int32_t ret = 0;
+        struct br_scanfs *fsscan = &child->fsscan;
+
+        ret = br_fsscan_reschedule (this, child, fsscan, fsscrub, _gf_true);
+        if (ret) {
+                gf_log (this->name, GF_LOG_ERROR,
+                        "Could not reschedule scrubber for brick: %s. "
+                        "Scubbing will continue according to old frequency.",
+                        child->brick_path);
+        }
+}
+
+static int
+br_reconfigure_scrubber (xlator_t *this, dict_t *options)
+{
+        int                 i       = 0;
+        int32_t             ret     = -1;
+        br_child_t         *child   = NULL;
+        br_private_t       *priv    = NULL;
         struct br_scrubber *fsscrub = NULL;
 
         priv = this->private;
+        fsscrub = &priv->fsscrub;
 
-        if (!priv->iamscrubber) {
-                ret = br_signer_handle_options (this, priv, options);
-                if (ret)
-                        goto err;
-                return 0;
+        pthread_mutex_lock (&priv->lock);
+        {
+                ret = br_scrubber_handle_options (this, priv, options);
         }
+        pthread_mutex_unlock (&priv->lock);
 
-        ret = br_scrubber_handle_options (this, priv, options);
         if (ret)
                 goto err;
 
-        fsscrub = &priv->fsscrub;
-
         /* reschedule all _up_ subvolume(s) */
-        pthread_mutex_lock (&priv->lock);
-        {
-                for (; i < priv->child_count; i++) {
-                        child = &priv->children[i];
-                        if (!child->child_up) {
-                                gf_msg (this->name, GF_LOG_INFO, 0,
-                                        BRB_MSG_BRICK_INFO, "Brick %s is "
-                                        "offline, skipping rescheduling (scrub"
-                                        " would auto- schedule when brick is "
-                                        "back online).", child->brick_path);
-                                continue;
-                        }
+        for (; i < priv->child_count; i++) {
+                child = &priv->children[i];
 
-                        fsscan = &child->fsscan;
-                        ret = br_fsscan_reschedule (this, child,
-                                                    fsscan, fsscrub, _gf_true);
-                        if (ret) {
-                                gf_msg (this->name, GF_LOG_ERROR, 0,
-                                        BRB_MSG_RESCHEDULE_SCRUBBER_FAILED,
-                                        "Could not reschedule scrubber for "
-                                        "brick: %s. Scubbing will continue "
-                                        "according to old frequency.",
-                                        child->brick_path);
+                LOCK (&child->lock);
+                {
+                        if (_br_child_failed_conn (child)) {
+                                gf_log (this->name, GF_LOG_INFO,
+                                        "Scrubber for brick [%s] failed "
+                                        "initialization, rescheduling is "
+                                        "skipped", child->brick_path);
+                                goto unblock;
                         }
+
+                        if (_br_is_child_connected (child))
+                                br_reconfigure_child (this, child, fsscrub);
+
+                        /**
+                         * for the rest.. either the child is in initialization
+                         * phase or is disconnected. either way, updated values
+                         * would be reflected on successful connection.
+                         */
                 }
+        unblock:
+                UNLOCK (&child->lock);
         }
-        pthread_mutex_unlock (&priv->lock);
-
-        return 0;
 
  err:
-        return -1;
+        return ret;
+}
+
+static int
+br_reconfigure_signer (xlator_t *this, dict_t *options)
+{
+        br_private_t *priv = this->private;
+
+        return br_signer_handle_options (this, priv, options);
+}
+
+int
+reconfigure (xlator_t *this, dict_t *options)
+{
+        int ret = 0;
+        br_private_t *priv = NULL;
+
+        priv = this->private;
+
+        if (priv->iamscrubber)
+                ret = br_reconfigure_scrubber (this, options);
+        else
+                ret = br_reconfigure_signer (this, options);
+
+        return ret;
 }
 
 struct xlator_fops fops;
diff --git a/xlators/features/bit-rot/src/bitd/bit-rot.h b/xlators/features/bit-rot/src/bitd/bit-rot.h
index 562f17e12bc..b8d9e3b1183 100644
--- a/xlators/features/bit-rot/src/bitd/bit-rot.h
+++ b/xlators/features/bit-rot/src/bitd/bit-rot.h
@@ -76,7 +76,18 @@ struct br_scanfs {
         struct gf_tw_timer_list *timer;
 };
 
+/* just need three states to track child status */
+typedef enum br_child_state {
+        BR_CHILD_STATE_CONNECTED = 1,
+        BR_CHILD_STATE_INITIALIZING,
+        BR_CHILD_STATE_CONNFAILED,
+        BR_CHILD_STATE_DISCONNECTED,
+} br_child_state_t;
+
 struct br_child {
+        gf_lock_t lock;
+        br_child_state_t c_state;
+
         char child_up;                /* Indicates whether this child is
                                          up or not */
         xlator_t *xl;                 /* client xlator corresponding to
@@ -140,8 +151,8 @@ typedef struct br_obj_n_workers br_obj_n_workers_t;
 struct br_private {
         pthread_mutex_t lock;
 
-        struct list_head bricks;          /* list of bricks from which CHILD_UP
-                                             has been received */
+        struct list_head bricks;          /* list of bricks from which enents
+                                             have been received */
 
         struct list_head signing;
 
@@ -207,5 +218,22 @@ br_prepare_loc (xlator_t *, br_child_t *, loc_t *, gf_dirent_t *, loc_t *);
 gf_boolean_t
 bitd_is_bad_file (xlator_t *, br_child_t *, loc_t *, fd_t *);
 
+static inline void
+_br_set_child_state (br_child_t *child, br_child_state_t state)
+{
+        child->c_state = state;
+}
+
+static inline int
+_br_is_child_connected (br_child_t *child)
+{
+        return (child->c_state == BR_CHILD_STATE_CONNECTED);
+}
+
+static inline int
+_br_child_failed_conn (br_child_t *child)
+{
+        return (child->c_state == BR_CHILD_STATE_CONNFAILED);
+}
 
 #endif /* __BIT_ROT_H__ */
diff --git a/xlators/features/bit-rot/src/stub/bit-rot-stub-mem-types.h b/xlators/features/bit-rot/src/stub/bit-rot-stub-mem-types.h
index fbb69ce7ea8..f70fafbca49 100644
--- a/xlators/features/bit-rot/src/stub/bit-rot-stub-mem-types.h
+++ b/xlators/features/bit-rot/src/stub/bit-rot-stub-mem-types.h
@@ -31,6 +31,7 @@ enum br_mem_types {
         gf_br_stub_mt_br_stub_fd_t,
         gf_br_stub_mt_br_scanner_freq_t,
         gf_br_stub_mt_sigstub_t,
+        gf_br_mt_br_child_event_t,
         gf_br_stub_mt_end,
 };