cluster/afr: Pick gfid from poststat during fresh lookup for read child calculation

Change-Id: I12c1e4f67f4ec4affbe13d7daf871044a8a2a12e
BUG: 1235216
Signed-off-by: Krutika Dhananjay <kdhananj@redhat.com>
Reviewed-on: http://review.gluster.org/11373
Reviewed-by: Pranith Kumar Karampuri <pkarampu@redhat.com>
Tested-by: NetBSD Build System <jenkins@build.gluster.org>

8 files changed, 93 insertions, 41 deletions

diff --git a/libglusterfs/src/inode.c b/libglusterfs/src/inode.c
index f0bb5238d56..f3dd48653df 100644
--- a/libglusterfs/src/inode.c
+++ b/libglusterfs/src/inode.c
@@ -2169,6 +2169,28 @@ inode_ctx_reset0 (inode_t *inode, xlator_t *xlator, uint64_t *value1_p)
         return ret;
 }
 
+int
+inode_is_linked (inode_t *inode)
+{
+        int            ret   = 0;
+        inode_table_t *table = NULL;
+
+        if (!inode) {
+                gf_msg_callingfn (THIS->name, GF_LOG_WARNING, 0,
+                                  LG_MSG_INODE_NOT_FOUND, "inode not found");
+                return 0;
+        }
+
+        table = inode->table;
+
+        pthread_mutex_lock (&table->lock);
+        {
+                ret = __is_inode_hashed (inode);
+        }
+        pthread_mutex_unlock (&table->lock);
+
+        return ret;
+}
 
 void
 inode_dump (inode_t *inode, char *prefix)
diff --git a/libglusterfs/src/inode.h b/libglusterfs/src/inode.h
index 83646ebda5a..3a7aa063920 100644
--- a/libglusterfs/src/inode.h
+++ b/libglusterfs/src/inode.h
@@ -264,4 +264,7 @@ inode_table_set_lru_limit (inode_table_t *table, uint32_t lru_limit);
 void
 inode_ctx_merge (fd_t *fd, inode_t *inode, inode_t *linked_inode);
 
+int
+inode_is_linked (inode_t *inode);
+
 #endif /* _INODE_H */
diff --git a/xlators/cluster/afr/src/afr-common.c b/xlators/cluster/afr/src/afr-common.c
index 2bbf0e3570e..f644c9dc200 100644
--- a/xlators/cluster/afr/src/afr-common.c
+++ b/xlators/cluster/afr/src/afr-common.c
@@ -982,7 +982,7 @@ out:
 
 
 int
-afr_hash_child (inode_t *inode, int32_t child_count, int hashmode)
+afr_hash_child (afr_read_subvol_args_t *args, int32_t child_count, int hashmode)
 {
         uuid_t gfid_copy = {0,};
         pid_t pid;
@@ -991,11 +991,9 @@ afr_hash_child (inode_t *inode, int32_t child_count, int hashmode)
                 return -1;
         }
 
-        if (inode) {
-               gf_uuid_copy (gfid_copy, inode->gfid);
-        }
+        gf_uuid_copy (gfid_copy, args->gfid);
 
-        if (hashmode > 1 && inode->ia_type != IA_IFDIR) {
+        if ((hashmode > 1) && (args->ia_type != IA_IFDIR)) {
                 /*
                  * Why getpid?  Because it's one of the cheapest calls
                  * available - faster than gethostname etc. - and returns a
@@ -1016,32 +1014,41 @@ afr_hash_child (inode_t *inode, int32_t child_count, int hashmode)
 
 int
 afr_read_subvol_select_by_policy (inode_t *inode, xlator_t *this,
-				  unsigned char *readable)
+				  unsigned char *readable,
+                                  afr_read_subvol_args_t *args)
 {
-	afr_private_t *priv = NULL;
-	int read_subvol = -1;
-	int i = 0;
+	int             i           = 0;
+	int             read_subvol = -1;
+	afr_private_t  *priv        = NULL;
+        afr_read_subvol_args_t local_args = {0,};
 
 	priv = this->private;
 
 	/* first preference - explicitly specified or local subvolume */
 	if (priv->read_child >= 0 && readable[priv->read_child])
-		return priv->read_child;
+                return priv->read_child;
+
+        if (inode_is_linked (inode)) {
+                gf_uuid_copy (local_args.gfid, inode->gfid);
+                local_args.ia_type = inode->ia_type;
+        } else if (args) {
+                local_args = *args;
+        }
 
 	/* second preference - use hashed mode */
-	read_subvol = afr_hash_child (inode, priv->child_count,
-				      priv->hash_mode);
+	read_subvol = afr_hash_child (&local_args, priv->child_count,
+                                      priv->hash_mode);
 	if (read_subvol >= 0 && readable[read_subvol])
-		return read_subvol;
+                return read_subvol;
 
 	for (i = 0; i < priv->child_count; i++) {
-		if (readable[i])
-			return i;
+                if (readable[i])
+                return i;
 	}
 
-	/* no readable subvolumes, either split brain or all subvols down */
+        /* no readable subvolumes, either split brain or all subvols down */
 
-	return -1;
+        return -1;
 }
 
 
@@ -1064,7 +1071,8 @@ afr_inode_read_subvol_type_get (inode_t *inode, xlator_t *this,
 
 int
 afr_read_subvol_get (inode_t *inode, xlator_t *this, int *subvol_p,
-		     int *event_p, afr_transaction_type type)
+		     int *event_p, afr_transaction_type type,
+                     afr_read_subvol_args_t *args)
 {
 	afr_private_t *priv = NULL;
 	unsigned char *data_readable = NULL;
@@ -1091,10 +1099,10 @@ afr_read_subvol_get (inode_t *inode, xlator_t *this, int *subvol_p,
 
 	if (AFR_COUNT (intersection, priv->child_count) > 0)
 		subvol = afr_read_subvol_select_by_policy (inode, this,
-							   intersection);
+							   intersection, args);
 	else
 		subvol = afr_read_subvol_select_by_policy (inode, this,
-							   readable);
+							   readable, args);
 	if (subvol_p)
 		*subvol_p = subvol;
 	if (event_p)
@@ -1408,7 +1416,8 @@ afr_get_parent_read_subvol (xlator_t *this, inode_t *parent,
         priv = this->private;
 
         if (parent)
-                par_read_subvol = afr_data_subvol_get (parent, this, 0, 0);
+                par_read_subvol = afr_data_subvol_get (parent, this, 0, 0,
+                                                       NULL);
 
         for (i = 0; i < priv->child_count; i++) {
                 if (!replies[i].valid)
@@ -1457,6 +1466,8 @@ afr_lookup_done (call_frame_t *frame, xlator_t *this)
 	gf_boolean_t        can_interpret = _gf_true;
         inode_t            *parent = NULL;
         int                 spb_choice = -1;
+        ia_type_t           ia_type = IA_INVAL;
+        afr_read_subvol_args_t args = {0,};
 
         priv  = this->private;
         local = frame->local;
@@ -1504,6 +1515,7 @@ afr_lookup_done (call_frame_t *frame, xlator_t *this)
 		if (read_subvol == -1 || !readable[read_subvol]) {
 			read_subvol = i;
 			gf_uuid_copy (read_gfid, replies[i].poststat.ia_gfid);
+                        ia_type = replies[i].poststat.ia_type;
 			local->op_ret = 0;
 		}
 	}
@@ -1549,14 +1561,16 @@ afr_lookup_done (call_frame_t *frame, xlator_t *this)
 		   a response from all the UP subvolumes and all of them resolved
 		   to the same GFID
 		*/
+                gf_uuid_copy (args.gfid, read_gfid);
+                args.ia_type = ia_type;
 		if (afr_replies_interpret (frame, this, local->inode)) {
 			read_subvol = afr_data_subvol_get (local->inode, this,
-							   0, 0);
+							   0, 0, &args);
 			afr_inode_read_subvol_reset (local->inode, this);
 			goto cant_interpret;
 		} else {
 			read_subvol = afr_data_subvol_get (local->inode, this,
-							   0, 0);
+							   0, 0, &args);
 		}
 	} else {
 	cant_interpret:
@@ -1974,7 +1988,7 @@ afr_discover_done (call_frame_t *frame, xlator_t *this)
 
 	afr_replies_interpret (frame, this, local->inode);
 
-	read_subvol = afr_data_subvol_get (local->inode, this, 0, 0);
+	read_subvol = afr_data_subvol_get (local->inode, this, 0, 0, NULL);
 	if (read_subvol == -1) {
 		gf_log (this->name, GF_LOG_WARNING, "no read subvols for %s",
 			local->loc.path);
@@ -2137,7 +2151,7 @@ afr_discover (call_frame_t *frame, xlator_t *this, loc_t *loc, dict_t *xattr_req
 	}
 
 	afr_read_subvol_get (loc->inode, this, NULL, &event,
-			     AFR_DATA_TRANSACTION);
+			     AFR_DATA_TRANSACTION, NULL);
 
 	if (event != local->event_generation)
 		afr_inode_refresh (frame, this, loc->inode, afr_discover_do);
@@ -2283,7 +2297,7 @@ afr_lookup (call_frame_t *frame, xlator_t *this, loc_t *loc, dict_t *xattr_req)
         }
 
 	afr_read_subvol_get (loc->parent, this, NULL, &event,
-			     AFR_DATA_TRANSACTION);
+			     AFR_DATA_TRANSACTION, NULL);
 
 	if (event != local->event_generation)
 		afr_inode_refresh (frame, this, loc->parent, afr_lookup_do);
@@ -2603,7 +2617,7 @@ afr_fsync_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
 
         local = frame->local;
 
-	read_subvol = afr_data_subvol_get (local->inode, this, 0, 0);
+	read_subvol = afr_data_subvol_get (local->inode, this, 0, 0, NULL);
 
         LOCK (&frame->lock);
         {
diff --git a/xlators/cluster/afr/src/afr-dir-read.c b/xlators/cluster/afr/src/afr-dir-read.c
index 1237b261b14..fae5661d021 100644
--- a/xlators/cluster/afr/src/afr-dir-read.c
+++ b/xlators/cluster/afr/src/afr-dir-read.c
@@ -148,7 +148,12 @@ afr_validate_read_subvol (inode_t *inode, xlator_t *this, int par_read_subvol)
         if (!priv->consistent_metadata)
                 return 0;
 
-        entry_read_subvol = afr_data_subvol_get (inode, this, 0, 0);
+        /* For an inode fetched through readdirp which is yet to be linked,
+         * inode ctx would not be initialised (yet). So this function returns
+         * -1 above due to gen being 0, which is why it is OK to pass NULL for
+         *  read_subvol_args here.
+         */
+        entry_read_subvol = afr_data_subvol_get (inode, this, 0, 0, NULL);
         if (entry_read_subvol != par_read_subvol)
                 return -1;
 
diff --git a/xlators/cluster/afr/src/afr-dir-write.c b/xlators/cluster/afr/src/afr-dir-write.c
index 05b4cc02923..8f525bee7e1 100644
--- a/xlators/cluster/afr/src/afr-dir-write.c
+++ b/xlators/cluster/afr/src/afr-dir-write.c
@@ -90,14 +90,14 @@ __afr_dir_write_finalize (call_frame_t *frame, xlator_t *this)
 	if (local->inode) {
 		afr_replies_interpret (frame, this, local->inode);
 		inode_read_subvol = afr_data_subvol_get (local->inode, this,
-							 NULL, NULL);
+							 NULL, NULL, NULL);
 	}
 	if (local->parent)
 		parent_read_subvol = afr_data_subvol_get (local->parent, this,
-							  NULL, NULL);
+							  NULL, NULL, NULL);
 	if (local->parent2)
 		parent2_read_subvol = afr_data_subvol_get (local->parent2, this,
-							   NULL, NULL);
+							   NULL, NULL, NULL);
 
 	local->op_ret = -1;
 	local->op_errno = afr_final_errno (local, priv);
diff --git a/xlators/cluster/afr/src/afr-inode-write.c b/xlators/cluster/afr/src/afr-inode-write.c
index 931785c2c5b..866294ea20e 100644
--- a/xlators/cluster/afr/src/afr-inode-write.c
+++ b/xlators/cluster/afr/src/afr-inode-write.c
@@ -48,10 +48,11 @@ __afr_inode_write_finalize (call_frame_t *frame, xlator_t *this)
 	if (local->inode) {
 		if (local->transaction.type == AFR_METADATA_TRANSACTION)
 			read_subvol = afr_metadata_subvol_get (local->inode, this,
-							       NULL, NULL);
+							       NULL, NULL,
+                                                               NULL);
 		else
 			read_subvol = afr_data_subvol_get (local->inode, this,
-							   NULL, NULL);
+							   NULL, NULL, NULL);
 	}
 
 	local->op_ret = -1;
diff --git a/xlators/cluster/afr/src/afr-read-txn.c b/xlators/cluster/afr/src/afr-read-txn.c
index 0ec1d912675..6121108872f 100644
--- a/xlators/cluster/afr/src/afr-read-txn.c
+++ b/xlators/cluster/afr/src/afr-read-txn.c
@@ -90,7 +90,7 @@ afr_read_txn_refresh_done (call_frame_t *frame, xlator_t *this, int err)
                 memcpy (local->readable, local->child_up, priv->child_count);
 
 	read_subvol = afr_read_subvol_select_by_policy (inode, this,
-							local->readable);
+							local->readable, NULL);
 	if (read_subvol == -1)
                 AFR_READ_TXN_SET_ERROR_AND_GOTO (-1, EIO, -1, readfn);
 
@@ -232,7 +232,7 @@ afr_read_txn (call_frame_t *frame, xlator_t *this, inode_t *inode,
 		goto refresh;
 
 	read_subvol = afr_read_subvol_select_by_policy (inode, this,
-							local->readable);
+							local->readable, NULL);
 
 	if (read_subvol < 0 || read_subvol > priv->child_count) {
 		gf_msg (this->name, GF_LOG_WARNING, 0, AFR_MSG_SPLIT_BRAIN,
diff --git a/xlators/cluster/afr/src/afr.h b/xlators/cluster/afr/src/afr.h
index 3d28ac22ab1..0bb184c78ae 100644
--- a/xlators/cluster/afr/src/afr.h
+++ b/xlators/cluster/afr/src/afr.h
@@ -750,6 +750,11 @@ typedef struct afr_spbc_timeout {
         int          spb_child_index;
 } afr_spbc_timeout_t;
 
+typedef struct afr_read_subvol_args {
+        ia_type_t ia_type;
+        uuid_t gfid;
+} afr_read_subvol_args_t;
+
 /* did a call fail due to a child failing? */
 #define child_went_down(op_ret, op_errno) (((op_ret) < 0) &&            \
                                            ((op_errno == ENOTCONN) ||   \
@@ -782,7 +787,8 @@ afr_inode_read_subvol_reset (inode_t *inode, xlator_t *this);
 
 int
 afr_read_subvol_select_by_policy (inode_t *inode, xlator_t *this,
-				  unsigned char *readable);
+				  unsigned char *readable,
+                                  afr_read_subvol_args_t *args);
 
 int
 afr_inode_read_subvol_type_get (inode_t *inode, xlator_t *this,
@@ -790,13 +796,14 @@ afr_inode_read_subvol_type_get (inode_t *inode, xlator_t *this,
 				int type);
 int
 afr_read_subvol_get (inode_t *inode, xlator_t *this, int *subvol_p,
-		     int *event_p, afr_transaction_type type);
+		     int *event_p, afr_transaction_type type,
+                     afr_read_subvol_args_t *args);
 
-#define afr_data_subvol_get(i, t, s, e) \
-	afr_read_subvol_get(i, t, s, e, AFR_DATA_TRANSACTION)
+#define afr_data_subvol_get(i, t, s, e, a) \
+	afr_read_subvol_get(i, t, s, e, AFR_DATA_TRANSACTION, a)
 
-#define afr_metadata_subvol_get(i, t, s, e) \
-	afr_read_subvol_get(i, t, s, e, AFR_METADATA_TRANSACTION)
+#define afr_metadata_subvol_get(i, t, s, e, a) \
+	afr_read_subvol_get(i, t, s, e, AFR_METADATA_TRANSACTION, a)
 
 int
 afr_inode_refresh (call_frame_t *frame, xlator_t *this, inode_t *inode,