cluster/afr: Handle lookups when self-heal is off

Change-Id: Ibc23fef417bcf613850e03dc4dadcc88f89e2b6f
BUG: 2586
Reviewed-on: http://review.gluster.com/59
Tested-by: Gluster Build System <jenkins@build.gluster.com>
Reviewed-by: Vijay Bellur <vijay@gluster.com>

9 files changed, 983 insertions, 590 deletions

diff --git a/xlators/cluster/afr/src/afr-common.c b/xlators/cluster/afr/src/afr-common.c
index a435a38b1..f49d8c55e 100644
--- a/xlators/cluster/afr/src/afr-common.c
+++ b/xlators/cluster/afr/src/afr-common.c
@@ -328,6 +328,8 @@ afr_local_sh_cleanup (afr_local_t *local, xlator_t *this)
 
         if (sh->linkname)
                 GF_FREE ((char *)sh->linkname);
+        if (sh->child_success)
+                GF_FREE (sh->child_success);
 
         loc_wipe (&sh->parent_loc);
 }
@@ -417,6 +419,18 @@ afr_local_cleanup (afr_local_t *local, xlator_t *this)
                 if (local->cont.lookup.inode) {
                         inode_unref (local->cont.lookup.inode);
                 }
+
+                if (local->cont.lookup.postparents)
+                        GF_FREE (local->cont.lookup.postparents);
+
+                if (local->cont.lookup.bufs)
+                        GF_FREE (local->cont.lookup.bufs);
+
+                if (local->cont.lookup.child_success)
+                        GF_FREE (local->cont.lookup.child_success);
+
+                if (local->cont.lookup.sources)
+                        GF_FREE (local->cont.lookup.sources);
         }
 
         { /* getxattr */
@@ -509,6 +523,22 @@ afr_up_children_count (int child_count, unsigned char *child_up)
         return ret;
 }
 
+gf_boolean_t
+afr_is_fresh_lookup (loc_t *loc, xlator_t *this)
+{
+        uint64_t          ctx = 0;
+        int32_t           ret = 0;
+
+        GF_ASSERT (loc);
+        GF_ASSERT (this);
+        GF_ASSERT (loc->inode);
+
+        ret = inode_ctx_get (loc->inode, this, &ctx);
+        if (0 == ret)
+                return _gf_false;
+        return _gf_true;
+}
+
 void
 afr_update_loc_gfids (loc_t *loc, struct iatt *buf, struct iatt *postparent)
 {
@@ -567,68 +597,96 @@ afr_self_heal_lookup_unwind (call_frame_t *frame, xlator_t *this)
         }
 
         AFR_STACK_UNWIND (lookup, frame, local->op_ret, local->op_errno,
-                          local->cont.lookup.inode,
-                          &local->cont.lookup.buf,
+                          local->cont.lookup.inode, &local->cont.lookup.buf,
                           local->cont.lookup.xattr,
                           &local->cont.lookup.postparent);
 
         return 0;
 }
 
+void
+afr_lookup_build_response_params (afr_local_t *local, xlator_t *this)
+{
+        int32_t         read_child = -1;
+        struct iatt     *buf = NULL;
+        struct iatt     *postparent = NULL;
+        dict_t          **xattr = NULL;
 
-static void
-afr_lookup_collect_xattr (afr_local_t *local, xlator_t *this,
-                          int child_index, dict_t *xattr)
+        GF_ASSERT (local);
+        GF_ASSERT (local->cont.lookup.read_child >= 0);
+
+        buf = &local->cont.lookup.buf;
+        postparent = &local->cont.lookup.postparent;
+        xattr = &local->cont.lookup.xattr;
+
+        read_child = local->cont.lookup.read_child;
+        *xattr = dict_ref (local->cont.lookup.xattrs[read_child]);
+        *buf = local->cont.lookup.bufs[read_child];
+        *postparent = local->cont.lookup.postparents[read_child];
+
+        if (IA_INVAL == local->cont.lookup.inode->ia_type) {
+                /* fix for RT #602 */
+                local->cont.lookup.inode->ia_type = buf->ia_type;
+        }
+}
+
+
+ static void
+afr_lookup_update_lk_counts (afr_local_t *local, xlator_t *this,
+                            int child_index, dict_t *xattr)
 {
         uint32_t inodelk_count = 0;
         uint32_t entrylk_count = 0;
-        int      ret           = 0;
+        int      ret           = -1;
+
+        GF_ASSERT (local);
+        GF_ASSERT (this);
+        GF_ASSERT (xattr);
+        GF_ASSERT (child_index >= 0);
+
+        ret = dict_get_uint32 (xattr, GLUSTERFS_INODELK_COUNT,
+                               &inodelk_count);
+        if (ret == 0)
+                local->inodelk_count += inodelk_count;
 
-        if (afr_sh_has_metadata_pending (xattr, child_index, this)) {
+        ret = dict_get_uint32 (xattr, GLUSTERFS_ENTRYLK_COUNT,
+                               &entrylk_count);
+        if (ret == 0)
+                local->entrylk_count += entrylk_count;
+}
+
+static void
+afr_lookup_detect_self_heal_by_xattr (afr_local_t *local, xlator_t *this,
+                                      dict_t *xattr)
+{
+        GF_ASSERT (local);
+        GF_ASSERT (this);
+        GF_ASSERT (xattr);
+
+        if (afr_sh_has_metadata_pending (xattr, this)) {
                 local->self_heal.need_metadata_self_heal = _gf_true;
                 gf_log(this->name, GF_LOG_DEBUG,
                        "metadata self-heal is pending for %s.",
                        local->loc.path);
         }
 
-        if (afr_sh_has_entry_pending (xattr, child_index, this)) {
+        if (afr_sh_has_entry_pending (xattr, this)) {
                 local->self_heal.need_entry_self_heal = _gf_true;
                 gf_log(this->name, GF_LOG_DEBUG,
                        "entry self-heal is pending for %s.", local->loc.path);
         }
 
-        if (afr_sh_has_data_pending (xattr, child_index, this)) {
+        if (afr_sh_has_data_pending (xattr, this)) {
                 local->self_heal.need_data_self_heal = _gf_true;
                 gf_log(this->name, GF_LOG_DEBUG,
                        "data self-heal is pending for %s.", local->loc.path);
         }
-
-        ret = dict_get_uint32 (xattr, GLUSTERFS_INODELK_COUNT,
-                               &inodelk_count);
-        if (ret == 0)
-                local->inodelk_count += inodelk_count;
-
-        ret = dict_get_uint32 (xattr, GLUSTERFS_ENTRYLK_COUNT,
-                               &entrylk_count);
-        if (ret == 0)
-                local->entrylk_count += entrylk_count;
 }
 
-
 static void
-afr_lookup_self_heal_check (xlator_t *this, afr_local_t *local,
+afr_detect_self_heal_by_iatt (afr_local_t *local, xlator_t *this,
                             struct iatt *buf, struct iatt *lookup_buf)
 {
-        if (FILETYPE_DIFFERS (buf, lookup_buf)) {
-                /* mismatching filetypes with same name
-                 */
-
-                gf_log (this->name, GF_LOG_INFO,
-                        "filetype differs for %s ", local->loc.path);
-
-                local->govinda_gOvinda = 1;
-        }
-
         if (PERMISSION_DIFFERS (buf, lookup_buf)) {
                 /* mismatching permissions */
                 gf_log (this->name, GF_LOG_INFO,
@@ -674,157 +732,299 @@ out:
         return valid;
 }
 
-void
-afr_lookup_set_read_child (xlator_t *this, afr_local_t *local)
+static void
+afr_detect_self_heal_by_lookup_status (afr_local_t *local, xlator_t *this)
 {
-        ia_type_t               ia_type = IA_INVAL;
-        afr_transaction_type    transaction_type = AFR_DATA_TRANSACTION;
-        afr_private_t           *priv = NULL;
-        int32_t                 read_child = -1;
-        afr_self_heal_t         *sh = NULL;
+        GF_ASSERT (local);
+        GF_ASSERT (this);
 
-        priv = this->private;
-        sh   = &local->self_heal;
+        if ((local->success_count > 0) && (local->enoent_count > 0)) {
+                local->self_heal.need_metadata_self_heal = _gf_true;
+                local->self_heal.need_data_self_heal     = _gf_true;
+                local->self_heal.need_entry_self_heal    = _gf_true;
+                gf_log(this->name, GF_LOG_INFO,
+                       "entries are missing in lookup of %s.",
+                       local->loc.path);
+                //If all self-heals are needed no need to check for other rules
+                goto out;
+        }
 
-        ia_type = local->cont.lookup.inode->ia_type;
-        if (IA_ISREG (ia_type)) {
-                transaction_type = AFR_DATA_TRANSACTION;
-        } else if IA_ISDIR (ia_type) {
-                transaction_type = AFR_ENTRY_TRANSACTION;
-        } else {
-                transaction_type = AFR_METADATA_TRANSACTION;
+        if (local->success_count > 0) {
+                if (afr_is_split_brain (this, local->cont.lookup.inode) &&
+                    IA_ISREG (local->cont.lookup.inode->ia_type)) {
+                        local->self_heal.need_data_self_heal = _gf_true;
+                        gf_log (this->name, GF_LOG_WARNING,
+                                "split brain detected during lookup of %s.",
+                                local->loc.path);
+                }
         }
-        afr_self_heal_find_sources (this, local,
-                                    local->cont.lookup.xattrs,
-                                    transaction_type);
-        if (!sh->sources)
-                goto out;
 
-        read_child = local->read_child_index;
-        if (afr_is_valid_read_child (sh->sources, priv->child_count,
-                                     read_child))
-                goto out;
+out:
+        return;
+}
+
+gf_boolean_t
+afr_can_self_heal_proceed (afr_self_heal_t *sh, afr_private_t *priv)
+{
+        GF_ASSERT (sh);
+        GF_ASSERT (priv);
+
+        return ((priv->data_self_heal && sh->need_data_self_heal)
+                || (priv->metadata_self_heal && sh->need_metadata_self_heal)
+                || (priv->entry_self_heal && sh->need_entry_self_heal));
+}
 
-        read_child = afr_read_child (this, local->loc.inode);
-        if (afr_is_valid_read_child (sh->sources, priv->child_count,
-                                     read_child))
+gf_boolean_t
+afr_is_self_heal_enabled (afr_private_t *priv)
+{
+        GF_ASSERT (priv);
+
+        return (priv->data_self_heal || priv->metadata_self_heal
+                || priv->entry_self_heal);
+}
+
+int
+afr_lookup_select_read_child (afr_local_t *local, xlator_t *this,
+                              int32_t *read_child)
+{
+        int32_t                 source = -1;
+        ia_type_t               ia_type = 0;
+        int                     ret = -1;
+        afr_transaction_type    type = AFR_METADATA_TRANSACTION;
+        dict_t                  **xattrs = NULL;
+        int32_t                 *child_success = NULL;
+        struct iatt             *bufs = NULL;
+
+        GF_ASSERT (local);
+        GF_ASSERT (this);
+
+        bufs = local->cont.lookup.bufs;
+        child_success = local->cont.lookup.child_success;
+        ia_type = local->cont.lookup.bufs[child_success[0]].ia_type;
+        if (IA_ISDIR (ia_type)) {
+                type = AFR_ENTRY_TRANSACTION;
+        } else if (IA_ISREG (ia_type)) {
+                type = AFR_DATA_TRANSACTION;
+        }
+        xattrs = local->cont.lookup.xattrs;
+        source = afr_lookup_select_read_child_by_txn_type (this, local, xattrs,
+                                                           type);
+        if (source < 0)
                 goto out;
 
-        read_child = afr_sh_select_source (sh->sources, priv->child_count);
+        *read_child = source;
+        ret = 0;
 out:
-        if (read_child >= 0) {
-                afr_set_read_child (this,
-                                    local->cont.lookup.inode,
-                                    read_child);
-        }
+        return ret;
+}
+
+static inline gf_boolean_t
+afr_is_self_heal_running (afr_local_t *local)
+{
+        GF_ASSERT (local);
+        return ((local->inodelk_count > 0) || (local->entrylk_count > 0));
 }
 
 static void
-afr_lookup_done (call_frame_t *frame, xlator_t *this, struct iatt *lookup_buf)
+afr_launch_self_heal (call_frame_t *frame, xlator_t *this,
+                      gf_boolean_t is_background, ia_type_t ia_type,
+                      int (*unwind) (call_frame_t *frame, xlator_t *this))
 {
-        int                     unwind = 1;
-        int                     up_count = 0;
-        char                    sh_type_str[256] = {0,};
-        afr_private_t           *priv  = NULL;
         afr_local_t             *local = NULL;
+        char                    sh_type_str[256] = {0,};
+
+        GF_ASSERT (frame);
+        GF_ASSERT (this);
 
-        priv  = this->private;
         local = frame->local;
+        local->self_heal.background = is_background;
+        local->self_heal.type       = ia_type;
+        local->self_heal.unwind     = unwind;
 
-        if (local->op_ret != 0)
-                goto unwind;
+        afr_self_heal_type_str_get (&local->self_heal,
+                                    sh_type_str,
+                                    sizeof (sh_type_str));
 
-        local->cont.lookup.postparent.ia_ino  = local->cont.lookup.parent_ino;
+        gf_log (this->name, GF_LOG_INFO,
+                "background %s self-heal triggered. path: %s",
+                sh_type_str, local->loc.path);
 
-        if (local->cont.lookup.ino) {
-                local->cont.lookup.buf.ia_ino = local->cont.lookup.ino;
+        afr_self_heal (frame, this);
+}
+
+static void
+afr_lookup_detect_self_heal (afr_local_t *local, xlator_t *this)
+{
+        int                     i = 0;
+        struct iatt             *bufs = NULL;
+        dict_t                  **xattr = NULL;
+        afr_private_t           *priv = NULL;
+        int32_t                 child1 = -1;
+        int32_t                 child2 = -1;
+
+        afr_detect_self_heal_by_lookup_status (local, this);
+
+        bufs = local->cont.lookup.bufs;
+        for (i = 1; i < local->success_count; i++) {
+                child1 = local->cont.lookup.child_success[i-1];
+                child2 = local->cont.lookup.child_success[i];;
+                afr_detect_self_heal_by_iatt (local, this,
+                                              &bufs[child1], &bufs[child2]);
         }
 
-        if (local->op_ret == 0) {
-                /* KLUDGE: assuming DHT will not itransform in
-                   revalidate */
-                if (local->cont.lookup.inode->ino) {
-                        local->cont.lookup.buf.ia_ino =
-                                local->cont.lookup.inode->ino;
-                }
+        xattr = local->cont.lookup.xattrs;
+        priv  = this->private;
+        for (i = 0; i < local->success_count; i++) {
+                child1 = local->cont.lookup.child_success[i];;
+                afr_lookup_detect_self_heal_by_xattr (local, this,
+                                                      xattr[child1]);
         }
+}
+
+static void
+afr_lookup_perform_self_heal_if_needed (call_frame_t *frame, xlator_t *this,
+                                        gf_boolean_t *sh_launched)
+{
+        size_t              up_count = 0;
+        afr_private_t       *priv    = NULL;
+        afr_local_t         *local   = NULL;
+
+        GF_ASSERT (sh_launched);
+        *sh_launched = _gf_false;
+        priv         = this->private;
+        local        = frame->local;
 
-        afr_lookup_set_read_child (this, local);
-        up_count = afr_up_children_count (priv->child_count, priv->child_up);
+        up_count  = afr_up_children_count (priv->child_count, local->child_up);
         if (up_count == 1) {
                 gf_log (this->name, GF_LOG_DEBUG,
                         "Only 1 child up - do not attempt to detect self heal");
+                goto out;
+        }
 
-                goto unwind;
+        if (_gf_false == afr_is_self_heal_enabled (priv)) {
+                gf_log (this->name, GF_LOG_DEBUG,
+                        "Self heal is not enabled");
+                goto out;
         }
 
-        if (local->success_count && local->enoent_count) {
-                local->self_heal.need_metadata_self_heal = _gf_true;
-                local->self_heal.need_data_self_heal     = _gf_true;
-                local->self_heal.need_entry_self_heal    = _gf_true;
-                gf_log(this->name, GF_LOG_INFO,
-                       "entries are missing in lookup of %s.",
-                       local->loc.path);
+        afr_lookup_detect_self_heal (local, this);
+        if (afr_can_self_heal_proceed (&local->self_heal, priv)) {
+                if  (afr_is_self_heal_running (local)) {
+                        goto out;
+                }
+
+                afr_launch_self_heal (frame, this, _gf_true,
+                                      local->cont.lookup.buf.ia_type,
+                                      afr_self_heal_lookup_unwind);
+                *sh_launched = _gf_true;
         }
+out:
+        return;
+}
 
-        if (local->success_count) {
-                /* check for split-brain case in previous lookup */
-                if (afr_is_split_brain (this, local->cont.lookup.inode)) {
-                        local->self_heal.need_data_self_heal = _gf_true;
-                        gf_log(this->name, GF_LOG_WARNING,
-                               "split brain detected during lookup of %s.",
-                               local->loc.path);
+static gf_boolean_t
+afr_lookup_split_brain (afr_local_t *local, xlator_t *this)
+{
+        int             i              = 0;
+        gf_boolean_t    symptom        = _gf_false;
+        struct iatt     *bufs          = NULL;
+        int32_t         *child_success = NULL;
+        struct iatt     *child1        = NULL;
+        struct iatt     *child2        = NULL;
+        const char      *path          = NULL;
+
+        bufs = local->cont.lookup.bufs;
+        child_success = local->cont.lookup.child_success;
+        for (i = 1; i < local->success_count; i++) {
+                child1 = &bufs[child_success[i-1]];
+                child2 = &bufs[child_success[i]];
+                /*
+                 * TODO: gfid self-heal
+                 * if (uuid_compare (child1->ia_gfid, child2->ia_gfid)) {
+                 *        gf_log (this->name, GF_LOG_WARNING, "%s: gfid differs"
+                 *                " on subvolumes (%d, %d)", local->loc.path,
+                 *                child_success[i-1], child_success[i]);
+                 *        symptom = _gf_true;
+                 * }
+                 */
+
+                if (FILETYPE_DIFFERS (child1, child2)) {
+                        path = local->loc.path;
+                        gf_log (this->name, GF_LOG_WARNING, "%s: filetype "
+                                "differs on subvolumes (%d, %d)", path,
+                                child_success[i-1], child_success[i]);
+                        symptom = _gf_true;
+                        local->govinda_gOvinda = 1;
                 }
+                if (symptom)
+                        break;
         }
+        return symptom;
+}
 
-        if (local->self_heal.need_metadata_self_heal
-             || local->self_heal.need_data_self_heal
-             || local->self_heal.need_entry_self_heal)
-             {
+static int
+afr_lookup_set_read_child (afr_local_t *local, xlator_t *this, int32_t read_child)
+{
+        GF_ASSERT (read_child >= 0);
 
-                if (local->inodelk_count || local->entrylk_count) {
+        afr_set_read_child (this, local->cont.lookup.inode, read_child);
+        local->cont.lookup.read_child = read_child;
 
-                        /* Someone else is doing self-heal on this file.
-                           return */
+        return 0;
+}
 
-                        goto unwind;
-                }
+static void
+afr_lookup_done (call_frame_t *frame, xlator_t *this)
+{
+        int                 unwind = 1;
+        afr_private_t       *priv  = NULL;
+        afr_local_t         *local = NULL;
+        int                 ret = -1;
+        gf_boolean_t        sh_launched = _gf_false;
+        int32_t             read_child = -1;
 
-                if (!local->cont.lookup.inode->ia_type) {
-                        /* fix for RT #602 */
-                        local->cont.lookup.inode->ia_type =
-                                lookup_buf->ia_type;
-                }
+        priv  = this->private;
+        local = frame->local;
 
-                local->self_heal.background = _gf_true;
-                local->self_heal.type       = local->cont.lookup.buf.ia_type;
-                local->self_heal.unwind     = afr_self_heal_lookup_unwind;
+        if (local->op_ret < 0)
+                goto unwind;
 
-                unwind = 0;
+        if (_gf_true == afr_lookup_split_brain (local, this)) {
+                local->op_ret = -1;
+                local->op_errno = EIO;
+                goto unwind;
+        }
 
-                afr_self_heal_type_str_get(&local->self_heal,
-                                           sh_type_str,
-                                           sizeof(sh_type_str));
+        ret = afr_lookup_select_read_child (local, this, &read_child);
+        if (ret) {
+                local->op_ret = -1;
+                local->op_errno = EIO;
+                goto unwind;
+        }
 
-                gf_log (this->name, GF_LOG_INFO,
-                        "background %s self-heal triggered. path: %s",
-                        sh_type_str, local->loc.path);
+        ret = afr_lookup_set_read_child (local, this, read_child);
+        if (ret)
+                goto unwind;
 
-                afr_self_heal (frame, this);
+        afr_lookup_build_response_params (local, this);
+        if (afr_is_fresh_lookup (&local->loc, this)) {
+                afr_update_loc_gfids (&local->loc, &local->cont.lookup.buf,
+                                      &local->cont.lookup.postparent);
         }
 
-unwind:
-        if (unwind) {
-                AFR_STACK_UNWIND (lookup, frame, local->op_ret,
-                                  local->op_errno,
-                                  local->cont.lookup.inode,
-                                  &local->cont.lookup.buf,
-                                  local->cont.lookup.xattr,
-                                  &local->cont.lookup.postparent);
+        afr_lookup_perform_self_heal_if_needed (frame, this, &sh_launched);
+        if (sh_launched)
+                unwind = 0;
+ unwind:
+         if (unwind) {
+                 AFR_STACK_UNWIND (lookup, frame, local->op_ret,
+                                  local->op_errno, local->cont.lookup.inode,
+                                   &local->cont.lookup.buf,
+                                   local->cont.lookup.xattr,
+                                   &local->cont.lookup.postparent);
         }
 }
 
-
 /*
  * During a lookup, some errors are more "important" than
  * others in that they must be given higher priority while
@@ -850,259 +1050,169 @@ __error_more_important (int32_t old_errno, int32_t new_errno)
         return ret;
 }
 
-
-int
-afr_fresh_lookup_cbk (call_frame_t *frame, void *cookie,
-                      xlator_t *this,  int32_t op_ret,  int32_t op_errno,
-                      inode_t *inode,   struct iatt *buf, dict_t *xattr,
-                      struct iatt *postparent)
+static void
+afr_lookup_handle_error (afr_local_t *local, int32_t op_ret,  int32_t op_errno)
 {
-        afr_local_t *   local = NULL;
-        afr_private_t * priv  = NULL;
-        struct iatt *   lookup_buf = NULL;
-        int             call_count      = -1;
-        int             child_index     = -1;
-        int             first_up_child  = -1;
+        GF_ASSERT (local);
+        if (op_errno == ENOENT)
+                local->enoent_count++;
 
-        child_index = (long) cookie;
-        priv = this->private;
-
-        LOCK (&frame->lock);
-        {
-                local = frame->local;
-
-                lookup_buf = &local->cont.lookup.buf;
-
-                if (op_ret == -1) {
-                        if (op_errno == ENOENT)
-                                local->enoent_count++;
-
-                        if (__error_more_important (local->op_errno, op_errno))
-                                local->op_errno = op_errno;
-
-                        if (local->op_errno == ESTALE) {
-                                local->op_ret = -1;
-                        }
-
-                        goto unlock;
-                }
-
-                afr_lookup_collect_xattr (local, this, child_index, xattr);
-
-                first_up_child = afr_first_up_child (priv);
-
-                if (child_index == first_up_child) {
-                        local->cont.lookup.ino =
-                                afr_itransform (buf->ia_ino,
-                                                priv->child_count,
-                                                first_up_child);
-                }
-
-                if (local->success_count == 0) {
-                        if (local->op_errno != ESTALE)
-                                local->op_ret = op_ret;
-
-                        local->cont.lookup.inode               = inode_ref (inode);
-                        local->cont.lookup.xattr               = dict_ref (xattr);
-                        local->cont.lookup.xattrs[child_index] = dict_ref (xattr);
-                        local->cont.lookup.postparent          = *postparent;
-
-                        if (priv->first_lookup && inode->ino == 1) {
-                                gf_log (this->name, GF_LOG_INFO,
-                                        "added root inode");
-                                priv->root_inode = inode_ref (inode);
-                                priv->first_lookup = 0;
-                        }
-
-                        *lookup_buf = *buf;
-
-                        uuid_copy (local->loc.gfid, buf->ia_gfid);
-                        uuid_copy (local->loc.pargfid,
-                                   postparent->ia_gfid);
-
-                        lookup_buf->ia_ino = afr_itransform (buf->ia_ino,
-                                                             priv->child_count,
-                                                             child_index);
-                        if (priv->read_child >= 0) {
-                                afr_set_read_child (this,
-                                                    local->cont.lookup.inode,
-                                                    priv->read_child);
-                        } else {
-                                afr_set_read_child (this,
-                                                    local->cont.lookup.inode,
-                                                    child_index);
-                                //prefer fast child
-                                local->read_child_index = child_index;
-                        }
-
-                } else {
-                        afr_lookup_self_heal_check (this, local, buf, lookup_buf);
+        if (__error_more_important (local->op_errno, op_errno))
+                local->op_errno = op_errno;
+        if (local->op_errno == ESTALE) {
+                local->op_ret = -1;
+        }
+}
 
-                        local->cont.lookup.xattrs[child_index] = dict_ref (xattr);
-                        if (child_index == local->read_child_index) {
-                                /*
-                                  lookup has succeeded on the read child.
-                                  So use its inode number
-                                */
-                                if (local->cont.lookup.xattr)
-                                        dict_unref (local->cont.lookup.xattr);
+static void
+afr_set_root_inode_on_first_lookup (afr_local_t *local, xlator_t *this,
+                                    inode_t *inode)
+{
+        afr_private_t           *priv = NULL;
+        GF_ASSERT (inode);
 
-                                local->cont.lookup.xattr = dict_ref (xattr);
-                                local->cont.lookup.postparent          = *postparent;
+        if (inode->ino != 1)
+                goto out;
+        if (!afr_is_fresh_lookup (&local->loc, this))
+                goto out;
+        priv = this->private;
+        if ((priv->first_lookup)) {
+                gf_log (this->name, GF_LOG_INFO, "added root inode");
+                priv->root_inode = inode_ref (inode);
+                priv->first_lookup = 0;
+        }
+out:
+        return;
+}
 
-                                *lookup_buf = *buf;
+static void
+afr_lookup_cache_args (afr_local_t *local, int child_index, dict_t *xattr,
+                       struct iatt *buf, struct iatt *postparent)
+{
+        GF_ASSERT (child_index >= 0);
+        local->cont.lookup.xattrs[child_index] = dict_ref (xattr);
+        local->cont.lookup.postparents[child_index] = *postparent;
+        local->cont.lookup.bufs[child_index] = *buf;
+}
 
-                                uuid_copy (local->loc.gfid, buf->ia_gfid);
-                                uuid_copy (local->loc.pargfid,
-                                           postparent->ia_gfid);
-                        }
+static void
+afr_lookup_handle_first_success (afr_local_t *local, xlator_t *this,
+                                 inode_t *inode, struct iatt *buf)
+{
+        local->cont.lookup.inode      = inode_ref (inode);
+        local->cont.lookup.buf        = *buf;
+        afr_set_root_inode_on_first_lookup (local, this, inode);
+}
 
+static void
+afr_lookup_handle_success (afr_local_t *local, xlator_t *this, int32_t child_index,
+                           int32_t op_ret, int32_t op_errno, inode_t *inode,
+                           struct iatt *buf, dict_t *xattr,
+                           struct iatt *postparent)
+{
+        if (local->success_count == 0) {
+                if (local->op_errno != ESTALE) {
+                        local->op_ret = op_ret;
+                        local->op_errno = 0;
                 }
-
-                local->success_count++;
-        }
-unlock:
-        UNLOCK (&frame->lock);
-
-        call_count = afr_frame_return (frame);
-
-        if (call_count == 0) {
-                afr_lookup_done (frame, this, lookup_buf);
+                afr_lookup_handle_first_success (local, this, inode, buf);
         }
+        afr_lookup_update_lk_counts (local, this,
+                                     child_index, xattr);
 
-        return 0;
+        afr_lookup_cache_args (local, child_index, xattr,
+                               buf, postparent);
+        local->cont.lookup.child_success[local->success_count] = child_index;
+        local->success_count++;
 }
 
-
 int
-afr_revalidate_lookup_cbk (call_frame_t *frame, void *cookie,
-                           xlator_t *this, int32_t op_ret, int32_t op_errno,
-                           inode_t *inode, struct iatt *buf, dict_t *xattr,
-                           struct iatt *postparent)
+afr_lookup_cbk (call_frame_t *frame, void *cookie,
+                xlator_t *this,  int32_t op_ret,  int32_t op_errno,
+                inode_t *inode,   struct iatt *buf, dict_t *xattr,
+                struct iatt *postparent)
 {
         afr_local_t *   local = NULL;
-        afr_private_t * priv  = NULL;
-        struct iatt *   lookup_buf = NULL;
         int             call_count      = -1;
         int             child_index     = -1;
-        int             first_up_child  = -1;
 
-        child_index = (long) cookie;
-        priv = this->private;
+         child_index = (long) cookie;
 
         LOCK (&frame->lock);
         {
                 local = frame->local;
 
-                lookup_buf = &local->cont.lookup.buf;
-
                 if (op_ret == -1) {
-                        if (op_errno == ENOENT)
-                                local->enoent_count++;
-
-                        if (__error_more_important (local->op_errno, op_errno))
-                                local->op_errno = op_errno;
-
-                        if (local->op_errno == ESTALE) {
-                                local->op_ret = -1;
-                        }
-
+                        afr_lookup_handle_error (local, op_ret, op_errno);
                         goto unlock;
                 }
+                afr_lookup_handle_success (local, this, child_index, op_ret,
+                                           op_errno, inode, buf, xattr,
+                                           postparent);
 
-                afr_lookup_collect_xattr (local, this, child_index, xattr);
-
-                first_up_child = afr_first_up_child (priv);
-
-                if (child_index == first_up_child) {
-                        local->cont.lookup.ino =
-                                afr_itransform (buf->ia_ino,
-                                                priv->child_count,
-                                                first_up_child);
-                }
-
-                /* in case of revalidate, we need to send stat of the
-                 * child whose stat was sent during the first lookup.
-                 * (so that time stamp does not vary with revalidate.
-                 * in case it is down, stat of the fist success will
-                 * be replied */
-
-                /* inode number should be preserved across revalidates */
-
-                if (local->success_count == 0) {
-                        if (local->op_errno != ESTALE)
-                                local->op_ret = op_ret;
-
-                        local->cont.lookup.inode               = inode_ref (inode);
-                        local->cont.lookup.xattr               = dict_ref (xattr);
-                        local->cont.lookup.xattrs[child_index] = dict_ref (xattr);
-                        local->cont.lookup.postparent          = *postparent;
-
-                        *lookup_buf = *buf;
-
-                        lookup_buf->ia_ino = afr_itransform (buf->ia_ino,
-                                                             priv->child_count,
-                                                             child_index);
-
-                        if (priv->read_child >= 0) {
-                                afr_set_read_child (this,
-                                                    local->cont.lookup.inode,
-                                                    priv->read_child);
-                        } else {
-                                afr_set_read_child (this,
-                                                    local->cont.lookup.inode,
-                                                    child_index);
-                        }
-
-                } else {
-                        afr_lookup_self_heal_check (this, local, buf, lookup_buf);
-
-                        local->cont.lookup.xattrs[child_index] = dict_ref (xattr);
-                        if (child_index == local->read_child_index) {
+         }
+unlock:
+        UNLOCK (&frame->lock);
 
-                                /*
-                                  lookup has succeeded on the read child.
-                                  So use its inode number
-                                */
+        call_count = afr_frame_return (frame);
+        if (call_count == 0) {
+               afr_lookup_done (frame, this);
+        }
 
-                                if (local->cont.lookup.xattr)
-                                        dict_unref (local->cont.lookup.xattr);
+         return 0;
+}
 
-                                local->cont.lookup.xattr               = dict_ref (xattr);
-                                local->cont.lookup.postparent          = *postparent;
+int
+afr_lookup_cont_init (afr_local_t *local, unsigned int child_count)
+{
+        int               ret            = -ENOMEM;
+        int32_t           *child_success = NULL;
+        struct iatt       *iatts         = NULL;
+        int               i              = 0;
 
-                                *lookup_buf = *buf;
-                        }
+        GF_ASSERT (local);
+        local->cont.lookup.xattrs = GF_CALLOC (child_count,
+                                               sizeof (*local->cont.lookup.xattr),
+                                               gf_afr_mt_dict_t);
+        if (NULL == local->cont.lookup.xattrs)
+                goto out;
 
-                }
+        iatts = GF_CALLOC (child_count, sizeof (*iatts), gf_afr_mt_iatt);
+        if (NULL == iatts)
+                goto out;
+        local->cont.lookup.postparents = iatts;
 
-                local->success_count++;
-        }
-unlock:
-        UNLOCK (&frame->lock);
+        iatts = GF_CALLOC (child_count, sizeof (*iatts), gf_afr_mt_iatt);
+        if (NULL == iatts)
+                goto out;
+        local->cont.lookup.bufs = iatts;
 
-        call_count = afr_frame_return (frame);
+        child_success = GF_CALLOC (child_count, sizeof (*child_success),
+                                   gf_afr_mt_char);
+        if (NULL == child_success)
+                goto out;
+        for (i = 0; i < child_count; i++)
+                child_success[i] = -1;
 
-        if (call_count == 0) {
-                afr_lookup_done (frame, this, lookup_buf);
-        }
+        local->cont.lookup.child_success = child_success;
 
-        return 0;
+        local->cont.lookup.read_child = -1;
+        ret = 0;
+out:
+        return ret;
 }
 
-
 int
 afr_lookup (call_frame_t *frame, xlator_t *this,
             loc_t *loc, dict_t *xattr_req)
 {
-        afr_private_t    *priv       = NULL;
-        afr_local_t      *local      = NULL;
-        int               ret        = -1;
-        int               i          = 0;
-        fop_lookup_cbk_t  callback   = NULL;
-        int               call_count = 0;
-        uint64_t          ctx        = 0;
-        int32_t           op_errno   = 0;
+        afr_private_t    *priv           = NULL;
+        afr_local_t      *local          = NULL;
+        int               ret            = -1;
+        int               i              = 0;
+        int               call_count     = 0;
+        uint64_t          ctx            = 0;
+        int32_t           op_errno       = 0;
 
         priv = this->private;
 
@@ -1123,14 +1233,9 @@ afr_lookup (call_frame_t *frame, xlator_t *this,
         if (ret == 0) {
                 /* lookup is a revalidate */
 
-                callback = afr_revalidate_lookup_cbk;
-
-                local->cont.lookup.is_revalidate = _gf_true;
                 local->read_child_index          = afr_read_child (this,
                                                                    loc->inode);
         } else {
-                callback = afr_fresh_lookup_cbk;
-
                 LOCK (&priv->read_child_lock);
                 {
                         local->read_child_index = (++priv->read_child_rr)
@@ -1143,10 +1248,16 @@ afr_lookup (call_frame_t *frame, xlator_t *this,
                 local->cont.lookup.parent_ino = loc->parent->ino;
 
         local->child_up = memdup (priv->child_up, priv->child_count);
+        if (NULL == local->child_up) {
+                op_errno = ENOMEM;
+                goto out;
+        }
 
-        local->cont.lookup.xattrs = GF_CALLOC (priv->child_count,
-                                               sizeof (*local->cont.lookup.xattr),
-                                               gf_afr_mt_dict_t);
+        ret = afr_lookup_cont_init (local, priv->child_count);
+        if (ret < 0) {
+                op_errno = -ret;
+                goto out;
+        }
 
         local->call_count = afr_up_children_count (priv->child_count,
                                                    local->child_up);
@@ -1192,7 +1303,8 @@ afr_lookup (call_frame_t *frame, xlator_t *this,
 
         for (i = 0; i < priv->child_count; i++) {
                 if (local->child_up[i]) {
-                        STACK_WIND_COOKIE (frame, callback, (void *) (long) i,
+                        STACK_WIND_COOKIE (frame, afr_lookup_cbk,
+                                           (void *) (long) i,
                                            priv->children[i],
                                            priv->children[i]->fops->lookup,
                                            loc, local->xattr_req);
diff --git a/xlators/cluster/afr/src/afr-dir-read.c b/xlators/cluster/afr/src/afr-dir-read.c
index e626a6f1a..3f056b686 100644
--- a/xlators/cluster/afr/src/afr-dir-read.c
+++ b/xlators/cluster/afr/src/afr-dir-read.c
@@ -70,15 +70,19 @@ gf_boolean_t
 __checksums_differ (uint32_t *checksum, int child_count,
                     unsigned char *child_up)
 {
-        int      ret   = _gf_false;
-        int      i     = 0;
-        uint32_t cksum = 0;
-
-        cksum = checksum[0];
+        int          ret            = _gf_false;
+        int          i              = 0;
+        uint32_t     cksum          = 0;
+        gf_boolean_t activate_check = _gf_false;
 
         for (i = 0; i < child_count; i++) {
                 if (!child_up[i])
                         continue;
+                if (_gf_false == activate_check) {
+                        cksum          = checksum[i];
+                        activate_check = _gf_true;
+                        continue;
+                }
 
                 if (cksum != checksum[i]) {
                         ret = _gf_true;
diff --git a/xlators/cluster/afr/src/afr-self-heal-common.c b/xlators/cluster/afr/src/afr-self-heal-common.c
index 4f50f3577..b29deb8bc 100644
--- a/xlators/cluster/afr/src/afr-self-heal-common.c
+++ b/xlators/cluster/afr/src/afr-self-heal-common.c
@@ -70,23 +70,6 @@ afr_sh_source_count (int sources[], int child_count)
         return nsource;
 }
 
-
-int
-afr_sh_supress_errenous_children (int sources[], int child_errno[],
-                                  int child_count)
-{
-        int i = 0;
-
-        for (i = 0; i < child_count; i++) {
-                if (child_errno[i] && sources[i]) {
-                        sources[i] = 0;
-                }
-        }
-
-        return 0;
-}
-
-
 void
 afr_sh_print_pending_matrix (int32_t *pending_matrix[], xlator_t *this)
 {
@@ -113,11 +96,46 @@ afr_sh_print_pending_matrix (int32_t *pending_matrix[], xlator_t *this)
         GF_FREE (buf);
 }
 
+void
+afr_init_pending_matrix (int32_t **pending_matrix, size_t child_count)
+{
+        int             i   = 0;
+        int             j   = 0;
+
+        GF_ASSERT (pending_matrix);
+
+        for (i = 0; i < child_count; i++) {
+                for (j = 0; j < child_count; j++) {
+                        pending_matrix[i][j] = 0;
+                }
+        }
+}
 
 void
-afr_sh_build_pending_matrix (afr_private_t *priv,
-                             int32_t *pending_matrix[], dict_t *xattr[],
-                             int child_count, afr_transaction_type type)
+afr_mark_ignorant_subvols_as_pending (int32_t **pending_matrix,
+                                      unsigned char *ignorant_subvols,
+                                      size_t  child_count)
+{
+        int            i                = 0;
+        int            j                = 0;
+
+        GF_ASSERT (pending_matrix);
+        GF_ASSERT (ignorant_subvols);
+
+        for (i = 0; i < child_count; i++) {
+                if (ignorant_subvols[i]) {
+                        for (j = 0; j < child_count; j++) {
+                                if (!ignorant_subvols[j])
+                                        pending_matrix[j][i] += 1;
+                        }
+                }
+        }
+}
+
+int
+afr_build_pending_matrix (char **pending_key, int32_t **pending_matrix,
+                          dict_t *xattr[], afr_transaction_type type,
+                          size_t child_count)
 {
         /* Indexable by result of afr_index_for_transaction_type(): 0 -- 2. */
         int32_t        pending[3]       = {0,};
@@ -130,19 +148,16 @@ afr_sh_build_pending_matrix (afr_private_t *priv,
 
         ignorant_subvols = GF_CALLOC (sizeof (*ignorant_subvols), child_count,
                                       gf_afr_mt_char);
+        if (NULL == ignorant_subvols)
+                goto out;
 
-        /* start clean */
-        for (i = 0; i < child_count; i++) {
-                for (j = 0; j < child_count; j++) {
-                        pending_matrix[i][j] = 0;
-                }
-        }
+        afr_init_pending_matrix (pending_matrix, child_count);
 
         for (i = 0; i < child_count; i++) {
                 pending_raw = NULL;
 
                 for (j = 0; j < child_count; j++) {
-                        ret = dict_get_ptr (xattr[i], priv->pending_key[j],
+                        ret = dict_get_ptr (xattr[i], pending_key[j],
                                             &pending_raw);
 
                         if (ret != 0) {
@@ -163,21 +178,12 @@ afr_sh_build_pending_matrix (afr_private_t *priv,
                 }
         }
 
-        /*
-         * Make all non-ignorant subvols point towards the ignorant
-         * subvolumes.
-         */
-
-        for (i = 0; i < child_count; i++) {
-                if (ignorant_subvols[i]) {
-                        for (j = 0; j < child_count; j++) {
-                                if (!ignorant_subvols[j])
-                                        pending_matrix[j][i] += 1;
-                        }
-                }
-        }
-
+        afr_mark_ignorant_subvols_as_pending (pending_matrix,
+                                              ignorant_subvols,
+                                              child_count);
         GF_FREE (ignorant_subvols);
+out:
+        return ret;
 }
 
 
@@ -208,7 +214,8 @@ afr_sh_build_pending_matrix (afr_private_t *priv,
 typedef enum {
         AFR_NODE_INNOCENT,
         AFR_NODE_FOOL,
-        AFR_NODE_WISE
+        AFR_NODE_WISE,
+        AFR_NODE_INVALID = -1,
 } afr_node_type;
 
 typedef struct {
@@ -353,182 +360,276 @@ afr_sh_mark_wisest_as_sources (int sources[],
         return nsources;
 }
 
-
-static int
-afr_sh_mark_if_size_differs (afr_self_heal_t *sh, int child_count)
+static void
+afr_compute_witness_of_fools (int32_t *witnesses, int32_t **pending_matrix,
+                              afr_node_character *characters,
+                              int32_t child_count)
 {
-        int32_t ** pending_matrix = NULL;
-        int        i              = 0;
-        int        j              = 0;
-        int        size_differs   = 0;
+        int i       = 0;
+        int j       = 0;
+        int witness = 0;
 
-        pending_matrix = sh->pending_matrix;
+        GF_ASSERT (witnesses);
+        GF_ASSERT (pending_matrix);
+        GF_ASSERT (characters);
+        GF_ASSERT (child_count > 0);
 
         for (i = 0; i < child_count; i++) {
+                if (characters[i].type != AFR_NODE_FOOL)
+                        continue;
+
+                witness = 0;
                 for (j = 0; j < child_count; j++) {
-                        if (!sh->buf)
-                                break;
+                        if (i == j)
+                                continue;
+                        witness += pending_matrix[i][j];
+                }
+                witnesses[i] = witness;
+        }
+}
 
-                        if (SIZE_DIFFERS (&sh->buf[i], &sh->buf[j])
-                            && (pending_matrix[i][j] == 0)
-                            && (pending_matrix[j][i] == 0)) {
+static int32_t
+afr_find_biggest_witness_among_fools (int32_t *witnesses,
+                                      afr_node_character *characters,
+                                      int32_t child_count)
+{
+        int i               = 0;
+        int biggest_witness = -1;
 
-                                pending_matrix[i][j] = 1;
-                                pending_matrix[j][i] = 1;
+        GF_ASSERT (witnesses);
+        GF_ASSERT (characters);
+        GF_ASSERT (child_count > 0);
 
-                                size_differs = 1;
-                        }
-                }
-        }
+        for (i = 0; i < child_count; i++) {
+                if (characters[i].type != AFR_NODE_FOOL)
+                        continue;
 
-        return size_differs;
+                if (biggest_witness < witnesses[i])
+                        biggest_witness = witnesses[i];
+        }
+        return biggest_witness;
 }
 
-
-static int
-afr_sh_mark_biggest_fool_as_source (afr_self_heal_t *sh,
+int
+afr_mark_fool_as_source_by_witness (int32_t *sources, int32_t *witnesses,
                                     afr_node_character *characters,
-                                    int child_count)
+                                    int32_t child_count, int32_t witness)
 {
-        int i       = 0;
-        int biggest = 0;
+        int i        = 0;
+        int nsources = 0;
 
-        for (i = 0; i < child_count; i++) {
-                if (characters[i].type == AFR_NODE_FOOL) {
-                        biggest = i;
-                        break;
-                }
-        }
+        GF_ASSERT (sources);
+        GF_ASSERT (witnesses);
+        GF_ASSERT (characters);
+        GF_ASSERT (child_count > 0);
 
         for (i = 0; i < child_count; i++) {
                 if (characters[i].type != AFR_NODE_FOOL)
                         continue;
 
-                if (!sh->buf)
-                        break;
-
-                if (SIZE_GREATER (&sh->buf[i], &sh->buf[biggest])) {
-                        biggest = i;
+                if (witness == witnesses[i]) {
+                        sources[i] = 1;
+                        nsources++;
                 }
         }
+        return nsources;
+}
+
+static int
+afr_mark_biggest_of_fools_as_source (int32_t *sources, int32_t **pending_matrix,
+                                     afr_node_character *characters,
+                                     int child_count)
+{
+        int32_t       biggest_witness = 0;
+        int           nsources        = 0;
+        int32_t       *witnesses      = NULL;
 
-        sh->sources[biggest] = 1;
+        GF_ASSERT (child_count > 0);
 
-        return 1;
-}
+        witnesses = GF_CALLOC (child_count, sizeof (*witnesses),
+                               gf_afr_mt_int32_t);
+        if (NULL == witnesses) {
+                nsources = -1;
+                goto out;
+        }
 
+        afr_compute_witness_of_fools (witnesses, pending_matrix, characters,
+                                      child_count);
+        biggest_witness = afr_find_biggest_witness_among_fools (witnesses,
+                                                                characters,
+                                                                child_count);
+        nsources = afr_mark_fool_as_source_by_witness (sources, witnesses,
+                                                       characters, child_count,
+                                                       biggest_witness);
+out:
+        if (witnesses)
+                GF_FREE (witnesses);
+        return nsources;
+}
 
-static int
-afr_sh_mark_biggest_as_source (afr_self_heal_t *sh, int child_count)
+int
+afr_mark_child_as_source_by_uid (int32_t *sources, struct iatt *bufs,
+                                 int32_t *valid_children, int child_count,
+                                 uint32_t uid)
 {
-        int biggest = 0;
-        int i       = 0;
+        int     i        = 0;
+        int     nsources = 0;
+        int     child    = 0;
+
+        GF_ASSERT (bufs);
+        GF_ASSERT (valid_children);
+        GF_ASSERT (sources);
+        GF_ASSERT (child_count > 0);
 
         for (i = 0; i < child_count; i++) {
-                if (!sh->buf)
-                        break;
+                if (-1 == valid_children[i])
+                        continue;
 
-                if (SIZE_GREATER (&sh->buf[i], &sh->buf[biggest])) {
-                        biggest = i;
+                child = valid_children[i];
+                if (uid == bufs[child].ia_uid) {
+                        sources[child] = 1;
+                        nsources++;
                 }
         }
+        return nsources;
+}
 
-        sh->sources[biggest] = 1;
+int
+afr_get_child_with_lowest_uid (struct iatt *bufs, int32_t *valid_children,
+                               int child_count)
+{
+        int     i        = 0;
+        int     smallest = -1;
+        int     child    = 0;
 
-        return 1;
-}
+        GF_ASSERT (bufs);
+        GF_ASSERT (valid_children);
+        GF_ASSERT (child_count > 0);
 
+        for (i = 0; i < child_count; i++) {
+                if (-1 == valid_children[i])
+                        continue;
+                child = valid_children[i];
+                if ((smallest == -1) ||
+                    (bufs[child].ia_uid < bufs[smallest].ia_uid)) {
+                        smallest = child;
+                }
+        }
+        return smallest;
+}
 
 static int
-afr_sh_mark_loweia_uid_as_source (afr_self_heal_t *sh, int child_count)
+afr_sh_mark_lowest_uid_as_source (struct iatt *bufs, int32_t *valid_children,
+                                  int child_count, int32_t *sources)
 {
-        uid_t smallest = 0;
-        int   i        = 0;
+        int   nsources              = 0;
+        int   smallest              = 0;
 
-        for (i = 0; i < child_count; i++) {
-                if (!sh->buf)
-                        break;
-
-                if (sh->buf[i].ia_uid < sh->buf[smallest].ia_uid) {
-                        smallest = i;
-                }
+        smallest = afr_get_child_with_lowest_uid (bufs, valid_children,
+                                                  child_count);
+        if (smallest < 0) {
+                nsources = -1;
+                goto out;
         }
+        nsources = afr_mark_child_as_source_by_uid (sources, bufs,
+                                                    valid_children, child_count,
+                                                    bufs[smallest].ia_uid);
+out:
+        return nsources;
+}
 
-        sh->sources[smallest] = 1;
+char *
+afr_get_character_str (afr_node_type type)
+{
+        char *character = NULL;
 
-        return 1;
+        switch (type) {
+        case AFR_NODE_INNOCENT:
+                character = "innocent";
+                break;
+        case AFR_NODE_FOOL:
+                character = "fool";
+                break;
+        case AFR_NODE_WISE:
+                character = "wise";
+                break;
+        default:
+                character = "invalid";
+                break;
+        }
+        return character;
 }
 
+afr_node_type
+afr_find_child_character_type (int32_t *pending_row, int32_t child,
+                               int32_t child_count, const char *xlator_name)
+{
+        afr_node_type type = AFR_NODE_INVALID;
+
+        GF_ASSERT (pending_row);
+        GF_ASSERT (child_count > 0);
+        GF_ASSERT ((child >= 0) && (child < child_count));
+
+        if (afr_sh_is_innocent (pending_row, child_count))
+                type = AFR_NODE_INNOCENT;
+        else if (afr_sh_is_fool (pending_row, child, child_count))
+                type = AFR_NODE_FOOL;
+        else if (afr_sh_is_wise (pending_row, child, child_count))
+                type = AFR_NODE_WISE;
+        else
+                GF_ASSERT (0);
+
+        gf_log (xlator_name, GF_LOG_DEBUG, "child %d character %s",
+                child, afr_get_character_str (type));
+        return type;
+}
 
 int
-afr_sh_mark_sources (afr_self_heal_t *sh, int child_count,
-                     afr_self_heal_type type)
+afr_mark_sources (int32_t *sources, int32_t **pending_matrix, struct iatt *bufs,
+                  int32_t child_count, afr_self_heal_type type,
+                  int32_t *valid_children, const char *xlator_name)
 {
         /* stores the 'characters' (innocent, fool, wise) of the nodes */
-        afr_node_character *characters =  NULL;
 
+        afr_node_character *characters =  NULL;
         int            i              = 0;
-        int32_t **     pending_matrix = NULL;
-        int *          sources        = NULL;
-        int            size_differs   = 0;
-        int            nsources       = 0;
+        int            nsources       = -1;
         xlator_t      *this           = NULL;
-        afr_private_t *priv           = NULL;
 
         characters = GF_CALLOC (sizeof (afr_node_character),
-                                        child_count,
-                                        gf_afr_mt_afr_node_character) ;
+                                child_count, gf_afr_mt_afr_node_character);
         if (!characters)
                 goto out;
 
         this = THIS;
-        priv = this->private;
-        pending_matrix = sh->pending_matrix;
-        sources        = sh->sources;
 
         /* start clean */
         for (i = 0; i < child_count; i++) {
                 sources[i] = 0;
         }
 
+        nsources = 0;
         for (i = 0; i < child_count; i++) {
-                if (afr_sh_is_innocent (pending_matrix[i], child_count)) {
-                        characters[i].type = AFR_NODE_INNOCENT;
-
-                } else if (afr_sh_is_fool (pending_matrix[i], i, child_count)) {
-                        characters[i].type = AFR_NODE_FOOL;
-
-                } else if (afr_sh_is_wise (pending_matrix[i], i, child_count)) {
-                        characters[i].type = AFR_NODE_WISE;
-
-                } else {
-                        gf_log (this->name, GF_LOG_CRITICAL,
-                                "Could not determine the state of subvolume %s!"
-                                " (This message should never appear."
-                                " Please file a bug report to "
-                                "<gluster-devel@nongnu.org>.)",
-                                priv->children[i]->name);
-                }
-        }
-
-        if (type == AFR_SELF_HEAL_DATA) {
-                size_differs = afr_sh_mark_if_size_differs (sh, child_count);
+                characters[i].type =
+                        afr_find_child_character_type (pending_matrix[i], i,
+                                                       child_count,
+                                                       xlator_name);
+                if (AFR_NODE_INVALID == characters[i].type)
+                        gf_log (xlator_name, GF_LOG_WARNING,
+                                "child %d had invalid xattrs", i);
         }
 
         if ((type == AFR_SELF_HEAL_METADATA)
             && afr_sh_all_nodes_innocent (characters, child_count)) {
 
-                nsources = afr_sh_mark_loweia_uid_as_source (sh, child_count);
+                nsources = afr_sh_mark_lowest_uid_as_source (bufs,
+                                                             valid_children,
+                                                             child_count,
+                                                             sources);
                 goto out;
         }
 
-        if (afr_sh_all_nodes_innocent (characters, child_count)) {
-                if (size_differs) {
-                        nsources = afr_sh_mark_biggest_as_source (sh,
-                                                                  child_count);
-                }
-
-        } else if (afr_sh_wise_nodes_exist (characters, child_count)) {
+        if (afr_sh_wise_nodes_exist (characters, child_count)) {
                 afr_sh_compute_wisdom (pending_matrix, characters, child_count);
 
                 if (afr_sh_wise_nodes_conflict (characters, child_count)) {
@@ -536,7 +637,6 @@ afr_sh_mark_sources (afr_self_heal_t *sh, int child_count,
                         gf_log (this->name, GF_LOG_INFO,
                                 "split-brain possible, no source detected");
                         nsources = -1;
-                        goto out;
 
                 } else {
                         nsources = afr_sh_mark_wisest_as_sources (sources,
@@ -544,18 +644,26 @@ afr_sh_mark_sources (afr_self_heal_t *sh, int child_count,
                                                                   child_count);
                 }
         } else {
-                nsources = afr_sh_mark_biggest_fool_as_source (sh, characters,
-                                                               child_count);
+                nsources = afr_mark_biggest_of_fools_as_source (sources,
+                                                                pending_matrix,
+                                                                characters,
+                                                                child_count);
         }
 
 out:
+        if (nsources == 0) {
+                for (i = 0; i < child_count; i++) {
+                        if (valid_children[i] != -1)
+                                sources[valid_children[i]] = 1;
+                }
+        }
         if (characters)
                 GF_FREE (characters);
 
+        gf_log (this->name, GF_LOG_DEBUG, "Number of sources: %d", nsources);
         return nsources;
 }
 
-
 void
 afr_sh_pending_to_delta (afr_private_t *priv, dict_t **xattr,
                          int32_t *delta_matrix[], int success[],
@@ -643,7 +751,7 @@ afr_sh_delta_to_xattr (afr_private_t *priv,
 
 
 int
-afr_sh_has_metadata_pending (dict_t *xattr, int child_count, xlator_t *this)
+afr_sh_has_metadata_pending (dict_t *xattr, xlator_t *this)
 {
         /* Indexable by result of afr_index_for_transaction_type(): 0 -- 2. */
         int32_t        pending[3]  = {0,};
@@ -674,7 +782,7 @@ afr_sh_has_metadata_pending (dict_t *xattr, int child_count, xlator_t *this)
 
 
 int
-afr_sh_has_data_pending (dict_t *xattr, int child_count, xlator_t *this)
+afr_sh_has_data_pending (dict_t *xattr, xlator_t *this)
 {
         /* Indexable by result of afr_index_for_transaction_type(): 0 -- 2. */
         int32_t        pending[3]  = {0,};
@@ -705,7 +813,7 @@ afr_sh_has_data_pending (dict_t *xattr, int child_count, xlator_t *this)
 
 
 int
-afr_sh_has_entry_pending (dict_t *xattr, int child_count, xlator_t *this)
+afr_sh_has_entry_pending (dict_t *xattr, xlator_t *this)
 {
         /* Indexable by result of afr_index_for_transaction_type(): 0 -- 2. */
         int32_t        pending[3]  = {0,};
@@ -1655,6 +1763,9 @@ afr_self_heal (call_frame_t *frame, xlator_t *this)
                                                  priv->child_count,
                                                  gf_afr_mt_int32_t);
         }
+        sh->child_success = GF_CALLOC (sizeof (*sh->child_success),
+                                       priv->child_count, gf_afr_mt_int32_t);
+
 
         FRAME_SU_DO (sh_frame, afr_local_t);
         if (local->success_count && local->enoent_count) {
@@ -1688,3 +1799,25 @@ afr_self_heal_type_str_get (afr_self_heal_t *self_heal_p, char *str,
                 snprintf(str + strlen(str), size - strlen(str), " entry");
         }
 }
+
+afr_self_heal_type
+afr_self_heal_type_for_transaction (afr_transaction_type type)
+{
+        afr_self_heal_type sh_type = AFR_SELF_HEAL_INVALID;
+
+        switch (type) {
+        case AFR_DATA_TRANSACTION:
+                sh_type = AFR_SELF_HEAL_DATA;
+                break;
+        case AFR_METADATA_TRANSACTION:
+                sh_type = AFR_SELF_HEAL_METADATA;
+                break;
+        case AFR_ENTRY_TRANSACTION:
+                sh_type = AFR_SELF_HEAL_ENTRY;
+                break;
+        case AFR_ENTRY_RENAME_TRANSACTION:
+                GF_ASSERT (0);
+                break;
+        }
+        return sh_type;
+}
diff --git a/xlators/cluster/afr/src/afr-self-heal-common.h b/xlators/cluster/afr/src/afr-self-heal-common.h
index 87856761c..676e933ae 100644
--- a/xlators/cluster/afr/src/afr-self-heal-common.h
+++ b/xlators/cluster/afr/src/afr-self-heal-common.h
@@ -26,6 +26,7 @@ typedef enum {
         AFR_SELF_HEAL_ENTRY,
         AFR_SELF_HEAL_METADATA,
         AFR_SELF_HEAL_DATA,
+        AFR_SELF_HEAL_INVALID = -1,
 } afr_self_heal_type;
 
 int
@@ -37,17 +38,13 @@ afr_sh_sink_count (int sources[], int child_count);
 int
 afr_sh_source_count (int sources[], int child_count);
 
-int
-afr_sh_supress_errenous_children (int sources[], int child_errno[],
-				  int child_count);
-
 void
 afr_sh_print_pending_matrix (int32_t *pending_matrix[], xlator_t *this);
 
-void
-afr_sh_build_pending_matrix (afr_private_t *priv,
-                             int32_t *pending_matrix[], dict_t *xattr[],
-			     int child_count, afr_transaction_type type);
+int
+afr_build_pending_matrix (char **pending_key, int32_t **pending_matrix,
+                          dict_t *xattr[], afr_transaction_type type,
+                          size_t child_count);
 
 void
 afr_sh_pending_to_delta (afr_private_t *priv, dict_t **xattr,
@@ -55,8 +52,9 @@ afr_sh_pending_to_delta (afr_private_t *priv, dict_t **xattr,
                          int child_count, afr_transaction_type type);
 
 int
-afr_sh_mark_sources (afr_self_heal_t *sh, int child_count,
-                     afr_self_heal_type type);
+afr_mark_sources (int32_t *sources, int32_t **pending_matrix, struct iatt *bufs,
+                  int32_t child_count, afr_self_heal_type type,
+                  int32_t *valid_children, const char *xlator_name);
 
 int
 afr_sh_delta_to_xattr (afr_private_t *priv,
@@ -70,4 +68,7 @@ void
 afr_self_heal_type_str_get (afr_self_heal_t *self_heal_p, char *str,
                             size_t size);
 
+afr_self_heal_type
+afr_self_heal_type_for_transaction (afr_transaction_type type);
+
 #endif /* __AFR_SELF_HEAL_COMMON_H__ */
diff --git a/xlators/cluster/afr/src/afr-self-heal-data.c b/xlators/cluster/afr/src/afr-self-heal-data.c
index 0b9e4218c..38799db70 100644
--- a/xlators/cluster/afr/src/afr-self-heal-data.c
+++ b/xlators/cluster/afr/src/afr-self-heal-data.c
@@ -594,16 +594,15 @@ afr_sh_data_fix (call_frame_t *frame, xlator_t *this)
         sh = &local->self_heal;
         priv = this->private;
 
-        afr_sh_build_pending_matrix (priv, sh->pending_matrix, sh->xattr,
-                                     priv->child_count, AFR_DATA_TRANSACTION);
+        afr_build_pending_matrix (priv->pending_key, sh->pending_matrix,
+                                  sh->xattr, AFR_DATA_TRANSACTION,
+                                  priv->child_count);
 
         afr_sh_print_pending_matrix (sh->pending_matrix, this);
 
-        nsources = afr_sh_mark_sources (sh, priv->child_count,
-                                        AFR_SELF_HEAL_DATA);
-
-        afr_sh_supress_errenous_children (sh->sources, sh->child_errno,
-                                          priv->child_count);
+        nsources = afr_mark_sources (sh->sources, sh->pending_matrix, sh->buf,
+                                     priv->child_count, AFR_SELF_HEAL_DATA,
+                                     sh->child_success, this->name);
 
         if (nsources == 0) {
                 gf_log (this->name, GF_LOG_TRACE,
@@ -692,53 +691,164 @@ afr_sh_data_fix (call_frame_t *frame, xlator_t *this)
         return 0;
 }
 
+gf_boolean_t
+afr_is_fresh_read_child (int32_t *sources, int32_t child_count,
+                         int32_t read_child)
+{
+        gf_boolean_t             is_fresh_child = _gf_false;
+
+        GF_ASSERT (read_child < child_count);
 
-void
-afr_self_heal_find_sources (xlator_t *this, afr_local_t *local, dict_t **xattr,
-                            afr_transaction_type transaction_type)
+        if ((read_child >= 0) && (read_child < child_count) &&
+             sources[read_child]) {
+                is_fresh_child = _gf_true;
+        }
+        return is_fresh_child;
+}
+
+static int
+afr_select_read_child_from_policy (int32_t *sources, int32_t child_count,
+                                   int32_t prev_read_child,
+                                   int32_t config_read_child,
+                                   int32_t *valid_children)
 {
-        afr_self_heal_t *sh   = NULL;
-        afr_private_t   *priv = NULL;
-        int              i = 0;
-        afr_self_heal_type sh_type = AFR_SELF_HEAL_DATA;
-        int              nsources = 0;
+        int32_t                  read_child = -1;
+        int                      i          = 0;
 
-        sh   = &local->self_heal;
-        priv = this->private;
+        GF_ASSERT (sources);
 
-        sh->pending_matrix = GF_CALLOC (sizeof (int32_t *), priv->child_count,
-                                        gf_afr_mt_int32_t);
-        for (i = 0; i < priv->child_count; i++) {
-                sh->pending_matrix[i] = GF_CALLOC (sizeof (int32_t),
-                                                   priv->child_count,
-                                                   gf_afr_mt_int32_t);
+        read_child = prev_read_child;
+        if (_gf_true == afr_is_fresh_read_child (sources, child_count,
+                                                 read_child))
+                goto out;
+
+        read_child = config_read_child;
+        if (_gf_true == afr_is_fresh_read_child (sources, child_count,
+                                                 read_child))
+                goto out;
+
+        for (i = 0; i < child_count; i++) {
+                read_child = valid_children[i];
+                if (read_child < 0)
+                        break;
+                if (_gf_true == afr_is_fresh_read_child (sources, child_count,
+                                                         read_child))
+                        goto out;
         }
+        read_child = -1;
+
+out:
+        return read_child;
+}
 
-        sh->sources = GF_CALLOC (priv->child_count, sizeof (*sh->sources),
-                                 gf_afr_mt_int32_t);
-
-        afr_sh_build_pending_matrix (priv, sh->pending_matrix, xattr,
-                                     priv->child_count, transaction_type);
-
-        switch (transaction_type) {
-        case AFR_DATA_TRANSACTION:
-                sh_type = AFR_SELF_HEAL_DATA;
-                break;
-        case AFR_ENTRY_TRANSACTION:
-                sh_type = AFR_SELF_HEAL_ENTRY;
-                break;
-        case AFR_METADATA_TRANSACTION:
-                sh_type = AFR_SELF_HEAL_METADATA;
-                break;
-        default:
-                sh_type = AFR_SELF_HEAL_METADATA;
-                break;
+static void
+afr_destroy_pending_matrix (int32_t **pending_matrix, int32_t child_count)
+{
+        int             i = 0;
+        GF_ASSERT (child_count > 0);
+        if (pending_matrix) {
+                for (i = 0; i < child_count; i++) {
+                        if (pending_matrix[i])
+                                GF_FREE (pending_matrix[i]);
+                }
+                GF_FREE (pending_matrix);
         }
-        nsources = afr_sh_mark_sources (sh, priv->child_count, sh_type);
-        if (nsources == 0) {
-                for (i = 0; i < priv->child_count; i++)
-                        sh->sources[i] = 1;
+}
+
+static int32_t**
+afr_create_pending_matrix (int32_t child_count)
+{
+        gf_boolean_t            cleanup = _gf_false;
+        int32_t                 **pending_matrix = NULL;
+        int                     i = 0;
+
+        GF_ASSERT (child_count > 0);
+
+        pending_matrix = GF_CALLOC (sizeof (*pending_matrix), child_count,
+                                    gf_afr_mt_int32_t);
+        if (NULL == pending_matrix)
+                goto out;
+        for (i = 0; i < child_count; i++) {
+                pending_matrix[i] = GF_CALLOC (sizeof (**pending_matrix),
+                                               child_count,
+                                               gf_afr_mt_int32_t);
+                if (NULL == pending_matrix[i]) {
+                        cleanup = _gf_true;
+                        goto out;
+                }
+        }
+out:
+        if (_gf_true == cleanup) {
+                afr_destroy_pending_matrix (pending_matrix, child_count);
+                pending_matrix = NULL;
+        }
+        return pending_matrix;
+}
+
+int
+afr_lookup_select_read_child_by_txn_type (xlator_t *this, afr_local_t *local,
+                                          dict_t **xattr,
+                                          afr_transaction_type txn_type)
+{
+        afr_private_t            *priv      = NULL;
+        int                      read_child = -1;
+        int                      ret        = -1;
+        afr_self_heal_type       sh_type    = AFR_SELF_HEAL_INVALID;
+        int32_t                  **pending_matrix = NULL;
+        int32_t                  *sources         = NULL;
+        int32_t                  *valid_children  = NULL;
+        struct iatt              *bufs            = NULL;
+        int32_t                  nsources         = 0;
+        int32_t                  prev_read_child  = -1;
+        int32_t                  config_read_child = -1;
+        afr_self_heal_t          *sh = NULL;
+
+        priv = this->private;
+        bufs = local->cont.lookup.bufs;
+        valid_children = local->cont.lookup.child_success;
+        sh = &local->self_heal;
+
+        pending_matrix = afr_create_pending_matrix (priv->child_count);
+        if (NULL == pending_matrix)
+                goto out;
+
+        sources = GF_CALLOC (sizeof (*sources), priv->child_count,
+                             gf_afr_mt_int32_t);
+        if (NULL == sources)
+                goto out;
+
+        afr_build_pending_matrix (priv->pending_key, pending_matrix,
+                                  xattr, txn_type, priv->child_count);
+
+        sh_type = afr_self_heal_type_for_transaction (txn_type);
+        if (AFR_SELF_HEAL_INVALID == sh_type)
+                goto out;
+
+        nsources = afr_mark_sources (sources, pending_matrix, bufs,
+                                     priv->child_count, sh_type,
+                                     valid_children, this->name);
+        if (nsources < 0) {
+                ret = -1;
+                goto out;
+        }
+
+        prev_read_child = local->read_child_index;
+        config_read_child = priv->read_child;
+        read_child = afr_select_read_child_from_policy (sources,
+                                                        priv->child_count,
+                                                        prev_read_child,
+                                                        config_read_child,
+                                                        valid_children);
+        ret = 0;
+        local->cont.lookup.sources = sources;
+out:
+        afr_destroy_pending_matrix (pending_matrix, priv->child_count);
+        if (-1 == ret) {
+                if (sources)
+                        GF_FREE (sources);
         }
+        gf_log (this->name, GF_LOG_DEBUG, "returning read_child: %d", read_child);
+        return read_child;
 }
 
 
@@ -766,6 +876,8 @@ afr_sh_data_fstat_cbk (call_frame_t *frame, void *cookie,
                                 priv->children[child_index]->name);
 
                         sh->buf[child_index] = *buf;
+                        sh->child_success[sh->success_count] = child_index;
+                        sh->success_count++;
                 }
         }
         UNLOCK (&frame->lock);
@@ -798,6 +910,9 @@ afr_sh_data_fstat (call_frame_t *frame, xlator_t *this)
 
         local->call_count = call_count;
 
+        for (i = 0; i < priv->child_count; i++)
+                sh->child_success[i] = -1;
+        sh->success_count = 0;
         for (i = 0; i < priv->child_count; i++) {
                 if (local->child_up[i]) {
                         STACK_WIND_COOKIE (frame, afr_sh_data_fstat_cbk,
diff --git a/xlators/cluster/afr/src/afr-self-heal-entry.c b/xlators/cluster/afr/src/afr-self-heal-entry.c
index 0504da17c..8c619ff45 100644
--- a/xlators/cluster/afr/src/afr-self-heal-entry.c
+++ b/xlators/cluster/afr/src/afr-self-heal-entry.c
@@ -2159,13 +2159,15 @@ afr_sh_entry_fix (call_frame_t *frame, xlator_t *this)
                 goto heal;
         }
 
-        afr_sh_build_pending_matrix (priv, sh->pending_matrix, sh->xattr,
-                                     priv->child_count, AFR_ENTRY_TRANSACTION);
+        afr_build_pending_matrix (priv->pending_key, sh->pending_matrix,
+                                  sh->xattr, AFR_ENTRY_TRANSACTION,
+                                  priv->child_count);
 
         afr_sh_print_pending_matrix (sh->pending_matrix, this);
 
-        nsources = afr_sh_mark_sources (sh, priv->child_count,
-                                        AFR_SELF_HEAL_ENTRY);
+        nsources = afr_mark_sources (sh->sources, sh->pending_matrix, sh->buf,
+                                     priv->child_count, AFR_SELF_HEAL_ENTRY,
+                                     sh->child_success, this->name);
 
         if (nsources == 0) {
                 gf_log (this->name, GF_LOG_TRACE,
@@ -2176,9 +2178,6 @@ afr_sh_entry_fix (call_frame_t *frame, xlator_t *this)
                 return 0;
         }
 
-        afr_sh_supress_errenous_children (sh->sources, sh->child_errno,
-                                          priv->child_count);
-
         source = afr_sh_select_source (sh->sources, priv->child_count);
 
         sh->source = source;
@@ -2211,6 +2210,8 @@ afr_sh_entry_lookup_cbk (call_frame_t *frame, void *cookie,
                 if (op_ret != -1) {
                         sh->xattr[child_index] = dict_ref (xattr);
                         sh->buf[child_index] = *buf;
+                        sh->child_success[sh->success_count] = child_index;
+                        sh->success_count++;
                 }
         }
         UNLOCK (&frame->lock);
@@ -2235,9 +2236,11 @@ afr_sh_entry_lookup (call_frame_t *frame, xlator_t *this)
         int ret = 0;
         int call_count = 0;
         int i = 0;
+        afr_self_heal_t *sh = NULL;
 
         priv  = this->private;
         local = frame->local;
+        sh = &local->self_heal;
 
         call_count = afr_up_children_count (priv->child_count,
                                             local->child_up);
@@ -2257,6 +2260,9 @@ afr_sh_entry_lookup (call_frame_t *frame, xlator_t *this)
                 }
         }
 
+        for (i = 0; i < priv->child_count; i++)
+                sh->child_success[i] = -1;
+        sh->success_count = 0;
         for (i = 0; i < priv->child_count; i++) {
                 if (local->child_up[i]) {
                         STACK_WIND_COOKIE (frame,
diff --git a/xlators/cluster/afr/src/afr-self-heal-metadata.c b/xlators/cluster/afr/src/afr-self-heal-metadata.c
index a4f037c04..1214eefe2 100644
--- a/xlators/cluster/afr/src/afr-self-heal-metadata.c
+++ b/xlators/cluster/afr/src/afr-self-heal-metadata.c
@@ -475,17 +475,15 @@ afr_sh_metadata_fix (call_frame_t *frame, xlator_t *this)
         sh = &local->self_heal;
         priv = this->private;
 
-        afr_sh_build_pending_matrix (priv, sh->pending_matrix, sh->xattr,
-                                     priv->child_count,
-                                     AFR_METADATA_TRANSACTION);
+        afr_build_pending_matrix (priv->pending_key, sh->pending_matrix,
+                                  sh->xattr, AFR_METADATA_TRANSACTION,
+                                  priv->child_count);
 
         afr_sh_print_pending_matrix (sh->pending_matrix, this);
 
-        nsources = afr_sh_mark_sources (sh, priv->child_count,
-                                        AFR_SELF_HEAL_METADATA);
-
-        afr_sh_supress_errenous_children (sh->sources, sh->child_errno,
-                                          priv->child_count);
+        nsources = afr_mark_sources (sh->sources, sh->pending_matrix, sh->buf,
+                                     priv->child_count, AFR_SELF_HEAL_METADATA,
+                                     sh->child_success, this->name);
 
         if (nsources == 0) {
                 gf_log (this->name, GF_LOG_TRACE,
@@ -584,6 +582,8 @@ afr_sh_metadata_lookup_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                         sh->buf[child_index] = *buf;
                         if (xattr)
                                 sh->xattr[child_index] = dict_ref (xattr);
+                        sh->child_success[sh->success_count] = child_index;
+                        sh->success_count++;
                 } else {
                         gf_log (this->name, GF_LOG_INFO,
                                 "path %s on subvolume %s => -1 (%s)",
@@ -614,9 +614,11 @@ afr_sh_metadata_lookup (call_frame_t *frame, xlator_t *this)
         int              call_count = 0;
         dict_t          *xattr_req = NULL;
         int              ret = 0;
+        afr_self_heal_t *sh = NULL;
 
         local = frame->local;
         priv = this->private;
+        sh = &local->self_heal;
 
         call_count = afr_up_children_count (priv->child_count,
                                             local->child_up);
@@ -635,6 +637,9 @@ afr_sh_metadata_lookup (call_frame_t *frame, xlator_t *this)
                 }
         }
 
+        for (i = 0; i < priv->child_count; i++)
+                sh->child_success[i] = -1;
+        sh->success_count = 0;
         for (i = 0; i < priv->child_count; i++) {
                 if (local->child_up[i]) {
                         gf_log (this->name, GF_LOG_TRACE,
diff --git a/xlators/cluster/afr/src/afr-self-heal.h b/xlators/cluster/afr/src/afr-self-heal.h
index 8bf484b48..00e9a1b1e 100644
--- a/xlators/cluster/afr/src/afr-self-heal.h
+++ b/xlators/cluster/afr/src/afr-self-heal.h
@@ -30,11 +30,11 @@
 #define SIZE_GREATER(buf1,buf2) ((buf1)->ia_size > (buf2)->ia_size)
 
 int
-afr_sh_has_metadata_pending (dict_t *xattr, int child_count, xlator_t *this);
+afr_sh_has_metadata_pending (dict_t *xattr, xlator_t *this);
 int
-afr_sh_has_entry_pending (dict_t *xattr, int child_count, xlator_t *this);
+afr_sh_has_entry_pending (dict_t *xattr, xlator_t *this);
 int
-afr_sh_has_data_pending (dict_t *xattr, int child_count, xlator_t *this);
+afr_sh_has_data_pending (dict_t *xattr, xlator_t *this);
 
 int
 afr_self_heal_entry (call_frame_t *frame, xlator_t *this);
@@ -52,4 +52,11 @@ afr_self_heal_find_sources (xlator_t *this, afr_local_t *local, dict_t **xattr,
 int
 afr_self_heal (call_frame_t *frame, xlator_t *this);
 
+gf_boolean_t
+afr_is_fresh_read_child (int32_t *sources, int32_t child_count,
+                         int32_t read_child);
+int
+afr_lookup_select_read_child_by_txn_type (xlator_t *this, afr_local_t *local,
+                                          dict_t **xattr,
+                                          afr_transaction_type txn_type);
 #endif /* __AFR_SELF_HEAL_H__ */
diff --git a/xlators/cluster/afr/src/afr.h b/xlators/cluster/afr/src/afr.h
index 6f40ded12..8f7f54faf 100644
--- a/xlators/cluster/afr/src/afr.h
+++ b/xlators/cluster/afr/src/afr.h
@@ -138,6 +138,10 @@ typedef struct {
         /* array of xattr's, one for each child */
         dict_t **xattr;
 
+        /* array containing if the lookups succeeded in the order of response
+         */
+        int32_t *child_success;
+        int     success_count;
         /* array of errno's, one for each child */
         int *child_errno;
 
@@ -341,14 +345,17 @@ typedef struct _afr_local {
                 struct {
                         inode_t *inode;
                         struct iatt buf;
-                        struct iatt read_child_buf;
                         struct iatt postparent;
                         ino_t ino;
                         uint64_t gen;
                         ino_t parent_ino;
-                        dict_t *xattr;
                         dict_t **xattrs;
-                        gf_boolean_t is_revalidate;
+                        dict_t *xattr;
+                        struct iatt *postparents;
+                        struct iatt *bufs;
+                        int32_t read_child;
+                        int32_t *child_success;//in the order of response
+                        int32_t *sources;
                 } lookup;
 
                 struct {
@@ -737,6 +744,9 @@ afr_build_parent_loc (loc_t *parent, loc_t *child);
 int
 afr_up_children_count (int child_count, unsigned char *child_up);
 
+gf_boolean_t
+afr_is_fresh_lookup (loc_t *loc, xlator_t *this);
+
 void
 afr_update_loc_gfids (loc_t *loc, struct iatt *buf, struct iatt *postparent);