Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / fs / ubifs / commit.c
diff --git a/kernel/fs/ubifs/commit.c b/kernel/fs/ubifs/commit.c
new file mode 100644 (file)
index 0000000..63f5661
--- /dev/null
@@ -0,0 +1,734 @@
+/*
+ * This file is part of UBIFS.
+ *
+ * Copyright (C) 2006-2008 Nokia Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 as published by
+ * the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program; if not, write to the Free Software Foundation, Inc., 51
+ * Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
+ *
+ * Authors: Adrian Hunter
+ *          Artem Bityutskiy (Битюцкий Артём)
+ */
+
+/*
+ * This file implements functions that manage the running of the commit process.
+ * Each affected module has its own functions to accomplish their part in the
+ * commit and those functions are called here.
+ *
+ * The commit is the process whereby all updates to the index and LEB properties
+ * are written out together and the journal becomes empty. This keeps the
+ * file system consistent - at all times the state can be recreated by reading
+ * the index and LEB properties and then replaying the journal.
+ *
+ * The commit is split into two parts named "commit start" and "commit end".
+ * During commit start, the commit process has exclusive access to the journal
+ * by holding the commit semaphore down for writing. As few I/O operations as
+ * possible are performed during commit start, instead the nodes that are to be
+ * written are merely identified. During commit end, the commit semaphore is no
+ * longer held and the journal is again in operation, allowing users to continue
+ * to use the file system while the bulk of the commit I/O is performed. The
+ * purpose of this two-step approach is to prevent the commit from causing any
+ * latency blips. Note that in any case, the commit does not prevent lookups
+ * (as permitted by the TNC mutex), or access to VFS data structures e.g. page
+ * cache.
+ */
+
+#include <linux/freezer.h>
+#include <linux/kthread.h>
+#include <linux/slab.h>
+#include "ubifs.h"
+
+/*
+ * nothing_to_commit - check if there is nothing to commit.
+ * @c: UBIFS file-system description object
+ *
+ * This is a helper function which checks if there is anything to commit. It is
+ * used as an optimization to avoid starting the commit if it is not really
+ * necessary. Indeed, the commit operation always assumes flash I/O (e.g.,
+ * writing the commit start node to the log), and it is better to avoid doing
+ * this unnecessarily. E.g., 'ubifs_sync_fs()' runs the commit, but if there is
+ * nothing to commit, it is more optimal to avoid any flash I/O.
+ *
+ * This function has to be called with @c->commit_sem locked for writing -
+ * this function does not take LPT/TNC locks because the @c->commit_sem
+ * guarantees that we have exclusive access to the TNC and LPT data structures.
+ *
+ * This function returns %1 if there is nothing to commit and %0 otherwise.
+ */
+static int nothing_to_commit(struct ubifs_info *c)
+{
+       /*
+        * During mounting or remounting from R/O mode to R/W mode we may
+        * commit for various recovery-related reasons.
+        */
+       if (c->mounting || c->remounting_rw)
+               return 0;
+
+       /*
+        * If the root TNC node is dirty, we definitely have something to
+        * commit.
+        */
+       if (c->zroot.znode && ubifs_zn_dirty(c->zroot.znode))
+               return 0;
+
+       /*
+        * Even though the TNC is clean, the LPT tree may have dirty nodes. For
+        * example, this may happen if the budgeting subsystem invoked GC to
+        * make some free space, and the GC found an LEB with only dirty and
+        * free space. In this case GC would just change the lprops of this
+        * LEB (by turning all space into free space) and unmap it.
+        */
+       if (c->nroot && test_bit(DIRTY_CNODE, &c->nroot->flags))
+               return 0;
+
+       ubifs_assert(atomic_long_read(&c->dirty_zn_cnt) == 0);
+       ubifs_assert(c->dirty_pn_cnt == 0);
+       ubifs_assert(c->dirty_nn_cnt == 0);
+
+       return 1;
+}
+
+/**
+ * do_commit - commit the journal.
+ * @c: UBIFS file-system description object
+ *
+ * This function implements UBIFS commit. It has to be called with commit lock
+ * locked. Returns zero in case of success and a negative error code in case of
+ * failure.
+ */
+static int do_commit(struct ubifs_info *c)
+{
+       int err, new_ltail_lnum, old_ltail_lnum, i;
+       struct ubifs_zbranch zroot;
+       struct ubifs_lp_stats lst;
+
+       dbg_cmt("start");
+       ubifs_assert(!c->ro_media && !c->ro_mount);
+
+       if (c->ro_error) {
+               err = -EROFS;
+               goto out_up;
+       }
+
+       if (nothing_to_commit(c)) {
+               up_write(&c->commit_sem);
+               err = 0;
+               goto out_cancel;
+       }
+
+       /* Sync all write buffers (necessary for recovery) */
+       for (i = 0; i < c->jhead_cnt; i++) {
+               err = ubifs_wbuf_sync(&c->jheads[i].wbuf);
+               if (err)
+                       goto out_up;
+       }
+
+       c->cmt_no += 1;
+       err = ubifs_gc_start_commit(c);
+       if (err)
+               goto out_up;
+       err = dbg_check_lprops(c);
+       if (err)
+               goto out_up;
+       err = ubifs_log_start_commit(c, &new_ltail_lnum);
+       if (err)
+               goto out_up;
+       err = ubifs_tnc_start_commit(c, &zroot);
+       if (err)
+               goto out_up;
+       err = ubifs_lpt_start_commit(c);
+       if (err)
+               goto out_up;
+       err = ubifs_orphan_start_commit(c);
+       if (err)
+               goto out_up;
+
+       ubifs_get_lp_stats(c, &lst);
+
+       up_write(&c->commit_sem);
+
+       err = ubifs_tnc_end_commit(c);
+       if (err)
+               goto out;
+       err = ubifs_lpt_end_commit(c);
+       if (err)
+               goto out;
+       err = ubifs_orphan_end_commit(c);
+       if (err)
+               goto out;
+       err = dbg_check_old_index(c, &zroot);
+       if (err)
+               goto out;
+
+       c->mst_node->cmt_no      = cpu_to_le64(c->cmt_no);
+       c->mst_node->log_lnum    = cpu_to_le32(new_ltail_lnum);
+       c->mst_node->root_lnum   = cpu_to_le32(zroot.lnum);
+       c->mst_node->root_offs   = cpu_to_le32(zroot.offs);
+       c->mst_node->root_len    = cpu_to_le32(zroot.len);
+       c->mst_node->ihead_lnum  = cpu_to_le32(c->ihead_lnum);
+       c->mst_node->ihead_offs  = cpu_to_le32(c->ihead_offs);
+       c->mst_node->index_size  = cpu_to_le64(c->bi.old_idx_sz);
+       c->mst_node->lpt_lnum    = cpu_to_le32(c->lpt_lnum);
+       c->mst_node->lpt_offs    = cpu_to_le32(c->lpt_offs);
+       c->mst_node->nhead_lnum  = cpu_to_le32(c->nhead_lnum);
+       c->mst_node->nhead_offs  = cpu_to_le32(c->nhead_offs);
+       c->mst_node->ltab_lnum   = cpu_to_le32(c->ltab_lnum);
+       c->mst_node->ltab_offs   = cpu_to_le32(c->ltab_offs);
+       c->mst_node->lsave_lnum  = cpu_to_le32(c->lsave_lnum);
+       c->mst_node->lsave_offs  = cpu_to_le32(c->lsave_offs);
+       c->mst_node->lscan_lnum  = cpu_to_le32(c->lscan_lnum);
+       c->mst_node->empty_lebs  = cpu_to_le32(lst.empty_lebs);
+       c->mst_node->idx_lebs    = cpu_to_le32(lst.idx_lebs);
+       c->mst_node->total_free  = cpu_to_le64(lst.total_free);
+       c->mst_node->total_dirty = cpu_to_le64(lst.total_dirty);
+       c->mst_node->total_used  = cpu_to_le64(lst.total_used);
+       c->mst_node->total_dead  = cpu_to_le64(lst.total_dead);
+       c->mst_node->total_dark  = cpu_to_le64(lst.total_dark);
+       if (c->no_orphs)
+               c->mst_node->flags |= cpu_to_le32(UBIFS_MST_NO_ORPHS);
+       else
+               c->mst_node->flags &= ~cpu_to_le32(UBIFS_MST_NO_ORPHS);
+
+       old_ltail_lnum = c->ltail_lnum;
+       err = ubifs_log_end_commit(c, new_ltail_lnum);
+       if (err)
+               goto out;
+
+       err = ubifs_log_post_commit(c, old_ltail_lnum);
+       if (err)
+               goto out;
+       err = ubifs_gc_end_commit(c);
+       if (err)
+               goto out;
+       err = ubifs_lpt_post_commit(c);
+       if (err)
+               goto out;
+
+out_cancel:
+       spin_lock(&c->cs_lock);
+       c->cmt_state = COMMIT_RESTING;
+       wake_up(&c->cmt_wq);
+       dbg_cmt("commit end");
+       spin_unlock(&c->cs_lock);
+       return 0;
+
+out_up:
+       up_write(&c->commit_sem);
+out:
+       ubifs_err(c, "commit failed, error %d", err);
+       spin_lock(&c->cs_lock);
+       c->cmt_state = COMMIT_BROKEN;
+       wake_up(&c->cmt_wq);
+       spin_unlock(&c->cs_lock);
+       ubifs_ro_mode(c, err);
+       return err;
+}
+
+/**
+ * run_bg_commit - run background commit if it is needed.
+ * @c: UBIFS file-system description object
+ *
+ * This function runs background commit if it is needed. Returns zero in case
+ * of success and a negative error code in case of failure.
+ */
+static int run_bg_commit(struct ubifs_info *c)
+{
+       spin_lock(&c->cs_lock);
+       /*
+        * Run background commit only if background commit was requested or if
+        * commit is required.
+        */
+       if (c->cmt_state != COMMIT_BACKGROUND &&
+           c->cmt_state != COMMIT_REQUIRED)
+               goto out;
+       spin_unlock(&c->cs_lock);
+
+       down_write(&c->commit_sem);
+       spin_lock(&c->cs_lock);
+       if (c->cmt_state == COMMIT_REQUIRED)
+               c->cmt_state = COMMIT_RUNNING_REQUIRED;
+       else if (c->cmt_state == COMMIT_BACKGROUND)
+               c->cmt_state = COMMIT_RUNNING_BACKGROUND;
+       else
+               goto out_cmt_unlock;
+       spin_unlock(&c->cs_lock);
+
+       return do_commit(c);
+
+out_cmt_unlock:
+       up_write(&c->commit_sem);
+out:
+       spin_unlock(&c->cs_lock);
+       return 0;
+}
+
+/**
+ * ubifs_bg_thread - UBIFS background thread function.
+ * @info: points to the file-system description object
+ *
+ * This function implements various file-system background activities:
+ * o when a write-buffer timer expires it synchronizes the appropriate
+ *   write-buffer;
+ * o when the journal is about to be full, it starts in-advance commit.
+ *
+ * Note, other stuff like background garbage collection may be added here in
+ * future.
+ */
+int ubifs_bg_thread(void *info)
+{
+       int err;
+       struct ubifs_info *c = info;
+
+       ubifs_msg(c, "background thread \"%s\" started, PID %d",
+                 c->bgt_name, current->pid);
+       set_freezable();
+
+       while (1) {
+               if (kthread_should_stop())
+                       break;
+
+               if (try_to_freeze())
+                       continue;
+
+               set_current_state(TASK_INTERRUPTIBLE);
+               /* Check if there is something to do */
+               if (!c->need_bgt) {
+                       /*
+                        * Nothing prevents us from going sleep now and
+                        * be never woken up and block the task which
+                        * could wait in 'kthread_stop()' forever.
+                        */
+                       if (kthread_should_stop())
+                               break;
+                       schedule();
+                       continue;
+               } else
+                       __set_current_state(TASK_RUNNING);
+
+               c->need_bgt = 0;
+               err = ubifs_bg_wbufs_sync(c);
+               if (err)
+                       ubifs_ro_mode(c, err);
+
+               run_bg_commit(c);
+               cond_resched();
+       }
+
+       ubifs_msg(c, "background thread \"%s\" stops", c->bgt_name);
+       return 0;
+}
+
+/**
+ * ubifs_commit_required - set commit state to "required".
+ * @c: UBIFS file-system description object
+ *
+ * This function is called if a commit is required but cannot be done from the
+ * calling function, so it is just flagged instead.
+ */
+void ubifs_commit_required(struct ubifs_info *c)
+{
+       spin_lock(&c->cs_lock);
+       switch (c->cmt_state) {
+       case COMMIT_RESTING:
+       case COMMIT_BACKGROUND:
+               dbg_cmt("old: %s, new: %s", dbg_cstate(c->cmt_state),
+                       dbg_cstate(COMMIT_REQUIRED));
+               c->cmt_state = COMMIT_REQUIRED;
+               break;
+       case COMMIT_RUNNING_BACKGROUND:
+               dbg_cmt("old: %s, new: %s", dbg_cstate(c->cmt_state),
+                       dbg_cstate(COMMIT_RUNNING_REQUIRED));
+               c->cmt_state = COMMIT_RUNNING_REQUIRED;
+               break;
+       case COMMIT_REQUIRED:
+       case COMMIT_RUNNING_REQUIRED:
+       case COMMIT_BROKEN:
+               break;
+       }
+       spin_unlock(&c->cs_lock);
+}
+
+/**
+ * ubifs_request_bg_commit - notify the background thread to do a commit.
+ * @c: UBIFS file-system description object
+ *
+ * This function is called if the journal is full enough to make a commit
+ * worthwhile, so background thread is kicked to start it.
+ */
+void ubifs_request_bg_commit(struct ubifs_info *c)
+{
+       spin_lock(&c->cs_lock);
+       if (c->cmt_state == COMMIT_RESTING) {
+               dbg_cmt("old: %s, new: %s", dbg_cstate(c->cmt_state),
+                       dbg_cstate(COMMIT_BACKGROUND));
+               c->cmt_state = COMMIT_BACKGROUND;
+               spin_unlock(&c->cs_lock);
+               ubifs_wake_up_bgt(c);
+       } else
+               spin_unlock(&c->cs_lock);
+}
+
+/**
+ * wait_for_commit - wait for commit.
+ * @c: UBIFS file-system description object
+ *
+ * This function sleeps until the commit operation is no longer running.
+ */
+static int wait_for_commit(struct ubifs_info *c)
+{
+       dbg_cmt("pid %d goes sleep", current->pid);
+
+       /*
+        * The following sleeps if the condition is false, and will be woken
+        * when the commit ends. It is possible, although very unlikely, that we
+        * will wake up and see the subsequent commit running, rather than the
+        * one we were waiting for, and go back to sleep.  However, we will be
+        * woken again, so there is no danger of sleeping forever.
+        */
+       wait_event(c->cmt_wq, c->cmt_state != COMMIT_RUNNING_BACKGROUND &&
+                             c->cmt_state != COMMIT_RUNNING_REQUIRED);
+       dbg_cmt("commit finished, pid %d woke up", current->pid);
+       return 0;
+}
+
+/**
+ * ubifs_run_commit - run or wait for commit.
+ * @c: UBIFS file-system description object
+ *
+ * This function runs commit and returns zero in case of success and a negative
+ * error code in case of failure.
+ */
+int ubifs_run_commit(struct ubifs_info *c)
+{
+       int err = 0;
+
+       spin_lock(&c->cs_lock);
+       if (c->cmt_state == COMMIT_BROKEN) {
+               err = -EROFS;
+               goto out;
+       }
+
+       if (c->cmt_state == COMMIT_RUNNING_BACKGROUND)
+               /*
+                * We set the commit state to 'running required' to indicate
+                * that we want it to complete as quickly as possible.
+                */
+               c->cmt_state = COMMIT_RUNNING_REQUIRED;
+
+       if (c->cmt_state == COMMIT_RUNNING_REQUIRED) {
+               spin_unlock(&c->cs_lock);
+               return wait_for_commit(c);
+       }
+       spin_unlock(&c->cs_lock);
+
+       /* Ok, the commit is indeed needed */
+
+       down_write(&c->commit_sem);
+       spin_lock(&c->cs_lock);
+       /*
+        * Since we unlocked 'c->cs_lock', the state may have changed, so
+        * re-check it.
+        */
+       if (c->cmt_state == COMMIT_BROKEN) {
+               err = -EROFS;
+               goto out_cmt_unlock;
+       }
+
+       if (c->cmt_state == COMMIT_RUNNING_BACKGROUND)
+               c->cmt_state = COMMIT_RUNNING_REQUIRED;
+
+       if (c->cmt_state == COMMIT_RUNNING_REQUIRED) {
+               up_write(&c->commit_sem);
+               spin_unlock(&c->cs_lock);
+               return wait_for_commit(c);
+       }
+       c->cmt_state = COMMIT_RUNNING_REQUIRED;
+       spin_unlock(&c->cs_lock);
+
+       err = do_commit(c);
+       return err;
+
+out_cmt_unlock:
+       up_write(&c->commit_sem);
+out:
+       spin_unlock(&c->cs_lock);
+       return err;
+}
+
+/**
+ * ubifs_gc_should_commit - determine if it is time for GC to run commit.
+ * @c: UBIFS file-system description object
+ *
+ * This function is called by garbage collection to determine if commit should
+ * be run. If commit state is @COMMIT_BACKGROUND, which means that the journal
+ * is full enough to start commit, this function returns true. It is not
+ * absolutely necessary to commit yet, but it feels like this should be better
+ * then to keep doing GC. This function returns %1 if GC has to initiate commit
+ * and %0 if not.
+ */
+int ubifs_gc_should_commit(struct ubifs_info *c)
+{
+       int ret = 0;
+
+       spin_lock(&c->cs_lock);
+       if (c->cmt_state == COMMIT_BACKGROUND) {
+               dbg_cmt("commit required now");
+               c->cmt_state = COMMIT_REQUIRED;
+       } else
+               dbg_cmt("commit not requested");
+       if (c->cmt_state == COMMIT_REQUIRED)
+               ret = 1;
+       spin_unlock(&c->cs_lock);
+       return ret;
+}
+
+/*
+ * Everything below is related to debugging.
+ */
+
+/**
+ * struct idx_node - hold index nodes during index tree traversal.
+ * @list: list
+ * @iip: index in parent (slot number of this indexing node in the parent
+ *       indexing node)
+ * @upper_key: all keys in this indexing node have to be less or equivalent to
+ *             this key
+ * @idx: index node (8-byte aligned because all node structures must be 8-byte
+ *       aligned)
+ */
+struct idx_node {
+       struct list_head list;
+       int iip;
+       union ubifs_key upper_key;
+       struct ubifs_idx_node idx __aligned(8);
+};
+
+/**
+ * dbg_old_index_check_init - get information for the next old index check.
+ * @c: UBIFS file-system description object
+ * @zroot: root of the index
+ *
+ * This function records information about the index that will be needed for the
+ * next old index check i.e. 'dbg_check_old_index()'.
+ *
+ * This function returns %0 on success and a negative error code on failure.
+ */
+int dbg_old_index_check_init(struct ubifs_info *c, struct ubifs_zbranch *zroot)
+{
+       struct ubifs_idx_node *idx;
+       int lnum, offs, len, err = 0;
+       struct ubifs_debug_info *d = c->dbg;
+
+       d->old_zroot = *zroot;
+       lnum = d->old_zroot.lnum;
+       offs = d->old_zroot.offs;
+       len = d->old_zroot.len;
+
+       idx = kmalloc(c->max_idx_node_sz, GFP_NOFS);
+       if (!idx)
+               return -ENOMEM;
+
+       err = ubifs_read_node(c, idx, UBIFS_IDX_NODE, len, lnum, offs);
+       if (err)
+               goto out;
+
+       d->old_zroot_level = le16_to_cpu(idx->level);
+       d->old_zroot_sqnum = le64_to_cpu(idx->ch.sqnum);
+out:
+       kfree(idx);
+       return err;
+}
+
+/**
+ * dbg_check_old_index - check the old copy of the index.
+ * @c: UBIFS file-system description object
+ * @zroot: root of the new index
+ *
+ * In order to be able to recover from an unclean unmount, a complete copy of
+ * the index must exist on flash. This is the "old" index. The commit process
+ * must write the "new" index to flash without overwriting or destroying any
+ * part of the old index. This function is run at commit end in order to check
+ * that the old index does indeed exist completely intact.
+ *
+ * This function returns %0 on success and a negative error code on failure.
+ */
+int dbg_check_old_index(struct ubifs_info *c, struct ubifs_zbranch *zroot)
+{
+       int lnum, offs, len, err = 0, uninitialized_var(last_level), child_cnt;
+       int first = 1, iip;
+       struct ubifs_debug_info *d = c->dbg;
+       union ubifs_key uninitialized_var(lower_key), upper_key, l_key, u_key;
+       unsigned long long uninitialized_var(last_sqnum);
+       struct ubifs_idx_node *idx;
+       struct list_head list;
+       struct idx_node *i;
+       size_t sz;
+
+       if (!dbg_is_chk_index(c))
+               return 0;
+
+       INIT_LIST_HEAD(&list);
+
+       sz = sizeof(struct idx_node) + ubifs_idx_node_sz(c, c->fanout) -
+            UBIFS_IDX_NODE_SZ;
+
+       /* Start at the old zroot */
+       lnum = d->old_zroot.lnum;
+       offs = d->old_zroot.offs;
+       len = d->old_zroot.len;
+       iip = 0;
+
+       /*
+        * Traverse the index tree preorder depth-first i.e. do a node and then
+        * its subtrees from left to right.
+        */
+       while (1) {
+               struct ubifs_branch *br;
+
+               /* Get the next index node */
+               i = kmalloc(sz, GFP_NOFS);
+               if (!i) {
+                       err = -ENOMEM;
+                       goto out_free;
+               }
+               i->iip = iip;
+               /* Keep the index nodes on our path in a linked list */
+               list_add_tail(&i->list, &list);
+               /* Read the index node */
+               idx = &i->idx;
+               err = ubifs_read_node(c, idx, UBIFS_IDX_NODE, len, lnum, offs);
+               if (err)
+                       goto out_free;
+               /* Validate index node */
+               child_cnt = le16_to_cpu(idx->child_cnt);
+               if (child_cnt < 1 || child_cnt > c->fanout) {
+                       err = 1;
+                       goto out_dump;
+               }
+               if (first) {
+                       first = 0;
+                       /* Check root level and sqnum */
+                       if (le16_to_cpu(idx->level) != d->old_zroot_level) {
+                               err = 2;
+                               goto out_dump;
+                       }
+                       if (le64_to_cpu(idx->ch.sqnum) != d->old_zroot_sqnum) {
+                               err = 3;
+                               goto out_dump;
+                       }
+                       /* Set last values as though root had a parent */
+                       last_level = le16_to_cpu(idx->level) + 1;
+                       last_sqnum = le64_to_cpu(idx->ch.sqnum) + 1;
+                       key_read(c, ubifs_idx_key(c, idx), &lower_key);
+                       highest_ino_key(c, &upper_key, INUM_WATERMARK);
+               }
+               key_copy(c, &upper_key, &i->upper_key);
+               if (le16_to_cpu(idx->level) != last_level - 1) {
+                       err = 3;
+                       goto out_dump;
+               }
+               /*
+                * The index is always written bottom up hence a child's sqnum
+                * is always less than the parents.
+                */
+               if (le64_to_cpu(idx->ch.sqnum) >= last_sqnum) {
+                       err = 4;
+                       goto out_dump;
+               }
+               /* Check key range */
+               key_read(c, ubifs_idx_key(c, idx), &l_key);
+               br = ubifs_idx_branch(c, idx, child_cnt - 1);
+               key_read(c, &br->key, &u_key);
+               if (keys_cmp(c, &lower_key, &l_key) > 0) {
+                       err = 5;
+                       goto out_dump;
+               }
+               if (keys_cmp(c, &upper_key, &u_key) < 0) {
+                       err = 6;
+                       goto out_dump;
+               }
+               if (keys_cmp(c, &upper_key, &u_key) == 0)
+                       if (!is_hash_key(c, &u_key)) {
+                               err = 7;
+                               goto out_dump;
+                       }
+               /* Go to next index node */
+               if (le16_to_cpu(idx->level) == 0) {
+                       /* At the bottom, so go up until can go right */
+                       while (1) {
+                               /* Drop the bottom of the list */
+                               list_del(&i->list);
+                               kfree(i);
+                               /* No more list means we are done */
+                               if (list_empty(&list))
+                                       goto out;
+                               /* Look at the new bottom */
+                               i = list_entry(list.prev, struct idx_node,
+                                              list);
+                               idx = &i->idx;
+                               /* Can we go right */
+                               if (iip + 1 < le16_to_cpu(idx->child_cnt)) {
+                                       iip = iip + 1;
+                                       break;
+                               } else
+                                       /* Nope, so go up again */
+                                       iip = i->iip;
+                       }
+               } else
+                       /* Go down left */
+                       iip = 0;
+               /*
+                * We have the parent in 'idx' and now we set up for reading the
+                * child pointed to by slot 'iip'.
+                */
+               last_level = le16_to_cpu(idx->level);
+               last_sqnum = le64_to_cpu(idx->ch.sqnum);
+               br = ubifs_idx_branch(c, idx, iip);
+               lnum = le32_to_cpu(br->lnum);
+               offs = le32_to_cpu(br->offs);
+               len = le32_to_cpu(br->len);
+               key_read(c, &br->key, &lower_key);
+               if (iip + 1 < le16_to_cpu(idx->child_cnt)) {
+                       br = ubifs_idx_branch(c, idx, iip + 1);
+                       key_read(c, &br->key, &upper_key);
+               } else
+                       key_copy(c, &i->upper_key, &upper_key);
+       }
+out:
+       err = dbg_old_index_check_init(c, zroot);
+       if (err)
+               goto out_free;
+
+       return 0;
+
+out_dump:
+       ubifs_err(c, "dumping index node (iip=%d)", i->iip);
+       ubifs_dump_node(c, idx);
+       list_del(&i->list);
+       kfree(i);
+       if (!list_empty(&list)) {
+               i = list_entry(list.prev, struct idx_node, list);
+               ubifs_err(c, "dumping parent index node");
+               ubifs_dump_node(c, &i->idx);
+       }
+out_free:
+       while (!list_empty(&list)) {
+               i = list_entry(list.next, struct idx_node, list);
+               list_del(&i->list);
+               kfree(i);
+       }
+       ubifs_err(c, "failed, error %d", err);
+       if (err > 0)
+               err = -EINVAL;
+       return err;
+}