Update bcachefs sources to fbb669e9de bcachefs: Kill btree_node_iter_large

.bcachefs_revision

@@ -1 +1 @@
-d372ddcbfabef5fcfd29bad150865cccc3faf172
+fbb669e9dec85dc63cdcae71746187d5562dc287

libbcachefs/bcachefs.h

@@ -425,7 +425,6 @@ struct bch_dev {
 	 */
 	alloc_fifo		free[RESERVE_NR];
 	alloc_fifo		free_inc;
-	spinlock_t		freelist_lock;
 
 	u8			open_buckets_partial[OPEN_BUCKETS_COUNT];
 	unsigned		open_buckets_partial_nr;

libbcachefs/bkey_methods.c

@@ -75,10 +75,10 @@ static void key_type_inline_data_to_text(struct printbuf *out, struct bch_fs *c,
 	pr_buf(out, "(%zu bytes)", bkey_val_bytes(k.k));
 }
 
-static const struct bkey_ops bch2_bkey_ops_inline_data = {
+#define bch2_bkey_ops_inline_data (struct bkey_ops) {	\
-	.key_invalid	= key_type_inline_data_invalid,
+	.key_invalid	= key_type_inline_data_invalid,	\
-	.val_to_text	= key_type_inline_data_to_text,
+	.val_to_text	= key_type_inline_data_to_text,	\
-};
+}
 
 static const struct bkey_ops bch2_bkey_ops[] = {
 #define x(name, nr) [KEY_TYPE_##name]	= bch2_bkey_ops_##name,

libbcachefs/bkey_on_stack.h

@@ -19,6 +19,14 @@ static inline void bkey_on_stack_realloc(struct bkey_on_stack *s,
 	}
 }
 
+static inline void bkey_on_stack_reassemble(struct bkey_on_stack *s,
+					    struct bch_fs *c,
+					    struct bkey_s_c k)
+{
+	bkey_on_stack_realloc(s, c, k.k->u64s);
+	bkey_reassemble(s->k, k);
+}
+
 static inline void bkey_on_stack_init(struct bkey_on_stack *s)
 {
 	s->k = (void *) s->onstack;

libbcachefs/bkey_sort.c

@@ -5,90 +5,15 @@
 #include "bset.h"
 #include "extents.h"
 
-/* too many iterators, need to clean this up */
-
-/* btree_node_iter_large: */
-
-#define btree_node_iter_cmp_heap(h, _l, _r) btree_node_iter_cmp(b, _l, _r)
-
-static inline bool
-bch2_btree_node_iter_large_end(struct btree_node_iter_large *iter)
-{
-	return !iter->used;
-}
-
-static inline struct bkey_packed *
-bch2_btree_node_iter_large_peek_all(struct btree_node_iter_large *iter,
-				    struct btree *b)
-{
-	return bch2_btree_node_iter_large_end(iter)
-		? NULL
-		: __btree_node_offset_to_key(b, iter->data->k);
-}
-
-static void
-bch2_btree_node_iter_large_advance(struct btree_node_iter_large *iter,
-				   struct btree *b)
-{
-	iter->data->k += __btree_node_offset_to_key(b, iter->data->k)->u64s;
-
-	EBUG_ON(!iter->used);
-	EBUG_ON(iter->data->k > iter->data->end);
-
-	if (iter->data->k == iter->data->end)
-		heap_del(iter, 0, btree_node_iter_cmp_heap, NULL);
-	else
-		heap_sift_down(iter, 0, btree_node_iter_cmp_heap, NULL);
-}
-
-static inline struct bkey_packed *
-bch2_btree_node_iter_large_next_all(struct btree_node_iter_large *iter,
-				    struct btree *b)
-{
-	struct bkey_packed *ret = bch2_btree_node_iter_large_peek_all(iter, b);
-
-	if (ret)
-		bch2_btree_node_iter_large_advance(iter, b);
-
-	return ret;
-}
-
-void bch2_btree_node_iter_large_push(struct btree_node_iter_large *iter,
-				     struct btree *b,
-				     const struct bkey_packed *k,
-				     const struct bkey_packed *end)
-{
-	if (k != end) {
-		struct btree_node_iter_set n =
-			((struct btree_node_iter_set) {
-				 __btree_node_key_to_offset(b, k),
-				 __btree_node_key_to_offset(b, end)
-			 });
-
-		__heap_add(iter, n, btree_node_iter_cmp_heap, NULL);
-	}
-}
-
-static void sort_key_next(struct btree_node_iter_large *iter,
-			  struct btree *b,
-			  struct btree_node_iter_set *i)
-{
-	i->k += __btree_node_offset_to_key(b, i->k)->u64s;
-
-	while (i->k != i->end &&
-	       !__btree_node_offset_to_key(b, i->k)->u64s)
-		i->k++;
-
-	if (i->k == i->end)
-		*i = iter->data[--iter->used];
-}
-
-/* regular sort_iters */
-
 typedef int (*sort_cmp_fn)(struct btree *,
 			   struct bkey_packed *,
 			   struct bkey_packed *);
 
+static inline bool sort_iter_end(struct sort_iter *iter)
+{
+	return !iter->used;
+}
+
 static inline void __sort_iter_sift(struct sort_iter *iter,
 				    unsigned from,
 				    sort_cmp_fn cmp)
@@ -118,19 +43,29 @@ static inline void sort_iter_sort(struct sort_iter *iter, sort_cmp_fn cmp)
 
 static inline struct bkey_packed *sort_iter_peek(struct sort_iter *iter)
 {
-	return iter->used ? iter->data->k : NULL;
+	return !sort_iter_end(iter) ? iter->data->k : NULL;
+}
+
+static inline void __sort_iter_advance(struct sort_iter *iter,
+				       unsigned idx, sort_cmp_fn cmp)
+{
+	struct sort_iter_set *i = iter->data + idx;
+
+	BUG_ON(idx >= iter->used);
+
+	i->k = bkey_next_skip_noops(i->k, i->end);
+
+	BUG_ON(i->k > i->end);
+
+	if (i->k == i->end)
+		array_remove_item(iter->data, iter->used, idx);
+	else
+		__sort_iter_sift(iter, idx, cmp);
 }
 
 static inline void sort_iter_advance(struct sort_iter *iter, sort_cmp_fn cmp)
 {
-	iter->data->k = bkey_next_skip_noops(iter->data->k, iter->data->end);
+	__sort_iter_advance(iter, 0, cmp);
-
-	BUG_ON(iter->data->k > iter->data->end);
-
-	if (iter->data->k == iter->data->end)
-		array_remove_item(iter->data, iter->used, 0);
-	else
-		sort_iter_sift(iter, cmp);
 }
 
 static inline struct bkey_packed *sort_iter_next(struct sort_iter *iter,
@@ -145,70 +80,50 @@ static inline struct bkey_packed *sort_iter_next(struct sort_iter *iter,
 }
 
 /*
- * Returns true if l > r - unless l == r, in which case returns true if l is
+ * If keys compare equal, compare by pointer order:
- * older than r.
- *
- * Necessary for btree_sort_fixup() - if there are multiple keys that compare
- * equal in different sets, we have to process them newest to oldest.
  */
-#define key_sort_cmp(h, l, r)						\
+static inline int key_sort_fix_overlapping_cmp(struct btree *b,
-({									\
+					       struct bkey_packed *l,
-	bkey_cmp_packed(b,						\
+					       struct bkey_packed *r)
-			__btree_node_offset_to_key(b, (l).k),		\
-			__btree_node_offset_to_key(b, (r).k))		\
-									\
-	?: (l).k - (r).k;						\
-})
-
-static inline bool should_drop_next_key(struct btree_node_iter_large *iter,
-					struct btree *b)
 {
-	struct btree_node_iter_set *l = iter->data, *r = iter->data + 1;
+	return bkey_cmp_packed(b, l, r) ?:
-	struct bkey_packed *k = __btree_node_offset_to_key(b, l->k);
+		cmp_int((unsigned long) l, (unsigned long) r);
-
-	if (bkey_whiteout(k))
-		return true;
-
-	if (iter->used < 2)
-		return false;
-
-	if (iter->used > 2 &&
-	    key_sort_cmp(iter, r[0], r[1]) >= 0)
-		r++;
-
-	/*
-	 * key_sort_cmp() ensures that when keys compare equal the older key
-	 * comes first; so if l->k compares equal to r->k then l->k is older and
-	 * should be dropped.
-	 */
-	return !bkey_cmp_packed(b,
-				__btree_node_offset_to_key(b, l->k),
-				__btree_node_offset_to_key(b, r->k));
 }
 
-struct btree_nr_keys bch2_key_sort_fix_overlapping(struct bset *dst,
+static inline bool should_drop_next_key(struct sort_iter *iter)
-					struct btree *b,
+{
-					struct btree_node_iter_large *iter)
+	/*
+	 * key_sort_cmp() ensures that when keys compare equal the older key
+	 * comes first; so if l->k compares equal to r->k then l->k is older
+	 * and should be dropped.
+	 */
+	return iter->used >= 2 &&
+		!bkey_cmp_packed(iter->b,
+				 iter->data[0].k,
+				 iter->data[1].k);
+}
+
+struct btree_nr_keys
+bch2_key_sort_fix_overlapping(struct bch_fs *c, struct bset *dst,
+			      struct sort_iter *iter)
 {
 	struct bkey_packed *out = dst->start;
+	struct bkey_packed *k;
 	struct btree_nr_keys nr;
 
 	memset(&nr, 0, sizeof(nr));
 
-	heap_resort(iter, key_sort_cmp, NULL);
+	sort_iter_sort(iter, key_sort_fix_overlapping_cmp);
-
-	while (!bch2_btree_node_iter_large_end(iter)) {
-		if (!should_drop_next_key(iter, b)) {
-			struct bkey_packed *k =
-				__btree_node_offset_to_key(b, iter->data->k);
 
+	while ((k = sort_iter_peek(iter))) {
+		if (!bkey_whiteout(k) &&
+		    !should_drop_next_key(iter)) {
 			bkey_copy(out, k);
 			btree_keys_account_key_add(&nr, 0, out);
 			out = bkey_next(out);
 		}
 
-		sort_key_next(iter, b, iter->data);
+		sort_iter_advance(iter, key_sort_fix_overlapping_cmp);
-		heap_sift_down(iter, 0, key_sort_cmp, NULL);
 	}
 
 	dst->u64s = cpu_to_le16((u64 *) out - dst->_data);
@@ -221,29 +136,16 @@ struct btree_nr_keys bch2_key_sort_fix_overlapping(struct bset *dst,
  * Necessary for sort_fix_overlapping() - if there are multiple keys that
  * compare equal in different sets, we have to process them newest to oldest.
  */
-#define extent_sort_cmp(h, l, r)					\
+static inline int extent_sort_fix_overlapping_cmp(struct btree *b,
-({									\
+						  struct bkey_packed *l,
-	struct bkey _ul = bkey_unpack_key(b,				\
+						  struct bkey_packed *r)
-				__btree_node_offset_to_key(b, (l).k));	\
-	struct bkey _ur = bkey_unpack_key(b,				\
-				__btree_node_offset_to_key(b, (r).k));	\
-									\
-	bkey_cmp(bkey_start_pos(&_ul),					\
-		 bkey_start_pos(&_ur)) ?: (r).k - (l).k;		\
-})
-
-static inline void extent_sort_sift(struct btree_node_iter_large *iter,
-				    struct btree *b, size_t i)
 {
-	heap_sift_down(iter, i, extent_sort_cmp, NULL);
+	struct bkey ul = bkey_unpack_key(b, l);
-}
+	struct bkey ur = bkey_unpack_key(b, r);
 
-static inline void extent_sort_next(struct btree_node_iter_large *iter,
+	return bkey_cmp(bkey_start_pos(&ul),
-				    struct btree *b,
+			bkey_start_pos(&ur)) ?:
-				    struct btree_node_iter_set *i)
+		cmp_int((unsigned long) r, (unsigned long) l);
-{
-	sort_key_next(iter, b, i);
-	heap_sift_down(iter, i - iter->data, extent_sort_cmp, NULL);
 }
 
 static void extent_sort_advance_prev(struct bkey_format *f,
@@ -286,14 +188,14 @@ static void extent_sort_append(struct bch_fs *c,
 	bkey_reassemble((void *) *prev, k.s_c);
 }
 
-struct btree_nr_keys bch2_extent_sort_fix_overlapping(struct bch_fs *c,
+struct btree_nr_keys
-					struct bset *dst,
+bch2_extent_sort_fix_overlapping(struct bch_fs *c, struct bset *dst,
-					struct btree *b,
+				 struct sort_iter *iter)
-					struct btree_node_iter_large *iter)
 {
+	struct btree *b = iter->b;
 	struct bkey_format *f = &b->format;
-	struct btree_node_iter_set *_l = iter->data, *_r;
+	struct sort_iter_set *_l = iter->data, *_r = iter->data + 1;
-	struct bkey_packed *prev = NULL, *lk, *rk;
+	struct bkey_packed *prev = NULL;
 	struct bkey l_unpacked, r_unpacked;
 	struct bkey_s l, r;
 	struct btree_nr_keys nr;
@@ -302,36 +204,32 @@ struct btree_nr_keys bch2_extent_sort_fix_overlapping(struct bch_fs *c,
 	memset(&nr, 0, sizeof(nr));
 	bkey_on_stack_init(&split);
 
-	heap_resort(iter, extent_sort_cmp, NULL);
+	sort_iter_sort(iter, extent_sort_fix_overlapping_cmp);
 
-	while (!bch2_btree_node_iter_large_end(iter)) {
+	while (!sort_iter_end(iter)) {
-		lk = __btree_node_offset_to_key(b, _l->k);
+		l = __bkey_disassemble(b, _l->k, &l_unpacked);
-		l = __bkey_disassemble(b, lk, &l_unpacked);
 
 		if (iter->used == 1) {
 			extent_sort_append(c, f, &nr, dst->start, &prev, l);
-			extent_sort_next(iter, b, _l);
+			sort_iter_advance(iter,
+					  extent_sort_fix_overlapping_cmp);
 			continue;
 		}
 
-		_r = iter->data + 1;
+		r = __bkey_disassemble(b, _r->k, &r_unpacked);
-		if (iter->used > 2 &&
-		    extent_sort_cmp(iter, _r[0], _r[1]) >= 0)
-			_r++;
-
-		rk = __btree_node_offset_to_key(b, _r->k);
-		r = __bkey_disassemble(b, rk, &r_unpacked);
 
 		/* If current key and next key don't overlap, just append */
 		if (bkey_cmp(l.k->p, bkey_start_pos(r.k)) <= 0) {
 			extent_sort_append(c, f, &nr, dst->start, &prev, l);
-			extent_sort_next(iter, b, _l);
+			sort_iter_advance(iter,
+					  extent_sort_fix_overlapping_cmp);
 			continue;
 		}
 
 		/* Skip 0 size keys */
 		if (!r.k->size) {
-			extent_sort_next(iter, b, _r);
+			__sort_iter_advance(iter, 1,
+					    extent_sort_fix_overlapping_cmp);
 			continue;
 		}
 
@@ -348,32 +246,33 @@ struct btree_nr_keys bch2_extent_sort_fix_overlapping(struct bch_fs *c,
 		if (_l->k > _r->k) {
 			/* l wins, trim r */
 			if (bkey_cmp(l.k->p, r.k->p) >= 0) {
-				sort_key_next(iter, b, _r);
+				__sort_iter_advance(iter, 1,
+					 extent_sort_fix_overlapping_cmp);
 			} else {
 				bch2_cut_front_s(l.k->p, r);
-				extent_save(b, rk, r.k);
+				extent_save(b, _r->k, r.k);
+				__sort_iter_sift(iter, 1,
+					 extent_sort_fix_overlapping_cmp);
 			}
-
-			extent_sort_sift(iter, b, _r - iter->data);
 		} else if (bkey_cmp(l.k->p, r.k->p) > 0) {
-			bkey_on_stack_realloc(&split, c, l.k->u64s);
 
 			/*
 			 * r wins, but it overlaps in the middle of l - split l:
 			 */
-			bkey_reassemble(split.k, l.s_c);
+			bkey_on_stack_reassemble(&split, c, l.s_c);
 			bch2_cut_back(bkey_start_pos(r.k), split.k);
 
 			bch2_cut_front_s(r.k->p, l);
-			extent_save(b, lk, l.k);
+			extent_save(b, _l->k, l.k);
 
-			extent_sort_sift(iter, b, 0);
+			__sort_iter_sift(iter, 0,
+					 extent_sort_fix_overlapping_cmp);
 
 			extent_sort_append(c, f, &nr, dst->start,
 					   &prev, bkey_i_to_s(split.k));
 		} else {
 			bch2_cut_back_s(bkey_start_pos(r.k), l);
-			extent_save(b, lk, l.k);
+			extent_save(b, _l->k, l.k);
 		}
 	}
 
@@ -531,28 +430,6 @@ unsigned bch2_sort_extents(struct bkey_packed *dst,
 	return (u64 *) out - (u64 *) dst;
 }
 
-static inline int sort_key_whiteouts_cmp(struct btree *b,
-					 struct bkey_packed *l,
-					 struct bkey_packed *r)
-{
-	return bkey_cmp_packed(b, l, r);
-}
-
-unsigned bch2_sort_key_whiteouts(struct bkey_packed *dst,
-				 struct sort_iter *iter)
-{
-	struct bkey_packed *in, *out = dst;
-
-	sort_iter_sort(iter, sort_key_whiteouts_cmp);
-
-	while ((in = sort_iter_next(iter, sort_key_whiteouts_cmp))) {
-		bkey_copy(out, in);
-		out = bkey_next(out);
-	}
-
-	return (u64 *) out - (u64 *) dst;
-}
-
 static inline int sort_extent_whiteouts_cmp(struct btree *b,
 					    struct bkey_packed *l,
 					    struct bkey_packed *r)

libbcachefs/bkey_sort.h

@@ -2,20 +2,10 @@
 #ifndef _BCACHEFS_BKEY_SORT_H
 #define _BCACHEFS_BKEY_SORT_H
 
-struct btree_node_iter_large {
-	u16		used;
-
-	struct btree_node_iter_set data[MAX_BSETS];
-};
-
-void bch2_btree_node_iter_large_push(struct btree_node_iter_large *,
-				     struct btree *,
-				     const struct bkey_packed *,
-				     const struct bkey_packed *);
-
 struct sort_iter {
-	struct btree	*b;
+	struct btree		*b;
 	unsigned		used;
+	unsigned		size;
 
 	struct sort_iter_set {
 		struct bkey_packed *k, *end;
@@ -24,27 +14,27 @@ struct sort_iter {
 
 static inline void sort_iter_init(struct sort_iter *iter, struct btree *b)
 {
-	memset(iter, 0, sizeof(*iter));
 	iter->b = b;
+	iter->used = 0;
+	iter->size = ARRAY_SIZE(iter->data);
 }
 
 static inline void sort_iter_add(struct sort_iter *iter,
 				 struct bkey_packed *k,
 				 struct bkey_packed *end)
 {
-	BUG_ON(iter->used >= ARRAY_SIZE(iter->data));
+	BUG_ON(iter->used >= iter->size);
 
 	if (k != end)
 		iter->data[iter->used++] = (struct sort_iter_set) { k, end };
 }
 
 struct btree_nr_keys
-bch2_key_sort_fix_overlapping(struct bset *, struct btree *,
+bch2_key_sort_fix_overlapping(struct bch_fs *, struct bset *,
-			      struct btree_node_iter_large *);
+			      struct sort_iter *);
 struct btree_nr_keys
 bch2_extent_sort_fix_overlapping(struct bch_fs *, struct bset *,
-				 struct btree *,
+				 struct sort_iter *);
-				 struct btree_node_iter_large *);
 
 struct btree_nr_keys
 bch2_sort_repack(struct bset *, struct btree *,
@@ -61,8 +51,6 @@ unsigned bch2_sort_keys(struct bkey_packed *,
 unsigned bch2_sort_extents(struct bkey_packed *,
 			   struct sort_iter *, bool);
 
-unsigned bch2_sort_key_whiteouts(struct bkey_packed *,
-				 struct sort_iter *);
 unsigned bch2_sort_extent_whiteouts(struct bkey_packed *,
 				    struct sort_iter *);
 

libbcachefs/bset.c

@@ -255,8 +255,7 @@ void bch2_verify_insert_pos(struct btree *b, struct bkey_packed *where,
 		panic("prev > insert:\n"
 		      "prev    key %5u %s\n"
 		      "insert  key %5u %s\n",
-		       __btree_node_key_to_offset(b, prev), buf1,
+		      buf1, buf2);
-		       __btree_node_key_to_offset(b, insert), buf2);
 	}
 #endif
 #if 0
@@ -275,10 +274,9 @@ void bch2_verify_insert_pos(struct btree *b, struct bkey_packed *where,
 		bch2_bkey_to_text(&PBUF(buf2), &k2);
 
 		panic("insert > next:\n"
-		      "insert  key %5u %s\n"
+		      "insert  key %s\n"
-		      "next    key %5u %s\n",
+		      "next    key %s\n",
-		       __btree_node_key_to_offset(b, insert), buf1,
+		      buf1, buf2);
-		       __btree_node_key_to_offset(b, next), buf2);
 	}
 #endif
 }

libbcachefs/btree_cache.c

@@ -557,7 +557,6 @@ out:
 	b->sib_u64s[0]		= 0;
 	b->sib_u64s[1]		= 0;
 	b->whiteout_u64s	= 0;
-	b->uncompacted_whiteout_u64s = 0;
 	bch2_btree_keys_init(b, &c->expensive_debug_checks);
 
 	bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],

libbcachefs/btree_io.c

@@ -80,27 +80,102 @@ static void *btree_bounce_alloc(struct bch_fs *c, unsigned order,
 	return mempool_alloc(&c->btree_bounce_pool, GFP_NOIO);
 }
 
-static unsigned should_compact_bset(struct btree *b, struct bset_tree *t,
+static void sort_bkey_ptrs(const struct btree *bt,
-				    bool compacting,
+			   struct bkey_packed **ptrs, unsigned nr)
-				    enum compact_mode mode)
 {
-	unsigned bset_u64s = le16_to_cpu(bset(b, t)->u64s);
+	unsigned n = nr, a = nr / 2, b, c, d;
-	unsigned dead_u64s = bset_u64s - b->nr.bset_u64s[t - b->set];
 
-	if (mode == COMPACT_LAZY) {
+	if (!a)
-		if (should_compact_bset_lazy(b, t) ||
+		return;
-		    (compacting && !bset_written(b, bset(b, t))))
+
-			return dead_u64s;
+	/* Heap sort: see lib/sort.c: */
-	} else {
+	while (1) {
-		if (bset_written(b, bset(b, t)))
+		if (a)
-			return dead_u64s;
+			a--;
+		else if (--n)
+			swap(ptrs[0], ptrs[n]);
+		else
+			break;
+
+		for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
+			b = bkey_cmp_packed(bt,
+					    ptrs[c],
+					    ptrs[d]) >= 0 ? c : d;
+		if (d == n)
+			b = c;
+
+		while (b != a &&
+		       bkey_cmp_packed(bt,
+				       ptrs[a],
+				       ptrs[b]) >= 0)
+			b = (b - 1) / 2;
+		c = b;
+		while (b != a) {
+			b = (b - 1) / 2;
+			swap(ptrs[b], ptrs[c]);
+		}
 	}
-
-	return 0;
 }
 
-bool __bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
+static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
-			     enum compact_mode mode)
+{
+	struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
+	bool used_mempool = false;
+	unsigned order;
+
+	if (!b->whiteout_u64s)
+		return;
+
+	order = get_order(b->whiteout_u64s * sizeof(u64));
+
+	new_whiteouts = btree_bounce_alloc(c, order, &used_mempool);
+
+	ptrs = ptrs_end = ((void *) new_whiteouts + (PAGE_SIZE << order));
+
+	for (k = unwritten_whiteouts_start(c, b);
+	     k != unwritten_whiteouts_end(c, b);
+	     k = bkey_next(k))
+		*--ptrs = k;
+
+	sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
+
+	k = new_whiteouts;
+
+	while (ptrs != ptrs_end) {
+		bkey_copy(k, *ptrs);
+		k = bkey_next(k);
+		ptrs++;
+	}
+
+	verify_no_dups(b, new_whiteouts,
+		       (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
+
+	memcpy_u64s(unwritten_whiteouts_start(c, b),
+		    new_whiteouts, b->whiteout_u64s);
+
+	btree_bounce_free(c, order, used_mempool, new_whiteouts);
+}
+
+static bool should_compact_bset(struct btree *b, struct bset_tree *t,
+				bool compacting, enum compact_mode mode)
+{
+	if (!bset_dead_u64s(b, t))
+		return false;
+
+	switch (mode) {
+	case COMPACT_LAZY:
+		return should_compact_bset_lazy(b, t) ||
+			(compacting && !bset_written(b, bset(b, t)));
+	case COMPACT_ALL:
+		return true;
+	default:
+		BUG();
+	}
+}
+
+static bool bch2_compact_extent_whiteouts(struct bch_fs *c,
+					  struct btree *b,
+					  enum compact_mode mode)
 {
 	const struct bkey_format *f = &b->format;
 	struct bset_tree *t;
@@ -110,13 +185,17 @@ bool __bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 	unsigned order, whiteout_u64s = 0, u64s;
 	bool used_mempool, compacting = false;
 
+	BUG_ON(!btree_node_is_extents(b));
+
 	for_each_bset(b, t)
-		whiteout_u64s += should_compact_bset(b, t,
+		if (should_compact_bset(b, t, whiteout_u64s != 0, mode))
-					whiteout_u64s != 0, mode);
+			whiteout_u64s += bset_dead_u64s(b, t);
 
 	if (!whiteout_u64s)
 		return false;
 
+	bch2_sort_whiteouts(c, b);
+
 	sort_iter_init(&sort_iter, b);
 
 	whiteout_u64s += b->whiteout_u64s;
@@ -139,9 +218,12 @@ bool __bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 		if (t != b->set && !bset_written(b, i)) {
 			src = container_of(i, struct btree_node_entry, keys);
 			dst = max(write_block(b),
-				  (void *) btree_bkey_last(b, t -1));
+				  (void *) btree_bkey_last(b, t - 1));
 		}
 
+		if (src != dst)
+			compacting = true;
+
 		if (!should_compact_bset(b, t, compacting, mode)) {
 			if (src != dst) {
 				memmove(dst, src, sizeof(*src) +
@@ -169,18 +251,21 @@ bool __bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 		for (k = start; k != end; k = n) {
 			n = bkey_next_skip_noops(k, end);
 
-			if (bkey_deleted(k) && btree_node_is_extents(b))
+			if (bkey_deleted(k))
 				continue;
 
+			BUG_ON(bkey_whiteout(k) &&
+			       k->needs_whiteout &&
+			       bkey_written(b, k));
+
 			if (bkey_whiteout(k) && !k->needs_whiteout)
 				continue;
 
 			if (bkey_whiteout(k)) {
-				unreserve_whiteout(b, k);
 				memcpy_u64s(u_pos, k, bkeyp_key_u64s(f, k));
 				set_bkeyp_val_u64s(f, u_pos, 0);
 				u_pos = bkey_next(u_pos);
-			} else if (mode != COMPACT_WRITTEN_NO_WRITE_LOCK) {
+			} else {
 				bkey_copy(out, k);
 				out = bkey_next(out);
 			}
@@ -188,11 +273,9 @@ bool __bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 
 		sort_iter_add(&sort_iter, u_start, u_pos);
 
-		if (mode != COMPACT_WRITTEN_NO_WRITE_LOCK) {
+		i->u64s = cpu_to_le16((u64 *) out - i->_data);
-			i->u64s = cpu_to_le16((u64 *) out - i->_data);
+		set_btree_bset_end(b, t);
-			set_btree_bset_end(b, t);
+		bch2_bset_set_no_aux_tree(b, t);
-			bch2_bset_set_no_aux_tree(b, t);
-		}
 	}
 
 	b->whiteout_u64s = (u64 *) u_pos - (u64 *) whiteouts;
@@ -200,13 +283,10 @@ bool __bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 	BUG_ON((void *) unwritten_whiteouts_start(c, b) <
 	       (void *) btree_bkey_last(b, bset_tree_last(b)));
 
-	u64s = (btree_node_is_extents(b)
+	u64s = bch2_sort_extent_whiteouts(unwritten_whiteouts_start(c, b),
-		? bch2_sort_extent_whiteouts
+					  &sort_iter);
-		: bch2_sort_key_whiteouts)(unwritten_whiteouts_start(c, b),
-					   &sort_iter);
 
 	BUG_ON(u64s > b->whiteout_u64s);
-	BUG_ON(u64s != b->whiteout_u64s && !btree_node_is_extents(b));
 	BUG_ON(u_pos != whiteouts && !u64s);
 
 	if (u64s != b->whiteout_u64s) {
@@ -222,8 +302,7 @@ bool __bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 
 	btree_bounce_free(c, order, used_mempool, whiteouts);
 
-	if (mode != COMPACT_WRITTEN_NO_WRITE_LOCK)
+	bch2_btree_build_aux_trees(b);
-		bch2_btree_build_aux_trees(b);
 
 	bch_btree_keys_u64s_remaining(c, b);
 	bch2_verify_btree_nr_keys(b);
@@ -231,7 +310,7 @@ bool __bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 	return true;
 }
 
-static bool bch2_drop_whiteouts(struct btree *b)
+static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
 {
 	struct bset_tree *t;
 	bool ret = false;
@@ -239,21 +318,34 @@ static bool bch2_drop_whiteouts(struct btree *b)
 	for_each_bset(b, t) {
 		struct bset *i = bset(b, t);
 		struct bkey_packed *k, *n, *out, *start, *end;
+		struct btree_node_entry *src = NULL, *dst = NULL;
 
-		if (!should_compact_bset(b, t, true, COMPACT_WRITTEN))
+		if (t != b->set && !bset_written(b, i)) {
+			src = container_of(i, struct btree_node_entry, keys);
+			dst = max(write_block(b),
+				  (void *) btree_bkey_last(b, t - 1));
+		}
+
+		if (src != dst)
+			ret = true;
+
+		if (!should_compact_bset(b, t, ret, mode)) {
+			if (src != dst) {
+				memmove(dst, src, sizeof(*src) +
+					le16_to_cpu(src->keys.u64s) *
+					sizeof(u64));
+				i = &dst->keys;
+				set_btree_bset(b, t, i);
+			}
 			continue;
+		}
 
 		start	= btree_bkey_first(b, t);
 		end	= btree_bkey_last(b, t);
 
-		if (!bset_written(b, i) &&
+		if (src != dst) {
-		    t != b->set) {
+			memmove(dst, src, sizeof(*src));
-			struct bset *dst =
+			i = &dst->keys;
-			       max_t(struct bset *, write_block(b),
-				     (void *) btree_bkey_last(b, t -1));
-
-			memmove(dst, i, sizeof(struct bset));
-			i = dst;
 			set_btree_bset(b, t, i);
 		}
 
@@ -265,19 +357,32 @@ static bool bch2_drop_whiteouts(struct btree *b)
 			if (!bkey_whiteout(k)) {
 				bkey_copy(out, k);
 				out = bkey_next(out);
+			} else {
+				BUG_ON(k->needs_whiteout);
 			}
 		}
 
 		i->u64s = cpu_to_le16((u64 *) out - i->_data);
+		set_btree_bset_end(b, t);
 		bch2_bset_set_no_aux_tree(b, t);
 		ret = true;
 	}
 
 	bch2_verify_btree_nr_keys(b);
 
+	bch2_btree_build_aux_trees(b);
+
 	return ret;
 }
 
+bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
+			    enum compact_mode mode)
+{
+	return !btree_node_is_extents(b)
+		? bch2_drop_whiteouts(b, mode)
+		: bch2_compact_extent_whiteouts(c, b, mode);
+}
+
 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 			    struct btree_iter *iter,
 			    unsigned start_idx,
@@ -758,7 +863,7 @@ fsck_err:
 int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry)
 {
 	struct btree_node_entry *bne;
-	struct btree_node_iter_large *iter;
+	struct sort_iter *iter;
 	struct btree_node *sorted;
 	struct bkey_packed *k;
 	struct bset *i;
@@ -767,7 +872,8 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry
 	int ret, retry_read = 0, write = READ;
 
 	iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
-	iter->used = 0;
+	sort_iter_init(iter, b);
+	iter->size = (btree_blocks(c) + 1) * 2;
 
 	if (bch2_meta_read_fault("btree"))
 		btree_err(BTREE_ERR_MUST_RETRY, c, b, NULL,
@@ -846,13 +952,12 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry
 		if (blacklisted && !first)
 			continue;
 
-		bch2_btree_node_iter_large_push(iter, b,
+		sort_iter_add(iter, i->start,
-					   i->start,
+			      vstruct_idx(i, whiteout_u64s));
-					   vstruct_idx(i, whiteout_u64s));
 
-		bch2_btree_node_iter_large_push(iter, b,
+		sort_iter_add(iter,
-					   vstruct_idx(i, whiteout_u64s),
+			      vstruct_idx(i, whiteout_u64s),
-					   vstruct_last(i));
+			      vstruct_last(i));
 	}
 
 	for (bne = write_block(b);
@@ -867,9 +972,9 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry
 
 	set_btree_bset(b, b->set, &b->data->keys);
 
-	b->nr = btree_node_is_extents(b)
+	b->nr = (btree_node_is_extents(b)
-		? bch2_extent_sort_fix_overlapping(c, &sorted->keys, b, iter)
+		 ? bch2_extent_sort_fix_overlapping
-		: bch2_key_sort_fix_overlapping(&sorted->keys, b, iter);
+		 : bch2_key_sort_fix_overlapping)(c, &sorted->keys, iter);
 
 	u64s = le16_to_cpu(sorted->keys.u64s);
 	*sorted = *b->data;
@@ -1343,21 +1448,7 @@ void __bch2_btree_node_write(struct bch_fs *c, struct btree *b,
 	BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
 	BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
 
-	/*
+	bch2_sort_whiteouts(c, b);
-	 * We can't block on six_lock_write() here; another thread might be
-	 * trying to get a journal reservation with read locks held, and getting
-	 * a journal reservation might be blocked on flushing the journal and
-	 * doing btree writes:
-	 */
-	if (lock_type_held == SIX_LOCK_intent &&
-	    six_trylock_write(&b->lock)) {
-		__bch2_compact_whiteouts(c, b, COMPACT_WRITTEN);
-		six_unlock_write(&b->lock);
-	} else {
-		__bch2_compact_whiteouts(c, b, COMPACT_WRITTEN_NO_WRITE_LOCK);
-	}
-
-	BUG_ON(b->uncompacted_whiteout_u64s);
 
 	sort_iter_init(&sort_iter, b);
 
@@ -1545,7 +1636,6 @@ bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
 		return false;
 
 	BUG_ON(b->whiteout_u64s);
-	BUG_ON(b->uncompacted_whiteout_u64s);
 
 	clear_btree_node_just_written(b);
 
@@ -1566,7 +1656,7 @@ bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
 		btree_node_sort(c, b, NULL, 0, b->nsets, true);
 		invalidated_iter = true;
 	} else {
-		invalidated_iter = bch2_drop_whiteouts(b);
+		invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
 	}
 
 	for_each_bset(b, t)

libbcachefs/btree_io.h

@@ -54,16 +54,17 @@ static inline bool btree_node_may_write(struct btree *b)
 
 enum compact_mode {
 	COMPACT_LAZY,
-	COMPACT_WRITTEN,
+	COMPACT_ALL,
-	COMPACT_WRITTEN_NO_WRITE_LOCK,
 };
 
-bool __bch2_compact_whiteouts(struct bch_fs *, struct btree *, enum compact_mode);
+bool bch2_compact_whiteouts(struct bch_fs *, struct btree *,
+			    enum compact_mode);
 
-static inline unsigned should_compact_bset_lazy(struct btree *b, struct bset_tree *t)
+static inline bool should_compact_bset_lazy(struct btree *b,
+					    struct bset_tree *t)
 {
 	unsigned total_u64s = bset_u64s(t);
-	unsigned dead_u64s = total_u64s - b->nr.bset_u64s[t - b->set];
+	unsigned dead_u64s = bset_dead_u64s(b, t);
 
 	return dead_u64s > 64 && dead_u64s * 3 > total_u64s;
 }
@@ -74,7 +75,7 @@ static inline bool bch2_maybe_compact_whiteouts(struct bch_fs *c, struct btree *
 
 	for_each_bset(b, t)
 		if (should_compact_bset_lazy(b, t))
-			return __bch2_compact_whiteouts(c, b, COMPACT_LAZY);
+			return bch2_compact_whiteouts(c, b, COMPACT_LAZY);
 
 	return false;
 }

libbcachefs/btree_types.h

@@ -94,7 +94,6 @@ struct btree {
 	struct btree_nr_keys	nr;
 	u16			sib_u64s[2];
 	u16			whiteout_u64s;
-	u16			uncompacted_whiteout_u64s;
 	u8			page_order;
 	u8			unpack_fn_len;
 
@@ -421,6 +420,11 @@ static inline unsigned bset_u64s(struct bset_tree *t)
 		sizeof(struct bset) / sizeof(u64);
 }
 
+static inline unsigned bset_dead_u64s(struct btree *b, struct bset_tree *t)
+{
+	return bset_u64s(t) - b->nr.bset_u64s[t - b->set];
+}
+
 static inline unsigned bset_byte_offset(struct btree *b, void *i)
 {
 	return i - (void *) b->data;

libbcachefs/btree_update_interior.h

@@ -251,8 +251,7 @@ static inline ssize_t __bch_btree_u64s_remaining(struct bch_fs *c,
 						 void *end)
 {
 	ssize_t used = bset_byte_offset(b, end) / sizeof(u64) +
-		b->whiteout_u64s +
+		b->whiteout_u64s;
-		b->uncompacted_whiteout_u64s;
 	ssize_t total = c->opts.btree_node_size << 6;
 
 	return total - used;
@@ -302,23 +301,19 @@ static inline struct btree_node_entry *want_new_bset(struct bch_fs *c,
 	return NULL;
 }
 
-static inline void unreserve_whiteout(struct btree *b, struct bkey_packed *k)
+static inline void push_whiteout(struct bch_fs *c, struct btree *b,
+				 struct bkey_packed *k)
 {
-	if (bkey_written(b, k)) {
+	unsigned u64s = bkeyp_key_u64s(&b->format, k);
-		EBUG_ON(b->uncompacted_whiteout_u64s <
+	struct bkey_packed *dst;
-			bkeyp_key_u64s(&b->format, k));
-		b->uncompacted_whiteout_u64s -=
-			bkeyp_key_u64s(&b->format, k);
-	}
-}
 
-static inline void reserve_whiteout(struct btree *b, struct bkey_packed *k)
+	BUG_ON(u64s > bch_btree_keys_u64s_remaining(c, b));
-{
+
-	if (bkey_written(b, k)) {
+	b->whiteout_u64s += bkeyp_key_u64s(&b->format, k);
-		BUG_ON(!k->needs_whiteout);
+	dst = unwritten_whiteouts_start(c, b);
-		b->uncompacted_whiteout_u64s +=
+	memcpy_u64s(dst, k, u64s);
-			bkeyp_key_u64s(&b->format, k);
+	dst->u64s = u64s;
-	}
+	dst->type = KEY_TYPE_deleted;
 }
 
 /*

libbcachefs/btree_update_leaf.c

@@ -104,38 +104,43 @@ bool bch2_btree_bset_insert_key(struct btree_iter *iter,
 			return true;
 		}
 
-		insert->k.needs_whiteout = k->needs_whiteout;
-
 		btree_account_key_drop(b, k);
 
+		if (bkey_whiteout(&insert->k)) {
+			unsigned clobber_u64s = k->u64s, new_u64s = k->u64s;
+
+			k->type = KEY_TYPE_deleted;
+
+			if (k->needs_whiteout) {
+				push_whiteout(iter->trans->c, b, k);
+				k->needs_whiteout = false;
+			}
+
+			if (k >= btree_bset_last(b)->start) {
+				bch2_bset_delete(b, k, clobber_u64s);
+				new_u64s = 0;
+			}
+
+			bch2_btree_node_iter_fix(iter, b, node_iter, k,
+						 clobber_u64s, new_u64s);
+			return true;
+
+		}
+
 		if (k >= btree_bset_last(b)->start) {
 			clobber_u64s = k->u64s;
-
-			/*
-			 * If we're deleting, and the key we're deleting doesn't
-			 * need a whiteout (it wasn't overwriting a key that had
-			 * been written to disk) - just delete it:
-			 */
-			if (bkey_whiteout(&insert->k) && !k->needs_whiteout) {
-				bch2_bset_delete(b, k, clobber_u64s);
-				bch2_btree_node_iter_fix(iter, b, node_iter,
-							 k, clobber_u64s, 0);
-				return true;
-			}
-
 			goto overwrite;
 		}
 
+		insert->k.needs_whiteout = k->needs_whiteout;
+		k->needs_whiteout = false;
 		k->type = KEY_TYPE_deleted;
+		/*
+		 * XXX: we should be able to do this without two calls to
+		 * bch2_btree_node_iter_fix:
+		 */
 		bch2_btree_node_iter_fix(iter, b, node_iter, k,
 					 k->u64s, k->u64s);
-
-		if (bkey_whiteout(&insert->k)) {
-			reserve_whiteout(b, k);
-			return true;
-		} else {
-			k->needs_whiteout = false;
-		}
 	} else {
 		/*
 		 * Deleting, but the key to delete wasn't found - nothing to do:
@@ -863,9 +868,6 @@ retry:
 	       bkey_cmp(iter->pos, end) < 0) {
 		struct bkey_i delete;
 
-		bch2_trans_unlink_iters(trans);
-		trans->iters_touched &= trans->iters_live;
-
 		bkey_init(&delete.k);
 
 		/*

libbcachefs/ec.c

@@ -807,8 +807,7 @@ static int ec_stripe_update_ptrs(struct bch_fs *c,
 
 		dev = s->key.v.ptrs[idx].dev;
 
-		bkey_on_stack_realloc(&sk, c, k.k->u64s);
+		bkey_on_stack_reassemble(&sk, c, k);
-		bkey_reassemble(sk.k, k);
 		e = bkey_i_to_s_extent(sk.k);
 
 		extent_for_each_ptr(e, ptr) {

libbcachefs/extent_update.c

@@ -171,49 +171,51 @@ bch2_extent_can_insert(struct btree_trans *trans,
 {
 	struct btree_iter_level *l = &insert->iter->l[0];
 	struct btree_node_iter node_iter = l->iter;
-	enum bch_extent_overlap overlap;
 	struct bkey_packed *_k;
 	struct bkey unpacked;
-	struct bkey_s_c k;
 	int sectors;
 
-	/*
+	while ((_k = bch2_btree_node_iter_peek_filter(&node_iter, l->b,
-	 * We avoid creating whiteouts whenever possible when deleting, but
+						      KEY_TYPE_discard))) {
-	 * those optimizations mean we may potentially insert two whiteouts
+		struct bkey_s_c k = bkey_disassemble(l->b, _k, &unpacked);
-	 * instead of one (when we overlap with the front of one extent and the
+		enum bch_extent_overlap overlap =
-	 * back of another):
+			bch2_extent_overlap(&insert->k->k, k.k);
-	 */
-	if (bkey_whiteout(&insert->k->k))
-		*u64s += BKEY_U64s;
 
-	_k = bch2_btree_node_iter_peek_filter(&node_iter, l->b,
+		if (bkey_cmp(bkey_start_pos(k.k), insert->k->k.p) >= 0)
-					      KEY_TYPE_discard);
-	if (!_k)
-		return BTREE_INSERT_OK;
-
-	k = bkey_disassemble(l->b, _k, &unpacked);
-
-	overlap = bch2_extent_overlap(&insert->k->k, k.k);
-
-	/* account for having to split existing extent: */
-	if (overlap == BCH_EXTENT_OVERLAP_MIDDLE)
-		*u64s += _k->u64s;
-
-	if (overlap == BCH_EXTENT_OVERLAP_MIDDLE &&
-	    (sectors = bch2_bkey_sectors_compressed(k))) {
-		int flags = trans->flags & BTREE_INSERT_NOFAIL
-			? BCH_DISK_RESERVATION_NOFAIL : 0;
-
-		switch (bch2_disk_reservation_add(trans->c,
-				trans->disk_res,
-				sectors, flags)) {
-		case 0:
 			break;
-		case -ENOSPC:
+
-			return BTREE_INSERT_ENOSPC;
+		overlap = bch2_extent_overlap(&insert->k->k, k.k);
-		default:
+
-			BUG();
+		if (bkey_written(l->b, _k) &&
+		    overlap != BCH_EXTENT_OVERLAP_ALL)
+			*u64s += _k->u64s;
+
+		/* account for having to split existing extent: */
+		if (overlap == BCH_EXTENT_OVERLAP_MIDDLE)
+			*u64s += _k->u64s;
+
+		if (overlap == BCH_EXTENT_OVERLAP_MIDDLE &&
+		    (sectors = bch2_bkey_sectors_compressed(k))) {
+			int flags = trans->flags & BTREE_INSERT_NOFAIL
+				? BCH_DISK_RESERVATION_NOFAIL : 0;
+
+			switch (bch2_disk_reservation_add(trans->c,
+					trans->disk_res,
+					sectors, flags)) {
+			case 0:
+				break;
+			case -ENOSPC:
+				return BTREE_INSERT_ENOSPC;
+			default:
+				BUG();
+			}
 		}
+
+		if (overlap == BCH_EXTENT_OVERLAP_FRONT ||
+		    overlap == BCH_EXTENT_OVERLAP_MIDDLE)
+			break;
+
+		bch2_btree_node_iter_advance(&node_iter, l->b);
 	}
 
 	return BTREE_INSERT_OK;
@@ -284,6 +286,30 @@ static void extent_bset_insert(struct bch_fs *c, struct btree_iter *iter,
 	bch2_btree_node_iter_fix(iter, l->b, &l->iter, k, 0, k->u64s);
 }
 
+static void
+extent_drop(struct bch_fs *c, struct btree_iter *iter,
+	    struct bkey_packed *_k, struct bkey_s k)
+{
+	struct btree_iter_level *l = &iter->l[0];
+
+	if (!bkey_whiteout(k.k))
+		btree_account_key_drop(l->b, _k);
+
+	k.k->size = 0;
+	k.k->type = KEY_TYPE_deleted;
+	k.k->needs_whiteout = false;
+
+	if (_k >= btree_bset_last(l->b)->start) {
+		unsigned u64s = _k->u64s;
+
+		bch2_bset_delete(l->b, _k, _k->u64s);
+		bch2_btree_node_iter_fix(iter, l->b, &l->iter, _k, u64s, 0);
+	} else {
+		extent_save(l->b, _k, k.k);
+		bch2_btree_iter_fix_key_modified(iter, l->b, _k);
+	}
+}
+
 static void
 extent_squash(struct bch_fs *c, struct btree_iter *iter,
 	      struct bkey_i *insert,
@@ -291,96 +317,76 @@ extent_squash(struct bch_fs *c, struct btree_iter *iter,
 	      enum bch_extent_overlap overlap)
 {
 	struct btree_iter_level *l = &iter->l[0];
-	int u64s_delta;
+	struct bkey_on_stack tmp, split;
+
+	bkey_on_stack_init(&tmp);
+	bkey_on_stack_init(&split);
 
 	switch (overlap) {
 	case BCH_EXTENT_OVERLAP_FRONT:
-		/* insert overlaps with start of k: */
+		if (bkey_written(l->b, _k)) {
-		u64s_delta = bch2_cut_front_s(insert->k.p, k);
+			bkey_on_stack_reassemble(&tmp, c, k.s_c);
-		btree_keys_account_val_delta(l->b, _k, u64s_delta);
+			bch2_cut_front(insert->k.p, tmp.k);
 
-		EBUG_ON(bkey_deleted(k.k));
+			extent_drop(c, iter, _k, k);
-		extent_save(l->b, _k, k.k);
+			extent_bset_insert(c, iter, tmp.k);
-		bch2_btree_iter_fix_key_modified(iter, l->b, _k);
-		break;
-
-	case BCH_EXTENT_OVERLAP_BACK:
-		/* insert overlaps with end of k: */
-		u64s_delta = bch2_cut_back_s(bkey_start_pos(&insert->k), k);
-		btree_keys_account_val_delta(l->b, _k, u64s_delta);
-
-		EBUG_ON(bkey_deleted(k.k));
-		extent_save(l->b, _k, k.k);
-
-		/*
-		 * As the auxiliary tree is indexed by the end of the
-		 * key and we've just changed the end, update the
-		 * auxiliary tree.
-		 */
-		bch2_bset_fix_invalidated_key(l->b, _k);
-		bch2_btree_node_iter_fix(iter, l->b, &l->iter,
-					 _k, _k->u64s, _k->u64s);
-		break;
-
-	case BCH_EXTENT_OVERLAP_ALL: {
-		/* The insert key completely covers k, invalidate k */
-		if (!bkey_whiteout(k.k))
-			btree_account_key_drop(l->b, _k);
-
-		k.k->size = 0;
-		k.k->type = KEY_TYPE_deleted;
-
-		if (_k >= btree_bset_last(l->b)->start) {
-			unsigned u64s = _k->u64s;
-
-			bch2_bset_delete(l->b, _k, _k->u64s);
-			bch2_btree_node_iter_fix(iter, l->b, &l->iter,
-						 _k, u64s, 0);
 		} else {
+			btree_keys_account_val_delta(l->b, _k,
+				bch2_cut_front_s(insert->k.p, k));
+
+			extent_save(l->b, _k, k.k);
+			/*
+			 * No need to call bset_fix_invalidated_key, start of
+			 * extent changed but extents are indexed by where they
+			 * end
+			 */
+			bch2_btree_iter_fix_key_modified(iter, l->b, _k);
+		}
+		break;
+	case BCH_EXTENT_OVERLAP_BACK:
+		if (bkey_written(l->b, _k)) {
+			bkey_on_stack_reassemble(&tmp, c, k.s_c);
+			bch2_cut_back(bkey_start_pos(&insert->k), tmp.k);
+
+			extent_drop(c, iter, _k, k);
+			extent_bset_insert(c, iter, tmp.k);
+		} else {
+			btree_keys_account_val_delta(l->b, _k,
+				bch2_cut_back_s(bkey_start_pos(&insert->k), k));
+			extent_save(l->b, _k, k.k);
+
+			bch2_bset_fix_invalidated_key(l->b, _k);
+			bch2_btree_node_iter_fix(iter, l->b, &l->iter,
+						 _k, _k->u64s, _k->u64s);
+		}
+		break;
+	case BCH_EXTENT_OVERLAP_ALL:
+		extent_drop(c, iter, _k, k);
+		break;
+	case BCH_EXTENT_OVERLAP_MIDDLE:
+		bkey_on_stack_reassemble(&split, c, k.s_c);
+		bch2_cut_back(bkey_start_pos(&insert->k), split.k);
+
+		if (bkey_written(l->b, _k)) {
+			bkey_on_stack_reassemble(&tmp, c, k.s_c);
+			bch2_cut_front(insert->k.p, tmp.k);
+
+			extent_drop(c, iter, _k, k);
+			extent_bset_insert(c, iter, tmp.k);
+		} else {
+			btree_keys_account_val_delta(l->b, _k,
+				bch2_cut_front_s(insert->k.p, k));
+
 			extent_save(l->b, _k, k.k);
 			bch2_btree_iter_fix_key_modified(iter, l->b, _k);
 		}
 
-		break;
-	}
-	case BCH_EXTENT_OVERLAP_MIDDLE: {
-		struct bkey_on_stack split;
-
-		bkey_on_stack_init(&split);
-		bkey_on_stack_realloc(&split, c, k.k->u64s);
-
-		/*
-		 * The insert key falls 'in the middle' of k
-		 * The insert key splits k in 3:
-		 * - start only in k, preserve
-		 * - middle common section, invalidate in k
-		 * - end only in k, preserve
-		 *
-		 * We update the old key to preserve the start,
-		 * insert will be the new common section,
-		 * we manually insert the end that we are preserving.
-		 *
-		 * modify k _before_ doing the insert (which will move
-		 * what k points to)
-		 */
-		bkey_reassemble(split.k, k.s_c);
-		split.k->k.needs_whiteout |= bkey_written(l->b, _k);
-
-		bch2_cut_back(bkey_start_pos(&insert->k), split.k);
-		BUG_ON(bkey_deleted(&split.k->k));
-
-		u64s_delta = bch2_cut_front_s(insert->k.p, k);
-		btree_keys_account_val_delta(l->b, _k, u64s_delta);
-
-		BUG_ON(bkey_deleted(k.k));
-		extent_save(l->b, _k, k.k);
-		bch2_btree_iter_fix_key_modified(iter, l->b, _k);
-
 		extent_bset_insert(c, iter, split.k);
-		bkey_on_stack_exit(&split, c);
 		break;
 	}
-	}
+
+	bkey_on_stack_exit(&split, c);
+	bkey_on_stack_exit(&tmp, c);
 }
 
 /**
@@ -430,10 +436,7 @@ void bch2_insert_fixup_extent(struct btree_trans *trans,
 	struct bkey_i *insert	= insert_entry->k;
 	struct btree_iter_level *l = &iter->l[0];
 	struct btree_node_iter node_iter = l->iter;
-	bool deleting		= bkey_whiteout(&insert->k);
+	bool do_update		= !bkey_whiteout(&insert->k);
-	bool update_journal	= !deleting;
-	bool update_btree	= !deleting;
-	struct bkey_i whiteout	= *insert;
 	struct bkey_packed *_k;
 	struct bkey unpacked;
 
@@ -444,7 +447,6 @@ void bch2_insert_fixup_extent(struct btree_trans *trans,
 	while ((_k = bch2_btree_node_iter_peek_filter(&l->iter, l->b,
 						      KEY_TYPE_discard))) {
 		struct bkey_s k = __bkey_disassemble(l->b, _k, &unpacked);
-		struct bpos cur_end = bpos_min(insert->k.p, k.k->p);
 		enum bch_extent_overlap overlap =
 			bch2_extent_overlap(&insert->k, k.k);
 
@@ -452,52 +454,18 @@ void bch2_insert_fixup_extent(struct btree_trans *trans,
 			break;
 
 		if (!bkey_whiteout(k.k))
-			update_journal = true;
+			do_update = true;
+
+		if (!do_update) {
+			struct bpos cur_end = bpos_min(insert->k.p, k.k->p);
 
-		if (!update_journal) {
 			bch2_cut_front(cur_end, insert);
-			bch2_cut_front(cur_end, &whiteout);
 			bch2_btree_iter_set_pos_same_leaf(iter, cur_end);
-			goto next;
+		} else {
+			insert->k.needs_whiteout |= k.k->needs_whiteout;
+			extent_squash(c, iter, insert, _k, k, overlap);
 		}
 
-		/*
-		 * When deleting, if possible just do it by switching the type
-		 * of the key we're deleting, instead of creating and inserting
-		 * a new whiteout:
-		 */
-		if (deleting &&
-		    !update_btree &&
-		    !bkey_cmp(insert->k.p, k.k->p) &&
-		    !bkey_cmp(bkey_start_pos(&insert->k), bkey_start_pos(k.k))) {
-			if (!bkey_whiteout(k.k)) {
-				btree_account_key_drop(l->b, _k);
-				_k->type = KEY_TYPE_discard;
-				reserve_whiteout(l->b, _k);
-				bch2_btree_iter_fix_key_modified(iter,
-								 l->b, _k);
-			}
-			break;
-		}
-
-		if (k.k->needs_whiteout || bkey_written(l->b, _k)) {
-			insert->k.needs_whiteout = true;
-			update_btree = true;
-		}
-
-		if (update_btree &&
-		    overlap == BCH_EXTENT_OVERLAP_ALL &&
-		    bkey_whiteout(k.k) &&
-		    k.k->needs_whiteout) {
-			unreserve_whiteout(l->b, _k);
-			_k->needs_whiteout = false;
-		}
-
-		extent_squash(c, iter, insert, _k, k, overlap);
-
-		if (!update_btree)
-			bch2_cut_front(cur_end, insert);
-next:
 		node_iter = l->iter;
 
 		if (overlap == BCH_EXTENT_OVERLAP_FRONT ||
@@ -508,24 +476,12 @@ next:
 	l->iter = node_iter;
 	bch2_btree_iter_set_pos_same_leaf(iter, insert->k.p);
 
-	if (update_btree) {
+	if (do_update) {
-		if (deleting)
+		if (insert->k.type == KEY_TYPE_deleted)
 			insert->k.type = KEY_TYPE_discard;
 
-		EBUG_ON(bkey_deleted(&insert->k) || !insert->k.size);
-
 		extent_bset_insert(c, iter, insert);
-	}
+		bch2_btree_journal_key(trans, iter, insert);
-
-	if (update_journal) {
-		struct bkey_i *k = !deleting ? insert : &whiteout;
-
-		if (deleting)
-			k->k.type = KEY_TYPE_discard;
-
-		EBUG_ON(bkey_deleted(&k->k) || !k->k.size);
-
-		bch2_btree_journal_key(trans, iter, k);
 	}
 
 	bch2_cut_front(insert->k.p, insert);

libbcachefs/fs-io.c

@@ -602,7 +602,7 @@ int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
 	EBUG_ON(!PageLocked(page));
 	EBUG_ON(!PageLocked(newpage));
 
-	ret = migrate_page_move_mapping(mapping, newpage, page, mode, 0);
+	ret = migrate_page_move_mapping(mapping, newpage, page, 0);
 	if (ret != MIGRATEPAGE_SUCCESS)
 		return ret;
 
@@ -837,8 +837,7 @@ retry:
 		if (ret)
 			break;
 
-		bkey_on_stack_realloc(&sk, c, k.k->u64s);
+		bkey_on_stack_reassemble(&sk, c, k);
-		bkey_reassemble(sk.k, k);
 		k = bkey_i_to_s_c(sk.k);
 
 		offset_into_extent = iter->pos.offset -
@@ -1239,7 +1238,7 @@ do_io:
 
 		if (w->io &&
 		    (w->io->op.res.nr_replicas != nr_replicas_this_write ||
-		     bio_full(&w->io->op.wbio.bio) ||
+		     bio_full(&w->io->op.wbio.bio, PAGE_SIZE) ||
 		     w->io->op.wbio.bio.bi_iter.bi_size >= (256U << 20) ||
 		     bio_end_sector(&w->io->op.wbio.bio) != sector))
 			bch2_writepage_do_io(w);
@@ -2504,8 +2503,7 @@ static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
 		    bkey_cmp(k.k->p, POS(inode->v.i_ino, offset >> 9)) <= 0)
 			break;
 reassemble:
-		bkey_on_stack_realloc(&copy, c, k.k->u64s);
+		bkey_on_stack_reassemble(&copy, c, k);
-		bkey_reassemble(copy.k, k);
 
 		if (insert &&
 		    bkey_cmp(bkey_start_pos(k.k), move_pos) < 0) {

libbcachefs/fs.c

@@ -971,7 +971,10 @@ static const struct file_operations bch_file_operations = {
 	.open		= generic_file_open,
 	.fsync		= bch2_fsync,
 	.splice_read	= generic_file_splice_read,
+	/*
+	 * Broken, on v5.3:
 	.splice_write	= iter_file_splice_write,
+	*/
 	.fallocate	= bch2_fallocate_dispatch,
 	.unlocked_ioctl = bch2_fs_file_ioctl,
 #ifdef CONFIG_COMPAT

libbcachefs/io.c

@@ -1139,6 +1139,8 @@ static void bch2_write_data_inline(struct bch_write_op *op, unsigned data_len)
 	unsigned sectors;
 	int ret;
 
+	bch2_check_set_feature(op->c, BCH_FEATURE_INLINE_DATA);
+
 	ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys,
 				   ARRAY_SIZE(op->inline_keys),
 				   BKEY_U64s + DIV_ROUND_UP(data_len, 8));
@@ -1220,7 +1222,8 @@ void bch2_write(struct closure *cl)
 	data_len = min_t(u64, bio->bi_iter.bi_size,
 			 op->new_i_size - (op->pos.offset << 9));
 
-	if (data_len <= min(block_bytes(c) / 2, 1024U)) {
+	if (c->opts.inline_data &&
+	    data_len <= min(block_bytes(c) / 2, 1024U)) {
 		bch2_write_data_inline(op, data_len);
 		return;
 	}
@@ -1536,8 +1539,7 @@ retry:
 	if (bkey_err(k))
 		goto err;
 
-	bkey_on_stack_realloc(&sk, c, k.k->u64s);
+	bkey_on_stack_reassemble(&sk, c, k);
-	bkey_reassemble(sk.k, k);
 	k = bkey_i_to_s_c(sk.k);
 	bch2_trans_unlock(&trans);
 
@@ -1588,8 +1590,7 @@ retry:
 			   BTREE_ITER_SLOTS, k, ret) {
 		unsigned bytes, sectors, offset_into_extent;
 
-		bkey_on_stack_realloc(&sk, c, k.k->u64s);
+		bkey_on_stack_reassemble(&sk, c, k);
-		bkey_reassemble(sk.k, k);
 		k = bkey_i_to_s_c(sk.k);
 
 		offset_into_extent = iter->pos.offset -
@@ -1712,8 +1713,7 @@ retry:
 	if (IS_ERR_OR_NULL(k.k))
 		goto out;
 
-	bkey_on_stack_realloc(&new, c, k.k->u64s);
+	bkey_on_stack_reassemble(&new, c, k);
-	bkey_reassemble(new.k, k);
 	k = bkey_i_to_s_c(new.k);
 
 	if (bversion_cmp(k.k->version, rbio->version) ||
@@ -2220,8 +2220,7 @@ retry:
 			bkey_start_offset(k.k);
 		sectors = k.k->size - offset_into_extent;
 
-		bkey_on_stack_realloc(&sk, c, k.k->u64s);
+		bkey_on_stack_reassemble(&sk, c, k);
-		bkey_reassemble(sk.k, k);
 		k = bkey_i_to_s_c(sk.k);
 
 		ret = bch2_read_indirect_extent(&trans,

libbcachefs/migrate.c

@@ -60,8 +60,7 @@ static int __bch2_dev_usrdata_drop(struct bch_fs *c, unsigned dev_idx, int flags
 			continue;
 		}
 
-		bkey_on_stack_realloc(&sk, c, k.k->u64s);
+		bkey_on_stack_reassemble(&sk, c, k);
-		bkey_reassemble(sk.k, k);
 
 		ret = drop_dev_ptrs(c, bkey_i_to_s(sk.k),
 				    dev_idx, flags, false);

libbcachefs/move.c

@@ -579,8 +579,7 @@ peek:
 		}
 
 		/* unlock before doing IO: */
-		bkey_on_stack_realloc(&sk, c, k.k->u64s);
+		bkey_on_stack_reassemble(&sk, c, k);
-		bkey_reassemble(sk.k, k);
 		k = bkey_i_to_s_c(sk.k);
 		bch2_trans_unlock(&trans);
 

libbcachefs/movinggc.c

@@ -107,10 +107,10 @@ static bool have_copygc_reserve(struct bch_dev *ca)
 {
 	bool ret;
 
-	spin_lock(&ca->freelist_lock);
+	spin_lock(&ca->fs->freelist_lock);
 	ret = fifo_full(&ca->free[RESERVE_MOVINGGC]) ||
 		ca->allocator_state != ALLOCATOR_RUNNING;
-	spin_unlock(&ca->freelist_lock);
+	spin_unlock(&ca->fs->freelist_lock);
 
 	return ret;
 }

libbcachefs/opts.c

@@ -299,15 +299,8 @@ int bch2_opt_check_may_set(struct bch_fs *c, int id, u64 v)
 		ret = bch2_check_set_has_compressed_data(c, v);
 		break;
 	case Opt_erasure_code:
-		if (v &&
+		if (v)
-		    !(c->sb.features & (1ULL << BCH_FEATURE_EC))) {
+			bch2_check_set_feature(c, BCH_FEATURE_EC);
-			mutex_lock(&c->sb_lock);
-			c->disk_sb.sb->features[0] |=
-				cpu_to_le64(1ULL << BCH_FEATURE_EC);
-
-			bch2_write_super(c);
-			mutex_unlock(&c->sb_lock);
-		}
 		break;
 	}
 

libbcachefs/opts.h

@@ -181,6 +181,11 @@ enum opt_type {
 	  OPT_BOOL(),							\
 	  BCH_SB_128_BIT_MACS,		false,				\
 	  NULL,		"Store full 128 bits of cryptographic MACs, instead of 80")\
+	x(inline_data,			u8,				\
+	  OPT_MOUNT|OPT_RUNTIME,					\
+	  OPT_BOOL(),							\
+	  NO_SB_OPT,			false,				\
+	  NULL,		"Enable inline data extents")			\
 	x(acl,				u8,				\
 	  OPT_FORMAT|OPT_MOUNT,						\
 	  OPT_BOOL(),							\

libbcachefs/recovery.c

@@ -913,12 +913,6 @@ int bch2_fs_recovery(struct bch_fs *c)
 		write_sb = true;
 	}
 
-	if (!(c->sb.features & (1ULL << BCH_FEATURE_INLINE_DATA))) {
-		c->disk_sb.sb->features[0] |=
-			cpu_to_le64(1ULL << BCH_FEATURE_INLINE_DATA);
-		write_sb = true;
-	}
-
 	if (!test_bit(BCH_FS_ERROR, &c->flags)) {
 		c->disk_sb.sb->compat[0] |= 1ULL << BCH_COMPAT_FEAT_ALLOC_INFO;
 		write_sb = true;

libbcachefs/reflink.c

@@ -171,16 +171,7 @@ s64 bch2_remap_range(struct bch_fs *c,
 	if (!percpu_ref_tryget(&c->writes))
 		return -EROFS;
 
-	if (!(c->sb.features & (1ULL << BCH_FEATURE_REFLINK))) {
+	bch2_check_set_feature(c, BCH_FEATURE_REFLINK);
-		mutex_lock(&c->sb_lock);
-		if (!(c->sb.features & (1ULL << BCH_FEATURE_REFLINK))) {
-			c->disk_sb.sb->features[0] |=
-				cpu_to_le64(1ULL << BCH_FEATURE_REFLINK);
-
-			bch2_write_super(c);
-		}
-		mutex_unlock(&c->sb_lock);
-	}
 
 	dst_end.offset += remap_sectors;
 	src_end.offset += remap_sectors;
@@ -225,8 +216,7 @@ s64 bch2_remap_range(struct bch_fs *c,
 			break;
 
 		if (src_k.k->type == KEY_TYPE_extent) {
-			bkey_on_stack_realloc(&new_src, c, src_k.k->u64s);
+			bkey_on_stack_reassemble(&new_src, c, src_k);
-			bkey_reassemble(new_src.k, src_k);
 			src_k = bkey_i_to_s_c(new_src.k);
 
 			bch2_cut_front(src_iter->pos,	new_src.k);

libbcachefs/super-io.c

@@ -795,6 +795,17 @@ out:
 	return ret;
 }
 
+void __bch2_check_set_feature(struct bch_fs *c, unsigned feat)
+{
+	mutex_lock(&c->sb_lock);
+	if (!(c->sb.features & (1ULL << feat))) {
+		c->disk_sb.sb->features[0] |= cpu_to_le64(1ULL << feat);
+
+		bch2_write_super(c);
+	}
+	mutex_unlock(&c->sb_lock);
+}
+
 /* BCH_SB_FIELD_journal: */
 
 static int u64_cmp(const void *_l, const void *_r)

libbcachefs/super-io.h

@@ -43,26 +43,6 @@ struct bch_sb_field_ops {
 				   struct bch_sb_field *);
 };
 
-static inline bool bch2_sb_test_feature(struct bch_sb *sb,
-					enum bch_sb_features f)
-{
-	unsigned w = f / 64;
-	unsigned b = f % 64;
-
-	return le64_to_cpu(sb->features[w]) & (1ULL << b);
-}
-
-static inline void bch2_sb_set_feature(struct bch_sb *sb,
-				       enum bch_sb_features f)
-{
-	if (!bch2_sb_test_feature(sb, f)) {
-		unsigned w = f / 64;
-		unsigned b = f % 64;
-
-		le64_add_cpu(&sb->features[w], 1ULL << b);
-	}
-}
-
 static inline __le64 bch2_sb_magic(struct bch_fs *c)
 {
 	__le64 ret;
@@ -90,6 +70,13 @@ const char *bch2_sb_validate(struct bch_sb_handle *);
 
 int bch2_read_super(const char *, struct bch_opts *, struct bch_sb_handle *);
 int bch2_write_super(struct bch_fs *);
+void __bch2_check_set_feature(struct bch_fs *, unsigned);
+
+static inline void bch2_check_set_feature(struct bch_fs *c, unsigned feat)
+{
+	if (!(c->sb.features & (1ULL << feat)))
+		__bch2_check_set_feature(c, feat);
+}
 
 /* BCH_SB_FIELD_journal: */
 

libbcachefs/super.c

@@ -735,9 +735,9 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	if (bch2_fs_init_fault("fs_alloc"))
 		goto err;
 
-	iter_size = sizeof(struct btree_node_iter_large) +
+	iter_size = sizeof(struct sort_iter) +
 		(btree_blocks(c) + 1) * 2 *
-		sizeof(struct btree_node_iter_set);
+		sizeof(struct sort_iter_set);
 
 	if (!(c->wq = alloc_workqueue("bcachefs",
 				WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE, 1)) ||
@@ -1092,7 +1092,6 @@ static struct bch_dev *__bch2_dev_alloc(struct bch_fs *c,
 
 	writepoint_init(&ca->copygc_write_point, BCH_DATA_USER);
 
-	spin_lock_init(&ca->freelist_lock);
 	bch2_dev_copygc_init(ca);
 
 	INIT_WORK(&ca->io_error_work, bch2_io_error_work);

libbcachefs/sysfs.c

@@ -775,7 +775,7 @@ static ssize_t show_reserve_stats(struct bch_dev *ca, char *buf)
 	struct printbuf out = _PBUF(buf, PAGE_SIZE);
 	enum alloc_reserve i;
 
-	spin_lock(&ca->freelist_lock);
+	spin_lock(&ca->fs->freelist_lock);
 
 	pr_buf(&out, "free_inc:\t%zu\t%zu\n",
 	       fifo_used(&ca->free_inc),
@@ -786,7 +786,7 @@ static ssize_t show_reserve_stats(struct bch_dev *ca, char *buf)
 		       fifo_used(&ca->free[i]),
 		       ca->free[i].size);
 
-	spin_unlock(&ca->freelist_lock);
+	spin_unlock(&ca->fs->freelist_lock);
 
 	return out.pos - buf;
 }