bcachefs: Allocation code refactoring

bch2_alloc_sectors_start() was a nightmare to work with - it's got some
tricky stuff to do, since it wants to use the buckets the writepoint
already has, unless they're not in the target it wants to write to,
unless it can't allocate from any other devices in which case it will
use those buckets if it has to - et cetera.

This restructures the code to start with a new empty list of open
buckets we're going to use for the new allocation, pulling buckets from
the write point's list as we decide that we really are going to use
them - making the code somewhat more functional and drastically easier
to understand.

Also fixes a bug where we could end up waiting on c->freelist_wait
(because allocating from one device failed) but return success from
bch2_bucket_alloc(), because allocating from a different device
succeeded.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

fs/bcachefs/alloc_background.c

@@ -1101,7 +1101,7 @@ void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
 		struct btree_alloc *a =
 			&c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 
-		bch2_open_bucket_put_refs(c, &a->ob.nr, a->ob.refs);
+		bch2_open_buckets_put(c, &a->ob);
 	}
 	mutex_unlock(&c->btree_reserve_cache_lock);
 

fs/bcachefs/alloc_foreground.c

@@ -70,10 +70,9 @@
 #include <linux/rcupdate.h>
 
 enum bucket_alloc_ret {
-	ALLOC_SUCCESS		= 0,
-	OPEN_BUCKETS_EMPTY	= -1,
-	FREELIST_EMPTY		= -2,	/* Allocator thread not keeping up */
-	NO_DEVICES		= -3,	/* -EROFS */
+	ALLOC_SUCCESS,
+	OPEN_BUCKETS_EMPTY,
+	FREELIST_EMPTY,		/* Allocator thread not keeping up */
 };
 
 /*
@@ -129,6 +128,43 @@ static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
 	return ob;
 }
 
+static void open_bucket_free_unused(struct bch_fs *c,
+				    struct write_point *wp,
+				    struct open_bucket *ob)
+{
+	struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
+
+	BUG_ON(ca->open_buckets_partial_nr >=
+	       ARRAY_SIZE(ca->open_buckets_partial));
+
+	if (wp->type == BCH_DATA_USER) {
+		spin_lock(&c->freelist_lock);
+		ob->on_partial_list = true;
+		ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
+			ob - c->open_buckets;
+		spin_unlock(&c->freelist_lock);
+
+		closure_wake_up(&c->open_buckets_wait);
+		closure_wake_up(&c->freelist_wait);
+	} else {
+		bch2_open_bucket_put(c, ob);
+	}
+}
+
+static void verify_not_stale(struct bch_fs *c, const struct open_buckets *obs)
+{
+#ifdef CONFIG_BCACHEFS_DEBUG
+	struct open_bucket *ob;
+	unsigned i;
+
+	open_bucket_for_each(c, obs, ob, i) {
+		struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
+
+		BUG_ON(ptr_stale(ca, &ob->ptr));
+	}
+#endif
+}
+
 /* _only_ for allocating the journal on a new device: */
 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
 {
@@ -164,7 +200,7 @@ static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
  *
  * Returns index of bucket on success, 0 on failure
  * */
-int bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
+struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 				      enum alloc_reserve reserve,
 				      bool may_alloc_partial,
 				      struct closure *cl)
@@ -177,10 +213,11 @@ int bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 
 	if (may_alloc_partial &&
 	    ca->open_buckets_partial_nr) {
-		int ret = ca->open_buckets_partial[--ca->open_buckets_partial_nr];
-		c->open_buckets[ret].on_partial_list = false;
+		ob = c->open_buckets +
+			ca->open_buckets_partial[--ca->open_buckets_partial_nr];
+		ob->on_partial_list = false;
 		spin_unlock(&c->freelist_lock);
-		return ret;
+		return ob;
 	}
 
 	if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
@@ -188,7 +225,7 @@ int bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 			closure_wait(&c->open_buckets_wait, cl);
 		spin_unlock(&c->freelist_lock);
 		trace_open_bucket_alloc_fail(ca, reserve);
-		return OPEN_BUCKETS_EMPTY;
+		return ERR_PTR(-OPEN_BUCKETS_EMPTY);
 	}
 
 	if (likely(fifo_pop(&ca->free[RESERVE_NONE], bucket)))
@@ -219,7 +256,7 @@ int bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 	spin_unlock(&c->freelist_lock);
 
 	trace_bucket_alloc_fail(ca, reserve);
-	return FREELIST_EMPTY;
+	return ERR_PTR(-FREELIST_EMPTY);
 out:
 	verify_not_on_freelist(c, ca, bucket);
 
@@ -245,7 +282,7 @@ int bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 	bch2_wake_allocator(ca);
 
 	trace_bucket_alloc(ca, reserve);
-	return ob - c->open_buckets;
+	return ob;
 }
 
 static int __dev_alloc_cmp(struct write_point *wp,
@@ -292,155 +329,114 @@ void bch2_wp_rescale(struct bch_fs *c, struct bch_dev *ca,
 		*v = *v < scale ? 0 : *v - scale;
 }
 
-static enum bucket_alloc_ret bch2_bucket_alloc_set(struct bch_fs *c,
+static int bch2_bucket_alloc_set(struct bch_fs *c,
+				 struct open_buckets *ptrs,
 				 struct write_point *wp,
+				 struct bch_devs_mask *devs_may_alloc,
 				 unsigned nr_replicas,
+				 unsigned *nr_effective,
+				 bool *have_cache,
 				 enum alloc_reserve reserve,
-					struct bch_devs_mask *devs,
 				 struct closure *cl)
 {
-	enum bucket_alloc_ret ret = NO_DEVICES;
-	struct dev_alloc_list devs_sorted;
+	struct dev_alloc_list devs_sorted =
+		bch2_wp_alloc_list(c, wp, devs_may_alloc);
 	struct bch_dev *ca;
-	unsigned i, nr_ptrs_effective = 0;
-	bool have_cache_dev = false;
-
-	BUG_ON(nr_replicas > ARRAY_SIZE(wp->ptrs));
-
-	for (i = wp->first_ptr; i < wp->nr_ptrs; i++) {
-		ca = bch_dev_bkey_exists(c, wp->ptrs[i]->ptr.dev);
-
-		nr_ptrs_effective += ca->mi.durability;
-		have_cache_dev |= !ca->mi.durability;
-	}
-
-	if (nr_ptrs_effective >= nr_replicas)
-		return ALLOC_SUCCESS;
+	bool alloc_failure = false;
+	unsigned i;
 
-	devs_sorted = bch2_wp_alloc_list(c, wp, devs);
+	BUG_ON(*nr_effective >= nr_replicas);
 
 	for (i = 0; i < devs_sorted.nr; i++) {
-		int ob;
+		struct open_bucket *ob;
 
 		ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
 		if (!ca)
 			continue;
 
 		if (!ca->mi.durability &&
-		    (have_cache_dev ||
+		    (*have_cache ||
 		     wp->type != BCH_DATA_USER))
 			continue;
 
 		ob = bch2_bucket_alloc(c, ca, reserve,
 				       wp->type == BCH_DATA_USER, cl);
-		if (ob < 0) {
-			ret = ob;
+		if (IS_ERR(ob)) {
+			enum bucket_alloc_ret ret = -PTR_ERR(ob);
+
+			WARN_ON(reserve == RESERVE_MOVINGGC &&
+				ret != OPEN_BUCKETS_EMPTY);
+
+			if (cl)
+				return -EAGAIN;
 			if (ret == OPEN_BUCKETS_EMPTY)
-				break;
+				return -ENOSPC;
+			alloc_failure = true;
 			continue;
 		}
 
-		BUG_ON(ob <= 0 || ob > U8_MAX);
-		BUG_ON(wp->nr_ptrs >= ARRAY_SIZE(wp->ptrs));
+		__clear_bit(ca->dev_idx, devs_may_alloc->d);
+		*nr_effective	+= ca->mi.durability;
+		*have_cache	|= !ca->mi.durability;
 
-		wp->ptrs[wp->nr_ptrs++] = c->open_buckets + ob;
+		ob_push(c, ptrs, ob);
 
 		bch2_wp_rescale(c, ca, wp);
 
-		nr_ptrs_effective += ca->mi.durability;
-		have_cache_dev |= !ca->mi.durability;
-
-		__clear_bit(ca->dev_idx, devs->d);
-
-		if (nr_ptrs_effective >= nr_replicas) {
-			ret = ALLOC_SUCCESS;
-			break;
-		}
-	}
-
-	EBUG_ON(reserve == RESERVE_MOVINGGC &&
-		ret != ALLOC_SUCCESS &&
-		ret != OPEN_BUCKETS_EMPTY);
-
-	switch (ret) {
-	case ALLOC_SUCCESS:
+		if (*nr_effective >= nr_replicas)
 			return 0;
-	case NO_DEVICES:
-		return -EROFS;
-	case FREELIST_EMPTY:
-	case OPEN_BUCKETS_EMPTY:
-		return cl ? -EAGAIN : -ENOSPC;
-	default:
-		BUG();
 	}
+
+	return alloc_failure ? -ENOSPC : -EROFS;
 }
 
 /* Sector allocator */
 
-static void bch2_writepoint_drop_ptr(struct bch_fs *c,
+static int get_buckets_from_writepoint(struct bch_fs *c,
+				       struct open_buckets *ptrs,
 				       struct write_point *wp,
-				     unsigned i)
-{
-	struct open_bucket *ob = wp->ptrs[i];
-	struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
-
-	BUG_ON(ca->open_buckets_partial_nr >=
-	       ARRAY_SIZE(ca->open_buckets_partial));
-
-	if (wp->type == BCH_DATA_USER) {
-		spin_lock(&c->freelist_lock);
-		ob->on_partial_list = true;
-		ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
-			ob - c->open_buckets;
-		spin_unlock(&c->freelist_lock);
-
-		closure_wake_up(&c->open_buckets_wait);
-		closure_wake_up(&c->freelist_wait);
-	} else {
-		bch2_open_bucket_put(c, ob);
-	}
-
-	array_remove_item(wp->ptrs, wp->nr_ptrs, i);
-
-	if (i < wp->first_ptr)
-		wp->first_ptr--;
-}
-
-void bch2_writepoint_drop_ptrs(struct bch_fs *c,
-			       struct write_point *wp,
-			       u16 target, bool in_target)
-{
-	int i;
-
-	for (i = wp->first_ptr - 1; i >= 0; --i)
-		if (bch2_dev_in_target(c, wp->ptrs[i]->ptr.dev,
-				       target) == in_target)
-			bch2_writepoint_drop_ptr(c, wp, i);
-}
-
-static void verify_not_stale(struct bch_fs *c, const struct write_point *wp)
+				       struct bch_devs_mask *devs_may_alloc,
+				       unsigned nr_replicas,
+				       unsigned *nr_effective,
+				       bool *have_cache)
 {
-#ifdef CONFIG_BCACHEFS_DEBUG
+	struct open_buckets ptrs_skip = { .nr = 0 };
 	struct open_bucket *ob;
 	unsigned i;
 
-	writepoint_for_each_ptr_all(wp, ob, i) {
+	open_bucket_for_each(c, &wp->ptrs, ob, i) {
 		struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 
-		BUG_ON(ptr_stale(ca, &ob->ptr));
+		if (*nr_effective < nr_replicas &&
+		    test_bit(ob->ptr.dev, devs_may_alloc->d) &&
+		    (ca->mi.durability ||
+		     (wp->type == BCH_DATA_USER && !*have_cache))) {
+			__clear_bit(ob->ptr.dev, devs_may_alloc->d);
+			*nr_effective	+= ca->mi.durability;
+			*have_cache	|= !ca->mi.durability;
+
+			ob_push(c, ptrs, ob);
+		} else {
+			ob_push(c, &ptrs_skip, ob);
 		}
-#endif
+	}
+	wp->ptrs = ptrs_skip;
+
+	return *nr_effective < nr_replicas ? -ENOSPC : 0;
 }
 
 static int open_bucket_add_buckets(struct bch_fs *c,
-				   u16 target,
+				   struct open_buckets *ptrs,
 				   struct write_point *wp,
 				   struct bch_devs_list *devs_have,
+				   u16 target,
 				   unsigned nr_replicas,
+				   unsigned *nr_effective,
+				   bool *have_cache,
 				   enum alloc_reserve reserve,
 				   struct closure *cl)
 {
-	struct bch_devs_mask devs = c->rw_devs[wp->type];
+	struct bch_devs_mask devs;
 	const struct bch_devs_mask *t;
 	struct open_bucket *ob;
 	unsigned i;
@@ -449,19 +445,38 @@ static int open_bucket_add_buckets(struct bch_fs *c,
 	percpu_down_read(&c->usage_lock);
 	rcu_read_lock();
 
+	devs = c->rw_devs[wp->type];
+
 	/* Don't allocate from devices we already have pointers to: */
 	for (i = 0; i < devs_have->nr; i++)
 		__clear_bit(devs_have->devs[i], devs.d);
 
-	writepoint_for_each_ptr_all(wp, ob, i)
+	open_bucket_for_each(c, ptrs, ob, i)
 		__clear_bit(ob->ptr.dev, devs.d);
 
 	t = bch2_target_to_mask(c, target);
 	if (t)
 		bitmap_and(devs.d, devs.d, t->d, BCH_SB_MEMBERS_MAX);
 
-	ret = bch2_bucket_alloc_set(c, wp, nr_replicas, reserve, &devs, cl);
+	ret = get_buckets_from_writepoint(c, ptrs, wp, &devs,
+				nr_replicas, nr_effective, have_cache);
+	if (!ret)
+		goto out;
 
+	/*
+	 * Try nonblocking first, so that if one device is full we'll try from
+	 * other devices:
+	 */
+	ret = bch2_bucket_alloc_set(c, ptrs, wp, &devs,
+				nr_replicas, nr_effective, have_cache,
+				reserve, NULL);
+	if (!ret || ret == -EROFS || !cl)
+		goto out;
+
+	ret = bch2_bucket_alloc_set(c, ptrs, wp, &devs,
+				nr_replicas, nr_effective, have_cache,
+				reserve, cl);
+out:
 	rcu_read_unlock();
 	percpu_up_read(&c->usage_lock);
 
@@ -471,13 +486,18 @@ static int open_bucket_add_buckets(struct bch_fs *c,
 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 			  struct write_point *wp)
 {
-	struct bch_devs_mask not_self;
-
-	bitmap_complement(not_self.d, ca->self.d, BCH_SB_MEMBERS_MAX);
+	struct open_buckets ptrs = { .nr = 0 };
+	struct open_bucket *ob;
+	unsigned i;
 
 	mutex_lock(&wp->lock);
-	wp->first_ptr = wp->nr_ptrs;
-	bch2_writepoint_drop_ptrs(c, wp, dev_to_target(ca->dev_idx), true);
+	open_bucket_for_each(c, &wp->ptrs, ob, i)
+		if (ob->ptr.dev == ca->dev_idx)
+			open_bucket_free_unused(c, wp, ob);
+		else
+			ob_push(c, &ptrs, ob);
+
+	wp->ptrs = ptrs;
 	mutex_unlock(&wp->lock);
 }
 
@@ -558,134 +578,64 @@ struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
 {
 	struct write_point *wp;
 	struct open_bucket *ob;
-	struct bch_dev *ca;
-	unsigned nr_ptrs_have, nr_ptrs_effective;
-	int ret, i, cache_idx = -1;
+	unsigned nr_effective = 0;
+	struct open_buckets ptrs = { .nr = 0 };
+	bool have_cache = false;
+	int ret = 0, i;
 
 	BUG_ON(!nr_replicas || !nr_replicas_required);
 
 	wp = writepoint_find(c, write_point.v);
 
-	wp->first_ptr = 0;
-
-	/* does writepoint have ptrs we can't use? */
-	writepoint_for_each_ptr(wp, ob, i)
-		if (bch2_dev_list_has_dev(*devs_have, ob->ptr.dev)) {
-			swap(wp->ptrs[i], wp->ptrs[wp->first_ptr]);
-			wp->first_ptr++;
-		}
-
-	nr_ptrs_have = wp->first_ptr;
-
-	/* does writepoint have ptrs we don't want to use? */
-	if (target)
-		writepoint_for_each_ptr(wp, ob, i)
-			if (!bch2_dev_in_target(c, ob->ptr.dev, target)) {
-				swap(wp->ptrs[i], wp->ptrs[wp->first_ptr]);
-				wp->first_ptr++;
-			}
-
-	if (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS) {
-		ret = open_bucket_add_buckets(c, target, wp, devs_have,
-					      nr_replicas, reserve, cl);
+	if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
+		ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have, target,
+					      nr_replicas, &nr_effective,
+					      &have_cache, reserve, cl);
 	} else {
-		ret = open_bucket_add_buckets(c, target, wp, devs_have,
-					      nr_replicas, reserve, NULL);
+		ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have, target,
+					      nr_replicas, &nr_effective,
+					      &have_cache, reserve, NULL);
 		if (!ret)
 			goto alloc_done;
 
-		wp->first_ptr = nr_ptrs_have;
-
-		ret = open_bucket_add_buckets(c, 0, wp, devs_have,
-					      nr_replicas, reserve, cl);
+		ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have, 0,
+					      nr_replicas, &nr_effective,
+					      &have_cache, reserve, cl);
 	}
-
-	if (ret && ret != -EROFS)
-		goto err;
 alloc_done:
-	/* check for more than one cache: */
-	for (i = wp->nr_ptrs - 1; i >= wp->first_ptr; --i) {
-		ca = bch_dev_bkey_exists(c, wp->ptrs[i]->ptr.dev);
-
-		if (ca->mi.durability)
-			continue;
-
-		/*
-		 * if we ended up with more than one cache device, prefer the
-		 * one in the target we want:
-		 */
-		if (cache_idx >= 0) {
-			if (!bch2_dev_in_target(c, wp->ptrs[i]->ptr.dev,
-						target)) {
-				bch2_writepoint_drop_ptr(c, wp, i);
-			} else {
-				bch2_writepoint_drop_ptr(c, wp, cache_idx);
-				cache_idx = i;
-			}
-		} else {
-			cache_idx = i;
-		}
-	}
-
-	/* we might have more effective replicas than required: */
-	nr_ptrs_effective = 0;
-	writepoint_for_each_ptr(wp, ob, i) {
-		ca = bch_dev_bkey_exists(c, ob->ptr.dev);
-		nr_ptrs_effective += ca->mi.durability;
-	}
+	BUG_ON(!ret && nr_effective < nr_replicas);
 
 	if (ret == -EROFS &&
-	    nr_ptrs_effective >= nr_replicas_required)
+	    nr_effective >= nr_replicas_required)
 		ret = 0;
 
 	if (ret)
 		goto err;
 
-	if (nr_ptrs_effective > nr_replicas) {
-		writepoint_for_each_ptr(wp, ob, i) {
-			ca = bch_dev_bkey_exists(c, ob->ptr.dev);
+	/* Free buckets we didn't use: */
+	open_bucket_for_each(c, &wp->ptrs, ob, i)
+		open_bucket_free_unused(c, wp, ob);
 
-			if (ca->mi.durability &&
-			    ca->mi.durability <= nr_ptrs_effective - nr_replicas &&
-			    !bch2_dev_in_target(c, ob->ptr.dev, target)) {
-				swap(wp->ptrs[i], wp->ptrs[wp->first_ptr]);
-				wp->first_ptr++;
-				nr_ptrs_effective -= ca->mi.durability;
-			}
-		}
-	}
-
-	if (nr_ptrs_effective > nr_replicas) {
-		writepoint_for_each_ptr(wp, ob, i) {
-			ca = bch_dev_bkey_exists(c, ob->ptr.dev);
-
-			if (ca->mi.durability &&
-			    ca->mi.durability <= nr_ptrs_effective - nr_replicas) {
-				swap(wp->ptrs[i], wp->ptrs[wp->first_ptr]);
-				wp->first_ptr++;
-				nr_ptrs_effective -= ca->mi.durability;
-			}
-		}
-	}
-
-	/* Remove pointers we don't want to use: */
-	if (target)
-		bch2_writepoint_drop_ptrs(c, wp, target, false);
-
-	BUG_ON(wp->first_ptr >= wp->nr_ptrs);
-	BUG_ON(nr_ptrs_effective < nr_replicas_required);
+	wp->ptrs = ptrs;
 
 	wp->sectors_free = UINT_MAX;
 
-	writepoint_for_each_ptr(wp, ob, i)
+	open_bucket_for_each(c, &wp->ptrs, ob, i)
 		wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
 
 	BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
 
-	verify_not_stale(c, wp);
+	verify_not_stale(c, &wp->ptrs);
 
 	return wp;
 err:
+	open_bucket_for_each(c, &wp->ptrs, ob, i)
+		if (ptrs.nr < ARRAY_SIZE(ptrs.v))
+			ob_push(c, &ptrs, ob);
+		else
+			open_bucket_free_unused(c, wp, ob);
+	wp->ptrs = ptrs;
+
 	mutex_unlock(&wp->lock);
 	return ERR_PTR(ret);
 }
@@ -703,7 +653,7 @@ void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
 	BUG_ON(sectors > wp->sectors_free);
 	wp->sectors_free -= sectors;
 
-	writepoint_for_each_ptr(wp, ob, i) {
+	open_bucket_for_each(c, &wp->ptrs, ob, i) {
 		struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 		struct bch_extent_ptr tmp = ob->ptr;
 
@@ -726,16 +676,15 @@ void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
  */
 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
 {
-	int i;
+	struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
+	struct open_bucket *ob;
+	unsigned i;
 
-	for (i = wp->nr_ptrs - 1; i >= 0; --i) {
-		struct open_bucket *ob = wp->ptrs[i];
-
-		if (!ob->sectors_free) {
-			array_remove_item(wp->ptrs, wp->nr_ptrs, i);
-			bch2_open_bucket_put(c, ob);
-		}
-	}
+	open_bucket_for_each(c, &wp->ptrs, ob, i)
+		ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
+	wp->ptrs = keep;
 
 	mutex_unlock(&wp->lock);
+
+	bch2_open_buckets_put(c, &ptrs);
 }

fs/bcachefs/alloc_foreground.h

@@ -23,19 +23,23 @@ void bch2_wp_rescale(struct bch_fs *, struct bch_dev *,
 
 long bch2_bucket_alloc_new_fs(struct bch_dev *);
 
-int bch2_bucket_alloc(struct bch_fs *, struct bch_dev *, enum alloc_reserve, bool,
+struct open_bucket *bch2_bucket_alloc(struct bch_fs *, struct bch_dev *,
+				      enum alloc_reserve, bool,
 				      struct closure *);
 
-#define __writepoint_for_each_ptr(_wp, _ob, _i, _start)			\
-	for ((_i) = (_start);						\
-	     (_i) < (_wp)->nr_ptrs && ((_ob) = (_wp)->ptrs[_i], true);	\
-	     (_i)++)
+static inline void ob_push(struct bch_fs *c, struct open_buckets *obs,
+			   struct open_bucket *ob)
+{
+	BUG_ON(obs->nr >= ARRAY_SIZE(obs->v));
 
-#define writepoint_for_each_ptr_all(_wp, _ob, _i)			\
-	__writepoint_for_each_ptr(_wp, _ob, _i, 0)
+	obs->v[obs->nr++] = ob - c->open_buckets;
+}
 
-#define writepoint_for_each_ptr(_wp, _ob, _i)				\
-	__writepoint_for_each_ptr(_wp, _ob, _i, wp->first_ptr)
+#define open_bucket_for_each(_c, _obs, _ob, _i)				\
+	for ((_i) = 0;							\
+	     (_i) < (_obs)->nr &&					\
+	     ((_ob) = (_c)->open_buckets + (_obs)->v[_i], true);	\
+	     (_i)++)
 
 void __bch2_open_bucket_put(struct bch_fs *, struct open_bucket *);
 
@@ -45,26 +49,27 @@ static inline void bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob
 		__bch2_open_bucket_put(c, ob);
 }
 
-static inline void bch2_open_bucket_put_refs(struct bch_fs *c, u8 *nr, u8 *refs)
+static inline void bch2_open_buckets_put(struct bch_fs *c,
+					 struct open_buckets *ptrs)
 {
+	struct open_bucket *ob;
 	unsigned i;
 
-	for (i = 0; i < *nr; i++)
-		bch2_open_bucket_put(c, c->open_buckets + refs[i]);
-
-	*nr = 0;
+	open_bucket_for_each(c, ptrs, ob, i)
+		bch2_open_bucket_put(c, ob);
+	ptrs->nr = 0;
 }
 
 static inline void bch2_open_bucket_get(struct bch_fs *c,
 					struct write_point *wp,
-					u8 *nr, u8 *refs)
+					struct open_buckets *ptrs)
 {
 	struct open_bucket *ob;
 	unsigned i;
 
-	writepoint_for_each_ptr(wp, ob, i) {
+	open_bucket_for_each(c, &wp->ptrs, ob, i) {
 		atomic_inc(&ob->pin);
-		refs[(*nr)++] = ob - c->open_buckets;
+		ob_push(c, ptrs, ob);
 	}
 }
 
@@ -84,9 +89,6 @@ void bch2_alloc_sectors_done(struct bch_fs *, struct write_point *);
 void bch2_writepoint_stop(struct bch_fs *, struct bch_dev *,
 			  struct write_point *);
 
-void bch2_writepoint_drop_ptrs(struct bch_fs *, struct write_point *,
-			       u16, bool);
-
 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 						 unsigned long write_point)
 {

fs/bcachefs/alloc_types.h

@@ -58,6 +58,13 @@ struct open_bucket {
 	struct bch_extent_ptr	ptr;
 };
 
+#define OPEN_BUCKET_LIST_MAX	15
+
+struct open_buckets {
+	u8			nr;
+	u8			v[OPEN_BUCKET_LIST_MAX];
+};
+
 struct write_point {
 	struct hlist_node	node;
 	struct mutex		lock;
@@ -65,13 +72,10 @@ struct write_point {
 	unsigned long		write_point;
 	enum bch_data_type	type;
 
-	u8			nr_ptrs;
-	u8			first_ptr;
-
 	/* calculated based on how many pointers we're actually going to use: */
 	unsigned		sectors_free;
 
-	struct open_bucket	*ptrs[BCH_REPLICAS_MAX * 2];
+	struct open_buckets	ptrs;
 	u64			next_alloc[BCH_SB_MEMBERS_MAX];
 };
 

fs/bcachefs/btree_gc.c

@@ -6,6 +6,7 @@
 
 #include "bcachefs.h"
 #include "alloc_background.h"
+#include "alloc_foreground.h"
 #include "bkey_methods.h"
 #include "btree_locking.h"
 #include "btree_update_interior.h"
@@ -803,7 +804,7 @@ static void bch2_coalesce_nodes(struct bch_fs *c, struct btree_iter *iter,
 	bch2_btree_iter_node_replace(iter, new_nodes[0]);
 
 	for (i = 0; i < nr_new_nodes; i++)
-		bch2_btree_open_bucket_put(c, new_nodes[i]);
+		bch2_open_buckets_put(c, &new_nodes[i]->ob);
 
 	/* Free the old nodes and update our sliding window */
 	for (i = 0; i < nr_old_nodes; i++) {

fs/bcachefs/btree_types.h

@@ -54,13 +54,8 @@ struct btree_write {
 	struct closure_waitlist		wait;
 };
 
-struct btree_ob_ref {
-	u8			nr;
-	u8			refs[BCH_REPLICAS_MAX];
-};
-
 struct btree_alloc {
-	struct btree_ob_ref	ob;
+	struct open_buckets	ob;
 	BKEY_PADDED(k);
 };
 
@@ -127,7 +122,7 @@ struct btree {
 	 */
 	unsigned long		will_make_reachable;
 
-	struct btree_ob_ref	ob;
+	struct open_buckets	ob;
 
 	/* lru list */
 	struct list_head	list;

fs/bcachefs/btree_update_interior.c

@@ -247,7 +247,7 @@ static void __btree_node_free(struct bch_fs *c, struct btree *b)
 
 void bch2_btree_node_free_never_inserted(struct bch_fs *c, struct btree *b)
 {
-	struct btree_ob_ref ob = b->ob;
+	struct open_buckets ob = b->ob;
 
 	btree_update_drop_new_node(c, b);
 
@@ -259,7 +259,7 @@ void bch2_btree_node_free_never_inserted(struct bch_fs *c, struct btree *b)
 	__btree_node_free(c, b);
 	six_unlock_write(&b->lock);
 
-	bch2_open_bucket_put_refs(c, &ob.nr, ob.refs);
+	bch2_open_buckets_put(c, &ob);
 }
 
 void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b,
@@ -300,11 +300,6 @@ static void bch2_btree_node_free_ondisk(struct bch_fs *c,
 	 */
 }
 
-void bch2_btree_open_bucket_put(struct bch_fs *c, struct btree *b)
-{
-	bch2_open_bucket_put_refs(c, &b->ob.nr, b->ob.refs);
-}
-
 static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
 					     struct disk_reservation *res,
 					     struct closure *cl,
@@ -314,7 +309,7 @@ static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
 	struct btree *b;
 	BKEY_PADDED(k) tmp;
 	struct bkey_i_extent *e;
-	struct btree_ob_ref ob;
+	struct open_buckets ob = { .nr = 0 };
 	struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 	unsigned nr_reserve;
 	enum alloc_reserve alloc_reserve;
@@ -356,7 +351,7 @@ static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
 		struct open_bucket *ob;
 		unsigned i;
 
-		writepoint_for_each_ptr(wp, ob, i)
+		open_bucket_for_each(c, &wp->ptrs, ob, i)
 			if (ob->sectors_free < c->opts.btree_node_size)
 				ob->sectors_free = 0;
 
@@ -367,8 +362,7 @@ static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
 	e = bkey_extent_init(&tmp.k);
 	bch2_alloc_sectors_append_ptrs(c, wp, e, c->opts.btree_node_size);
 
-	ob.nr = 0;
-	bch2_open_bucket_get(c, wp, &ob.nr, ob.refs);
+	bch2_open_bucket_get(c, wp, &ob);
 	bch2_alloc_sectors_done(c, wp);
 mem_alloc:
 	b = bch2_btree_node_mem_alloc(c);
@@ -489,7 +483,7 @@ static void bch2_btree_reserve_put(struct bch_fs *c, struct btree_reserve *reser
 			b->ob.nr = 0;
 			bkey_copy(&a->k, &b->key);
 		} else {
-			bch2_btree_open_bucket_put(c, b);
+			bch2_open_buckets_put(c, &b->ob);
 		}
 
 		btree_node_lock_type(c, b, SIX_LOCK_write);
@@ -1432,11 +1426,11 @@ static void btree_split(struct btree_update *as, struct btree *b,
 		bch2_btree_set_root(as, n1, iter);
 	}
 
-	bch2_btree_open_bucket_put(c, n1);
+	bch2_open_buckets_put(c, &n1->ob);
 	if (n2)
-		bch2_btree_open_bucket_put(c, n2);
+		bch2_open_buckets_put(c, &n2->ob);
 	if (n3)
-		bch2_btree_open_bucket_put(c, n3);
+		bch2_open_buckets_put(c, &n3->ob);
 
 	/*
 	 * Note - at this point other linked iterators could still have @b read
@@ -1751,7 +1745,7 @@ void __bch2_foreground_maybe_merge(struct bch_fs *c,
 
 	bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
 
-	bch2_btree_open_bucket_put(c, n);
+	bch2_open_buckets_put(c, &n->ob);
 	bch2_btree_node_free_inmem(c, b, iter);
 	bch2_btree_node_free_inmem(c, m, iter);
 	bch2_btree_iter_node_replace(iter, n);
@@ -1843,7 +1837,7 @@ static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
 		bch2_btree_set_root(as, n, iter);
 	}
 
-	bch2_btree_open_bucket_put(c, n);
+	bch2_open_buckets_put(c, &n->ob);
 
 	bch2_btree_node_free_inmem(c, b, iter);
 

fs/bcachefs/btree_update_interior.h

@@ -132,7 +132,6 @@ struct btree_update {
 void bch2_btree_node_free_inmem(struct bch_fs *, struct btree *,
 				struct btree_iter *);
 void bch2_btree_node_free_never_inserted(struct bch_fs *, struct btree *);
-void bch2_btree_open_bucket_put(struct bch_fs *, struct btree *);
 
 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *,
 						  struct btree *,

fs/bcachefs/io.c

@@ -369,7 +369,7 @@ static void __bch2_write_index(struct bch_write_op *op)
 		}
 	}
 out:
-	bch2_open_bucket_put_refs(c, &op->open_buckets_nr, op->open_buckets);
+	bch2_open_buckets_put(c, &op->open_buckets);
 	return;
 err:
 	keys->top = keys->keys;
@@ -816,8 +816,8 @@ static void __bch2_write(struct closure *cl)
 again:
 	do {
 		/* +1 for possible cache device: */
-		if (op->open_buckets_nr + op->nr_replicas + 1 >
-		    ARRAY_SIZE(op->open_buckets))
+		if (op->open_buckets.nr + op->nr_replicas + 1 >
+		    ARRAY_SIZE(op->open_buckets.v))
 			goto flush_io;
 
 		if (bch2_keylist_realloc(&op->insert_keys,
@@ -848,11 +848,7 @@ static void __bch2_write(struct closure *cl)
 
 		ret = bch2_write_extent(op, wp);
 
-		BUG_ON(op->open_buckets_nr + wp->nr_ptrs - wp->first_ptr >
-		       ARRAY_SIZE(op->open_buckets));
-		bch2_open_bucket_get(c, wp,
-				     &op->open_buckets_nr,
-				     op->open_buckets);
+		bch2_open_bucket_get(c, wp, &op->open_buckets);
 		bch2_alloc_sectors_done(c, wp);
 
 		if (ret < 0)

fs/bcachefs/io.h

@@ -75,7 +75,7 @@ static inline void bch2_write_op_init(struct bch_write_op *op, struct bch_fs *c,
 	op->nr_replicas		= 0;
 	op->nr_replicas_required = c->opts.data_replicas_required;
 	op->alloc_reserve	= RESERVE_NONE;
-	op->open_buckets_nr	= 0;
+	op->open_buckets.nr	= 0;
 	op->devs_have.nr	= 0;
 	op->target		= 0;
 	op->opts		= opts;

fs/bcachefs/io_types.h

@@ -106,7 +106,6 @@ struct bch_write_op {
 	unsigned		nr_replicas_required:4;
 	unsigned		alloc_reserve:4;
 
-	u8			open_buckets_nr;
 	struct bch_devs_list	devs_have;
 	u16			target;
 	u16			nonce;
@@ -123,7 +122,7 @@ struct bch_write_op {
 
 	struct disk_reservation	res;
 
-	u8			open_buckets[16];
+	struct open_buckets	open_buckets;
 
 	/*
 	 * If caller wants to flush but hasn't passed us a journal_seq ptr, we

fs/bcachefs/journal.c

@@ -717,13 +717,13 @@ static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
 				goto err;
 			}
 		} else {
-			int ob_idx = bch2_bucket_alloc(c, ca, RESERVE_ALLOC, false, cl);
-			if (ob_idx < 0) {
+			ob = bch2_bucket_alloc(c, ca, RESERVE_ALLOC,
+					       false, cl);
+			if (IS_ERR(ob)) {
 				ret = cl ? -EAGAIN : -ENOSPC;
 				goto err;
 			}
 
-			ob = c->open_buckets + ob_idx;
 			bucket = sector_to_bucket(ca, ob->ptr.offset);
 		}