cluster/stripe: implement the coalesce stripe file format

The coalesce file format for cluster/stripe condenses the striped
files to a contiguous layout. The elimination of holes in striped
files eliminates space wasted via local filesystem preallocation
heuristics and significantly improves read performance.

Coalesce mode is implemented with a new 'coalesce' xlator option,
which is user-configurable and disabled by default. The format of
newly created files is marked with a new 'stripe-coalesce' xattr.
Cluster/stripe handles/preserves the format of files regardless
of the current mode of operation (i.e., a volume can
simultaneously consist of coalesced and non-coalesced files).
Files without the stripe-coalesce attribute are assumed to have
the traditional format to provide backward compatibility.

extras/stripe-merge: support traditional and coalesce stripe formats

Update the stripe-merge recovery tool to handle the traditional
and coalesced file formats. The format of the file is detected
automatically (and verified) via the stripe-coalesce attributes.

BUG: 801887
Change-Id: I682f0b4e819f496ddb68c9a01c4de4688280fdf8
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-on: http://review.gluster.com/3282
Tested-by: Gluster Build System <jenkins@build.gluster.com>
Reviewed-by: Amar Tumballi <amarts@redhat.com>
Reviewed-by: Anand Avati <avati@redhat.com>

5 files changed, 409 insertions, 79 deletions

diff --git a/xlators/cluster/stripe/src/stripe-helpers.c b/xlators/cluster/stripe/src/stripe-helpers.c
index 1821832c2..336da793e 100644
--- a/xlators/cluster/stripe/src/stripe-helpers.c
+++ b/xlators/cluster/stripe/src/stripe-helpers.c
@@ -236,8 +236,6 @@ out:
         return block_size;
 }
 
-
-
 int32_t
 stripe_ctx_handle (xlator_t *this, call_frame_t *prev, stripe_local_t *local,
                    dict_t *dict)
@@ -246,7 +244,6 @@ stripe_ctx_handle (xlator_t *this, call_frame_t *prev, stripe_local_t *local,
         data_t         *data            = NULL;
         int32_t         index           = 0;
         stripe_private_t *priv          = NULL;
-        int32_t         ret             = -1;
 
         priv = this->private;
 
@@ -343,14 +340,31 @@ stripe_ctx_handle (xlator_t *this, call_frame_t *prev, stripe_local_t *local,
                 if (!local->fctx->xl_array[index])
                         local->fctx->xl_array[index] = prev->this;
         }
-        ret = 0;
+
+	sprintf(key, "trusted.%s.stripe-coalesce", this->name);
+	data = dict_get(dict, key);
+	if (!data) {
+		/*
+		 * The file was probably created prior to coalesce support.
+		 * Assume non-coalesce mode for this file to maintain backwards
+		 * compatibility.
+		 */
+		gf_log(this->name, GF_LOG_DEBUG, "missing stripe-coalesce "
+			"attr, assume non-coalesce mode");
+		local->fctx->stripe_coalesce = 0;
+	} else {
+		local->fctx->stripe_coalesce = data_to_int32(data);
+	}
+
+
 out:
-        return ret;
+        return 0;
 }
 
 int32_t
 stripe_xattr_request_build (xlator_t *this, dict_t *dict, uint64_t stripe_size,
-                            uint32_t stripe_count, uint32_t stripe_index)
+                            uint32_t stripe_count, uint32_t stripe_index,
+			    uint32_t stripe_coalesce)
 {
         char            key[256]       = {0,};
         int32_t         ret             = -1;
@@ -378,6 +392,14 @@ stripe_xattr_request_build (xlator_t *this, dict_t *dict, uint64_t stripe_size,
                         "failed to set %s in xattr_req dict", key);
                 goto out;
         }
+
+	sprintf(key, "trusted.%s.stripe-coalesce", this->name);
+	ret = dict_set_int32(dict, key, stripe_coalesce);
+	if (ret) {
+		gf_log(this->name, GF_LOG_WARNING,
+			"failed to set %s in xattr_req_dict", key);
+		goto out;
+	}
 out:
         return ret;
 }
@@ -508,3 +530,61 @@ stripe_iatt_merge (struct iatt *from, struct iatt *to)
                 to->ia_atime = from->ia_atime;
         return 0;
 }
+
+off_t
+coalesced_offset(off_t offset, uint64_t stripe_size, int stripe_count)
+{
+	size_t line_size = 0;
+	uint64_t stripe_num = 0;
+	off_t coalesced_offset = 0;
+
+	line_size = stripe_size * stripe_count;
+	stripe_num = offset / line_size;
+
+	coalesced_offset = (stripe_num * stripe_size) +
+		(offset % stripe_size);
+
+	return coalesced_offset;
+}
+
+off_t
+uncoalesced_size(off_t size, uint64_t stripe_size, int stripe_count,
+		int stripe_index)
+{
+        uint64_t nr_full_stripe_chunks = 0, mod = 0;
+
+        if (!size)
+                return size;
+
+	/*
+	 * Estimate the number of fully written stripes from the
+	 * local file size. Each stripe_size chunk corresponds to
+	 * a stripe.
+	 */
+        nr_full_stripe_chunks = (size / stripe_size) * stripe_count;
+        mod = size % stripe_size;
+
+        if (!mod) {
+                /*
+		 * There is no remainder, thus we could have overestimated
+		 * the size of the file in terms of chunks. Trim the number
+		 * of chunks by the following stripe members and leave it
+		 * up to those nodes to respond with a larger size (if
+		 * necessary).
+		 */
+                nr_full_stripe_chunks -= stripe_count -
+                        (stripe_index + 1);
+                size = nr_full_stripe_chunks * stripe_size;
+        } else {
+		/*
+		 * There is a remainder and thus we own the last chunk of the
+		 * file. Add the preceding stripe members of the final stripe
+		 * along with the remainder to calculate the exact size.
+		 */
+                nr_full_stripe_chunks += stripe_index;
+                size = nr_full_stripe_chunks * stripe_size + mod;
+        }
+
+        return size;
+}
+
diff --git a/xlators/cluster/stripe/src/stripe.c b/xlators/cluster/stripe/src/stripe.c
index a98e14e95..efee9444e 100644
--- a/xlators/cluster/stripe/src/stripe.c
+++ b/xlators/cluster/stripe/src/stripe.c
@@ -32,7 +32,6 @@
 
 struct volume_options options[];
 
-
 int32_t
 stripe_sh_chown_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                      int32_t op_ret, int32_t op_errno,
@@ -237,6 +236,8 @@ stripe_lookup_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                         local->stbuf_blocks      += buf->ia_blocks;
                         local->postparent_blocks += postparent->ia_blocks;
 
+			correct_file_size(buf, local->fctx, prev);
+
                         if (local->stbuf_size < buf->ia_size)
                                 local->stbuf_size = buf->ia_size;
                         if (local->postparent_size < postparent->ia_size)
@@ -326,9 +327,19 @@ stripe_lookup (call_frame_t *frame, xlator_t *this, loc_t *loc,
         /* get stripe-size xattr on lookup. This would be required for
          * open/read/write/pathinfo calls. Hence we send down the request
          * even when type == IA_INVAL */
+
+	/*
+	 * We aren't guaranteed to have xdata here. We need the format info for
+	 * the file, so allocate xdata if necessary.
+	 */
+	if (!xdata)
+		xdata = dict_new();
+	else
+		xdata = dict_ref(xdata);
+
         if (xdata && (IA_ISREG (loc->inode->ia_type) ||
             (loc->inode->ia_type == IA_INVAL))) {
-                ret = stripe_xattr_request_build (this, xdata, 8, 4, 4);
+                ret = stripe_xattr_request_build (this, xdata, 8, 4, 4, 0);
                 if (ret)
                         gf_log (this->name , GF_LOG_ERROR, "Failed to build"
                                 " xattr request for %s", loc->path);
@@ -344,6 +355,8 @@ stripe_lookup (call_frame_t *frame, xlator_t *this, loc_t *loc,
                 trav = trav->next;
         }
 
+	dict_unref(xdata);
+
         return 0;
 err:
         STRIPE_STACK_UNWIND (lookup, frame, -1, op_errno, NULL, NULL, NULL, NULL);
@@ -388,6 +401,9 @@ stripe_stat_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                         }
 
                         local->stbuf_blocks += buf->ia_blocks;
+
+			correct_file_size(buf, local->fctx, prev);
+
                         if (local->stbuf_size < buf->ia_size)
                                 local->stbuf_size = buf->ia_size;
                 }
@@ -416,6 +432,7 @@ stripe_stat (call_frame_t *frame, xlator_t *this, loc_t *loc, dict_t *xdata)
         xlator_list_t    *trav = NULL;
         stripe_local_t   *local = NULL;
         stripe_private_t *priv = NULL;
+	stripe_fd_ctx_t  *fctx = NULL;
         int32_t           op_errno = EINVAL;
 
         VALIDATE_OR_GOTO (frame, err);
@@ -442,6 +459,13 @@ stripe_stat (call_frame_t *frame, xlator_t *this, loc_t *loc, dict_t *xdata)
         frame->local = local;
         local->call_count = priv->child_count;
 
+	if (IA_ISREG(loc->inode->ia_type)) {
+		inode_ctx_get(loc->inode, this, (uint64_t *) &fctx);
+		if (!fctx)
+			goto err;
+		local->fctx = fctx;
+	}
+
         while (trav) {
                 STACK_WIND (frame, stripe_stat_cbk, trav->xlator,
                             trav->xlator->fops->stat, loc, NULL);
@@ -583,6 +607,9 @@ stripe_truncate_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                         local->prebuf_blocks  += prebuf->ia_blocks;
                         local->postbuf_blocks += postbuf->ia_blocks;
 
+			correct_file_size(prebuf, local->fctx, prev);
+			correct_file_size(postbuf, local->fctx, prev);
+
                         if (local->prebuf_size < prebuf->ia_size)
                                 local->prebuf_size = prebuf->ia_size;
 
@@ -614,10 +641,12 @@ out:
 int32_t
 stripe_truncate (call_frame_t *frame, xlator_t *this, loc_t *loc, off_t offset, dict_t *xdata)
 {
-        xlator_list_t    *trav = NULL;
         stripe_local_t   *local = NULL;
         stripe_private_t *priv = NULL;
+	stripe_fd_ctx_t  *fctx = NULL;
         int32_t           op_errno = EINVAL;
+	int		  i, eof_idx;
+	off_t		  dest_offset, tmp_offset;
 
         VALIDATE_OR_GOTO (frame, err);
         VALIDATE_OR_GOTO (this, err);
@@ -626,7 +655,6 @@ stripe_truncate (call_frame_t *frame, xlator_t *this, loc_t *loc, off_t offset,
         VALIDATE_OR_GOTO (loc->inode, err);
 
         priv = this->private;
-        trav = this->children;
 
         if (priv->first_child_down) {
                 op_errno = ENOTCONN;
@@ -643,11 +671,51 @@ stripe_truncate (call_frame_t *frame, xlator_t *this, loc_t *loc, off_t offset,
         frame->local = local;
         local->call_count = priv->child_count;
 
-        while (trav) {
-                STACK_WIND (frame, stripe_truncate_cbk, trav->xlator,
-                            trav->xlator->fops->truncate, loc, offset, NULL);
-                trav = trav->next;
-        }
+	inode_ctx_get(loc->inode, this, (uint64_t *) &fctx);
+	if (!fctx) {
+		gf_log(this->name, GF_LOG_ERROR, "no stripe context");
+		op_errno = EINVAL;
+		goto err;
+	}
+
+	local->fctx = fctx;
+	eof_idx = (offset / fctx->stripe_size) % fctx->stripe_count;
+
+	for (i = 0; i < fctx->stripe_count; i++) {
+		if (!fctx->xl_array[i]) {
+			gf_log(this->name, GF_LOG_ERROR,
+				"no xlator at index %d", i);
+			op_errno = EINVAL;
+			goto err;
+		}
+
+		if (fctx->stripe_coalesce) {
+			/*
+	 		 * The node that owns EOF is truncated to the exact
+	 		 * coalesced offset. Nodes prior to this index should
+	 		 * be rounded up to the size of the complete stripe,
+	 		 * while nodes after this index should be rounded down
+			 * to the size of the previous stripe.
+			 */
+			if (i < eof_idx)
+				tmp_offset = roof(offset, fctx->stripe_size *
+						fctx->stripe_count);
+			else if (i > eof_idx)
+				tmp_offset = floor(offset, fctx->stripe_size *
+						fctx->stripe_count);
+			else
+				tmp_offset = offset;
+
+			dest_offset = coalesced_offset(tmp_offset,
+					fctx->stripe_size, fctx->stripe_count);
+		} else {
+			dest_offset = offset;
+		}
+
+		STACK_WIND(frame, stripe_truncate_cbk, fctx->xl_array[i],
+			fctx->xl_array[i]->fops->truncate, loc, dest_offset,
+			NULL);
+	}
 
         return 0;
 err:
@@ -698,6 +766,9 @@ stripe_setattr_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                         local->prebuf_blocks  += preop->ia_blocks;
                         local->postbuf_blocks += postop->ia_blocks;
 
+			correct_file_size(preop, local->fctx, prev);
+			correct_file_size(postop, local->fctx, prev);
+
                         if (local->prebuf_size < preop->ia_size)
                                 local->prebuf_size = preop->ia_size;
                         if (local->postbuf_size < postop->ia_size)
@@ -733,6 +804,7 @@ stripe_setattr (call_frame_t *frame, xlator_t *this, loc_t *loc,
         xlator_list_t    *trav = NULL;
         stripe_local_t   *local = NULL;
         stripe_private_t *priv = NULL;
+	stripe_fd_ctx_t	 *fctx = NULL;
         int32_t           op_errno = EINVAL;
 
         VALIDATE_OR_GOTO (frame, err);
@@ -766,6 +838,13 @@ stripe_setattr (call_frame_t *frame, xlator_t *this, loc_t *loc,
                 return 0;
         }
 
+	if (IA_ISREG(loc->inode->ia_type)) {
+		inode_ctx_get(loc->inode, this, (uint64_t *) &fctx);
+		if (!fctx)
+			goto err;
+		local->fctx = fctx;
+	}
+
         local->call_count = priv->child_count;
         while (trav) {
                 STACK_WIND (frame, stripe_setattr_cbk,
@@ -862,6 +941,8 @@ stripe_stack_rename_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                         local->pre_buf.ia_blocks    += prenewparent->ia_blocks;
                         local->post_buf.ia_blocks   += postnewparent->ia_blocks;
 
+			correct_file_size(buf, local->fctx, prev);
+
                         if (local->stbuf.ia_size < buf->ia_size)
                                 local->stbuf.ia_size =  buf->ia_size;
 
@@ -947,6 +1028,7 @@ stripe_rename (call_frame_t *frame, xlator_t *this, loc_t *oldloc,
         stripe_private_t *priv = NULL;
         stripe_local_t   *local = NULL;
         xlator_list_t    *trav = NULL;
+	stripe_fd_ctx_t	 *fctx = NULL;
         int32_t           op_errno = EINVAL;
 
         VALIDATE_OR_GOTO (frame, err);
@@ -977,6 +1059,11 @@ stripe_rename (call_frame_t *frame, xlator_t *this, loc_t *oldloc,
 
         local->call_count = priv->child_count;
 
+	inode_ctx_get(oldloc->inode, this, (uint64_t *) &fctx);
+	if (!fctx)
+		goto err;
+	local->fctx = fctx;
+
         frame->local = local;
 
         STACK_WIND (frame, stripe_first_rename_cbk, trav->xlator,
@@ -1367,7 +1454,6 @@ stripe_mknod_ifreg_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
         stripe_private_t *priv = NULL;
         call_frame_t     *prev = NULL;
         xlator_list_t    *trav = NULL;
-        stripe_fd_ctx_t  *fctx = NULL;
 
         if (!this || !frame || !frame->local || !cookie) {
                 gf_log ("stripe", GF_LOG_DEBUG, "possible NULL deref");
@@ -1399,10 +1485,16 @@ stripe_mknod_ifreg_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                         if (uuid_is_null (local->ia_gfid))
                                 uuid_copy (local->ia_gfid, buf->ia_gfid);
 
+			if (stripe_ctx_handle(this, prev, local, xdata))
+				gf_log(this->name, GF_LOG_ERROR,
+					"Error getting fctx info from dict");
+
                         local->stbuf_blocks += buf->ia_blocks;
                         local->preparent_blocks  += preparent->ia_blocks;
                         local->postparent_blocks += postparent->ia_blocks;
 
+			correct_file_size(buf, local->fctx, prev);
+
                         if (local->stbuf_size < buf->ia_size)
                                 local->stbuf_size = buf->ia_size;
                         if (local->preparent_size < preparent->ia_size)
@@ -1441,23 +1533,10 @@ stripe_mknod_ifreg_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                         local->postparent.ia_size   = local->postparent_size;
                         local->stbuf.ia_size        = local->stbuf_size;
                         local->stbuf.ia_blocks      = local->stbuf_blocks;
-                        fctx = GF_CALLOC (1, sizeof (stripe_fd_ctx_t),
-                                          gf_stripe_mt_stripe_fd_ctx_t);
-                        if (!fctx) {
-                                local->op_ret = -1;
-                                local->op_errno = ENOMEM;
-                                goto unwind;
-                        }
-
-                        fctx->stripe_size  = local->stripe_size;
-                        fctx->stripe_count = priv->child_count;
-                        fctx->static_array = 1;
-                        fctx->xl_array = priv->xl_array;
                         inode_ctx_put (local->inode, this,
-                                       (uint64_t)(long)fctx);
+                                       (uint64_t)(long) local->fctx);
 
                 }
-unwind:
                 STRIPE_STACK_UNWIND (mknod, frame, local->op_ret, local->op_errno,
                                      local->inode, &local->stbuf,
                                      &local->preparent, &local->postparent, NULL);
@@ -1531,7 +1610,8 @@ stripe_mknod_first_ifreg_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
 
                         ret = stripe_xattr_request_build (this, dict,
                                                           local->stripe_size,
-                                                          priv->child_count, i);
+                                                          priv->child_count, i,
+							  priv->coalesce);
                         if (ret)
                                 gf_log (this->name, GF_LOG_ERROR,
                                         "Failed to build xattr request");
@@ -1579,9 +1659,6 @@ stripe_mknod (call_frame_t *frame, xlator_t *this, loc_t *loc, mode_t mode,
         stripe_local_t   *local          = NULL;
         int32_t           op_errno       = EINVAL;
         int32_t           i              = 0;
-        char              size_key[256]  = {0,};
-        char              index_key[256] = {0,};
-        char              count_key[256] = {0,};
         dict_t           *dict           = NULL;
         int               ret            = 0;
         int               need_unref     = 0;
@@ -1631,15 +1708,6 @@ stripe_mknod (call_frame_t *frame, xlator_t *this, loc_t *loc, mode_t mode,
                    be looked up */
                 local->call_count = priv->child_count;
 
-                /* Send a setxattr request to nodes where the
-                   files are created */
-                sprintf (size_key,
-                         "trusted.%s.stripe-size", this->name);
-                sprintf (count_key,
-                         "trusted.%s.stripe-count", this->name);
-                sprintf (index_key,
-                         "trusted.%s.stripe-index", this->name);
-
                 if (priv->xattr_supported) {
                         dict = dict_new ();
                         if (!dict) {
@@ -1653,7 +1721,7 @@ stripe_mknod (call_frame_t *frame, xlator_t *this, loc_t *loc, mode_t mode,
                         ret = stripe_xattr_request_build (this, dict,
                                                           local->stripe_size,
                                                           priv->child_count,
-                                                          i);
+                                                          i, priv->coalesce);
                         if (ret)
                                 gf_log (this->name, GF_LOG_ERROR,
                                         "failed to build xattr request");
@@ -1867,6 +1935,7 @@ stripe_link_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
         int32_t         callcnt = 0;
         stripe_local_t  *local   = NULL;
         call_frame_t    *prev = NULL;
+	stripe_fd_ctx_t *fctx = NULL;
 
         if (!this || !frame || !frame->local || !cookie) {
                 gf_log ("stripe", GF_LOG_DEBUG, "possible NULL deref");
@@ -1880,6 +1949,14 @@ stripe_link_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
         {
                 callcnt = --local->call_count;
 
+		inode_ctx_get(inode, this, (uint64_t *) &fctx);
+		if (!fctx) {
+			gf_log(this->name, GF_LOG_ERROR, "failed to get stripe "
+				"context");
+			op_ret = -1;
+			op_errno = EINVAL;
+		}
+
                 if (op_ret == -1) {
                         gf_log (this->name, GF_LOG_DEBUG,
                                 "%s returned error %s",
@@ -1903,6 +1980,8 @@ stripe_link_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                         local->preparent_blocks  += preparent->ia_blocks;
                         local->postparent_blocks += postparent->ia_blocks;
 
+			correct_file_size(buf, fctx, prev);
+
                         if (local->stbuf_size < buf->ia_size)
                                 local->stbuf_size = buf->ia_size;
                         if (local->preparent_size < preparent->ia_size)
@@ -2023,7 +2102,6 @@ stripe_create_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
         int32_t           callcnt = 0;
         stripe_local_t   *local = NULL;
         stripe_private_t *priv = NULL;
-        stripe_fd_ctx_t  *fctx = NULL;
         call_frame_t     *prev = NULL;
         xlator_list_t    *trav = NULL;
 
@@ -2049,12 +2127,21 @@ stripe_create_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                 }
 
                 if (op_ret >= 0) {
+			if (IA_ISREG(buf->ia_type)) {
+				if (stripe_ctx_handle(this, prev, local, xdata))
+					gf_log(this->name, GF_LOG_ERROR,
+						"Error getting fctx info from "
+						"dict");
+			}
+
                         local->op_ret = op_ret;
 
                         local->stbuf_blocks += buf->ia_blocks;
                         local->preparent_blocks  += preparent->ia_blocks;
                         local->postparent_blocks += postparent->ia_blocks;
 
+			correct_file_size(buf, local->fctx, prev);
+
                         if (local->stbuf_size < buf->ia_size)
                                 local->stbuf_size = buf->ia_size;
                         if (local->preparent_size < preparent->ia_size)
@@ -2092,23 +2179,13 @@ stripe_create_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                         local->stbuf.ia_size        = local->stbuf_size;
                         local->stbuf.ia_blocks      = local->stbuf_blocks;
 
-                        fctx = GF_CALLOC (1, sizeof (stripe_fd_ctx_t),
-                                          gf_stripe_mt_stripe_fd_ctx_t);
-                        if (!fctx) {
-                                local->op_ret = -1;
-                                local->op_errno = ENOMEM;
-                                goto unwind;
-                        }
-
-                        fctx->stripe_size  = local->stripe_size;
-                        fctx->stripe_count = priv->child_count;
-                        fctx->static_array = 1;
-                        fctx->xl_array = priv->xl_array;
-                        inode_ctx_put (local->inode, this,
-                                    (uint64_t)(long)fctx);
+			stripe_copy_xl_array(local->fctx->xl_array,
+				             priv->xl_array,
+					     local->fctx->stripe_count);
+			inode_ctx_put(local->inode, this,
+					(uint64_t) local->fctx);
                 }
 
-        unwind:
                 /* Create itself has failed.. so return
                    without setxattring */
                 STRIPE_STACK_UNWIND (create, frame, local->op_ret,
@@ -2214,14 +2291,14 @@ stripe_first_create_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                         ret = stripe_xattr_request_build (this, dict,
                                                           local->stripe_size,
                                                           priv->child_count,
-                                                          i);
+                                                          i, priv->coalesce);
                         if (ret)
                                 gf_log (this->name, GF_LOG_ERROR,
                                         "failed to build xattr request");
                 } else {
                         dict = local->xattr;
                 }
- 
+
                 STACK_WIND (frame, stripe_create_cbk, trav->xlator,
                             trav->xlator->fops->create, &local->loc,
                             local->flags, local->mode, local->umask, local->fd,
@@ -2310,7 +2387,7 @@ stripe_create (call_frame_t *frame, xlator_t *this, loc_t *loc,
                 ret = stripe_xattr_request_build (this, dict,
                                                   local->stripe_size,
                                                   priv->child_count,
-                                                  i);
+                                                  i, priv->coalesce);
                 if (ret)
                         gf_log (this->name, GF_LOG_ERROR,
                                 "failed to build xattr request");
@@ -2743,6 +2820,9 @@ stripe_fsync_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                         local->prebuf_blocks  += prebuf->ia_blocks;
                         local->postbuf_blocks += postbuf->ia_blocks;
 
+			correct_file_size(prebuf, local->fctx, prev);
+			correct_file_size(postbuf, local->fctx, prev);
+
                         if (local->prebuf_size < prebuf->ia_size)
                                 local->prebuf_size = prebuf->ia_size;
 
@@ -2777,6 +2857,7 @@ stripe_fsync (call_frame_t *frame, xlator_t *this, fd_t *fd, int32_t flags, dict
         stripe_local_t   *local = NULL;
         stripe_private_t *priv = NULL;
         xlator_list_t    *trav = NULL;
+	stripe_fd_ctx_t  *fctx = NULL;
         int32_t           op_errno = 1;
 
         VALIDATE_OR_GOTO (frame, err);
@@ -2793,6 +2874,14 @@ stripe_fsync (call_frame_t *frame, xlator_t *this, fd_t *fd, int32_t flags, dict
                 op_errno = ENOMEM;
                 goto err;
         }
+
+	inode_ctx_get(fd->inode, this, (uint64_t *) &fctx);
+	if (!fctx) {
+		op_errno = EINVAL;
+		goto err;
+	}
+	local->fctx = fctx;
+
         local->op_ret = -1;
         frame->local = local;
         local->call_count = priv->child_count;
@@ -2846,6 +2935,9 @@ stripe_fstat_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                                 local->stbuf = *buf;
 
                         local->stbuf_blocks += buf->ia_blocks;
+
+			correct_file_size(buf, local->fctx, prev);
+
                         if (local->stbuf_size < buf->ia_size)
                                 local->stbuf_size = buf->ia_size;
                 }
@@ -2877,6 +2969,7 @@ stripe_fstat (call_frame_t *frame,
         stripe_local_t   *local = NULL;
         stripe_private_t *priv = NULL;
         xlator_list_t    *trav = NULL;
+	stripe_fd_ctx_t  *fctx = NULL;
         int32_t           op_errno = 1;
 
         VALIDATE_OR_GOTO (frame, err);
@@ -2897,6 +2990,13 @@ stripe_fstat (call_frame_t *frame,
         frame->local = local;
         local->call_count = priv->child_count;
 
+	if (IA_ISREG(fd->inode->ia_type)) {
+		inode_ctx_get(fd->inode, this, (uint64_t *) &fctx);
+		if (!fctx)
+			goto err;
+		local->fctx = fctx;
+	}
+
         while (trav) {
                 STACK_WIND (frame, stripe_fstat_cbk, trav->xlator,
                             trav->xlator->fops->fstat, fd, NULL);
@@ -2915,8 +3015,10 @@ stripe_ftruncate (call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset, d
 {
         stripe_local_t   *local = NULL;
         stripe_private_t *priv = NULL;
-        xlator_list_t    *trav = NULL;
-        int32_t           op_errno = 1;
+	stripe_fd_ctx_t  *fctx = NULL;
+	int		  i, eof_idx;
+	off_t		  dest_offset, tmp_offset;
+        int32_t		  op_errno = 1;
 
         VALIDATE_OR_GOTO (frame, err);
         VALIDATE_OR_GOTO (this, err);
@@ -2924,7 +3026,6 @@ stripe_ftruncate (call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset, d
         VALIDATE_OR_GOTO (fd->inode, err);
 
         priv = this->private;
-        trav = this->children;
 
         /* Initialization */
         local = mem_get0 (this->local_pool);
@@ -2936,11 +3037,49 @@ stripe_ftruncate (call_frame_t *frame, xlator_t *this, fd_t *fd, off_t offset, d
         frame->local = local;
         local->call_count = priv->child_count;
 
-        while (trav) {
-                STACK_WIND (frame, stripe_truncate_cbk, trav->xlator,
-                            trav->xlator->fops->ftruncate, fd, offset, NULL);
-                trav = trav->next;
-        }
+	inode_ctx_get(fd->inode, this, (uint64_t *) &fctx);
+	if (!fctx) {
+		gf_log(this->name, GF_LOG_ERROR, "no stripe context");
+		op_errno = EINVAL;
+		goto err;
+	}
+	if (!fctx->stripe_count) {
+		gf_log(this->name, GF_LOG_ERROR, "no stripe count");
+		op_errno = EINVAL;
+		goto err;
+	}
+
+	local->fctx = fctx;
+	eof_idx = (offset / fctx->stripe_size) % fctx->stripe_count;
+
+	for (i = 0; i < fctx->stripe_count; i++) {
+		if (!fctx->xl_array[i]) {
+			gf_log(this->name, GF_LOG_ERROR, "no xlator at index "
+				"%d", i);
+			op_errno = EINVAL;
+			goto err;
+		}
+
+		if (fctx->stripe_coalesce) {
+			if (i < eof_idx)
+				tmp_offset = roof(offset, fctx->stripe_size *
+						fctx->stripe_count);
+			else if (i > eof_idx)
+				tmp_offset = floor(offset, fctx->stripe_size *
+						fctx->stripe_count);
+			else
+				tmp_offset = offset;
+
+			dest_offset = coalesced_offset(tmp_offset,
+				fctx->stripe_size, fctx->stripe_count);
+		} else {
+			dest_offset = offset;
+		}
+
+		STACK_WIND(frame, stripe_truncate_cbk, fctx->xl_array[i],
+			fctx->xl_array[i]->fops->ftruncate, fd, dest_offset,
+			NULL);
+	}
 
         return 0;
 err:
@@ -3045,6 +3184,7 @@ stripe_readv_fstat_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
         struct iatt     tmp_stbuf = {0,};
         struct iobref  *tmp_iobref = NULL;
         struct iobuf   *iobuf = NULL;
+	call_frame_t   *prev = NULL;
 
         if (!this || !frame || !frame->local) {
                 gf_log ("stripe", GF_LOG_DEBUG, "possible NULL deref");
@@ -3052,13 +3192,16 @@ stripe_readv_fstat_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
         }
 
         local = frame->local;
+	prev = cookie;
 
         LOCK (&frame->lock);
         {
                 callcnt = --local->call_count;
-                if (op_ret != -1)
+                if (op_ret != -1) {
+			correct_file_size(buf, local->fctx, prev);
                         if (local->stbuf_size < buf->ia_size)
                                 local->stbuf_size = buf->ia_size;
+		}
         }
         UNLOCK (&frame->lock);
 
@@ -3150,6 +3293,7 @@ stripe_readv_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
         struct iatt    *tmp_stbuf_p = NULL; //need it for a warning
         struct iobref  *tmp_iobref = NULL;
         stripe_fd_ctx_t  *fctx = NULL;
+	call_frame_t	*prev = NULL;
 
         if (!this || !frame || !frame->local || !cookie) {
                 gf_log ("stripe", GF_LOG_DEBUG, "possible NULL deref");
@@ -3158,6 +3302,7 @@ stripe_readv_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
 
         local  = frame->local;
         index  = local->node_index;
+	prev = cookie;
         mframe = local->orig_frame;
         if (!mframe)
                 goto out;
@@ -3177,6 +3322,9 @@ stripe_readv_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                         mlocal->replies[index].stbuf  = *stbuf;
                         mlocal->replies[index].count  = count;
                         mlocal->replies[index].vector = iov_dup (vector, count);
+
+			correct_file_size(stbuf, fctx, prev);
+
                         if (local->stbuf_size < stbuf->ia_size)
                                 local->stbuf_size = stbuf->ia_size;
                         local->stbuf_blocks += stbuf->ia_blocks;
@@ -3289,6 +3437,7 @@ stripe_readv (call_frame_t *frame, xlator_t *this, fd_t *fd,
         uint64_t          stripe_size = 0;
         off_t             rounded_start = 0;
         off_t             frame_offset = offset;
+	off_t		  dest_offset = 0;
         stripe_local_t   *local = NULL;
         call_frame_t     *rframe = NULL;
         stripe_local_t   *rlocal = NULL;
@@ -3361,9 +3510,16 @@ stripe_readv (call_frame_t *frame, xlator_t *this, fd_t *fd,
                 rlocal->readv_size = frame_size;
                 rframe->local = rlocal;
                 idx = (index % fctx->stripe_count);
+
+		if (fctx->stripe_coalesce)
+			dest_offset = coalesced_offset(frame_offset,
+				stripe_size, fctx->stripe_count);
+		else
+			dest_offset = frame_offset;
+
                 STACK_WIND (rframe, stripe_readv_cbk, fctx->xl_array[idx],
                             fctx->xl_array[idx]->fops->readv,
-                            fd, frame_size, frame_offset, flags, xdata);
+                            fd, frame_size, dest_offset, flags, xdata);
 
                 frame_offset += frame_size;
         }
@@ -3410,11 +3566,27 @@ stripe_writev_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                         local->op_ret += op_ret;
                         local->post_buf = *postbuf;
                         local->pre_buf = *prebuf;
+
+			local->prebuf_blocks  += prebuf->ia_blocks;
+			local->postbuf_blocks += postbuf->ia_blocks;
+
+			correct_file_size(prebuf, local->fctx, prev);
+			correct_file_size(postbuf, local->fctx, prev);
+
+			if (local->prebuf_size < prebuf->ia_size)
+				local->prebuf_size = prebuf->ia_size;
+			if (local->postbuf_size < postbuf->ia_size)
+				local->postbuf_size = postbuf->ia_size;
                 }
         }
         UNLOCK (&frame->lock);
 
         if ((callcnt == local->wind_count) && local->unwind) {
+		local->pre_buf.ia_size = local->prebuf_size;
+		local->pre_buf.ia_blocks = local->prebuf_blocks;
+		local->post_buf.ia_size = local->postbuf_size;
+		local->post_buf.ia_blocks = local->postbuf_blocks;
+
                 STRIPE_STACK_UNWIND (writev, frame, local->op_ret,
                                      local->op_errno, &local->pre_buf,
                                      &local->post_buf, NULL);
@@ -3440,6 +3612,7 @@ stripe_writev (call_frame_t *frame, xlator_t *this, fd_t *fd,
         off_t             fill_size = 0;
         uint64_t          stripe_size = 0;
         uint64_t          tmp_fctx = 0;
+	off_t		  dest_offset = 0;
 
         VALIDATE_OR_GOTO (frame, err);
         VALIDATE_OR_GOTO (this, err);
@@ -3469,6 +3642,7 @@ stripe_writev (call_frame_t *frame, xlator_t *this, fd_t *fd,
         }
         frame->local = local;
         local->stripe_size = stripe_size;
+	local->fctx = fctx;
 
         if (!stripe_size) {
                 gf_log (this->name, GF_LOG_DEBUG,
@@ -3505,9 +3679,15 @@ stripe_writev (call_frame_t *frame, xlator_t *this, fd_t *fd,
                 if (remaining_size == 0)
                         local->unwind = 1;
 
+		if (fctx->stripe_coalesce)
+			dest_offset = coalesced_offset(offset + offset_offset,
+				local->stripe_size, fctx->stripe_count);
+		else
+			dest_offset = offset + offset_offset;
+
                 STACK_WIND (frame, stripe_writev_cbk, fctx->xl_array[idx],
                             fctx->xl_array[idx]->fops->writev, fd, tmp_vec,
-                            tmp_count, offset + offset_offset, flags, iobref,
+                            tmp_count, dest_offset, flags, iobref,
                             xdata);
 
                 GF_FREE (tmp_vec);
@@ -3859,10 +4039,15 @@ stripe_readdirp_lookup_cbk (call_frame_t *frame, void *cookie,
                         local->op_ret = op_ret;
                         goto unlock;
                 }
+
+		if (stripe_ctx_handle(this, prev, local, xattr))
+			gf_log(this->name, GF_LOG_ERROR,
+				"Error getting fctx info from dict.");
+
+		correct_file_size(stbuf, local->fctx, prev);
+
                 stripe_iatt_merge (stbuf, &entry->d_stat);
                 local->stbuf_blocks += stbuf->ia_blocks;
-
-                stripe_ctx_handle (this, prev, local, xattr);
         }
 unlock:
         UNLOCK(&frame->lock);
@@ -3957,7 +4142,7 @@ unlock:
 
         xattrs = dict_new ();
         if (xattrs)
-                (void) stripe_xattr_request_build (this, xattrs, 0, 0, 0);
+                (void) stripe_xattr_request_build (this, xattrs, 0, 0, 0, 0);
         count = op_ret;
         list_for_each_entry_safe (local_entry, tmp_entry,
                                   (&local->entries.list), list) {
@@ -4165,6 +4350,9 @@ reconfigure (xlator_t *this, dict_t *options)
                                 goto unlock;
                         }
                 }
+
+		GF_OPTION_RECONF("coalesce", priv->coalesce, options, bool,
+				unlock);
         }
  unlock:
         UNLOCK (&priv->lock);
@@ -4285,6 +4473,8 @@ init (xlator_t *this)
         /* notify related */
         priv->nodes_down = priv->child_count;
 
+	GF_OPTION_INIT("coalesce", priv->coalesce, bool, out);
+
         this->local_pool = mem_pool_new (stripe_local_t, 128);
         if (!this->local_pool) {
                 ret = -1;
@@ -4768,5 +4958,12 @@ struct volume_options options[] = {
           .type = GF_OPTION_TYPE_BOOL,
           .default_value = "true"
         },
+	{ .key = {"coalesce"},
+	  .type = GF_OPTION_TYPE_BOOL,
+	  .default_value = "false",
+	  .description = "Enable coalesce mode to flatten striped files as "
+			 "stored on the server (i.e., eliminate holes caused "
+			 "by the traditional format)."
+	},
         { .key  = {NULL} },
 };
diff --git a/xlators/cluster/stripe/src/stripe.h b/xlators/cluster/stripe/src/stripe.h
index cb05eb56f..1b9e660c1 100644
--- a/xlators/cluster/stripe/src/stripe.h
+++ b/xlators/cluster/stripe/src/stripe.h
@@ -101,6 +101,7 @@ struct stripe_private {
         int8_t                  child_count;
         int8_t                 *state; /* Current state of child node */
         gf_boolean_t            xattr_supported;  /* default yes */
+	gf_boolean_t		coalesce;
         char                    vol_uuid[UUID_SIZE + 1];
 };
 
@@ -119,6 +120,7 @@ struct readv_replies {
 typedef struct _stripe_fd_ctx {
         off_t      stripe_size;
         int        stripe_count;
+	int	   stripe_coalesce;
         int        static_array;
         xlator_t **xl_array;
 } stripe_fd_ctx_t;
@@ -214,13 +216,41 @@ struct stripe_local {
 typedef struct stripe_local   stripe_local_t;
 typedef struct stripe_private stripe_private_t;
 
+/*
+ * Determine the stripe index of a particular frame based on the translator.
+ */
+static inline int32_t stripe_get_frame_index(stripe_fd_ctx_t *fctx,
+					     call_frame_t *prev)
+{
+	int32_t i, idx = -1;
+
+	for (i = 0; i < fctx->stripe_count; i++) {
+		if (fctx->xl_array[i] == prev->this) {
+			idx = i;
+			break;
+		}
+	}
+
+	return idx;
+}
+
+static inline void stripe_copy_xl_array(xlator_t **dst, xlator_t **src,
+					int count)
+{
+	int i;
+
+	for (i = 0; i < count; i++)
+		dst[i] = src[i];
+}
+
 void stripe_local_wipe (stripe_local_t *local);
 int32_t stripe_ctx_handle (xlator_t *this, call_frame_t *prev,
                            stripe_local_t *local, dict_t *dict);
 void stripe_aggregate_xattr (dict_t *dst, dict_t *src);
 int32_t stripe_xattr_request_build (xlator_t *this, dict_t *dict,
                                     uint64_t stripe_size, uint32_t stripe_count,
-                                    uint32_t stripe_index);
+                                    uint32_t stripe_index,
+				    uint32_t stripe_coalesce);
 int32_t stripe_get_matching_bs (const char *path, stripe_private_t *priv);
 int set_stripe_block_size (xlator_t *this, stripe_private_t *priv, char *data);
 int32_t stripe_iatt_merge (struct iatt *from, struct iatt *to);
@@ -229,5 +259,27 @@ int32_t stripe_fill_pathinfo_xattr (xlator_t *this, stripe_local_t *local,
 int32_t stripe_free_xattr_str (stripe_local_t *local);
 int32_t stripe_xattr_aggregate (char *buffer, stripe_local_t *local,
                                 int32_t *total);
+off_t coalesced_offset(off_t offset, uint64_t stripe_size, int stripe_count);
+off_t uncoalesced_size(off_t size, uint64_t stripe_size, int stripe_count,
+			int stripe_index);
+
+/*
+ * Adjust the size attribute for files if coalesce is enabled.
+ */
+static inline void correct_file_size(struct iatt *buf, stripe_fd_ctx_t *fctx,
+	call_frame_t *prev)
+{
+	int index;
+
+	if (!IA_ISREG(buf->ia_type))
+		return;
+
+	if (!fctx || !fctx->stripe_coalesce)
+		return;
+
+	index = stripe_get_frame_index(fctx, prev);
+	buf->ia_size = uncoalesced_size(buf->ia_size, fctx->stripe_size,
+		fctx->stripe_count, index);
+}
 
 #endif /* _STRIPE_H_ */
diff --git a/xlators/mgmt/glusterd/src/glusterd-volgen.c b/xlators/mgmt/glusterd/src/glusterd-volgen.c
index 2a97e4052..a01d2a773 100644
--- a/xlators/mgmt/glusterd/src/glusterd-volgen.c
+++ b/xlators/mgmt/glusterd/src/glusterd-volgen.c
@@ -141,6 +141,7 @@ static struct volopt_map_entry glusterd_volopt_map[] = {
         {"cluster.choose-local",                 "cluster/replicate",  NULL, NULL, DOC, 0},
 
         {"cluster.stripe-block-size",            "cluster/stripe",     "block-size", NULL, DOC, 0},
+	{"cluster.stripe-coalesce",		 "cluster/stripe",     "coalesce", NULL, DOC, 0},
 
         {VKEY_DIAG_LAT_MEASUREMENT,              "debug/io-stats",     "latency-measurement", "off", NO_DOC, 0},
         {"diagnostics.dump-fd-stats",            "debug/io-stats",     NULL, NULL, NO_DOC, 0},
diff --git a/xlators/storage/posix/src/posix.c b/xlators/storage/posix/src/posix.c
index e18654cc7..46a83ea2f 100644
--- a/xlators/storage/posix/src/posix.c
+++ b/xlators/storage/posix/src/posix.c
@@ -1810,7 +1810,7 @@ out:
 
         STACK_UNWIND_STRICT (create, frame, op_ret, op_errno,
                              fd, (loc)?loc->inode:NULL, &stbuf, &preparent,
-                             &postparent, NULL);
+                             &postparent, xdata);
 
         return 0;
 }