ext4: online defrag -- Exchange the blocks between two inodes
From: Akira Fujita <[email protected]>
For each page, exchange the extents between original inode
and destination inode, and then write the file data of
the original inode to destination inode.
Signed-off-by: Akira Fujita <[email protected]>
Signed-off-by: Takashi Sato <[email protected]>
Signed-off-by: Kazuya Mio <[email protected]>
---
fs/ext4/defrag.c | 754 ++++++++++++++++++++++++++++++++++++++++++++++++++++
1 files changed, 753 insertions(+), 1 deletions(-)
diff --git a/fs/ext4/defrag.c b/fs/ext4/defrag.c
index c9f903e..a281ff8 100644
--- a/fs/ext4/defrag.c
+++ b/fs/ext4/defrag.c
@@ -94,12 +94,764 @@ err:
return -EIO;
}
+/**
+ * ext4_defrag_insert_across_blocks - Insert extents across leaf block
+ *
+ * @handle: journal handle
+ * @org_inode: original inode
+ * @o_start: first original extent to be changed
+ * @o_end: last original extent to be changed
+ * @start_ext: first new extent to be inserted
+ * @new_ext: middle of new extent to be inserted
+ * @end_ext: last new extent to be inserted
+ *
+ * Allocate a new leaf block and insert extents into it. Return 0 on success,
+ * or a negative error value on failure.
+ */
+static int
+ext4_defrag_insert_across_blocks(handle_t *handle, struct inode *org_inode,
+ struct ext4_extent *o_start, struct ext4_extent *o_end,
+ struct ext4_extent *start_ext, struct ext4_extent *new_ext,
+ struct ext4_extent *end_ext)
+{
+ struct ext4_ext_path *org_path = NULL;
+ ext4_lblk_t eblock = 0;
+ int new_flag = 0;
+ int end_flag = 0;
+ int err;
+
+ if (ext4_ext_get_actual_len(start_ext) &&
+ ext4_ext_get_actual_len(new_ext) &&
+ ext4_ext_get_actual_len(end_ext)) {
+
+ if (o_start == o_end) {
+
+ /* start_ext new_ext end_ext
+ * dest |---------|-----------|--------|
+ * org |------------------------------|
+ */
+
+ end_flag = 1;
+ } else {
+
+ /* start_ext new_ext end_ext
+ * dest |---------|----------|---------|
+ * org |---------------|--------------|
+ */
+
+ o_end->ee_block = end_ext->ee_block;
+ o_end->ee_len =
+ cpu_to_le16(ext4_ext_get_actual_len(end_ext));
+ ext4_ext_store_pblock(o_end, ext_pblock(end_ext));
+ }
+
+ o_start->ee_len =
+ cpu_to_le16(ext4_ext_get_actual_len(start_ext));
+ new_flag = 1;
+
+ } else if (ext4_ext_get_actual_len(start_ext) &&
+ ext4_ext_get_actual_len(new_ext) &&
+ !ext4_ext_get_actual_len(end_ext) &&
+ o_start == o_end) {
+
+ /* start_ext new_ext
+ * dest |--------------|---------------|
+ * org |------------------------------|
+ */
+
+ o_start->ee_len =
+ cpu_to_le16(ext4_ext_get_actual_len(start_ext));
+ new_flag = 1;
+
+ } else if (!ext4_ext_get_actual_len(start_ext) &&
+ ext4_ext_get_actual_len(new_ext) &&
+ ext4_ext_get_actual_len(end_ext) &&
+ o_start == o_end) {
+
+ /* new_ext end_ext
+ * dest |--------------|---------------|
+ * org |------------------------------|
+ */
+
+ o_end->ee_block = end_ext->ee_block;
+ o_end->ee_len =
+ cpu_to_le16(ext4_ext_get_actual_len(end_ext));
+ ext4_ext_store_pblock(o_end, ext_pblock(end_ext));
+
+ /*
+ * Set 0 to the extent block if new_ext was
+ * the first block.
+ */
+ if (!new_ext->ee_block)
+ eblock = 0;
+ else
+ eblock = le32_to_cpu(new_ext->ee_block);
+
+ new_flag = 1;
+ } else {
+ printk(KERN_ERR "ext4 defrag: Unexpected insert case\n");
+ return -EIO;
+ }
+
+ if (new_flag) {
+ org_path = ext4_ext_find_extent(org_inode, eblock, NULL);
+ if (IS_ERR(org_path)) {
+ err = PTR_ERR(org_path);
+ org_path = NULL;
+ goto out;
+ }
+ err = ext4_ext_insert_extent(handle, org_inode,
+ org_path, new_ext);
+ if (err)
+ goto out;
+ }
+
+ if (end_flag) {
+ org_path = ext4_ext_find_extent(org_inode,
+ le32_to_cpu(end_ext->ee_block) - 1, org_path);
+ if (IS_ERR(org_path)) {
+ err = PTR_ERR(org_path);
+ org_path = NULL;
+ goto out;
+ }
+ err = ext4_ext_insert_extent(handle, org_inode,
+ org_path, end_ext);
+ if (err)
+ goto out;
+ }
+out:
+ if (org_path) {
+ ext4_ext_drop_refs(org_path);
+ kfree(org_path);
+ }
+
+ return err;
+
+}
+
+/**
+ * ext4_defrag_insert_inside_block - Insert new extent to the extent block
+ *
+ * @o_start: first original extent to be moved
+ * @o_end: last original extent to be moved
+ * @start_ext: first new extent to be inserted
+ * @new_ext: middle of new extent to be inserted
+ * @end_ext: last new extent to be inserted
+ * @eh: extent header of target leaf block
+ * @replaced: the number of blocks which will be replaced with new_ext
+ * @range_to_move: used to decide how to insert extent
+ *
+ * Insert extents into the leaf block. The extent (@o_start) is overwritten
+ * by inserted extents.
+ */
+static int
+ext4_defrag_insert_inside_block(struct ext4_extent *o_start,
+ struct ext4_extent *o_end, struct ext4_extent *start_ext,
+ struct ext4_extent *new_ext, struct ext4_extent *end_ext,
+ struct ext4_extent_header *eh, ext4_fsblk_t replaced,
+ int range_to_move)
+{
+ int i = 0;
+ unsigned len;
+
+ /* Move the existing extents */
+ if (range_to_move && o_end < EXT_LAST_EXTENT(eh)) {
+ len = (unsigned long)(EXT_LAST_EXTENT(eh) + 1) -
+ (unsigned long)(o_end + 1);
+ memmove(o_end + 1 + range_to_move, o_end + 1, len);
+ }
+
+ /* Insert start entry */
+ if (ext4_ext_get_actual_len(start_ext))
+ o_start[i++].ee_len =
+ cpu_to_le16(ext4_ext_get_actual_len(start_ext));
+
+ /* Insert new entry */
+ if (ext4_ext_get_actual_len(new_ext)) {
+ o_start[i].ee_block = new_ext->ee_block;
+ o_start[i].ee_len = cpu_to_le16(replaced);
+ ext4_ext_store_pblock(&o_start[i++], ext_pblock(new_ext));
+ }
+
+ /* Insert end entry */
+ if (ext4_ext_get_actual_len(end_ext))
+ o_start[i] = *end_ext;
+
+ /* Increment the total entries counter on the extent block */
+ le16_add_cpu(&eh->eh_entries, range_to_move);
+
+ return 0;
+}
+
+/**
+ * ext4_defrag_insert_extents - Insert new extent
+ *
+ * @handle: journal handle
+ * @org_inode: original inode
+ * @org_path: path indicates first extent to be defraged
+ * @o_start: first original extent to be defraged
+ * @o_end: last original extent to be defraged
+ * @start_ext: first new extent to be inserted
+ * @new_ext: middle of new extent to be inserted
+ * @end_ext: last new extent to be inserted
+ * @replaced: the number of blocks which will be replaced with new_ext
+ *
+ * Call the function to insert extents. If we cannot add more extents into
+ * the leaf block, we call ext4_defrag_insert_across_blocks() to create a
+ * new leaf block. Otherwise call ext4_defrag_insert_inside_block(). Return 0
+ * on success, or a negative error value on failure.
+ */
+static int
+ext4_defrag_insert_extents(handle_t *handle, struct inode *org_inode,
+ struct ext4_ext_path *org_path,
+ struct ext4_extent *o_start, struct ext4_extent *o_end,
+ struct ext4_extent *start_ext, struct ext4_extent *new_ext,
+ struct ext4_extent *end_ext, ext4_fsblk_t replaced)
+{
+ struct ext4_extent_header *eh;
+ unsigned need_slots, slots_range;
+ int range_to_move, depth, ret;
+
+ /*
+ * The extents need to be inserted
+ * start_extent + new_extent + end_extent.
+ */
+ need_slots = (ext4_ext_get_actual_len(start_ext) ? 1 : 0) +
+ (ext4_ext_get_actual_len(end_ext) ? 1 : 0) +
+ (ext4_ext_get_actual_len(new_ext) ? 1 : 0);
+
+ /* The number of slots between start and end */
+ slots_range = ((unsigned long)(o_end + 1) - (unsigned long)o_start + 1)
+ / sizeof(struct ext4_extent);
+
+ /* Range to move the end of extent */
+ range_to_move = need_slots - slots_range;
+ depth = org_path->p_depth;
+ org_path += depth;
+ eh = org_path->p_hdr;
+
+ if (depth) {
+ /* Register to journal */
+ ret = ext4_journal_get_write_access(handle, org_path->p_bh);
+ if (ret)
+ return ret;
+ }
+
+ /* Expansion */
+ if (range_to_move > 0 &&
+ (range_to_move > le16_to_cpu(eh->eh_max)
+ - le16_to_cpu(eh->eh_entries))) {
+
+ ret = ext4_defrag_insert_across_blocks(handle, org_inode,
+ o_start, o_end, start_ext, new_ext,
+ end_ext);
+ if (ret < 0)
+ return ret;
+ } else {
+ ret = ext4_defrag_insert_inside_block(o_start, o_end,
+ start_ext, new_ext, end_ext, eh,
+ replaced, range_to_move);
+ if (ret < 0)
+ return ret;
+ }
+
+ if (depth) {
+ ret = ext4_handle_dirty_metadata(handle, org_inode,
+ org_path->p_bh);
+ if (ret)
+ return ret;
+ } else {
+ ret = ext4_mark_inode_dirty(handle, org_inode);
+ if (ret < 0)
+ return ret;
+ }
+
+ return 0;
+
+}
+
+/**
+ * ext4_defrag_leaf_block - Defragmentation for one leaf extent block
+ *
+ * @handle: journal handle
+ * @org_inode: original inode
+ * @org_path: path indicates first extent to be defraged
+ * @dext: destination extent
+ * @from: start offset on the target file
+ *
+ * In order to insert extents into the leaf block, we must divide the extent
+ * in the leaf block into three extents. The one is located to be inserted
+ * extents, and the others are located around it.
+ *
+ * Therefore, this function creates structures to save extents of the leaf
+ * block, and inserts extents by calling ext4_defrag_insert_extents() with
+ * created extents. Return 0 on success, or a negative error value on failure.
+ */
+static int
+ext4_defrag_leaf_block(handle_t *handle, struct inode *org_inode,
+ struct ext4_ext_path *org_path, struct ext4_extent *dext,
+ ext4_lblk_t *from)
+{
+ struct ext4_extent *oext, *o_start = NULL, *o_end = NULL, *prev_ext;
+ struct ext4_extent new_ext, start_ext, end_ext;
+ ext4_fsblk_t replaced = 0;
+ ext4_lblk_t new_end, lblock;
+ unsigned long depth;
+ unsigned short len;
+ ext4_fsblk_t new_phys_end;
+ int ret;
+
+ depth = ext_depth(org_inode);
+ start_ext.ee_len = end_ext.ee_len = 0;
+ o_start = o_end = oext = org_path[depth].p_ext;
+ ext4_ext_store_pblock(&new_ext, ext_pblock(dext));
+ new_ext.ee_len =
+ cpu_to_le16(ext4_ext_get_actual_len(dext));
+ len = le16_to_cpu(new_ext.ee_len);
+ new_ext.ee_block = cpu_to_le32(*from);
+ lblock = le32_to_cpu(oext->ee_block);
+ new_end = le32_to_cpu(new_ext.ee_block)
+ + le16_to_cpu(new_ext.ee_len) - 1;
+ new_phys_end = ext_pblock(&new_ext)
+ + le16_to_cpu(new_ext.ee_len) - 1;
+
+ /*
+ * First original extent
+ * dest |---------------|
+ * org |---------------|
+ */
+ if (le32_to_cpu(new_ext.ee_block) >
+ le32_to_cpu(oext->ee_block) &&
+ le32_to_cpu(new_ext.ee_block) <
+ le32_to_cpu(oext->ee_block)
+ + ext4_ext_get_actual_len(oext)) {
+ start_ext.ee_len = cpu_to_le16(le32_to_cpu(new_ext.ee_block)
+ - le32_to_cpu(oext->ee_block));
+ replaced += ext4_ext_get_actual_len(oext)
+ - le16_to_cpu(start_ext.ee_len);
+ } else if (oext > EXT_FIRST_EXTENT(org_path[depth].p_hdr)) {
+ /* We can merge previous extent. */
+ prev_ext = oext - 1;
+ if (ext4_can_extents_be_merged(org_inode, prev_ext, &new_ext)) {
+ o_start = prev_ext;
+ start_ext.ee_len = cpu_to_le16(
+ ext4_ext_get_actual_len(prev_ext)
+ + le16_to_cpu(new_ext.ee_len));
+ new_ext.ee_len = 0;
+ }
+ }
+
+ for (;;) {
+ /* The extent for destination must be found. */
+ BUG_ON(!oext || lblock != le32_to_cpu(oext->ee_block));
+ lblock += ext4_ext_get_actual_len(oext);
+
+ /*
+ * Middle of original extent
+ * dest |-------------------|
+ * org |-----------------|
+ */
+ if (le32_to_cpu(new_ext.ee_block) <=
+ le32_to_cpu(oext->ee_block) &&
+ new_end >= le32_to_cpu(oext->ee_block)
+ + ext4_ext_get_actual_len(oext) - 1)
+ replaced += ext4_ext_get_actual_len(oext);
+
+ /*
+ * Last original extent
+ * dest |----------------|
+ * org |---------------|
+ */
+ if (new_end >= le32_to_cpu(oext->ee_block) &&
+ new_end < le32_to_cpu(oext->ee_block)
+ + ext4_ext_get_actual_len(oext) - 1) {
+ end_ext.ee_len
+ = cpu_to_le16(le32_to_cpu(oext->ee_block)
+ + ext4_ext_get_actual_len(oext) - 1 - new_end);
+ ext4_ext_store_pblock(&end_ext, (ext_pblock(o_end)
+ + ext4_ext_get_actual_len(oext)
+ - le16_to_cpu(end_ext.ee_len)));
+ end_ext.ee_block
+ = cpu_to_le32(le32_to_cpu(o_end->ee_block)
+ + ext4_ext_get_actual_len(oext)
+ - le16_to_cpu(end_ext.ee_len));
+ replaced += ext4_ext_get_actual_len(oext)
+ - le16_to_cpu(end_ext.ee_len);
+ }
+
+ /*
+ * Detected the block end, reached the number of replaced
+ * blocks to dext->ee_len. Then merge the extent.
+ */
+ if (oext == EXT_LAST_EXTENT(org_path[depth].p_hdr) ||
+ new_end <= le32_to_cpu(oext->ee_block)
+ + ext4_ext_get_actual_len(oext) - 1) {
+ ret = ext4_defrag_insert_extents(handle, org_inode,
+ org_path, o_start, o_end, &start_ext,
+ &new_ext, &end_ext, replaced);
+ if (ret < 0)
+ return ret;
+
+ /* All expected blocks are replaced */
+ if (le16_to_cpu(new_ext.ee_len) <= 0)
+ return 0;
+
+ /* Re-calculate new_ext */
+ le16_add_cpu(&new_ext.ee_len, -replaced);
+ le32_add_cpu(&new_ext.ee_block, replaced);
+ ext4_ext_store_pblock(&new_ext, ext_pblock(&new_ext)
+ + replaced);
+ replaced = 0;
+ start_ext.ee_len = end_ext.ee_len = 0;
+ o_start = NULL;
+
+ /* All expected blocks are replaced. */
+ if (le16_to_cpu(new_ext.ee_len) <= 0)
+ return 0;
+ }
+
+ /* Get the next extent for original. */
+ if (org_path)
+ ext4_ext_drop_refs(org_path);
+ org_path = ext4_ext_find_extent(org_inode, lblock, org_path);
+ if (IS_ERR(org_path)) {
+ ret = PTR_ERR(org_path);
+ org_path = NULL;
+ return ret;
+ }
+ depth = ext_depth(org_inode);
+ oext = org_path[depth].p_ext;
+ if (le32_to_cpu(oext->ee_block) + ext4_ext_get_actual_len(oext)
+ <= lblock)
+ return -ENOENT;
+
+ o_end = oext;
+ if (!o_start)
+ o_start = oext;
+ }
+}
+
+/**
+ * ext4_defrag_replace_branches - Replace original extents with new extents
+ *
+ * @handle: journal handle
+ * @org_inode: original inode
+ * @dest_inode: destination inode
+ * @from: block offset of org_inode
+ * @count: block count to be replaced
+ *
+ * Replace original inode extents and destination inode extents every page.
+ * We implement this replacement in the following three steps:
+ * 1. Save the block information of original and destination inodes into
+ * dummy extents.
+ * 2. Change the block information of original inode to point at the
+ * destination inode blocks.
+ * 3. Change the block information of destination inode to point at the saved
+ * original inode blocks in the dummy extents.
+ *
+ * Return 0 on success, or a negative error value on failure.
+ */
+static int
+ext4_defrag_replace_branches(handle_t *handle, struct inode *org_inode,
+ struct inode *dest_inode, ext4_lblk_t from,
+ ext4_lblk_t count)
+{
+ struct ext4_ext_path *org_path = NULL;
+ struct ext4_ext_path *dest_path = NULL;
+ struct ext4_extent *oext, *dext, *swap_ext;
+ struct ext4_extent tmp_ext, tmp_ext2;
+ ext4_lblk_t diff, org_diff, dest_off = from;
+ int err = 0;
+ int depth;
+ int replaced_count = 0;
+
+ /* Get the original extent for the block "from" */
+ org_path = ext4_ext_find_extent(org_inode, from, NULL);
+ if (IS_ERR(org_path)) {
+ err = PTR_ERR(org_path);
+ org_path = NULL;
+ goto out;
+ }
+
+ /* Get the destination extent for the head */
+ dest_path = ext4_ext_find_extent(dest_inode, dest_off, NULL);
+ if (IS_ERR(dest_path)) {
+ err = PTR_ERR(dest_path);
+ dest_path = NULL;
+ goto out;
+ }
+ depth = ext_depth(dest_inode);
+ dext = dest_path[depth].p_ext;
+ /* When dext is too large, pick up the target range. */
+ diff = dest_off - le32_to_cpu(dext->ee_block);
+ ext4_ext_store_pblock(&tmp_ext, ext_pblock(dext) + diff);
+ tmp_ext.ee_block = cpu_to_le32(le32_to_cpu(dext->ee_block) + diff);
+ tmp_ext.ee_len = cpu_to_le16(ext4_ext_get_actual_len(dext) - diff);
+ if (count < le16_to_cpu(tmp_ext.ee_len))
+ tmp_ext.ee_len = cpu_to_le16(count);
+ dext = &tmp_ext;
+
+ depth = ext_depth(org_inode);
+ oext = org_path[depth].p_ext;
+ org_diff = from - le32_to_cpu(oext->ee_block);
+ ext4_ext_store_pblock(&tmp_ext2, ext_pblock(oext) + org_diff);
+ tmp_ext2.ee_block = tmp_ext.ee_block;
+
+ /* Adjust extent length when blocksize != pagesize */
+ if (le16_to_cpu(tmp_ext.ee_len) <=
+ ext4_ext_get_actual_len(oext) - org_diff) {
+ tmp_ext2.ee_len = tmp_ext.ee_len;
+ } else {
+ tmp_ext2.ee_len = cpu_to_le16(ext4_ext_get_actual_len(oext)
+ - org_diff);
+ tmp_ext.ee_len = tmp_ext2.ee_len;
+ }
+ swap_ext = &tmp_ext2;
+
+ /* Loop for the destination extents */
+ while (1) {
+ /* The extent for destination must be found. */
+ BUG_ON(!dext || dest_off != le32_to_cpu(dext->ee_block));
+
+ /* Loop for the original extent blocks */
+ err = ext4_defrag_leaf_block(handle, org_inode,
+ org_path, dext, &from);
+ if (err < 0)
+ goto out;
+
+ /*
+ * We need the function which fixes extent information for
+ * inserting.
+ * e.g. ext4_defrag_inset_extents()
+ */
+ err = ext4_defrag_leaf_block(handle, dest_inode,
+ dest_path, swap_ext, &dest_off);
+ if (err < 0)
+ goto out;
+
+ replaced_count += ext4_ext_get_actual_len(dext);
+ dest_off += ext4_ext_get_actual_len(dext);
+ from += ext4_ext_get_actual_len(dext);
+
+ /* Already moved the expected blocks */
+ if (replaced_count >= count)
+ break;
+
+ if (org_path)
+ ext4_ext_drop_refs(org_path);
+ org_path = ext4_ext_find_extent(org_inode, from, NULL);
+ if (IS_ERR(org_path)) {
+ err = PTR_ERR(org_path);
+ org_path = NULL;
+ goto out;
+ }
+ depth = ext_depth(org_inode);
+ oext = org_path[depth].p_ext;
+ if (le32_to_cpu(oext->ee_block) + ext4_ext_get_actual_len(oext)
+ <= from) {
+ err = 0;
+ goto out;
+ }
+
+ if (dest_path)
+ ext4_ext_drop_refs(dest_path);
+ dest_path = ext4_ext_find_extent(dest_inode, dest_off, NULL);
+ if (IS_ERR(dest_path)) {
+ err = PTR_ERR(dest_path);
+ dest_path = NULL;
+ goto out;
+ }
+ depth = ext_depth(dest_inode);
+ dext = dest_path[depth].p_ext;
+ if (le32_to_cpu(dext->ee_block) + ext4_ext_get_actual_len(dext)
+ <= dest_off) {
+ err = 0;
+ goto out;
+ }
+
+ /* When dext is too large, pick up the target range. */
+ diff = dest_off - le32_to_cpu(dext->ee_block);
+ ext4_ext_store_pblock(&tmp_ext, ext_pblock(dext) + diff);
+ tmp_ext.ee_block =
+ cpu_to_le32(le32_to_cpu(dext->ee_block) + diff);
+ tmp_ext.ee_len = cpu_to_le16(ext4_ext_get_actual_len(dext)
+ - diff);
+
+ if (count - replaced_count < le16_to_cpu(tmp_ext.ee_len))
+ tmp_ext.ee_len = cpu_to_le16(count - replaced_count);
+
+ dext = &tmp_ext;
+
+ org_diff = from - le32_to_cpu(oext->ee_block);
+ ext4_ext_store_pblock(&tmp_ext2, ext_pblock(oext) + org_diff);
+ tmp_ext2.ee_block = tmp_ext.ee_block;
+
+ /* Adjust extent length when blocksize != pagesize */
+ if (le16_to_cpu(tmp_ext.ee_len) <=
+ ext4_ext_get_actual_len(oext) - org_diff) {
+ tmp_ext2.ee_len = tmp_ext.ee_len;
+ } else {
+ tmp_ext2.ee_len = cpu_to_le16(
+ ext4_ext_get_actual_len(oext)
+ - org_diff);
+ tmp_ext.ee_len = tmp_ext2.ee_len;
+ }
+ swap_ext = &tmp_ext2;
+ }
+
+out:
+ if (org_path) {
+ ext4_ext_drop_refs(org_path);
+ kfree(org_path);
+ }
+ if (dest_path) {
+ ext4_ext_drop_refs(dest_path);
+ kfree(dest_path);
+ }
+
+ return err;
+}
+
+/**
+ * ext4_defrag_partial - Defrag a file per page
+ *
+ * @o_filp: file structure of original file
+ * @dest_inode: destination inode
+ * @org_page_offset: page index on original file
+ * @data_offset_in_page: block index where data swapping starts
+ * @block_len_in_page: the number of blocks to be swapped
+ *
+ * Save the data in original inode blocks and replace original inode extents
+ * with destination inode extents by calling ext4_defrag_replace_branches().
+ * Finally, write out the saved data in new original inode blocks. Return 0
+ * on success, or a negative error value on failure.
+ */
static int
ext4_defrag_partial(struct file *o_filp, struct inode *dest_inode,
pgoff_t org_page_offset, int data_offset_in_page,
int block_len_in_page)
{
- return 0;
+ struct inode *org_inode = o_filp->f_dentry->d_inode;
+ struct address_space *mapping = org_inode->i_mapping;
+ struct buffer_head *bh;
+ struct page *page = NULL;
+ const struct address_space_operations *a_ops = mapping->a_ops;
+ handle_t *handle;
+ ext4_lblk_t org_blk_offset;
+ long long offs = org_page_offset << PAGE_CACHE_SHIFT;
+ unsigned long blocksize = org_inode->i_sb->s_blocksize;
+ unsigned int w_flags = 0;
+ unsigned int tmp_data_len;
+ unsigned data_len;
+ void *fsdata;
+ int ret, i, jblocks;
+ int blocks_per_page = PAGE_CACHE_SIZE >> org_inode->i_blkbits;
+
+ /*
+ * It needs twice the amount of ordinary journal buffers because
+ * inode and dest_inode may change each different metadata blocks.
+ */
+ jblocks = ext4_writepage_trans_blocks(org_inode) * 2;
+ handle = ext4_journal_start(org_inode, jblocks);
+ if (IS_ERR(handle)) {
+ ret = PTR_ERR(handle);
+ return ret;
+ }
+
+ if (segment_eq(get_fs(), KERNEL_DS))
+ w_flags |= AOP_FLAG_UNINTERRUPTIBLE;
+
+ org_blk_offset = org_page_offset * blocks_per_page +
+ data_offset_in_page;
+ offs = (long long)org_blk_offset << org_inode->i_blkbits;
+
+ /* Calculate data_len */
+ if ((org_blk_offset + block_len_in_page - 1) ==
+ ((org_inode->i_size - 1) >> org_inode->i_blkbits)) {
+ /* the case which we replace the last block */
+ tmp_data_len = org_inode->i_size & (blocksize - 1);
+ /*
+ * If data_len equal zero, it shows data_len is multiples of
+ * blocksize. So we set appropriate value.
+ */
+ if (tmp_data_len == 0)
+ tmp_data_len = blocksize;
+
+ data_len = tmp_data_len +
+ ((block_len_in_page - 1) << org_inode->i_blkbits);
+ } else {
+ data_len = block_len_in_page << org_inode->i_blkbits;
+ }
+
+ up_write(&EXT4_I(org_inode)->i_data_sem);
+ ret = a_ops->write_begin(o_filp, mapping, offs, data_len, w_flags,
+ &page, &fsdata);
+ down_write(&EXT4_I(org_inode)->i_data_sem);
+
+ if (unlikely(ret < 0))
+ goto out;
+
+ if (!PageUptodate(page)) {
+ up_write(&EXT4_I(org_inode)->i_data_sem);
+ mapping->a_ops->readpage(o_filp, page);
+ down_write(&EXT4_I(org_inode)->i_data_sem);
+ lock_page(page);
+ }
+
+ /*
+ * try_to_release_page() doesn't call releasepage in writeback mode.
+ * We should care about the order of writing to the same file
+ * by multiple defrag processes.
+ * It needs to call wait_on_page_writeback() to wait for the
+ * writeback of the page.
+ */
+ if (PageWriteback(page))
+ wait_on_page_writeback(page);
+
+ /* Release old bh and drop refs */
+ try_to_release_page(page, 0);
+ ret = ext4_defrag_replace_branches(handle, org_inode, dest_inode,
+ org_blk_offset,
+ block_len_in_page);
+ if (ret < 0)
+ goto out;
+
+ /* Clear the inode cache not to refer to the old data */
+ ext4_ext_invalidate_cache(org_inode);
+
+ if (!page_has_buffers(page))
+ create_empty_buffers(page, 1 << org_inode->i_blkbits, 0);
+
+ bh = page_buffers(page);
+ for (i = 0; i < data_offset_in_page; i++)
+ bh = bh->b_this_page;
+
+ for (i = 0; i < block_len_in_page; i++) {
+ up_write(&EXT4_I(org_inode)->i_data_sem);
+ ret = ext4_get_block(org_inode, (sector_t)(org_blk_offset + i),
+ bh, 0);
+ down_write(&EXT4_I(org_inode)->i_data_sem);
+
+ if (ret < 0)
+ goto out;
+
+ if (bh->b_this_page != NULL)
+ bh = bh->b_this_page;
+ }
+
+ ret = a_ops->write_end(o_filp, mapping, offs, data_len, data_len, page,
+ fsdata);
+ page = NULL;
+
+out:
+ if (unlikely(page)) {
+ if (PageLocked(page))
+ unlock_page(page);
+ page_cache_release(page);
+ }
+ ext4_journal_stop(handle);
+
+ return ret < 0 ? ret : 0;
}
/**
On Fri, Jan 30, 2009 at 03:13:37PM +0900, Akira Fujita wrote:
> ext4: online defrag -- Exchange the blocks between two inodes
>
> From: Akira Fujita <[email protected]>
>
> For each page, exchange the extents between original inode
> and destination inode, and then write the file data of
> the original inode to destination inode.
As I mentioned earlier, it would be better to merge this patch into
the previous once; we don't need to keep them broken apart.
> +/**
> + * ext4_defrag_replace_branches - Replace original extents with new extents
> + *
> + * @handle: journal handle
> + * @org_inode: original inode
> + * @dest_inode: destination inode
> + * @from: block offset of org_inode
> + * @count: block count to be replaced
It's really good that this function can support moving an arbitrarily
large block range. It's unfortunate that its caller is only moving a
4k page at a time. :-)
> +
> + up_write(&EXT4_I(org_inode)->i_data_sem);
> + ret = a_ops->write_begin(o_filp, mapping, offs, data_len, w_flags,
> + &page, &fsdata);
> + down_write(&EXT4_I(org_inode)->i_data_sem);
This is going to be a problem. Once we release i_data_sem, there is
the possibility that other processes which might be running and
accessing the file at the same time that the defragger is running
could be blocked waiting for i_data_sem to be released. The moment it
gets released, they will grab the lock then start to modify extent
tree --- either allocating new blocks to it, or worse, truncating or
unlinking the target inode.
This is going to be a mess to fix, since Linux doesn't have recursive
locking primitives. We do take i_mutex, which will protect us against
truncates, but it won't protect against a write() system call. Also,
if there inode has delayed allocation blocks pending, those could get
written out by the page cleaner, and i_mutex won't protect us against
changes to i_data_sem, either.
We could add special-case kludgery to wrap around all of the places
that takes or release the i_data_sem so that we get recursive locking
support --- but that would be very ugly indeed.
I'm not sure what's the best way to deal with this; maybe if we sleep
on it someone will come up with a better suggestion --- but it's
something that we have to figure out.
- Ted
Hi Ted,
Thank you for comment. :-)
Theodore Tso wrote:
>> +
>> + up_write(&EXT4_I(org_inode)->i_data_sem);
>> + ret = a_ops->write_begin(o_filp, mapping, offs, data_len, w_flags,
>> + &page, &fsdata);
>> + down_write(&EXT4_I(org_inode)->i_data_sem);
>>
>
> This is going to be a problem. Once we release i_data_sem, there is
> the possibility that other processes which might be running and
> accessing the file at the same time that the defragger is running
> could be blocked waiting for i_data_sem to be released. The moment it
> gets released, they will grab the lock then start to modify extent
> tree --- either allocating new blocks to it, or worse, truncating or
> unlinking the target inode.
>
> This is going to be a mess to fix, since Linux doesn't have recursive
> locking primitives. We do take i_mutex, which will protect us against
> truncates, but it won't protect against a write() system call. Also,
> if there inode has delayed allocation blocks pending, those could get
> written out by the page cleaner, and i_mutex won't protect us against
> changes to i_data_sem, either.
>
>
As you said, we take i_mutex at the start of ext4_defrag()
and hold it until the end of this function,
so orig file is protected against truncates and
it never be shrunk during defrag.
On the other hand, semaphore is released/taken around write_begin()
in ext4_defrag_partial() every page, so it does not protect orig file
against a write() system call from other process.
So that defrag result (fragmentation) might not be best,
but data corruption does not occur at least.
So I think it is not a serious problem.
As above, it is not necessary to lock the whole of ext4_defrag()
with semaphore, we should just lock only a necessary point.
Therefore defrag V1's lock seems have unneeded lock points.
I will change lock point and semaphore type in the next version.
Do I overlook something?
Regards,
Akira Fujita
> We could add special-case kludgery to wrap around all of the places
> that takes or release the i_data_sem so that we get recursive locking
> support --- but that would be very ugly indeed.
>
> I'm not sure what's the best way to deal with this; maybe if we sleep
> on it someone will come up with a better suggestion --- but it's
> something that we have to figure out.
>
> - Ted
>
>