bcachefs: Initial commit

Initially forked from drivers/md/bcache, bcachefs is a new copy-on-write
filesystem with every feature you could possibly want.

Website: https://bcachefs.org

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

1 files changed, 113 insertions, 0 deletions

diff --git a/fs/bcachefs/btree_gc.h b/fs/bcachefs/btree_gc.h
new file mode 100644
index 000000000000..9d2b9d5953d2
--- /dev/null
+++ b/fs/bcachefs/btree_gc.h
@@ -0,0 +1,113 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_BTREE_GC_H
+#define _BCACHEFS_BTREE_GC_H
+
+#include "btree_types.h"
+
+enum bkey_type;
+
+void bch2_coalesce(struct bch_fs *);
+void bch2_gc(struct bch_fs *);
+void bch2_gc_thread_stop(struct bch_fs *);
+int bch2_gc_thread_start(struct bch_fs *);
+int bch2_initial_gc(struct bch_fs *, struct list_head *);
+u8 bch2_btree_key_recalc_oldest_gen(struct bch_fs *, struct bkey_s_c);
+int bch2_btree_mark_key_initial(struct bch_fs *, enum bkey_type,
+				struct bkey_s_c);
+void bch2_mark_dev_superblock(struct bch_fs *, struct bch_dev *, unsigned);
+
+/*
+ * For concurrent mark and sweep (with other index updates), we define a total
+ * ordering of _all_ references GC walks:
+ *
+ * Note that some references will have the same GC position as others - e.g.
+ * everything within the same btree node; in those cases we're relying on
+ * whatever locking exists for where those references live, i.e. the write lock
+ * on a btree node.
+ *
+ * That locking is also required to ensure GC doesn't pass the updater in
+ * between the updater adding/removing the reference and updating the GC marks;
+ * without that, we would at best double count sometimes.
+ *
+ * That part is important - whenever calling bch2_mark_pointers(), a lock _must_
+ * be held that prevents GC from passing the position the updater is at.
+ *
+ * (What about the start of gc, when we're clearing all the marks? GC clears the
+ * mark with the gc pos seqlock held, and bch_mark_bucket checks against the gc
+ * position inside its cmpxchg loop, so crap magically works).
+ */
+
+/* Position of (the start of) a gc phase: */
+static inline struct gc_pos gc_phase(enum gc_phase phase)
+{
+	return (struct gc_pos) {
+		.phase	= phase,
+		.pos	= POS_MIN,
+		.level	= 0,
+	};
+}
+
+static inline int gc_pos_cmp(struct gc_pos l, struct gc_pos r)
+{
+	if (l.phase != r.phase)
+		return l.phase < r.phase ? -1 : 1;
+	if (bkey_cmp(l.pos, r.pos))
+		return bkey_cmp(l.pos, r.pos);
+	if (l.level != r.level)
+		return l.level < r.level ? -1 : 1;
+	return 0;
+}
+
+static inline struct gc_pos gc_pos_btree(enum btree_id id,
+					 struct bpos pos, unsigned level)
+{
+	return (struct gc_pos) {
+		.phase	= GC_PHASE_BTREE_EXTENTS + id,
+		.pos	= pos,
+		.level	= level,
+	};
+}
+
+/*
+ * GC position of the pointers within a btree node: note, _not_ for &b->key
+ * itself, that lives in the parent node:
+ */
+static inline struct gc_pos gc_pos_btree_node(struct btree *b)
+{
+	return gc_pos_btree(b->btree_id, b->key.k.p, b->level);
+}
+
+/*
+ * GC position of the pointer to a btree root: we don't use
+ * gc_pos_pointer_to_btree_node() here to avoid a potential race with
+ * btree_split() increasing the tree depth - the new root will have level > the
+ * old root and thus have a greater gc position than the old root, but that
+ * would be incorrect since once gc has marked the root it's not coming back.
+ */
+static inline struct gc_pos gc_pos_btree_root(enum btree_id id)
+{
+	return gc_pos_btree(id, POS_MAX, BTREE_MAX_DEPTH);
+}
+
+static inline struct gc_pos gc_pos_alloc(struct bch_fs *c, struct open_bucket *ob)
+{
+	return (struct gc_pos) {
+		.phase	= GC_PHASE_ALLOC,
+		.pos	= POS(ob ? ob - c->open_buckets : 0, 0),
+	};
+}
+
+static inline bool gc_will_visit(struct bch_fs *c, struct gc_pos pos)
+{
+	unsigned seq;
+	bool ret;
+
+	do {
+		seq = read_seqcount_begin(&c->gc_pos_lock);
+		ret = gc_pos_cmp(c->gc_pos, pos) < 0;
+	} while (read_seqcount_retry(&c->gc_pos_lock, seq));
+
+	return ret;
+}
+
+#endif /* _BCACHEFS_BTREE_GC_H */