cluster/dht - rebalance: handle the rebalance @ inode level (!fd level)

* migrate all the fd's on an inode to newer subvol after rebalance
* use the migration in progress flag in inode, so all the operations
  on the inode can make use of it

Change-Id: Ib807a46e927a1062688fc15119c916797c52a350
BUG: 1013456
Signed-off-by: Amar Tumballi <amarts@redhat.com>
Reviewed-on: http://review.gluster.org/5891
Tested-by: Gluster Build System <jenkins@build.gluster.com>
Reviewed-by: Anand Avati <avati@redhat.com>

4 files changed, 148 insertions, 128 deletions

diff --git a/xlators/cluster/dht/src/dht-common.h b/xlators/cluster/dht/src/dht-common.h
index 8c3449f0b3b..5ccd66799ae 100644
--- a/xlators/cluster/dht/src/dht-common.h
+++ b/xlators/cluster/dht/src/dht-common.h
@@ -781,4 +781,7 @@ dht_priv_dump (xlator_t *this);
 int32_t
 dht_inodectx_dump (xlator_t *this, inode_t *inode);
 
+int
+dht_inode_ctx_get1 (xlator_t *this, inode_t *inode, xlator_t **subvol);
+
 #endif/* _DHT_H */
diff --git a/xlators/cluster/dht/src/dht-helper.c b/xlators/cluster/dht/src/dht-helper.c
index 1cf343d24ac..311a481126d 100644
--- a/xlators/cluster/dht/src/dht-helper.c
+++ b/xlators/cluster/dht/src/dht-helper.c
@@ -18,6 +18,28 @@
 #include "xlator.h"
 #include "dht-common.h"
 
+static inline int
+dht_inode_ctx_set1 (xlator_t *this, inode_t *inode, xlator_t *subvol)
+{
+        uint64_t tmp_subvol = 0;
+
+        tmp_subvol = (long)subvol;
+        return inode_ctx_set1 (inode, this, &tmp_subvol);
+}
+
+int
+dht_inode_ctx_get1 (xlator_t *this, inode_t *inode, xlator_t **subvol)
+{
+        int ret = -1;
+        uint64_t tmp_subvol = 0;
+
+        ret =  inode_ctx_get1 (inode, this, &tmp_subvol);
+        if (tmp_subvol && subvol)
+                *subvol = (xlator_t *)tmp_subvol;
+
+        return ret;
+}
+
 
 int
 dht_frame_return (call_frame_t *frame)
@@ -340,20 +362,6 @@ out:
         return local;
 }
 
-
-char *
-basestr (const char *str)
-{
-        char *basestr = NULL;
-
-        basestr = strrchr (str, '/');
-        if (basestr)
-                basestr ++;
-
-        return basestr;
-}
-
-
 xlator_t *
 dht_first_up_subvol (xlator_t *this)
 {
@@ -727,6 +735,10 @@ dht_migration_complete_check_task (void *data)
         loc_t         tmp_loc  = {0,};
         char         *path     = NULL;
         dht_conf_t   *conf     = NULL;
+        inode_t      *inode    = NULL;
+        fd_t         *iter_fd  = NULL;
+        uint64_t      tmp_subvol = 0;
+        int           open_failed = 0;
 
         this  = THIS;
         frame = data;
@@ -738,6 +750,8 @@ dht_migration_complete_check_task (void *data)
         if (!local->loc.inode && !local->fd)
                 goto out;
 
+        inode = (!local->fd) ? local->loc.inode : local->fd->inode;
+
         /* getxattr on cached_subvol for 'linkto' value. Do path based getxattr
          * as root:root. If a fd is already open, access check wont be done*/
 
@@ -800,10 +814,7 @@ dht_migration_complete_check_task (void *data)
         /* update inode ctx (the layout) */
         dht_layout_unref (this, local->layout);
 
-        if (!local->loc.inode)
-                ret = dht_layout_preset (this, dst_node, local->fd->inode);
-        else
-                ret = dht_layout_preset (this, dst_node, local->loc.inode);
+        ret = dht_layout_preset (this, dst_node, inode);
         if (ret != 0) {
                 gf_log (this->name, GF_LOG_DEBUG,
                         "%s: could not set preset layout for subvol %s",
@@ -821,10 +832,7 @@ dht_migration_complete_check_task (void *data)
                 goto out;
         }
 
-        if (!local->loc.inode)
-                ret = dht_layout_set (this, local->fd->inode, layout);
-        else
-                ret = dht_layout_set (this, local->loc.inode, layout);
+        ret = dht_layout_set (this, inode, layout);
         if (ret) {
                 gf_log (this->name, GF_LOG_ERROR,
                         "%s: failed to set the new layout",
@@ -835,41 +843,46 @@ dht_migration_complete_check_task (void *data)
         local->cached_subvol = dst_node;
         ret = 0;
 
-        if (!local->fd)
+        /* once we detect the migration complete, the inode-ctx2 is no more
+           required.. delete the ctx and also, it means, open() already
+           done on all the fd of inode */
+        ret = inode_ctx_reset1 (inode, this, &tmp_subvol);
+        if (tmp_subvol)
                 goto out;
-        /* once we detect the migration complete, the fd-ctx is no more
-           required.. delete the ctx */
-        ret = fd_ctx_del (local->fd, this, NULL);
-        if (!ret)
+
+        if (list_empty (&inode->fd_list))
                 goto out;
 
         /* perform open as root:root. There is window between linkfile
          * creation(root:root) and setattr with the correct uid/gid
          */
         SYNCTASK_SETID(0, 0);
-        /* if 'local->fd' (ie, fd based operation), send a 'open()' on
-           destination if not already done */
-        if (local->loc.inode) {
-                ret = syncop_open (dst_node, &local->loc,
-                                   local->fd->flags, local->fd);
-        } else {
-                tmp_loc.inode = local->fd->inode;
-                inode_path (local->fd->inode, NULL, &path);
-                if (path)
-                        tmp_loc.path = path;
-                ret = syncop_open (dst_node, &tmp_loc,
-                                   local->fd->flags, local->fd);
-                GF_FREE (path);
 
+        /* perform 'open()' on all the fd's present on the inode */
+        tmp_loc.inode = inode;
+        inode_path (inode, NULL, &path);
+        if (path)
+                tmp_loc.path = path;
+        list_for_each_entry (iter_fd, &inode->fd_list, inode_list) {
+                if (fd_is_anonymous (iter_fd))
+                        continue;
+
+                ret = syncop_open (dst_node, &tmp_loc,
+                                   iter_fd->flags, iter_fd);
+                if (ret == -1) {
+                        gf_log (this->name, GF_LOG_ERROR, "failed to open "
+                                "the fd (%p, flags=0%o) on file %s @ %s",
+                                iter_fd, iter_fd->flags, path, dst_node->name);
+                        open_failed = 1;
+                }
         }
+        GF_FREE (path);
+
         SYNCTASK_SETID (frame->root->uid, frame->root->gid);
-        if (ret == -1) {
-                gf_log (this->name, GF_LOG_ERROR,
-                        "%s: failed to send open() on target file at %s",
-                        local->loc.path, dst_node->name);
+        if (open_failed) {
+                ret = -1;
                 goto out;
         }
-
         ret = 0;
 out:
 
@@ -914,6 +927,9 @@ dht_rebalance_inprogress_task (void *data)
         struct iatt   stbuf    = {0,};
         loc_t         tmp_loc  = {0,};
         dht_conf_t   *conf     = NULL;
+        inode_t      *inode    = NULL;
+        fd_t         *iter_fd  = NULL;
+        int           open_failed = 0;
 
         this  = THIS;
         frame = data;
@@ -925,6 +941,8 @@ dht_rebalance_inprogress_task (void *data)
         if (!local->loc.inode && !local->fd)
                 goto out;
 
+        inode = (!local->fd) ? local->loc.inode : local->fd->inode;
+
         /* getxattr on cached_subvol for 'linkto' value. Do path based getxattr
          * as root:root. If a fd is already open, access check wont be done*/
         if (local->loc.inode) {
@@ -976,35 +994,47 @@ dht_rebalance_inprogress_task (void *data)
         }
 
         ret = 0;
+
+        if (list_empty (&inode->fd_list))
+                goto done;
+
         /* perform open as root:root. There is window between linkfile
          * creation(root:root) and setattr with the correct uid/gid
          */
         SYNCTASK_SETID (0, 0);
-        if (local->loc.inode) {
-                ret = syncop_open (dst_node, &local->loc,
-                                   local->fd->flags, local->fd);
-        } else {
-                tmp_loc.inode = local->fd->inode;
-                inode_path (local->fd->inode, NULL, &path);
-                if (path)
-                        tmp_loc.path = path;
+
+        tmp_loc.inode = inode;
+        inode_path (inode, NULL, &path);
+        if (path)
+                tmp_loc.path = path;
+
+        list_for_each_entry (iter_fd, &inode->fd_list, inode_list) {
+                if (fd_is_anonymous (iter_fd))
+                        continue;
+
                 ret = syncop_open (dst_node, &tmp_loc,
-                                   local->fd->flags, local->fd);
-                GF_FREE (path);
+                                   iter_fd->flags, iter_fd);
+                if (ret == -1) {
+                        gf_log (this->name, GF_LOG_ERROR, "failed to send open "
+                                "the fd (%p, flags=0%o) on file %s @ %s",
+                                iter_fd, iter_fd->flags, path, dst_node->name);
+                        open_failed = 1;
+                }
         }
+        GF_FREE (path);
 
-        if (ret == -1) {
-                gf_log (this->name, GF_LOG_ERROR,
-                        "%s: failed to send open() on target file at %s",
-                        local->loc.path, dst_node->name);
+        SYNCTASK_SETID (frame->root->uid, frame->root->gid);
+
+        if (open_failed) {
+                ret = -1;
                 goto out;
         }
 
-        SYNCTASK_SETID (frame->root->uid, frame->root->gid);
-        ret = fd_ctx_set (local->fd, this, (uint64_t)(long)dst_node);
+done:
+        ret = dht_inode_ctx_set1 (this, inode, dst_node);
         if (ret) {
                 gf_log (this->name, GF_LOG_ERROR,
-                        "%s: failed to set fd-ctx target file at %s",
+                        "%s: failed to set inode-ctx target file at %s",
                         local->loc.path, dst_node->name);
                 goto out;
         }
diff --git a/xlators/cluster/dht/src/dht-inode-read.c b/xlators/cluster/dht/src/dht-inode-read.c
index 10f59ab26a1..ece84151adb 100644
--- a/xlators/cluster/dht/src/dht-inode-read.c
+++ b/xlators/cluster/dht/src/dht-inode-read.c
@@ -130,10 +130,11 @@ int
 dht_file_attr_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                    int op_ret, int op_errno, struct iatt *stbuf, dict_t *xdata)
 {
-        uint64_t      tmp_subvol = 0;
+        xlator_t     *subvol = 0;
         dht_local_t  *local = NULL;
         call_frame_t *prev = NULL;
         int           ret = -1;
+        inode_t      *inode = NULL;
 
         GF_VALIDATE_OR_GOTO ("dht", frame, err);
         GF_VALIDATE_OR_GOTO ("dht", this, out);
@@ -157,19 +158,20 @@ dht_file_attr_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
         local->op_errno = op_errno;
         /* Check if the rebalance phase2 is true */
         if ((op_ret == -1) || IS_DHT_MIGRATION_PHASE2 (stbuf)) {
-                if (local->fd)
-                        ret = fd_ctx_get (local->fd, this, &tmp_subvol);
-                if (ret) {
+                inode = (local->fd) ? local->fd->inode : local->loc.inode;
+                ret = dht_inode_ctx_get1 (this, inode, &subvol);
+                if (!subvol) {
                         /* Phase 2 of migration */
                         local->rebalance.target_op_fn = dht_attr2;
                         ret = dht_rebalance_complete_check (this, frame);
+                        if (!ret)
+                                return 0;
                 } else {
                         /* value is already set in fd_ctx, that means no need
                            to check for whether its complete or not. */
                         dht_attr2 (this, frame, 0);
-                }
-                if (!ret)
                         return 0;
+                }
         }
 
 out:
@@ -382,6 +384,8 @@ dht_readv_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
 {
         dht_local_t *local      = NULL;
         int          ret        = 0;
+        inode_t     *inode      = NULL;
+        xlator_t    *subvol = 0;
 
         local = frame->local;
         if (!local) {
@@ -400,17 +404,18 @@ dht_readv_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
         local->op_errno = op_errno;
         if ((op_ret == -1) || IS_DHT_MIGRATION_PHASE2 (stbuf)) {
                 /* File would be migrated to other node */
-                ret = fd_ctx_get (local->fd, this, NULL);
-                if (ret) {
+                ret = dht_inode_ctx_get1 (this, inode, &subvol);
+                if (!subvol) {
                         local->rebalance.target_op_fn = dht_readv2;
                         ret = dht_rebalance_complete_check (this, frame);
+                        if (!ret)
+                                return 0;
                 } else {
                         /* value is already set in fd_ctx, that means no need
                            to check for whether its complete or not. */
                         dht_readv2 (this, frame, 0);
-                }
-                if (!ret)
                         return 0;
+                }
         }
 
 out:
@@ -616,8 +621,9 @@ int
 dht_flush_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                int op_ret, int op_errno, dict_t *xdata)
 {
-        dht_local_t  *local = NULL;
-        int           ret = -1;
+        dht_local_t  *local  = NULL;
+        inode_t      *inode  = NULL;
+        xlator_t     *subvol = 0;
 
         local = frame->local;
 
@@ -627,8 +633,8 @@ dht_flush_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                 goto out;
 
         /* If context is set, then send flush() it to the destination */
-        ret = fd_ctx_get (local->fd, this, NULL);
-        if (!ret) {
+        dht_inode_ctx_get1 (this, inode, &subvol);
+        if (subvol) {
                 dht_flush2 (this, frame, 0);
                 return 0;
         }
@@ -644,14 +650,10 @@ dht_flush2 (xlator_t *this, call_frame_t *frame, int op_ret)
 {
         dht_local_t  *local  = NULL;
         xlator_t     *subvol = NULL;
-        uint64_t      tmp_subvol = 0;
-        int           ret = -1;
 
         local = frame->local;
 
-        ret = fd_ctx_get (local->fd, this, &tmp_subvol);
-        if (!ret)
-                subvol = (xlator_t *)(long)tmp_subvol;
+        dht_inode_ctx_get1 (this, local->fd->inode, &subvol);
 
         if (!subvol)
                 subvol = local->cached_subvol;
@@ -713,6 +715,8 @@ dht_fsync_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int op_ret,
         dht_local_t  *local = NULL;
         call_frame_t *prev = NULL;
         int           ret = -1;
+        inode_t      *inode = NULL;
+        xlator_t     *subvol = 0;
 
         local = frame->local;
         prev = cookie;
@@ -734,8 +738,8 @@ dht_fsync_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int op_ret,
         }
 
         local->op_errno = op_errno;
-        ret = fd_ctx_get (local->fd, this, NULL);
-        if (ret) {
+        dht_inode_ctx_get1 (this, inode, &subvol);
+        if (!subvol) {
                 local->rebalance.target_op_fn = dht_fsync2;
 
                 /* Check if the rebalance phase1 is true */
@@ -750,11 +754,12 @@ dht_fsync_cbk (call_frame_t *frame, void *cookie, xlator_t *this, int op_ret,
                 if (IS_DHT_MIGRATION_PHASE2 (postbuf)) {
                         ret = dht_rebalance_complete_check (this, frame);
                 }
+                if (!ret)
+                        return 0;
         } else {
                 dht_fsync2 (this, frame, 0);
-        }
-        if (!ret)
                 return 0;
+        }
 
 out:
         DHT_STRIP_PHASE1_FLAGS (postbuf);
@@ -770,15 +775,10 @@ dht_fsync2 (xlator_t *this, call_frame_t *frame, int op_ret)
 {
         dht_local_t  *local  = NULL;
         xlator_t     *subvol = NULL;
-        uint64_t      tmp_subvol = 0;
-        int           ret = -1;
 
         local = frame->local;
 
-        ret = fd_ctx_get (local->fd, this, &tmp_subvol);
-        if (!ret)
-                subvol = (xlator_t *)(long)tmp_subvol;
-
+        dht_inode_ctx_get1 (this, local->fd->inode, &subvol);
         if (!subvol)
                 subvol = local->cached_subvol;
 
diff --git a/xlators/cluster/dht/src/dht-inode-write.c b/xlators/cluster/dht/src/dht-inode-write.c
index 26db8a533c1..4b3f3a0496b 100644
--- a/xlators/cluster/dht/src/dht-inode-write.c
+++ b/xlators/cluster/dht/src/dht-inode-write.c
@@ -30,6 +30,7 @@ dht_writev_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
 {
         dht_local_t *local = NULL;
         int          ret   = -1;
+        xlator_t    *subvol = NULL;
 
         if (op_ret == -1 && (op_errno != ENOENT)) {
                 goto out;
@@ -66,8 +67,8 @@ dht_writev_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
                 dht_iatt_merge (this, &local->stbuf, postbuf, NULL);
                 dht_iatt_merge (this, &local->prebuf, prebuf, NULL);
 
-                ret = fd_ctx_get (local->fd, this, NULL);
-                if (!ret) {
+                ret = dht_inode_ctx_get1 (this, local->fd->inode, &subvol);
+                if (subvol) {
                         dht_writev2 (this, frame, 0);
                         return 0;
                 }
@@ -91,14 +92,10 @@ dht_writev2 (xlator_t *this, call_frame_t *frame, int op_ret)
 {
         dht_local_t  *local  = NULL;
         xlator_t     *subvol = NULL;
-        uint64_t      tmp_subvol = 0;
-        int           ret = -1;
 
         local = frame->local;
 
-        ret = fd_ctx_get (local->fd, this, &tmp_subvol);
-        if (!ret)
-                subvol = (xlator_t *)(long)tmp_subvol;
+        dht_inode_ctx_get1 (this, local->fd->inode, &subvol);
 
         if (!subvol)
                 subvol = local->cached_subvol;
@@ -173,6 +170,8 @@ dht_truncate_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
         dht_local_t  *local = NULL;
         call_frame_t *prev = NULL;
         int           ret = -1;
+        xlator_t    *subvol = NULL;
+        inode_t      *inode = NULL;
 
         GF_VALIDATE_OR_GOTO ("dht", frame, err);
         GF_VALIDATE_OR_GOTO ("dht", this, out);
@@ -214,8 +213,9 @@ dht_truncate_cbk (call_frame_t *frame, void *cookie, xlator_t *this,
         if (IS_DHT_MIGRATION_PHASE1 (postbuf)) {
                 dht_iatt_merge (this, &local->stbuf, postbuf, NULL);
                 dht_iatt_merge (this, &local->prebuf, prebuf, NULL);
-                ret = fd_ctx_get (local->fd, this, NULL);
-                if (!ret || !local->fd) {
+                inode = (local->fd) ? local->fd->inode : local->loc.inode;
+                dht_inode_ctx_get1 (this, inode, &subvol);
+                if (subvol) {
                         dht_truncate2 (this, frame, 0);
                         return 0;
                 }
@@ -239,16 +239,13 @@ dht_truncate2 (xlator_t *this, call_frame_t *frame, int op_ret)
 {
         dht_local_t  *local  = NULL;
         xlator_t     *subvol = NULL;
-        uint64_t      tmp_subvol = 0;
-        int           ret = -1;
+        inode_t      *inode = NULL;
 
         local = frame->local;
 
-        if (local->fd)
-                ret = fd_ctx_get (local->fd, this, &tmp_subvol);
-        if (!ret)
-                subvol = (xlator_t *)(long)tmp_subvol;
+        inode = local->fd ? local->fd->inode : local->loc.inode;
 
+        dht_inode_ctx_get1 (this, inode, &subvol);
         if (!subvol)
                 subvol = local->cached_subvol;
 
@@ -360,6 +357,7 @@ dht_fallocate_cbk(call_frame_t *frame, void *cookie, xlator_t *this,
         dht_local_t  *local = NULL;
         call_frame_t *prev = NULL;
         int           ret = -1;
+        xlator_t    *subvol = NULL;
 
         GF_VALIDATE_OR_GOTO ("dht", frame, err);
         GF_VALIDATE_OR_GOTO ("dht", this, out);
@@ -399,8 +397,8 @@ dht_fallocate_cbk(call_frame_t *frame, void *cookie, xlator_t *this,
         if (IS_DHT_MIGRATION_PHASE1 (postbuf)) {
                 dht_iatt_merge (this, &local->stbuf, postbuf, NULL);
                 dht_iatt_merge (this, &local->prebuf, prebuf, NULL);
-                ret = fd_ctx_get (local->fd, this, NULL);
-                if (!ret) {
+                dht_inode_ctx_get1 (this, local->fd->inode, &subvol);
+                if (subvol) {
                         dht_fallocate2 (this, frame, 0);
                         return 0;
                 }
@@ -423,15 +421,10 @@ dht_fallocate2(xlator_t *this, call_frame_t *frame, int op_ret)
 {
         dht_local_t  *local  = NULL;
         xlator_t     *subvol = NULL;
-        uint64_t      tmp_subvol = 0;
-        int           ret = -1;
 
         local = frame->local;
 
-        if (local->fd)
-                ret = fd_ctx_get (local->fd, this, &tmp_subvol);
-        if (!ret)
-                subvol = (xlator_t *)(long)tmp_subvol;
+        dht_inode_ctx_get1 (this, local->fd->inode, &subvol);
 
         if (!subvol)
                 subvol = local->cached_subvol;
@@ -498,6 +491,7 @@ dht_discard_cbk(call_frame_t *frame, void *cookie, xlator_t *this,
         dht_local_t  *local = NULL;
         call_frame_t *prev = NULL;
         int           ret = -1;
+        xlator_t    *subvol = NULL;
 
         GF_VALIDATE_OR_GOTO ("dht", frame, err);
         GF_VALIDATE_OR_GOTO ("dht", this, out);
@@ -537,8 +531,8 @@ dht_discard_cbk(call_frame_t *frame, void *cookie, xlator_t *this,
         if (IS_DHT_MIGRATION_PHASE1 (postbuf)) {
                 dht_iatt_merge (this, &local->stbuf, postbuf, NULL);
                 dht_iatt_merge (this, &local->prebuf, prebuf, NULL);
-                ret = fd_ctx_get (local->fd, this, NULL);
-                if (!ret) {
+                dht_inode_ctx_get1 (this, local->fd->inode, &subvol);
+                if (subvol) {
                         dht_discard2 (this, frame, 0);
                         return 0;
                 }
@@ -561,15 +555,10 @@ dht_discard2(xlator_t *this, call_frame_t *frame, int op_ret)
 {
         dht_local_t  *local  = NULL;
         xlator_t     *subvol = NULL;
-        uint64_t      tmp_subvol = 0;
-        int           ret = -1;
 
         local = frame->local;
 
-        if (local->fd)
-                ret = fd_ctx_get (local->fd, this, &tmp_subvol);
-        if (!ret)
-                subvol = (xlator_t *)(long)tmp_subvol;
+        dht_inode_ctx_get1 (this, local->fd->inode, &subvol);
 
         if (!subvol)
                 subvol = local->cached_subvol;
@@ -811,15 +800,13 @@ dht_setattr2 (xlator_t *this, call_frame_t *frame, int op_ret)
 {
         dht_local_t  *local  = NULL;
         xlator_t     *subvol = NULL;
-        uint64_t      tmp_subvol = 0;
-        int           ret = -1;
+        inode_t      *inode = NULL;
 
         local = frame->local;
 
-        if (local->fd)
-                ret = fd_ctx_get (local->fd, this, &tmp_subvol);
-        if (!ret)
-                subvol = (xlator_t *)(long)tmp_subvol;
+        inode = (local->fd) ? local->fd->inode : local->loc.inode;
+
+        dht_inode_ctx_get1 (this, inode, &subvol);
 
         if (!subvol)
                 subvol = local->cached_subvol;