Btrfs: Cache free inode numbers in memory

Currently btrfs stores the highest objectid of the fs tree, and it always
returns (highest+1) inode number when we create a file, so inode numbers
won't be reclaimed when we delete files, so we'll run out of inode numbers
as we keep create/delete files in 32bits machines.

This fixes it, and it works similarly to how we cache free space in block
cgroups.

We start a kernel thread to read the file tree. By scanning inode items,
we know which chunks of inode numbers are free, and we cache them in
an rb-tree.

Because we are searching the commit root, we have to carefully handle the
cross-transaction case.

The rb-tree is a hybrid extent+bitmap tree, so if we have too many small
chunks of inode numbers, we'll use bitmaps. Initially we allow 16K ram
of extents, and a bitmap will be used if we exceed this threshold. The
extents threshold is adjusted in runtime.

Signed-off-by: Li Zefan <lizf@cn.fujitsu.com>
diff --git a/fs/btrfs/inode-map.c b/fs/btrfs/inode-map.c
index c05a08f..5be62df 100644
--- a/fs/btrfs/inode-map.c
+++ b/fs/btrfs/inode-map.c
@@ -16,11 +16,343 @@
  * Boston, MA 021110-1307, USA.
  */
 
+#include <linux/delay.h>
+#include <linux/kthread.h>
+#include <linux/pagemap.h>
+
 #include "ctree.h"
 #include "disk-io.h"
+#include "free-space-cache.h"
+#include "inode-map.h"
 #include "transaction.h"
 
-int btrfs_find_highest_inode(struct btrfs_root *root, u64 *objectid)
+static int caching_kthread(void *data)
+{
+	struct btrfs_root *root = data;
+	struct btrfs_fs_info *fs_info = root->fs_info;
+	struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
+	struct btrfs_key key;
+	struct btrfs_path *path;
+	struct extent_buffer *leaf;
+	u64 last = (u64)-1;
+	int slot;
+	int ret;
+
+	path = btrfs_alloc_path();
+	if (!path)
+		return -ENOMEM;
+
+	/* Since the commit root is read-only, we can safely skip locking. */
+	path->skip_locking = 1;
+	path->search_commit_root = 1;
+	path->reada = 2;
+
+	key.objectid = BTRFS_FIRST_FREE_OBJECTID;
+	key.offset = 0;
+	key.type = BTRFS_INODE_ITEM_KEY;
+again:
+	/* need to make sure the commit_root doesn't disappear */
+	mutex_lock(&root->fs_commit_mutex);
+
+	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
+	if (ret < 0)
+		goto out;
+
+	while (1) {
+		smp_mb();
+		if (fs_info->closing > 1)
+			goto out;
+
+		leaf = path->nodes[0];
+		slot = path->slots[0];
+		if (path->slots[0] >= btrfs_header_nritems(leaf)) {
+			ret = btrfs_next_leaf(root, path);
+			if (ret < 0)
+				goto out;
+			else if (ret > 0)
+				break;
+
+			if (need_resched() ||
+			    btrfs_transaction_in_commit(fs_info)) {
+				leaf = path->nodes[0];
+
+				if (btrfs_header_nritems(leaf) == 0) {
+					WARN_ON(1);
+					break;
+				}
+
+				/*
+				 * Save the key so we can advances forward
+				 * in the next search.
+				 */
+				btrfs_item_key_to_cpu(leaf, &key, 0);
+				btrfs_release_path(root, path);
+				root->cache_progress = last;
+				mutex_unlock(&root->fs_commit_mutex);
+				schedule_timeout(1);
+				goto again;
+			} else
+				continue;
+		}
+
+		btrfs_item_key_to_cpu(leaf, &key, slot);
+
+		if (key.type != BTRFS_INODE_ITEM_KEY)
+			goto next;
+
+		if (key.objectid >= BTRFS_LAST_FREE_OBJECTID)
+			break;
+
+		if (last != (u64)-1 && last + 1 != key.objectid) {
+			__btrfs_add_free_space(ctl, last + 1,
+					       key.objectid - last - 1);
+			wake_up(&root->cache_wait);
+		}
+
+		last = key.objectid;
+next:
+		path->slots[0]++;
+	}
+
+	if (last < BTRFS_LAST_FREE_OBJECTID - 1) {
+		__btrfs_add_free_space(ctl, last + 1,
+				       BTRFS_LAST_FREE_OBJECTID - last - 1);
+	}
+
+	spin_lock(&root->cache_lock);
+	root->cached = BTRFS_CACHE_FINISHED;
+	spin_unlock(&root->cache_lock);
+
+	root->cache_progress = (u64)-1;
+	btrfs_unpin_free_ino(root);
+out:
+	wake_up(&root->cache_wait);
+	mutex_unlock(&root->fs_commit_mutex);
+
+	btrfs_free_path(path);
+
+	return ret;
+}
+
+static void start_caching(struct btrfs_root *root)
+{
+	struct task_struct *tsk;
+
+	spin_lock(&root->cache_lock);
+	if (root->cached != BTRFS_CACHE_NO) {
+		spin_unlock(&root->cache_lock);
+		return;
+	}
+
+	root->cached = BTRFS_CACHE_STARTED;
+	spin_unlock(&root->cache_lock);
+
+	tsk = kthread_run(caching_kthread, root, "btrfs-ino-cache-%llu\n",
+			  root->root_key.objectid);
+	BUG_ON(IS_ERR(tsk));
+}
+
+int btrfs_find_free_ino(struct btrfs_root *root, u64 *objectid)
+{
+again:
+	*objectid = btrfs_find_ino_for_alloc(root);
+
+	if (*objectid != 0)
+		return 0;
+
+	start_caching(root);
+
+	wait_event(root->cache_wait,
+		   root->cached == BTRFS_CACHE_FINISHED ||
+		   root->free_ino_ctl->free_space > 0);
+
+	if (root->cached == BTRFS_CACHE_FINISHED &&
+	    root->free_ino_ctl->free_space == 0)
+		return -ENOSPC;
+	else
+		goto again;
+}
+
+void btrfs_return_ino(struct btrfs_root *root, u64 objectid)
+{
+	struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
+	struct btrfs_free_space_ctl *pinned = root->free_ino_pinned;
+again:
+	if (root->cached == BTRFS_CACHE_FINISHED) {
+		__btrfs_add_free_space(ctl, objectid, 1);
+	} else {
+		/*
+		 * If we are in the process of caching free ino chunks,
+		 * to avoid adding the same inode number to the free_ino
+		 * tree twice due to cross transaction, we'll leave it
+		 * in the pinned tree until a transaction is committed
+		 * or the caching work is done.
+		 */
+
+		mutex_lock(&root->fs_commit_mutex);
+		spin_lock(&root->cache_lock);
+		if (root->cached == BTRFS_CACHE_FINISHED) {
+			spin_unlock(&root->cache_lock);
+			mutex_unlock(&root->fs_commit_mutex);
+			goto again;
+		}
+		spin_unlock(&root->cache_lock);
+
+		start_caching(root);
+
+		if (objectid <= root->cache_progress)
+			__btrfs_add_free_space(ctl, objectid, 1);
+		else
+			__btrfs_add_free_space(pinned, objectid, 1);
+
+		mutex_unlock(&root->fs_commit_mutex);
+	}
+}
+
+/*
+ * When a transaction is committed, we'll move those inode numbers which
+ * are smaller than root->cache_progress from pinned tree to free_ino tree,
+ * and others will just be dropped, because the commit root we were
+ * searching has changed.
+ *
+ * Must be called with root->fs_commit_mutex held
+ */
+void btrfs_unpin_free_ino(struct btrfs_root *root)
+{
+	struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
+	struct rb_root *rbroot = &root->free_ino_pinned->free_space_offset;
+	struct btrfs_free_space *info;
+	struct rb_node *n;
+	u64 count;
+
+	while (1) {
+		n = rb_first(rbroot);
+		if (!n)
+			break;
+
+		info = rb_entry(n, struct btrfs_free_space, offset_index);
+		BUG_ON(info->bitmap);
+
+		if (info->offset > root->cache_progress)
+			goto free;
+		else if (info->offset + info->bytes > root->cache_progress)
+			count = root->cache_progress - info->offset + 1;
+		else
+			count = info->bytes;
+
+		__btrfs_add_free_space(ctl, info->offset, count);
+free:
+		rb_erase(&info->offset_index, rbroot);
+		kfree(info);
+	}
+}
+
+#define INIT_THRESHOLD	(((1024 * 32) / 2) / sizeof(struct btrfs_free_space))
+#define INODES_PER_BITMAP (PAGE_CACHE_SIZE * 8)
+
+/*
+ * The goal is to keep the memory used by the free_ino tree won't
+ * exceed the memory if we use bitmaps only.
+ */
+static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
+{
+	struct btrfs_free_space *info;
+	struct rb_node *n;
+	int max_ino;
+	int max_bitmaps;
+
+	n = rb_last(&ctl->free_space_offset);
+	if (!n) {
+		ctl->extents_thresh = INIT_THRESHOLD;
+		return;
+	}
+	info = rb_entry(n, struct btrfs_free_space, offset_index);
+
+	/*
+	 * Find the maximum inode number in the filesystem. Note we
+	 * ignore the fact that this can be a bitmap, because we are
+	 * not doing precise calculation.
+	 */
+	max_ino = info->bytes - 1;
+
+	max_bitmaps = ALIGN(max_ino, INODES_PER_BITMAP) / INODES_PER_BITMAP;
+	if (max_bitmaps <= ctl->total_bitmaps) {
+		ctl->extents_thresh = 0;
+		return;
+	}
+
+	ctl->extents_thresh = (max_bitmaps - ctl->total_bitmaps) *
+				PAGE_CACHE_SIZE / sizeof(*info);
+}
+
+/*
+ * We don't fall back to bitmap, if we are below the extents threshold
+ * or this chunk of inode numbers is a big one.
+ */
+static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
+		       struct btrfs_free_space *info)
+{
+	if (ctl->free_extents < ctl->extents_thresh ||
+	    info->bytes > INODES_PER_BITMAP / 10)
+		return false;
+
+	return true;
+}
+
+static struct btrfs_free_space_op free_ino_op = {
+	.recalc_thresholds	= recalculate_thresholds,
+	.use_bitmap		= use_bitmap,
+};
+
+static void pinned_recalc_thresholds(struct btrfs_free_space_ctl *ctl)
+{
+}
+
+static bool pinned_use_bitmap(struct btrfs_free_space_ctl *ctl,
+			      struct btrfs_free_space *info)
+{
+	/*
+	 * We always use extents for two reasons:
+	 *
+	 * - The pinned tree is only used during the process of caching
+	 *   work.
+	 * - Make code simpler. See btrfs_unpin_free_ino().
+	 */
+	return false;
+}
+
+static struct btrfs_free_space_op pinned_free_ino_op = {
+	.recalc_thresholds	= pinned_recalc_thresholds,
+	.use_bitmap		= pinned_use_bitmap,
+};
+
+void btrfs_init_free_ino_ctl(struct btrfs_root *root)
+{
+	struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
+	struct btrfs_free_space_ctl *pinned = root->free_ino_pinned;
+
+	spin_lock_init(&ctl->tree_lock);
+	ctl->unit = 1;
+	ctl->start = 0;
+	ctl->private = NULL;
+	ctl->op = &free_ino_op;
+
+	/*
+	 * Initially we allow to use 16K of ram to cache chunks of
+	 * inode numbers before we resort to bitmaps. This is somewhat
+	 * arbitrary, but it will be adjusted in runtime.
+	 */
+	ctl->extents_thresh = INIT_THRESHOLD;
+
+	spin_lock_init(&pinned->tree_lock);
+	pinned->unit = 1;
+	pinned->start = 0;
+	pinned->private = NULL;
+	pinned->extents_thresh = 0;
+	pinned->op = &pinned_free_ino_op;
+}
+
+static int btrfs_find_highest_objectid(struct btrfs_root *root, u64 *objectid)
 {
 	struct btrfs_path *path;
 	int ret;
@@ -55,15 +387,14 @@
 	return ret;
 }
 
-int btrfs_find_free_objectid(struct btrfs_trans_handle *trans,
-			     struct btrfs_root *root,
-			     u64 dirid, u64 *objectid)
+int btrfs_find_free_objectid(struct btrfs_root *root, u64 *objectid)
 {
 	int ret;
 	mutex_lock(&root->objectid_mutex);
 
 	if (unlikely(root->highest_objectid < BTRFS_FIRST_FREE_OBJECTID)) {
-		ret = btrfs_find_highest_inode(root, &root->highest_objectid);
+		ret = btrfs_find_highest_objectid(root,
+						  &root->highest_objectid);
 		if (ret)
 			goto out;
 	}