1 files changed, 29 insertions, 42 deletions

diff --git a/xlators/cluster/afr/src/afr-read-txn.c b/xlators/cluster/afr/src/afr-read-txn.c
index 1cd5c2eee3b..6fc2c75145c 100644
--- a/xlators/cluster/afr/src/afr-read-txn.c
+++ b/xlators/cluster/afr/src/afr-read-txn.c
@@ -30,27 +30,6 @@ afr_pending_read_decrement(afr_private_t *priv, int child_index)
     GF_ATOMIC_DEC(priv->pending_reads[child_index]);
 }
 
-static gf_boolean_t
-afr_ta_dict_contains_pending_xattr(dict_t *dict, afr_private_t *priv, int child)
-{
-    int *pending = NULL;
-    int ret = 0;
-    int i = 0;
-
-    ret = dict_get_ptr(dict, priv->pending_key[child], (void *)&pending);
-    if (ret == 0) {
-        for (i = 0; i < AFR_NUM_CHANGE_LOGS; i++) {
-            /* Not doing a ntoh32(pending) as we just want to check
-             * if it is non-zero or not. */
-            if (pending[i]) {
-                return _gf_true;
-            }
-        }
-    }
-
-    return _gf_false;
-}
-
 void
 afr_read_txn_wind(call_frame_t *frame, xlator_t *this, int subvol)
 {
@@ -114,7 +93,7 @@ afr_ta_read_txn(void *opaque)
     call_frame_t *frame = NULL;
     xlator_t *this = NULL;
     int read_subvol = -1;
-    int up_child = AFR_CHILD_UNKNOWN;
+    int query_child = AFR_CHILD_UNKNOWN;
     int possible_bad_child = AFR_CHILD_UNKNOWN;
     int ret = 0;
     int op_errno = ENOMEM;
@@ -134,18 +113,18 @@ afr_ta_read_txn(void *opaque)
     this = frame->this;
     local = frame->local;
     priv = this->private;
+    query_child = local->read_txn_query_child;
 
-    if (local->child_up[AFR_CHILD_ZERO]) {
-        up_child = AFR_CHILD_ZERO;
+    if (query_child == AFR_CHILD_ZERO) {
         possible_bad_child = AFR_CHILD_ONE;
-    } else if (local->child_up[AFR_CHILD_ONE]) {
-        up_child = AFR_CHILD_ONE;
+    } else if (query_child == AFR_CHILD_ONE) {
         possible_bad_child = AFR_CHILD_ZERO;
+    } else {
+        /*read_txn_query_child is AFR_CHILD_UNKNOWN*/
+        goto out;
     }
 
-    GF_ASSERT(up_child != AFR_CHILD_UNKNOWN);
-
-    /* Query the up_child to see if it blames the down one. */
+    /* Ask the query_child to see if it blames the possibly bad one. */
     xdata_req = dict_new();
     if (!xdata_req)
         goto out;
@@ -159,30 +138,33 @@ afr_ta_read_txn(void *opaque)
         goto out;
 
     if (local->fd) {
-        ret = syncop_fxattrop(priv->children[up_child], local->fd,
+        ret = syncop_fxattrop(priv->children[query_child], local->fd,
                               GF_XATTROP_ADD_ARRAY, xdata_req, NULL, &xdata_rsp,
                               NULL);
     } else {
-        ret = syncop_xattrop(priv->children[up_child], &local->loc,
+        ret = syncop_xattrop(priv->children[query_child], &local->loc,
                              GF_XATTROP_ADD_ARRAY, xdata_req, NULL, &xdata_rsp,
                              NULL);
     }
     if (ret || !xdata_rsp) {
         gf_msg(this->name, GF_LOG_ERROR, -ret, AFR_MSG_THIN_ARB,
                "Failed xattrop for gfid %s on %s",
-               uuid_utoa(local->inode->gfid), priv->children[up_child]->name);
+               uuid_utoa(local->inode->gfid),
+               priv->children[query_child]->name);
         op_errno = -ret;
         goto out;
     }
 
     if (afr_ta_dict_contains_pending_xattr(xdata_rsp, priv,
                                            possible_bad_child)) {
-        read_subvol = up_child;
+        read_subvol = query_child;
         goto out;
     }
     dict_unref(xdata_rsp);
-    /* Query thin-arbiter to see if it blames any data brick. */
-    ret = afr_fill_ta_loc(this, &loc);
+    xdata_rsp = NULL;
+
+    /* It doesn't. So query thin-arbiter to see if it blames any data brick. */
+    ret = afr_fill_ta_loc(this, &loc, _gf_true);
     if (ret) {
         gf_msg(this->name, GF_LOG_ERROR, -ret, AFR_MSG_THIN_ARB,
                "Failed to populate thin-arbiter loc for: %s.", loc.name);
@@ -211,8 +193,8 @@ afr_ta_read_txn(void *opaque)
         goto unlock;
     }
 
-    if (!afr_ta_dict_contains_pending_xattr(xdata_rsp, priv, up_child)) {
-        read_subvol = up_child;
+    if (!afr_ta_dict_contains_pending_xattr(xdata_rsp, priv, query_child)) {
+        read_subvol = query_child;
     } else {
         gf_msg(this->name, GF_LOG_ERROR, EIO, AFR_MSG_THIN_ARB,
                "Failing read for gfid %s since good brick %s is down",
@@ -290,7 +272,7 @@ afr_read_txn_refresh_done(call_frame_t *frame, xlator_t *this, int err)
     int read_subvol = -1;
     inode_t *inode = NULL;
     int ret = -1;
-    int spb_choice = -1;
+    int spb_subvol = -1;
 
     local = frame->local;
     inode = local->inode;
@@ -321,9 +303,9 @@ afr_read_txn_refresh_done(call_frame_t *frame, xlator_t *this, int err)
     local->read_attempted[read_subvol] = 1;
 readfn:
     if (read_subvol == -1) {
-        ret = afr_inode_split_brain_choice_get(inode, this, &spb_choice);
-        if ((ret == 0) && spb_choice >= 0)
-            read_subvol = spb_choice;
+        ret = afr_split_brain_read_subvol_get(inode, this, frame, &spb_subvol);
+        if ((ret == 0) && spb_subvol >= 0)
+            read_subvol = spb_subvol;
     }
 
     if (read_subvol == -1) {
@@ -431,7 +413,7 @@ afr_read_txn(call_frame_t *frame, xlator_t *this, inode_t *inode,
     local->is_read_txn = _gf_true;
     local->transaction.type = type;
 
-    if (priv->quorum_count && !afr_has_quorum(local->child_up, this)) {
+    if (priv->quorum_count && !afr_has_quorum(local->child_up, this, NULL)) {
         local->op_ret = -1;
         local->op_errno = afr_quorum_errno(priv);
         goto read;
@@ -450,6 +432,11 @@ afr_read_txn(call_frame_t *frame, xlator_t *this, inode_t *inode,
 
     if (priv->thin_arbiter_count &&
         AFR_COUNT(local->child_up, priv->child_count) != priv->child_count) {
+        if (local->child_up[0]) {
+            local->read_txn_query_child = AFR_CHILD_ZERO;
+        } else if (local->child_up[1]) {
+            local->read_txn_query_child = AFR_CHILD_ONE;
+        }
         afr_ta_read_txn_synctask(frame, this);
         return 0;
     }