cluster/afr: refactor

- Remove client side self-healing completely (opendir, openfd, lookup)
- Re-work readdir-failover to work reliably in case of NFS
- Remove unused/dead lock recovery code
- Consistently use xdata in both calls and callbacks in all FOPs
- Per-inode event generation, used to force inode ctx refresh
- Implement dirty flag support (in place of pending counts)
- Eliminate inode ctx structure, use read subvol bits + event_generation
- Implement inode ctx refreshing based on event generation
- Provide backward compatibility in transactions
- remove unused variables and functions
- make code more consistent in style and pattern
- regularize and clean up inode-write transaction code
- regularize and clean up dir-write transaction code
- regularize and clean up common FOPs
- reorganize transaction framework code
- skip setting xattrs in pending dict if nothing is pending
- re-write self-healing code using syncops
- re-write simpler self-heal-daemon

Change-Id: I1e4080c9796c8a2815c2dab4be3073f389d614a8
BUG: 1021686
Signed-off-by: Anand Avati <avati@redhat.com>
Reviewed-on: http://review.gluster.org/6010
Tested-by: Gluster Build System <jenkins@build.gluster.com>
Reviewed-by: Vijay Bellur <vbellur@redhat.com>

1 files changed, 751 insertions, 1858 deletions

diff --git a/xlators/cluster/afr/src/afr-inode-write.c b/xlators/cluster/afr/src/afr-inode-write.c
index d62847def..3dacfc8dd 100644
--- a/xlators/cluster/afr/src/afr-inode-write.c
+++ b/xlators/cluster/afr/src/afr-inode-write.c
@@ -37,46 +37,128 @@
 
 #include "afr.h"
 #include "afr-transaction.h"
-#include "afr-self-heal-common.h"
+//#include "afr-self-heal-common.h"
 
-void
-__inode_write_fop_cbk (call_frame_t *frame, int child_index, int read_child,
-                       xlator_t *this, int32_t *op_ret, int32_t *op_errno,
-                       struct iatt *prebuf, struct iatt *postbuf, dict_t *xdata)
+
+static void
+__afr_inode_write_finalize (call_frame_t *frame, xlator_t *this)
+{
+	afr_local_t *local = NULL;
+	afr_private_t *priv = NULL;
+	int read_subvol = 0;
+	int i = 0;
+
+	local = frame->local;
+	priv = this->private;
+
+	if (local->inode) {
+		if (local->transaction.type == AFR_METADATA_TRANSACTION)
+			read_subvol = afr_metadata_subvol_get (local->inode, this,
+							       NULL, NULL);
+		else
+			read_subvol = afr_data_subvol_get (local->inode, this,
+							   NULL, NULL);
+	}
+
+	local->op_ret = -1;
+	local->op_errno = afr_final_errno (local, priv);
+
+	for (i = 0; i < priv->child_count; i++) {
+		if (!local->replies[i].valid)
+			continue;
+		if (local->replies[i].op_ret < 0) {
+			afr_inode_read_subvol_reset (local->inode, this);
+			continue;
+		}
+
+		/* Order of checks in the compound conditional
+		   below is important.
+
+		   - Highest precedence: largest op_ret
+		   - Next precendence: if all op_rets are equal, read subvol
+		   - Least precedence: any succeeded subvol
+		*/
+		if ((local->op_ret < local->replies[i].op_ret) ||
+		    ((local->op_ret == local->replies[i].op_ret) &&
+		     (i == read_subvol))) {
+
+			local->op_ret = local->replies[i].op_ret;
+			local->op_errno = local->replies[i].op_errno;
+
+			local->cont.inode_wfop.prebuf =
+				local->replies[i].prestat;
+			local->cont.inode_wfop.postbuf =
+				local->replies[i].poststat;
+
+			if (local->replies[i].xdata) {
+				if (local->xdata_rsp)
+					dict_unref (local->xdata_rsp);
+				local->xdata_rsp =
+					dict_ref (local->replies[i].xdata);
+			}
+		}
+	}
+}
+
+
+static void
+__afr_inode_write_fill (call_frame_t *frame, xlator_t *this, int child_index,
+			int op_ret, int op_errno,
+			struct iatt *prebuf, struct iatt *postbuf, dict_t *xdata)
 {
-        afr_local_t     *local = NULL;
+        afr_local_t *local = NULL;
 
         local = frame->local;
 
-        if (afr_fop_failed (*op_ret, *op_errno)) {
-                local->child_errno[child_index] = *op_errno;
-
-                switch (local->op) {
-                case GF_FOP_TRUNCATE:
-                case GF_FOP_FTRUNCATE:
-                        if (*op_errno != EFBIG)
-                                afr_transaction_fop_failed (frame, this,
-                                                            child_index);
-                break;
-                default:
-                        afr_transaction_fop_failed (frame, this, child_index);
-                break;
-                }
-                local->op_errno = *op_errno;
-                goto out;
+	local->replies[child_index].valid = 1;
+	local->replies[child_index].op_ret = op_ret;
+	local->replies[child_index].op_errno = op_errno;
+
+	if (op_ret >= 0) {
+		if (prebuf)
+			local->replies[child_index].prestat = *prebuf;
+		if (postbuf)
+			local->replies[child_index].poststat = *postbuf;
+		if (xdata)
+			local->replies[child_index].xdata = dict_ref (xdata);
+	} else {
+		afr_transaction_fop_failed (frame, this, child_index);
+	}
+
+        return;
+}
+
+
+static int
+__afr_inode_write_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
+                       int32_t op_ret, int32_t op_errno, struct iatt *prebuf,
+                       struct iatt *postbuf, dict_t *xdata)
+{
+        afr_local_t *local = NULL;
+        int child_index = (long) cookie;
+        int call_count = -1;
+
+        local = frame->local;
+
+        LOCK (&frame->lock);
+        {
+                __afr_inode_write_fill (frame, this, child_index, op_ret,
+					op_errno, prebuf, postbuf, xdata);
         }
+        UNLOCK (&frame->lock);
 
-        if ((local->success_count == 0) || (read_child == child_index)) {
-                local->op_ret              = *op_ret;
-                if (prebuf)
-                        local->cont.inode_wfop.prebuf  = *prebuf;
-                if (postbuf)
-                        local->cont.inode_wfop.postbuf = *postbuf;
+        call_count = afr_frame_return (frame);
+
+        if (call_count == 0) {
+		__afr_inode_write_finalize (frame, this);
+
+		if (afr_txn_nothing_failed (frame, this))
+			local->transaction.unwind (frame, this);
+
+                local->transaction.resume (frame, this);
         }
 
-        local->success_count++;
-out:
-        return;
+        return 0;
 }
 
 /* {{{ writev */
@@ -94,6 +176,8 @@ afr_writev_copy_outvars (call_frame_t *src_frame, call_frame_t *dst_frame)
         dst_local->op_errno = src_local->op_errno;
         dst_local->cont.inode_wfop.prebuf = src_local->cont.inode_wfop.prebuf;
         dst_local->cont.inode_wfop.postbuf = src_local->cont.inode_wfop.postbuf;
+	if (src_local->xdata_rsp)
+		dst_local->xdata_rsp = dict_ref (src_local->xdata_rsp);
 }
 
 void
@@ -106,26 +190,9 @@ afr_writev_unwind (call_frame_t *frame, xlator_t *this)
                           local->op_ret, local->op_errno,
                           &local->cont.inode_wfop.prebuf,
                           &local->cont.inode_wfop.postbuf,
-                          NULL);
+                          local->xdata_rsp);
 }
 
-call_frame_t*
-afr_transaction_detach_fop_frame (call_frame_t *frame)
-{
-        afr_local_t *   local = NULL;
-        call_frame_t   *fop_frame = NULL;
-
-        local = frame->local;
-
-        LOCK (&frame->lock);
-        {
-                fop_frame = local->transaction.main_frame;
-                local->transaction.main_frame = NULL;
-        }
-        UNLOCK (&frame->lock);
-
-        return fop_frame;
-}
 
 int
 afr_transaction_writev_unwind (call_frame_t *frame, xlator_t *this)
@@ -173,82 +240,60 @@ afr_writev_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                      struct iatt *postbuf, dict_t *xdata)
 {
         afr_local_t *   local = NULL;
-        afr_private_t  *priv  = NULL;
         call_frame_t    *fop_frame = NULL;
         int child_index = (long) cookie;
         int call_count  = -1;
-        int read_child  = 0;
-        int      ret = 0;
+        int ret = 0;
         uint32_t open_fd_count = 0;
         uint32_t write_is_append = 0;
 
         local = frame->local;
-        priv  = this->private;
-
-        read_child = afr_inode_get_read_ctx (this, local->fd->inode, NULL);
 
         LOCK (&frame->lock);
         {
-                if (child_index == read_child) {
-                        local->read_child_returned = _gf_true;
-                }
-
-                __inode_write_fop_cbk (frame, child_index, read_child, this,
-                                       &op_ret, &op_errno, prebuf, postbuf,
-                                       xdata);
-
-		local->replies[child_index].valid = 1;
-		local->replies[child_index].op_ret = op_ret;
-		local->replies[child_index].op_errno = op_errno;
-
-
-		/* stage the best case return value for unwind */
-                if ((local->success_count == 0) || (op_ret > local->op_ret)) {
-                        local->op_ret              = op_ret;
-			local->op_errno		   = op_errno;
-		}
-
-		if (op_ret != -1) {
-                        if (xdata) {
-                                ret = dict_get_uint32 (xdata,
-                                                       GLUSTERFS_OPEN_FD_COUNT,
-                                                       &open_fd_count);
-                                if ((ret == 0) &&
-                                    (open_fd_count > local->open_fd_count)) {
-                                        local->open_fd_count = open_fd_count;
-                                        local->update_open_fd_count = _gf_true;
-                                }
-
-				write_is_append = 0;
-                                ret = dict_get_uint32 (xdata,
-                                                       GLUSTERFS_WRITE_IS_APPEND,
-                                                       &write_is_append);
-                                if (ret || !write_is_append)
-					local->append_write = _gf_false;
-                        }
-
+                __afr_inode_write_fill (frame, this, child_index, op_ret,
+					op_errno, prebuf, postbuf, xdata);
+		if (op_ret == -1 || !xdata)
+			goto unlock;
+
+		write_is_append = 0;
+		ret = dict_get_uint32 (xdata, GLUSTERFS_WRITE_IS_APPEND,
+				       &write_is_append);
+		if (ret || !write_is_append)
+			local->append_write = _gf_false;
+
+		ret = dict_get_uint32 (xdata, GLUSTERFS_OPEN_FD_COUNT,
+				       &open_fd_count);
+		if (ret == -1)
+			goto unlock;
+		if ((open_fd_count > local->open_fd_count)) {
+			local->open_fd_count = open_fd_count;
+			local->update_open_fd_count = _gf_true;
 		}
         }
+unlock:
         UNLOCK (&frame->lock);
 
         call_count = afr_frame_return (frame);
 
         if (call_count == 0) {
-
-                if (local->update_open_fd_count)
-                        afr_handle_open_fd_count (frame, this);
-
-                if (!local->stable_write && !local->append_write)
+		if (!local->stable_write && !local->append_write)
 			/* An appended write removes the necessity to
 			   fsync() the file. This is because self-heal
 			   has the logic to check for larger file when
 			   the xattrs are not reliably pointing at
 			   a stale file.
 			*/
-                        afr_fd_report_unstable_write (this, local->fd);
+			afr_fd_report_unstable_write (this, local->fd);
+
+		__afr_inode_write_finalize (frame, this);
 
                 afr_writev_handle_short_writes (frame, this);
-                if (afr_any_fops_failed (local, priv)) {
+
+                if (local->update_open_fd_count)
+                        afr_handle_open_fd_count (frame, this);
+
+                if (!afr_txn_nothing_failed (frame, this)) {
                         //Don't unwind until post-op is complete
                         local->transaction.resume (frame, this);
                 } else {
@@ -272,91 +317,23 @@ afr_writev_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
         return 0;
 }
 
+
 int
-afr_writev_wind (call_frame_t *frame, xlator_t *this)
+afr_writev_wind (call_frame_t *frame, xlator_t *this, int subvol)
 {
         afr_local_t *local = NULL;
         afr_private_t *priv = NULL;
-        int i = 0;
-        int call_count = -1;
-        dict_t *xdata = NULL;
-        GF_UNUSED int     ret = 0;
 
         local = frame->local;
         priv = this->private;
 
-        call_count = afr_pre_op_done_children_count (local->transaction.pre_op,
-                                                     priv->child_count);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-                return 0;
-        }
-
-        local->call_count = call_count;
-	local->replies = GF_CALLOC(priv->child_count, sizeof(*local->replies),
-				   gf_afr_mt_reply_t);
-	if (!local->replies) {
-		local->op_ret = -1;
-		local->op_errno = ENOMEM;
-		local->transaction.unwind(frame, this);
-		local->transaction.resume(frame, this);
-		return 0;
-	}
-
-        xdata = dict_new ();
-        if (xdata) {
-                ret = dict_set_uint32 (xdata, GLUSTERFS_OPEN_FD_COUNT,
-                                       sizeof (uint32_t));
-		ret = dict_set_uint32 (xdata, GLUSTERFS_WRITE_IS_APPEND,
-				       0);
-		/* Set append_write to be true speculatively. If on any
-		   server it turns not be true, we unset it in the
-		   callback.
-		*/
-		local->append_write = _gf_true;
-        }
-
-        for (i = 0; i < priv->child_count; i++) {
-                if (local->transaction.pre_op[i]) {
-                        STACK_WIND_COOKIE (frame, afr_writev_wind_cbk,
-                                           (void *) (long) i,
-                                           priv->children[i],
-                                           priv->children[i]->fops->writev,
-                                           local->fd,
-                                           local->cont.writev.vector,
-                                           local->cont.writev.count,
-                                           local->cont.writev.offset,
-                                           local->cont.writev.flags,
-                                           local->cont.writev.iobref,
-                                           xdata);
-
-                        if (!--call_count)
-                                break;
-                }
-        }
-
-        if (xdata)
-                dict_unref (xdata);
-
-        return 0;
-}
-
-
-int
-afr_writev_done (call_frame_t *frame, xlator_t *this)
-{
-        afr_local_t *local = NULL;
-
-        local = frame->local;
-
-        iobref_unref (local->cont.writev.iobref);
-        local->cont.writev.iobref = NULL;
-
-        local->transaction.unwind (frame, this);
-
-        AFR_STACK_DESTROY (frame);
-
+	STACK_WIND_COOKIE (frame, afr_writev_wind_cbk, (void *) (long) subvol,
+			   priv->children[subvol],
+			   priv->children[subvol]->fops->writev,
+			   local->fd, local->cont.writev.vector,
+			   local->cont.writev.count, local->cont.writev.offset,
+			   local->cont.writev.flags, local->cont.writev.iobref,
+			   local->xdata_req);
         return 0;
 }
 
@@ -366,29 +343,29 @@ afr_do_writev (call_frame_t *frame, xlator_t *this)
 {
         call_frame_t    *transaction_frame = NULL;
         afr_local_t     *local             = NULL;
-        int             op_ret   = -1;
-        int             op_errno = 0;
-
-        local = frame->local;
+        int             ret   = -1;
+        int             op_errno = ENOMEM;
 
         transaction_frame = copy_frame (frame);
-        if (!transaction_frame) {
-                op_errno = ENOMEM;
+        if (!transaction_frame)
                 goto out;
-        }
 
+        local = frame->local;
         transaction_frame->local = local;
-        AFR_LOCAL_ALLOC_OR_GOTO (frame->local, out);
+	frame->local = NULL;
 
-        local->op = GF_FOP_WRITE;
+	if (!AFR_FRAME_INIT (frame, op_errno))
+		goto out;
 
-        local->success_count      = 0;
+        local->op = GF_FOP_WRITE;
 
-        local->transaction.fop    = afr_writev_wind;
-        local->transaction.done   = afr_writev_done;
+        local->transaction.wind   = afr_writev_wind;
+        local->transaction.fop    = __afr_txn_write_fop;
+        local->transaction.done   = __afr_txn_write_done;
         local->transaction.unwind = afr_transaction_writev_unwind;
 
         local->transaction.main_frame = frame;
+
         if (local->fd->flags & O_APPEND) {
                /*
                 * Backend vfs ignores the 'offset' for append mode fd so
@@ -405,179 +382,86 @@ afr_do_writev (call_frame_t *frame, xlator_t *this)
                                                          local->cont.writev.count);
         }
 
-        op_ret = afr_transaction (transaction_frame, this, AFR_DATA_TRANSACTION);
-        if (op_ret < 0) {
-            op_errno = -op_ret;
+        ret = afr_transaction (transaction_frame, this, AFR_DATA_TRANSACTION);
+        if (ret < 0) {
+            op_errno = -ret;
             goto out;
         }
 
-        op_ret = 0;
+	return 0;
 out:
-        if (op_ret < 0) {
-                if (transaction_frame)
-                        AFR_STACK_DESTROY (transaction_frame);
-                AFR_STACK_UNWIND (writev, frame, op_ret, op_errno, NULL, NULL, NULL);
-        }
+	if (transaction_frame)
+		AFR_STACK_DESTROY (transaction_frame);
 
+	AFR_STACK_UNWIND (writev, frame, -1, op_errno, NULL, NULL, NULL);
         return 0;
 }
 
-static void
-afr_trigger_open_fd_self_heal (fd_t *fd, xlator_t *this)
-{
-        call_frame_t    *frame   = NULL;
-        afr_local_t     *local   = NULL;
-        afr_self_heal_t *sh      = NULL;
-        char            *reason  = NULL;
-        int32_t         op_errno = 0;
-        int             ret      = 0;
-
-        if (!fd || !fd->inode || uuid_is_null (fd->inode->gfid)) {
-                gf_log_callingfn (this->name, GF_LOG_ERROR, "Invalid args: "
-                                  "fd: %p, inode: %p", fd,
-                                  fd ? fd->inode : NULL);
-                goto out;
-        }
-
-        frame = create_frame (this, this->ctx->pool);
-        if (!frame)
-                goto out;
-
-        AFR_LOCAL_ALLOC_OR_GOTO (frame->local, out);
-        local = frame->local;
-        ret = afr_local_init (local, this->private, &op_errno);
-        if (ret < 0)
-                goto out;
-
-        local->loc.inode = inode_ref (fd->inode);
-        ret = loc_path (&local->loc, NULL);
-        if (ret < 0)
-                goto out;
-
-        sh    = &local->self_heal;
-        sh->do_metadata_self_heal = _gf_true;
-        if (fd->inode->ia_type == IA_IFREG)
-                sh->do_data_self_heal = _gf_true;
-        else if (fd->inode->ia_type == IA_IFDIR)
-                sh->do_entry_self_heal = _gf_true;
-
-        reason = "subvolume came online";
-        afr_launch_self_heal (frame, this, fd->inode, _gf_true,
-                              fd->inode->ia_type, reason, NULL, NULL);
-        return;
-out:
-        AFR_STACK_DESTROY (frame);
-}
-
-void
-afr_open_fd_fix (fd_t *fd, xlator_t *this)
-{
-        int           ret             = 0;
-        int           i               = 0;
-        afr_fd_ctx_t  *fd_ctx         = NULL;
-        gf_boolean_t  need_self_heal  = _gf_false;
-        int           *need_open      = NULL;
-        size_t        need_open_count = 0;
-        afr_private_t *priv           = NULL;
-
-        priv  = this->private;
-
-        if (!afr_is_fd_fixable (fd))
-                goto out;
-
-        fd_ctx = afr_fd_ctx_get (fd, this);
-        if (!fd_ctx)
-                goto out;
-
-        LOCK (&fd->lock);
-        {
-                if (fd_ctx->up_count < priv->up_count) {
-                        need_self_heal = _gf_true;
-                        fd_ctx->up_count   = priv->up_count;
-                        fd_ctx->down_count = priv->down_count;
-                }
-
-                need_open = alloca (priv->child_count * sizeof (*need_open));
-                for (i = 0; i < priv->child_count; i++) {
-                        need_open[i] = 0;
-                        if (fd_ctx->opened_on[i] != AFR_FD_NOT_OPENED)
-                                continue;
-
-                        if (!priv->child_up[i])
-                                continue;
-
-                        fd_ctx->opened_on[i] = AFR_FD_OPENING;
-
-                        need_open[i] = 1;
-                        need_open_count++;
-                }
-        }
-        UNLOCK (&fd->lock);
-        if (ret)
-                goto out;
-
-        if (need_self_heal)
-                afr_trigger_open_fd_self_heal (fd, this);
-
-        if (!need_open_count)
-                goto out;
-
-        afr_fix_open (this, fd, need_open_count, need_open);
-out:
-        return;
-}
 
 int
 afr_writev (call_frame_t *frame, xlator_t *this, fd_t *fd,
             struct iovec *vector, int32_t count, off_t offset,
             uint32_t flags, struct iobref *iobref, dict_t *xdata)
 {
-        afr_private_t * priv  = NULL;
-        afr_local_t   * local = NULL;
-        int ret = -1;
-        int op_errno = 0;
-
-        VALIDATE_OR_GOTO (frame, out);
-        VALIDATE_OR_GOTO (this, out);
-        VALIDATE_OR_GOTO (this->private, out);
+        afr_private_t *priv = NULL;
+        afr_local_t *local = NULL;
+        int op_errno = ENOMEM;
 
         priv = this->private;
 
-        if (afr_is_split_brain (this, fd->inode)) {
-                op_errno = EIO;
-                goto out;
-        }
-
         QUORUM_CHECK(writev,out);
 
-        AFR_LOCAL_ALLOC_OR_GOTO (frame->local, out);
-        local = frame->local;
-
-        ret = afr_local_init (local, priv, &op_errno);
-        if (ret < 0)
-                goto out;
+	local = AFR_FRAME_INIT (frame, op_errno);
+	if (!local)
+		goto out;
 
-        local->cont.writev.vector     = iov_dup (vector, count);
+        local->cont.writev.vector = iov_dup (vector, count);
+	if (!local->cont.writev.vector)
+		goto out;
         local->cont.writev.count      = count;
         local->cont.writev.offset     = offset;
         local->cont.writev.flags      = flags;
         local->cont.writev.iobref     = iobref_ref (iobref);
 
-        local->fd                = fd_ref (fd);
+	if (xdata)
+		local->xdata_req = dict_copy_with_ref (xdata, NULL);
+	else
+		local->xdata_req = dict_new ();
+
+	if (!local->xdata_req)
+		goto out;
+
+        local->fd = fd_ref (fd);
+	local->inode = inode_ref (fd->inode);
+
+	if (dict_set_uint32 (local->xdata_req, GLUSTERFS_OPEN_FD_COUNT, 4)) {
+		op_errno = ENOMEM;
+		goto out;
+	}
+
+	if (dict_set_uint32 (local->xdata_req, GLUSTERFS_WRITE_IS_APPEND, 4)) {
+		op_errno = ENOMEM;
+		goto out;
+	}
+
+	/* Set append_write to be true speculatively. If on any
+	   server it turns not be true, we unset it in the
+	   callback.
+	*/
+	local->append_write = _gf_true;
 
 	/* detect here, but set it in writev_wind_cbk *after* the unstable
 	   write is performed
 	*/
 	local->stable_write = !!((fd->flags|flags)&(O_SYNC|O_DSYNC));
 
-        afr_open_fd_fix (fd, this);
+        afr_fix_open (fd, this);
 
         afr_do_writev (frame, this);
 
-        ret = 0;
+	return 0;
 out:
-        if (ret < 0)
-                AFR_STACK_UNWIND (writev, frame, -1, op_errno, NULL, NULL, NULL);
+	AFR_STACK_UNWIND (writev, frame, -1, op_errno, NULL, NULL, NULL);
 
         return 0;
 }
@@ -595,22 +479,13 @@ afr_truncate_unwind (call_frame_t *frame, xlator_t *this)
 
         local = frame->local;
 
-        LOCK (&frame->lock);
-        {
-                if (local->transaction.main_frame)
-                        main_frame = local->transaction.main_frame;
-                local->transaction.main_frame = NULL;
-        }
-        UNLOCK (&frame->lock);
-
-        if (main_frame) {
-                AFR_STACK_UNWIND (truncate, main_frame, local->op_ret,
-                                  local->op_errno,
-                                  &local->cont.inode_wfop.prebuf,
-                                  &local->cont.inode_wfop.postbuf,
-                                  NULL);
-        }
+	main_frame = afr_transaction_detach_fop_frame (frame);
+	if (!main_frame)
+		return 0;
 
+	AFR_STACK_UNWIND (truncate, main_frame, local->op_ret, local->op_errno,
+			  &local->cont.inode_wfop.prebuf,
+			  &local->cont.inode_wfop.postbuf, local->xdata_rsp);
         return 0;
 }
 
@@ -620,96 +495,32 @@ afr_truncate_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                        int32_t op_ret, int32_t op_errno, struct iatt *prebuf,
                        struct iatt *postbuf, dict_t *xdata)
 {
-        afr_local_t *   local = NULL;
-        int child_index = (long) cookie;
-        int read_child  = 0;
-        int call_count  = -1;
-
-        local = frame->local;
-
-        read_child = afr_inode_get_read_ctx (this, local->loc.inode, NULL);
+	afr_local_t *local = NULL;
 
-        LOCK (&frame->lock);
-        {
-                if (child_index == read_child) {
-                        local->read_child_returned = _gf_true;
-                }
+	local = frame->local;
 
-                if (op_ret != -1) {
-			if (prebuf->ia_size != postbuf->ia_size)
-				local->stable_write = _gf_false;
-                }
-                __inode_write_fop_cbk (frame, child_index, read_child, this,
-                                       &op_ret, &op_errno, prebuf, postbuf,
-                                       xdata);
-        }
-        UNLOCK (&frame->lock);
+	if (op_ret == 0 && prebuf->ia_size != postbuf->ia_size)
+		local->stable_write = _gf_false;
 
-        call_count = afr_frame_return (frame);
-
-        if (call_count == 0) {
-		if (local->stable_write && afr_txn_nothing_failed (frame, this))
-			local->transaction.unwind (frame, this);
-
-                local->transaction.resume (frame, this);
-        }
-
-        return 0;
+	return __afr_inode_write_cbk (frame, cookie, this, op_ret, op_errno,
+				      prebuf, postbuf, xdata);
 }
 
 
-int32_t
-afr_truncate_wind (call_frame_t *frame, xlator_t *this)
+int
+afr_truncate_wind (call_frame_t *frame, xlator_t *this, int subvol)
 {
         afr_local_t *local = NULL;
         afr_private_t *priv = NULL;
-        int call_count = -1;
-        int i = 0;
 
         local = frame->local;
         priv = this->private;
 
-        call_count = afr_pre_op_done_children_count (local->transaction.pre_op,
-                                                     priv->child_count);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-                return 0;
-        }
-
-        local->call_count = call_count;
-	local->stable_write = _gf_true;
-
-        for (i = 0; i < priv->child_count; i++) {
-                if (local->transaction.pre_op[i]) {
-                        STACK_WIND_COOKIE (frame, afr_truncate_wind_cbk,
-                                           (void *) (long) i,
-                                           priv->children[i],
-                                           priv->children[i]->fops->truncate,
-                                           &local->loc,
-                                           local->cont.truncate.offset,
-                                           NULL);
-
-                        if (!--call_count)
-                                break;
-                }
-        }
-
-        return 0;
-}
-
-
-int
-afr_truncate_done (call_frame_t *frame, xlator_t *this)
-{
-        afr_local_t *local = NULL;
-
-        local = frame->local;
-
-        local->transaction.unwind (frame, this);
-
-        AFR_STACK_DESTROY (frame);
-
+	STACK_WIND_COOKIE (frame, afr_truncate_wind_cbk, (void *) (long) subvol,
+			   priv->children[subvol],
+			   priv->children[subvol]->fops->truncate,
+			   &local->loc, local->cont.truncate.offset,
+			   local->xdata_req);
         return 0;
 }
 
@@ -721,56 +532,60 @@ afr_truncate (call_frame_t *frame, xlator_t *this,
         afr_private_t * priv  = NULL;
         afr_local_t   * local = NULL;
         call_frame_t   *transaction_frame = NULL;
-        int ret = -1;
-        int op_errno = 0;
-
-        VALIDATE_OR_GOTO (frame, out);
-        VALIDATE_OR_GOTO (this, out);
-        VALIDATE_OR_GOTO (this->private, out);
+	int ret = -1;
+        int op_errno = ENOMEM;
 
         priv = this->private;
 
         QUORUM_CHECK(truncate,out);
 
         transaction_frame = copy_frame (frame);
-        if (!transaction_frame) {
-                op_errno = ENOMEM;
+        if (!transaction_frame)
                 goto out;
-        }
 
-        AFR_LOCAL_ALLOC_OR_GOTO (transaction_frame->local, out);
-        local = transaction_frame->local;
-
-        ret = afr_local_init (local, priv, &op_errno);
-        if (ret < 0)
-                goto out;
+	local = AFR_FRAME_INIT (transaction_frame, op_errno);
+	if (!local)
+		goto out;
 
         local->cont.truncate.offset  = offset;
+	if (xdata)
+		local->xdata_req = dict_copy_with_ref (xdata, NULL);
+	else
+		local->xdata_req = dict_new ();
+
+	if (!local->xdata_req)
+		goto out;
 
-        local->transaction.fop    = afr_truncate_wind;
-        local->transaction.done   = afr_truncate_done;
+	local->transaction.wind   = afr_truncate_wind;
+        local->transaction.fop    = __afr_txn_write_fop;
+        local->transaction.done   = __afr_txn_write_done;
         local->transaction.unwind = afr_truncate_unwind;
 
         loc_copy (&local->loc, loc);
+	local->inode = inode_ref (loc->inode);
+
+        local->op = GF_FOP_TRUNCATE;
 
         local->transaction.main_frame = frame;
         local->transaction.start   = offset;
         local->transaction.len     = 0;
 
+	/* Set it true speculatively, will get reset in afr_truncate_wind_cbk
+	   if truncate was not a NOP */
+	local->stable_write = _gf_true;
+
         ret = afr_transaction (transaction_frame, this, AFR_DATA_TRANSACTION);
         if (ret < 0) {
-            op_errno = -ret;
-            goto out;
+		op_errno = -ret;
+		goto out;
         }
 
-        ret = 0;
+	return 0;
 out:
-        if (ret < 0) {
-                if (transaction_frame)
-                        AFR_STACK_DESTROY (transaction_frame);
-                AFR_STACK_UNWIND (truncate, frame, -1, op_errno, NULL, NULL, NULL);
-        }
+	if (transaction_frame)
+		AFR_STACK_DESTROY (transaction_frame);
 
+	AFR_STACK_UNWIND (truncate, frame, -1, op_errno, NULL, NULL, NULL);
         return 0;
 }
 
@@ -788,21 +603,13 @@ afr_ftruncate_unwind (call_frame_t *frame, xlator_t *this)
 
         local = frame->local;
 
-        LOCK (&frame->lock);
-        {
-                if (local->transaction.main_frame)
-                        main_frame = local->transaction.main_frame;
-                local->transaction.main_frame = NULL;
-        }
-        UNLOCK (&frame->lock);
+	main_frame = afr_transaction_detach_fop_frame (frame);
+	if (!main_frame)
+		return 0;
 
-        if (main_frame) {
-                AFR_STACK_UNWIND (ftruncate, main_frame, local->op_ret,
-                                  local->op_errno,
-                                  &local->cont.inode_wfop.prebuf,
-                                  &local->cont.inode_wfop.postbuf,
-                                  NULL);
-        }
+	AFR_STACK_UNWIND (ftruncate, main_frame, local->op_ret, local->op_errno,
+			  &local->cont.inode_wfop.prebuf,
+			  &local->cont.inode_wfop.postbuf, local->xdata_rsp);
         return 0;
 }
 
@@ -812,122 +619,75 @@ afr_ftruncate_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                         int32_t op_ret, int32_t op_errno, struct iatt *prebuf,
                         struct iatt *postbuf, dict_t *xdata)
 {
-        afr_local_t *   local = NULL;
-        int child_index = (long) cookie;
-        int call_count  = -1;
-        int read_child  = 0;
-
-        local = frame->local;
-
-        read_child = afr_inode_get_read_ctx (this, local->fd->inode, NULL);
-
-        LOCK (&frame->lock);
-        {
-                if (child_index == read_child) {
-                        local->read_child_returned = _gf_true;
-                }
-
-                if (op_ret != -1) {
-			if (prebuf->ia_size != postbuf->ia_size)
-				local->stable_write = _gf_false;
-                }
-                __inode_write_fop_cbk (frame, child_index, read_child, this,
-                                       &op_ret, &op_errno, prebuf, postbuf,
-                                       xdata);
-        }
-        UNLOCK (&frame->lock);
+	afr_local_t *local = NULL;
 
-        call_count = afr_frame_return (frame);
+	local = frame->local;
 
-        if (call_count == 0) {
-		if (local->stable_write && afr_txn_nothing_failed (frame, this))
-			local->transaction.unwind (frame, this);
+	if (op_ret == 0 && prebuf->ia_size != postbuf->ia_size)
+		local->stable_write = _gf_false;
 
-                local->transaction.resume (frame, this);
-        }
-
-        return 0;
+	return __afr_inode_write_cbk (frame, cookie, this, op_ret, op_errno,
+				      prebuf, postbuf, xdata);
 }
 
 
 int
-afr_ftruncate_wind (call_frame_t *frame, xlator_t *this)
+afr_ftruncate_wind (call_frame_t *frame, xlator_t *this, int subvol)
 {
         afr_local_t *local = NULL;
         afr_private_t *priv = NULL;
-        int call_count = -1;
-        int i = 0;
-
-        local = frame->local;
-        priv = this->private;
-
-        call_count = afr_pre_op_done_children_count (local->transaction.pre_op,
-                                                     priv->child_count);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-                return 0;
-        }
 
-        local->call_count = call_count;
-	local->stable_write = _gf_true;
-
-        for (i = 0; i < priv->child_count; i++) {
-                if (local->transaction.pre_op[i]) {
-                        STACK_WIND_COOKIE (frame, afr_ftruncate_wind_cbk,
-                                           (void *) (long) i,
-                                           priv->children[i],
-                                           priv->children[i]->fops->ftruncate,
-                                           local->fd,
-                                           local->cont.ftruncate.offset,
-                                           NULL);
-
-                        if (!--call_count)
-                                break;
-                }
-        }
+	local = frame->local;
+	priv = this->private;
 
+	STACK_WIND_COOKIE (frame, afr_ftruncate_wind_cbk, (void *) (long) subvol,
+			   priv->children[subvol],
+			   priv->children[subvol]->fops->ftruncate,
+			   local->fd, local->cont.ftruncate.offset,
+			   local->xdata_req);
         return 0;
 }
 
 
 int
-afr_ftruncate_done (call_frame_t *frame, xlator_t *this)
+afr_ftruncate (call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset,
+	       dict_t *xdata)
 {
+        afr_private_t *priv = NULL;
         afr_local_t *local = NULL;
+	call_frame_t *transaction_frame = NULL;
+	int ret = -1;
+        int op_errno = ENOMEM;
 
-        local = frame->local;
-
-        local->transaction.unwind (frame, this);
+        priv = this->private;
 
-        AFR_STACK_DESTROY (frame);
+        QUORUM_CHECK(ftruncate,out);
 
-        return 0;
-}
+	transaction_frame = copy_frame (frame);
+	if (!frame)
+		goto out;
 
+        local = AFR_FRAME_INIT (transaction_frame, op_errno);
+        if (!local)
+		goto out;
 
-int
-afr_do_ftruncate (call_frame_t *frame, xlator_t *this)
-{
-        call_frame_t * transaction_frame = NULL;
-        afr_local_t *  local             = NULL;
-        int op_ret   = -1;
-        int op_errno = 0;
+        local->cont.ftruncate.offset  = offset;
+	if (xdata)
+		local->xdata_req = dict_copy_with_ref (xdata, NULL);
+	else
+		local->xdata_req = dict_new ();
 
-        local = frame->local;
+	if (!local->xdata_req)
+		goto out;
 
-        transaction_frame = copy_frame (frame);
-        if (!transaction_frame) {
-                goto out;
-        }
-
-        transaction_frame->local = local;
-        frame->local = NULL;
+        local->fd = fd_ref (fd);
+	local->inode = inode_ref (fd->inode);
 
         local->op = GF_FOP_FTRUNCATE;
 
-        local->transaction.fop    = afr_ftruncate_wind;
-        local->transaction.done   = afr_ftruncate_done;
+	local->transaction.wind   = afr_ftruncate_wind;
+        local->transaction.fop    = __afr_txn_write_fop;
+        local->transaction.done   = __afr_txn_write_done;
         local->transaction.unwind = afr_ftruncate_unwind;
 
         local->transaction.main_frame = frame;
@@ -935,69 +695,21 @@ afr_do_ftruncate (call_frame_t *frame, xlator_t *this)
         local->transaction.start   = local->cont.ftruncate.offset;
         local->transaction.len     = 0;
 
-        op_ret = afr_transaction (transaction_frame, this, AFR_DATA_TRANSACTION);
-        if (op_ret < 0) {
-            op_errno = -op_ret;
-            goto out;
-        }
+        afr_fix_open (fd, this);
 
-        op_ret = 0;
-out:
-        if (op_ret < 0) {
-                if (transaction_frame)
-                        AFR_STACK_DESTROY (transaction_frame);
-                AFR_STACK_UNWIND (ftruncate, frame, op_ret, op_errno, NULL,
-                                  NULL, NULL);
-        }
-
-        return 0;
-}
-
-
-int
-afr_ftruncate (call_frame_t *frame, xlator_t *this,
-               fd_t *fd, off_t offset, dict_t *xdata)
-{
-        afr_private_t * priv  = NULL;
-        afr_local_t   * local = NULL;
-        call_frame_t   *transaction_frame = NULL;
-        int ret = -1;
-        int op_errno = 0;
-
-        VALIDATE_OR_GOTO (frame, out);
-        VALIDATE_OR_GOTO (this, out);
-        VALIDATE_OR_GOTO (this->private, out);
-
-        priv = this->private;
+	/* Set it true speculatively, will get reset in afr_ftruncate_wind_cbk
+	   if truncate was not a NOP */
+	local->stable_write = _gf_true;
 
-        if (afr_is_split_brain (this, fd->inode)) {
-                op_errno = EIO;
-                goto out;
+        ret = afr_transaction (transaction_frame, this, AFR_DATA_TRANSACTION);
+        if (ret < 0) {
+		op_errno = -ret;
+		goto out;
         }
-        QUORUM_CHECK(ftruncate,out);
-
-        AFR_LOCAL_ALLOC_OR_GOTO (frame->local, out);
-        local = frame->local;
-
-        ret = afr_local_init (local, priv, &op_errno);
-        if (ret < 0)
-                goto out;
 
-        local->cont.ftruncate.offset  = offset;
-
-        local->fd = fd_ref (fd);
-
-        afr_open_fd_fix (fd, this);
-
-        afr_do_ftruncate (frame, this);
-
-        ret = 0;
+	return 0;
 out:
-        if (ret < 0) {
-                if (transaction_frame)
-                        AFR_STACK_DESTROY (transaction_frame);
-                AFR_STACK_UNWIND (ftruncate, frame, -1, op_errno, NULL, NULL, NULL);
-        }
+	AFR_STACK_UNWIND (ftruncate, frame, -1, op_errno, NULL, NULL, NULL);
 
         return 0;
 }
@@ -1009,173 +721,92 @@ out:
 int
 afr_setattr_unwind (call_frame_t *frame, xlator_t *this)
 {
-        afr_local_t *   local = NULL;
-        call_frame_t   *main_frame = NULL;
+        afr_local_t *local = NULL;
+        call_frame_t *main_frame = NULL;
 
         local = frame->local;
 
-        LOCK (&frame->lock);
-        {
-                if (local->transaction.main_frame)
-                        main_frame = local->transaction.main_frame;
-                local->transaction.main_frame = NULL;
-        }
-        UNLOCK (&frame->lock);
-
-        if (main_frame) {
-                AFR_STACK_UNWIND (setattr, main_frame, local->op_ret,
-                                  local->op_errno,
-                                  &local->cont.inode_wfop.prebuf,
-                                  &local->cont.inode_wfop.postbuf,
-                                  NULL);
-        }
+	main_frame = afr_transaction_detach_fop_frame (frame);
+	if (!main_frame)
+		return 0;
 
+	AFR_STACK_UNWIND (setattr, main_frame, local->op_ret, local->op_errno,
+			  &local->cont.inode_wfop.prebuf,
+			  &local->cont.inode_wfop.postbuf,
+			  local->xdata_rsp);
         return 0;
 }
 
 
 int
 afr_setattr_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
-                      int32_t op_ret, int32_t op_errno,
+                      int op_ret, int op_errno,
                       struct iatt *preop, struct iatt *postop, dict_t *xdata)
 {
-        afr_local_t *   local = NULL;
-        afr_private_t * priv  = NULL;
-        int child_index = (long) cookie;
-        int read_child  = 0;
-        int call_count  = -1;
-        int need_unwind = 0;
-
-        local = frame->local;
-        priv  = this->private;
-
-        read_child = afr_inode_get_read_ctx (this, local->loc.inode, NULL);
-
-        LOCK (&frame->lock);
-        {
-                if (child_index == read_child) {
-                        local->read_child_returned = _gf_true;
-                }
-
-                __inode_write_fop_cbk (frame, child_index, read_child, this,
-                                       &op_ret, &op_errno, preop, postop,
-                                       xdata);
-
-                if ((local->success_count >= priv->wait_count)
-                    && local->read_child_returned) {
-                        need_unwind = 1;
-                }
-        }
-        UNLOCK (&frame->lock);
-
-        if (need_unwind)
-                local->transaction.unwind (frame, this);
-
-        call_count = afr_frame_return (frame);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-        }
-
-        return 0;
+	return __afr_inode_write_cbk (frame, cookie, this, op_ret, op_errno,
+				      preop, postop, xdata);
 }
 
 
-int32_t
-afr_setattr_wind (call_frame_t *frame, xlator_t *this)
+int
+afr_setattr_wind (call_frame_t *frame, xlator_t *this, int subvol)
 {
         afr_local_t *local = NULL;
         afr_private_t *priv = NULL;
-        int call_count = -1;
-        int i = 0;
 
         local = frame->local;
         priv = this->private;
 
-        call_count = afr_pre_op_done_children_count (local->transaction.pre_op,
-                                                     priv->child_count);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-                return 0;
-        }
-
-        local->call_count = call_count;
-
-        for (i = 0; i < priv->child_count; i++) {
-                if (local->transaction.pre_op[i]) {
-                        STACK_WIND_COOKIE (frame, afr_setattr_wind_cbk,
-                                           (void *) (long) i,
-                                           priv->children[i],
-                                           priv->children[i]->fops->setattr,
-                                           &local->loc,
-                                           &local->cont.setattr.in_buf,
-                                           local->cont.setattr.valid,
-                                           NULL);
-
-                        if (!--call_count)
-                                break;
-                }
-        }
-
+	STACK_WIND_COOKIE (frame, afr_setattr_wind_cbk, (void *) (long) subvol,
+			   priv->children[subvol],
+			   priv->children[subvol]->fops->setattr,
+			   &local->loc, &local->cont.setattr.in_buf,
+			   local->cont.setattr.valid, local->xdata_req);
         return 0;
 }
 
 
 int
-afr_setattr_done (call_frame_t *frame, xlator_t *this)
+afr_setattr (call_frame_t *frame, xlator_t *this, loc_t *loc, struct iatt *buf,
+	     int32_t valid, dict_t *xdata)
 {
+        afr_private_t *priv = NULL;
         afr_local_t *local = NULL;
-
-        local = frame->local;
-
-        local->transaction.unwind (frame, this);
-
-        AFR_STACK_DESTROY (frame);
-
-        return 0;
-}
-
-
-int
-afr_setattr (call_frame_t *frame, xlator_t *this,
-             loc_t *loc, struct iatt *buf, int32_t valid, dict_t *xdata)
-{
-        afr_private_t * priv  = NULL;
-        afr_local_t   * local = NULL;
-        call_frame_t   *transaction_frame = NULL;
+        call_frame_t *transaction_frame = NULL;
         int ret = -1;
-        int op_errno = 0;
-
-        VALIDATE_OR_GOTO (frame, out);
-        VALIDATE_OR_GOTO (this, out);
-        VALIDATE_OR_GOTO (this->private, out);
+        int op_errno = ENOMEM;
 
         priv = this->private;
 
         QUORUM_CHECK(setattr,out);
 
         transaction_frame = copy_frame (frame);
-        if (!transaction_frame) {
-                op_errno = ENOMEM;
+        if (!transaction_frame)
                 goto out;
-        }
 
-        AFR_LOCAL_ALLOC_OR_GOTO (transaction_frame->local, out);
-        local = transaction_frame->local;
-
-        ret = afr_local_init (local, priv, &op_errno);
-        if (ret < 0)
-                goto out;
+	local = AFR_FRAME_INIT (transaction_frame, op_errno);
+	if (!local)
+		goto out;
 
         local->cont.setattr.in_buf = *buf;
         local->cont.setattr.valid  = valid;
+	if (xdata)
+		local->xdata_req = dict_copy_with_ref (xdata, NULL);
+	else
+		local->xdata_req = dict_new ();
 
-        local->transaction.fop    = afr_setattr_wind;
-        local->transaction.done   = afr_setattr_done;
+	if (!local->xdata_req)
+		goto out;
+
+        local->transaction.wind   = afr_setattr_wind;
+        local->transaction.fop    = __afr_txn_write_fop;
+        local->transaction.done   = __afr_txn_write_done;
         local->transaction.unwind = afr_setattr_unwind;
 
         loc_copy (&local->loc, loc);
+	local->inode = inode_ref (loc->inode);
+
+	local->op = GF_FOP_SETATTR;
 
         local->transaction.main_frame = frame;
         local->transaction.start   = LLONG_MAX - 1;
@@ -1183,18 +814,16 @@ afr_setattr (call_frame_t *frame, xlator_t *this,
 
         ret = afr_transaction (transaction_frame, this, AFR_METADATA_TRANSACTION);
         if (ret < 0) {
-            op_errno = -ret;
-            goto out;
+		op_errno = -ret;
+		goto out;
         }
 
-        ret = 0;
+	return 0;
 out:
-        if (ret < 0) {
-                if (transaction_frame)
-                        AFR_STACK_DESTROY (transaction_frame);
-                AFR_STACK_UNWIND (setattr, frame, -1, op_errno, NULL, NULL, NULL);
-        }
+	if (transaction_frame)
+		AFR_STACK_DESTROY (transaction_frame);
 
+	AFR_STACK_UNWIND (setattr, frame, -1, op_errno, NULL, NULL, NULL);
         return 0;
 }
 
@@ -1208,22 +837,13 @@ afr_fsetattr_unwind (call_frame_t *frame, xlator_t *this)
 
         local = frame->local;
 
-        LOCK (&frame->lock);
-        {
-                if (local->transaction.main_frame)
-                        main_frame = local->transaction.main_frame;
-                local->transaction.main_frame = NULL;
-        }
-        UNLOCK (&frame->lock);
-
-        if (main_frame) {
-                AFR_STACK_UNWIND (fsetattr, main_frame, local->op_ret,
-                                  local->op_errno,
-                                  &local->cont.inode_wfop.prebuf,
-                                  &local->cont.inode_wfop.postbuf,
-                                  NULL);
-        }
+	main_frame = afr_transaction_detach_fop_frame (frame);
+        if (!main_frame)
+		return 0;
 
+	AFR_STACK_UNWIND (fsetattr, main_frame, local->op_ret, local->op_errno,
+			  &local->cont.inode_wfop.prebuf,
+			  &local->cont.inode_wfop.postbuf, local->xdata_rsp);
         return 0;
 }
 
@@ -1233,149 +853,72 @@ afr_fsetattr_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                        int32_t op_ret, int32_t op_errno,
                        struct iatt *preop, struct iatt *postop, dict_t *xdata)
 {
-        afr_local_t *   local = NULL;
-        afr_private_t * priv  = NULL;
-        int child_index = (long) cookie;
-        int read_child  = 0;
-        int call_count  = -1;
-        int need_unwind = 0;
-
-        local = frame->local;
-        priv  = this->private;
-
-        read_child = afr_inode_get_read_ctx (this, local->fd->inode, NULL);
-
-        LOCK (&frame->lock);
-        {
-                if (child_index == read_child) {
-                        local->read_child_returned = _gf_true;
-                }
-
-                __inode_write_fop_cbk (frame, child_index, read_child, this,
-                                       &op_ret, &op_errno, preop, postop,
-                                       xdata);
-
-                if ((local->success_count >= priv->wait_count)
-                    && local->read_child_returned) {
-                        need_unwind = 1;
-                }
-        }
-        UNLOCK (&frame->lock);
-
-        if (need_unwind)
-                local->transaction.unwind (frame, this);
-
-        call_count = afr_frame_return (frame);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-        }
-
-        return 0;
+	return __afr_inode_write_cbk (frame, cookie, this, op_ret, op_errno,
+				      preop, postop, xdata);
 }
 
 
-int32_t
-afr_fsetattr_wind (call_frame_t *frame, xlator_t *this)
+int
+afr_fsetattr_wind (call_frame_t *frame, xlator_t *this, int subvol)
 {
         afr_local_t *local = NULL;
         afr_private_t *priv = NULL;
-        int call_count = -1;
-        int i = 0;
 
         local = frame->local;
         priv = this->private;
 
-        call_count = afr_pre_op_done_children_count (local->transaction.pre_op,
-                                                     priv->child_count);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-                return 0;
-        }
-
-        local->call_count = call_count;
-
-        for (i = 0; i < priv->child_count; i++) {
-                if (local->transaction.pre_op[i]) {
-                        STACK_WIND_COOKIE (frame, afr_fsetattr_wind_cbk,
-                                           (void *) (long) i,
-                                           priv->children[i],
-                                           priv->children[i]->fops->fsetattr,
-                                           local->fd,
-                                           &local->cont.fsetattr.in_buf,
-                                           local->cont.fsetattr.valid,
-                                           NULL);
-
-                        if (!--call_count)
-                                break;
-                }
-        }
-
+	STACK_WIND_COOKIE (frame, afr_fsetattr_wind_cbk, (void *) (long) subvol,
+			   priv->children[subvol],
+			   priv->children[subvol]->fops->fsetattr,
+			   local->fd, &local->cont.fsetattr.in_buf,
+			   local->cont.fsetattr.valid, local->xdata_req);
         return 0;
 }
 
 
 int
-afr_fsetattr_done (call_frame_t *frame, xlator_t *this)
-{
-        afr_local_t *local = NULL;
-
-        local = frame->local;
-
-        local->transaction.unwind (frame, this);
-
-        AFR_STACK_DESTROY (frame);
-
-        return 0;
-}
-
-int
 afr_fsetattr (call_frame_t *frame, xlator_t *this,
               fd_t *fd, struct iatt *buf, int32_t valid, dict_t *xdata)
 {
-        afr_private_t * priv  = NULL;
-        afr_local_t   * local = NULL;
-        call_frame_t   *transaction_frame = NULL;
+        afr_private_t *priv = NULL;
+        afr_local_t *local = NULL;
+        call_frame_t *transaction_frame = NULL;
         int ret = -1;
-        int op_errno = 0;
-
-        VALIDATE_OR_GOTO (frame, out);
-        VALIDATE_OR_GOTO (this, out);
-        VALIDATE_OR_GOTO (this->private, out);
+        int op_errno = ENOMEM;
 
         priv = this->private;
 
-        if (afr_is_split_brain (this, fd->inode)) {
-                op_errno = EIO;
-                goto out;
-        }
-
         QUORUM_CHECK(fsetattr,out);
 
         transaction_frame = copy_frame (frame);
-        if (!transaction_frame) {
-                op_errno = ENOMEM;
+        if (!transaction_frame)
                 goto out;
-        }
 
-        AFR_LOCAL_ALLOC_OR_GOTO (transaction_frame->local, out);
-        local = transaction_frame->local;
-
-        ret = afr_local_init (local, priv, &op_errno);
-        if (ret < 0)
-                goto out;
+	local = AFR_FRAME_INIT (transaction_frame, op_errno);
+	if (!local)
+		goto out;
 
         local->cont.fsetattr.in_buf = *buf;
         local->cont.fsetattr.valid  = valid;
+	if (xdata)
+		local->xdata_req = dict_copy_with_ref (xdata, NULL);
+	else
+		local->xdata_req = dict_new ();
 
-        local->transaction.fop    = afr_fsetattr_wind;
-        local->transaction.done   = afr_fsetattr_done;
+	if (!local->xdata_req)
+		goto out;
+
+        local->transaction.wind   = afr_fsetattr_wind;
+        local->transaction.fop    = __afr_txn_write_fop;
+        local->transaction.done   = __afr_txn_write_done;
         local->transaction.unwind = afr_fsetattr_unwind;
 
         local->fd                 = fd_ref (fd);
+	local->inode = inode_ref (fd->inode);
+
+	local->op = GF_FOP_FSETATTR;
 
-        afr_open_fd_fix (fd, this);
+        afr_fix_open (fd, this);
 
         local->transaction.main_frame = frame;
         local->transaction.start   = LLONG_MAX - 1;
@@ -1383,18 +926,16 @@ afr_fsetattr (call_frame_t *frame, xlator_t *this,
 
         ret = afr_transaction (transaction_frame, this, AFR_METADATA_TRANSACTION);
         if (ret < 0) {
-            op_errno = -ret;
-            goto out;
+		op_errno = -ret;
+		goto out;
         }
 
-        ret = 0;
+	return 0;
 out:
-        if (ret < 0) {
-                if (transaction_frame)
-                        AFR_STACK_DESTROY (transaction_frame);
-                AFR_STACK_UNWIND (fsetattr, frame, -1, op_errno, NULL, NULL, NULL);
-        }
+	if (transaction_frame)
+		AFR_STACK_DESTROY (transaction_frame);
 
+	AFR_STACK_UNWIND (fsetattr, frame, -1, op_errno, NULL, NULL, NULL);
         return 0;
 }
 
@@ -1410,19 +951,12 @@ afr_setxattr_unwind (call_frame_t *frame, xlator_t *this)
 
         local = frame->local;
 
-        LOCK (&frame->lock);
-        {
-                if (local->transaction.main_frame)
-                        main_frame = local->transaction.main_frame;
-                local->transaction.main_frame = NULL;
-        }
-        UNLOCK (&frame->lock);
+	main_frame = afr_transaction_detach_fop_frame (frame);
+	if (!main_frame)
+		return 0;
 
-        if (main_frame) {
-                AFR_STACK_UNWIND (setxattr, main_frame,
-                                  local->op_ret, local->op_errno,
-                                  NULL);
-        }
+	AFR_STACK_UNWIND (setxattr, main_frame, local->op_ret, local->op_errno,
+			  local->xdata_rsp);
         return 0;
 }
 
@@ -1431,95 +965,32 @@ int
 afr_setxattr_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                        int32_t op_ret, int32_t op_errno, dict_t *xdata)
 {
-        afr_local_t   *local      = NULL;
-        afr_private_t *priv       = NULL;
-        int           call_count  = -1;
-        int           need_unwind = 0;
-        int           child_index = (long) cookie;
-
-        local = frame->local;
-        priv = this->private;
-
-        LOCK (&frame->lock);
-        {
-                __inode_write_fop_cbk (frame, child_index, -1, this,
-                                       &op_ret, &op_errno, NULL, NULL,
-                                       xdata);
-                if (local->success_count == priv->child_count) {
-                        need_unwind = 1;
-                }
-        }
-        UNLOCK (&frame->lock);
-
-        if (need_unwind)
-                local->transaction.unwind (frame, this);
-
-        call_count = afr_frame_return (frame);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-        }
-
-        return 0;
+        return __afr_inode_write_cbk (frame, cookie, this, op_ret, op_errno,
+				      NULL, NULL, xdata);
 }
 
 
 int
-afr_setxattr_wind (call_frame_t *frame, xlator_t *this)
+afr_setxattr_wind (call_frame_t *frame, xlator_t *this, int subvol)
 {
         afr_local_t      *local         = NULL;
         afr_private_t    *priv          = NULL;
-        int               call_count    = -1;
-        int               i             = 0;
 
         local = frame->local;
         priv = this->private;
 
-        call_count = afr_pre_op_done_children_count (local->transaction.pre_op,
-                                                     priv->child_count);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-                return 0;
-        }
-
-        local->call_count = call_count;
-
-        for (i = 0; i < priv->child_count; i++) {
-                if (local->transaction.pre_op[i]) {
-                        STACK_WIND_COOKIE (frame, afr_setxattr_wind_cbk,
-                                           (void *) (long) i,
-                                           priv->children[i],
-                                           priv->children[i]->fops->setxattr,
-                                           &local->loc,
-                                           local->cont.setxattr.dict,
-                                           local->cont.setxattr.flags,
-                                           NULL);
-
-                        if (!--call_count)
-                                break;
-                }
-        }
-
+	STACK_WIND_COOKIE (frame, afr_setxattr_wind_cbk, (void *) (long) subvol,
+			   priv->children[subvol],
+			   priv->children[subvol]->fops->setxattr,
+			   &local->loc, local->cont.setxattr.dict,
+			   local->cont.setxattr.flags, local->xdata_req);
         return 0;
 }
 
 
 int
-afr_setxattr_done (call_frame_t *frame, xlator_t *this)
-{
-        afr_local_t   *local    = frame->local;
-
-        local->transaction.unwind (frame, this);
-
-        AFR_STACK_DESTROY (frame);
-
-        return 0;
-}
-
-int
-afr_setxattr (call_frame_t *frame, xlator_t *this,
-              loc_t *loc, dict_t *dict, int32_t flags, dict_t *xdata)
+afr_setxattr (call_frame_t *frame, xlator_t *this, loc_t *loc, dict_t *dict,
+	      int32_t flags, dict_t *xdata)
 {
         afr_private_t  *priv              = NULL;
         afr_local_t    *local             = NULL;
@@ -1527,59 +998,60 @@ afr_setxattr (call_frame_t *frame, xlator_t *this,
         int             ret               = -1;
         int             op_errno          = EINVAL;
 
-        VALIDATE_OR_GOTO (this, out);
-
         GF_IF_INTERNAL_XATTR_GOTO ("trusted.afr.*", dict,
                                    op_errno, out);
 
         GF_IF_INTERNAL_XATTR_GOTO ("trusted.glusterfs.afr.*", dict,
                                    op_errno, out);
 
-        VALIDATE_OR_GOTO (frame, out);
-        VALIDATE_OR_GOTO (this->private, out);
-
         priv = this->private;
 
         QUORUM_CHECK(setxattr,out);
+
         transaction_frame = copy_frame (frame);
-        if (!transaction_frame) {
-                op_errno = ENOMEM;
+        if (!transaction_frame)
                 goto out;
-        }
-
-        AFR_LOCAL_ALLOC_OR_GOTO (transaction_frame->local, out);
-        local = transaction_frame->local;
 
-        ret = afr_local_init (local, priv, &op_errno);
-        if (ret < 0)
+	local = AFR_FRAME_INIT (transaction_frame, op_errno);
+	if (!local)
                 goto out;
 
         local->cont.setxattr.dict  = dict_ref (dict);
         local->cont.setxattr.flags = flags;
+	if (xdata)
+		local->xdata_req = dict_copy_with_ref (xdata, NULL);
+	else
+		local->xdata_req = dict_new ();
 
-        local->transaction.fop    = afr_setxattr_wind;
-        local->transaction.done   = afr_setxattr_done;
+	if (!local->xdata_req)
+		goto out;
+
+        local->transaction.wind   = afr_setxattr_wind;
+        local->transaction.fop    = __afr_txn_write_fop;
+        local->transaction.done   = __afr_txn_write_done;
         local->transaction.unwind = afr_setxattr_unwind;
 
         loc_copy (&local->loc, loc);
+	local->inode = inode_ref (loc->inode);
 
         local->transaction.main_frame = frame;
         local->transaction.start   = LLONG_MAX - 1;
         local->transaction.len     = 0;
 
+	local->op = GF_FOP_SETXATTR;
+
         ret = afr_transaction (transaction_frame, this, AFR_METADATA_TRANSACTION);
         if (ret < 0) {
-            op_errno = -ret;
-            goto out;
+		op_errno = -ret;
+		goto out;
         }
 
-        ret = 0;
+	return 0;
 out:
-        if (ret < 0) {
-                if (transaction_frame)
-                        AFR_STACK_DESTROY (transaction_frame);
-                AFR_STACK_UNWIND (setxattr, frame, -1, op_errno, NULL);
-        }
+	if (transaction_frame)
+		AFR_STACK_DESTROY (transaction_frame);
+
+	AFR_STACK_UNWIND (setxattr, frame, -1, op_errno, NULL);
 
         return 0;
 }
@@ -1595,19 +1067,12 @@ afr_fsetxattr_unwind (call_frame_t *frame, xlator_t *this)
 
         local = frame->local;
 
-        LOCK (&frame->lock);
-        {
-                if (local->transaction.main_frame)
-                        main_frame = local->transaction.main_frame;
-                local->transaction.main_frame = NULL;
-        }
-        UNLOCK (&frame->lock);
+	main_frame = afr_transaction_detach_fop_frame (frame);
+        if (!main_frame)
+		return 0;
 
-        if (main_frame) {
-                AFR_STACK_UNWIND (fsetxattr, main_frame,
-                                  local->op_ret, local->op_errno,
-                                  NULL);
-        }
+	AFR_STACK_UNWIND (fsetxattr, main_frame, local->op_ret, local->op_errno,
+			  local->xdata_rsp);
         return 0;
 }
 
@@ -1616,94 +1081,30 @@ int
 afr_fsetxattr_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                         int32_t op_ret, int32_t op_errno, dict_t *xdata)
 {
-        afr_local_t   *local      = NULL;
-        afr_private_t *priv       = NULL;
-        int           call_count  = -1;
-        int           need_unwind = 0;
-        int           child_index = (long) cookie;
-
-        local = frame->local;
-        priv = this->private;
-
-        LOCK (&frame->lock);
-        {
-
-                __inode_write_fop_cbk (frame, child_index, -1, this,
-                                       &op_ret, &op_errno, NULL, NULL,
-                                       xdata);
-                if (local->success_count == priv->child_count) {
-                        need_unwind = 1;
-                }
-        }
-        UNLOCK (&frame->lock);
-
-        if (need_unwind)
-                local->transaction.unwind (frame, this);
-
-        call_count = afr_frame_return (frame);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-        }
-
-        return 0;
+        return __afr_inode_write_cbk (frame, cookie, this, op_ret, op_errno,
+				      NULL, NULL, xdata);
 }
 
 
 int
-afr_fsetxattr_wind (call_frame_t *frame, xlator_t *this)
+afr_fsetxattr_wind (call_frame_t *frame, xlator_t *this, int subvol)
 {
         afr_local_t        *local       = NULL;
         afr_private_t      *priv        = NULL;
-        int                 call_count  = -1;
-        int                 i           = 0;
 
         local = frame->local;
         priv = this->private;
 
-        call_count = afr_pre_op_done_children_count (local->transaction.pre_op,
-                                                     priv->child_count);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-                return 0;
-        }
-
-        local->call_count = call_count;
-
-        for (i = 0; i < priv->child_count; i++) {
-                if (local->transaction.pre_op[i]) {
-                        STACK_WIND_COOKIE (frame, afr_fsetxattr_wind_cbk,
-                                           (void *) (long) i,
-                                           priv->children[i],
-                                           priv->children[i]->fops->fsetxattr,
-                                           local->fd,
-                                           local->cont.fsetxattr.dict,
-                                           local->cont.fsetxattr.flags,
-                                           NULL);
-
-                        if (!--call_count)
-                                break;
-                }
-        }
-
+	STACK_WIND_COOKIE (frame, afr_fsetxattr_wind_cbk, (void *) (long) subvol,
+			   priv->children[subvol],
+			   priv->children[subvol]->fops->fsetxattr,
+			   local->fd, local->cont.fsetxattr.dict,
+			   local->cont.fsetxattr.flags, local->xdata_req);
         return 0;
 }
 
 
 int
-afr_fsetxattr_done (call_frame_t *frame, xlator_t *this)
-{
-        afr_local_t   *local   = frame->local;
-
-        local->transaction.unwind (frame, this);
-
-        AFR_STACK_DESTROY (frame);
-
-        return 0;
-}
-
-int
 afr_fsetxattr (call_frame_t *frame, xlator_t *this,
                fd_t *fd, dict_t *dict, int32_t flags, dict_t *xdata)
 {
@@ -1711,11 +1112,7 @@ afr_fsetxattr (call_frame_t *frame, xlator_t *this,
         afr_local_t      *local             = NULL;
         call_frame_t     *transaction_frame = NULL;
         int               ret               = -1;
-        int               op_errno          = EINVAL;
-
-        VALIDATE_OR_GOTO (frame, out);
-        VALIDATE_OR_GOTO (this, out);
-        VALIDATE_OR_GOTO (this->private, out);
+        int               op_errno          = ENOMEM;
 
         GF_IF_INTERNAL_XATTR_GOTO ("trusted.afr.*", dict,
                                    op_errno, out);
@@ -1725,36 +1122,36 @@ afr_fsetxattr (call_frame_t *frame, xlator_t *this,
 
         priv = this->private;
 
-        if (afr_is_split_brain (this, fd->inode)) {
-                op_errno = EIO;
-                goto out;
-        }
-
         QUORUM_CHECK(fsetxattr,out);
 
-        AFR_LOCAL_ALLOC_OR_GOTO (local, out);
-
-        ret = afr_local_init (local, priv, &op_errno);
-        if (ret < 0)
-                goto out;
-
         transaction_frame = copy_frame (frame);
-        if (!transaction_frame) {
+        if (!transaction_frame)
                 goto out;
-        }
-
-        transaction_frame->local = local;
 
-        local->op_ret = -1;
+	local = AFR_FRAME_INIT (transaction_frame, op_errno);
+	if (!local)
+		goto out;
 
         local->cont.fsetxattr.dict  = dict_ref (dict);
         local->cont.fsetxattr.flags = flags;
 
-        local->transaction.fop    = afr_fsetxattr_wind;
-        local->transaction.done   = afr_fsetxattr_done;
+	if (xdata)
+		local->xdata_req = dict_copy_with_ref (xdata, NULL);
+	else
+		local->xdata_req = dict_new ();
+
+	if (!local->xdata_req)
+		goto out;
+
+        local->transaction.wind   = afr_fsetxattr_wind;
+        local->transaction.fop    = __afr_txn_write_fop;
+        local->transaction.done   = __afr_txn_write_done;
         local->transaction.unwind = afr_fsetxattr_unwind;
 
         local->fd                 = fd_ref (fd);
+	local->inode = inode_ref (fd->inode);
+
+	local->op = GF_FOP_FSETXATTR;
 
         local->transaction.main_frame = frame;
         local->transaction.start  = LLONG_MAX - 1;
@@ -1762,18 +1159,16 @@ afr_fsetxattr (call_frame_t *frame, xlator_t *this,
 
         ret = afr_transaction (transaction_frame, this, AFR_METADATA_TRANSACTION);
         if (ret < 0) {
-            op_errno = -ret;
-            goto out;
+		op_errno = -ret;
+		goto out;
         }
 
-        ret = 0;
+	return 0;
 out:
-        if (ret < 0) {
-                if (transaction_frame)
-                        AFR_STACK_DESTROY (transaction_frame);
-                AFR_STACK_UNWIND (fsetxattr, frame, -1, op_errno, NULL);
-        }
+	if (transaction_frame)
+		AFR_STACK_DESTROY (transaction_frame);
 
+	AFR_STACK_UNWIND (fsetxattr, frame, -1, op_errno, NULL);
         return 0;
 }
 
@@ -1791,19 +1186,12 @@ afr_removexattr_unwind (call_frame_t *frame, xlator_t *this)
 
         local = frame->local;
 
-        LOCK (&frame->lock);
-        {
-                if (local->transaction.main_frame)
-                        main_frame = local->transaction.main_frame;
-                local->transaction.main_frame = NULL;
-        }
-        UNLOCK (&frame->lock);
+	main_frame = afr_transaction_detach_fop_frame (frame);
+	if (!main_frame)
+		return 0;
 
-        if (main_frame) {
-                AFR_STACK_UNWIND (removexattr, main_frame,
-                                  local->op_ret, local->op_errno,
-                                  NULL);
-        }
+	AFR_STACK_UNWIND (removexattr, main_frame, local->op_ret, local->op_errno,
+			  local->xdata_rsp);
         return 0;
 }
 
@@ -1812,88 +1200,25 @@ int
 afr_removexattr_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                           int32_t op_ret, int32_t op_errno, dict_t *xdata)
 {
-        afr_local_t   *local      = NULL;
-        afr_private_t *priv       = NULL;
-        int           call_count  = -1;
-        int           need_unwind = 0;
-        int           child_index = (long) cookie;
-
-        local = frame->local;
-        priv = this->private;
-
-        LOCK (&frame->lock);
-        {
-                __inode_write_fop_cbk (frame, child_index, -1, this,
-                                       &op_ret, &op_errno, NULL, NULL,
-                                       xdata);
-                if (local->success_count == priv->wait_count) {
-                        need_unwind = 1;
-                }
-        }
-        UNLOCK (&frame->lock);
-
-        if (need_unwind)
-                local->transaction.unwind (frame, this);
-
-        call_count = afr_frame_return (frame);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-        }
-
-        return 0;
+        return __afr_inode_write_cbk (frame, cookie, this, op_ret, op_errno,
+				      NULL, NULL, xdata);
 }
 
 
-int32_t
-afr_removexattr_wind (call_frame_t *frame, xlator_t *this)
+int
+afr_removexattr_wind (call_frame_t *frame, xlator_t *this, int subvol)
 {
         afr_local_t *local = NULL;
         afr_private_t *priv = NULL;
-        int call_count = -1;
-        int i = 0;
-
-        local = frame->local;
-        priv = this->private;
-
-        call_count = afr_pre_op_done_children_count (local->transaction.pre_op,
-                                                     priv->child_count);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-                return 0;
-        }
-
-        local->call_count = call_count;
 
-        for (i = 0; i < priv->child_count; i++) {
-                if (local->transaction.pre_op[i]) {
-                        STACK_WIND_COOKIE (frame, afr_removexattr_wind_cbk,
-                                           (void *) (long) i,
-                                           priv->children[i],
-                                           priv->children[i]->fops->removexattr,
-                                           &local->loc,
-                                           local->cont.removexattr.name,
-                                           NULL);
-
-                        if (!--call_count)
-                                break;
-                }
-        }
-
-        return 0;
-}
-
-
-int
-afr_removexattr_done (call_frame_t *frame, xlator_t *this)
-{
-        afr_local_t * local = frame->local;
-
-        local->transaction.unwind (frame, this);
-
-        AFR_STACK_DESTROY (frame);
+	local = frame->local;
+	priv = this->private;
 
+	STACK_WIND_COOKIE (frame, afr_removexattr_wind_cbk, (void *) (long) subvol,
+			   priv->children[subvol],
+			   priv->children[subvol]->fops->removexattr,
+			   &local->loc, local->cont.removexattr.name,
+			   local->xdata_req);
         return 0;
 }
 
@@ -1906,9 +1231,7 @@ afr_removexattr (call_frame_t *frame, xlator_t *this,
         afr_local_t     *local             = NULL;
         call_frame_t    *transaction_frame = NULL;
         int              ret               = -1;
-        int              op_errno          = 0;
-
-        VALIDATE_OR_GOTO (this, out);
+        int              op_errno          = ENOMEM;
 
         GF_IF_NATIVE_XATTR_GOTO ("trusted.afr.*",
                                  name, op_errno, out);
@@ -1916,34 +1239,37 @@ afr_removexattr (call_frame_t *frame, xlator_t *this,
         GF_IF_NATIVE_XATTR_GOTO ("trusted.glusterfs.afr.*",
                                  name, op_errno, out);
 
-        VALIDATE_OR_GOTO (frame, out);
-        VALIDATE_OR_GOTO (this->private, out);
-        VALIDATE_OR_GOTO (loc, out);
-
         priv = this->private;
 
         QUORUM_CHECK(removexattr,out);
 
         transaction_frame = copy_frame (frame);
-        if (!transaction_frame) {
-                op_errno = ENOMEM;
+        if (!transaction_frame)
                 goto out;
-        }
 
-        AFR_LOCAL_ALLOC_OR_GOTO (transaction_frame->local, out);
-        local = transaction_frame->local;
-
-        ret = afr_local_init (local, priv, &op_errno);
-        if (ret < 0)
-                goto out;
+	local = AFR_FRAME_INIT (transaction_frame, op_errno);
+	if (!local)
+		goto out;
 
         local->cont.removexattr.name = gf_strdup (name);
 
-        local->transaction.fop    = afr_removexattr_wind;
-        local->transaction.done   = afr_removexattr_done;
+	if (xdata)
+		local->xdata_req = dict_copy_with_ref (xdata, NULL);
+	else
+		local->xdata_req = dict_new ();
+
+	if (!local->xdata_req)
+		goto out;
+
+        local->transaction.wind   = afr_removexattr_wind;
+        local->transaction.fop    = __afr_txn_write_fop;
+        local->transaction.done   = __afr_txn_write_done;
         local->transaction.unwind = afr_removexattr_unwind;
 
         loc_copy (&local->loc, loc);
+	local->inode = inode_ref (loc->inode);
+
+	local->op = GF_FOP_REMOVEXATTR;
 
         local->transaction.main_frame = frame;
         local->transaction.start   = LLONG_MAX - 1;
@@ -1951,18 +1277,16 @@ afr_removexattr (call_frame_t *frame, xlator_t *this,
 
         ret = afr_transaction (transaction_frame, this, AFR_METADATA_TRANSACTION);
         if (ret < 0) {
-            op_errno = -ret;
-            goto out;
+		op_errno = -ret;
+		goto out;
         }
 
-        ret = 0;
+	return 0;
 out:
-        if (ret < 0) {
-                if (transaction_frame)
-                        AFR_STACK_DESTROY (transaction_frame);
-                AFR_STACK_UNWIND (removexattr, frame, -1, op_errno, NULL);
-        }
+	if (transaction_frame)
+		AFR_STACK_DESTROY (transaction_frame);
 
+	AFR_STACK_UNWIND (removexattr, frame, -1, op_errno, NULL);
         return 0;
 }
 
@@ -1975,19 +1299,12 @@ afr_fremovexattr_unwind (call_frame_t *frame, xlator_t *this)
 
         local = frame->local;
 
-        LOCK (&frame->lock);
-        {
-                if (local->transaction.main_frame)
-                        main_frame = local->transaction.main_frame;
-                local->transaction.main_frame = NULL;
-        }
-        UNLOCK (&frame->lock);
+	main_frame = afr_transaction_detach_fop_frame (frame);
+        if (!main_frame)
+		return 0;
 
-        if (main_frame) {
-                AFR_STACK_UNWIND (fremovexattr, main_frame,
-                                  local->op_ret, local->op_errno,
-                                  NULL);
-        }
+	AFR_STACK_UNWIND (fremovexattr, main_frame, local->op_ret, local->op_errno,
+			  local->xdata_rsp);
         return 0;
 }
 
@@ -1996,105 +1313,38 @@ int
 afr_fremovexattr_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                           int32_t op_ret, int32_t op_errno, dict_t *xdata)
 {
-        afr_local_t *   local       = NULL;
-        afr_private_t * priv        = NULL;
-        int             call_count  = -1;
-        int             need_unwind = 0;
-        int             child_index = (long) cookie;
-
-        local = frame->local;
-        priv = this->private;
-
-        LOCK (&frame->lock);
-        {
-                __inode_write_fop_cbk (frame, child_index, -1, this,
-                                       &op_ret, &op_errno, NULL, NULL,
-                                       xdata);
-
-                if (local->success_count == priv->wait_count) {
-                        need_unwind = 1;
-                }
-        }
-        UNLOCK (&frame->lock);
-
-        if (need_unwind)
-                local->transaction.unwind (frame, this);
-
-        call_count = afr_frame_return (frame);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-        }
-
-        return 0;
+	return __afr_inode_write_cbk (frame, cookie, this, op_ret, op_errno,
+				      NULL, NULL, xdata);
 }
 
 
-int32_t
-afr_fremovexattr_wind (call_frame_t *frame, xlator_t *this)
+int
+afr_fremovexattr_wind (call_frame_t *frame, xlator_t *this, int subvol)
 {
         afr_local_t *local = NULL;
         afr_private_t *priv = NULL;
-        int call_count = -1;
-        int i = 0;
 
         local = frame->local;
         priv = this->private;
 
-        call_count = afr_pre_op_done_children_count (local->transaction.pre_op,
-                                                     priv->child_count);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-                return 0;
-        }
-
-        local->call_count = call_count;
-
-        for (i = 0; i < priv->child_count; i++) {
-                if (local->transaction.pre_op[i]) {
-                        STACK_WIND_COOKIE (frame, afr_fremovexattr_wind_cbk,
-                                           (void *) (long) i,
-                                           priv->children[i],
-                                           priv->children[i]->fops->fremovexattr,
-                                           local->fd,
-                                           local->cont.removexattr.name,
-                                           NULL);
-
-                        if (!--call_count)
-                                break;
-                }
-        }
-
+	STACK_WIND_COOKIE (frame, afr_fremovexattr_wind_cbk, (void *) (long) subvol,
+			   priv->children[subvol],
+			   priv->children[subvol]->fops->fremovexattr,
+			   local->fd, local->cont.removexattr.name,
+			   local->xdata_req);
         return 0;
 }
 
 
 int
-afr_fremovexattr_done (call_frame_t *frame, xlator_t *this)
+afr_fremovexattr (call_frame_t *frame, xlator_t *this, fd_t *fd,
+		  const char *name, dict_t *xdata)
 {
-        afr_local_t * local = frame->local;
-
-        local->transaction.unwind (frame, this);
-
-        AFR_STACK_DESTROY (frame);
-
-        return 0;
-}
-
-
-int
-afr_fremovexattr (call_frame_t *frame, xlator_t *this,
-                  fd_t *fd, const char *name, dict_t *xdata)
-{
-        afr_private_t * priv  = NULL;
-        afr_local_t   * local = NULL;
-        call_frame_t   *transaction_frame = NULL;
+	afr_private_t *priv = NULL;
+        afr_local_t *local = NULL;
+        call_frame_t *transaction_frame = NULL;
         int ret = -1;
-        int op_ret   = -1;
-        int op_errno = 0;
-
-        VALIDATE_OR_GOTO (this, out);
+        int op_errno = ENOMEM;
 
         GF_IF_NATIVE_XATTR_GOTO ("trusted.afr.*",
                                  name, op_errno, out);
@@ -2102,64 +1352,59 @@ afr_fremovexattr (call_frame_t *frame, xlator_t *this,
         GF_IF_NATIVE_XATTR_GOTO ("trusted.glusterfs.afr.*",
                                  name, op_errno, out);
 
-        VALIDATE_OR_GOTO (frame, out);
-        VALIDATE_OR_GOTO (this->private, out);
-
-        priv = this->private;
-        if (afr_is_split_brain (this, fd->inode)) {
-                op_errno = EIO;
-                goto out;
-        }
+	priv = this->private;
 
         QUORUM_CHECK(fremovexattr, out);
 
         transaction_frame = copy_frame (frame);
-        if (!transaction_frame) {
+        if (!transaction_frame)
                 goto out;
-        }
 
-        AFR_LOCAL_ALLOC_OR_GOTO (local, out);
-
-        ret = afr_local_init (local, priv, &op_errno);
-        if (ret < 0) {
-                op_errno = -ret;
+	local = AFR_FRAME_INIT (transaction_frame, op_errno);
+	if (!local)
                 goto out;
-        }
-
-        transaction_frame->local = local;
-
-        local->op_ret = -1;
 
         local->cont.removexattr.name = gf_strdup (name);
+	if (xdata)
+		local->xdata_req = dict_copy_with_ref (xdata, NULL);
+	else
+		local->xdata_req = dict_new ();
 
-        local->transaction.fop    = afr_fremovexattr_wind;
-        local->transaction.done   = afr_fremovexattr_done;
+	if (!local->xdata_req)
+		goto out;
+
+        local->transaction.wind   = afr_fremovexattr_wind;
+        local->transaction.fop    = __afr_txn_write_fop;
+        local->transaction.done   = __afr_txn_write_done;
         local->transaction.unwind = afr_fremovexattr_unwind;
 
         local->fd = fd_ref (fd);
+	local->inode = inode_ref (fd->inode);
+
+	local->op = GF_FOP_FREMOVEXATTR;
 
         local->transaction.main_frame = frame;
         local->transaction.start   = LLONG_MAX - 1;
         local->transaction.len     = 0;
 
-        op_ret = afr_transaction (transaction_frame, this, AFR_METADATA_TRANSACTION);
-        if (op_ret < 0) {
-            op_errno = -op_ret;
-            goto out;
+        ret = afr_transaction (transaction_frame, this, AFR_METADATA_TRANSACTION);
+        if (ret < 0) {
+		op_errno = -ret;
+		goto out;
         }
 
-        op_ret = 0;
+	return 0;
 out:
-        if (op_ret < 0) {
-                if (transaction_frame)
-                        AFR_STACK_DESTROY (transaction_frame);
-                AFR_STACK_UNWIND (fremovexattr, frame, op_ret, op_errno, NULL);
-        }
+	if (transaction_frame)
+		AFR_STACK_DESTROY (transaction_frame);
+
+	AFR_STACK_UNWIND (fremovexattr, frame, -1, op_errno, NULL);
 
         return 0;
 }
 
-static int
+
+int
 afr_fallocate_unwind (call_frame_t *frame, xlator_t *this)
 {
         afr_local_t *   local = NULL;
@@ -2167,147 +1412,88 @@ afr_fallocate_unwind (call_frame_t *frame, xlator_t *this)
 
         local = frame->local;
 
-        LOCK (&frame->lock);
-        {
-                if (local->transaction.main_frame)
-                        main_frame = local->transaction.main_frame;
-                local->transaction.main_frame = NULL;
-        }
-        UNLOCK (&frame->lock);
+	main_frame = afr_transaction_detach_fop_frame (frame);
+	if (!main_frame)
+		return 0;
 
-        if (main_frame) {
-                AFR_STACK_UNWIND (fallocate, main_frame, local->op_ret,
-                                  local->op_errno,
-                                  &local->cont.inode_wfop.prebuf,
-                                  &local->cont.inode_wfop.postbuf,
-                                  NULL);
-        }
+	AFR_STACK_UNWIND (fallocate, main_frame, local->op_ret, local->op_errno,
+			  &local->cont.inode_wfop.prebuf,
+			  &local->cont.inode_wfop.postbuf, local->xdata_rsp);
         return 0;
 }
 
-static int
+
+int
 afr_fallocate_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                         int32_t op_ret, int32_t op_errno, struct iatt *prebuf,
                         struct iatt *postbuf, dict_t *xdata)
 {
-        afr_local_t *   local = NULL;
-        afr_private_t * priv  = NULL;
-        int child_index = (long) cookie;
-        int call_count  = -1;
-        int need_unwind = 0;
-        int read_child  = 0;
-
-        local = frame->local;
-        priv  = this->private;
-
-        read_child = afr_inode_get_read_ctx (this, local->fd->inode, NULL);
-
-        LOCK (&frame->lock);
-        {
-                if (child_index == read_child) {
-                        local->read_child_returned = _gf_true;
-                }
-
-                __inode_write_fop_cbk (frame, child_index, read_child, this,
-                                       &op_ret, &op_errno, prebuf, postbuf,
-                                       xdata);
-
-                if ((local->success_count >= priv->wait_count)
-                    && local->read_child_returned) {
-                        need_unwind = 1;
-                }
-        }
-        UNLOCK (&frame->lock);
-
-        if (need_unwind)
-                local->transaction.unwind (frame, this);
-
-        call_count = afr_frame_return (frame);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-        }
-
-        return 0;
+        return __afr_inode_write_cbk (frame, cookie, this, op_ret, op_errno,
+				      prebuf, postbuf, xdata);
 }
 
-static int
-afr_fallocate_wind (call_frame_t *frame, xlator_t *this)
+
+int
+afr_fallocate_wind (call_frame_t *frame, xlator_t *this, int subvol)
 {
         afr_local_t *local = NULL;
         afr_private_t *priv = NULL;
-        int call_count = -1;
-        int i = 0;
 
         local = frame->local;
         priv = this->private;
 
-        call_count = afr_pre_op_done_children_count (local->transaction.pre_op,
-                                                     priv->child_count);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-                return 0;
-        }
-
-        local->call_count = call_count;
-
-        for (i = 0; i < priv->child_count; i++) {
-                if (local->transaction.pre_op[i]) {
-                        STACK_WIND_COOKIE (frame, afr_fallocate_wind_cbk,
-                                           (void *) (long) i,
-                                           priv->children[i],
-                                           priv->children[i]->fops->fallocate,
-                                           local->fd,
-                                           local->cont.fallocate.mode,
-                                           local->cont.fallocate.offset,
-                                           local->cont.fallocate.len,
-                                           NULL);
-
-                        if (!--call_count)
-                                break;
-                }
-        }
-
+	STACK_WIND_COOKIE (frame, afr_fallocate_wind_cbk, (void *) (long) subvol,
+			   priv->children[subvol],
+			   priv->children[subvol]->fops->fallocate,
+			   local->fd, local->cont.fallocate.mode,
+			   local->cont.fallocate.offset,
+			   local->cont.fallocate.len, local->xdata_req);
         return 0;
 }
 
-static int
-afr_fallocate_done (call_frame_t *frame, xlator_t *this)
+
+int
+afr_fallocate (call_frame_t *frame, xlator_t *this, fd_t *fd, int32_t mode,
+               off_t offset, size_t len, dict_t *xdata)
 {
+	afr_private_t *priv = NULL;
+        call_frame_t *transaction_frame = NULL;
         afr_local_t *local = NULL;
+        int ret = -1;
+        int op_errno = ENOMEM;
 
-        local = frame->local;
+	priv = this->private;
 
-        local->transaction.unwind (frame, this);
+        QUORUM_CHECK(fallocate,out);
 
-        AFR_STACK_DESTROY (frame);
+        transaction_frame = copy_frame (frame);
+        if (!transaction_frame)
+                goto out;
 
-        return 0;
-}
+	local = AFR_FRAME_INIT (transaction_frame, op_errno);
+	if (!local)
+		goto out;
 
-static int
-afr_do_fallocate (call_frame_t *frame, xlator_t *this)
-{
-        call_frame_t * transaction_frame = NULL;
-        afr_local_t *  local             = NULL;
-        int op_ret   = -1;
-        int op_errno = 0;
+        local->cont.fallocate.mode = mode;
+        local->cont.fallocate.offset  = offset;
+        local->cont.fallocate.len = len;
 
-        local = frame->local;
+        local->fd = fd_ref (fd);
+	local->inode = inode_ref (fd->inode);
 
-        transaction_frame = copy_frame (frame);
-        if (!transaction_frame) {
-                goto out;
-        }
+	if (xdata)
+		local->xdata_req = dict_copy_with_ref (xdata, NULL);
+	else
+		local->xdata_req = dict_new ();
 
-        transaction_frame->local = local;
-        frame->local = NULL;
+	if (!local->xdata_req)
+		goto out;
 
         local->op = GF_FOP_FALLOCATE;
 
-        local->transaction.fop    = afr_fallocate_wind;
-        local->transaction.done   = afr_fallocate_done;
+        local->transaction.wind   = afr_fallocate_wind;
+        local->transaction.fop    = __afr_txn_write_fop;
+        local->transaction.done   = __afr_txn_write_done;
         local->transaction.unwind = afr_fallocate_unwind;
 
         local->transaction.main_frame = frame;
@@ -2315,80 +1501,29 @@ afr_do_fallocate (call_frame_t *frame, xlator_t *this)
         local->transaction.start   = local->cont.fallocate.offset;
         local->transaction.len     = 0;
 
-        /* fallocate can modify the file size */
-        op_ret = afr_transaction (transaction_frame, this, AFR_DATA_TRANSACTION);
-        if (op_ret < 0) {
-            op_errno = -op_ret;
-            goto out;
-        }
-
-        op_ret = 0;
-out:
-        if (op_ret < 0) {
-                if (transaction_frame)
-                        AFR_STACK_DESTROY (transaction_frame);
-                AFR_STACK_UNWIND (fallocate, frame, op_ret, op_errno, NULL,
-                                  NULL, NULL);
-        }
-
-        return 0;
-}
-
-int
-afr_fallocate (call_frame_t *frame, xlator_t *this, fd_t *fd, int32_t mode,
-               off_t offset, size_t len, dict_t *xdata)
-{
-        afr_private_t * priv  = NULL;
-        afr_local_t   * local = NULL;
-        call_frame_t   *transaction_frame = NULL;
-        int ret = -1;
-        int op_errno = 0;
-
-        VALIDATE_OR_GOTO (frame, out);
-        VALIDATE_OR_GOTO (this, out);
-        VALIDATE_OR_GOTO (this->private, out);
+        afr_fix_open (fd, this);
 
-        priv = this->private;
-
-        if (afr_is_split_brain (this, fd->inode)) {
-                op_errno = EIO;
-                goto out;
+        ret = afr_transaction (transaction_frame, this, AFR_DATA_TRANSACTION);
+        if (ret < 0) {
+		op_errno = -ret;
+		goto out;
         }
-        QUORUM_CHECK(fallocate,out);
-
-        AFR_LOCAL_ALLOC_OR_GOTO (frame->local, out);
-        local = frame->local;
-
-        ret = afr_local_init (local, priv, &op_errno);
-        if (ret < 0)
-                goto out;
 
-        local->cont.fallocate.mode = mode;
-        local->cont.fallocate.offset  = offset;
-        local->cont.fallocate.len = len;
-
-        local->fd = fd_ref (fd);
-
-        afr_open_fd_fix (fd, this);
-
-        afr_do_fallocate (frame, this);
-
-        ret = 0;
+	return 0;
 out:
-        if (ret < 0) {
-                if (transaction_frame)
-                        AFR_STACK_DESTROY (transaction_frame);
-                AFR_STACK_UNWIND (fallocate, frame, -1, op_errno, NULL, NULL, NULL);
-        }
+	if (transaction_frame)
+		AFR_STACK_DESTROY (transaction_frame);
 
+	AFR_STACK_UNWIND (fallocate, frame, -1, op_errno, NULL, NULL, NULL);
         return 0;
 }
 
+
 /* }}} */
 
 /* {{{ discard */
 
-static int
+int
 afr_discard_unwind (call_frame_t *frame, xlator_t *this)
 {
         afr_local_t *   local = NULL;
@@ -2396,146 +1531,86 @@ afr_discard_unwind (call_frame_t *frame, xlator_t *this)
 
         local = frame->local;
 
-        LOCK (&frame->lock);
-        {
-                if (local->transaction.main_frame)
-                        main_frame = local->transaction.main_frame;
-                local->transaction.main_frame = NULL;
-        }
-        UNLOCK (&frame->lock);
+	main_frame = afr_transaction_detach_fop_frame (frame);
+        if (!main_frame)
+		return 0;
 
-        if (main_frame) {
-                AFR_STACK_UNWIND (discard, main_frame, local->op_ret,
-                                  local->op_errno,
-                                  &local->cont.inode_wfop.prebuf,
-                                  &local->cont.inode_wfop.postbuf,
-                                  NULL);
-        }
+	AFR_STACK_UNWIND (discard, main_frame, local->op_ret, local->op_errno,
+			  &local->cont.inode_wfop.prebuf,
+			  &local->cont.inode_wfop.postbuf, local->xdata_rsp);
         return 0;
 }
 
-static int
+
+int
 afr_discard_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                       int32_t op_ret, int32_t op_errno, struct iatt *prebuf,
                       struct iatt *postbuf, dict_t *xdata)
 {
-        afr_local_t *   local = NULL;
-        afr_private_t * priv  = NULL;
-        int child_index = (long) cookie;
-        int call_count  = -1;
-        int need_unwind = 0;
-        int read_child  = 0;
-
-        local = frame->local;
-        priv  = this->private;
-
-        read_child = afr_inode_get_read_ctx (this, local->fd->inode, NULL);
-
-        LOCK (&frame->lock);
-        {
-                if (child_index == read_child) {
-                        local->read_child_returned = _gf_true;
-                }
-
-                __inode_write_fop_cbk (frame, child_index, read_child, this,
-                                       &op_ret, &op_errno, prebuf, postbuf,
-                                       xdata);
-
-                if ((local->success_count >= priv->wait_count)
-                    && local->read_child_returned) {
-                        need_unwind = 1;
-                }
-        }
-        UNLOCK (&frame->lock);
-
-        if (need_unwind)
-                local->transaction.unwind (frame, this);
-
-        call_count = afr_frame_return (frame);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-        }
-
-        return 0;
+	return __afr_inode_write_cbk (frame, cookie, this, op_ret, op_errno,
+				      prebuf, postbuf, xdata);
 }
 
-static int
-afr_discard_wind (call_frame_t *frame, xlator_t *this)
+
+int
+afr_discard_wind (call_frame_t *frame, xlator_t *this, int subvol)
 {
         afr_local_t *local = NULL;
         afr_private_t *priv = NULL;
-        int call_count = -1;
-        int i = 0;
 
         local = frame->local;
         priv = this->private;
 
-        call_count = afr_pre_op_done_children_count (local->transaction.pre_op,
-                                                     priv->child_count);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-                return 0;
-        }
-
-        local->call_count = call_count;
-
-        for (i = 0; i < priv->child_count; i++) {
-                if (local->transaction.pre_op[i]) {
-                        STACK_WIND_COOKIE (frame, afr_discard_wind_cbk,
-                                           (void *) (long) i,
-                                           priv->children[i],
-                                           priv->children[i]->fops->discard,
-                                           local->fd,
-                                           local->cont.discard.offset,
-                                           local->cont.discard.len,
-                                           NULL);
-
-                        if (!--call_count)
-                                break;
-                }
-        }
-
+	STACK_WIND_COOKIE (frame, afr_discard_wind_cbk, (void *) (long) subvol,
+			   priv->children[subvol],
+			   priv->children[subvol]->fops->discard,
+			   local->fd, local->cont.discard.offset,
+			   local->cont.discard.len, local->xdata_req);
         return 0;
 }
 
-static int
-afr_discard_done (call_frame_t *frame, xlator_t *this)
+
+int
+afr_discard (call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset,
+             size_t len, dict_t *xdata)
 {
+        afr_private_t *priv  = NULL;
         afr_local_t *local = NULL;
+        call_frame_t *transaction_frame = NULL;
+        int ret = -1;
+        int op_errno = ENOMEM;
 
-        local = frame->local;
+        priv = this->private;
 
-        local->transaction.unwind (frame, this);
+        QUORUM_CHECK(discard, out);
 
-        AFR_STACK_DESTROY (frame);
+	transaction_frame = copy_frame (frame);
+	if (!transaction_frame)
+		goto out;
 
-        return 0;
-}
+	local = AFR_FRAME_INIT (frame, op_errno);
+	if (!local)
+		goto out;
 
-static int
-afr_do_discard (call_frame_t *frame, xlator_t *this)
-{
-        call_frame_t * transaction_frame = NULL;
-        afr_local_t *  local             = NULL;
-        int op_ret   = -1;
-        int op_errno = 0;
+        local->cont.discard.offset  = offset;
+        local->cont.discard.len = len;
 
-        local = frame->local;
+        local->fd = fd_ref (fd);
+	local->inode = inode_ref (fd->inode);
 
-        transaction_frame = copy_frame (frame);
-        if (!transaction_frame) {
-                goto out;
-        }
+	if (xdata)
+		local->xdata_req = dict_copy_with_ref (xdata, NULL);
+	else
+		local->xdata_req = dict_new ();
 
-        transaction_frame->local = local;
-        frame->local = NULL;
+	if (!local->xdata_req)
+		goto out;
 
         local->op = GF_FOP_DISCARD;
 
-        local->transaction.fop    = afr_discard_wind;
-        local->transaction.done   = afr_discard_done;
+        local->transaction.wind   = afr_discard_wind;
+        local->transaction.fop    = __afr_txn_write_fop;
+        local->transaction.done   = __afr_txn_write_done;
         local->transaction.unwind = afr_discard_unwind;
 
         local->transaction.main_frame = frame;
@@ -2543,316 +1618,134 @@ afr_do_discard (call_frame_t *frame, xlator_t *this)
         local->transaction.start   = local->cont.discard.offset;
         local->transaction.len     = 0;
 
-        op_ret = afr_transaction (transaction_frame, this, AFR_DATA_TRANSACTION);
-        if (op_ret < 0) {
-            op_errno = -op_ret;
-            goto out;
-        }
+        afr_fix_open (fd, this);
 
-        op_ret = 0;
-out:
-        if (op_ret < 0) {
-                if (transaction_frame)
-                        AFR_STACK_DESTROY (transaction_frame);
-                AFR_STACK_UNWIND (discard, frame, op_ret, op_errno, NULL,
-                                  NULL, NULL);
-        }
-
-        return 0;
-}
-
-int
-afr_discard (call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset,
-             size_t len, dict_t *xdata)
-{
-        afr_private_t * priv  = NULL;
-        afr_local_t   * local = NULL;
-        call_frame_t   *transaction_frame = NULL;
-        int ret = -1;
-        int op_errno = 0;
-
-        VALIDATE_OR_GOTO (frame, out);
-        VALIDATE_OR_GOTO (this, out);
-        VALIDATE_OR_GOTO (this->private, out);
-
-        priv = this->private;
-
-        if (afr_is_split_brain (this, fd->inode)) {
-                op_errno = EIO;
-                goto out;
+        ret = afr_transaction (transaction_frame, this, AFR_DATA_TRANSACTION);
+        if (ret < 0) {
+		op_errno = -ret;
+		goto out;
         }
-        QUORUM_CHECK(discard, out);
-
-        AFR_LOCAL_ALLOC_OR_GOTO (frame->local, out);
-        local = frame->local;
-
-        ret = afr_local_init (local, priv, &op_errno);
-        if (ret < 0)
-                goto out;
 
-        local->cont.discard.offset  = offset;
-        local->cont.discard.len = len;
-
-        local->fd = fd_ref (fd);
-
-        afr_open_fd_fix (fd, this);
-
-        afr_do_discard(frame, this);
-
-        ret = 0;
+	return 0;
 out:
-        if (ret < 0) {
-                if (transaction_frame)
-                        AFR_STACK_DESTROY (transaction_frame);
-                AFR_STACK_UNWIND (discard, frame, -1, op_errno, NULL, NULL, NULL);
-        }
+	if (transaction_frame)
+		AFR_STACK_DESTROY (transaction_frame);
 
+	AFR_STACK_UNWIND (discard, frame, -1, op_errno, NULL, NULL, NULL);
         return 0;
 }
 
 
 /* {{{ zerofill */
 
-static int
+int
 afr_zerofill_unwind (call_frame_t *frame, xlator_t *this)
 {
-        afr_local_t     *local            = NULL;
-        call_frame_t    *main_frame       = NULL;
+        afr_local_t *   local = NULL;
+        call_frame_t   *main_frame = NULL;
 
         local = frame->local;
 
-        LOCK (&frame->lock);
-        {
-                if (local->transaction.main_frame) {
-                        main_frame = local->transaction.main_frame;
-                }
-                local->transaction.main_frame = NULL;
-        }
-        UNLOCK (&frame->lock);
+	main_frame = afr_transaction_detach_fop_frame (frame);
+        if (!main_frame)
+		return 0;
 
-        if (main_frame) {
-                AFR_STACK_UNWIND (zerofill, main_frame, local->op_ret,
-                                  local->op_errno,
-                                  &local->cont.zerofill.prebuf,
-                                  &local->cont.zerofill.postbuf,
-                                  NULL);
-        }
+	AFR_STACK_UNWIND (discard, main_frame, local->op_ret, local->op_errno,
+			  &local->cont.inode_wfop.prebuf,
+			  &local->cont.inode_wfop.postbuf, local->xdata_rsp);
         return 0;
 }
 
-static int
-afr_zerofill_wind_cbk(call_frame_t *frame, void *cookie, xlator_t *this,
-                     int32_t op_ret, int32_t op_errno, struct iatt *prebuf,
-                     struct iatt *postbuf, dict_t *xdata)
-{
-        afr_local_t       *local             = NULL;
-        afr_private_t     *priv              = NULL;
-        int                child_index       = (long) cookie;
-        int                call_count        = -1;
-        int                need_unwind       = 0;
-        int                read_child        = 0;
-
-        local = frame->local;
-        priv  = this->private;
-
-        read_child = afr_inode_get_read_ctx (this, local->fd->inode, NULL);
-
-        LOCK (&frame->lock);
-        {
-                if (child_index == read_child) {
-                        local->read_child_returned = _gf_true;
-                }
-
-                if (afr_fop_failed (op_ret, op_errno)) {
-                        afr_transaction_fop_failed (frame, this, child_index);
-                }
 
-                if (op_ret != -1) {
-                        if (local->success_count == 0) {
-                                local->op_ret = op_ret;
-                                local->cont.zerofill.prebuf  = *prebuf;
-                                local->cont.zerofill.postbuf = *postbuf;
-                        }
-
-                        if (child_index == read_child) {
-                                local->cont.zerofill.prebuf  = *prebuf;
-                                local->cont.zerofill.postbuf = *postbuf;
-                        }
-
-                        local->success_count++;
-
-                        if ((local->success_count >= priv->wait_count)
-                            && local->read_child_returned) {
-                                need_unwind = 1;
-                        }
-                }
-                local->op_errno = op_errno;
-        }
-        UNLOCK (&frame->lock);
-
-        if (need_unwind) {
-                local->transaction.unwind (frame, this);
-        }
-        call_count = afr_frame_return (frame);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-        }
-
-        return 0;
+int
+afr_zerofill_wind_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
+                      int32_t op_ret, int32_t op_errno, struct iatt *prebuf,
+                      struct iatt *postbuf, dict_t *xdata)
+{
+	return __afr_inode_write_cbk (frame, cookie, this, op_ret, op_errno,
+				      prebuf, postbuf, xdata);
 }
 
-static int
-afr_zerofill_wind (call_frame_t *frame, xlator_t *this)
+
+int
+afr_zerofill_wind (call_frame_t *frame, xlator_t *this, int subvol)
 {
-        afr_local_t    *local         = NULL;
-        afr_private_t  *priv          = NULL;
-        int             call_count    = -1;
-        int             i             = 0;
+        afr_local_t *local = NULL;
+        afr_private_t *priv = NULL;
 
         local = frame->local;
         priv = this->private;
 
-        call_count = afr_pre_op_done_children_count (local->transaction.pre_op,
-                                                     priv->child_count);
-
-        if (call_count == 0) {
-                local->transaction.resume (frame, this);
-                return 0;
-        }
-
-        local->call_count = call_count;
-
-        for (i = 0; i < priv->child_count; i++) {
-                if (local->transaction.pre_op[i]) {
-                        STACK_WIND_COOKIE (frame, afr_zerofill_wind_cbk,
-                                           (void *) (long) i,
-                                           priv->children[i],
-                                           priv->children[i]->fops->zerofill,
-                                           local->fd,
-                                           local->cont.zerofill.offset,
-                                           local->cont.zerofill.len,
-                                           NULL);
-
-                        if (!--call_count)
-                                break;
-                }
-        }
-
+	STACK_WIND_COOKIE (frame, afr_zerofill_wind_cbk, (void *) (long) subvol,
+			   priv->children[subvol],
+			   priv->children[subvol]->fops->zerofill,
+			   local->fd, local->cont.zerofill.offset,
+			   local->cont.zerofill.len, local->xdata_req);
         return 0;
 }
 
-static int
-afr_zerofill_done (call_frame_t *frame, xlator_t *this)
+int
+afr_zerofill (call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset,
+             size_t len, dict_t *xdata)
 {
+        afr_private_t *priv  = NULL;
         afr_local_t *local = NULL;
+        call_frame_t *transaction_frame = NULL;
+        int ret = -1;
+        int op_errno = ENOMEM;
 
-        local = frame->local;
+        priv = this->private;
 
-        local->transaction.unwind (frame, this);
+        QUORUM_CHECK(discard, out);
 
-        AFR_STACK_DESTROY (frame);
+	transaction_frame = copy_frame (frame);
+	if (!transaction_frame)
+		goto out;
 
-        return 0;
-}
+	local = AFR_FRAME_INIT (frame, op_errno);
+	if (!local)
+		goto out;
 
-static int
-afr_do_zerofill(call_frame_t *frame, xlator_t *this)
-{
-        call_frame_t  *transaction_frame = NULL;
-        afr_local_t   *local             = NULL;
-        int            op_ret            = -1;
-        int            op_errno          = 0;
+        local->cont.zerofill.offset  = offset;
+        local->cont.zerofill.len = len;
 
-        local = frame->local;
+        local->fd = fd_ref (fd);
+	local->inode = inode_ref (fd->inode);
 
-        transaction_frame = copy_frame (frame);
-        if (!transaction_frame) {
-                goto out;
-        }
+	if (xdata)
+		local->xdata_req = dict_copy_with_ref (xdata, NULL);
+	else
+		local->xdata_req = dict_new ();
 
-        transaction_frame->local = local;
-        frame->local = NULL;
+	if (!local->xdata_req)
+		goto out;
 
         local->op = GF_FOP_ZEROFILL;
 
-        local->transaction.fop    = afr_zerofill_wind;
-        local->transaction.done   = afr_zerofill_done;
+        local->transaction.wind   = afr_zerofill_wind;
+        local->transaction.fop    = __afr_txn_write_fop;
+        local->transaction.done   = __afr_txn_write_done;
         local->transaction.unwind = afr_zerofill_unwind;
 
         local->transaction.main_frame = frame;
 
-        local->transaction.start   = local->cont.zerofill.offset;
-        local->transaction.len     = 0;
-
-        op_ret = afr_transaction (transaction_frame, this,
-                                  AFR_DATA_TRANSACTION);
-        if (op_ret < 0) {
-                op_errno = -op_ret;
-                goto out;
-        }
-
-        op_ret = 0;
-out:
-        if (op_ret < 0) {
-                if (transaction_frame) {
-                        AFR_STACK_DESTROY (transaction_frame);
-                }
-                AFR_STACK_UNWIND (zerofill, frame, op_ret, op_errno, NULL,
-                                  NULL, NULL);
-        }
-
-        return 0;
-}
-
-int
-afr_zerofill (call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset,
-              off_t len, dict_t *xdata)
-{
-        afr_private_t   *priv               = NULL;
-        afr_local_t     *local              = NULL;
-        call_frame_t    *transaction_frame  = NULL;
-        int              ret                = -1;
-        int              op_errno           = 0;
-
-        VALIDATE_OR_GOTO (frame, out);
-        VALIDATE_OR_GOTO (this, out);
-        VALIDATE_OR_GOTO (this->private, out);
-
-        priv = this->private;
-
-        if (afr_is_split_brain (this, fd->inode)) {
-                op_errno = EIO;
-                goto out;
-        }
-        QUORUM_CHECK(zerofill, out);
+        local->transaction.start   = local->cont.discard.offset;
+        local->transaction.len     = len;
 
-        AFR_LOCAL_ALLOC_OR_GOTO (frame->local, out);
-        local = frame->local;
+        afr_fix_open (fd, this);
 
-        ret = afr_local_init (local, priv, &op_errno);
+        ret = afr_transaction (transaction_frame, this, AFR_DATA_TRANSACTION);
         if (ret < 0) {
-                goto out;
+		op_errno = -ret;
+		goto out;
         }
-        local->cont.zerofill.offset  = offset;
-        local->cont.zerofill.len = len;
-
-        local->fd = fd_ref (fd);
-
-        afr_open_fd_fix (fd, this);
 
-        afr_do_zerofill(frame, this);
-
-        ret = 0;
+	return 0;
 out:
-        if (ret < 0) {
-                if (transaction_frame) {
-                        AFR_STACK_DESTROY (transaction_frame);
-                }
-                AFR_STACK_UNWIND (zerofill, frame, -1, op_errno, NULL,
-                                  NULL, NULL);
-        }
+	if (transaction_frame)
+		AFR_STACK_DESTROY (transaction_frame);
 
+	AFR_STACK_UNWIND (zerofill, frame, -1, op_errno, NULL, NULL, NULL);
         return 0;
 }