alloc.c

linux 内核源代码

 *
 * ocfs2_shift_tree_depth() uses this to determine the # clusters
 * value for the new topmost tree record.
 */
static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list  *el)
{
	int i;

	i = le16_to_cpu(el->l_next_free_rec) - 1;

	return le32_to_cpu(el->l_recs[i].e_cpos) +
		ocfs2_rec_clusters(el, &el->l_recs[i]);
}

/*
 * Add an entire tree branch to our inode. eb_bh is the extent block
 * to start at, if we don't want to start the branch at the dinode
 * structure.
 *
 * last_eb_bh is required as we have to update it's next_leaf pointer
 * for the new last extent block.
 *
 * the new branch will be 'empty' in the sense that every block will
 * contain a single record with cluster count == 0.
 */
static int ocfs2_add_branch(struct ocfs2_super *osb,
			    handle_t *handle,
			    struct inode *inode,
			    struct buffer_head *fe_bh,
			    struct buffer_head *eb_bh,
			    struct buffer_head **last_eb_bh,
			    struct ocfs2_alloc_context *meta_ac)
{
	int status, new_blocks, i;
	u64 next_blkno, new_last_eb_blk;
	struct buffer_head *bh;
	struct buffer_head **new_eb_bhs = NULL;
	struct ocfs2_dinode *fe;
	struct ocfs2_extent_block *eb;
	struct ocfs2_extent_list  *eb_el;
	struct ocfs2_extent_list  *el;
	u32 new_cpos;

	mlog_entry_void();

	BUG_ON(!last_eb_bh || !*last_eb_bh);

	fe = (struct ocfs2_dinode *) fe_bh->b_data;

	if (eb_bh) {
		eb = (struct ocfs2_extent_block *) eb_bh->b_data;
		el = &eb->h_list;
	} else
		el = &fe->id2.i_list;

	/* we never add a branch to a leaf. */
	BUG_ON(!el->l_tree_depth);

	new_blocks = le16_to_cpu(el->l_tree_depth);

	/* allocate the number of new eb blocks we need */
	new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
			     GFP_KERNEL);
	if (!new_eb_bhs) {
		status = -ENOMEM;
		mlog_errno(status);
		goto bail;
	}

	status = ocfs2_create_new_meta_bhs(osb, handle, inode, new_blocks,
					   meta_ac, new_eb_bhs);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}

	eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
	new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);

	/* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
	 * linked with the rest of the tree.
	 * conversly, new_eb_bhs[0] is the new bottommost leaf.
	 *
	 * when we leave the loop, new_last_eb_blk will point to the
	 * newest leaf, and next_blkno will point to the topmost extent
	 * block. */
	next_blkno = new_last_eb_blk = 0;
	for(i = 0; i < new_blocks; i++) {
		bh = new_eb_bhs[i];
		eb = (struct ocfs2_extent_block *) bh->b_data;
		if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
			OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
			status = -EIO;
			goto bail;
		}
		eb_el = &eb->h_list;

		status = ocfs2_journal_access(handle, inode, bh,
					      OCFS2_JOURNAL_ACCESS_CREATE);
		if (status < 0) {
			mlog_errno(status);
			goto bail;
		}

		eb->h_next_leaf_blk = 0;
		eb_el->l_tree_depth = cpu_to_le16(i);
		eb_el->l_next_free_rec = cpu_to_le16(1);
		/*
		 * This actually counts as an empty extent as
		 * c_clusters == 0
		 */
		eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
		eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
		/*
		 * eb_el isn't always an interior node, but even leaf
		 * nodes want a zero'd flags and reserved field so
		 * this gets the whole 32 bits regardless of use.
		 */
		eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
		if (!eb_el->l_tree_depth)
			new_last_eb_blk = le64_to_cpu(eb->h_blkno);

		status = ocfs2_journal_dirty(handle, bh);
		if (status < 0) {
			mlog_errno(status);
			goto bail;
		}

		next_blkno = le64_to_cpu(eb->h_blkno);
	}

	/* This is a bit hairy. We want to update up to three blocks
	 * here without leaving any of them in an inconsistent state
	 * in case of error. We don't have to worry about
	 * journal_dirty erroring as it won't unless we've aborted the
	 * handle (in which case we would never be here) so reserving
	 * the write with journal_access is all we need to do. */
	status = ocfs2_journal_access(handle, inode, *last_eb_bh,
				      OCFS2_JOURNAL_ACCESS_WRITE);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}
	status = ocfs2_journal_access(handle, inode, fe_bh,
				      OCFS2_JOURNAL_ACCESS_WRITE);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}
	if (eb_bh) {
		status = ocfs2_journal_access(handle, inode, eb_bh,
					      OCFS2_JOURNAL_ACCESS_WRITE);
		if (status < 0) {
			mlog_errno(status);
			goto bail;
		}
	}

	/* Link the new branch into the rest of the tree (el will
	 * either be on the fe, or the extent block passed in. */
	i = le16_to_cpu(el->l_next_free_rec);
	el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
	el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
	el->l_recs[i].e_int_clusters = 0;
	le16_add_cpu(&el->l_next_free_rec, 1);

	/* fe needs a new last extent block pointer, as does the
	 * next_leaf on the previously last-extent-block. */
	fe->i_last_eb_blk = cpu_to_le64(new_last_eb_blk);

	eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
	eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);

	status = ocfs2_journal_dirty(handle, *last_eb_bh);
	if (status < 0)
		mlog_errno(status);
	status = ocfs2_journal_dirty(handle, fe_bh);
	if (status < 0)
		mlog_errno(status);
	if (eb_bh) {
		status = ocfs2_journal_dirty(handle, eb_bh);
		if (status < 0)
			mlog_errno(status);
	}

	/*
	 * Some callers want to track the rightmost leaf so pass it
	 * back here.
	 */
	brelse(*last_eb_bh);
	get_bh(new_eb_bhs[0]);
	*last_eb_bh = new_eb_bhs[0];

	status = 0;
bail:
	if (new_eb_bhs) {
		for (i = 0; i < new_blocks; i++)
			if (new_eb_bhs[i])
				brelse(new_eb_bhs[i]);
		kfree(new_eb_bhs);
	}

	mlog_exit(status);
	return status;
}

/*
 * adds another level to the allocation tree.
 * returns back the new extent block so you can add a branch to it
 * after this call.
 */
static int ocfs2_shift_tree_depth(struct ocfs2_super *osb,
				  handle_t *handle,
				  struct inode *inode,
				  struct buffer_head *fe_bh,
				  struct ocfs2_alloc_context *meta_ac,
				  struct buffer_head **ret_new_eb_bh)
{
	int status, i;
	u32 new_clusters;
	struct buffer_head *new_eb_bh = NULL;
	struct ocfs2_dinode *fe;
	struct ocfs2_extent_block *eb;
	struct ocfs2_extent_list  *fe_el;
	struct ocfs2_extent_list  *eb_el;

	mlog_entry_void();

	status = ocfs2_create_new_meta_bhs(osb, handle, inode, 1, meta_ac,
					   &new_eb_bh);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}

	eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
	if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
		OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
		status = -EIO;
		goto bail;
	}

	eb_el = &eb->h_list;
	fe = (struct ocfs2_dinode *) fe_bh->b_data;
	fe_el = &fe->id2.i_list;

	status = ocfs2_journal_access(handle, inode, new_eb_bh,
				      OCFS2_JOURNAL_ACCESS_CREATE);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}

	/* copy the fe data into the new extent block */
	eb_el->l_tree_depth = fe_el->l_tree_depth;
	eb_el->l_next_free_rec = fe_el->l_next_free_rec;
	for(i = 0; i < le16_to_cpu(fe_el->l_next_free_rec); i++)
		eb_el->l_recs[i] = fe_el->l_recs[i];

	status = ocfs2_journal_dirty(handle, new_eb_bh);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}

	status = ocfs2_journal_access(handle, inode, fe_bh,
				      OCFS2_JOURNAL_ACCESS_WRITE);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}

	new_clusters = ocfs2_sum_rightmost_rec(eb_el);

	/* update fe now */
	le16_add_cpu(&fe_el->l_tree_depth, 1);
	fe_el->l_recs[0].e_cpos = 0;
	fe_el->l_recs[0].e_blkno = eb->h_blkno;
	fe_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
	for(i = 1; i < le16_to_cpu(fe_el->l_next_free_rec); i++)
		memset(&fe_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
	fe_el->l_next_free_rec = cpu_to_le16(1);

	/* If this is our 1st tree depth shift, then last_eb_blk
	 * becomes the allocated extent block */
	if (fe_el->l_tree_depth == cpu_to_le16(1))
		fe->i_last_eb_blk = eb->h_blkno;

	status = ocfs2_journal_dirty(handle, fe_bh);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}

	*ret_new_eb_bh = new_eb_bh;
	new_eb_bh = NULL;
	status = 0;
bail:
	if (new_eb_bh)
		brelse(new_eb_bh);

	mlog_exit(status);
	return status;
}

/*
 * Should only be called when there is no space left in any of the
 * leaf nodes. What we want to do is find the lowest tree depth
 * non-leaf extent block with room for new records. There are three
 * valid results of this search:
 *
 * 1) a lowest extent block is found, then we pass it back in
 *    *lowest_eb_bh and return '0'
 *
 * 2) the search fails to find anything, but the dinode has room. We
 *    pass NULL back in *lowest_eb_bh, but still return '0'
 *
 * 3) the search fails to find anything AND the dinode is full, in
 *    which case we return > 0
 *
 * return status < 0 indicates an error.
 */
static int ocfs2_find_branch_target(struct ocfs2_super *osb,
				    struct inode *inode,
				    struct buffer_head *fe_bh,
				    struct buffer_head **target_bh)
{
	int status = 0, i;
	u64 blkno;
	struct ocfs2_dinode *fe;
	struct ocfs2_extent_block *eb;
	struct ocfs2_extent_list  *el;
	struct buffer_head *bh = NULL;
	struct buffer_head *lowest_bh = NULL;

	mlog_entry_void();

	*target_bh = NULL;

	fe = (struct ocfs2_dinode *) fe_bh->b_data;
	el = &fe->id2.i_list;

	while(le16_to_cpu(el->l_tree_depth) > 1) {
		if (le16_to_cpu(el->l_next_free_rec) == 0) {
			ocfs2_error(inode->i_sb, "Dinode %llu has empty "
				    "extent list (next_free_rec == 0)",
				    (unsigned long long)OCFS2_I(inode)->ip_blkno);
			status = -EIO;
			goto bail;
		}
		i = le16_to_cpu(el->l_next_free_rec) - 1;
		blkno = le64_to_cpu(el->l_recs[i].e_blkno);
		if (!blkno) {
			ocfs2_error(inode->i_sb, "Dinode %llu has extent "
				    "list where extent # %d has no physical "
				    "block start",
				    (unsigned long long)OCFS2_I(inode)->ip_blkno, i);
			status = -EIO;
			goto bail;
		}

		if (bh) {
			brelse(bh);
			bh = NULL;
		}

		status = ocfs2_read_block(osb, blkno, &bh, OCFS2_BH_CACHED,
					  inode);
		if (status < 0) {
			mlog_errno(status);
			goto bail;
		}

		eb = (struct ocfs2_extent_block *) bh->b_data;
		if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
			OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
			status = -EIO;
			goto bail;
		}
		el = &eb->h_list;

		if (le16_to_cpu(el->l_next_free_rec) <
		    le16_to_cpu(el->l_count)) {
			if (lowest_bh)
				brelse(lowest_bh);
			lowest_bh = bh;
			get_bh(lowest_bh);
		}
	}

	/* If we didn't find one and the fe doesn't have any room,
	 * then return '1' */
	if (!lowest_bh
	    && (fe->id2.i_list.l_next_free_rec == fe->id2.i_list.l_count))
		status = 1;

	*target_bh = lowest_bh;
bail:
	if (bh)
		brelse(bh);

	mlog_exit(status);
	return status;
}

/*
 * Grow a b-tree so that it has more records.
 *
 * We might shift the tree depth in which case existing paths should
 * be considered invalid.
 *
 * Tree depth after the grow is returned via *final_depth.
 *
 * *last_eb_bh will be updated by ocfs2_add_branch().
 */
static int ocfs2_grow_tree(struct inode *inode, handle_t *handle,
			   struct buffer_head *di_bh, int *final_depth,
			   struct buffer_head **last_eb_bh,
			   struct ocfs2_alloc_context *meta_ac)
{
	int ret, shift;
	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
	int depth = le16_to_cpu(di->id2.i_list.l_tree_depth);
	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
	struct buffer_head *bh = NULL;

	BUG_ON(meta_ac == NULL);

	shift = ocfs2_find_branch_target(osb, inode, di_bh, &bh);
	if (shift < 0) {
		ret = shift;
		mlog_errno(ret);
		goto out;
	}

	/* We traveled all the way to the bottom of the allocation tree
	 * and didn't find room for any more extents - we need to add
	 * another tree level */
	if (shift) {
		BUG_ON(bh);
		mlog(0, "need to shift tree depth (current = %d)\n", depth);

		/* ocfs2_shift_tree_depth will return us a buffer with
		 * the new extent block (so we can pass that to
		 * ocfs2_add_branch). */
		ret = ocfs2_shift_tree_depth(osb, handle, inode, di_bh,
					     meta_ac, &bh);
		if (ret < 0) {
			mlog_errno(ret);
			goto out;
		}
		depth++;
		if (depth == 1) {
			/*
			 * Special case: we have room now if we shifted from
			 * tree_depth 0, so no more work needs to be done.
			 *
			 * We won't be calling add_branch, so pass
			 * back *last_eb_bh as the new leaf. At depth
			 * zero, it should always be null so there's
			 * no reason to brelse.
			 */
			BUG_ON(*last_eb_bh);
			get_bh(bh);
			*last_eb_bh = bh;
			goto out;
		}
	}

	/* call ocfs2_add_branch to add the final part of the tree with
	 * the new data. */
	mlog(0, "add branch. bh = %p\n", bh);
	ret = ocfs2_add_branch(osb, handle, inode, di_bh, bh, last_eb_bh,
			       meta_ac);
	if (ret < 0) {
		mlog_errno(ret);
		goto out;
	}

out:
	if (final_depth)
		*final_depth = depth;
	brelse(bh);
	return ret;
}

/*
 * This is only valid for leaf nodes, which are the only ones that can
 * have empty extents anyway.
 */
static inline int ocfs2_is_empty_extent(struct ocfs2_extent_rec *rec)
{
	return !rec->e_leaf_clusters;
}

/*
 * This function will discard the rightmost extent record.
 */
static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
{
	int next_free = le16_to_cpu(el->l_next_free_rec);
	int count = le16_to_cpu(el->l_count);
	unsigned int num_bytes;

	BUG_ON(!next_free);