Update bcachefs sources to 99750eab4d bcachefs: Persist stripe blocks_used

37 files changed, 916 insertions, 401 deletions

diff --git a/.bcachefs_revision b/.bcachefs_revision
index 088f645c..8eca0593 100644
--- a/.bcachefs_revision
+++ b/.bcachefs_revision
@@ -1 +1 @@
-bcca1c557b1897ecc3aeb1f89ab91865487d91ab
+99750eab4d583132cf61f071082c7cf21f5295c0
diff --git a/include/asm/page.h b/include/asm/page.h
new file mode 100644
index 00000000..e69de29b
--- /dev/null
+++ b/include/asm/page.h
diff --git a/include/linux/atomic.h b/include/linux/atomic.h
index 7471bd97..38a364c0 100644
--- a/include/linux/atomic.h
+++ b/include/linux/atomic.h
@@ -37,6 +37,7 @@ typedef struct {
 #define xchg_acquire(p, v)		uatomic_xchg(p, v)
 #define cmpxchg(p, old, new)		uatomic_cmpxchg(p, old, new)
 #define cmpxchg_acquire(p, old, new)	uatomic_cmpxchg(p, old, new)
+#define cmpxchg_release(p, old, new)	uatomic_cmpxchg(p, old, new)
 
 #define smp_mb__before_atomic()		cmm_smp_mb__before_uatomic_add()
 #define smp_mb__after_atomic()		cmm_smp_mb__after_uatomic_add()
@@ -77,6 +78,16 @@ typedef struct {
 	__old;							\
 })
 
+#define cmpxchg_release(p, old, new)				\
+({								\
+	typeof(*(p)) __old = (old);				\
+								\
+	__atomic_compare_exchange_n((p), &__old, new, false,	\
+				    __ATOMIC_RELEASE,		\
+				    __ATOMIC_RELEASE);		\
+	__old;							\
+})
+
 #define smp_mb__before_atomic()	__atomic_thread_fence(__ATOMIC_SEQ_CST)
 #define smp_mb__after_atomic()	__atomic_thread_fence(__ATOMIC_SEQ_CST)
 #define smp_wmb()		__atomic_thread_fence(__ATOMIC_SEQ_CST)
diff --git a/include/linux/generic-radix-tree.h b/include/linux/generic-radix-tree.h
index 7f637e17..3a91130a 100644
--- a/include/linux/generic-radix-tree.h
+++ b/include/linux/generic-radix-tree.h
@@ -1,34 +1,60 @@
 #ifndef _LINUX_GENERIC_RADIX_TREE_H
 #define _LINUX_GENERIC_RADIX_TREE_H
 
-/*
- * Generic radix trees/sparse arrays:
+/**
+ * DOC: Generic radix trees/sparse arrays:
+ *
+ * Very simple and minimalistic, supporting arbitrary size entries up to
+ * PAGE_SIZE.
+ *
+ * A genradix is defined with the type it will store, like so:
+ *
+ * static GENRADIX(struct foo) foo_genradix;
+ *
+ * The main operations are:
+ *
+ * - genradix_init(radix) - initialize an empty genradix
+ *
+ * - genradix_free(radix) - free all memory owned by the genradix and
+ *   reinitialize it
+ *
+ * - genradix_ptr(radix, idx) - gets a pointer to the entry at idx, returning
+ *   NULL if that entry does not exist
  *
- * A generic radix tree has all nodes of size PAGE_SIZE - both leaves and
- * interior nodes.
+ * - genradix_ptr_alloc(radix, idx, gfp) - gets a pointer to an entry,
+ *   allocating it if necessary
+ *
+ * - genradix_for_each(radix, iter, p) - iterate over each entry in a genradix
+ *
+ * The radix tree allocates one page of entries at a time, so entries may exist
+ * that were never explicitly allocated - they will be initialized to all
+ * zeroes.
+ *
+ * Internally, a genradix is just a radix tree of pages, and indexing works in
+ * terms of byte offsets. The wrappers in this header file use sizeof on the
+ * type the radix contains to calculate a byte offset from the index - see
+ * __idx_to_offset.
  */
 
+#include <asm/page.h>
 #include <linux/bug.h>
 #include <linux/kernel.h>
 #include <linux/log2.h>
 
-struct genradix_node;
+struct genradix_root;
 
 struct __genradix {
-	struct genradix_node		*root;
-	size_t				depth;
+	struct genradix_root __rcu	*root;
 };
 
 /*
- * NOTE: currently, sizeof(_type) must be a power of two and not larger than
- * PAGE_SIZE:
+ * NOTE: currently, sizeof(_type) must not be larger than PAGE_SIZE:
  */
 
 #define __GENRADIX_INITIALIZER					\
 	{							\
 		.tree = {					\
 			.root = NULL,				\
-			.depth = 0,				\
 		}						\
 	}
 
@@ -49,6 +75,12 @@ struct {							\
 #define DEFINE_GENRADIX(_name, _type)				\
 	GENRADIX(_type) _name = __GENRADIX_INITIALIZER
 
+/**
+ * genradix_init - initialize a genradix
+ * @_radix:	genradix to initialize
+ *
+ * Does not fail
+ */
 #define genradix_init(_radix)					\
 do {								\
 	*(_radix) = (typeof(*_radix)) __GENRADIX_INITIALIZER;	\
@@ -56,11 +88,20 @@ do {								\
 
 void __genradix_free(struct __genradix *);
 
+/**
+ * genradix_free: free all memory owned by a genradix
+ * @_radix: the genradix to free
+ *
+ * After freeing, @_radix will be reinitialized and empty
+ */
 #define genradix_free(_radix)	__genradix_free(&(_radix)->tree)
 
 static inline size_t __idx_to_offset(size_t idx, size_t obj_size)
 {
-	BUILD_BUG_ON(obj_size > PAGE_SIZE);
+	if (__builtin_constant_p(obj_size))
+		BUILD_BUG_ON(obj_size > PAGE_SIZE);
+	else
+		BUG_ON(obj_size > PAGE_SIZE);
 
 	if (!is_power_of_2(obj_size)) {
 		size_t objs_per_page = PAGE_SIZE / obj_size;
@@ -79,7 +120,13 @@ static inline size_t __idx_to_offset(size_t idx, size_t obj_size)
 
 void *__genradix_ptr(struct __genradix *, size_t);
 
-/* Returns a pointer to element at @_idx */
+/**
+ * genradix_ptr - get a pointer to a genradix entry
+ * @_radix:	genradix to access
+ * @_idx:	index to fetch
+ *
+ * Returns a pointer to entry at @_idx, or NULL if that entry does not exist.
+ */
 #define genradix_ptr(_radix, _idx)				\
 	(__genradix_cast(_radix)				\
 	 __genradix_ptr(&(_radix)->tree,			\
@@ -87,7 +134,15 @@ void *__genradix_ptr(struct __genradix *, size_t);
 
 void *__genradix_ptr_alloc(struct __genradix *, size_t, gfp_t);
 
-/* Returns a pointer to element at @_idx, allocating it if necessary */
+/**
+ * genradix_ptr_alloc - get a pointer to a genradix entry, allocating it
+ *			if necessary
+ * @_radix:	genradix to access
+ * @_idx:	index to fetch
+ * @_gfp:	gfp mask
+ *
+ * Returns a pointer to entry at @_idx, or NULL on allocation failure
+ */
 #define genradix_ptr_alloc(_radix, _idx, _gfp)			\
 	(__genradix_cast(_radix)				\
 	 __genradix_ptr_alloc(&(_radix)->tree,			\
@@ -99,6 +154,11 @@ struct genradix_iter {
 	size_t			pos;
 };
 
+/**
+ * genradix_iter_init - initialize a genradix_iter
+ * @_radix:	genradix that will be iterated over
+ * @_idx:	index to start iterating from
+ */
 #define genradix_iter_init(_radix, _idx)			\
 	((struct genradix_iter) {				\
 		.pos	= (_idx),				\
@@ -107,6 +167,14 @@ struct genradix_iter {
 
 void *__genradix_iter_peek(struct genradix_iter *, struct __genradix *, size_t);
 
+/**
+ * genradix_iter_peek - get first entry at or above iterator's current
+ *			position
+ * @_iter:	a genradix_iter
+ * @_radix:	genradix being iterated over
+ *
+ * If no more entries exist at or above @_iter's current position, returns NULL
+ */
 #define genradix_iter_peek(_iter, _radix)			\
 	(__genradix_cast(_radix)				\
 	 __genradix_iter_peek(_iter, &(_radix)->tree,		\
@@ -127,4 +195,37 @@ static inline void __genradix_iter_advance(struct genradix_iter *iter,
 #define genradix_iter_advance(_iter, _radix)			\
 	__genradix_iter_advance(_iter, __genradix_obj_size(_radix))
 
+#define genradix_for_each_from(_radix, _iter, _p, _start)	\
+	for (_iter = genradix_iter_init(_radix, _start);	\
+	     (_p = genradix_iter_peek(&_iter, _radix)) != NULL;	\
+	     genradix_iter_advance(&_iter, _radix))
+
+/**
+ * genradix_for_each - iterate over entry in a genradix
+ * @_radix:	genradix to iterate over
+ * @_iter:	a genradix_iter to track current position
+ * @_p:		pointer to genradix entry type
+ *
+ * On every iteration, @_p will point to the current entry, and @_iter.pos
+ * will be the current entry's index.
+ */
+#define genradix_for_each(_radix, _iter, _p)			\
+	genradix_for_each_from(_radix, _iter, _p, 0)
+
+int __genradix_prealloc(struct __genradix *, size_t, gfp_t);
+
+/**
+ * genradix_prealloc - preallocate entries in a generic radix tree
+ * @_radix:	genradix to preallocate
+ * @_nr:	number of entries to preallocate
+ * @_gfp:	gfp mask
+ *
+ * Returns 0 on success, -ENOMEM on failure
+ */
+#define genradix_prealloc(_radix, _nr, _gfp)			\
+	 __genradix_prealloc(&(_radix)->tree,			\
+			__genradix_idx_to_offset(_radix, _nr + 1),\
+			_gfp)
+
+
 #endif /* _LINUX_GENERIC_RADIX_TREE_H */
diff --git a/libbcachefs/alloc_background.c b/libbcachefs/alloc_background.c
index 6de6e263..2552d457 100644
--- a/libbcachefs/alloc_background.c
+++ b/libbcachefs/alloc_background.c
@@ -249,6 +249,9 @@ int bch2_alloc_read(struct bch_fs *c, struct list_head *journal_replay_list)
 				bch2_alloc_read_key(c, bkey_i_to_s_c(k));
 	}
 
+	for_each_member_device(ca, c, i)
+		bch2_dev_usage_from_buckets(c, ca);
+
 	mutex_lock(&c->bucket_clock[READ].lock);
 	for_each_member_device(ca, c, i) {
 		down_read(&ca->bucket_lock);
@@ -280,35 +283,51 @@ static int __bch2_alloc_write_key(struct bch_fs *c, struct bch_dev *ca,
 #endif
 	struct bkey_i_alloc *a = bkey_alloc_init(&alloc_key.k);
 	struct bucket *g;
-	struct bucket_mark m;
+	struct bucket_mark m, new;
 	int ret;
 
 	BUG_ON(BKEY_ALLOC_VAL_U64s_MAX > 8);
 
 	a->k.p = POS(ca->dev_idx, b);
 
+	bch2_btree_iter_set_pos(iter, a->k.p);
+
+	ret = bch2_btree_iter_traverse(iter);
+	if (ret)
+		return ret;
+
 	percpu_down_read_preempt_disable(&c->mark_lock);
 	g = bucket(ca, b);
-	m = bucket_cmpxchg(g, m, m.dirty = false);
+	m = READ_ONCE(g->mark);
+
+	if (!m.dirty) {
+		percpu_up_read_preempt_enable(&c->mark_lock);
+		return 0;
+	}
 
 	__alloc_write_key(a, g, m);
 	percpu_up_read_preempt_enable(&c->mark_lock);
 
 	bch2_btree_iter_cond_resched(iter);
 
-	bch2_btree_iter_set_pos(iter, a->k.p);
-
 	ret = bch2_btree_insert_at(c, NULL, journal_seq,
+				   BTREE_INSERT_NOCHECK_RW|
 				   BTREE_INSERT_NOFAIL|
 				   BTREE_INSERT_USE_RESERVE|
 				   BTREE_INSERT_USE_ALLOC_RESERVE|
 				   flags,
 				   BTREE_INSERT_ENTRY(iter, &a->k_i));
+	if (ret)
+		return ret;
+
+	new = m;
+	new.dirty = false;
+	atomic64_cmpxchg(&g->_mark.v, m.v.counter, new.v.counter);
 
-	if (!ret && ca->buckets_written)
+	if (ca->buckets_written)
 		set_bit(b, ca->buckets_written);
 
-	return ret;
+	return 0;
 }
 
 int bch2_alloc_replay_key(struct bch_fs *c, struct bkey_i *k)
@@ -898,10 +917,19 @@ static int push_invalidated_bucket(struct bch_fs *c, struct bch_dev *ca, size_t
 		for (i = 0; i < RESERVE_NR; i++)
 			if (fifo_push(&ca->free[i], bucket)) {
 				fifo_pop(&ca->free_inc, bucket);
+
 				closure_wake_up(&c->freelist_wait);
+				ca->allocator_blocked_full = false;
+
 				spin_unlock(&c->freelist_lock);
 				goto out;
 			}
+
+		if (!ca->allocator_blocked_full) {
+			ca->allocator_blocked_full = true;
+			closure_wake_up(&c->freelist_wait);
+		}
+
 		spin_unlock(&c->freelist_lock);
 
 		if ((current->flags & PF_KTHREAD) &&
@@ -1226,6 +1254,11 @@ void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
 			set_bit(ca->dev_idx, c->rw_devs[i].d);
 }
 
+void bch2_dev_allocator_quiesce(struct bch_fs *c, struct bch_dev *ca)
+{
+	closure_wait_event(&c->freelist_wait, ca->allocator_blocked_full);
+}
+
 /* stop allocator thread: */
 void bch2_dev_allocator_stop(struct bch_dev *ca)
 {
@@ -1333,6 +1366,24 @@ static void allocator_start_issue_discards(struct bch_fs *c)
 					     ca->mi.bucket_size, GFP_NOIO, 0);
 }
 
+static int resize_free_inc(struct bch_dev *ca)
+{
+	alloc_fifo free_inc;
+
+	if (!fifo_full(&ca->free_inc))
+		return 0;
+
+	if (!init_fifo(&free_inc,
+		       ca->free_inc.size * 2,
+		       GFP_KERNEL))
+		return -ENOMEM;
+
+	fifo_move(&free_inc, &ca->free_inc);
+	swap(free_inc, ca->free_inc);
+	free_fifo(&free_inc);
+	return 0;
+}
+
 static int __bch2_fs_allocator_start(struct bch_fs *c)
 {
 	struct bch_dev *ca;
@@ -1408,6 +1459,12 @@ not_enough:
 
 			while (!fifo_full(&ca->free[RESERVE_BTREE]) &&
 			       (bu = next_alloc_bucket(ca)) >= 0) {
+				ret = resize_free_inc(ca);
+				if (ret) {
+					percpu_ref_put(&ca->io_ref);
+					return ret;
+				}
+
 				bch2_invalidate_one_bucket(c, ca, bu,
 							   &journal_seq);
 
diff --git a/libbcachefs/alloc_background.h b/libbcachefs/alloc_background.h
index a0c08e34..26561b3b 100644
--- a/libbcachefs/alloc_background.h
+++ b/libbcachefs/alloc_background.h
@@ -51,6 +51,7 @@ void bch2_recalc_capacity(struct bch_fs *);
 void bch2_dev_allocator_remove(struct bch_fs *, struct bch_dev *);
 void bch2_dev_allocator_add(struct bch_fs *, struct bch_dev *);
 
+void bch2_dev_allocator_quiesce(struct bch_fs *, struct bch_dev *);
 void bch2_dev_allocator_stop(struct bch_dev *);
 int bch2_dev_allocator_start(struct bch_dev *);
 
diff --git a/libbcachefs/alloc_foreground.c b/libbcachefs/alloc_foreground.c
index 14e6453b..f2f9015d 100644
--- a/libbcachefs/alloc_foreground.c
+++ b/libbcachefs/alloc_foreground.c
@@ -106,6 +106,7 @@ void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
 	bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr),
 			       false, gc_pos_alloc(c, ob), 0);
 	ob->valid = false;
+	ob->type = 0;
 
 	spin_unlock(&ob->lock);
 	percpu_up_read_preempt_enable(&c->mark_lock);
@@ -141,6 +142,7 @@ static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
 	ob = c->open_buckets + c->open_buckets_freelist;
 	c->open_buckets_freelist = ob->freelist;
 	atomic_set(&ob->pin, 1);
+	ob->type = 0;
 
 	c->open_buckets_nr_free--;
 	return ob;
@@ -209,9 +211,9 @@ static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
 	case RESERVE_ALLOC:
 		return 0;
 	case RESERVE_BTREE:
-		return BTREE_NODE_RESERVE / 2;
+		return BTREE_NODE_OPEN_BUCKET_RESERVE;
 	default:
-		return BTREE_NODE_RESERVE;
+		return BTREE_NODE_OPEN_BUCKET_RESERVE * 2;
 	}
 }
 
@@ -837,15 +839,17 @@ struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
 {
 	struct write_point *wp;
 	struct open_bucket *ob;
-	unsigned nr_effective = 0;
-	struct open_buckets ptrs = { .nr = 0 };
-	bool have_cache = false;
-	unsigned write_points_nr;
-	int ret = 0, i;
+	struct open_buckets ptrs;
+	unsigned nr_effective, write_points_nr;
+	bool have_cache;
+	int ret, i;
 
 	BUG_ON(!nr_replicas || !nr_replicas_required);
 retry:
+	ptrs.nr		= 0;
+	nr_effective	= 0;
 	write_points_nr = c->write_points_nr;
+	have_cache	= false;
 
 	wp = writepoint_find(c, write_point.v);
 
diff --git a/libbcachefs/alloc_foreground.h b/libbcachefs/alloc_foreground.h
index b0e44f75..5224a52f 100644
--- a/libbcachefs/alloc_foreground.h
+++ b/libbcachefs/alloc_foreground.h
@@ -85,6 +85,7 @@ static inline void bch2_open_bucket_get(struct bch_fs *c,
 	unsigned i;
 
 	open_bucket_for_each(c, &wp->ptrs, ob, i) {
+		ob->type = wp->type;
 		atomic_inc(&ob->pin);
 		ob_push(c, ptrs, ob);
 	}
diff --git a/libbcachefs/alloc_types.h b/libbcachefs/alloc_types.h
index 6f17f094..66457fc7 100644
--- a/libbcachefs/alloc_types.h
+++ b/libbcachefs/alloc_types.h
@@ -55,9 +55,10 @@ struct open_bucket {
 	spinlock_t		lock;
 	atomic_t		pin;
 	u8			freelist;
-	bool			valid;
-	bool			on_partial_list;
 	u8			ec_idx;
+	u8			type;
+	unsigned		valid:1;
+	unsigned		on_partial_list:1;
 	unsigned		sectors_free;
 	struct bch_extent_ptr	ptr;
 	struct ec_stripe_new	*ec;
diff --git a/libbcachefs/bcachefs.h b/libbcachefs/bcachefs.h
index 449eb0c1..f42b2f90 100644
--- a/libbcachefs/bcachefs.h
+++ b/libbcachefs/bcachefs.h
@@ -330,6 +330,8 @@ enum bch_time_stats {
 /* Size of the freelist we allocate btree nodes from: */
 #define BTREE_NODE_RESERVE	BTREE_RESERVE_MAX
 
+#define BTREE_NODE_OPEN_BUCKET_RESERVE	(BTREE_RESERVE_MAX * BCH_REPLICAS_MAX)
+
 struct btree;
 
 enum gc_phase {
@@ -426,7 +428,13 @@ struct bch_dev {
 
 	size_t			inc_gen_needs_gc;
 	size_t			inc_gen_really_needs_gc;
+
+	/*
+	 * XXX: this should be an enum for allocator state, so as to include
+	 * error state
+	 */
 	bool			allocator_blocked;
+	bool			allocator_blocked_full;
 
 	alloc_heap		alloc_heap;
 
@@ -597,6 +605,7 @@ struct bch_fs {
 	struct workqueue_struct	*wq;
 	/* copygc needs its own workqueue for index updates.. */
 	struct workqueue_struct	*copygc_wq;
+	struct workqueue_struct	*journal_reclaim_wq;
 
 	/* ALLOCATION */
 	struct delayed_work	pd_controllers_update;
diff --git a/libbcachefs/bkey.c b/libbcachefs/bkey.c
index 25725e42..40ce33a4 100644
--- a/libbcachefs/bkey.c
+++ b/libbcachefs/bkey.c
@@ -1010,11 +1010,8 @@ static inline int __bkey_cmp_bits(const u64 *l, const u64 *r,
 			nr_key_bits -= 64;
 		}
 
-		if (l_v != r_v)
-			return l_v < r_v ? -1 : 1;
-
-		if (!nr_key_bits)
-			return 0;
+		if (!nr_key_bits || l_v != r_v)
+			break;
 
 		l = next_word(l);
 		r = next_word(r);
@@ -1022,6 +1019,8 @@ static inline int __bkey_cmp_bits(const u64 *l, const u64 *r,
 		l_v = *l;
 		r_v = *r;
 	}
+
+	return (l_v > r_v) - (l_v < r_v);
 }
 #endif
 
diff --git a/libbcachefs/btree_gc.c b/libbcachefs/btree_gc.c
index 23013fbb..433e8f22 100644
--- a/libbcachefs/btree_gc.c
+++ b/libbcachefs/btree_gc.c
@@ -483,31 +483,6 @@ static void bch2_gc_free(struct bch_fs *c)
 	percpu_up_write(&c->mark_lock);
 }
 
-/*
- * Accumulate percpu counters onto one cpu's copy - only valid when access
- * against any percpu counter is guarded against
- */
-static u64 *acc_percpu_u64s(u64 __percpu *p, unsigned nr)
-{
-	u64 *ret;
-	int cpu;
-
-	preempt_disable();
-	ret = this_cpu_ptr(p);
-	preempt_enable();
-
-	for_each_possible_cpu(cpu) {
-		u64 *i = per_cpu_ptr(p, cpu);
-
-		if (i != ret) {
-			acc_u64s(ret, i, nr);
-			memset(i, 0, nr * sizeof(u64));
-		}
-	}
-
-	return ret;
-}
-
 static void bch2_gc_done_nocheck(struct bch_fs *c)
 {
 	struct bch_dev *ca;
@@ -543,9 +518,9 @@ static void bch2_gc_done_nocheck(struct bch_fs *c)
 	for_each_member_device(ca, c, i) {
 		unsigned nr = sizeof(struct bch_dev_usage) / sizeof(u64);
 		struct bch_dev_usage *dst = (void *)
-			acc_percpu_u64s((void *) ca->usage[0], nr);
+			bch2_acc_percpu_u64s((void *) ca->usage[0], nr);
 		struct bch_dev_usage *src = (void *)
-			acc_percpu_u64s((void *) ca->usage[1], nr);
+			bch2_acc_percpu_u64s((void *) ca->usage[1], nr);
 
 		*dst = *src;
 	}
@@ -554,9 +529,9 @@ static void bch2_gc_done_nocheck(struct bch_fs *c)
 		unsigned nr = sizeof(struct bch_fs_usage) / sizeof(u64) +
 			c->replicas.nr;
 		struct bch_fs_usage *dst = (void *)
-			acc_percpu_u64s((void *) c->usage[0], nr);
+			bch2_acc_percpu_u64s((void *) c->usage[0], nr);
 		struct bch_fs_usage *src = (void *)
-			acc_percpu_u64s((void *) c->usage[1], nr);
+			bch2_acc_percpu_u64s((void *) c->usage[1], nr);
 
 		memcpy(&dst->s.gc_start[0],
 		       &src->s.gc_start[0],
@@ -582,6 +557,7 @@ static void bch2_gc_done(struct bch_fs *c, bool initial)
 			dst_iter.pos, ##__VA_ARGS__,			\
 			dst->_f, src->_f);				\
 		dst->_f = src->_f;					\
+		dst->dirty = true;					\
 	}
 #define copy_bucket_field(_f)						\
 	if (dst->b[b].mark._f != src->b[b].mark._f) {			\
@@ -612,16 +588,18 @@ static void bch2_gc_done(struct bch_fs *c, bool initial)
 
 		while ((dst = genradix_iter_peek(&dst_iter, &c->stripes[0])) &&
 		       (src = genradix_iter_peek(&src_iter, &c->stripes[1]))) {
+			BUG_ON(src_iter.pos != dst_iter.pos);
+
 			copy_stripe_field(alive,	"alive");
 			copy_stripe_field(sectors,	"sectors");
 			copy_stripe_field(algorithm,	"algorithm");
 			copy_stripe_field(nr_blocks,	"nr_blocks");
 			copy_stripe_field(nr_redundant,	"nr_redundant");
-			copy_stripe_field(blocks_nonempty.counter,
+			copy_stripe_field(blocks_nonempty,
 					  "blocks_nonempty");
 
 			for (i = 0; i < ARRAY_SIZE(dst->block_sectors); i++)
-				copy_stripe_field(block_sectors[i].counter,
+				copy_stripe_field(block_sectors[i],
 						  "block_sectors[%u]", i);
 
 			if (dst->alive)
@@ -656,9 +634,9 @@ static void bch2_gc_done(struct bch_fs *c, bool initial)
 	for_each_member_device(ca, c, i) {
 		unsigned nr = sizeof(struct bch_dev_usage) / sizeof(u64);
 		struct bch_dev_usage *dst = (void *)
-			acc_percpu_u64s((void *) ca->usage[0], nr);
+			bch2_acc_percpu_u64s((void *) ca->usage[0], nr);
 		struct bch_dev_usage *src = (void *)
-			acc_percpu_u64s((void *) ca->usage[1], nr);
+			bch2_acc_percpu_u64s((void *) ca->usage[1], nr);
 		unsigned b;
 
 		for (b = 0; b < BCH_DATA_NR; b++)
@@ -678,9 +656,9 @@ static void bch2_gc_done(struct bch_fs *c, bool initial)
 		unsigned nr = sizeof(struct bch_fs_usage) / sizeof(u64) +
 			c->replicas.nr;
 		struct bch_fs_usage *dst = (void *)
-			acc_percpu_u64s((void *) c->usage[0], nr);
+			bch2_acc_percpu_u64s((void *) c->usage[0], nr);
 		struct bch_fs_usage *src = (void *)
-			acc_percpu_u64s((void *) c->usage[1], nr);
+			bch2_acc_percpu_u64s((void *) c->usage[1], nr);
 
 		copy_fs_field(s.hidden,		"hidden");
 		copy_fs_field(s.data,		"data");
diff --git a/libbcachefs/btree_gc.h b/libbcachefs/btree_gc.h
index 8af5f841..1905acfa 100644
--- a/libbcachefs/btree_gc.h
+++ b/libbcachefs/btree_gc.h
@@ -109,7 +109,7 @@ static inline bool gc_visited(struct bch_fs *c, struct gc_pos pos)
 
 	do {
 		seq = read_seqcount_begin(&c->gc_pos_lock);
-		ret = gc_pos_cmp(pos, c->gc_pos) <= 0;
+		ret = gc_pos_cmp(pos, c->gc_pos) < 0;
 	} while (read_seqcount_retry(&c->gc_pos_lock, seq));
 
 	return ret;
diff --git a/libbcachefs/btree_update.h b/libbcachefs/btree_update.h
index dd9d2559..4bd07258 100644
--- a/libbcachefs/btree_update.h
+++ b/libbcachefs/btree_update.h
@@ -77,6 +77,7 @@ enum {
 	__BTREE_INSERT_ATOMIC,
 	__BTREE_INSERT_NOUNLOCK,
 	__BTREE_INSERT_NOFAIL,
+	__BTREE_INSERT_NOCHECK_RW,
 	__BTREE_INSERT_USE_RESERVE,
 	__BTREE_INSERT_USE_ALLOC_RESERVE,
 	__BTREE_INSERT_JOURNAL_REPLAY,
@@ -100,6 +101,8 @@ enum {
 /* Don't check for -ENOSPC: */
 #define BTREE_INSERT_NOFAIL		(1 << __BTREE_INSERT_NOFAIL)
 
+#define BTREE_INSERT_NOCHECK_RW		(1 << __BTREE_INSERT_NOCHECK_RW)
+
 /* for copygc, or when merging btree nodes */
 #define BTREE_INSERT_USE_RESERVE	(1 << __BTREE_INSERT_USE_RESERVE)
 #define BTREE_INSERT_USE_ALLOC_RESERVE	(1 << __BTREE_INSERT_USE_ALLOC_RESERVE)
diff --git a/libbcachefs/btree_update_leaf.c b/libbcachefs/btree_update_leaf.c
index 7eca9203..0df894fc 100644
--- a/libbcachefs/btree_update_leaf.c
+++ b/libbcachefs/btree_update_leaf.c
@@ -628,7 +628,8 @@ int __bch2_btree_insert_at(struct btree_insert *trans)
 	trans_for_each_entry(trans, i)
 		btree_insert_entry_checks(c, i);
 
-	if (unlikely(!percpu_ref_tryget(&c->writes)))
+	if (unlikely(!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
+		     !percpu_ref_tryget(&c->writes)))
 		return -EROFS;
 retry:
 	trans_for_each_iter(trans, i) {
@@ -658,7 +659,8 @@ retry:
 	trans_for_each_iter(trans, i)
 		bch2_btree_iter_downgrade(i->iter);
 out:
-	percpu_ref_put(&c->writes);
+	if (unlikely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
+		percpu_ref_put(&c->writes);
 
 	/* make sure we didn't drop or screw up locks: */
 	trans_for_each_iter(trans, i) {
diff --git a/libbcachefs/buckets.c b/libbcachefs/buckets.c
index d33d0bf0..ea71acb5 100644
--- a/libbcachefs/buckets.c
+++ b/libbcachefs/buckets.c
@@ -151,7 +151,6 @@ retry:
 	acc_u64s_percpu((u64 *) ret,
 			(u64 __percpu *) c->usage[0],
 			sizeof(*ret) / sizeof(u64) + nr);
-	percpu_up_read_preempt_enable(&c->mark_lock);
 
 	return ret;
 }
@@ -223,13 +222,14 @@ static bool bucket_became_unavailable(struct bucket_mark old,
 	       !is_available_bucket(new);
 }
 
-void bch2_fs_usage_apply(struct bch_fs *c,
-			 struct bch_fs_usage *fs_usage,
-			 struct disk_reservation *disk_res,
-			 struct gc_pos gc_pos)
+int bch2_fs_usage_apply(struct bch_fs *c,
+			struct bch_fs_usage *fs_usage,
+			struct disk_reservation *disk_res,
+			struct gc_pos gc_pos)
 {
 	s64 added = fs_usage->s.data + fs_usage->s.reserved;
 	s64 should_not_have_added;
+	int ret = 0;
 
 	percpu_rwsem_assert_held(&c->mark_lock);
 
@@ -242,6 +242,7 @@ void bch2_fs_usage_apply(struct bch_fs *c,
 		      "disk usage increased without a reservation")) {
 		atomic64_sub(should_not_have_added, &c->sectors_available);
 		added -= should_not_have_added;
+		ret = -1;
 	}
 
 	if (added > 0) {
@@ -259,6 +260,8 @@ void bch2_fs_usage_apply(struct bch_fs *c,
 			 (u64 *) fs_usage,
 			 sizeof(*fs_usage) / sizeof(u64) + c->replicas.nr);
 	}
+
+	return ret;
 }
 
 static inline void account_bucket(struct bch_fs_usage *fs_usage,
@@ -363,10 +366,7 @@ static inline void update_cached_sectors(struct bch_fs *c,
 {
 	struct bch_replicas_padded r;
 
-	r.e.data_type	= BCH_DATA_CACHED;
-	r.e.nr_devs	= 1;
-	r.e.nr_required	= 1;
-	r.e.devs[0]	= dev;
+	bch2_replicas_entry_cached(&r.e, dev);
 
 	update_replicas(c, fs_usage, &r.e, sectors);
 }
@@ -382,7 +382,8 @@ static void __bch2_invalidate_bucket(struct bch_fs *c, struct bch_dev *ca,
 	*old = bucket_data_cmpxchg(c, ca, fs_usage, g, new, ({
 		BUG_ON(!is_available_bucket(new));
 
-		new.owned_by_allocator	= 1;
+		new.owned_by_allocator	= true;
+		new.dirty		= true;
 		new.data_type		= 0;
 		new.cached_sectors	= 0;
 		new.dirty_sectors	= 0;
@@ -455,6 +456,7 @@ static void __bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 	       type != BCH_DATA_JOURNAL);
 
 	bucket_data_cmpxchg(c, ca, fs_usage, g, new, ({
+		new.dirty	= true;
 		new.data_type	= type;
 		checked_add(new.dirty_sectors, sectors);
 	}));
@@ -480,13 +482,14 @@ void bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 						    true);
 	} else {
 		struct bucket *g;
-		struct bucket_mark old, new;
+		struct bucket_mark new;
 
 		rcu_read_lock();
 
 		g = bucket(ca, b);
-		old = bucket_cmpxchg(g, new, ({
-			new.data_type = type;
+		bucket_cmpxchg(g, new, ({
+			new.dirty	= true;
+			new.data_type	= type;
 			checked_add(new.dirty_sectors, sectors);
 		}));
 
@@ -537,6 +540,8 @@ static void bch2_mark_pointer(struct bch_fs *c,
 	do {
 		new.v.counter = old.v.counter = v;
 
+		new.dirty = true;
+
 		/*
 		 * Check this after reading bucket mark to guard against
 		 * the allocator invalidating a bucket after we've already
@@ -591,9 +596,14 @@ static int bch2_mark_stripe_ptr(struct bch_fs *c,
 	int blocks_nonempty_delta;
 	s64 parity_sectors;
 
+	BUG_ON(!sectors);
+
 	m = genradix_ptr(&c->stripes[gc], p.idx);
 
+	spin_lock(&c->ec_stripes_heap_lock);
+
 	if (!m || !m->alive) {
+		spin_unlock(&c->ec_stripes_heap_lock);
 		bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
 				    (u64) p.idx);
 		return -1;
@@ -609,19 +619,21 @@ static int bch2_mark_stripe_ptr(struct bch_fs *c,
 		parity_sectors = -parity_sectors;
 	sectors += parity_sectors;
 
-	new = atomic_add_return(sectors, &m->block_sectors[p.block]);
-	old = new - sectors;
+	old = m->block_sectors[p.block];
+	m->block_sectors[p.block] += sectors;
+	new = m->block_sectors[p.block];
 
 	blocks_nonempty_delta = (int) !!new - (int) !!old;
-	if (!blocks_nonempty_delta)
-		return 0;
+	if (blocks_nonempty_delta) {
+		m->blocks_nonempty += blocks_nonempty_delta;
 
-	atomic_add(blocks_nonempty_delta, &m->blocks_nonempty);
+		if (!gc)
+			bch2_stripes_heap_update(c, m, p.idx);
+	}
 
-	BUG_ON(atomic_read(&m->blocks_nonempty) < 0);
+	m->dirty = true;
 
-	if (!gc)
-		bch2_stripes_heap_update(c, m, p.idx);
+	spin_unlock(&c->ec_stripes_heap_lock);
 
 	update_replicas(c, fs_usage, &m->r.e, sectors);
 
@@ -629,8 +641,7 @@ static int bch2_mark_stripe_ptr(struct bch_fs *c,
 }
 
 static int bch2_mark_extent(struct bch_fs *c, struct bkey_s_c k,
-			    s64 sectors,
-			    enum bch_data_type data_type,
+			    s64 sectors, enum bch_data_type data_type,
 			    struct bch_fs_usage *fs_usage,
 			    unsigned journal_seq, unsigned flags,
 			    bool gc)
@@ -701,14 +712,13 @@ static void bucket_set_stripe(struct bch_fs *c,
 		BUG_ON(ptr_stale(ca, ptr));
 
 		old = bucket_data_cmpxchg(c, ca, fs_usage, g, new, ({
+			new.dirty			= true;
 			new.stripe			= enabled;
 			if (journal_seq) {
 				new.journal_seq_valid	= 1;
 				new.journal_seq		= journal_seq;
 			}
 		}));
-
-		BUG_ON(old.stripe == enabled);
 	}
 }
 
@@ -723,22 +733,19 @@ static int bch2_mark_stripe(struct bch_fs *c, struct bkey_s_c k,
 	struct stripe *m = genradix_ptr(&c->stripes[gc], idx);
 	unsigned i;
 
+	spin_lock(&c->ec_stripes_heap_lock);
+
 	if (!m || (!inserting && !m->alive)) {
+		spin_unlock(&c->ec_stripes_heap_lock);
 		bch_err_ratelimited(c, "error marking nonexistent stripe %zu",
 				    idx);
 		return -1;
 	}
 
-	if (inserting && m->alive) {
-		bch_err_ratelimited(c, "error marking stripe %zu: already exists",
-				    idx);
-		return -1;
-	}
-
-	BUG_ON(atomic_read(&m->blocks_nonempty));
+	if (m->alive)
+		bch2_stripes_heap_del(c, m, idx);
 
-	for (i = 0; i < EC_STRIPE_MAX; i++)
-		BUG_ON(atomic_read(&m->block_sectors[i]));
+	memset(m, 0, sizeof(*m));
 
 	if (inserting) {
 		m->sectors	= le16_to_cpu(s.v->sectors);
@@ -754,7 +761,6 @@ static int bch2_mark_stripe(struct bch_fs *c, struct bkey_s_c k,
 
 		for (i = 0; i < s.v->nr_blocks; i++)
 			m->r.e.devs[i] = s.v->ptrs[i].dev;
-	}
 
 	/*
 	 * XXX: account for stripes somehow here
@@ -763,15 +769,23 @@ static int bch2_mark_stripe(struct bch_fs *c, struct bkey_s_c k,
 	update_replicas(c, fs_usage, &m->r.e, stripe_sectors);
 #endif
 
-	if (!gc) {
-		if (inserting)
+		/* gc recalculates these fields: */
+		if (!(flags & BCH_BUCKET_MARK_GC)) {
+			for (i = 0; i < s.v->nr_blocks; i++) {
+				m->block_sectors[i] =
+					stripe_blockcount_get(s.v, i);
+				m->blocks_nonempty += !!m->block_sectors[i];
+			}
+		}
+
+		if (!gc)
 			bch2_stripes_heap_insert(c, m, idx);
 		else
-			bch2_stripes_heap_del(c, m, idx);
-	} else {
-		m->alive = inserting;
+			m->alive = true;
 	}
 
+	spin_unlock(&c->ec_stripes_heap_lock);
+
 	bucket_set_stripe(c, s.v, inserting, fs_usage, 0, gc);
 	return 0;
 }
@@ -879,6 +893,8 @@ void bch2_mark_update(struct btree_insert *trans,
 	struct bch_fs_usage	*fs_usage;
 	struct gc_pos		pos = gc_pos_btree_node(b);
 	struct bkey_packed	*_k;
+	u64 disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
+	static int warned_disk_usage = 0;
 
 	if (!btree_node_type_needs_gc(iter->btree_id))
 		return;
@@ -939,7 +955,37 @@ void bch2_mark_update(struct btree_insert *trans,
 		bch2_btree_node_iter_advance(&node_iter, b);
 	}
 
-	bch2_fs_usage_apply(c, fs_usage, trans->disk_res, pos);
+	if (bch2_fs_usage_apply(c, fs_usage, trans->disk_res, pos) &&
+	    !warned_disk_usage &&
+	    !xchg(&warned_disk_usage, 1)) {
+		char buf[200];
+
+		pr_err("disk usage increased more than %llu sectors reserved", disk_res_sectors);
+
+		pr_err("while inserting");
+		bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(insert->k));
+		pr_err("%s", buf);
+		pr_err("overlapping with");
+
+		node_iter = iter->l[0].iter;
+		while ((_k = bch2_btree_node_iter_peek_filter(&node_iter, b,
+							      KEY_TYPE_discard))) {
+			struct bkey		unpacked;
+			struct bkey_s_c		k;
+
+			k = bkey_disassemble(b, _k, &unpacked);
+
+			if (btree_node_is_extents(b)
+			    ? bkey_cmp(insert->k->k.p, bkey_start_pos(k.k)) <= 0
+			    : bkey_cmp(insert->k->k.p, k.k->p))
+				break;
+
+			bch2_bkey_val_to_text(&PBUF(buf), c, k);
+			pr_err("%s", buf);
+
+			bch2_btree_node_iter_advance(&node_iter, b);
+		}
+	}
 
 	percpu_up_read_preempt_enable(&c->mark_lock);
 }
diff --git a/libbcachefs/buckets.h b/libbcachefs/buckets.h
index ebd39e85..6f368172 100644
--- a/libbcachefs/buckets.h
+++ b/libbcachefs/buckets.h
@@ -181,6 +181,8 @@ static inline bool bucket_needs_journal_commit(struct bucket_mark m,
 
 struct bch_dev_usage bch2_dev_usage_read(struct bch_fs *, struct bch_dev *);
 
+void bch2_dev_usage_from_buckets(struct bch_fs *, struct bch_dev *);
+
 static inline u64 __dev_buckets_available(struct bch_dev *ca,
 					  struct bch_dev_usage stats)
 {
@@ -264,8 +266,8 @@ int bch2_mark_key(struct bch_fs *, struct bkey_s_c,
 		  bool, s64, struct gc_pos,
 		  struct bch_fs_usage *, u64, unsigned);
 void bch2_mark_update(struct btree_insert *, struct btree_insert_entry *);
-void bch2_fs_usage_apply(struct bch_fs *, struct bch_fs_usage *,
-			 struct disk_reservation *, struct gc_pos);
+int bch2_fs_usage_apply(struct bch_fs *, struct bch_fs_usage *,
+			struct disk_reservation *, struct gc_pos);
 
 /* disk reservations: */
 
diff --git a/libbcachefs/chardev.c b/libbcachefs/chardev.c
index 56ceb260..b84ae5c9 100644
--- a/libbcachefs/chardev.c
+++ b/libbcachefs/chardev.c
@@ -402,6 +402,8 @@ static long bch2_ioctl_usage(struct bch_fs *c,
 		if (!src)
 			return -ENOMEM;
 
+		percpu_up_read_preempt_enable(&c->mark_lock);
+
 		dst.used		= bch2_fs_sectors_used(c, *src);
 		dst.online_reserved	= src->s.online_reserved;
 
diff --git a/libbcachefs/ec.c b/libbcachefs/ec.c
index 755a2603..8018c2bc 100644
--- a/libbcachefs/ec.c
+++ b/libbcachefs/ec.c
@@ -11,6 +11,7 @@
 #include "ec.h"
 #include "error.h"
 #include "io.h"
+#include "journal_io.h"
 #include "keylist.h"
 #include "super-io.h"
 #include "util.h"
@@ -98,40 +99,6 @@ struct ec_bio {
 
 /* Stripes btree keys: */
 
-static unsigned stripe_csums_per_device(const struct bch_stripe *s)
-{
-	return DIV_ROUND_UP(le16_to_cpu(s->sectors),
-			    1 << s->csum_granularity_bits);
-}
-
-static unsigned stripe_csum_offset(const struct bch_stripe *s,
-				   unsigned dev, unsigned csum_idx)
-{
-	unsigned csum_bytes = bch_crc_bytes[s->csum_type];
-
-	return sizeof(struct bch_stripe) +
-		sizeof(struct bch_extent_ptr) * s->nr_blocks +
-		(dev * stripe_csums_per_device(s) + csum_idx) * csum_bytes;
-}
-
-static unsigned stripe_blockcount_offset(const struct bch_stripe *s,
-					 unsigned idx)
-{
-	return stripe_csum_offset(s, s->nr_blocks, 0) +
-		sizeof(16) * idx;
-}
-
-static unsigned stripe_val_u64s(const struct bch_stripe *s)
-{
-	return DIV_ROUND_UP(stripe_blockcount_offset(s, s->nr_blocks),
-			    sizeof(u64));
-}
-
-static void *stripe_csum(struct bch_stripe *s, unsigned dev, unsigned csum_idx)
-{
-	return (void *) s + stripe_csum_offset(s, dev, csum_idx);
-}
-
 const char *bch2_stripe_invalid(const struct bch_fs *c, struct bkey_s_c k)
 {
 	const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
@@ -164,8 +131,9 @@ void bch2_stripe_to_text(struct printbuf *out, struct bch_fs *c,
 	       1U << s->csum_granularity_bits);
 
 	for (i = 0; i < s->nr_blocks; i++)
-		pr_buf(out, " %u:%llu", s->ptrs[i].dev,
-		       (u64) s->ptrs[i].offset);
+		pr_buf(out, " %u:%llu:%u", s->ptrs[i].dev,
+		       (u64) s->ptrs[i].offset,
+		       stripe_blockcount_get(s, i));
 }
 
 static int ptr_matches_stripe(struct bch_fs *c,
@@ -609,29 +577,15 @@ static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
 	BUG_ON(h->data[m->heap_idx].idx != idx);
 }
 
-static inline unsigned stripe_entry_blocks(struct stripe *m)
-{
-	return atomic_read(&m->blocks_nonempty);
-}
-
 void bch2_stripes_heap_update(struct bch_fs *c,
 			      struct stripe *m, size_t idx)
 {
 	ec_stripes_heap *h = &c->ec_stripes_heap;
-	bool queue_delete;
 	size_t i;
 
-	spin_lock(&c->ec_stripes_heap_lock);
-
-	if (!m->alive) {
-		spin_unlock(&c->ec_stripes_heap_lock);
-		return;
-	}
-
 	heap_verify_backpointer(c, idx);
 
-	h->data[m->heap_idx].blocks_nonempty =
-		stripe_entry_blocks(m);
+	h->data[m->heap_idx].blocks_nonempty = m->blocks_nonempty;
 
 	i = m->heap_idx;
 	heap_sift_up(h,	  i, ec_stripes_heap_cmp,
@@ -641,44 +595,35 @@ void bch2_stripes_heap_update(struct bch_fs *c,
 
 	heap_verify_backpointer(c, idx);
 
-	queue_delete = stripe_idx_to_delete(c) >= 0;
-	spin_unlock(&c->ec_stripes_heap_lock);
-
-	if (queue_delete)
+	if (stripe_idx_to_delete(c) >= 0)
 		schedule_work(&c->ec_stripe_delete_work);
 }
 
 void bch2_stripes_heap_del(struct bch_fs *c,
 			   struct stripe *m, size_t idx)
 {
-	spin_lock(&c->ec_stripes_heap_lock);
 	heap_verify_backpointer(c, idx);
 
 	m->alive = false;
 	heap_del(&c->ec_stripes_heap, m->heap_idx,
 		 ec_stripes_heap_cmp,
 		 ec_stripes_heap_set_backpointer);
-	spin_unlock(&c->ec_stripes_heap_lock);
 }
 
 void bch2_stripes_heap_insert(struct bch_fs *c,
 			      struct stripe *m, size_t idx)
 {
-	spin_lock(&c->ec_stripes_heap_lock);
-
 	BUG_ON(heap_full(&c->ec_stripes_heap));
 
 	heap_add(&c->ec_stripes_heap, ((struct ec_stripe_heap_entry) {
 			.idx = idx,
-			.blocks_nonempty = stripe_entry_blocks(m),
+			.blocks_nonempty = m->blocks_nonempty,
 		}),
 		 ec_stripes_heap_cmp,
 		 ec_stripes_heap_set_backpointer);
 	m->alive = true;
 
 	heap_verify_backpointer(c, idx);
-
-	spin_unlock(&c->ec_stripes_heap_lock);
 }
 
 /* stripe deletion */
@@ -1217,6 +1162,116 @@ unlock:
 	mutex_unlock(&c->ec_new_stripe_lock);
 }
 
+static int __bch2_stripe_write_key(struct bch_fs *c,
+				   struct btree_iter *iter,
+				   struct stripe *m,
+				   size_t idx,
+				   struct bkey_i_stripe *new_key,
+				   unsigned flags)
+{
+	struct bkey_s_c k;
+	unsigned i;
+	int ret;
+
+	bch2_btree_iter_set_pos(iter, POS(0, idx));
+
+	k = bch2_btree_iter_peek_slot(iter);
+	ret = btree_iter_err(k);
+	if (ret)
+		return ret;
+
+	if (k.k->type != KEY_TYPE_stripe)
+		return -EIO;
+
+	bkey_reassemble(&new_key->k_i, k);
+
+	spin_lock(&c->ec_stripes_heap_lock);
+
+	for (i = 0; i < new_key->v.nr_blocks; i++)
+		stripe_blockcount_set(&new_key->v, i,
+				      m->block_sectors[i]);
+	m->dirty = false;
+
+	spin_unlock(&c->ec_stripes_heap_lock);
+
+	return bch2_btree_insert_at(c, NULL, NULL,
+				   BTREE_INSERT_NOFAIL|flags,
+				   BTREE_INSERT_ENTRY(iter, &new_key->k_i));
+}
+
+int bch2_stripes_write(struct bch_fs *c, bool *wrote)
+{
+	struct btree_iter iter;
+	struct genradix_iter giter;
+	struct bkey_i_stripe *new_key;
+	struct stripe *m;
+	int ret = 0;
+
+	new_key = kmalloc(255 * sizeof(u64), GFP_KERNEL);
+	BUG_ON(!new_key);
+
+	bch2_btree_iter_init(&iter, c, BTREE_ID_EC, POS_MIN,
+			     BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
+
+	genradix_for_each(&c->stripes[0], giter, m) {
+		if (!m->dirty)
+			continue;
+
+		ret = __bch2_stripe_write_key(c, &iter, m, giter.pos,
+					new_key, BTREE_INSERT_NOCHECK_RW);
+		if (ret)
+			break;
+
+		*wrote = true;
+	}
+
+	bch2_btree_iter_unlock(&iter);
+
+	kfree(new_key);
+
+	return ret;
+}
+
+static void bch2_stripe_read_key(struct bch_fs *c, struct bkey_s_c k)
+{
+
+	struct gc_pos pos = { 0 };
+
+	bch2_mark_key(c, k, true, 0, pos, NULL, 0, 0);
+}
+
+int bch2_stripes_read(struct bch_fs *c, struct list_head *journal_replay_list)
+{
+	struct journal_replay *r;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret;
+
+	ret = bch2_fs_ec_start(c);
+	if (ret)
+		return ret;
+
+	for_each_btree_key(&iter, c, BTREE_ID_EC, POS_MIN, 0, k) {
+		bch2_stripe_read_key(c, k);
+		bch2_btree_iter_cond_resched(&iter);
+	}
+
+	ret = bch2_btree_iter_unlock(&iter);
+	if (ret)
+		return ret;
+
+	list_for_each_entry(r, journal_replay_list, list) {
+		struct bkey_i *k, *n;
+		struct jset_entry *entry;
+
+		for_each_jset_key(k, n, entry, &r->j)
+			if (entry->btree_id == BTREE_ID_EC)
+				bch2_stripe_read_key(c, bkey_i_to_s_c(k));
+	}
+
+	return 0;
+}
+
 int bch2_ec_mem_alloc(struct bch_fs *c, bool gc)
 {
 	struct btree_iter iter;
diff --git a/libbcachefs/ec.h b/libbcachefs/ec.h
index c728c52c..28178330 100644
--- a/libbcachefs/ec.h
+++ b/libbcachefs/ec.h
@@ -13,6 +13,55 @@ void bch2_stripe_to_text(struct printbuf *, struct bch_fs *,
 	.val_to_text	= bch2_stripe_to_text,		\
 }
 
+static inline unsigned stripe_csums_per_device(const struct bch_stripe *s)
+{
+	return DIV_ROUND_UP(le16_to_cpu(s->sectors),
+			    1 << s->csum_granularity_bits);
+}
+
+static inline unsigned stripe_csum_offset(const struct bch_stripe *s,
+					  unsigned dev, unsigned csum_idx)
+{
+	unsigned csum_bytes = bch_crc_bytes[s->csum_type];
+
+	return sizeof(struct bch_stripe) +
+		sizeof(struct bch_extent_ptr) * s->nr_blocks +
+		(dev * stripe_csums_per_device(s) + csum_idx) * csum_bytes;
+}
+
+static inline unsigned stripe_blockcount_offset(const struct bch_stripe *s,
+						unsigned idx)
+{
+	return stripe_csum_offset(s, s->nr_blocks, 0) +
+		sizeof(u16) * idx;
+}
+
+static inline unsigned stripe_blockcount_get(const struct bch_stripe *s,
+					     unsigned idx)
+{
+	return le16_to_cpup((void *) s + stripe_blockcount_offset(s, idx));
+}
+
+static inline void stripe_blockcount_set(struct bch_stripe *s,
+					 unsigned idx, unsigned v)
+{
+	__le16 *p = (void *) s + stripe_blockcount_offset(s, idx);
+
+	*p = cpu_to_le16(v);
+}
+
+static inline unsigned stripe_val_u64s(const struct bch_stripe *s)
+{
+	return DIV_ROUND_UP(stripe_blockcount_offset(s, s->nr_blocks),
+			    sizeof(u64));
+}
+
+static inline void *stripe_csum(struct bch_stripe *s,
+				unsigned dev, unsigned csum_idx)
+{
+	return (void *) s + stripe_csum_offset(s, dev, csum_idx);
+}
+
 struct bch_read_bio;
 
 struct ec_stripe_buf {
@@ -100,6 +149,9 @@ void bch2_ec_stop_dev(struct bch_fs *, struct bch_dev *);
 
 void bch2_ec_flush_new_stripes(struct bch_fs *);
 
+int bch2_stripes_read(struct bch_fs *, struct list_head *);
+int bch2_stripes_write(struct bch_fs *, bool *);
+
 int bch2_ec_mem_alloc(struct bch_fs *, bool);
 
 int bch2_fs_ec_start(struct bch_fs *);
diff --git a/libbcachefs/ec_types.h b/libbcachefs/ec_types.h
index 44c5d382..b4d37705 100644
--- a/libbcachefs/ec_types.h
+++ b/libbcachefs/ec_types.h
@@ -19,9 +19,10 @@ struct stripe {
 	u8			nr_blocks;
 	u8			nr_redundant;
 
-	u8			alive;
-	atomic_t		blocks_nonempty;
-	atomic_t		block_sectors[EC_STRIPE_MAX];
+	unsigned		alive:1;
+	unsigned		dirty:1;
+	u8			blocks_nonempty;
+	u16			block_sectors[EC_STRIPE_MAX];
 
 	struct bch_replicas_padded r;
 };
diff --git a/libbcachefs/extents.c b/libbcachefs/extents.c
index 29804168..0f075fa1 100644
--- a/libbcachefs/extents.c
+++ b/libbcachefs/extents.c
@@ -1664,12 +1664,13 @@ static bool bch2_extent_merge_inline(struct bch_fs *c,
 	return ret == BCH_MERGE_MERGE;
 }
 
-int bch2_check_range_allocated(struct bch_fs *c, struct bpos pos, u64 size)
+bool bch2_check_range_allocated(struct bch_fs *c, struct bpos pos, u64 size,
+			       unsigned nr_replicas)
 {
 	struct btree_iter iter;
 	struct bpos end = pos;
 	struct bkey_s_c k;
-	int ret = 0;
+	bool ret = true;
 
 	end.offset += size;
 
@@ -1678,8 +1679,8 @@ int bch2_check_range_allocated(struct bch_fs *c, struct bpos pos, u64 size)
 		if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
 			break;
 
-		if (!bch2_extent_is_fully_allocated(k)) {
-			ret = -ENOSPC;
+		if (nr_replicas > bch2_bkey_nr_ptrs_allocated(k)) {
+			ret = false;
 			break;
 		}
 	}
@@ -1688,6 +1689,29 @@ int bch2_check_range_allocated(struct bch_fs *c, struct bpos pos, u64 size)
 	return ret;
 }
 
+unsigned bch2_bkey_nr_ptrs_allocated(struct bkey_s_c k)
+{
+	unsigned ret = 0;
+
+	switch (k.k->type) {
+	case KEY_TYPE_extent: {
+		struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
+		const union bch_extent_entry *entry;
+		struct extent_ptr_decoded p;
+
+		extent_for_each_ptr_decode(e, p, entry)
+			ret += !p.ptr.cached &&
+				p.crc.compression_type == BCH_COMPRESSION_NONE;
+		break;
+	}
+	case KEY_TYPE_reservation:
+		ret = bkey_s_c_to_reservation(k).v->nr_replicas;
+		break;
+	}
+
+	return ret;
+}
+
 /* KEY_TYPE_reservation: */
 
 const char *bch2_reservation_invalid(const struct bch_fs *c, struct bkey_s_c k)
diff --git a/libbcachefs/extents.h b/libbcachefs/extents.h
index 0e6f4a0b..698b2581 100644
--- a/libbcachefs/extents.h
+++ b/libbcachefs/extents.h
@@ -571,6 +571,7 @@ static inline void extent_save(struct btree *b, struct bkey_packed *dst,
 		BUG_ON(!bch2_bkey_pack_key(dst, src, f));
 }
 
-int bch2_check_range_allocated(struct bch_fs *, struct bpos, u64);
+bool bch2_check_range_allocated(struct bch_fs *, struct bpos, u64, unsigned);
+unsigned bch2_bkey_nr_ptrs_allocated(struct bkey_s_c);
 
 #endif /* _BCACHEFS_EXTENTS_H */
diff --git a/libbcachefs/eytzinger.h b/libbcachefs/eytzinger.h
index 66fa227c..d19d809c 100644
--- a/libbcachefs/eytzinger.h
+++ b/libbcachefs/eytzinger.h
@@ -262,18 +262,20 @@ static inline ssize_t eytzinger0_find_le(void *base, size_t nr, size_t size,
 	}
 }
 
-static inline size_t eytzinger0_find(void *base, size_t nr, size_t size,
-				     eytzinger_cmp_fn cmp, const void *search)
-{
-	size_t i = 0;
-	int res;
-
-	while (i < nr &&
-	       (res = cmp(search, base + i * size, size)))
-		i = eytzinger0_child(i, res > 0);
-
-	return i;
-}
+#define eytzinger0_find(base, nr, size, _cmp, search)			\
+({									\
+	void *_base	= (base);					\
+	void *_search	= (search);					\
+	size_t _nr	= (nr);						\
+	size_t _size	= (size);					\
+	size_t _i	= 0;						\
+	int _res;							\
+									\
+	while (_i < _nr &&						\
+	       (_res = _cmp(_search, _base + _i * _size, _size)))	\
+		_i = eytzinger0_child(_i, _res > 0);			\
+	_i;								\
+})
 
 void eytzinger0_sort(void *, size_t, size_t,
 		    int (*cmp_func)(const void *, const void *, size_t),
diff --git a/libbcachefs/fs-io.c b/libbcachefs/fs-io.c
index c1739f53..2cfc2d9e 100644
--- a/libbcachefs/fs-io.c
+++ b/libbcachefs/fs-io.c
@@ -253,7 +253,9 @@ static s64 sum_sector_overwrites(struct bkey_i *new, struct btree_iter *_iter,
 		BUG_ON(btree_iter_err(old));
 
 		if (allocating &&
-		    !bch2_extent_is_fully_allocated(old))
+		    !*allocating &&
+		    bch2_bkey_nr_ptrs_allocated(old) <
+		    bch2_bkey_nr_dirty_ptrs(bkey_i_to_s_c(new)))
 			*allocating = true;
 
 		delta += (min(new->k.p.offset,
@@ -858,9 +860,7 @@ static void bch2_add_page_sectors(struct bio *bio, struct bkey_s_c k)
 {
 	struct bvec_iter iter;
 	struct bio_vec bv;
-	unsigned nr_ptrs = !bch2_extent_is_compressed(k)
-		? bch2_bkey_nr_dirty_ptrs(k)
-		: 0;
+	unsigned nr_ptrs = bch2_bkey_nr_ptrs_allocated(k);
 
 	bio_for_each_segment(bv, bio, iter) {
 		/* brand new pages, don't need to be locked: */
@@ -1759,6 +1759,7 @@ static long bch2_dio_write_loop(struct dio_write *dio)
 	struct bch_inode_info *inode = dio->iop.inode;
 	struct bio *bio = &dio->iop.op.wbio.bio;
 	struct bio_vec *bv;
+	loff_t offset;
 	bool sync;
 	long ret;
 	int i;
@@ -1770,12 +1771,16 @@ static long bch2_dio_write_loop(struct dio_write *dio)
 	__pagecache_block_get(&mapping->add_lock);
 
 	/* Write and invalidate pagecache range that we're writing to: */
-	ret = write_invalidate_inode_pages_range(mapping, req->ki_pos,
-				req->ki_pos + iov_iter_count(&dio->iter) - 1);
+	offset = req->ki_pos + (dio->iop.op.written << 9);
+	ret = write_invalidate_inode_pages_range(mapping,
+					offset,
+					offset + iov_iter_count(&dio->iter) - 1);
 	if (unlikely(ret))
 		goto err;
 
 	while (1) {
+		offset = req->ki_pos + (dio->iop.op.written << 9);
+
 		BUG_ON(current->pagecache_lock);
 		current->pagecache_lock = &mapping->add_lock;
 		if (kthread)
@@ -1792,13 +1797,12 @@ static long bch2_dio_write_loop(struct dio_write *dio)
 
 		/* gup might have faulted pages back in: */
 		ret = write_invalidate_inode_pages_range(mapping,
-				req->ki_pos + (dio->iop.op.written << 9),
-				req->ki_pos + iov_iter_count(&dio->iter) - 1);
+				offset,
+				offset + bio->bi_iter.bi_size - 1);
 		if (unlikely(ret))
 			goto err;
 
-		dio->iop.op.pos = POS(inode->v.i_ino,
-				(req->ki_pos >> 9) + dio->iop.op.written);
+		dio->iop.op.pos = POS(inode->v.i_ino, offset >> 9);
 
 		task_io_account_write(bio->bi_iter.bi_size);
 
@@ -1878,7 +1882,6 @@ static int bch2_direct_IO_write(struct kiocb *req,
 	struct bch_fs *c = inode->v.i_sb->s_fs_info;
 	struct dio_write *dio;
 	struct bio *bio;
-	loff_t offset = req->ki_pos;
 	ssize_t ret;
 
 	lockdep_assert_held(&inode->v.i_rwsem);
@@ -1886,7 +1889,7 @@ static int bch2_direct_IO_write(struct kiocb *req,
 	if (unlikely(!iter->count))
 		return 0;
 
-	if (unlikely((offset|iter->count) & (block_bytes(c) - 1)))
+	if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
 		return -EINVAL;
 
 	bio = bio_alloc_bioset(GFP_KERNEL,
@@ -1898,7 +1901,7 @@ static int bch2_direct_IO_write(struct kiocb *req,
 	dio->mm			= current->mm;
 	dio->loop		= false;
 	dio->sync		= is_sync_kiocb(req) ||
-		offset + iter->count > inode->v.i_size;
+		req->ki_pos + iter->count > inode->v.i_size;
 	dio->free_iov		= false;
 	dio->quota_res.sectors	= 0;
 	dio->iter		= *iter;
@@ -1915,19 +1918,20 @@ static int bch2_direct_IO_write(struct kiocb *req,
 	if (unlikely(ret))
 		goto err;
 
+	dio->iop.op.nr_replicas	= dio->iop.op.opts.data_replicas;
+
 	ret = bch2_disk_reservation_get(c, &dio->iop.op.res, iter->count >> 9,
 					dio->iop.op.opts.data_replicas, 0);
 	if (unlikely(ret)) {
-		if (bch2_check_range_allocated(c, POS(inode->v.i_ino,
-						      offset >> 9),
-					       iter->count >> 9))
+		if (!bch2_check_range_allocated(c, POS(inode->v.i_ino,
+						       req->ki_pos >> 9),
+						iter->count >> 9,
+						dio->iop.op.opts.data_replicas))
 			goto err;
 
 		dio->iop.unalloc = true;
 	}
 
-	dio->iop.op.nr_replicas	= dio->iop.op.res.nr_replicas;
-
 	return bch2_dio_write_loop(dio);
 err:
 	bch2_disk_reservation_put(c, &dio->iop.op.res);
diff --git a/libbcachefs/journal_io.c b/libbcachefs/journal_io.c
index 52498627..5cc0651c 100644
--- a/libbcachefs/journal_io.c
+++ b/libbcachefs/journal_io.c
@@ -694,6 +694,11 @@ int bch2_journal_read(struct bch_fs *c, struct list_head *list)
 	}
 
 	list_for_each_entry(i, list, list) {
+		struct bch_replicas_padded replicas;
+		char buf[80];
+
+		bch2_devlist_to_replicas(&replicas.e, BCH_DATA_JOURNAL, i->devs);
+
 		ret = jset_validate_entries(c, &i->j, READ);
 		if (ret)
 			goto fsck_err;
@@ -705,11 +710,11 @@ int bch2_journal_read(struct bch_fs *c, struct list_head *list)
 
 		if (!degraded &&
 		    (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
-		     fsck_err_on(!bch2_replicas_marked(c, BCH_DATA_JOURNAL,
-						       i->devs, false), c,
-				 "superblock not marked as containing replicas (type %u)",
-				 BCH_DATA_JOURNAL))) {
-			ret = bch2_mark_replicas(c, BCH_DATA_JOURNAL, i->devs);
+		     fsck_err_on(!bch2_replicas_marked(c, &replicas.e, false), c,
+				 "superblock not marked as containing replicas %s",
+				 (bch2_replicas_entry_to_text(&PBUF(buf),
+							      &replicas.e), buf)))) {
+			ret = bch2_mark_replicas(c, &replicas.e);
 			if (ret)
 				return ret;
 		}
@@ -1108,6 +1113,7 @@ static void journal_write_done(struct closure *cl)
 	struct journal_buf *w = journal_prev_buf(j);
 	struct bch_devs_list devs =
 		bch2_bkey_devs(bkey_i_to_s_c(&w->key));
+	struct bch_replicas_padded replicas;
 	u64 seq = le64_to_cpu(w->data->seq);
 	u64 last_seq = le64_to_cpu(w->data->last_seq);
 
@@ -1118,7 +1124,9 @@ static void journal_write_done(struct closure *cl)
 		goto err;
 	}
 
-	if (bch2_mark_replicas(c, BCH_DATA_JOURNAL, devs))
+	bch2_devlist_to_replicas(&replicas.e, BCH_DATA_JOURNAL, devs);
+
+	if (bch2_mark_replicas(c, &replicas.e))
 		goto err;
 
 	spin_lock(&j->lock);
diff --git a/libbcachefs/journal_reclaim.c b/libbcachefs/journal_reclaim.c
index 4a997366..a795e888 100644
--- a/libbcachefs/journal_reclaim.c
+++ b/libbcachefs/journal_reclaim.c
@@ -335,7 +335,7 @@ void bch2_journal_reclaim_work(struct work_struct *work)
 		mutex_unlock(&j->reclaim_lock);
 
 	if (!test_bit(BCH_FS_RO, &c->flags))
-		queue_delayed_work(system_freezable_wq, &j->reclaim_work,
+		queue_delayed_work(c->journal_reclaim_wq, &j->reclaim_work,
 				   msecs_to_jiffies(j->reclaim_delay_ms));
 }
 
@@ -387,7 +387,6 @@ int bch2_journal_flush_device_pins(struct journal *j, int dev_idx)
 {
 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
 	struct journal_entry_pin_list *p;
-	struct bch_devs_list devs;
 	u64 iter, seq = 0;
 	int ret = 0;
 
@@ -412,12 +411,15 @@ int bch2_journal_flush_device_pins(struct journal *j, int dev_idx)
 
 	spin_lock(&j->lock);
 	while (!ret && seq < j->pin.back) {
+		struct bch_replicas_padded replicas;
+
 		seq = max(seq, journal_last_seq(j));
-		devs = journal_seq_pin(j, seq)->devs;
+		bch2_devlist_to_replicas(&replicas.e, BCH_DATA_JOURNAL,
+					 journal_seq_pin(j, seq)->devs);
 		seq++;
 
 		spin_unlock(&j->lock);
-		ret = bch2_mark_replicas(c, BCH_DATA_JOURNAL, devs);
+		ret = bch2_mark_replicas(c, &replicas.e);
 		spin_lock(&j->lock);
 	}
 	spin_unlock(&j->lock);
diff --git a/libbcachefs/migrate.c b/libbcachefs/migrate.c
index b2198651..bb425d88 100644
--- a/libbcachefs/migrate.c
+++ b/libbcachefs/migrate.c
@@ -4,6 +4,7 @@
 
 #include "bcachefs.h"
 #include "btree_update.h"
+#include "btree_update_interior.h"
 #include "buckets.h"
 #include "extents.h"
 #include "io.h"
@@ -152,6 +153,16 @@ retry:
 		bch2_btree_iter_unlock(&iter);
 	}
 
+	/* flush relevant btree updates */
+	while (1) {
+		closure_wait_event(&c->btree_interior_update_wait,
+				   !bch2_btree_interior_updates_nr_pending(c) ||
+				   c->btree_roots_dirty);
+		if (!bch2_btree_interior_updates_nr_pending(c))
+			break;
+		bch2_journal_meta(&c->journal);
+	}
+
 	ret = 0;
 out:
 	ret = bch2_replicas_gc_end(c, ret);
diff --git a/libbcachefs/move.c b/libbcachefs/move.c
index 80909ae4..98cfcefd 100644
--- a/libbcachefs/move.c
+++ b/libbcachefs/move.c
@@ -3,6 +3,7 @@
 #include "alloc_foreground.h"
 #include "btree_gc.h"
 #include "btree_update.h"
+#include "btree_update_interior.h"
 #include "buckets.h"
 #include "disk_groups.h"
 #include "inode.h"
@@ -763,6 +764,16 @@ int bch2_data_job(struct bch_fs *c,
 		ret = bch2_journal_flush_device_pins(&c->journal, -1);
 
 		ret = bch2_move_btree(c, rereplicate_pred, c, stats) ?: ret;
+
+		while (1) {
+			closure_wait_event(&c->btree_interior_update_wait,
+					   !bch2_btree_interior_updates_nr_pending(c) ||
+					   c->btree_roots_dirty);
+			if (!bch2_btree_interior_updates_nr_pending(c))
+				break;
+			bch2_journal_meta(&c->journal);
+		}
+
 		ret = bch2_gc_btree_replicas(c) ?: ret;
 
 		ret = bch2_move_data(c, NULL,
diff --git a/libbcachefs/recovery.c b/libbcachefs/recovery.c
index eae38ea7..f5f3f94e 100644
--- a/libbcachefs/recovery.c
+++ b/libbcachefs/recovery.c
@@ -214,12 +214,12 @@ int bch2_fs_recovery(struct bch_fs *c)
 	if (ret)
 		goto err;
 
-	set_bit(BCH_FS_ALLOC_READ_DONE, &c->flags);
-
-	err = "cannot allocate memory";
-	ret = bch2_fs_ec_start(c);
+	ret = bch2_stripes_read(c, &journal);
 	if (ret)
 		goto err;
+	pr_info("stripes_read done");
+
+	set_bit(BCH_FS_ALLOC_READ_DONE, &c->flags);
 
 	bch_verbose(c, "starting mark and sweep:");
 	err = "error in recovery";
diff --git a/libbcachefs/replicas.c b/libbcachefs/replicas.c
index 66ca13aa..230f807b 100644
--- a/libbcachefs/replicas.c
+++ b/libbcachefs/replicas.c
@@ -13,6 +13,16 @@ static inline int u8_cmp(u8 l, u8 r)
 	return (l > r) - (l < r);
 }
 
+static void verify_replicas_entry_sorted(struct bch_replicas_entry *e)
+{
+#ifdef CONFIG_BCACHES_DEBUG
+	unsigned i;
+
+	for (i = 0; i + 1 < e->nr_devs; i++)
+		BUG_ON(e->devs[i] >= e->devs[i + 1]);
+#endif
+}
+
 static void replicas_entry_sort(struct bch_replicas_entry *e)
 {
 	bubble_sort(e->devs, e->nr_devs, u8_cmp);
@@ -23,19 +33,13 @@ static void replicas_entry_sort(struct bch_replicas_entry *e)
 	     (void *) (_i) < (void *) (_r)->entries + (_r)->nr * (_r)->entry_size;\
 	     _i = (void *) (_i) + (_r)->entry_size)
 
-static inline struct bch_replicas_entry *
-cpu_replicas_entry(struct bch_replicas_cpu *r, unsigned i)
-{
-	return (void *) r->entries + r->entry_size * i;
-}
-
 static void bch2_cpu_replicas_sort(struct bch_replicas_cpu *r)
 {
 	eytzinger0_sort(r->entries, r->nr, r->entry_size, memcmp, NULL);
 }
 
-static void replicas_entry_to_text(struct printbuf *out,
-				  struct bch_replicas_entry *e)
+void bch2_replicas_entry_to_text(struct printbuf *out,
+				 struct bch_replicas_entry *e)
 {
 	unsigned i;
 
@@ -60,7 +64,7 @@ void bch2_cpu_replicas_to_text(struct printbuf *out,
 			pr_buf(out, " ");
 		first = false;
 
-		replicas_entry_to_text(out, e);
+		bch2_replicas_entry_to_text(out, e);
 	}
 }
 
@@ -100,8 +104,8 @@ static void stripe_to_replicas(struct bkey_s_c k,
 		r->devs[r->nr_devs++] = ptr->dev;
 }
 
-static void bkey_to_replicas(struct bkey_s_c k,
-			     struct bch_replicas_entry *e)
+static void bkey_to_replicas(struct bch_replicas_entry *e,
+			     struct bkey_s_c k)
 {
 	e->nr_devs = 0;
 
@@ -119,11 +123,13 @@ static void bkey_to_replicas(struct bkey_s_c k,
 		stripe_to_replicas(k, e);
 		break;
 	}
+
+	replicas_entry_sort(e);
 }
 
-static inline void devlist_to_replicas(struct bch_devs_list devs,
-				       enum bch_data_type data_type,
-				       struct bch_replicas_entry *e)
+void bch2_devlist_to_replicas(struct bch_replicas_entry *e,
+			      enum bch_data_type data_type,
+			      struct bch_devs_list devs)
 {
 	unsigned i;
 
@@ -137,6 +143,8 @@ static inline void devlist_to_replicas(struct bch_devs_list devs,
 
 	for (i = 0; i < devs.nr; i++)
 		e->devs[e->nr_devs++] = devs.devs[i];
+
+	replicas_entry_sort(e);
 }
 
 static struct bch_replicas_cpu
@@ -150,6 +158,9 @@ cpu_replicas_add_entry(struct bch_replicas_cpu *old,
 					replicas_entry_bytes(new_entry)),
 	};
 
+	BUG_ON(!new_entry->data_type);
+	verify_replicas_entry_sorted(new_entry);
+
 	new.entries = kcalloc(new.nr, new.entry_size, GFP_NOIO);
 	if (!new.entries)
 		return new;
@@ -175,13 +186,12 @@ static inline int __replicas_entry_idx(struct bch_replicas_cpu *r,
 	if (unlikely(entry_size > r->entry_size))
 		return -1;
 
-	replicas_entry_sort(search);
-
-	while (entry_size < r->entry_size)
-		((char *) search)[entry_size++] = 0;
+	verify_replicas_entry_sorted(search);
 
+#define entry_cmp(_l, _r, size)	memcmp(_l, _r, entry_size)
 	idx = eytzinger0_find(r->entries, r->nr, r->entry_size,
-			      memcmp, search);
+			      entry_cmp, search);
+#undef entry_cmp
 
 	return idx < r->nr ? idx : -1;
 }
@@ -189,6 +199,8 @@ static inline int __replicas_entry_idx(struct bch_replicas_cpu *r,
 int bch2_replicas_entry_idx(struct bch_fs *c,
 			    struct bch_replicas_entry *search)
 {
+	replicas_entry_sort(search);
+
 	return __replicas_entry_idx(&c->replicas, search);
 }
 
@@ -198,12 +210,17 @@ static bool __replicas_has_entry(struct bch_replicas_cpu *r,
 	return __replicas_entry_idx(r, search) >= 0;
 }
 
-static bool replicas_has_entry(struct bch_fs *c,
-			       struct bch_replicas_entry *search,
-			       bool check_gc_replicas)
+bool bch2_replicas_marked(struct bch_fs *c,
+			  struct bch_replicas_entry *search,
+			  bool check_gc_replicas)
 {
 	bool marked;
 
+	if (!search->nr_devs)
+		return true;
+
+	verify_replicas_entry_sorted(search);
+
 	percpu_down_read_preempt_disable(&c->mark_lock);
 	marked = __replicas_has_entry(&c->replicas, search) &&
 		(!check_gc_replicas ||
@@ -214,35 +231,31 @@ static bool replicas_has_entry(struct bch_fs *c,
 	return marked;
 }
 
-static void __replicas_table_update(struct bch_fs_usage __percpu *dst,
+static void __replicas_table_update(struct bch_fs_usage __percpu *dst_p,
 				    struct bch_replicas_cpu *dst_r,
-				    struct bch_fs_usage __percpu *src,
+				    struct bch_fs_usage __percpu *src_p,
 				    struct bch_replicas_cpu *src_r)
 {
-	int src_idx, dst_idx, cpu;
+	unsigned src_nr = sizeof(struct bch_fs_usage) / sizeof(u64) + src_r->nr;
+	struct bch_fs_usage *dst, *src = (void *)
+		bch2_acc_percpu_u64s((void *) src_p, src_nr);
+	int src_idx, dst_idx;
 
-	for (src_idx = 0; src_idx < src_r->nr; src_idx++) {
-		u64 *dst_v, src_v = 0;
+	preempt_disable();
+	dst = this_cpu_ptr(dst_p);
+	preempt_enable();
 
-		for_each_possible_cpu(cpu)
-			src_v += *per_cpu_ptr(&src->data[src_idx], cpu);
+	*dst = *src;
 
-		dst_idx = __replicas_entry_idx(dst_r,
-				cpu_replicas_entry(src_r, src_idx));
-
-		if (dst_idx < 0) {
-			BUG_ON(src_v);
+	for (src_idx = 0; src_idx < src_r->nr; src_idx++) {
+		if (!src->data[src_idx])
 			continue;
-		}
-
-		preempt_disable();
 
-		dst_v = this_cpu_ptr(&dst->data[dst_idx]);
-		BUG_ON(*dst_v);
-
-		*dst_v = src_v;
+		dst_idx = __replicas_entry_idx(dst_r,
+				cpu_replicas_entry(src_r, src_idx));
+		BUG_ON(dst_idx < 0);
 
-		preempt_enable();
+		dst->data[dst_idx] = src->data[src_idx];
 	}
 }
 
@@ -344,30 +357,32 @@ err:
 	return ret;
 }
 
-static int __bch2_mark_replicas(struct bch_fs *c,
-				struct bch_replicas_entry *devs)
+int bch2_mark_replicas(struct bch_fs *c,
+		       struct bch_replicas_entry *r)
 {
-	return likely(replicas_has_entry(c, devs, true))
+	return likely(bch2_replicas_marked(c, r, true))
 		? 0
-		: bch2_mark_replicas_slowpath(c, devs);
+		: bch2_mark_replicas_slowpath(c, r);
 }
 
-int bch2_mark_replicas(struct bch_fs *c,
-		       enum bch_data_type data_type,
-		       struct bch_devs_list devs)
+bool bch2_bkey_replicas_marked(struct bch_fs *c,
+			       struct bkey_s_c k,
+			       bool check_gc_replicas)
 {
 	struct bch_replicas_padded search;
+	struct bch_devs_list cached = bch2_bkey_cached_devs(k);
+	unsigned i;
 
-	if (!devs.nr)
-		return 0;
-
-	memset(&search, 0, sizeof(search));
+	for (i = 0; i < cached.nr; i++) {
+		bch2_replicas_entry_cached(&search.e, cached.devs[i]);
 
-	BUG_ON(devs.nr >= BCH_REPLICAS_MAX);
+		if (!bch2_replicas_marked(c, &search.e, check_gc_replicas))
+			return false;
+	}
 
-	devlist_to_replicas(devs, data_type, &search.e);
+	bkey_to_replicas(&search.e, k);
 
-	return __bch2_mark_replicas(c, &search.e);
+	return bch2_replicas_marked(c, &search.e, check_gc_replicas);
 }
 
 int bch2_mark_bkey_replicas(struct bch_fs *c, struct bkey_s_c k)
@@ -377,18 +392,17 @@ int bch2_mark_bkey_replicas(struct bch_fs *c, struct bkey_s_c k)
 	unsigned i;
 	int ret;
 
-	memset(&search, 0, sizeof(search));
+	for (i = 0; i < cached.nr; i++) {
+		bch2_replicas_entry_cached(&search.e, cached.devs[i]);
 
-	for (i = 0; i < cached.nr; i++)
-		if ((ret = bch2_mark_replicas(c, BCH_DATA_CACHED,
-					      bch2_dev_list_single(cached.devs[i]))))
+		ret = bch2_mark_replicas(c, &search.e);
+		if (ret)
 			return ret;
+	}
 
-	bkey_to_replicas(k, &search.e);
+	bkey_to_replicas(&search.e, k);
 
-	return search.e.nr_devs
-		? __bch2_mark_replicas(c, &search.e)
-		: 0;
+	return bch2_mark_replicas(c, &search.e);
 }
 
 int bch2_replicas_gc_end(struct bch_fs *c, int ret)
@@ -749,7 +763,7 @@ static void bch2_sb_replicas_to_text(struct printbuf *out,
 			pr_buf(out, " ");
 		first = false;
 
-		replicas_entry_to_text(out, e);
+		bch2_replicas_entry_to_text(out, e);
 	}
 }
 
@@ -798,46 +812,6 @@ const struct bch_sb_field_ops bch_sb_field_ops_replicas_v0 = {
 
 /* Query replicas: */
 
-bool bch2_replicas_marked(struct bch_fs *c,
-			  enum bch_data_type data_type,
-			  struct bch_devs_list devs,
-			  bool check_gc_replicas)
-{
-	struct bch_replicas_padded search;
-
-	if (!devs.nr)
-		return true;
-
-	memset(&search, 0, sizeof(search));
-
-	devlist_to_replicas(devs, data_type, &search.e);
-
-	return replicas_has_entry(c, &search.e, check_gc_replicas);
-}
-
-bool bch2_bkey_replicas_marked(struct bch_fs *c,
-			       struct bkey_s_c k,
-			       bool check_gc_replicas)
-{
-	struct bch_replicas_padded search;
-	struct bch_devs_list cached = bch2_bkey_cached_devs(k);
-	unsigned i;
-
-	memset(&search, 0, sizeof(search));
-
-	for (i = 0; i < cached.nr; i++)
-		if (!bch2_replicas_marked(c, BCH_DATA_CACHED,
-					  bch2_dev_list_single(cached.devs[i]),
-					  check_gc_replicas))
-			return false;
-
-	bkey_to_replicas(k, &search.e);
-
-	return search.e.nr_devs
-		? replicas_has_entry(c, &search.e, check_gc_replicas)
-		: true;
-}
-
 struct replicas_status __bch2_replicas_status(struct bch_fs *c,
 					      struct bch_devs_mask online_devs)
 {
diff --git a/libbcachefs/replicas.h b/libbcachefs/replicas.h
index fc833653..0ac2b8e0 100644
--- a/libbcachefs/replicas.h
+++ b/libbcachefs/replicas.h
@@ -4,17 +4,39 @@
 #include "eytzinger.h"
 #include "replicas_types.h"
 
+void bch2_replicas_entry_to_text(struct printbuf *,
+				 struct bch_replicas_entry *);
+void bch2_cpu_replicas_to_text(struct printbuf *, struct bch_replicas_cpu *);
+
+static inline struct bch_replicas_entry *
+cpu_replicas_entry(struct bch_replicas_cpu *r, unsigned i)
+{
+	return (void *) r->entries + r->entry_size * i;
+}
+
 int bch2_replicas_entry_idx(struct bch_fs *,
 			    struct bch_replicas_entry *);
-bool bch2_replicas_marked(struct bch_fs *, enum bch_data_type,
-			  struct bch_devs_list, bool);
+
+void bch2_devlist_to_replicas(struct bch_replicas_entry *,
+			      enum bch_data_type,
+			      struct bch_devs_list);
+bool bch2_replicas_marked(struct bch_fs *,
+			  struct bch_replicas_entry *, bool);
+int bch2_mark_replicas(struct bch_fs *,
+		       struct bch_replicas_entry *);
+
 bool bch2_bkey_replicas_marked(struct bch_fs *,
 			       struct bkey_s_c, bool);
-int bch2_mark_replicas(struct bch_fs *, enum bch_data_type,
-		       struct bch_devs_list);
 int bch2_mark_bkey_replicas(struct bch_fs *, struct bkey_s_c);
 
-void bch2_cpu_replicas_to_text(struct printbuf *, struct bch_replicas_cpu *);
+static inline void bch2_replicas_entry_cached(struct bch_replicas_entry *e,
+					      unsigned dev)
+{
+	e->data_type	= BCH_DATA_CACHED;
+	e->nr_devs	= 1;
+	e->nr_required	= 1;
+	e->devs[0]	= dev;
+}
 
 struct replicas_status {
 	struct {
diff --git a/libbcachefs/super.c b/libbcachefs/super.c
index a539f2a8..1835b535 100644
--- a/libbcachefs/super.c
+++ b/libbcachefs/super.c
@@ -205,7 +205,9 @@ int bch2_congested(void *data, int bdi_bits)
 static void __bch2_fs_read_only(struct bch_fs *c)
 {
 	struct bch_dev *ca;
+	bool wrote;
 	unsigned i;
+	int ret;
 
 	bch2_rebalance_stop(c);
 
@@ -220,23 +222,42 @@ static void __bch2_fs_read_only(struct bch_fs *c)
 	 */
 	bch2_journal_flush_all_pins(&c->journal);
 
-	for_each_member_device(ca, c, i)
-		bch2_dev_allocator_stop(ca);
+	do {
+		ret = bch2_alloc_write(c, false, &wrote);
+		if (ret) {
+			bch2_fs_inconsistent(c, "error writing out alloc info %i", ret);
+			break;
+		}
 
-	bch2_journal_flush_all_pins(&c->journal);
+		ret = bch2_stripes_write(c, &wrote);
+		if (ret) {
+			bch2_fs_inconsistent(c, "error writing out stripes");
+			break;
+		}
 
-	/*
-	 * We need to explicitly wait on btree interior updates to complete
-	 * before stopping the journal, flushing all journal pins isn't
-	 * sufficient, because in the BTREE_INTERIOR_UPDATING_ROOT case btree
-	 * interior updates have to drop their journal pin before they're
-	 * fully complete:
-	 */
-	closure_wait_event(&c->btree_interior_update_wait,
-			   !bch2_btree_interior_updates_nr_pending(c));
+		for_each_member_device(ca, c, i)
+			bch2_dev_allocator_quiesce(c, ca);
+
+		bch2_journal_flush_all_pins(&c->journal);
+
+		/*
+		 * We need to explicitly wait on btree interior updates to complete
+		 * before stopping the journal, flushing all journal pins isn't
+		 * sufficient, because in the BTREE_INTERIOR_UPDATING_ROOT case btree
+		 * interior updates have to drop their journal pin before they're
+		 * fully complete:
+		 */
+		closure_wait_event(&c->btree_interior_update_wait,
+				   !bch2_btree_interior_updates_nr_pending(c));
+	} while (wrote);
+
+	for_each_member_device(ca, c, i)
+		bch2_dev_allocator_stop(ca);
 
 	bch2_fs_journal_stop(&c->journal);
 
+	/* XXX: mark super that alloc info is persistent */
+
 	/*
 	 * the journal kicks off btree writes via reclaim - wait for in flight
 	 * writes after stopping journal:
@@ -420,6 +441,8 @@ static void bch2_fs_free(struct bch_fs *c)
 	kfree(c->replicas_gc.entries);
 	kfree(rcu_dereference_protected(c->disk_groups, 1));
 
+	if (c->journal_reclaim_wq)
+		destroy_workqueue(c->journal_reclaim_wq);
 	if (c->copygc_wq)
 		destroy_workqueue(c->copygc_wq);
 	if (c->wq)
@@ -638,6 +661,8 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 				WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_HIGHPRI, 1)) ||
 	    !(c->copygc_wq = alloc_workqueue("bcache_copygc",
 				WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_HIGHPRI, 1)) ||
+	    !(c->journal_reclaim_wq = alloc_workqueue("bcache_journal",
+				WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_HIGHPRI, 1)) ||
 	    percpu_ref_init(&c->writes, bch2_writes_disabled, 0, GFP_KERNEL) ||
 	    mempool_init_kmalloc_pool(&c->btree_reserve_pool, 1,
 				      sizeof(struct btree_reserve)) ||
@@ -1297,8 +1322,8 @@ int bch2_dev_remove(struct bch_fs *c, struct bch_dev *ca, int flags)
 	if (data) {
 		char data_has_str[100];
 
-		bch2_string_opt_to_text(&PBUF(data_has_str),
-					bch2_data_types, data);
+		bch2_flags_to_text(&PBUF(data_has_str),
+				   bch2_data_types, data);
 		bch_err(ca, "Remove failed, still has data (%s)", data_has_str);
 		ret = -EBUSY;
 		goto err;
diff --git a/libbcachefs/sysfs.c b/libbcachefs/sysfs.c
index 2e6e9bd5..40384e7e 100644
--- a/libbcachefs/sysfs.c
+++ b/libbcachefs/sysfs.c
@@ -234,17 +234,17 @@ static ssize_t show_fs_alloc_debug(struct bch_fs *c, char *buf)
 {
 	struct printbuf out = _PBUF(buf, PAGE_SIZE);
 	struct bch_fs_usage *fs_usage = bch2_fs_usage_read(c);
-	unsigned replicas;
+	unsigned i;
 
 	if (!fs_usage)
 		return -ENOMEM;
 
 	pr_buf(&out, "capacity:\t\t%llu\n", c->capacity);
 
-	for (replicas = 0;
-	     replicas < ARRAY_SIZE(fs_usage->persistent_reserved);
-	     replicas++) {
-		pr_buf(&out, "%u replicas:\n", replicas + 1);
+	for (i = 0;
+	     i < ARRAY_SIZE(fs_usage->persistent_reserved);
+	     i++) {
+		pr_buf(&out, "%u replicas:\n", i + 1);
 #if 0
 		for (type = BCH_DATA_SB; type < BCH_DATA_NR; type++)
 			pr_buf(&out, "\t%s:\t\t%llu\n",
@@ -254,12 +254,23 @@ static ssize_t show_fs_alloc_debug(struct bch_fs *c, char *buf)
 		       stats.replicas[replicas].ec_data);
 #endif
 		pr_buf(&out, "\treserved:\t%llu\n",
-		       fs_usage->persistent_reserved[replicas]);
+		       fs_usage->persistent_reserved[i]);
 	}
 
 	pr_buf(&out, "online reserved:\t%llu\n",
 	       fs_usage->s.online_reserved);
 
+	for (i = 0; i < c->replicas.nr; i++) {
+		struct bch_replicas_entry *e =
+			cpu_replicas_entry(&c->replicas, i);
+
+		pr_buf(&out, "\t");
+		bch2_replicas_entry_to_text(&out, e);
+		pr_buf(&out, ":\t%llu\n", fs_usage->data[i]);
+	}
+
+	percpu_up_read_preempt_enable(&c->mark_lock);
+
 	kfree(fs_usage);
 
 	return out.pos - buf;
@@ -797,6 +808,12 @@ static ssize_t show_dev_alloc_debug(struct bch_dev *ca, char *buf)
 {
 	struct bch_fs *c = ca->fs;
 	struct bch_dev_usage stats = bch2_dev_usage_read(c, ca);
+	unsigned i, nr[BCH_DATA_NR];
+
+	memset(nr, 0, sizeof(nr));
+
+	for (i = 0; i < ARRAY_SIZE(c->open_buckets); i++)
+		nr[c->open_buckets[i].type]++;
 
 	return scnprintf(buf, PAGE_SIZE,
 		"free_inc:               %zu/%zu\n"
@@ -823,7 +840,10 @@ static ssize_t show_dev_alloc_debug(struct bch_dev *ca, char *buf)
 		"    copygc threshold:   %llu\n"
 		"freelist_wait:          %s\n"
 		"open buckets:           %u/%u (reserved %u)\n"
-		"open_buckets_wait:      %s\n",
+		"open_buckets_wait:      %s\n"
+		"open_buckets_btree:     %u\n"
+		"open_buckets_user:      %u\n"
+		"btree reserve cache:    %u\n",
 		fifo_used(&ca->free_inc),		ca->free_inc.size,
 		fifo_used(&ca->free[RESERVE_BTREE]),	ca->free[RESERVE_BTREE].size,
 		fifo_used(&ca->free[RESERVE_MOVINGGC]),	ca->free[RESERVE_MOVINGGC].size,
@@ -845,8 +865,12 @@ static ssize_t show_dev_alloc_debug(struct bch_dev *ca, char *buf)
 		stats.sectors_fragmented,
 		ca->copygc_threshold,
 		c->freelist_wait.list.first		? "waiting" : "empty",
-		c->open_buckets_nr_free, OPEN_BUCKETS_COUNT, BTREE_NODE_RESERVE,
-		c->open_buckets_wait.list.first		? "waiting" : "empty");
+		c->open_buckets_nr_free, OPEN_BUCKETS_COUNT,
+		BTREE_NODE_OPEN_BUCKET_RESERVE,
+		c->open_buckets_wait.list.first		? "waiting" : "empty",
+		nr[BCH_DATA_BTREE],
+		nr[BCH_DATA_USER],
+		c->btree_reserve_cache_nr);
 }
 
 static const char * const bch2_rw[] = {
diff --git a/libbcachefs/util.c b/libbcachefs/util.c
index 5c060e77..fea80e24 100644
--- a/libbcachefs/util.c
+++ b/libbcachefs/util.c
@@ -133,6 +133,7 @@ void bch2_flags_to_text(struct printbuf *out,
 			const char * const list[], u64 flags)
 {
 	unsigned bit, nr = 0;
+	bool first = true;
 
 	if (out->pos != out->end)
 		*out->pos = '\0';
@@ -141,7 +142,10 @@ void bch2_flags_to_text(struct printbuf *out,
 		nr++;
 
 	while (flags && (bit = __ffs(flags)) < nr) {
-		pr_buf(out, "%s,", list[bit]);
+		pr_buf(out, "%s", list[bit]);
+		if (!first)
+			pr_buf(out, ",");
+		first = false;
 		flags ^= 1 << bit;
 	}
 }
@@ -894,3 +898,28 @@ void eytzinger0_find_test(void)
 	kfree(test_array);
 }
 #endif
+
+/*
+ * Accumulate percpu counters onto one cpu's copy - only valid when access
+ * against any percpu counter is guarded against
+ */
+u64 *bch2_acc_percpu_u64s(u64 __percpu *p, unsigned nr)
+{
+	u64 *ret;
+	int cpu;
+
+	preempt_disable();
+	ret = this_cpu_ptr(p);
+	preempt_enable();
+
+	for_each_possible_cpu(cpu) {
+		u64 *i = per_cpu_ptr(p, cpu);
+
+		if (i != ret) {
+			acc_u64s(ret, i, nr);
+			memset(i, 0, nr * sizeof(u64));
+		}
+	}
+
+	return ret;
+}
diff --git a/libbcachefs/util.h b/libbcachefs/util.h
index 25d67509..fbfb2085 100644
--- a/libbcachefs/util.h
+++ b/libbcachefs/util.h
@@ -715,4 +715,6 @@ static inline void acc_u64s_percpu(u64 *acc, const u64 __percpu *src,
 		acc_u64s(acc, per_cpu_ptr(src, cpu), nr);
 }
 
+u64 *bch2_acc_percpu_u64s(u64 __percpu *, unsigned);
+
 #endif /* _BCACHEFS_UTIL_H */
diff --git a/linux/generic-radix-tree.c b/linux/generic-radix-tree.c
index 5c4a275e..4f43d0bb 100644
--- a/linux/generic-radix-tree.c
+++ b/linux/generic-radix-tree.c
@@ -1,4 +1,5 @@
 
+#include <linux/atomic.h>
 #include <linux/export.h>
 #include <linux/generic-radix-tree.h>
 #include <linux/gfp.h>
@@ -16,7 +17,7 @@ struct genradix_node {
 	};
 };
 
-static inline unsigned genradix_depth_shift(unsigned depth)
+static inline int genradix_depth_shift(unsigned depth)
 {
 	return PAGE_SHIFT + GENRADIX_ARY_SHIFT * depth;
 }
@@ -29,16 +30,34 @@ static inline size_t genradix_depth_size(unsigned depth)
 	return 1UL << genradix_depth_shift(depth);
 }
 
+/* depth that's needed for a genradix that can address up to ULONG_MAX: */
+#define GENRADIX_MAX_DEPTH	\
+	DIV_ROUND_UP(BITS_PER_LONG - PAGE_SHIFT, GENRADIX_ARY_SHIFT)
+
+#define GENRADIX_DEPTH_MASK				\
+	((unsigned long) (roundup_pow_of_two(GENRADIX_MAX_DEPTH + 1) - 1))
+
+unsigned genradix_root_to_depth(struct genradix_root *r)
+{
+	return (unsigned long) r & GENRADIX_DEPTH_MASK;
+}
+
+struct genradix_node *genradix_root_to_node(struct genradix_root *r)
+{
+	return (void *) ((unsigned long) r & ~GENRADIX_DEPTH_MASK);
+}
+
 /*
  * Returns pointer to the specified byte @offset within @radix, or NULL if not
  * allocated
  */
 void *__genradix_ptr(struct __genradix *radix, size_t offset)
 {
-	size_t level = radix->depth;
-	struct genradix_node *n = radix->root;
+	struct genradix_root *r = READ_ONCE(radix->root);
+	struct genradix_node *n = genradix_root_to_node(r);
+	unsigned level		= genradix_root_to_depth(r);
 
-	if (offset >= genradix_depth_size(radix->depth))
+	if (ilog2(offset) >= genradix_depth_shift(level))
 		return NULL;
 
 	while (1) {
@@ -64,43 +83,60 @@ EXPORT_SYMBOL(__genradix_ptr);
 void *__genradix_ptr_alloc(struct __genradix *radix, size_t offset,
 			   gfp_t gfp_mask)
 {
-	struct genradix_node **n;
-	size_t level;
+	struct genradix_root *v = READ_ONCE(radix->root);
+	struct genradix_node *n, *new_node = NULL;
+	unsigned level;
 
 	/* Increase tree depth if necessary: */
+	while (1) {
+		struct genradix_root *r = v, *new_root;
 
-	while (offset >= genradix_depth_size(radix->depth)) {
-		struct genradix_node *new_root =
-			(void *) __get_free_page(gfp_mask|__GFP_ZERO);
-
-		if (!new_root)
-			return NULL;
-
-		new_root->children[0] = radix->root;
-		radix->root = new_root;
-		radix->depth++;
-	}
+		n	= genradix_root_to_node(r);
+		level	= genradix_root_to_depth(r);
 
-	n = &radix->root;
-	level = radix->depth;
+		if (n && ilog2(offset) < genradix_depth_shift(level))
+			break;
 
-	while (1) {
-		if (!*n) {
-			*n = (void *) __get_free_page(gfp_mask|__GFP_ZERO);
-			if (!*n)
+		if (!new_node) {
+			new_node = (void *)
+				__get_free_page(gfp_mask|__GFP_ZERO);
+			if (!new_node)
 				return NULL;
 		}
 
-		if (!level)
-			break;
+		new_node->children[0] = n;
+		new_root = ((struct genradix_root *)
+			    ((unsigned long) new_node | (n ? level + 1 : 0)));
 
-		level--;
+		if ((v = cmpxchg_release(&radix->root, r, new_root)) == r) {
+			v = new_root;
+			new_node = NULL;
+		}
+	}
 
-		n = &(*n)->children[offset >> genradix_depth_shift(level)];
+	while (level--) {
+		struct genradix_node **p =
+			&n->children[offset >> genradix_depth_shift(level)];
 		offset &= genradix_depth_size(level) - 1;
+
+		n = READ_ONCE(*p);
+		if (!n) {
+			if (!new_node) {
+				new_node = (void *)
+					__get_free_page(gfp_mask|__GFP_ZERO);
+				if (!new_node)
+					return NULL;
+			}
+
+			if (!(n = cmpxchg_release(p, NULL, new_node)))
+				swap(n, new_node);
+		}
 	}
 
-	return &(*n)->data[offset];
+	if (new_node)
+		free_page((unsigned long) new_node);
+
+	return &n->data[offset];
 }
 EXPORT_SYMBOL(__genradix_ptr_alloc);
 
@@ -108,17 +144,19 @@ void *__genradix_iter_peek(struct genradix_iter *iter,
 			   struct __genradix *radix,
 			   size_t objs_per_page)
 {
+	struct genradix_root *r;
 	struct genradix_node *n;
-	size_t level, i;
-
-	if (!radix->root)
-		return NULL;
+	unsigned level, i;
 restart:
-	if (iter->offset >= genradix_depth_size(radix->depth))
+	r = READ_ONCE(radix->root);
+	if (!r)
 		return NULL;
 
-	n	= radix->root;
-	level	= radix->depth;
+	n	= genradix_root_to_node(r);
+	level	= genradix_root_to_depth(r);
+
+	if (ilog2(iter->offset) >= genradix_depth_shift(level))
+		return NULL;
 
 	while (level) {
 		level--;
@@ -157,11 +195,24 @@ static void genradix_free_recurse(struct genradix_node *n, unsigned level)
 	free_page((unsigned long) n);
 }
 
+int __genradix_prealloc(struct __genradix *radix, size_t size,
+			gfp_t gfp_mask)
+{
+	size_t offset;
+
+	for (offset = 0; offset < size; offset += PAGE_SIZE)
+		if (!__genradix_ptr_alloc(radix, offset, gfp_mask))
+			return -ENOMEM;
+
+	return 0;
+}
+EXPORT_SYMBOL(__genradix_prealloc);
+
 void __genradix_free(struct __genradix *radix)
 {
-	genradix_free_recurse(radix->root, radix->depth);
+	struct genradix_root *r = xchg(&radix->root, NULL);
 
-	radix->root = NULL;
-	radix->depth = 0;
+	genradix_free_recurse(genradix_root_to_node(r),
+			      genradix_root_to_depth(r));
 }
 EXPORT_SYMBOL(__genradix_free);