[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

[dm-devel] [Bcache v15 11/16] bcache: Core btree code



Signed-off-by: Kent Overstreet <koverstreet google com>
---
 drivers/md/bcache/bcache.h | 1142 ++++++++++++++++++++
 drivers/md/bcache/btree.c  | 2508 ++++++++++++++++++++++++++++++++++++++++++++
 drivers/md/bcache/btree.h  |  423 ++++++++
 3 files changed, 4073 insertions(+), 0 deletions(-)
 create mode 100644 drivers/md/bcache/bcache.h
 create mode 100644 drivers/md/bcache/btree.c
 create mode 100644 drivers/md/bcache/btree.h

diff --git a/drivers/md/bcache/bcache.h b/drivers/md/bcache/bcache.h
new file mode 100644
index 0000000..462d0ea
--- /dev/null
+++ b/drivers/md/bcache/bcache.h
@@ -0,0 +1,1142 @@
+#ifndef _BCACHE_H
+#define _BCACHE_H
+
+/*
+ * SOME HIGH LEVEL CODE DOCUMENTATION:
+ *
+ * Bcache mostly works with cache sets, cache devices, and backing devices.
+ *
+ * Support for multiple cache devices hasn't quite been finished off yet, but
+ * it's about 95% plumbed through. A cache set and its cache devices is sort of
+ * like a md raid array and its component devices. Most of the code doesn't care
+ * about individual cache devices, the main abstraction is the cache set.
+ *
+ * Multiple cache devices is intended to give us the ability to mirror dirty
+ * cached data and metadata, without mirroring clean cached data.
+ *
+ * Backing devices are different, in that they have a lifetime independent of a
+ * cache set. When you register a newly formatted backing device it'll come up
+ * in passthrough mode, and then you can attach and detach a backing device from
+ * a cache set at runtime - while it's mounted and in use. Detaching implicitly
+ * invalidates any cached data for that backing device.
+ *
+ * A cache set can have multiple (many) backing devices attached to it.
+ *
+ * There's also flash only volumes - this is the reason for the distinction
+ * between struct cached_dev and struct bcache_device. A flash only volume
+ * works much like a bcache device that has a backing device, except the
+ * "cached" data is always dirty. The end result is that we get thin
+ * provisioning with very little additional code.
+ *
+ * Flash only volumes work but they're not production ready because the moving
+ * garbage collector needs more work. More on that later.
+ *
+ * BUCKETS/ALLOCATION:
+ *
+ * Bcache is primarily designed for caching, which means that in normal
+ * operation all of our available space will be allocated. Thus, we need an
+ * efficient way of deleting things from the cache so we can write new things to
+ * it.
+ *
+ * To do this, we first divide the cache device up into buckets. A bucket is the
+ * unit of allocation; they're typically around 1 mb - anywhere from 128k to 2M+
+ * works efficiently.
+ *
+ * Each bucket has a 16 bit priority, and an 8 bit generation associated with
+ * it. The gens and priorities for all the buckets are stored contiguously and
+ * packed on disk (in a linked list of buckets - aside from the superblock, all
+ * of bcache's metadata is stored in buckets).
+ *
+ * The priority is used to implement an LRU. We reset a bucket's priority when
+ * we allocate it or on cache it, and every so often we decrement the priority
+ * of each bucket. It could be used to implement something more sophisticated,
+ * if anyone ever gets around to it.
+ *
+ * The generation is used for invalidating buckets. Each pointer also has an 8
+ * bit generation embedded in it; for a pointer to be considered valid, its gen
+ * must match the gen of the bucket it points into.  Thus, to reuse a bucket all
+ * we have to do is increment its gen (and write its new gen to disk; we batch
+ * this up).
+ *
+ * Bcache is entirely COW - we never write twice to a bucket, even buckets that
+ * contain metadata (including btree nodes).
+ *
+ * THE BTREE:
+ *
+ * Bcache is in large part design around the btree.
+ *
+ * At a high level, the btree is just an index of key -> ptr tuples.
+ *
+ * Keys represent extents, and thus have a size field. Keys also have a variable
+ * number of pointers attached to them (potentially zero, which is handy for
+ * invalidating the cache).
+ *
+ * The key itself is an inode:offset pair. The inode number corresponds to a
+ * backing device or a flash only volume. The offset is the ending offset of the
+ * extent within the inode - not the starting offset; this makes lookups
+ * slightly more convenient.
+ *
+ * Pointers contain the cache device id, the offset on that device, and an 8 bit
+ * generation number. More on the gen later.
+ *
+ * Index lookups are not fully abstracted - cache lookups in particular are
+ * still somewhat mixed in with the btree code, but things are headed in that
+ * direction.
+ *
+ * Updates are fairly well abstracted, though. There are two different ways of
+ * updating the btree; insert and replace.
+ *
+ * BTREE_INSERT will just take a list of keys and insert them into the btree -
+ * overwriting (possibly only partially) any extents they overlap with. This is
+ * used to update the index after a write.
+ *
+ * BTREE_REPLACE is really cmpxchg(); it inserts a key into the btree iff it is
+ * overwriting a key that matches another given key. This is used for inserting
+ * data into the cache after a cache miss, and for background writeback, and for
+ * the moving garbage collector.
+ *
+ * There is no "delete" operation; deleting things from the index is
+ * accomplished by either by invalidating pointers (by incrementing a bucket's
+ * gen) or by inserting a key with 0 pointers - which will overwrite anything
+ * previously present at that location in the index.
+ *
+ * This means that there are always stale/invalid keys in the btree. They're
+ * filtered out by the code that iterates through a btree node, and removed when
+ * a btree node is rewritten.
+ *
+ * BTREE NODES:
+ *
+ * Our unit of allocation is a bucket, and we we can't arbitrarily allocate and
+ * free smaller than a bucket - so, that's how big our btree nodes are.
+ *
+ * (If buckets are really big we'll only use part of the bucket for a btree node
+ * - no less than 1/4th - but a bucket still contains no more than a single
+ * btree node. I'd actually like to change this, but for now we rely on the
+ * bucket's gen for deleting btree nodes when we rewrite/split a node.)
+ *
+ * Anyways, btree nodes are big - big enough to be inefficient with a textbook
+ * btree implementation.
+ *
+ * The way this is solved is that btree nodes are internally log structured; we
+ * can append new keys to an existing btree node without rewriting it. This
+ * means each set of keys we write is sorted, but the node is not.
+ *
+ * We maintain this log structure in memory - keeping 1Mb of keys sorted would
+ * be expensive, and we have to distinguish between the keys we have written and
+ * the keys we haven't. So to do a lookup in a btree node, we have to search
+ * each sorted set. But we do merge written sets together lazily, so the cost of
+ * these extra searches is quite low (normally most of the keys in a btree node
+ * will be in one big set, and then there'll be one or two sets that are much
+ * smaller).
+ *
+ * This log structure makes bcache's btree more of a hybrid between a
+ * conventional btree and a compacting data structure, with some of the
+ * advantages of both.
+ *
+ * GARBAGE COLLECTION:
+ *
+ * We can't just invalidate any bucket - it might contain dirty data or
+ * metadata. If it once contained dirty data, other writes might overwrite it
+ * later, leaving no valid pointers into that bucket in the index.
+ *
+ * Thus, the primary purpose of garbage collection is to find buckets to reuse.
+ * It also counts how much valid data it each bucket currently contains, so that
+ * allocation can reuse buckets sooner when they've been mostly overwritten.
+ *
+ * It also does some things that are really internal to the btree
+ * implementation. If a btree node contains pointers that are stale by more than
+ * some threshold, it rewrites the btree node to avoid the bucket's generation
+ * wrapping around. It also merges adjacent btree nodes if they're empty enough.
+ *
+ * THE JOURNAL:
+ *
+ * Bcache's journal is not necessary for consistency; we always strictly
+ * order metadata writes so that the btree and everything else is consistent on
+ * disk in the event of an unclean shutdown, and in fact bcache had writeback
+ * caching (with recovery from unclean shutdown) before journalling was
+ * implemented.
+ *
+ * Rather, the journal is purely a performance optimization; we can't complete a
+ * write until we've updated the index on disk, otherwise the cache would be
+ * inconsistent in the event of an unclean shutdown. This means that without the
+ * journal, on random write workloads we constantly have to update all the leaf
+ * nodes in the btree, and those writes will be mostly empty (appending at most
+ * a few keys each) - highly inefficient in terms of amount of metadata writes,
+ * and it puts more strain on the various btree resorting/compacting code.
+ *
+ * The journal is just a log of keys we've inserted; on startup we just reinsert
+ * all the keys in the open journal entries. That means that when we're updating
+ * a node in the btree, we can wait until a 4k block of keys fills up before
+ * writing them out.
+ *
+ * For simplicity, we only journal updates to leaf nodes; updates to parent
+ * nodes are rare enough (since our leaf nodes are huge) that it wasn't worth
+ * the complexity to deal with journalling them (in particular, journal replay)
+ * - updates to non leaf nodes just happen synchronously (see btree_split()).
+ */
+
+#define pr_fmt(fmt) "bcache: %s() " fmt "\n", __func__
+
+#include <linux/bio.h>
+#include <linux/blktrace_api.h>
+#include <linux/closure.h>
+#include <linux/kobject.h>
+#include <linux/list.h>
+#include <linux/mutex.h>
+#include <linux/rbtree.h>
+#include <linux/rwsem.h>
+#include <linux/types.h>
+#include <linux/workqueue.h>
+
+#include "util.h"
+
+struct bucket {
+	atomic_t	pin;
+	uint16_t	prio;
+	uint8_t		gen;
+	uint8_t		disk_gen;
+	uint8_t		last_gc; /* Most out of date gen in the btree */
+	uint8_t		gc_gen;
+	uint16_t	gc_mark;
+};
+
+/*
+ * I'd use bitfields for these, but I don't trust the compiler not to screw me
+ * as multiple threads touch struct bucket without locking
+ */
+
+BITMASK(GC_MARK,	 struct bucket, gc_mark, 0, 2);
+#define GC_MARK_RECLAIMABLE	0
+#define GC_MARK_DIRTY		1
+#define GC_MARK_BTREE		2
+BITMASK(GC_SECTORS_USED, struct bucket, gc_mark, 2, 14);
+
+struct bkey {
+	uint64_t	high;
+	uint64_t	low;
+	uint64_t	ptr[];
+};
+
+/* Enough for a key with 6 pointers */
+#define BKEY_PAD		8
+
+#define BKEY_PADDED(key)					\
+	union { struct bkey key; uint64_t key ## _pad[BKEY_PAD]; }
+
+/* Version 1: Backing device
+ * Version 2: Seed pointer into btree node checksum
+ * Version 3: New UUID format
+ */
+#define BCACHE_SB_VERSION	3
+
+#define SB_SECTOR		8
+#define SB_SIZE			4096
+#define SB_LABEL_SIZE		32
+#define SB_JOURNAL_BUCKETS	256
+/* SB_JOURNAL_BUCKETS must be divisible by BITS_PER_LONG */
+#define MAX_CACHES_PER_SET	8
+
+#define BDEV_DATA_START		16	/* sectors */
+
+struct cache_sb {
+	uint64_t		csum;
+	uint64_t		offset;	/* sector where this sb was written */
+	uint64_t		version;
+#define CACHE_BACKING_DEV	1
+
+	uint8_t			magic[16];
+
+	uint8_t			uuid[16];
+	union {
+		uint8_t		set_uuid[16];
+		uint64_t	set_magic;
+	};
+	uint8_t			label[SB_LABEL_SIZE];
+
+	uint64_t		flags;
+	uint64_t		seq;
+	uint64_t		pad[8];
+
+	uint64_t		nbuckets;	/* device size */
+	uint16_t		block_size;	/* sectors */
+	uint16_t		bucket_size;	/* sectors */
+
+	uint16_t		nr_in_set;
+	uint16_t		nr_this_dev;
+
+	uint32_t		last_mount;	/* time_t */
+
+	uint16_t		first_bucket;
+	union {
+		uint16_t	njournal_buckets;
+		uint16_t	keys;
+	};
+	uint64_t		d[SB_JOURNAL_BUCKETS];	/* journal buckets */
+};
+
+BITMASK(CACHE_SYNC,		struct cache_sb, flags, 0, 1);
+BITMASK(CACHE_DISCARD,		struct cache_sb, flags, 1, 1);
+BITMASK(CACHE_REPLACEMENT,	struct cache_sb, flags, 2, 3);
+#define CACHE_REPLACEMENT_LRU	0U
+#define CACHE_REPLACEMENT_FIFO	1U
+#define CACHE_REPLACEMENT_RANDOM 2U
+
+BITMASK(BDEV_CACHE_MODE,	struct cache_sb, flags, 0, 4);
+#define CACHE_MODE_WRITETHROUGH	0U
+#define CACHE_MODE_WRITEBACK	1U
+#define CACHE_MODE_WRITEAROUND	2U
+#define CACHE_MODE_NONE		3U
+BITMASK(BDEV_STATE,		struct cache_sb, flags, 61, 2);
+#define BDEV_STATE_NONE		0U
+#define BDEV_STATE_CLEAN	1U
+#define BDEV_STATE_DIRTY	2U
+#define BDEV_STATE_STALE	3U
+
+/* Version 1: Seed pointer into btree node checksum
+ */
+#define BCACHE_BSET_VERSION	1
+
+/*
+ * This is the on disk format for btree nodes - a btree node on disk is a list
+ * of these; within each set the keys are sorted
+ */
+struct bset {
+	uint64_t		csum;
+	uint64_t		magic;
+	uint64_t		seq;
+	uint32_t		version;
+	uint32_t		keys;
+
+	union {
+		struct bkey	start[0];
+		uint64_t	d[0];
+	};
+};
+
+/*
+ * On disk format for priorities and gens - see super.c near prio_write() for
+ * more.
+ */
+struct prio_set {
+	uint64_t		csum;
+	uint64_t		magic;
+	uint64_t		seq;
+	uint32_t		version;
+	uint32_t		pad;
+
+	uint64_t		next_bucket;
+
+	struct bucket_disk {
+		uint16_t	prio;
+		uint8_t		gen;
+	} __attribute((packed)) data[];
+};
+
+#include "journal.h"
+#include "stats.h"
+struct search;
+struct btree;
+struct keybuf;
+
+struct keybuf_key {
+	struct rb_node		node;
+	BKEY_PADDED(key);
+	void			*private;
+};
+
+typedef bool (keybuf_pred_fn)(struct keybuf *, struct bkey *);
+
+struct keybuf {
+	keybuf_pred_fn		*key_predicate;
+
+	struct bkey		last_scanned;
+	spinlock_t		lock;
+
+	/*
+	 * Beginning and end of range in rb tree - so that we can skip taking
+	 * lock and checking the rb tree when we need to check for overlapping
+	 * keys.
+	 */
+	struct bkey		start;
+	struct bkey		end;
+
+	struct rb_root		keys;
+
+#define KEYBUF_NR		100
+	DECLARE_ARRAY_ALLOCATOR(struct keybuf_key, freelist, KEYBUF_NR);
+};
+
+struct bcache_device {
+	struct closure		cl;
+
+	struct kobject		kobj;
+
+	struct cache_set	*c;
+	unsigned		id;
+#define BCACHEDEVNAME_SIZE	12
+	char			name[BCACHEDEVNAME_SIZE];
+
+	struct gendisk		*disk;
+
+	/* If nonzero, we're closing */
+	atomic_t		closing;
+
+	/* If nonzero, we're detaching/unregistering from cache set */
+	atomic_t		detaching;
+
+	atomic_long_t		sectors_dirty;
+	unsigned long		sectors_dirty_gc;
+	unsigned long		sectors_dirty_last;
+	long			sectors_dirty_derivative;
+
+	mempool_t		*unaligned_bvec;
+	struct bio_set		*bio_split;
+
+	unsigned		data_csum:1;
+
+	int (*cache_miss)(struct btree *, struct search *, struct bio *, unsigned);
+	int (*ioctl) (struct bcache_device *, fmode_t, unsigned, unsigned long);
+};
+
+struct io {
+	/* Used to track sequential IO so it can be skipped */
+	struct hlist_node	hash;
+	struct list_head	lru;
+
+	unsigned long		jiffies;
+	unsigned		sequential;
+	sector_t		last;
+};
+
+struct cached_dev {
+	struct list_head	list;
+	struct bcache_device	disk;
+	struct block_device	*bdev;
+
+	struct cache_sb		sb;
+	struct bio		sb_bio;
+	struct bio_vec		sb_bv[1];
+	struct closure_with_waitlist sb_write;
+
+	/* Refcount on the cache set. Always nonzero when we're caching. */
+	atomic_t		count;
+	struct work_struct	detach;
+
+	/*
+	 * Device might not be running if it's dirty and the cache set hasn't
+	 * showed up yet.
+	 */
+	atomic_t		running;
+
+	/*
+	 * Writes take a shared lock from start to finish; scanning for dirty
+	 * data to refill the rb tree requires an exclusive lock.
+	 */
+	struct rw_semaphore	writeback_lock;
+
+	/*
+	 * Nonzero, and writeback has a refcount (d->count), iff there is dirty
+	 * data in the cache. Protected by writeback_lock; must have an
+	 * shared lock to set and exclusive lock to clear.
+	 */
+	atomic_t		has_dirty;
+
+	struct ratelimit	writeback_rate;
+	struct delayed_work	writeback_rate_update;
+
+	/*
+	 * Internal to the writeback code, so read_dirty() can keep track of
+	 * where it's at.
+	 */
+	sector_t		last_read;
+
+	/* Number of writeback bios in flight */
+	atomic_t		in_flight;
+	struct closure_with_timer writeback;
+	struct closure_waitlist	writeback_wait;
+
+	struct keybuf		writeback_keys;
+
+	/* For tracking sequential IO */
+#define RECENT_IO_BITS	7
+#define RECENT_IO	(1 << RECENT_IO_BITS)
+	struct io		io[RECENT_IO];
+	struct hlist_head	io_hash[RECENT_IO + 1];
+	struct list_head	io_lru;
+	spinlock_t		io_lock;
+
+	struct cache_accounting	accounting;
+
+	/* The rest of this all shows up in sysfs */
+	unsigned		sequential_cutoff;
+	unsigned		readahead;
+
+	unsigned		sequential_merge:1;
+	unsigned		verify:1;
+
+	unsigned		writeback_metadata:1;
+	unsigned		writeback_running:1;
+	unsigned char		writeback_percent;
+	unsigned		writeback_delay;
+
+	int			writeback_rate_change;
+	int64_t			writeback_rate_derivative;
+	uint64_t		writeback_rate_target;
+
+	unsigned		writeback_rate_update_seconds;
+	unsigned		writeback_rate_d_term;
+	unsigned		writeback_rate_p_term_inverse;
+	unsigned		writeback_rate_d_smooth;
+};
+
+struct cache {
+	struct cache_set	*set;
+	struct cache_sb		sb;
+	struct bio		sb_bio;
+	struct bio_vec		sb_bv[1];
+
+	struct kobject		kobj;
+	struct block_device	*bdev;
+
+	struct closure		prio;
+	struct prio_set		*disk_buckets;
+
+	/*
+	 * When allocating new buckets, prio_write() gets first dibs - since we
+	 * may not be allocate at all without writing priorities and gens.
+	 * prio_buckets[] contains the last buckets we wrote priorities to (so
+	 * gc can mark them as metadata), prio_next[] contains the buckets
+	 * allocated for the next prio write.
+	 */
+	uint64_t		*prio_buckets;
+	uint64_t		*prio_next;
+	unsigned		prio_write;
+	unsigned		prio_alloc;
+
+	/* > 0: buckets in free_inc have been marked as free
+	 * = 0: buckets in free_inc can't be used until priorities are written
+	 * < 0: priority write in progress
+	 */
+	atomic_t		prio_written;
+
+	/*
+	 * free: Buckets that are ready to be used
+	 *
+	 * free_inc: Incoming buckets - these are buckets that currently have
+	 * cached data in them, and we can't reuse them until after we write
+	 * their new gen to disk. After prio_write() finishes writing the new
+	 * gens/prios, they'll be moved to the free list (and possibly discarded
+	 * in the process)
+	 *
+	 * unused: GC found nothing pointing into these buckets (possibly
+	 * because all the data they contained was overwritten), so we only
+	 * need to discard them before they can be moved to the free list.
+	 */
+	DECLARE_FIFO(long, free);
+	DECLARE_FIFO(long, free_inc);
+	DECLARE_FIFO(long, unused);
+
+	size_t			fifo_last_bucket;
+
+	/* Allocation stuff: */
+	struct bucket		*buckets;
+
+	DECLARE_HEAP(struct bucket *, heap);
+
+	/*
+	 * max(gen - disk_gen) for all buckets. When it gets too big we have to
+	 * call prio_write() to keep gens from wrapping.
+	 */
+	uint8_t			need_save_prio;
+	unsigned		gc_move_threshold;
+
+	/*
+	 * If nonzero, we know we aren't going to find any buckets to invalidate
+	 * until a gc finishes - otherwise we could pointlessly burn a ton of
+	 * cpu
+	 */
+	unsigned		invalidate_needs_gc:1;
+
+	bool			discard; /* Get rid of? */
+
+	/*
+	 * We preallocate structs for issuing discards to buckets, and keep them
+	 * on this list when they're not in use; do_discard() issues discards
+	 * whenever there's work to do and is called by free_some_buckets() and
+	 * when a discard finishes.
+	 */
+	struct list_head	discards;
+	struct page		*discard_page;
+
+	struct journal_device	journal;
+
+	/* The rest of this all shows up in sysfs */
+#define IO_ERROR_SHIFT		20
+	atomic_t		io_errors;
+	atomic_t		io_count;
+
+	atomic_long_t		meta_sectors_written;
+	atomic_long_t		btree_sectors_written;
+	atomic_long_t		sectors_written;
+};
+
+struct gc_stat {
+	size_t			nodes;
+	size_t			key_bytes;
+
+	size_t			nkeys;
+	uint64_t		data;	/* sectors */
+	uint64_t		dirty;	/* sectors */
+	unsigned		in_use; /* percent */
+};
+
+struct cache_set {
+	struct closure		cl;
+
+	struct list_head	list;
+	struct kobject		kobj;
+	struct kobject		internal;
+	struct dentry		*debug;
+	struct cache_accounting accounting;
+
+	/*
+	 * If nonzero, we're trying to detach from all the devices we're
+	 * caching; otherwise we're merely closing
+	 */
+	atomic_t		unregistering;
+	atomic_t		closing;
+
+	struct cache_sb		sb;
+
+	struct cache		*cache[MAX_CACHES_PER_SET];
+	struct cache		*cache_by_alloc[MAX_CACHES_PER_SET];
+	int			caches_loaded;
+
+	struct bcache_device	**devices;
+	struct list_head	cached_devs;
+	uint64_t		cached_dev_sectors;
+	struct closure		caching;
+
+	struct closure_with_waitlist sb_write;
+
+	mempool_t		*search;
+	mempool_t		*bio_meta;
+	struct bio_set		*bio_split;
+
+	/* For the btree cache */
+	struct shrinker		shrink;
+
+	/* For the btree cache and anything allocation related */
+	struct mutex		bucket_lock;
+
+	/* log2(bucket_size), in sectors */
+	unsigned short		bucket_bits;
+
+	/* log2(block_size), in sectors */
+	unsigned short		block_bits;
+
+	/*
+	 * Default number of pages for a new btree node - may be less than a
+	 * full bucket
+	 */
+	unsigned		btree_pages;
+
+	/*
+	 * Lists of struct btrees; lru is the list for structs that have memory
+	 * allocated for actual btree node, freed is for structs that do not.
+	 *
+	 * We never free a struct btree, except on shutdown - we just put it on
+	 * the btree_cache_freed list and reuse it later. This simplifies the
+	 * code, and it doesn't cost us much memory as the memory usage is
+	 * dominated by buffers that hold the actual btree node data and those
+	 * can be freed - and the number of struct btrees allocated is
+	 * effectively bounded.
+	 *
+	 * btree_cache_freeable effectively is a small cache - we use it because
+	 * high order page allocations can be rather expensive, and it's quite
+	 * common to delete and allocate btree nodes in quick succession. It
+	 * should never grow past ~2-3 nodes in practice.
+	 */
+	struct list_head	btree_cache;
+	struct list_head	btree_cache_freeable;
+	struct list_head	btree_cache_freed;
+
+	/* Number of elements in btree_cache + btree_cache_freeable lists */
+	unsigned		bucket_cache_used;
+
+	/*
+	 * If we need to allocate memory for a new btree node and that
+	 * allocation fails, we can cannibalize another node in the btree cache
+	 * to satisfy the allocation. However, only one thread can be doing this
+	 * at a time, for obvious reasons - try_harder and try_wait are
+	 * basically a lock for this that we can wait on asynchronously. The
+	 * btree_root() macro releases the lock when it returns.
+	 */
+	struct closure		*try_harder;
+	struct closure_waitlist	try_wait;
+	uint64_t		try_harder_start;
+
+	/*
+	 * When we free a btree node, we increment the gen of the bucket the
+	 * node is in - but we can't rewrite the prios and gens until we
+	 * finished whatever it is we were doing, otherwise after a crash the
+	 * btree node would be freed but for say a split, we might not have the
+	 * pointers to the new nodes inserted into the btree yet.
+	 *
+	 * This is a refcount that blocks prio_write() until the new keys are
+	 * written.
+	 */
+	atomic_t		prio_blocked;
+	struct closure_waitlist	bucket_wait;
+
+	/*
+	 * For any bio we don't skip we subtract the number of sectors from
+	 * rescale; when it hits 0 we rescale all the bucket priorities.
+	 */
+	atomic_t		rescale;
+	/*
+	 * When we invalidate buckets, we use both the priority and the amount
+	 * of good data to determine which buckets to reuse first - to weight
+	 * those together consistently we keep track of the smallest nonzero
+	 * priority of any bucket.
+	 */
+	uint16_t		min_prio;
+
+	/*
+	 * max(gen - gc_gen) for all buckets. When it gets too big we have to gc
+	 * to keep gens from wrapping around.
+	 */
+	uint8_t			need_gc;
+	struct gc_stat		gc_stats;
+	size_t			nbuckets;
+
+	struct closure_with_waitlist gc;
+	/* Where in the btree gc currently is */
+	struct bkey		gc_done;
+
+	/*
+	 * The allocation code needs gc_mark in struct bucket to be correct, but
+	 * it's not while a gc is in progress. Protected by bucket_lock.
+	 */
+	int			gc_mark_valid;
+
+	/* Counts how many sectors bio_insert has added to the cache */
+	atomic_t		sectors_to_gc;
+
+	struct closure		moving_gc;
+	struct closure_waitlist	moving_gc_wait;
+	struct keybuf		moving_gc_keys;
+	/* Number of moving GC bios in flight */
+	atomic_t		in_flight;
+
+	struct btree		*root;
+
+#ifdef CONFIG_BCACHE_DEBUG
+	struct btree		*verify_data;
+	struct mutex		verify_lock;
+#endif
+
+	unsigned		nr_uuids;
+	struct uuid_entry	*uuids;
+	BKEY_PADDED(uuid_bucket);
+	struct closure_with_waitlist uuid_write;
+
+	/*
+	 * A btree node on disk could have too many bsets for an iterator to fit
+	 * on the stack - this is a single element mempool for btree_read_work()
+	 */
+	struct mutex		fill_lock;
+	struct btree_iter	*fill_iter;
+
+	/*
+	 * btree_sort() is a merge sort and requires temporary space - single
+	 * element mempool
+	 */
+	struct mutex		sort_lock;
+	struct bset		*sort;
+
+	/* List of buckets we're currently writing data to */
+	struct list_head	data_buckets;
+	spinlock_t		data_bucket_lock;
+
+	struct journal		journal;
+
+#define CONGESTED_MAX		1024
+	unsigned		congested_last_us;
+	atomic_t		congested;
+
+	/* The rest of this all shows up in sysfs */
+	unsigned		congested_read_threshold_us;
+	unsigned		congested_write_threshold_us;
+
+	spinlock_t		sort_time_lock;
+	struct time_stats	sort_time;
+	struct time_stats	btree_gc_time;
+	struct time_stats	btree_split_time;
+	spinlock_t		btree_read_time_lock;
+	struct time_stats	btree_read_time;
+	struct time_stats	try_harder_time;
+
+	atomic_long_t		cache_read_races;
+	atomic_long_t		writeback_keys_done;
+	atomic_long_t		writeback_keys_failed;
+	unsigned		error_limit;
+	unsigned		error_decay;
+	unsigned short		journal_delay_ms;
+	unsigned		verify:1;
+	unsigned		key_merging_disabled:1;
+	unsigned		gc_always_rewrite:1;
+	unsigned		shrinker_disabled:1;
+	unsigned		copy_gc_enabled:1;
+
+#define BUCKET_HASH_BITS	12
+	struct hlist_head	bucket_hash[1 << BUCKET_HASH_BITS];
+};
+
+static inline bool key_merging_disabled(struct cache_set *c)
+{
+#ifdef CONFIG_BCACHE_DEBUG
+	return c->key_merging_disabled;
+#else
+	return 0;
+#endif
+}
+
+struct bbio {
+	unsigned		submit_time_us;
+	union {
+		struct bkey	key;
+		uint64_t	_pad[3];
+		/*
+		 * We only need pad = 3 here because we only ever carry around a
+		 * single pointer - i.e. the pointer we're doing io to/from.
+		 */
+	};
+	struct bio		bio;
+};
+
+static inline unsigned local_clock_us(void)
+{
+	return local_clock() >> 10;
+}
+
+#define MAX_BSETS		4
+
+#define BTREE_PRIO		USHRT_MAX
+#define INITIAL_PRIO		32768
+
+#define btree_bytes(c)		((c)->btree_pages * PAGE_SIZE)
+#define btree_blocks(b)							\
+	((unsigned) (KEY_SIZE(&b->key) >> (b)->c->block_bits))
+
+#define btree_default_blocks(c)						\
+	((unsigned) ((PAGE_SECTORS * (c)->btree_pages) >> (c)->block_bits))
+
+#define bucket_pages(c)		((c)->sb.bucket_size / PAGE_SECTORS)
+#define bucket_bytes(c)		((c)->sb.bucket_size << 9)
+#define block_bytes(c)		((c)->sb.block_size << 9)
+
+#define __set_bytes(i, k)	(sizeof(*(i)) + (k) * sizeof(uint64_t))
+#define set_bytes(i)		__set_bytes(i, i->keys)
+
+#define __set_blocks(i, k, c)	DIV_ROUND_UP(__set_bytes(i, k), block_bytes(c))
+#define set_blocks(i, c)	__set_blocks(i, (i)->keys, c)
+
+#define node(i, j)		((struct bkey *) ((i)->d + (j)))
+#define end(i)			node(i, (i)->keys)
+
+#define index(i, b)							\
+	((size_t) (((void *) i - (void *) (b)->sets[0].data) /		\
+		   block_bytes(b->c)))
+
+#define btree_data_space(b)	(PAGE_SIZE << (b)->page_order)
+
+#define prios_per_bucket(c)				\
+	((bucket_bytes(c) - sizeof(struct prio_set)) /	\
+	 sizeof(struct bucket_disk))
+#define prio_buckets(c)					\
+	DIV_ROUND_UP((size_t) (c)->sb.nbuckets, prios_per_bucket(c))
+
+#define JSET_MAGIC		0x245235c1a3625032ULL
+#define PSET_MAGIC		0x6750e15f87337f91ULL
+#define BSET_MAGIC		0x90135c78b99e07f5ULL
+
+#define jset_magic(c)		((c)->sb.set_magic ^ JSET_MAGIC)
+#define pset_magic(c)		((c)->sb.set_magic ^ PSET_MAGIC)
+#define bset_magic(c)		((c)->sb.set_magic ^ BSET_MAGIC)
+
+/* Bkey fields: all units are in sectors */
+
+#define KEY_FIELD(name, field, offset, size)				\
+	BITMASK(name, struct bkey, field, offset, size)
+
+#define PTR_FIELD(name, offset, size)					\
+	static inline uint64_t name(const struct bkey *k, unsigned i)	\
+	{ return (k->ptr[i] >> offset) & ~(((uint64_t) ~0) << size); }	\
+									\
+	static inline void SET_##name(struct bkey *k, unsigned i, uint64_t v)\
+	{								\
+		k->ptr[i] &= ~(~((uint64_t) ~0 << size) << offset);	\
+		k->ptr[i] |= v << offset;				\
+	}
+
+KEY_FIELD(KEY_PTRS,	high, 60, 3)
+KEY_FIELD(HEADER_SIZE,	high, 58, 2)
+KEY_FIELD(KEY_CSUM,	high, 56, 2)
+KEY_FIELD(KEY_PINNED,	high, 55, 1)
+KEY_FIELD(KEY_DIRTY,	high, 36, 1)
+
+KEY_FIELD(KEY_SIZE,	high, 20, 16)
+KEY_FIELD(KEY_INODE,	high, 0,  20)
+
+/* Next time I change the on disk format, KEY_OFFSET() won't be 64 bits */
+
+static inline uint64_t KEY_OFFSET(const struct bkey *k)
+{
+	return k->low;
+}
+
+static inline void SET_KEY_OFFSET(struct bkey *k, uint64_t v)
+{
+	k->low = v;
+}
+
+PTR_FIELD(PTR_DEV,		51, 12)
+PTR_FIELD(PTR_OFFSET,		8,  43)
+PTR_FIELD(PTR_GEN,		0,  8)
+
+#define PTR_CHECK_DEV		((1 << 12) - 1)
+
+#define PTR(gen, offset, dev)						\
+	((((uint64_t) dev) << 51) | ((uint64_t) offset) << 8 | gen)
+
+static inline size_t sector_to_bucket(struct cache_set *c, sector_t s)
+{
+	return s >> c->bucket_bits;
+}
+
+static inline sector_t bucket_to_sector(struct cache_set *c, size_t b)
+{
+	return ((sector_t) b) << c->bucket_bits;
+}
+
+static inline sector_t bucket_remainder(struct cache_set *c, sector_t s)
+{
+	return s & (c->sb.bucket_size - 1);
+}
+
+static inline struct cache *PTR_CACHE(struct cache_set *c,
+				      const struct bkey *k,
+				      unsigned ptr)
+{
+	return c->cache[PTR_DEV(k, ptr)];
+}
+
+static inline size_t PTR_BUCKET_NR(struct cache_set *c,
+				   const struct bkey *k,
+				   unsigned ptr)
+{
+	return sector_to_bucket(c, PTR_OFFSET(k, ptr));
+}
+
+static inline struct bucket *PTR_BUCKET(struct cache_set *c,
+					const struct bkey *k,
+					unsigned ptr)
+{
+	return PTR_CACHE(c, k, ptr)->buckets + PTR_BUCKET_NR(c, k, ptr);
+}
+
+/* Btree key macros */
+
+/*
+ * The high bit being set is a relic from when we used it to do binary
+ * searches - it told you where a key started. It's not used anymore,
+ * and can probably be safely dropped.
+ */
+#define KEY(dev, sector, len)	(struct bkey)				\
+{									\
+	.high = (1ULL << 63) | ((uint64_t) (len) << 20) | (dev),	\
+	.low = (sector)							\
+}
+
+static inline void bkey_init(struct bkey *k)
+{
+	*k = KEY(0, 0, 0);
+}
+
+#define KEY_START(k)		(KEY_OFFSET(k) - KEY_SIZE(k))
+#define START_KEY(k)		KEY(KEY_INODE(k), KEY_START(k), 0)
+#define MAX_KEY			KEY(~(~0 << 20), ((uint64_t) ~0) >> 1, 0)
+#define ZERO_KEY		KEY(0, 0, 0)
+
+/*
+ * This is used for various on disk data structures - cache_sb, prio_set, bset,
+ * jset: The checksum is _always_ the first 8 bytes of these structs
+ */
+#define csum_set(i)							\
+	crc64(((void *) (i)) + sizeof(uint64_t),			\
+	      ((void *) end(i)) - (((void *) (i)) + sizeof(uint64_t)))
+
+/* Error handling macros */
+
+#define btree_bug(b, ...)						\
+do {									\
+	if (bch_cache_set_error((b)->c, __VA_ARGS__))			\
+		dump_stack();						\
+} while (0)
+
+#define cache_bug(c, ...)						\
+do {									\
+	if (bch_cache_set_error(c, __VA_ARGS__))			\
+		dump_stack();						\
+} while (0)
+
+#define btree_bug_on(cond, b, ...)					\
+do {									\
+	if (cond)							\
+		btree_bug(b, __VA_ARGS__);				\
+} while (0)
+
+#define cache_bug_on(cond, c, ...)					\
+do {									\
+	if (cond)							\
+		cache_bug(c, __VA_ARGS__);				\
+} while (0)
+
+#define cache_set_err_on(cond, c, ...)					\
+do {									\
+	if (cond)							\
+		bch_cache_set_error(c, __VA_ARGS__);			\
+} while (0)
+
+/* Looping macros */
+
+#define for_each_cache(ca, cs)						\
+	for (int _i = 0; ca = cs->cache[_i], _i < (cs)->sb.nr_in_set; _i++)
+
+#define for_each_bucket(b, ca)						\
+	for (b = (ca)->buckets + (ca)->sb.first_bucket;			\
+	     b < (ca)->buckets + (ca)->sb.nbuckets; b++)
+
+static inline void __bkey_put(struct cache_set *c, struct bkey *k)
+{
+	unsigned i;
+
+	for (i = 0; i < KEY_PTRS(k); i++)
+		atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
+}
+
+/* Blktrace macros */
+
+#define blktrace_msg(c, fmt, ...)					\
+do {									\
+	struct request_queue *q = bdev_get_queue(c->bdev);		\
+	if (q)								\
+		blk_add_trace_msg(q, fmt, ##__VA_ARGS__);		\
+} while (0)
+
+#define blktrace_msg_all(s, fmt, ...)					\
+do {									\
+	struct cache *_c;						\
+	for_each_cache(_c, (s))						\
+		blktrace_msg(_c, fmt, ##__VA_ARGS__);			\
+} while (0)
+
+#define err_printk(...)	printk(KERN_ERR "bcache: " __VA_ARGS__)
+
+static inline void cached_dev_put(struct cached_dev *dc)
+{
+	if (atomic_dec_and_test(&dc->count))
+		schedule_work(&dc->detach);
+}
+
+static inline bool cached_dev_get(struct cached_dev *dc)
+{
+	if (!atomic_inc_not_zero(&dc->count))
+		return false;
+
+	/* Paired with the mb in cached_dev_attach */
+	smp_mb__after_atomic_inc();
+	return true;
+}
+
+/*
+ * bucket_gc_gen() returns the difference between the bucket's current gen and
+ * the oldest gen of any pointer into that bucket in the btree (last_gc).
+ *
+ * bucket_disk_gen() returns the difference between the current gen and the gen
+ * on disk; they're both used to make sure gens don't wrap around.
+ */
+
+static inline uint8_t bucket_gc_gen(struct bucket *b)
+{
+	return b->gen - b->last_gc;
+}
+
+static inline uint8_t bucket_disk_gen(struct bucket *b)
+{
+	return b->gen - b->disk_gen;
+}
+
+#define BUCKET_GC_GEN_MAX	96U
+#define BUCKET_DISK_GEN_MAX	64U
+
+#define kobj_attribute_write(n, fn)					\
+	static struct kobj_attribute ksysfs_##n = __ATTR(n, S_IWUSR, NULL, fn)
+
+#define kobj_attribute_rw(n, show, store)				\
+	static struct kobj_attribute ksysfs_##n =			\
+		__ATTR(n, S_IWUSR|S_IRUSR, show, store)
+
+/* Forward declarations */
+
+void bch_writeback_queue(struct cached_dev *);
+void bch_writeback_add(struct cached_dev *, unsigned);
+
+void bch_count_io_errors(struct cache *, int, const char *);
+void bch_bbio_count_io_errors(struct cache_set *, struct bio *, int, const char *);
+void bch_bbio_endio(struct cache_set *, struct bio *, int, const char *);
+void bch_bbio_free(struct bio *, struct cache_set *);
+struct bio *bch_bbio_alloc(struct cache_set *);
+
+void __bch_submit_bbio(struct bio *, struct cache_set *);
+void bch_submit_bbio(struct bio *, struct cache_set *, struct bkey *, unsigned);
+
+uint8_t bch_inc_gen(struct cache *, struct bucket *);
+void bch_rescale_priorities(struct cache_set *, int);
+bool bch_bucket_add_unused(struct cache *, struct bucket *);
+bool bch_can_save_prios(struct cache *);
+void bch_free_some_buckets(struct cache *);
+void bch_bucket_free(struct cache_set *, struct bkey *);
+int __bch_bucket_alloc_set(struct cache_set *, int, uint16_t,
+			   struct bkey *, int, struct closure *);
+int bch_bucket_alloc_set(struct cache_set *, int, uint16_t,
+			 struct bkey *, int, struct closure *);
+
+__printf(2, 3)
+bool bch_cache_set_error(struct cache_set *, const char *, ...);
+
+void bch_prio_write(struct cache *);
+void bch_write_bdev_super(struct cached_dev *, struct closure *);
+
+extern struct workqueue_struct *bcache_wq, *bch_gc_wq;
+extern const char * const bch_cache_modes[];
+
+struct cache_set *bch_cache_set_alloc(struct cache_sb *);
+void bch_free_discards(struct cache *);
+int bch_alloc_discards(struct cache *);
+void bch_btree_cache_free(struct cache_set *);
+int bch_btree_cache_alloc(struct cache_set *);
+void bch_writeback_init_cached_dev(struct cached_dev *);
+void bch_moving_init_cache_set(struct cache_set *);
+
+void bch_debug_exit(void);
+int bch_debug_init(struct kobject *);
+void bch_writeback_exit(void);
+int bch_writeback_init(void);
+void bch_request_exit(void);
+int bch_request_init(void);
+void bch_btree_exit(void);
+int bch_btree_init(void);
+
+#endif /* _BCACHE_H */
diff --git a/drivers/md/bcache/btree.c b/drivers/md/bcache/btree.c
new file mode 100644
index 0000000..960ab44
--- /dev/null
+++ b/drivers/md/bcache/btree.c
@@ -0,0 +1,2508 @@
+/*
+ * Copyright (C) 2010 Kent Overstreet <kent overstreet gmail com>
+ *
+ * Uses a block device as cache for other block devices; optimized for SSDs.
+ * All allocation is done in buckets, which should match the erase block size
+ * of the device.
+ *
+ * Buckets containing cached data are kept on a heap sorted by priority;
+ * bucket priority is increased on cache hit, and periodically all the buckets
+ * on the heap have their priority scaled down. This currently is just used as
+ * an LRU but in the future should allow for more intelligent heuristics.
+ *
+ * Buckets have an 8 bit counter; freeing is accomplished by incrementing the
+ * counter. Garbage collection is used to remove stale pointers.
+ *
+ * Indexing is done via a btree; nodes are not necessarily fully sorted, rather
+ * as keys are inserted we only sort the pages that have not yet been written.
+ * When garbage collection is run, we resort the entire node.
+ *
+ * All configuration is done via sysfs; see Documentation/bcache.txt.
+ */
+
+#include "bcache.h"
+#include "btree.h"
+#include "debug.h"
+#include "request.h"
+
+#include <linux/slab.h>
+#include <linux/bitops.h>
+#include <linux/hash.h>
+#include <linux/random.h>
+#include <linux/rcupdate.h>
+#include <trace/events/bcache.h>
+
+/*
+ * Todo:
+ * register_bcache: Return errors out to userspace correctly
+ *
+ * Writeback: don't undirty key until after a cache flush
+ *
+ * Create an iterator for key pointers
+ *
+ * On btree write error, mark bucket such that it won't be freed from the cache
+ *
+ * Journalling:
+ *   Check for bad keys in replay
+ *   Propagate barriers
+ *   Refcount journal entries in journal_replay
+ *
+ * Garbage collection:
+ *   Finish incremental gc
+ *   Gc should free old UUIDs, data for invalid UUIDs
+ *
+ * Provide a way to list backing device UUIDs we have data cached for, and
+ * probably how long it's been since we've seen them, and a way to invalidate
+ * dirty data for devices that will never be attached again
+ *
+ * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so
+ * that based on that and how much dirty data we have we can keep writeback
+ * from being starved
+ *
+ * Add a tracepoint or somesuch to watch for writeback starvation
+ *
+ * When btree depth > 1 and splitting an interior node, we have to make sure
+ * alloc_bucket() cannot fail. This should be true but is not completely
+ * obvious.
+ *
+ * Make sure all allocations get charged to the root cgroup
+ *
+ * Plugging?
+ *
+ * If data write is less than hard sector size of ssd, round up offset in open
+ * bucket to the next whole sector
+ *
+ * Also lookup by cgroup in get_open_bucket()
+ *
+ * Superblock needs to be fleshed out for multiple cache devices
+ *
+ * Add a sysfs tunable for the number of writeback IOs in flight
+ *
+ * Add a sysfs tunable for the number of open data buckets
+ *
+ * IO tracking: Can we track when one process is doing io on behalf of another?
+ * IO tracking: Don't use just an average, weigh more recent stuff higher
+ *
+ * Test module load/unload
+ */
+
+static const char * const op_types[] = {
+	"insert", "replace"
+};
+
+static const char *op_type(struct btree_op *op)
+{
+	return op_types[op->type];
+}
+
+#define MAX_NEED_GC		64
+#define MAX_SAVE_PRIO		72
+
+#define PTR_DIRTY_BIT		(((uint64_t) 1 << 36))
+
+#define PTR_HASH(c, k)							\
+	(((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
+
+struct workqueue_struct *bch_gc_wq;
+static struct workqueue_struct *btree_io_wq;
+
+void bch_btree_op_init_stack(struct btree_op *op)
+{
+	memset(op, 0, sizeof(struct btree_op));
+	closure_init_stack(&op->cl);
+	op->lock = -1;
+	bch_keylist_init(&op->keys);
+}
+
+/* Btree key manipulation */
+
+static void bkey_put(struct cache_set *c, struct bkey *k, int level)
+{
+	if ((level && KEY_OFFSET(k)) || !level)
+		__bkey_put(c, k);
+}
+
+/* Btree IO */
+
+static uint64_t btree_csum_set(struct btree *b, struct bset *i)
+{
+	uint64_t crc = b->key.ptr[0];
+	void *data = (void *) i + 8, *end = end(i);
+
+	crc = crc64_update(crc, data, end - data);
+	return crc ^ 0xffffffffffffffff;
+}
+
+static void btree_bio_endio(struct bio *bio, int error)
+{
+	struct closure *cl = bio->bi_private;
+	struct btree *b = container_of(cl, struct btree, io.cl);
+
+	if (error)
+		set_btree_node_io_error(b);
+
+	bch_bbio_count_io_errors(b->c, bio, error, (bio->bi_rw & WRITE)
+				 ? "writing btree" : "reading btree");
+	closure_put(cl);
+}
+
+static void btree_bio_init(struct btree *b)
+{
+	BUG_ON(b->bio);
+	b->bio = bch_bbio_alloc(b->c);
+
+	b->bio->bi_end_io	= btree_bio_endio;
+	b->bio->bi_private	= &b->io.cl;
+}
+
+void bch_btree_read_done(struct closure *cl)
+{
+	struct btree *b = container_of(cl, struct btree, io.cl);
+	struct bset *i = b->sets[0].data;
+	struct btree_iter *iter = b->c->fill_iter;
+	const char *err = "bad btree header";
+	BUG_ON(b->nsets || b->written);
+
+	bch_bbio_free(b->bio, b->c);
+	b->bio = NULL;
+
+	mutex_lock(&b->c->fill_lock);
+	iter->used = 0;
+
+	if (btree_node_io_error(b) ||
+	    !i->seq)
+		goto err;
+
+	for (;
+	     b->written < btree_blocks(b) && i->seq == b->sets[0].data->seq;
+	     i = write_block(b)) {
+		err = "unsupported bset version";
+		if (i->version > BCACHE_BSET_VERSION)
+			goto err;
+
+		err = "bad btree header";
+		if (b->written + set_blocks(i, b->c) > btree_blocks(b))
+			goto err;
+
+		err = "bad magic";
+		if (i->magic != bset_magic(b->c))
+			goto err;
+
+		err = "bad checksum";
+		switch (i->version) {
+		case 0:
+			if (i->csum != csum_set(i))
+				goto err;
+			break;
+		case BCACHE_BSET_VERSION:
+			if (i->csum != btree_csum_set(b, i))
+				goto err;
+			break;
+		}
+
+		err = "empty set";
+		if (i != b->sets[0].data && !i->keys)
+			goto err;
+
+		bch_btree_iter_push(iter, i->start, end(i));
+
+		b->written += set_blocks(i, b->c);
+	}
+
+	err = "corrupted btree";
+	for (i = write_block(b);
+	     index(i, b) < btree_blocks(b);
+	     i = ((void *) i) + block_bytes(b->c))
+		if (i->seq == b->sets[0].data->seq)
+			goto err;
+
+	bch_btree_sort_and_fix_extents(b, iter);
+
+	i = b->sets[0].data;
+	err = "short btree key";
+	if (b->sets[0].size &&
+	    bkey_cmp(&b->key, &b->sets[0].end) < 0)
+		goto err;
+
+	if (b->written < btree_blocks(b))
+		bch_bset_init_next(b);
+out:
+
+	mutex_unlock(&b->c->fill_lock);
+
+	spin_lock(&b->c->btree_read_time_lock);
+	time_stats_update(&b->c->btree_read_time, b->io_start_time);
+	spin_unlock(&b->c->btree_read_time_lock);
+
+	smp_wmb(); /* read_done is our write lock */
+	set_btree_node_read_done(b);
+
+	closure_return(cl);
+err:
+	set_btree_node_io_error(b);
+	bch_cache_set_error(b->c, "%s at bucket %lu, block %zu, %u keys",
+			    err, PTR_BUCKET_NR(b->c, &b->key, 0),
+			    index(i, b), i->keys);
+	goto out;
+}
+
+void bch_btree_read(struct btree *b)
+{
+	BUG_ON(b->nsets || b->written);
+
+	if (!closure_trylock(&b->io.cl, &b->c->cl))
+		BUG();
+
+	b->io_start_time = local_clock();
+
+	btree_bio_init(b);
+	b->bio->bi_rw	= REQ_META|READ_SYNC;
+	b->bio->bi_size	= KEY_SIZE(&b->key) << 9;
+
+	bio_map(b->bio, b->sets[0].data);
+
+	pr_debug("%s", pbtree(b));
+	trace_bcache_btree_read(b->bio);
+	bch_submit_bbio(b->bio, b->c, &b->key, 0);
+
+	continue_at(&b->io.cl, bch_btree_read_done, system_wq);
+}
+
+static void btree_complete_write(struct btree *b, struct btree_write *w)
+{
+	if (w->prio_blocked &&
+	    !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
+		closure_wake_up(&b->c->bucket_wait);
+
+	if (w->journal) {
+		atomic_dec_bug(w->journal);
+		__closure_wake_up(&b->c->journal.wait);
+	}
+
+	if (w->owner)
+		closure_put(w->owner);
+
+	w->prio_blocked	= 0;
+	w->journal	= NULL;
+	w->owner	= NULL;
+}
+
+static void __btree_write_done(struct closure *cl)
+{
+	struct btree *b = container_of(cl, struct btree, io.cl);
+	struct btree_write *w = btree_prev_write(b);
+
+	bch_bbio_free(b->bio, b->c);
+	b->bio = NULL;
+	btree_complete_write(b, w);
+
+	if (btree_node_dirty(b))
+		queue_delayed_work(btree_io_wq, &b->work,
+				   msecs_to_jiffies(30000));
+
+	closure_return(cl);
+}
+
+static void btree_write_done(struct closure *cl)
+{
+	struct btree *b = container_of(cl, struct btree, io.cl);
+	struct bio_vec *bv;
+	int n;
+
+	__bio_for_each_segment(bv, b->bio, n, 0)
+		__free_page(bv->bv_page);
+
+	__btree_write_done(cl);
+}
+
+static void do_btree_write(struct btree *b)
+{
+	struct closure *cl = &b->io.cl;
+	struct bset *i = b->sets[b->nsets].data;
+	BKEY_PADDED(key) k;
+
+	i->version	= BCACHE_BSET_VERSION;
+	i->csum		= btree_csum_set(b, i);
+
+	btree_bio_init(b);
+	b->bio->bi_rw	= REQ_META|WRITE_SYNC;
+	b->bio->bi_size	= set_blocks(i, b->c) * block_bytes(b->c);
+	bio_map(b->bio, i);
+
+	bkey_copy(&k.key, &b->key);
+	SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) + bset_offset(b, i));
+
+	if (!bio_alloc_pages(b->bio, GFP_NOIO)) {
+		int j;
+		struct bio_vec *bv;
+		void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
+
+		bio_for_each_segment(bv, b->bio, j)
+			memcpy(page_address(bv->bv_page),
+			       base + j * PAGE_SIZE, PAGE_SIZE);
+
+		trace_bcache_btree_write(b->bio);
+		bch_submit_bbio(b->bio, b->c, &k.key, 0);
+
+		continue_at(cl, btree_write_done, NULL);
+	} else {
+		b->bio->bi_vcnt = 0;
+		bio_map(b->bio, i);
+
+		trace_bcache_btree_write(b->bio);
+		bch_submit_bbio(b->bio, b->c, &k.key, 0);
+
+		closure_sync(cl);
+		__btree_write_done(cl);
+	}
+}
+
+static void __btree_write(struct btree *b)
+{
+	struct bset *i = b->sets[b->nsets].data;
+
+	BUG_ON(current->bio_list);
+
+	closure_lock(&b->io, &b->c->cl);
+	__cancel_delayed_work(&b->work);
+
+	clear_bit(BTREE_NODE_dirty,	 &b->flags);
+	change_bit(BTREE_NODE_write_idx, &b->flags);
+
+	bch_check_key_order(b, i);
+	BUG_ON(b->written && !i->keys);
+
+	do_btree_write(b);
+
+	pr_debug("%s block %i keys %i", pbtree(b), b->written, i->keys);
+
+	b->written += set_blocks(i, b->c);
+	atomic_long_add(set_blocks(i, b->c) * b->c->sb.block_size,
+			&PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
+
+	bch_btree_sort_lazy(b);
+
+	if (b->written < btree_blocks(b))
+		bch_bset_init_next(b);
+}
+
+static void btree_write_work(struct work_struct *w)
+{
+	struct btree *b = container_of(to_delayed_work(w), struct btree, work);
+
+	down_write(&b->lock);
+
+	if (btree_node_dirty(b))
+		__btree_write(b);
+	up_write(&b->lock);
+}
+
+void bch_btree_write(struct btree *b, bool now, struct btree_op *op)
+{
+	struct bset *i = b->sets[b->nsets].data;
+	struct btree_write *w = btree_current_write(b);
+
+	BUG_ON(b->written &&
+	       (b->written >= btree_blocks(b) ||
+		i->seq != b->sets[0].data->seq ||
+		!i->keys));
+
+	if (!btree_node_dirty(b)) {
+		set_btree_node_dirty(b);
+		queue_delayed_work(btree_io_wq, &b->work,
+				   msecs_to_jiffies(30000));
+	}
+
+	w->prio_blocked += b->prio_blocked;
+	b->prio_blocked = 0;
+
+	if (op && op->journal && !b->level) {
+		if (w->journal &&
+		    journal_pin_cmp(b->c, w, op)) {
+			atomic_dec_bug(w->journal);
+			w->journal = NULL;
+		}
+
+		if (!w->journal) {
+			w->journal = op->journal;
+			atomic_inc(w->journal);
+		}
+	}
+
+	if (current->bio_list)
+		return;
+
+	/* Force write if set is too big */
+	if (now ||
+	    b->level ||
+	    set_bytes(i) > PAGE_SIZE - 48) {
+		if (op && now) {
+			/* Must wait on multiple writes */
+			BUG_ON(w->owner);
+			w->owner = &op->cl;
+			closure_get(&op->cl);
+		}
+
+		__btree_write(b);
+	}
+	BUG_ON(!b->written);
+}
+
+/*
+ * Btree in memory cache - allocation/freeing
+ * mca -> memory cache
+ */
+
+static void mca_reinit(struct btree *b)
+{
+	b->flags	= 0;
+	b->written	= 0;
+	b->nsets	= 0;
+
+	for (int i = 0; i < MAX_BSETS; i++)
+		b->sets[i].size = 0;
+	/*
+	 * Second loop starts at 1 because b->sets[0]->data is the memory we
+	 * allocated
+	 */
+	for (int i = 1; i < MAX_BSETS; i++)
+		b->sets[i].data = NULL;
+}
+
+#define mca_reserve(c)	((c->root ? c->root->level : 1) * 8 + 16)
+#define mca_can_free(c)						\
+	max_t(int, 0, c->bucket_cache_used - mca_reserve(c))
+
+static void mca_data_free(struct btree *b)
+{
+	struct bset_tree *t = b->sets;
+	BUG_ON(!closure_is_unlocked(&b->io.cl));
+
+	if (bset_prev_bytes(b) < PAGE_SIZE)
+		kfree(t->prev);
+	else
+		free_pages((unsigned long) t->prev,
+			   get_order(bset_prev_bytes(b)));
+
+	if (bset_tree_bytes(b) < PAGE_SIZE)
+		kfree(t->tree);
+	else
+		free_pages((unsigned long) t->tree,
+			   get_order(bset_tree_bytes(b)));
+
+	free_pages((unsigned long) t->data, b->page_order);
+
+	t->prev = NULL;
+	t->tree = NULL;
+	t->data = NULL;
+	list_move(&b->list, &b->c->btree_cache_freed);
+	b->c->bucket_cache_used--;
+}
+
+static void mca_bucket_free(struct btree *b)
+{
+	BUG_ON(btree_node_dirty(b));
+
+	b->key.ptr[0] = 0;
+	hlist_del_init_rcu(&b->hash);
+	list_move(&b->list, &b->c->btree_cache_freeable);
+}
+
+static unsigned btree_order(struct bkey *k)
+{
+	return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
+}
+
+static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
+{
+	struct bset_tree *t = b->sets;
+	BUG_ON(t->data);
+
+	b->page_order = max_t(unsigned,
+			      ilog2(b->c->btree_pages),
+			      btree_order(k));
+
+	t->data = (void *) __get_free_pages(gfp, b->page_order);
+	if (!t->data)
+		goto err;
+
+	t->tree = bset_tree_bytes(b) < PAGE_SIZE
+		? kmalloc(bset_tree_bytes(b), gfp)
+		: (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b)));
+	if (!t->tree)
+		goto err;
+
+	t->prev = bset_prev_bytes(b) < PAGE_SIZE
+		? kmalloc(bset_prev_bytes(b), gfp)
+		: (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b)));
+	if (!t->prev)
+		goto err;
+
+	list_move(&b->list, &b->c->btree_cache);
+	b->c->bucket_cache_used++;
+	return;
+err:
+	mca_data_free(b);
+}
+
+static struct btree *mca_bucket_alloc(struct cache_set *c,
+				      struct bkey *k, gfp_t gfp)
+{
+	struct btree *b = kzalloc(sizeof(struct btree), gfp);
+	if (!b)
+		return NULL;
+
+	init_rwsem(&b->lock);
+	lockdep_set_novalidate_class(&b->lock);
+	INIT_LIST_HEAD(&b->list);
+	INIT_DELAYED_WORK(&b->work, btree_write_work);
+	b->c = c;
+	closure_init_unlocked(&b->io);
+
+	mca_data_alloc(b, k, gfp);
+	return b;
+}
+
+static int mca_reap(struct btree *b, struct closure *cl, unsigned min_order)
+{
+	lockdep_assert_held(&b->c->bucket_lock);
+
+	if (!down_write_trylock(&b->lock))
+		return -ENOMEM;
+
+	if (b->page_order < min_order) {
+		rw_unlock(true, b);
+		return -ENOMEM;
+	}
+
+	BUG_ON(btree_node_dirty(b) && !b->sets[0].data);
+
+	if (cl && btree_node_dirty(b))
+		bch_btree_write(b, true, NULL);
+
+	if (cl)
+		closure_wait_event_async(&b->io.wait, cl,
+			 atomic_read(&b->io.cl.remaining) == -1);
+
+	if (btree_node_dirty(b) ||
+	    !closure_is_unlocked(&b->io.cl) ||
+	    work_pending(&b->work.work)) {
+		rw_unlock(true, b);
+		return -EAGAIN;
+	}
+
+	return 0;
+}
+
+static int bch_mca_shrink(struct shrinker *shrink,
+			      struct shrink_control *sc)
+{
+	struct cache_set *c = container_of(shrink, struct cache_set, shrink);
+	struct btree *b, *t;
+	unsigned i;
+	int nr, orig_nr = sc->nr_to_scan;
+
+	if (c->shrinker_disabled)
+		return 0;
+
+	/*
+	 * If nr == 0, we're supposed to return the number of items we have
+	 * cached. Not allowed to return -1.
+	 */
+	if (!orig_nr)
+		goto out;
+
+	/* Return -1 if we can't do anything right now */
+	if (!mutex_trylock(&c->bucket_lock))
+		return -1;
+
+	if (c->try_harder) {
+		mutex_unlock(&c->bucket_lock);
+		return -1;
+	}
+
+	if (list_empty(&c->btree_cache)) {
+		/*
+		 * Can happen right when we first start up, before we've read in
+		 * any btree nodes
+		 */
+		mutex_unlock(&c->bucket_lock);
+		return 0;
+	}
+
+	orig_nr /= c->btree_pages;
+	nr = orig_nr = min_t(int, orig_nr, mca_can_free(c));
+
+	i = 0;
+	list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
+		if (!nr)
+			break;
+
+		if (++i > 3 &&
+		    !mca_reap(b, NULL, 0)) {
+			mca_data_free(b);
+			rw_unlock(true, b);
+			--nr;
+		}
+	}
+
+	for (i = c->bucket_cache_used;
+	     i && nr;
+	     --i) {
+		b = list_first_entry(&c->btree_cache, struct btree, list);
+		list_rotate_left(&c->btree_cache);
+
+		if (!b->accessed &&
+		    !mca_reap(b, NULL, 0)) {
+			mca_bucket_free(b);
+			mca_data_free(b);
+			rw_unlock(true, b);
+			--nr;
+		} else
+			b->accessed = 0;
+	}
+
+	mutex_unlock(&c->bucket_lock);
+out:
+	return mca_can_free(c) * c->btree_pages;
+}
+
+void bch_btree_cache_free(struct cache_set *c)
+{
+	struct btree *b;
+	struct closure cl;
+	closure_init_stack(&cl);
+
+	if (c->shrink.list.next)
+		unregister_shrinker(&c->shrink);
+
+	mutex_lock(&c->bucket_lock);
+
+#ifdef CONFIG_BCACHE_DEBUG
+	if (c->verify_data)
+		list_move(&c->verify_data->list, &c->btree_cache);
+#endif
+
+	list_splice(&c->btree_cache_freeable,
+		    &c->btree_cache);
+
+	while (!list_empty(&c->btree_cache)) {
+		b = list_first_entry(&c->btree_cache, struct btree, list);
+
+		if (btree_node_dirty(b))
+			btree_complete_write(b, btree_current_write(b));
+		clear_bit(BTREE_NODE_dirty, &b->flags);
+
+		mca_data_free(b);
+	}
+
+	while (!list_empty(&c->btree_cache_freed)) {
+		b = list_first_entry(&c->btree_cache_freed,
+				     struct btree, list);
+		list_del(&b->list);
+		cancel_delayed_work_sync(&b->work);
+		kfree(b);
+	}
+
+	mutex_unlock(&c->bucket_lock);
+}
+
+int bch_btree_cache_alloc(struct cache_set *c)
+{
+	/* XXX: doesn't check for errors */
+
+	closure_init_unlocked(&c->gc);
+
+	for (int i = 0; i < mca_reserve(c); i++)
+		mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
+
+	list_splice_init(&c->btree_cache,
+			 &c->btree_cache_freeable);
+
+#ifdef CONFIG_BCACHE_DEBUG
+	mutex_init(&c->verify_lock);
+
+	c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
+
+	if (c->verify_data &&
+	    c->verify_data->sets[0].data)
+		list_del_init(&c->verify_data->list);
+	else
+		c->verify_data = NULL;
+#endif
+
+	c->shrink.shrink = bch_mca_shrink;
+	c->shrink.seeks = 3;
+	register_shrinker(&c->shrink);
+
+	return 0;
+}
+
+/* Btree in memory cache - hash table */
+
+static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
+{
+	return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
+}
+
+static struct btree *mca_find(struct cache_set *c, struct bkey *k)
+{
+	struct hlist_node *cursor;
+	struct btree *b;
+
+	rcu_read_lock();
+	hlist_for_each_entry_rcu(b, cursor, mca_hash(c, k), hash)
+		if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
+			goto out;
+	b = NULL;
+out:
+	rcu_read_unlock();
+	return b;
+}
+
+static struct btree *mca_cannibalize(struct cache_set *c, struct bkey *k,
+				     int level, struct closure *cl)
+{
+	int ret = -ENOMEM;
+	struct btree *i;
+
+	if (!cl)
+		return ERR_PTR(-ENOMEM);
+
+	/*
+	 * Trying to free up some memory - i.e. reuse some btree nodes - may
+	 * require initiating IO to flush the dirty part of the node. If we're
+	 * running under generic_make_request(), that IO will never finish and
+	 * we would deadlock. Returning -EAGAIN causes the cache lookup code to
+	 * punt to workqueue and retry.
+	 */
+	if (current->bio_list)
+		return ERR_PTR(-EAGAIN);
+
+	if (c->try_harder && c->try_harder != cl) {
+		closure_wait_event_async(&c->try_wait, cl, !c->try_harder);
+		return ERR_PTR(-EAGAIN);
+	}
+
+	/* XXX: tracepoint */
+	c->try_harder = cl;
+	c->try_harder_start = local_clock();
+retry:
+	list_for_each_entry_reverse(i, &c->btree_cache, list) {
+		int r = mca_reap(i, cl, btree_order(k));
+		if (!r)
+			return i;
+		if (r != -ENOMEM)
+			ret = r;
+	}
+
+	if (ret == -EAGAIN &&
+	    closure_blocking(cl)) {
+		mutex_unlock(&c->bucket_lock);
+		closure_sync(cl);
+		mutex_lock(&c->bucket_lock);
+		goto retry;
+	}
+
+	return ERR_PTR(ret);
+}
+
+/*
+ * We can only have one thread cannibalizing other cached btree nodes at a time,
+ * or we'll deadlock. We use an open coded mutex to ensure that, which a
+ * cannibalize_bucket() will take. This means every time we unlock the root of
+ * the btree, we need to release this lock if we have it held.
+ */
+void bch_cannibalize_unlock(struct cache_set *c, struct closure *cl)
+{
+	if (c->try_harder == cl) {
+		time_stats_update(&c->try_harder_time, c->try_harder_start);
+		c->try_harder = NULL;
+		__closure_wake_up(&c->try_wait);
+	}
+}
+
+static struct btree *mca_alloc(struct cache_set *c, struct bkey *k,
+			       int level, struct closure *cl)
+{
+	struct btree *b;
+
+	lockdep_assert_held(&c->bucket_lock);
+
+	if (mca_find(c, k))
+		return NULL;
+
+	/* btree_free() doesn't free memory; it sticks the node on the end of
+	 * the list. Check if there's any freed nodes there:
+	 */
+	list_for_each_entry(b, &c->btree_cache_freeable, list)
+		if (!mca_reap(b, NULL, btree_order(k)))
+			goto out;
+
+	/* We never free struct btree itself, just the memory that holds the on
+	 * disk node. Check the freed list before allocating a new one:
+	 */
+	list_for_each_entry(b, &c->btree_cache_freed, list)
+		if (!mca_reap(b, NULL, 0)) {
+			mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
+			if (!b->sets[0].data)
+				goto err;
+			else
+				goto out;
+		}
+
+	b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
+	if (!b)
+		goto err;
+
+	BUG_ON(!down_write_trylock(&b->lock));
+	if (!b->sets->data)
+		goto err;
+out:
+	BUG_ON(!closure_is_unlocked(&b->io.cl));
+
+	bkey_copy(&b->key, k);
+	list_move(&b->list, &c->btree_cache);
+	hlist_del_init_rcu(&b->hash);
+	hlist_add_head_rcu(&b->hash, mca_hash(c, k));
+
+	lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
+	b->level	= level;
+
+	mca_reinit(b);
+
+	return b;
+err:
+	if (b)
+		rw_unlock(true, b);
+
+	b = mca_cannibalize(c, k, level, cl);
+	if (!IS_ERR(b))
+		goto out;
+
+	return b;
+}
+
+/**
+ * bch_btree_node_get - find a btree node in the cache and lock it, reading it
+ * in from disk if necessary.
+ *
+ * If IO is necessary, it uses the closure embedded in struct btree_op to wait;
+ * if that closure is in non blocking mode, will return -EAGAIN.
+ *
+ * The btree node will have either a read or a write lock held, depending on
+ * level and op->lock.
+ */
+struct btree *bch_btree_node_get(struct cache_set *c, struct bkey *k,
+				 int level, struct btree_op *op)
+{
+	int i = 0;
+	bool write = level <= op->lock;
+	struct btree *b;
+
+	BUG_ON(level < 0);
+retry:
+	b = mca_find(c, k);
+
+	if (!b) {
+		mutex_lock(&c->bucket_lock);
+		b = mca_alloc(c, k, level, &op->cl);
+		mutex_unlock(&c->bucket_lock);
+
+		if (!b)
+			goto retry;
+		if (IS_ERR(b))
+			return b;
+
+		bch_btree_read(b);
+
+		if (!write)
+			downgrade_write(&b->lock);
+	} else {
+		rw_lock(write, b, level);
+		if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
+			rw_unlock(write, b);
+			goto retry;
+		}
+		BUG_ON(b->level != level);
+	}
+
+	b->accessed = 1;
+
+	for (; i <= b->nsets && b->sets[i].size; i++) {
+		prefetch(b->sets[i].tree);
+		prefetch(b->sets[i].data);
+	}
+
+	for (; i <= b->nsets; i++)
+		prefetch(b->sets[i].data);
+
+	if (!closure_wait_event(&b->io.wait, &op->cl,
+				btree_node_read_done(b))) {
+		rw_unlock(write, b);
+		b = ERR_PTR(-EAGAIN);
+	} else if (btree_node_io_error(b)) {
+		rw_unlock(write, b);
+		b = ERR_PTR(-EIO);
+	} else
+		BUG_ON(!b->written);
+
+	return b;
+}
+
+static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level)
+{
+	struct btree *b;
+
+	mutex_lock(&c->bucket_lock);
+	b = mca_alloc(c, k, level, NULL);
+	mutex_unlock(&c->bucket_lock);
+
+	if (!IS_ERR_OR_NULL(b)) {
+		bch_btree_read(b);
+		rw_unlock(true, b);
+	}
+}
+
+/* Btree alloc */
+
+static void btree_node_free(struct btree *b, struct btree_op *op)
+{
+	/*
+	 * The BUG_ON() in btree_node_get() implies that we must have a write
+	 * lock on parent to free or even invalidate a node
+	 */
+	BUG_ON(op->lock <= b->level);
+	BUG_ON(b == b->c->root);
+	pr_debug("bucket %s", pbtree(b));
+
+	if (btree_node_dirty(b))
+		btree_complete_write(b, btree_current_write(b));
+	clear_bit(BTREE_NODE_dirty, &b->flags);
+
+	if (b->prio_blocked &&
+	    !atomic_sub_return(b->prio_blocked, &b->c->prio_blocked))
+		closure_wake_up(&b->c->bucket_wait);
+
+	b->prio_blocked = 0;
+
+	__cancel_delayed_work(&b->work);
+
+	mutex_lock(&b->c->bucket_lock);
+
+	for (unsigned i = 0; i < KEY_PTRS(&b->key); i++) {
+		BUG_ON(atomic_read(&PTR_BUCKET(b->c, &b->key, i)->pin));
+
+		bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
+			    PTR_BUCKET(b->c, &b->key, i));
+	}
+
+	bch_bucket_free(b->c, &b->key);
+	mca_bucket_free(b);
+	mutex_unlock(&b->c->bucket_lock);
+}
+
+struct btree *bch_btree_node_alloc(struct cache_set *c, int level,
+				   struct closure *cl)
+{
+	BKEY_PADDED(key) k;
+	struct btree *b = ERR_PTR(-EAGAIN);
+
+	mutex_lock(&c->bucket_lock);
+retry:
+	if (__bch_bucket_alloc_set(c, GC_MARK_BTREE, 0, &k.key, 1, cl))
+		goto err;
+
+	SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
+
+	b = mca_alloc(c, &k.key, level, cl);
+	if (IS_ERR(b))
+		goto err_free;
+
+	if (!b) {
+		cache_bug(c, "Tried to allocate bucket"
+			  " that was in btree cache");
+		__bkey_put(c, &k.key);
+		goto retry;
+	}
+
+	set_btree_node_read_done(b);
+	b->accessed = 1;
+	bch_bset_init_next(b);
+
+	mutex_unlock(&c->bucket_lock);
+	return b;
+err_free:
+	bch_bucket_free(c, &k.key);
+	__bkey_put(c, &k.key);
+err:
+	mutex_unlock(&c->bucket_lock);
+	return b;
+}
+
+static struct btree *btree_node_alloc_replacement(struct btree *b,
+						  struct closure *cl)
+{
+	struct btree *n = bch_btree_node_alloc(b->c, b->level, cl);
+	if (!IS_ERR_OR_NULL(n))
+		bch_btree_sort_into(b, n);
+
+	return n;
+}
+
+/* Garbage collection */
+
+uint8_t __bch_btree_mark_key(struct cache_set *c, int level, struct bkey *k)
+{
+	uint8_t stale = 0;
+	struct bucket *g;
+
+	/*
+	 * ptr_invalid() can't return true for the keys that mark btree nodes as
+	 * freed, but since ptr_bad() returns true we'll never actually use them
+	 * for anything and thus we don't want mark their pointers here
+	 */
+	if (!bkey_cmp(k, &ZERO_KEY))
+		return stale;
+
+	for (unsigned i = 0; i < KEY_PTRS(k); i++) {
+		if (!ptr_available(c, k, i))
+			continue;
+
+		g = PTR_BUCKET(c, k, i);
+
+		if (gen_after(g->gc_gen, PTR_GEN(k, i)))
+			g->gc_gen = PTR_GEN(k, i);
+
+		if (ptr_stale(c, k, i)) {
+			stale = max(stale, ptr_stale(c, k, i));
+			continue;
+		}
+
+		cache_bug_on(GC_MARK(g) &&
+			     (GC_MARK(g) == GC_MARK_BTREE) != (level != 0), c,
+			     "inconsistent pointers: mark = %llu, level = %i",
+			     GC_MARK(g), level);
+
+		if (level)
+			SET_GC_MARK(g, GC_MARK_BTREE);
+		else if (KEY_DIRTY(k))
+			SET_GC_MARK(g, GC_MARK_DIRTY);
+
+		/* guard against overflow */
+		SET_GC_SECTORS_USED(g, min_t(unsigned,
+					     GC_SECTORS_USED(g) + KEY_SIZE(k),
+					     (1 << 14) - 1));
+
+		BUG_ON(!GC_SECTORS_USED(g));
+	}
+
+	return stale;
+}
+
+#define btree_mark_key(b, k)	__bch_btree_mark_key(b->c, b->level, k)
+
+static int btree_gc_mark_node(struct btree *b, unsigned *keys, struct gc_stat *gc)
+{
+	uint8_t stale = 0;
+	unsigned last_dev = -1;
+	struct bcache_device *d = NULL;
+	struct bkey *k;
+
+	struct btree_iter iter;
+	bch_btree_iter_init(b, &iter, NULL);
+
+	gc->nodes++;
+
+	while ((k = bch_btree_iter_next(&iter))) {
+		if (bch_ptr_invalid(b, k))
+			continue;
+
+		if (last_dev != KEY_INODE(k)) {
+			last_dev = KEY_INODE(k);
+
+			d = KEY_INODE(k) < b->c->nr_uuids
+				? b->c->devices[last_dev]
+				: NULL;
+		}
+
+		stale = max(stale, btree_mark_key(b, k));
+
+		if (bch_ptr_bad(b, k))
+			continue;
+
+		*keys += bkey_u64s(k);
+
+		gc->key_bytes += bkey_u64s(k);
+		gc->nkeys++;
+
+		gc->data += KEY_SIZE(k);
+		if (KEY_DIRTY(k)) {
+			gc->dirty += KEY_SIZE(k);
+			if (d)
+				d->sectors_dirty_gc += KEY_SIZE(k);
+		}
+	}
+
+	for (struct bset_tree *t = b->sets; t <= &b->sets[b->nsets]; t++)
+		btree_bug_on(t->size &&
+			     bset_written(b, t) &&
+			     bkey_cmp(&b->key, &t->end) < 0,
+			     b, "found short btree key in gc");
+
+	return stale;
+}
+
+static struct btree *btree_gc_alloc(struct btree *b, struct bkey *k,
+				    struct btree_op *op)
+{
+	/*
+	 * We block priorities from being written for the duration of garbage
+	 * collection, so we can't sleep in btree_alloc() ->
+	 * bch_bucket_alloc_set(), or we'd risk deadlock - so we don't pass it
+	 * our closure.
+	 */
+	struct btree *n = btree_node_alloc_replacement(b, NULL);
+
+	if (!IS_ERR_OR_NULL(n)) {
+		swap(b, n);
+
+		memcpy(k->ptr, b->key.ptr,
+		       sizeof(uint64_t) * KEY_PTRS(&b->key));
+
+		__bkey_put(b->c, &b->key);
+		atomic_inc(&b->c->prio_blocked);
+		b->prio_blocked++;
+
+		btree_node_free(n, op);
+		up_write(&n->lock);
+	}
+
+	return b;
+}
+
+/*
+ * Leaving this at 2 until we've got incremental garbage collection done; it
+ * could be higher (and has been tested with 4) except that garbage collection
+ * could take much longer, adversely affecting latency.
+ */
+#define GC_MERGE_NODES	2
+
+struct gc_merge_info {
+	struct btree	*b;
+	struct bkey	*k;
+	unsigned	keys;
+};
+
+static void btree_gc_coalesce(struct btree *b, struct btree_op *op,
+			      struct gc_stat *gc, struct gc_merge_info *r)
+{
+	unsigned nodes = 0, keys = 0, blocks;
+
+	while (nodes < GC_MERGE_NODES && r[nodes].b)
+		keys += r[nodes++].keys;
+
+	blocks = btree_default_blocks(b->c) * 2 / 3;
+
+	if (nodes < 2 ||
+	    __set_blocks(b->sets[0].data, keys, b->c) > blocks * (nodes - 1))
+		return;
+
+	for (int i = nodes - 1; i >= 0; --i) {
+		if (r[i].b->written)
+			r[i].b = btree_gc_alloc(r[i].b, r[i].k, op);
+
+		if (r[i].b->written)
+			return;
+	}
+
+	for (int i = nodes - 1; i > 0; --i) {
+		struct bset *n1 = r[i].b->sets->data;
+		struct bset *n2 = r[i - 1].b->sets->data;
+		struct bkey *last = NULL;
+
+		keys = 0;
+
+		if (i == 1) {
+			/*
+			 * Last node we're not getting rid of - we're getting
+			 * rid of the node at r[0]. Have to try and fit all of
+			 * the remaining keys into this node; we can't ensure
+			 * they will always fit due to rounding and variable
+			 * length keys (shouldn't be possible in practice,
+			 * though)
+			 */
+			if (__set_blocks(n1, n1->keys + r->keys,
+					 b->c) > btree_blocks(r[i].b))
+				return;
+
+			keys = n2->keys;
+			last = &r->b->key;
+		} else
+			for (struct bkey *k = n2->start;
+			     k < end(n2);
+			     k = bkey_next(k)) {
+				if (__set_blocks(n1, n1->keys + keys +
+						 bkey_u64s(k), b->c) > blocks)
+					break;
+
+				last = k;
+				keys += bkey_u64s(k);
+			}
+
+		BUG_ON(__set_blocks(n1, n1->keys + keys,
+				    b->c) > btree_blocks(r[i].b));
+
+		if (last) {
+			bkey_copy_key(&r[i].b->key, last);
+			bkey_copy_key(r[i].k, last);
+		}
+
+		memcpy(end(n1),
+		       n2->start,
+		       (void *) node(n2, keys) - (void *) n2->start);
+
+		n1->keys += keys;
+
+		memmove(n2->start,
+			node(n2, keys),
+			(void *) end(n2) - (void *) node(n2, keys));
+
+		n2->keys -= keys;
+
+		r[i].keys	= n1->keys;
+		r[i - 1].keys	= n2->keys;
+	}
+
+	btree_node_free(r->b, op);
+	up_write(&r->b->lock);
+
+	pr_debug("coalesced %u nodes", nodes);
+
+	gc->nodes--;
+	nodes--;
+
+	memmove(&r[0], &r[1], sizeof(struct gc_merge_info) * nodes);
+	memset(&r[nodes], 0, sizeof(struct gc_merge_info));
+}
+
+static int btree_gc_recurse(struct btree *b, struct btree_op *op,
+			    struct closure *writes, struct gc_stat *gc)
+{
+	void write(struct btree *r)
+	{
+		if (!r->written)
+			bch_btree_write(r, true, op);
+		else if (btree_node_dirty(r)) {
+			BUG_ON(btree_current_write(r)->owner);
+			btree_current_write(r)->owner = writes;
+			closure_get(writes);
+
+			bch_btree_write(r, true, NULL);
+		}
+
+		up_write(&r->lock);
+	}
+
+	int ret = 0, stale;
+	struct gc_merge_info r[GC_MERGE_NODES];
+
+	memset(r, 0, sizeof(r));
+
+	while ((r->k = bch_next_recurse_key(b, &b->c->gc_done))) {
+		r->b = bch_btree_node_get(b->c, r->k, b->level - 1, op);
+
+		if (IS_ERR(r->b)) {
+			ret = PTR_ERR(r->b);
+			break;
+		}
+
+		r->keys	= 0;
+		stale = btree_gc_mark_node(r->b, &r->keys, gc);
+
+		if (!b->written &&
+		    (r->b->level || stale > 10 ||
+		     b->c->gc_always_rewrite))
+			r->b = btree_gc_alloc(r->b, r->k, op);
+
+		if (r->b->level)
+			ret = btree_gc_recurse(r->b, op, writes, gc);
+
+		if (ret) {
+			write(r->b);
+			break;
+		}
+
+		bkey_copy_key(&b->c->gc_done, r->k);
+
+		if (!b->written)
+			btree_gc_coalesce(b, op, gc, r);
+
+		if (r[GC_MERGE_NODES - 1].b)
+			write(r[GC_MERGE_NODES - 1].b);
+
+		memmove(&r[1], &r[0],
+			sizeof(struct gc_merge_info) * (GC_MERGE_NODES - 1));
+
+		/* When we've got incremental GC working, we'll want to do
+		 * if (should_resched())
+		 *	return -EAGAIN;
+		 */
+		cond_resched();
+#if 0
+		if (need_resched()) {
+			ret = -EAGAIN;
+			break;
+		}
+#endif
+	}
+
+	for (unsigned i = 1; i < GC_MERGE_NODES && r[i].b; i++)
+		write(r[i].b);
+
+	/* Might have freed some children, must remove their keys */
+	if (!b->written)
+		bch_btree_sort(b);
+
+	return ret;
+}
+
+static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
+			     struct closure *writes, struct gc_stat *gc)
+{
+	struct btree *n = NULL;
+	unsigned keys = 0;
+	int ret = 0, stale = btree_gc_mark_node(b, &keys, gc);
+
+	if (b->level || stale > 10)
+		n = btree_node_alloc_replacement(b, NULL);
+
+	if (!IS_ERR_OR_NULL(n))
+		swap(b, n);
+
+	if (b->level)
+		ret = btree_gc_recurse(b, op, writes, gc);
+
+	if (!b->written || btree_node_dirty(b)) {
+		atomic_inc(&b->c->prio_blocked);
+		b->prio_blocked++;
+		bch_btree_write(b, true, n ? op : NULL);
+	}
+
+	if (!IS_ERR_OR_NULL(n)) {
+		closure_sync(&op->cl);
+		bch_btree_set_root(b);
+		btree_node_free(n, op);
+		rw_unlock(true, b);
+	}
+
+	return ret;
+}
+
+static void btree_gc_start(struct cache_set *c)
+{
+	struct cache *ca;
+	struct bucket *b;
+
+	if (!c->gc_mark_valid)
+		return;
+
+	mutex_lock(&c->bucket_lock);
+
+	for_each_cache(ca, c)
+		bch_free_some_buckets(ca);
+
+	c->gc_mark_valid = 0;
+	c->gc_done = ZERO_KEY;
+
+	for_each_cache(ca, c)
+		for_each_bucket(b, ca) {
+			b->gc_gen = b->gen;
+			if (!atomic_read(&b->pin))
+				SET_GC_MARK(b, GC_MARK_RECLAIMABLE);
+		}
+
+	for (struct bcache_device **d = c->devices;
+	     d < c->devices + c->nr_uuids;
+	     d++)
+		if (*d)
+			(*d)->sectors_dirty_gc = 0;
+
+	mutex_unlock(&c->bucket_lock);
+}
+
+size_t bch_btree_gc_finish(struct cache_set *c)
+{
+	size_t available = 0;
+	struct bucket *b;
+	struct cache *ca;
+	uint64_t *i;
+
+	mutex_lock(&c->bucket_lock);
+
+	set_gc_sectors(c);
+	c->gc_mark_valid = 1;
+	c->need_gc	= 0;
+
+	if (c->root)
+		for (unsigned i = 0; i < KEY_PTRS(&c->root->key); i++)
+			SET_GC_MARK(PTR_BUCKET(c, &c->root->key, i), GC_MARK_BTREE);
+
+	for (unsigned i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
+		SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i), GC_MARK_BTREE);
+
+	for_each_cache(ca, c) {
+		ca->invalidate_needs_gc = 0;
+
+		for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
+			SET_GC_MARK(ca->buckets + *i, GC_MARK_BTREE);
+
+		for (i = ca->prio_buckets;
+		     i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
+			SET_GC_MARK(ca->buckets + *i, GC_MARK_BTREE);
+
+		for_each_bucket(b, ca) {
+			b->last_gc	= b->gc_gen;
+			c->need_gc	= max(c->need_gc, bucket_gc_gen(b));
+
+			if (!atomic_read(&b->pin) &&
+			    GC_MARK(b) == GC_MARK_RECLAIMABLE) {
+				available++;
+				if (!GC_SECTORS_USED(b))
+					bch_bucket_add_unused(ca, b);
+			}
+		}
+	}
+
+	for (struct bcache_device **d = c->devices;
+	     d < c->devices + c->nr_uuids;
+	     d++)
+		if (*d) {
+			unsigned long last =
+				atomic_long_read(&((*d)->sectors_dirty));
+			long difference = (*d)->sectors_dirty_gc - last;
+
+			pr_debug("sectors dirty off by %li", difference);
+
+			(*d)->sectors_dirty_last += difference;
+
+			atomic_long_set(&((*d)->sectors_dirty),
+					(*d)->sectors_dirty_gc);
+		}
+
+	mutex_unlock(&c->bucket_lock);
+	return available;
+}
+
+static void btree_gc(struct closure *cl)
+{
+	struct cache_set *c = container_of(cl, struct cache_set, gc.cl);
+	int ret;
+	unsigned long available;
+	struct gc_stat stats;
+	struct closure writes;
+	struct btree_op op;
+
+	uint64_t start_time = local_clock();
+	trace_bcache_gc_start(c->sb.set_uuid);
+	blktrace_msg_all(c, "Starting gc");
+
+	memset(&stats, 0, sizeof(struct gc_stat));
+	closure_init_stack(&writes);
+	bch_btree_op_init_stack(&op);
+	op.lock = SHRT_MAX;
+
+	btree_gc_start(c);
+
+	ret = btree_root(gc_root, c, &op, &writes, &stats);
+	closure_sync(&op.cl);
+	closure_sync(&writes);
+
+	if (ret) {
+		blktrace_msg_all(c, "Stopped gc");
+		printk(KERN_WARNING "bcache: gc failed!\n");
+
+		continue_at(cl, btree_gc, bch_gc_wq);
+	}
+
+	/* Possibly wait for new UUIDs or whatever to hit disk */
+	bch_journal_meta(c, &op.cl);
+	closure_sync(&op.cl);
+
+	available = bch_btree_gc_finish(c);
+
+	time_stats_update(&c->btree_gc_time, start_time);
+
+	stats.key_bytes *= sizeof(uint64_t);
+	stats.dirty	<<= 9;
+	stats.data	<<= 9;
+	stats.in_use	= (c->nbuckets - available) * 100 / c->nbuckets;
+	memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
+	blktrace_msg_all(c, "Finished gc");
+
+	trace_bcache_gc_end(c->sb.set_uuid);
+	closure_wake_up(&c->bucket_wait);
+
+	continue_at(cl, bch_moving_gc, bch_gc_wq);
+}
+
+void bch_queue_gc(struct cache_set *c)
+{
+	if (closure_trylock(&c->gc.cl, &c->cl))
+		continue_at(&c->gc.cl, btree_gc, bch_gc_wq);
+}
+
+/* Initial partial gc */
+
+static int bch_btree_check_recurse(struct btree *b, struct btree_op *op,
+				   unsigned long **seen)
+{
+	int ret;
+	struct bkey *k;
+	struct bucket *g;
+
+	for_each_key_filter(b, k, bch_ptr_invalid) {
+		for (unsigned i = 0; i < KEY_PTRS(k); i++) {
+			if (!ptr_available(b->c, k, i))
+				continue;
+
+			g = PTR_BUCKET(b->c, k, i);
+
+			if (!__test_and_set_bit(PTR_BUCKET_NR(b->c, k, i),
+						seen[PTR_DEV(k, i)]) ||
+			    !ptr_stale(b->c, k, i)) {
+				g->gen = PTR_GEN(k, i);
+
+				if (b->level)
+					g->prio = BTREE_PRIO;
+				else if (g->prio == BTREE_PRIO)
+					g->prio = INITIAL_PRIO;
+			}
+		}
+
+		btree_mark_key(b, k);
+	}
+
+	if (b->level) {
+		k = bch_next_recurse_key(b, &ZERO_KEY);
+
+		while (k) {
+			struct bkey *p = bch_next_recurse_key(b, k);
+			if (p)
+				btree_node_prefetch(b->c, p, b->level - 1);
+
+			ret = btree(check_recurse, k, b, op, seen);
+			if (ret)
+				return ret;
+
+			k = p;
+		}
+	}
+
+	return 0;
+}
+
+int bch_btree_check(struct cache_set *c, struct btree_op *op)
+{
+	int ret = -ENOMEM;
+	unsigned long *seen[MAX_CACHES_PER_SET];
+
+	memset(seen, 0, sizeof(seen));
+
+	for (int i = 0; c->cache[i]; i++) {
+		size_t n = DIV_ROUND_UP(c->cache[i]->sb.nbuckets, 8);
+		seen[i] = kmalloc(n, GFP_KERNEL);
+		if (!seen[i])
+			goto err;
+
+		/* Disables the seen array until prio_read() uses it too */
+		memset(seen[i], 0xFF, n);
+	}
+
+	ret = btree_root(check_recurse, c, op, seen);
+err:
+	for (int i = 0; i < MAX_CACHES_PER_SET; i++)
+		kfree(seen[i]);
+	return ret;
+}
+
+/* Btree insertion */
+
+static void shift_keys(struct btree *b, struct bkey *where, struct bkey *insert)
+{
+	struct bset *i = b->sets[b->nsets].data;
+
+	memmove((uint64_t *) where + bkey_u64s(insert),
+		where,
+		(void *) end(i) - (void *) where);
+
+	i->keys += bkey_u64s(insert);
+	bkey_copy(where, insert);
+	bch_bset_fix_lookup_table(b, where);
+}
+
+static bool fix_overlapping_extents(struct btree *b,
+				    struct bkey *insert,
+				    struct btree_iter *iter,
+				    struct btree_op *op)
+{
+	void subtract_dirty(struct bkey *k, int sectors)
+	{
+		struct bcache_device *d = b->c->devices[KEY_INODE(k)];
+
+		if (KEY_DIRTY(k) && d)
+			atomic_long_sub(sectors, &d->sectors_dirty);
+	}
+
+	unsigned sectors_found = 0;
+
+	while (1) {
+		struct bkey *k = bch_btree_iter_next(iter);
+		if (!k ||
+		    bkey_cmp(&START_KEY(k), insert) >= 0)
+			break;
+
+		if (bkey_cmp(k, &START_KEY(insert)) <= 0)
+			continue;
+
+		/*
+		 * We might overlap with 0 size extents; we can't skip these
+		 * because if they're in the set we're inserting to we have to
+		 * adjust them so they don't overlap with the key we're
+		 * inserting. But we don't want to check them for BTREE_REPLACE
+		 * operations.
+		 */
+
+		if (op->type == BTREE_REPLACE &&
+		    KEY_SIZE(k)) {
+			/*
+			 * k might have been split since we inserted/found the
+			 * key we're replacing
+			 */
+			uint64_t offset = KEY_START(k) -
+				KEY_START(&op->replace);
+
+			/* But it must be a subset of the replace key */
+			if (KEY_START(k) < KEY_START(&op->replace) ||
+			    KEY_OFFSET(k) > KEY_OFFSET(&op->replace))
+				goto check_failed;
+
+			/* We didn't find a key that we were supposed to */
+			if (KEY_START(k) > KEY_START(insert) + sectors_found)
+				goto check_failed;
+
+			if (KEY_PTRS(&op->replace) != KEY_PTRS(k))
+				goto check_failed;
+
+			/* skip past gen */
+			offset <<= 8;
+
+			BUG_ON(!KEY_PTRS(&op->replace));
+
+			for (unsigned i = 0; i < KEY_PTRS(&op->replace); i++)
+				if (k->ptr[i] != op->replace.ptr[i] + offset)
+					goto check_failed;
+
+			sectors_found = KEY_OFFSET(k) - KEY_START(insert);
+		}
+
+		if (bkey_cmp(insert, k) < 0 &&
+		    bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) {
+			/*
+			 * We overlapped in the middle of an existing key: that
+			 * means we have to split the old key. But we have to do
+			 * slightly different things depending on whether the
+			 * old key has been written out yet.
+			 */
+
+			struct bkey *top;
+
+			subtract_dirty(k, KEY_SIZE(insert));
+
+			if (bkey_written(b, k)) {
+				/*
+				 * We insert a new key to cover the top of the
+				 * old key, and the old key is modified in place
+				 * to represent the bottom split.
+				 *
+				 * It's completely arbitrary whether the new key
+				 * is the top or the bottom, but it has to match
+				 * up with what btree_sort_fixup() does - it
+				 * doesn't check for this kind of overlap, it
+				 * depends on us inserting a new key for the top
+				 * here.
+				 */
+				top = bch_bset_search(b, &b->sets[b->nsets],
+						      insert);
+				shift_keys(b, top, k);
+			} else {
+				BKEY_PADDED(key) temp;
+				bkey_copy(&temp.key, k);
+				shift_keys(b, k, &temp.key);
+				top = bkey_next(k);
+			}
+
+			bch_cut_front(insert, top);
+			bch_cut_back(&START_KEY(insert), k);
+			bch_bset_fix_invalidated_key(b, k);
+			return false;
+		}
+
+		if (bkey_cmp(insert, k) < 0) {
+			subtract_dirty(k, KEY_OFFSET(insert) - KEY_START(k));
+
+			bch_cut_front(insert, k);
+		} else {
+			subtract_dirty(k, KEY_OFFSET(k) - KEY_START(insert));
+
+			if (bkey_written(b, k) &&
+			    bkey_cmp(&START_KEY(insert), &START_KEY(k)) <= 0)
+				/*
+				 * Completely overwrote, so we don't have to
+				 * invalidate the binary search tree
+				 */
+				bch_cut_front(k, k);
+			else {
+				__bch_cut_back(&START_KEY(insert), k);
+				bch_bset_fix_invalidated_key(b, k);
+			}
+		}
+	}
+
+check_failed:
+	if (op->type == BTREE_REPLACE) {
+		if (!sectors_found) {
+			op->insert_collision = true;
+			return true;
+		} else if (sectors_found < KEY_SIZE(insert)) {
+			SET_KEY_OFFSET(insert, KEY_OFFSET(insert) -
+				       (KEY_SIZE(insert) - sectors_found));
+			SET_KEY_SIZE(insert, sectors_found);
+		}
+	}
+
+	return false;
+}
+
+static bool btree_insert_key(struct btree *b, struct btree_op *op,
+			     struct bkey *k)
+{
+	struct bset *i = b->sets[b->nsets].data;
+	struct bkey *m, *prev;
+	const char *status = "insert";
+
+	BUG_ON(bkey_cmp(k, &b->key) > 0);
+	BUG_ON(b->level && !KEY_PTRS(k));
+	BUG_ON(!b->level && !KEY_OFFSET(k));
+
+	if (!b->level) {
+		struct btree_iter iter;
+		struct bkey search = KEY(KEY_INODE(k), KEY_START(k), 0);
+
+		/*
+		 * bset_search() returns the first key that is strictly greater
+		 * than the search key - but for back merging, we want to find
+		 * the first key that is greater than or equal to KEY_START(k) -
+		 * unless KEY_START(k) is 0.
+		 */
+		if (KEY_OFFSET(&search))
+			SET_KEY_OFFSET(&search, KEY_OFFSET(&search) - 1);
+
+		prev = NULL;
+		m = bch_btree_iter_init(b, &iter, &search);
+
+		if (fix_overlapping_extents(b, k, &iter, op))
+			return false;
+
+		while (m != end(i) &&
+		       bkey_cmp(k, &START_KEY(m)) > 0)
+			prev = m, m = bkey_next(m);
+
+		if (key_merging_disabled(b->c))
+			goto insert;
+
+		/* prev is in the tree, if we merge we're done */
+		status = "back merging";
+		if (prev &&
+		    bch_bkey_try_merge(b, prev, k))
+			goto merged;
+
+		status = "overwrote front";
+		if (m != end(i) &&
+		    KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m))
+			goto copy;
+
+		status = "front merge";
+		if (m != end(i) &&
+		    bch_bkey_try_merge(b, k, m))
+			goto copy;
+	} else
+		m = bch_bset_search(b, &b->sets[b->nsets], k);
+
+insert:	shift_keys(b, m, k);
+copy:	bkey_copy(m, k);
+merged:
+	bch_check_keys(b, "%s for %s at %s: %s", status,
+		       op_type(op), pbtree(b), pkey(k));
+	bch_check_key_order_msg(b, i, "%s for %s at %s: %s", status,
+				op_type(op), pbtree(b), pkey(k));
+
+	if (b->level && !KEY_OFFSET(k))
+		b->prio_blocked++;
+
+	pr_debug("%s for %s at %s: %s", status,
+		 op_type(op), pbtree(b), pkey(k));
+
+	return true;
+}
+
+bool bch_btree_insert_keys(struct btree *b, struct btree_op *op)
+{
+	bool ret = false;
+	struct bkey *k;
+	unsigned oldsize = bch_count_data(b);
+
+	while ((k = bch_keylist_pop(&op->keys))) {
+		bkey_put(b->c, k, b->level);
+		ret |= btree_insert_key(b, op, k);
+	}
+
+	BUG_ON(bch_count_data(b) < oldsize);
+	return ret;
+}
+
+bool bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
+				   struct bio *bio)
+{
+	bool ret = false;
+	uint64_t btree_ptr = b->key.ptr[0];
+	unsigned long seq = b->seq;
+	BKEY_PADDED(k) tmp;
+
+	rw_unlock(false, b);
+	rw_lock(true, b, b->level);
+
+	if (b->key.ptr[0] != btree_ptr ||
+	    b->seq != seq + 1 ||
+	    should_split(b))
+		goto out;
+
+	op->replace = KEY(op->inode, bio_end(bio), bio_sectors(bio));
+
+	SET_KEY_PTRS(&op->replace, 1);
+	get_random_bytes(&op->replace.ptr[0], sizeof(uint64_t));
+
+	SET_PTR_DEV(&op->replace, 0, PTR_CHECK_DEV);
+
+	bkey_copy(&tmp.k, &op->replace);
+
+	BUG_ON(op->type != BTREE_INSERT);
+	BUG_ON(!btree_insert_key(b, op, &tmp.k));
+	bch_btree_write(b, false, NULL);
+	ret = true;
+out:
+	downgrade_write(&b->lock);
+	return ret;
+}
+
+static int btree_split(struct btree *b, struct btree_op *op)
+{
+	bool split, root = b == b->c->root;
+	struct btree *n1, *n2 = NULL, *n3 = NULL;
+	uint64_t start_time = local_clock();
+
+	if (b->level)
+		set_closure_blocking(&op->cl);
+
+	n1 = btree_node_alloc_replacement(b, &op->cl);
+	if (IS_ERR(n1))
+		goto err;
+
+	split = set_blocks(n1->sets[0].data, n1->c) > (btree_blocks(b) * 4) / 5;
+
+	pr_debug("%ssplitting at %s keys %i", split ? "" : "not ",
+		 pbtree(b), n1->sets[0].data->keys);
+
+	if (split) {
+		unsigned keys = 0;
+
+		n2 = bch_btree_node_alloc(b->c, b->level, &op->cl);
+		if (IS_ERR(n2))
+			goto err_free1;
+
+		if (root) {
+			n3 = bch_btree_node_alloc(b->c, b->level + 1, &op->cl);
+			if (IS_ERR(n3))
+				goto err_free2;
+		}
+
+		bch_btree_insert_keys(n1, op);
+
+		/* Has to be a linear search because we don't have an auxiliary
+		 * search tree yet
+		 */
+
+		while (keys < (n1->sets[0].data->keys * 3) / 5)
+			keys += bkey_u64s(node(n1->sets[0].data, keys));
+
+		bkey_copy_key(&n1->key, node(n1->sets[0].data, keys));
+		keys += bkey_u64s(node(n1->sets[0].data, keys));
+
+		n2->sets[0].data->keys = n1->sets[0].data->keys - keys;
+		n1->sets[0].data->keys = keys;
+
+		memcpy(n2->sets[0].data->start,
+		       end(n1->sets[0].data),
+		       n2->sets[0].data->keys * sizeof(uint64_t));
+
+		bkey_copy_key(&n2->key, &b->key);
+
+		bch_keylist_add(&op->keys, &n2->key);
+		bch_btree_write(n2, true, op);
+		rw_unlock(true, n2);
+	} else
+		bch_btree_insert_keys(n1, op);
+
+	bch_keylist_add(&op->keys, &n1->key);
+	bch_btree_write(n1, true, op);
+
+	if (n3) {
+		bkey_copy_key(&n3->key, &MAX_KEY);
+		bch_btree_insert_keys(n3, op);
+		bch_btree_write(n3, true, op);
+
+		closure_sync(&op->cl);
+		bch_btree_set_root(n3);
+		rw_unlock(true, n3);
+	} else if (root) {
+		op->keys.top = op->keys.bottom;
+		closure_sync(&op->cl);
+		bch_btree_set_root(n1);
+	} else {
+		bkey_copy(op->keys.top, &b->key);
+		bkey_copy_key(op->keys.top, &ZERO_KEY);
+
+		for (unsigned i = 0; i < KEY_PTRS(&b->key); i++) {
+			uint8_t g = PTR_BUCKET(b->c, &b->key, i)->gen + 1;
+
+			SET_PTR_GEN(op->keys.top, i, g);
+		}
+
+		bch_keylist_push(&op->keys);
+		closure_sync(&op->cl);
+		atomic_inc(&b->c->prio_blocked);
+	}
+
+	rw_unlock(true, n1);
+	btree_node_free(b, op);
+
+	time_stats_update(&b->c->btree_split_time, start_time);
+
+	return 0;
+err_free2:
+	__bkey_put(n2->c, &n2->key);
+	btree_node_free(n2, op);
+	rw_unlock(true, n2);
+err_free1:
+	__bkey_put(n1->c, &n1->key);
+	btree_node_free(n1, op);
+	rw_unlock(true, n1);
+err:
+	if (n3 == ERR_PTR(-EAGAIN) ||
+	    n2 == ERR_PTR(-EAGAIN) ||
+	    n1 == ERR_PTR(-EAGAIN))
+		return -EAGAIN;
+
+	printk(KERN_WARNING "bcache: couldn't split");
+	return -ENOMEM;
+}
+
+static int bch_btree_insert_recurse(struct btree *b, struct btree_op *op,
+				    struct keylist *stack_keys)
+{
+	if (b->level) {
+		int ret;
+		struct bkey *insert = op->keys.bottom;
+		struct bkey *k = bch_next_recurse_key(b, &START_KEY(insert));
+
+		if (!k) {
+			btree_bug(b, "no key to recurse on at level %i/%i",
+				  b->level, b->c->root->level);
+
+			op->keys.top = op->keys.bottom;
+			return -EIO;
+		}
+
+		if (bkey_cmp(insert, k) > 0) {
+			if (op->type == BTREE_REPLACE) {
+				__bkey_put(b->c, insert);
+				op->keys.top = op->keys.bottom;
+				op->insert_collision = true;
+				return 0;
+			}
+
+			for (unsigned i = 0; i < KEY_PTRS(insert); i++)
+				atomic_inc(&PTR_BUCKET(b->c, insert, i)->pin);
+
+			bkey_copy(stack_keys->top, insert);
+
+			bch_cut_back(k, insert);
+			bch_cut_front(k, stack_keys->top);
+
+			bch_keylist_push(stack_keys);
+		}
+
+		ret = btree(insert_recurse, k, b, op, stack_keys);
+		if (ret)
+			return ret;
+	}
+
+	if (!bch_keylist_empty(&op->keys)) {
+		if (should_split(b)) {
+			if (op->lock <= b->c->root->level) {
+				BUG_ON(b->level);
+				op->lock = b->c->root->level + 1;
+				return -EINTR;
+			}
+			return btree_split(b, op);
+		}
+
+		BUG_ON(write_block(b) != b->sets[b->nsets].data);
+
+		if (bch_btree_insert_keys(b, op))
+			bch_btree_write(b, false, op);
+	}
+
+	return 0;
+}
+
+int bch_btree_insert(struct btree_op *op, struct cache_set *c)
+{
+	int ret = 0;
+	struct cache *ca;
+	struct keylist stack_keys;
+
+	/*
+	 * Don't want to block with the btree locked unless we have to,
+	 * otherwise we get deadlocks with try_harder and between split/gc
+	 */
+	clear_closure_blocking(&op->cl);
+
+	BUG_ON(bch_keylist_empty(&op->keys));
+	bch_keylist_copy(&stack_keys, &op->keys);
+	bch_keylist_init(&op->keys);
+
+	while (c->need_gc > MAX_NEED_GC) {
+		closure_lock(&c->gc, &c->cl);
+		btree_gc(&c->gc.cl);
+	}
+
+	for_each_cache(ca, c)
+		while (ca->need_save_prio > MAX_SAVE_PRIO) {
+			mutex_lock(&c->bucket_lock);
+			bch_free_some_buckets(ca);
+			mutex_unlock(&c->bucket_lock);
+
+			closure_wait_event_sync(&c->bucket_wait, &op->cl,
+				ca->need_save_prio <= MAX_SAVE_PRIO ||
+				bch_can_save_prios(ca));
+		}
+
+	while (!bch_keylist_empty(&stack_keys) ||
+	       !bch_keylist_empty(&op->keys)) {
+		if (bch_keylist_empty(&op->keys)) {
+			bch_keylist_add(&op->keys, bch_keylist_pop(&stack_keys));
+			op->lock = 0;
+		}
+
+		ret = btree_root(insert_recurse, c, op, &stack_keys);
+
+		if (ret == -EAGAIN) {
+			ret = 0;
+			closure_sync(&op->cl);
+		} else if (ret) {
+			struct bkey *k;
+
+			printk(KERN_WARNING "bcache: error %i trying to "
+			       "insert key for %s\n", ret, op_type(op));
+
+			while ((k = bch_keylist_pop(&stack_keys) ?:
+				    bch_keylist_pop(&op->keys)))
+				bkey_put(c, k, 0);
+		}
+	}
+
+	bch_keylist_free(&stack_keys);
+
+	if (op->journal)
+		atomic_dec_bug(op->journal);
+	op->journal = NULL;
+	return ret;
+}
+
+void bch_btree_set_root(struct btree *b)
+{
+	BUG_ON(!b->written);
+
+	for (unsigned i = 0; i < KEY_PTRS(&b->key); i++)
+		BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
+
+	mutex_lock(&b->c->bucket_lock);
+	list_del_init(&b->list);
+	mutex_unlock(&b->c->bucket_lock);
+
+	b->c->root = b;
+	__bkey_put(b->c, &b->key);
+
+	bch_journal_meta(b->c, NULL);
+	pr_debug("%s for %pf", pbtree(b), __builtin_return_address(0));
+}
+
+/* Cache lookup */
+
+static int submit_partial_cache_miss(struct btree *b, struct btree_op *op,
+				     struct bkey *k)
+{
+	struct search *s = container_of(op, struct search, op);
+	struct bio *bio = &s->bio.bio;
+	int ret = 0;
+
+	while (!ret &&
+	       !op->lookup_done) {
+		unsigned sectors = INT_MAX;
+
+		if (KEY_INODE(k) == op->inode) {
+			if (KEY_START(k) <= bio->bi_sector)
+				break;
+
+			sectors = min_t(uint64_t, sectors,
+					KEY_START(k) - bio->bi_sector);
+		}
+
+		ret = s->d->cache_miss(b, s, bio, sectors);
+	}
+
+	return ret;
+}
+
+/*
+ * Read from a single key, handling the initial cache miss if the key starts in
+ * the middle of the bio
+ */
+static int submit_partial_cache_hit(struct btree *b, struct btree_op *op,
+				    struct bkey *k)
+{
+	struct search *s = container_of(op, struct search, op);
+	struct bio *bio = &s->bio.bio;
+	unsigned ptr;
+	struct bio *n;
+
+	int ret = submit_partial_cache_miss(b, op, k);
+	if (ret || op->lookup_done)
+		return ret;
+
+	/* XXX: figure out best pointer - for multiple cache devices */
+	ptr = 0;
+
+	PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO;
+
+	while (!op->lookup_done &&
+	       KEY_INODE(k) == op->inode &&
+	       bio->bi_sector < KEY_OFFSET(k)) {
+		struct bkey *bio_key;
+		sector_t sector = PTR_OFFSET(k, ptr) +
+			(bio->bi_sector - KEY_START(k));
+		unsigned sectors = min_t(uint64_t, INT_MAX,
+					 KEY_OFFSET(k) - bio->bi_sector);
+
+		n = bio_split(bio, sectors, GFP_NOIO, s->d->bio_split);
+		if (!n)
+			return -EAGAIN;
+
+		if (n == bio)
+			op->lookup_done = true;
+
+		bio_key = &container_of(n, struct bbio, bio)->key;
+
+		/*
+		 * The bucket we're reading from might be reused while our bio
+		 * is in flight, and we could then end up reading the wrong
+		 * data.
+		 *
+		 * We guard against this by checking (in cache_read_endio()) if
+		 * the pointer is stale again; if so, we treat it as an error
+		 * and reread from the backing device (but we don't pass that
+		 * error up anywhere).
+		 */
+
+		bch_bkey_copy_single_ptr(bio_key, k, ptr);
+		SET_PTR_OFFSET(bio_key, 0, sector);
+
+		n->bi_end_io = bch_cache_read_endio;
+
+		trace_bcache_cache_hit(n);
+		__bch_submit_bbio(n, b->c);
+	}
+
+	return 0;
+}
+
+int bch_btree_search_recurse(struct btree *b, struct btree_op *op)
+{
+	struct search *s = container_of(op, struct search, op);
+	struct bio *bio = &s->bio.bio;
+
+	int ret = 0;
+	struct bkey *k;
+	struct btree_iter iter;
+	bch_btree_iter_init(b, &iter, &KEY(op->inode, bio->bi_sector, 0));
+
+	pr_debug("at %s searching for %u:%llu", pbtree(b), op->inode,
+		 (uint64_t) bio->bi_sector);
+
+	do {
+		k = bch_btree_iter_next(&iter);
+		if (!k) {
+			/*
+			 * b->key would be exactly what we want, except that
+			 * pointers to btree nodes have nonzero size - we
+			 * wouldn't go far enough
+			 */
+
+			ret = submit_partial_cache_miss(b, op,
+					&KEY(KEY_INODE(&b->key),
+					     KEY_OFFSET(&b->key), 0));
+			break;
+		}
+
+		if (bch_ptr_bad(b, k))
+			continue;
+
+		ret = b->level
+			? btree(search_recurse, k, b, op)
+			: submit_partial_cache_hit(b, op, k);
+	} while (!ret &&
+		 !op->lookup_done);
+
+	return ret;
+}
+
+/* Keybuf code */
+
+static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
+{
+	/* Overlapping keys compare equal */
+	if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
+		return -1;
+	if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
+		return 1;
+	return 0;
+}
+
+static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l, struct keybuf_key *r)
+{
+	return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
+}
+
+static int bch_btree_refill_keybuf(struct btree *b, struct btree_op *op,
+				   struct keybuf *buf, struct bkey *end)
+{
+	struct btree_iter iter;
+	bch_btree_iter_init(b, &iter, &buf->last_scanned);
+
+	while (!array_freelist_empty(&buf->freelist)) {
+		struct bkey *k = bch_btree_iter_next(&iter);
+
+		if (!b->level) {
+			if (!k) {
+				buf->last_scanned = b->key;
+				break;
+			}
+
+			buf->last_scanned = *k;
+			if (bkey_cmp(&buf->last_scanned, end) >= 0)
+				break;
+
+			if (bch_ptr_bad(b, k))
+				continue;
+
+			if (buf->key_predicate(buf, k)) {
+				struct keybuf_key *w;
+
+				pr_debug("%s", pkey(k));
+
+				spin_lock(&buf->lock);
+
+				w = array_alloc(&buf->freelist);
+
+				w->private = NULL;
+				bkey_copy(&w->key, k);
+
+				if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
+					array_free(&buf->freelist, w);
+
+				spin_unlock(&buf->lock);
+			}
+		} else {
+			if (!k)
+				break;
+
+			if (bch_ptr_bad(b, k))
+				continue;
+
+			btree(refill_keybuf, k, b, op, buf, end);
+			/*
+			 * Might get an error here, but can't really do anything
+			 * and it'll get logged elsewhere. Just read what we
+			 * can.
+			 */
+
+			if (bkey_cmp(&buf->last_scanned, end) >= 0)
+				break;
+
+			cond_resched();
+		}
+	}
+
+	return 0;
+}
+
+void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
+			  struct bkey *end)
+{
+	struct bkey start = buf->last_scanned;
+	struct btree_op op;
+	bch_btree_op_init_stack(&op);
+
+	btree_root(refill_keybuf, c, &op, buf, end);
+	closure_sync(&op.cl);
+
+	pr_debug("found %s keys from %llu:%llu to %llu:%llu",
+		 RB_EMPTY_ROOT(&buf->keys) ? "no" :
+		 array_freelist_empty(&buf->freelist) ? "some" : "a few",
+		 KEY_INODE(&start), KEY_OFFSET(&start),
+		 KEY_INODE(&buf->last_scanned), KEY_OFFSET(&buf->last_scanned));
+
+	spin_lock(&buf->lock);
+
+	if (!RB_EMPTY_ROOT(&buf->keys)) {
+		struct keybuf_key *w;
+		w = RB_FIRST(&buf->keys, struct keybuf_key, node);
+		buf->start	= START_KEY(&w->key);
+
+		w = RB_LAST(&buf->keys, struct keybuf_key, node);
+		buf->end	= w->key;
+	} else {
+		buf->start	= MAX_KEY;
+		buf->end	= MAX_KEY;
+	}
+
+	spin_unlock(&buf->lock);
+}
+
+static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
+{
+	rb_erase(&w->node, &buf->keys);
+	array_free(&buf->freelist, w);
+}
+
+void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
+{
+	spin_lock(&buf->lock);
+	__bch_keybuf_del(buf, w);
+	spin_unlock(&buf->lock);
+}
+
+bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
+				  struct bkey *end)
+{
+	bool ret = false;
+	struct keybuf_key *p, *w, s;
+	s.key = *start;
+
+	if (bkey_cmp(end, &buf->start) <= 0 ||
+	    bkey_cmp(start, &buf->end) >= 0)
+		return false;
+
+	spin_lock(&buf->lock);
+	w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
+
+	while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
+		p = w;
+		w = RB_NEXT(w, node);
+
+		if (p->private)
+			ret = true;
+		else
+			__bch_keybuf_del(buf, p);
+	}
+
+	spin_unlock(&buf->lock);
+	return ret;
+}
+
+struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
+{
+	struct keybuf_key *w;
+	spin_lock(&buf->lock);
+
+	w = RB_FIRST(&buf->keys, struct keybuf_key, node);
+
+	while (w && w->private)
+		w = RB_NEXT(w, node);
+
+	if (w)
+		w->private = ERR_PTR(-EINTR);
+
+	spin_unlock(&buf->lock);
+	return w;
+}
+
+struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
+					     struct keybuf *buf,
+					     struct bkey *end)
+{
+	struct keybuf_key *ret;
+
+	while (1) {
+		ret = bch_keybuf_next(buf);
+		if (ret)
+			break;
+
+		if (bkey_cmp(&buf->last_scanned, end) >= 0) {
+			pr_debug("scan finished");
+			break;
+		}
+
+		bch_refill_keybuf(c, buf, end);
+	}
+
+	return ret;
+}
+
+void bch_keybuf_init(struct keybuf *buf, keybuf_pred_fn *fn)
+{
+	buf->key_predicate	= fn;
+	buf->last_scanned	= MAX_KEY;
+	buf->keys		= RB_ROOT;
+
+	spin_lock_init(&buf->lock);
+	array_allocator_init(&buf->freelist);
+}
+
+void bch_btree_exit(void)
+{
+	if (btree_io_wq)
+		destroy_workqueue(btree_io_wq);
+	if (bch_gc_wq)
+		destroy_workqueue(bch_gc_wq);
+}
+
+int __init bch_btree_init(void)
+{
+	if (!(bch_gc_wq = create_singlethread_workqueue("bch_btree_gc")) ||
+	    !(btree_io_wq = create_singlethread_workqueue("bch_btree_io")))
+		return -ENOMEM;
+
+	return 0;
+}
diff --git a/drivers/md/bcache/btree.h b/drivers/md/bcache/btree.h
new file mode 100644
index 0000000..9746b8d
--- /dev/null
+++ b/drivers/md/bcache/btree.h
@@ -0,0 +1,423 @@
+#ifndef _BCACHE_BTREE_H
+#define _BCACHE_BTREE_H
+
+/*
+ * THE BTREE:
+ *
+ * At a high level, bcache's btree is relatively standard b+ tree. All keys and
+ * pointers are in the leaves; interior nodes only have pointers to the child
+ * nodes.
+ *
+ * In the interior nodes, a struct bkey always points to a child btree node, and
+ * the key is the highest key in the child node - except that the highest key in
+ * an interior node is always MAX_KEY. The size field refers to the size on disk
+ * of the child node - this would allow us to have variable sized btree nodes
+ * (handy for keeping the depth of the btree 1 by expanding just the root).
+ *
+ * Btree nodes are themselves log structured, but this is hidden fairly
+ * thoroughly. Btree nodes on disk will in practice have extents that overlap
+ * (because they were written at different times), but in memory we never have
+ * overlapping extents - when we read in a btree node from disk, the first thing
+ * we do is resort all the sets of keys with a mergesort, and in the same pass
+ * we check for overlapping extents and adjust them appropriately.
+ *
+ * struct btree_op is a central interface to the btree code. It's used for
+ * specifying read vs. write locking, and the embedded closure is used for
+ * waiting on IO or reserve memory.
+ *
+ * BTREE CACHE:
+ *
+ * Btree nodes are cached in memory; traversing the btree might require reading
+ * in btree nodes which is handled mostly transparently.
+ *
+ * bch_btree_node_get() looks up a btree node in the cache and reads it in from
+ * disk if necessary. This function is almost never called directly though - the
+ * btree() macro is used to get a btree node, call some function on it, and
+ * unlock the node after the function returns.
+ *
+ * The root is special cased - it's taken out of the cache's lru (thus pinning
+ * it in memory), so we can find the root of the btree by just dereferencing a
+ * pointer instead of looking it up in the cache. This makes locking a bit
+ * tricky, since the root pointer is protected by the lock in the btree node it
+ * points to - the btree_root() macro handles this.
+ *
+ * In various places we must be able to allocate memory for multiple btree nodes
+ * in order to make forward progress. To do this we use the btree cache itself
+ * as a reserve; if __get_free_pages() fails, we'll find a node in the btree
+ * cache we can reuse. We can't allow more than one thread to be doing this at a
+ * time, so there's a lock, implemented by a pointer to the btree_op closure -
+ * this allows the btree_root() macro to implicitly release this lock.
+ *
+ * BTREE IO:
+ *
+ * Btree nodes never have to be explicitly read in; bch_btree_node_get() handles
+ * this.
+ *
+ * For writing, we have two btree_write structs embeddded in struct btree - one
+ * write in flight, and one being set up, and we toggle between them.
+ *
+ * Writing is done with a single function -  bch_btree_write() really serves two
+ * different purposes and should be broken up into two different functions. When
+ * passing now = false, it merely indicates that the node is now dirty - calling
+ * it ensures that the dirty keys will be written at some point in the future.
+ *
+ * When passing now = true, bch_btree_write() causes a write to happen
+ * "immediately" (if there was already a write in flight, it'll cause the write
+ * to happen as soon as the previous write completes). It returns immediately
+ * though - but it takes a refcount on the closure in struct btree_op you passed
+ * to it, so a closure_sync() later can be used to wait for the write to
+ * complete.
+ *
+ * This is handy because btree_split() and garbage collection can issue writes
+ * in parallel, reducing the amount of time they have to hold write locks.
+ *
+ * LOCKING:
+ *
+ * When traversing the btree, we may need write locks starting at some level -
+ * inserting a key into the btree will typically only require a write lock on
+ * the leaf node.
+ *
+ * This is specified with the lock field in struct btree_op; lock = 0 means we
+ * take write locks at level <= 0, i.e. only leaf nodes. bch_btree_node_get()
+ * checks this field and returns the node with the appropriate lock held.
+ *
+ * If, after traversing the btree, the insertion code discovers it has to split
+ * then it must restart from the root and take new locks - to do this it changes
+ * the lock field and returns -EINTR, which causes the btree_root() macro to
+ * loop.
+ *
+ * Handling cache misses require a different mechanism for upgrading to a write
+ * lock. We do cache lookups with only a read lock held, but if we get a cache
+ * miss and we wish to insert this data into the cache, we have to insert a
+ * placeholder key to detect races - otherwise, we could race with a write and
+ * overwrite the data that was just written to the cache with stale data from
+ * the backing device.
+ *
+ * For this we use a sequence number that write locks and unlocks increment - to
+ * insert the check key it unlocks the btree node and then takes a write lock,
+ * and fails if the sequence number doesn't match.
+ */
+
+#include "bset.h"
+#include "debug.h"
+
+struct btree_write {
+	struct closure		*owner;
+	atomic_t		*journal;
+
+	/* If btree_split() frees a btree node, it writes a new pointer to that
+	 * btree node indicating it was freed; it takes a refcount on
+	 * c->prio_blocked because we can't write the gens until the new
+	 * pointer is on disk. This allows btree_write_endio() to release the
+	 * refcount that btree_split() took.
+	 */
+	int			prio_blocked;
+};
+
+struct btree {
+	/* Hottest entries first */
+	struct hlist_node	hash;
+
+	/* Key/pointer for this btree node */
+	BKEY_PADDED(key);
+
+	/* Single bit - set when accessed, cleared by shrinker */
+	unsigned long		accessed;
+	unsigned long		seq;
+	struct rw_semaphore	lock;
+	struct cache_set	*c;
+
+	unsigned long		flags;
+	uint16_t		written;	/* would be nice to kill */
+	uint8_t			level;
+	uint8_t			nsets;
+	uint8_t			page_order;
+
+	/*
+	 * Set of sorted keys - the real btree node - plus a binary search tree
+	 *
+	 * sets[0] is special; set[0]->tree, set[0]->prev and set[0]->data point
+	 * to the memory we have allocated for this btree node. Additionally,
+	 * set[0]->data points to the entire btree node as it exists on disk.
+	 */
+	struct bset_tree	sets[MAX_BSETS];
+
+	/* Used to refcount bio splits, also protects b->bio */
+	struct closure_with_waitlist	io;
+
+	/* Gets transferred to w->prio_blocked - see the comment there */
+	int			prio_blocked;
+
+	struct list_head	list;
+	struct delayed_work	work;
+
+	uint64_t		io_start_time;
+	struct btree_write	writes[2];
+	struct bio		*bio;
+};
+
+#define BTREE_FLAG(flag)						\
+static inline bool btree_node_ ## flag(struct btree *b)			\
+{	return test_bit(BTREE_NODE_ ## flag, &b->flags); }		\
+									\
+static inline void set_btree_node_ ## flag(struct btree *b)		\
+{	set_bit(BTREE_NODE_ ## flag, &b->flags); }			\
+
+enum btree_flags {
+	BTREE_NODE_read_done,
+	BTREE_NODE_io_error,
+	BTREE_NODE_dirty,
+	BTREE_NODE_write_idx,
+};
+
+BTREE_FLAG(read_done);
+BTREE_FLAG(io_error);
+BTREE_FLAG(dirty);
+BTREE_FLAG(write_idx);
+
+static inline struct btree_write *btree_current_write(struct btree *b)
+{
+	return b->writes + btree_node_write_idx(b);
+}
+
+static inline struct btree_write *btree_prev_write(struct btree *b)
+{
+	return b->writes + (btree_node_write_idx(b) ^ 1);
+}
+
+static inline unsigned bset_offset(struct btree *b, struct bset *i)
+{
+	return (((size_t) i) - ((size_t) b->sets->data)) >> 9;
+}
+
+static inline struct bset *write_block(struct btree *b)
+{
+	return ((void *) b->sets[0].data) + b->written * block_bytes(b->c);
+}
+
+static inline bool bset_written(struct btree *b, struct bset_tree *t)
+{
+	return t->data < write_block(b);
+}
+
+static inline bool bkey_written(struct btree *b, struct bkey *k)
+{
+	return k < write_block(b)->start;
+}
+
+static inline void set_gc_sectors(struct cache_set *c)
+{
+	atomic_set(&c->sectors_to_gc, c->sb.bucket_size * c->nbuckets / 8);
+}
+
+static inline bool bch_ptr_invalid(struct btree *b, const struct bkey *k)
+{
+	return __bch_ptr_invalid(b->c, b->level, k);
+}
+
+static inline struct bkey *bch_btree_iter_init(struct btree *b,
+					       struct btree_iter *iter,
+					       struct bkey *search)
+{
+	return __bch_btree_iter_init(b, iter, search, b->sets);
+}
+
+/* Looping macros */
+
+#define for_each_cached_btree(b, cursor, c)				\
+	for (unsigned _i = 0;						\
+	     _i < ARRAY_SIZE((c)->bucket_hash);				\
+	     _i++)							\
+		hlist_for_each_entry_rcu((b), cursor,			\
+					 (c)->bucket_hash + _i, hash)
+
+#define for_each_sorted_set_start(b, i, start)				\
+	for (int _i = start; i = (b)->sets[_i].data, _i <= (b)->nsets; _i++)
+
+#define for_each_sorted_set(b, i)	for_each_sorted_set_start(b, i, 0)
+
+#define bkey_filter(b, i, k, filter)					\
+({									\
+	while (k < end(i) && filter(b, k))				\
+		k = bkey_next(k);					\
+	k;								\
+})
+
+#define all_keys(b, k)		0
+
+#define for_each_key_filter(b, k, filter)				\
+	for (struct bset_tree *_t = (b)->sets;				\
+	     _t <= &(b)->sets[(b)->nsets];				\
+	     _t++)							\
+		for (k = _t->data->start;				\
+		     (k = bkey_filter(b, _t->data, k, filter))		\
+			< end(_t->data);				\
+		     k = bkey_next(k))
+
+#define for_each_key(b, k)	for_each_key_filter(b, k, all_keys)
+
+/* Recursing down the btree */
+
+struct btree_op {
+	struct closure		cl;
+	struct cache_set	*c;
+
+	/* Journal entry we have a refcount on */
+	atomic_t		*journal;
+
+	/* Bio to be inserted into the cache */
+	struct bio		*cache_bio;
+
+	unsigned		inode;
+
+	uint16_t		write_prio;
+
+	/* Btree level at which we start taking write locks */
+	short			lock;
+
+	/* Btree insertion type */
+	enum {
+		BTREE_INSERT,
+		BTREE_REPLACE
+	} type:8;
+
+	unsigned		csum:1;
+	unsigned		skip:1;
+	unsigned		flush_journal:1;
+
+	unsigned		insert_data_done:1;
+	unsigned		lookup_done:1;
+	unsigned		insert_collision:1;
+
+	/* Anything after this point won't get zeroed in do_bio_hook() */
+
+	/* Keys to be inserted */
+	struct keylist		keys;
+	BKEY_PADDED(replace);
+};
+
+void bch_btree_op_init_stack(struct btree_op *);
+
+static inline void rw_lock(bool w, struct btree *b, int level)
+{
+	w ? down_write_nested(&b->lock, level + 1)
+	  : down_read_nested(&b->lock, level + 1);
+	if (w)
+		b->seq++;
+}
+
+static inline void rw_unlock(bool w, struct btree *b)
+{
+#ifdef CONFIG_BCACHE_EDEBUG
+	unsigned i;
+
+	if (w &&
+	    b->key.ptr[0] &&
+	    btree_node_read_done(b))
+		for (i = 0; i <= b->nsets; i++)
+			bch_check_key_order(b, b->sets[i].data);
+#endif
+
+	if (w)
+		b->seq++;
+	(w ? up_write : up_read)(&b->lock);
+}
+
+#define insert_lock(s, b)	((b)->level <= (s)->lock)
+
+/*
+ * These macros are for recursing down the btree - they handle the details of
+ * locking and looking up nodes in the cache for you. They're best treated as
+ * mere syntax when reading code that uses them.
+ *
+ * op->lock determines whether we take a read or a write lock at a given depth.
+ * If you've got a read lock and find that you need a write lock (i.e. you're
+ * going to have to split), set op->lock and return -EINTR; btree_root() will
+ * call you again and you'll have the correct lock.
+ */
+
+/**
+ * btree - recurse down the btree on a specified key
+ * @fn:		function to call, which will be passed the child node
+ * @key:	key to recurse on
+ * @b:		parent btree node
+ * @op:		pointer to struct btree_op
+ */
+#define btree(fn, key, b, op, ...)					\
+({									\
+	int _r, l = (b)->level - 1;					\
+	bool _w = l <= (op)->lock;					\
+	struct btree *_b = bch_btree_node_get((b)->c, key, l, op);	\
+	if (!IS_ERR(_b)) {						\
+		_r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__);		\
+		rw_unlock(_w, _b);					\
+	} else								\
+		_r = PTR_ERR(_b);					\
+	_r;								\
+})
+
+/**
+ * btree_root - call a function on the root of the btree
+ * @fn:		function to call, which will be passed the child node
+ * @c:		cache set
+ * @op:		pointer to struct btree_op
+ */
+#define btree_root(fn, c, op, ...)					\
+({									\
+	int _r = -EINTR;						\
+	do {								\
+		struct btree *_b = (c)->root;				\
+		bool _w = insert_lock(op, _b);				\
+		rw_lock(_w, _b, _b->level);				\
+		if (_b == (c)->root &&					\
+		    _w == insert_lock(op, _b))				\
+			_r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__);	\
+		rw_unlock(_w, _b);					\
+		bch_cannibalize_unlock(c, &(op)->cl);		\
+	} while (_r == -EINTR);						\
+									\
+	_r;								\
+})
+
+static inline bool should_split(struct btree *b)
+{
+	struct bset *i = write_block(b);
+	return b->written >= btree_blocks(b) ||
+		(i->seq == b->sets[0].data->seq &&
+		 b->written + __set_blocks(i, i->keys + 15, b->c)
+		 > btree_blocks(b));
+}
+
+void bch_btree_read_done(struct closure *);
+void bch_btree_read(struct btree *);
+void bch_btree_write(struct btree *b, bool now, struct btree_op *op);
+
+void bch_cannibalize_unlock(struct cache_set *, struct closure *);
+void bch_btree_set_root(struct btree *);
+struct btree *bch_btree_node_alloc(struct cache_set *, int, struct closure *);
+struct btree *bch_btree_node_get(struct cache_set *, struct bkey *,
+				int, struct btree_op *);
+
+bool bch_btree_insert_keys(struct btree *, struct btree_op *);
+bool bch_btree_insert_check_key(struct btree *, struct btree_op *,
+				   struct bio *);
+int bch_btree_insert(struct btree_op *, struct cache_set *);
+
+int bch_btree_search_recurse(struct btree *, struct btree_op *);
+
+void bch_queue_gc(struct cache_set *);
+size_t bch_btree_gc_finish(struct cache_set *);
+void bch_moving_gc(struct closure *);
+int bch_btree_check(struct cache_set *, struct btree_op *);
+uint8_t __bch_btree_mark_key(struct cache_set *, int, struct bkey *);
+
+void bch_keybuf_init(struct keybuf *, keybuf_pred_fn *);
+void bch_refill_keybuf(struct cache_set *, struct keybuf *, struct bkey *);
+bool bch_keybuf_check_overlapping(struct keybuf *, struct bkey *,
+				  struct bkey *);
+void bch_keybuf_del(struct keybuf *, struct keybuf_key *);
+struct keybuf_key *bch_keybuf_next(struct keybuf *);
+struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *,
+					  struct keybuf *, struct bkey *);
+
+#endif
-- 
1.7.7.3


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]