bcachefs: Make copygc thread global

Per device copygc threads don't move data to different devices and they
make fragmentation works - they don't make much sense anymore.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>

10 files changed, 159 insertions, 172 deletions

diff --git a/fs/bcachefs/alloc_background.c b/fs/bcachefs/alloc_background.c
index 0cc63b430736..bb88ffb18e24 100644
--- a/fs/bcachefs/alloc_background.c
+++ b/fs/bcachefs/alloc_background.c
@@ -41,29 +41,26 @@ static void pd_controllers_update(struct work_struct *work)
 					   struct bch_fs,
 					   pd_controllers_update);
 	struct bch_dev *ca;
+	s64 free = 0, fragmented = 0;
 	unsigned i;
 
 	for_each_member_device(ca, c, i) {
 		struct bch_dev_usage stats = bch2_dev_usage_read(c, ca);
 
-		u64 free = bucket_to_sector(ca,
+		free += bucket_to_sector(ca,
 				__dev_buckets_free(ca, stats)) << 9;
 		/*
 		 * Bytes of internal fragmentation, which can be
 		 * reclaimed by copy GC
 		 */
-		s64 fragmented = (bucket_to_sector(ca,
+		fragmented += max_t(s64, 0, (bucket_to_sector(ca,
 					stats.buckets[BCH_DATA_user] +
 					stats.buckets[BCH_DATA_cached]) -
 				  (stats.sectors[BCH_DATA_user] +
-				   stats.sectors[BCH_DATA_cached])) << 9;
-
-		fragmented = max(0LL, fragmented);
-
-		bch2_pd_controller_update(&ca->copygc_pd,
-					 free, fragmented, -1);
+				   stats.sectors[BCH_DATA_cached])) << 9);
 	}
 
+	bch2_pd_controller_update(&c->copygc_pd, free, fragmented, -1);
 	schedule_delayed_work(&c->pd_controllers_update,
 			      c->pd_controllers_update_seconds * HZ);
 }
@@ -1191,7 +1188,7 @@ stop:
 void bch2_recalc_capacity(struct bch_fs *c)
 {
 	struct bch_dev *ca;
-	u64 capacity = 0, reserved_sectors = 0, gc_reserve;
+	u64 capacity = 0, reserved_sectors = 0, gc_reserve, copygc_threshold = 0;
 	unsigned bucket_size_max = 0;
 	unsigned long ra_pages = 0;
 	unsigned i, j;
@@ -1234,7 +1231,7 @@ void bch2_recalc_capacity(struct bch_fs *c)
 
 		dev_reserve *= ca->mi.bucket_size;
 
-		ca->copygc_threshold = dev_reserve;
+		copygc_threshold += dev_reserve;
 
 		capacity += bucket_to_sector(ca, ca->mi.nbuckets -
 					     ca->mi.first_bucket);
@@ -1253,6 +1250,7 @@ void bch2_recalc_capacity(struct bch_fs *c)
 
 	reserved_sectors = min(reserved_sectors, capacity);
 
+	c->copygc_threshold = copygc_threshold;
 	c->capacity = capacity - reserved_sectors;
 
 	c->bucket_size_max = bucket_size_max;
@@ -1312,7 +1310,7 @@ void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
 	for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
 		bch2_writepoint_stop(c, ca, &c->write_points[i]);
 
-	bch2_writepoint_stop(c, ca, &ca->copygc_write_point);
+	bch2_writepoint_stop(c, ca, &c->copygc_write_point);
 	bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
 	bch2_writepoint_stop(c, ca, &c->btree_write_point);
 
diff --git a/fs/bcachefs/alloc_foreground.c b/fs/bcachefs/alloc_foreground.c
index cc5e6d3d0012..10bc5bfaf2fa 100644
--- a/fs/bcachefs/alloc_foreground.c
+++ b/fs/bcachefs/alloc_foreground.c
@@ -956,8 +956,9 @@ void bch2_fs_allocator_foreground_init(struct bch_fs *c)
 		c->open_buckets_freelist = ob - c->open_buckets;
 	}
 
-	writepoint_init(&c->btree_write_point, BCH_DATA_btree);
-	writepoint_init(&c->rebalance_write_point, BCH_DATA_user);
+	writepoint_init(&c->btree_write_point,		BCH_DATA_btree);
+	writepoint_init(&c->rebalance_write_point,	BCH_DATA_user);
+	writepoint_init(&c->copygc_write_point,		BCH_DATA_user);
 
 	for (wp = c->write_points;
 	     wp < c->write_points + c->write_points_nr; wp++) {
diff --git a/fs/bcachefs/bcachefs.h b/fs/bcachefs/bcachefs.h
index 98acd64c33d9..e862b17f1930 100644
--- a/fs/bcachefs/bcachefs.h
+++ b/fs/bcachefs/bcachefs.h
@@ -452,13 +452,6 @@ struct bch_dev {
 
 	alloc_heap		alloc_heap;
 
-	/* Copying GC: */
-	struct task_struct	*copygc_thread;
-	copygc_heap		copygc_heap;
-	struct bch_pd_controller copygc_pd;
-	struct write_point	copygc_write_point;
-	u64			copygc_threshold;
-
 	atomic64_t		rebalance_work;
 
 	struct journal_device	journal;
@@ -753,6 +746,13 @@ struct bch_fs {
 	/* REBALANCE */
 	struct bch_fs_rebalance	rebalance;
 
+	/* COPYGC */
+	struct task_struct	*copygc_thread;
+	copygc_heap		copygc_heap;
+	struct bch_pd_controller copygc_pd;
+	struct write_point	copygc_write_point;
+	u64			copygc_threshold;
+
 	/* STRIPES: */
 	GENRADIX(struct stripe) stripes[2];
 
diff --git a/fs/bcachefs/buckets.c b/fs/bcachefs/buckets.c
index 0ca810d6ec56..2ba4ef8e1a50 100644
--- a/fs/bcachefs/buckets.c
+++ b/fs/bcachefs/buckets.c
@@ -2009,7 +2009,6 @@ int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
 	alloc_fifo	free[RESERVE_NR];
 	alloc_fifo	free_inc;
 	alloc_heap	alloc_heap;
-	copygc_heap	copygc_heap;
 
 	size_t btree_reserve	= DIV_ROUND_UP(BTREE_NODE_RESERVE,
 			     ca->mi.bucket_size / c->opts.btree_node_size);
@@ -2018,15 +2017,13 @@ int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
 	size_t copygc_reserve	= max_t(size_t, 2, nbuckets >> 7);
 	size_t free_inc_nr	= max(max_t(size_t, 1, nbuckets >> 12),
 				      btree_reserve * 2);
-	bool resize = ca->buckets[0] != NULL,
-	     start_copygc = ca->copygc_thread != NULL;
+	bool resize = ca->buckets[0] != NULL;
 	int ret = -ENOMEM;
 	unsigned i;
 
 	memset(&free,		0, sizeof(free));
 	memset(&free_inc,	0, sizeof(free_inc));
 	memset(&alloc_heap,	0, sizeof(alloc_heap));
-	memset(&copygc_heap,	0, sizeof(copygc_heap));
 
 	if (!(buckets		= kvpmalloc(sizeof(struct bucket_array) +
 					    nbuckets * sizeof(struct bucket),
@@ -2039,14 +2036,13 @@ int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
 		       copygc_reserve, GFP_KERNEL) ||
 	    !init_fifo(&free[RESERVE_NONE], reserve_none, GFP_KERNEL) ||
 	    !init_fifo(&free_inc,	free_inc_nr, GFP_KERNEL) ||
-	    !init_heap(&alloc_heap,	ALLOC_SCAN_BATCH(ca) << 1, GFP_KERNEL) ||
-	    !init_heap(&copygc_heap,	copygc_reserve, GFP_KERNEL))
+	    !init_heap(&alloc_heap,	ALLOC_SCAN_BATCH(ca) << 1, GFP_KERNEL))
 		goto err;
 
 	buckets->first_bucket	= ca->mi.first_bucket;
 	buckets->nbuckets	= nbuckets;
 
-	bch2_copygc_stop(ca);
+	bch2_copygc_stop(c);
 
 	if (resize) {
 		down_write(&c->gc_lock);
@@ -2089,21 +2085,13 @@ int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
 	/* with gc lock held, alloc_heap can't be in use: */
 	swap(ca->alloc_heap, alloc_heap);
 
-	/* and we shut down copygc: */
-	swap(ca->copygc_heap, copygc_heap);
-
 	nbuckets = ca->mi.nbuckets;
 
 	if (resize)
 		up_write(&ca->bucket_lock);
 
-	if (start_copygc &&
-	    bch2_copygc_start(c, ca))
-		bch_err(ca, "error restarting copygc thread");
-
 	ret = 0;
 err:
-	free_heap(&copygc_heap);
 	free_heap(&alloc_heap);
 	free_fifo(&free_inc);
 	for (i = 0; i < RESERVE_NR; i++)
@@ -2120,7 +2108,6 @@ void bch2_dev_buckets_free(struct bch_dev *ca)
 {
 	unsigned i;
 
-	free_heap(&ca->copygc_heap);
 	free_heap(&ca->alloc_heap);
 	free_fifo(&ca->free_inc);
 	for (i = 0; i < RESERVE_NR; i++)
diff --git a/fs/bcachefs/buckets_types.h b/fs/bcachefs/buckets_types.h
index b64b2fc9a896..0f7fcfe29e0e 100644
--- a/fs/bcachefs/buckets_types.h
+++ b/fs/bcachefs/buckets_types.h
@@ -121,6 +121,7 @@ struct disk_reservation {
 };
 
 struct copygc_heap_entry {
+	u8			dev;
 	u8			gen;
 	u32			sectors;
 	u64			offset;
diff --git a/fs/bcachefs/movinggc.c b/fs/bcachefs/movinggc.c
index 27e966edac23..4429b2bb2df8 100644
--- a/fs/bcachefs/movinggc.c
+++ b/fs/bcachefs/movinggc.c
@@ -43,13 +43,6 @@
 #define COPYGC_BUCKETS_PER_ITER(ca)					\
 	((ca)->free[RESERVE_MOVINGGC].size / 2)
 
-/*
- * Max sectors to move per iteration: Have to take into account internal
- * fragmentation from the multiple write points for each generation:
- */
-#define COPYGC_SECTORS_PER_ITER(ca)					\
-	((ca)->mi.bucket_size *	COPYGC_BUCKETS_PER_ITER(ca))
-
 static inline int sectors_used_cmp(copygc_heap *heap,
 				   struct copygc_heap_entry l,
 				   struct copygc_heap_entry r)
@@ -62,18 +55,22 @@ static int bucket_offset_cmp(const void *_l, const void *_r, size_t size)
 	const struct copygc_heap_entry *l = _l;
 	const struct copygc_heap_entry *r = _r;
 
-	return cmp_int(l->offset, r->offset);
+	return  cmp_int(l->dev,    r->dev) ?:
+		cmp_int(l->offset, r->offset);
 }
 
-static bool __copygc_pred(struct bch_dev *ca,
-			  struct bkey_s_c k)
+static int __copygc_pred(struct bch_fs *c, struct bkey_s_c k)
 {
-	copygc_heap *h = &ca->copygc_heap;
-	const struct bch_extent_ptr *ptr =
-		bch2_bkey_has_device(k, ca->dev_idx);
-
-	if (ptr) {
-		struct copygc_heap_entry search = { .offset = ptr->offset };
+	copygc_heap *h = &c->copygc_heap;
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	const struct bch_extent_ptr *ptr;
+
+	bkey_for_each_ptr(ptrs, ptr) {
+		struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
+		struct copygc_heap_entry search = {
+			.dev = ptr->dev,
+			.offset = ptr->offset
+		};
 
 		ssize_t i = eytzinger0_find_le(h->data, h->used,
 					       sizeof(h->data[0]),
@@ -89,12 +86,13 @@ static bool __copygc_pred(struct bch_dev *ca,
 
 		BUG_ON(i != j);
 #endif
-		return (i >= 0 &&
-			ptr->offset < h->data[i].offset + ca->mi.bucket_size &&
-			ptr->gen == h->data[i].gen);
+		if (i >= 0 &&
+		    ptr->offset < h->data[i].offset + ca->mi.bucket_size &&
+		    ptr->gen == h->data[i].gen)
+			return ptr->dev;
 	}
 
-	return false;
+	return -1;
 }
 
 static enum data_cmd copygc_pred(struct bch_fs *c, void *arg,
@@ -102,14 +100,14 @@ static enum data_cmd copygc_pred(struct bch_fs *c, void *arg,
 				 struct bch_io_opts *io_opts,
 				 struct data_opts *data_opts)
 {
-	struct bch_dev *ca = arg;
-
-	if (!__copygc_pred(ca, k))
+	int dev_idx = __copygc_pred(c, k);
+	if (dev_idx < 0)
 		return DATA_SKIP;
 
-	data_opts->target		= dev_to_target(ca->dev_idx);
+	/* XXX: use io_opts for this inode */
+	data_opts->target		= dev_to_target(dev_idx);
 	data_opts->btree_insert_flags	= BTREE_INSERT_USE_RESERVE;
-	data_opts->rewrite_dev		= ca->dev_idx;
+	data_opts->rewrite_dev		= dev_idx;
 	return DATA_REWRITE;
 }
 
@@ -125,20 +123,21 @@ static bool have_copygc_reserve(struct bch_dev *ca)
 	return ret;
 }
 
-static void bch2_copygc(struct bch_fs *c, struct bch_dev *ca)
+static void bch2_copygc(struct bch_fs *c)
 {
-	copygc_heap *h = &ca->copygc_heap;
+	copygc_heap *h = &c->copygc_heap;
 	struct copygc_heap_entry e, *i;
 	struct bucket_array *buckets;
 	struct bch_move_stats move_stats;
 	u64 sectors_to_move = 0, sectors_not_moved = 0;
+	u64 sectors_reserved = 0;
 	u64 buckets_to_move, buckets_not_moved = 0;
-	size_t b;
+	struct bch_dev *ca;
+	unsigned dev_idx;
+	size_t b, heap_size = 0;
 	int ret;
 
 	memset(&move_stats, 0, sizeof(move_stats));
-	closure_wait_event(&c->freelist_wait, have_copygc_reserve(ca));
-
 	/*
 	 * Find buckets with lowest sector counts, skipping completely
 	 * empty buckets, by building a maxheap sorted by sector count,
@@ -147,38 +146,51 @@ static void bch2_copygc(struct bch_fs *c, struct bch_dev *ca)
 	 */
 	h->used = 0;
 
-	/*
-	 * We need bucket marks to be up to date - gc can't be recalculating
-	 * them:
-	 */
-	down_read(&c->gc_lock);
-	down_read(&ca->bucket_lock);
-	buckets = bucket_array(ca);
-
-	for (b = buckets->first_bucket; b < buckets->nbuckets; b++) {
-		struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
-		struct copygc_heap_entry e;
-
-		if (m.owned_by_allocator ||
-		    m.data_type != BCH_DATA_user ||
-		    !bucket_sectors_used(m) ||
-		    bucket_sectors_used(m) >= ca->mi.bucket_size)
-			continue;
+	for_each_rw_member(ca, c, dev_idx)
+		heap_size += ca->mi.nbuckets >> 7;
 
-		e = (struct copygc_heap_entry) {
-			.gen		= m.gen,
-			.sectors	= bucket_sectors_used(m),
-			.offset		= bucket_to_sector(ca, b),
-		};
-		heap_add_or_replace(h, e, -sectors_used_cmp, NULL);
+	if (h->size < heap_size) {
+		free_heap(&c->copygc_heap);
+		if (!init_heap(&c->copygc_heap, heap_size, GFP_KERNEL)) {
+			bch_err(c, "error allocating copygc heap");
+			return;
+		}
+	}
+
+	for_each_rw_member(ca, c, dev_idx) {
+		closure_wait_event(&c->freelist_wait, have_copygc_reserve(ca));
+
+		spin_lock(&ca->fs->freelist_lock);
+		sectors_reserved += fifo_used(&ca->free[RESERVE_MOVINGGC]) * ca->mi.bucket_size;
+		spin_unlock(&ca->fs->freelist_lock);
+
+		down_read(&ca->bucket_lock);
+		buckets = bucket_array(ca);
+
+		for (b = buckets->first_bucket; b < buckets->nbuckets; b++) {
+			struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
+			struct copygc_heap_entry e;
+
+			if (m.owned_by_allocator ||
+			    m.data_type != BCH_DATA_user ||
+			    !bucket_sectors_used(m) ||
+			    bucket_sectors_used(m) >= ca->mi.bucket_size)
+				continue;
+
+			e = (struct copygc_heap_entry) {
+				.gen		= m.gen,
+				.sectors	= bucket_sectors_used(m),
+				.offset		= bucket_to_sector(ca, b),
+			};
+			heap_add_or_replace(h, e, -sectors_used_cmp, NULL);
+		}
+		up_read(&ca->bucket_lock);
 	}
-	up_read(&ca->bucket_lock);
-	up_read(&c->gc_lock);
 
 	for (i = h->data; i < h->data + h->used; i++)
 		sectors_to_move += i->sectors;
 
-	while (sectors_to_move > COPYGC_SECTORS_PER_ITER(ca)) {
+	while (sectors_to_move > sectors_reserved) {
 		BUG_ON(!heap_pop(h, e, -sectors_used_cmp, NULL));
 		sectors_to_move -= e.sectors;
 	}
@@ -192,24 +204,26 @@ static void bch2_copygc(struct bch_fs *c, struct bch_dev *ca)
 			sizeof(h->data[0]),
 			bucket_offset_cmp, NULL);
 
-	ret = bch2_move_data(c, &ca->copygc_pd.rate,
-			     writepoint_ptr(&ca->copygc_write_point),
+	ret = bch2_move_data(c, &c->copygc_pd.rate,
+			     writepoint_ptr(&c->copygc_write_point),
 			     POS_MIN, POS_MAX,
-			     copygc_pred, ca,
+			     copygc_pred, NULL,
 			     &move_stats);
 
-	down_read(&ca->bucket_lock);
-	buckets = bucket_array(ca);
-	for (i = h->data; i < h->data + h->used; i++) {
-		size_t b = sector_to_bucket(ca, i->offset);
-		struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
-
-		if (i->gen == m.gen && bucket_sectors_used(m)) {
-			sectors_not_moved += bucket_sectors_used(m);
-			buckets_not_moved++;
+	for_each_rw_member(ca, c, dev_idx) {
+		down_read(&ca->bucket_lock);
+		buckets = bucket_array(ca);
+		for (i = h->data; i < h->data + h->used; i++) {
+			size_t b = sector_to_bucket(ca, i->offset);
+			struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
+
+			if (i->gen == m.gen && bucket_sectors_used(m)) {
+				sectors_not_moved += bucket_sectors_used(m);
+				buckets_not_moved++;
+			}
 		}
+		up_read(&ca->bucket_lock);
 	}
-	up_read(&ca->bucket_lock);
 
 	if (sectors_not_moved && !ret)
 		bch_warn_ratelimited(c,
@@ -220,7 +234,7 @@ static void bch2_copygc(struct bch_fs *c, struct bch_dev *ca)
 			 atomic64_read(&move_stats.keys_raced),
 			 atomic64_read(&move_stats.sectors_raced));
 
-	trace_copygc(ca,
+	trace_copygc(c,
 		     atomic64_read(&move_stats.sectors_moved), sectors_not_moved,
 		     buckets_to_move, buckets_not_moved);
 }
@@ -239,20 +253,27 @@ static void bch2_copygc(struct bch_fs *c, struct bch_dev *ca)
  * often and continually reduce the amount of fragmented space as the device
  * fills up. So, we increase the threshold by half the current free space.
  */
-unsigned long bch2_copygc_wait_amount(struct bch_dev *ca)
+unsigned long bch2_copygc_wait_amount(struct bch_fs *c)
 {
-	struct bch_fs *c = ca->fs;
-	struct bch_dev_usage usage = bch2_dev_usage_read(c, ca);
-	u64 fragmented_allowed = ca->copygc_threshold +
-		((__dev_buckets_available(ca, usage) * ca->mi.bucket_size) >> 1);
+	struct bch_dev *ca;
+	unsigned dev_idx;
+	u64 fragmented_allowed = c->copygc_threshold;
+	u64 fragmented = 0;
+
+	for_each_rw_member(ca, c, dev_idx) {
+		struct bch_dev_usage usage = bch2_dev_usage_read(c, ca);
+
+		fragmented_allowed += ((__dev_buckets_available(ca, usage) *
+					ca->mi.bucket_size) >> 1);
+		fragmented += usage.sectors_fragmented;
+	}
 
-	return max_t(s64, 0, fragmented_allowed - usage.sectors_fragmented);
+	return max_t(s64, 0, fragmented_allowed - fragmented);
 }
 
 static int bch2_copygc_thread(void *arg)
 {
-	struct bch_dev *ca = arg;
-	struct bch_fs *c = ca->fs;
+	struct bch_fs *c = arg;
 	struct io_clock *clock = &c->io_clock[WRITE];
 	unsigned long last, wait;
 
@@ -263,7 +284,7 @@ static int bch2_copygc_thread(void *arg)
 			break;
 
 		last = atomic_long_read(&clock->now);
-		wait = bch2_copygc_wait_amount(ca);
+		wait = bch2_copygc_wait_amount(c);
 
 		if (wait > clock->max_slop) {
 			bch2_kthread_io_clock_wait(clock, last + wait,
@@ -271,29 +292,29 @@ static int bch2_copygc_thread(void *arg)
 			continue;
 		}
 
-		bch2_copygc(c, ca);
+		bch2_copygc(c);
 	}
 
 	return 0;
 }
 
-void bch2_copygc_stop(struct bch_dev *ca)
+void bch2_copygc_stop(struct bch_fs *c)
 {
-	ca->copygc_pd.rate.rate = UINT_MAX;
-	bch2_ratelimit_reset(&ca->copygc_pd.rate);
+	c->copygc_pd.rate.rate = UINT_MAX;
+	bch2_ratelimit_reset(&c->copygc_pd.rate);
 
-	if (ca->copygc_thread) {
-		kthread_stop(ca->copygc_thread);
-		put_task_struct(ca->copygc_thread);
+	if (c->copygc_thread) {
+		kthread_stop(c->copygc_thread);
+		put_task_struct(c->copygc_thread);
 	}
-	ca->copygc_thread = NULL;
+	c->copygc_thread = NULL;
 }
 
-int bch2_copygc_start(struct bch_fs *c, struct bch_dev *ca)
+int bch2_copygc_start(struct bch_fs *c)
 {
 	struct task_struct *t;
 
-	if (ca->copygc_thread)
+	if (c->copygc_thread)
 		return 0;
 
 	if (c->opts.nochanges)
@@ -302,21 +323,21 @@ int bch2_copygc_start(struct bch_fs *c, struct bch_dev *ca)
 	if (bch2_fs_init_fault("copygc_start"))
 		return -ENOMEM;
 
-	t = kthread_create(bch2_copygc_thread, ca,
-			   "bch_copygc[%s]", ca->name);
+	t = kthread_create(bch2_copygc_thread, c,
+			   "bch_copygc[%s]", c->name);
 	if (IS_ERR(t))
 		return PTR_ERR(t);
 
 	get_task_struct(t);
 
-	ca->copygc_thread = t;
-	wake_up_process(ca->copygc_thread);
+	c->copygc_thread = t;
+	wake_up_process(c->copygc_thread);
 
 	return 0;
 }
 
-void bch2_dev_copygc_init(struct bch_dev *ca)
+void bch2_fs_copygc_init(struct bch_fs *c)
 {
-	bch2_pd_controller_init(&ca->copygc_pd);
-	ca->copygc_pd.d_term = 0;
+	bch2_pd_controller_init(&c->copygc_pd);
+	c->copygc_pd.d_term = 0;
 }
diff --git a/fs/bcachefs/movinggc.h b/fs/bcachefs/movinggc.h
index dcd479632cf1..922738247d03 100644
--- a/fs/bcachefs/movinggc.h
+++ b/fs/bcachefs/movinggc.h
@@ -2,8 +2,8 @@
 #ifndef _BCACHEFS_MOVINGGC_H
 #define _BCACHEFS_MOVINGGC_H
 
-void bch2_copygc_stop(struct bch_dev *);
-int bch2_copygc_start(struct bch_fs *, struct bch_dev *);
-void bch2_dev_copygc_init(struct bch_dev *);
+void bch2_copygc_stop(struct bch_fs *);
+int bch2_copygc_start(struct bch_fs *);
+void bch2_fs_copygc_init(struct bch_fs *);
 
 #endif /* _BCACHEFS_MOVINGGC_H */
diff --git a/fs/bcachefs/super.c b/fs/bcachefs/super.c
index aa78f3219a92..4cfca0d04e77 100644
--- a/fs/bcachefs/super.c
+++ b/fs/bcachefs/super.c
@@ -182,10 +182,7 @@ static void __bch2_fs_read_only(struct bch_fs *c)
 	int ret;
 
 	bch2_rebalance_stop(c);
-
-	for_each_member_device(ca, c, i)
-		bch2_copygc_stop(ca);
-
+	bch2_copygc_stop(c);
 	bch2_gc_thread_stop(c);
 
 	/*
@@ -365,8 +362,6 @@ bool bch2_fs_emergency_read_only(struct bch_fs *c)
 
 static int bch2_fs_read_write_late(struct bch_fs *c)
 {
-	struct bch_dev *ca;
-	unsigned i;
 	int ret;
 
 	ret = bch2_gc_thread_start(c);
@@ -375,13 +370,10 @@ static int bch2_fs_read_write_late(struct bch_fs *c)
 		return ret;
 	}
 
-	for_each_rw_member(ca, c, i) {
-		ret = bch2_copygc_start(c, ca);
-		if (ret) {
-			bch_err(c, "error starting copygc threads");
-			percpu_ref_put(&ca->io_ref);
-			return ret;
-		}
+	ret = bch2_copygc_start(c);
+	if (ret) {
+		bch_err(c, "error starting copygc thread");
+		return ret;
 	}
 
 	ret = bch2_rebalance_start(c);
@@ -505,6 +497,7 @@ static void bch2_fs_free(struct bch_fs *c)
 	kfree(c->replicas_gc.entries);
 	kfree(rcu_dereference_protected(c->disk_groups, 1));
 	kfree(c->journal_seq_blacklist_table);
+	free_heap(&c->copygc_heap);
 
 	if (c->journal_reclaim_wq)
 		destroy_workqueue(c->journal_reclaim_wq);
@@ -653,6 +646,7 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	for (i = 0; i < BCH_TIME_STAT_NR; i++)
 		bch2_time_stats_init(&c->times[i]);
 
+	bch2_fs_copygc_init(c);
 	bch2_fs_btree_key_cache_init_early(&c->btree_key_cache);
 	bch2_fs_allocator_background_init(c);
 	bch2_fs_allocator_foreground_init(c);
@@ -1077,10 +1071,6 @@ static struct bch_dev *__bch2_dev_alloc(struct bch_fs *c,
 
 	init_rwsem(&ca->bucket_lock);
 
-	writepoint_init(&ca->copygc_write_point, BCH_DATA_user);
-
-	bch2_dev_copygc_init(ca);
-
 	INIT_WORK(&ca->io_error_work, bch2_io_error_work);
 
 	bch2_time_stats_init(&ca->io_latency[READ]);
@@ -1321,8 +1311,6 @@ static bool bch2_fs_may_start(struct bch_fs *c)
 
 static void __bch2_dev_read_only(struct bch_fs *c, struct bch_dev *ca)
 {
-	bch2_copygc_stop(ca);
-
 	/*
 	 * The allocator thread itself allocates btree nodes, so stop it first:
 	 */
@@ -1343,9 +1331,6 @@ static const char *__bch2_dev_read_write(struct bch_fs *c, struct bch_dev *ca)
 	if (bch2_dev_allocator_start(ca))
 		return "error starting allocator thread";
 
-	if (bch2_copygc_start(c, ca))
-		return "error starting copygc thread";
-
 	return NULL;
 }
 
diff --git a/fs/bcachefs/sysfs.c b/fs/bcachefs/sysfs.c
index ac8cf6dcec3d..058e2137f0c9 100644
--- a/fs/bcachefs/sysfs.c
+++ b/fs/bcachefs/sysfs.c
@@ -379,6 +379,7 @@ SHOW(bch2_fs)
 
 	sysfs_printf(rebalance_enabled,		"%i", c->rebalance.enabled);
 	sysfs_pd_controller_show(rebalance,	&c->rebalance.pd); /* XXX */
+	sysfs_pd_controller_show(copy_gc,	&c->copygc_pd);
 
 	if (attr == &sysfs_rebalance_work)
 		return bch2_rebalance_work_show(c, buf);
@@ -460,14 +461,11 @@ STORE(bch2_fs)
 	}
 
 	if (attr == &sysfs_copy_gc_enabled) {
-		struct bch_dev *ca;
-		unsigned i;
 		ssize_t ret = strtoul_safe(buf, c->copy_gc_enabled)
 			?: (ssize_t) size;
 
-		for_each_member_device(ca, c, i)
-			if (ca->copygc_thread)
-				wake_up_process(ca->copygc_thread);
+		if (c->copygc_thread)
+			wake_up_process(c->copygc_thread);
 		return ret;
 	}
 
@@ -482,6 +480,7 @@ STORE(bch2_fs)
 	sysfs_strtoul(pd_controllers_update_seconds,
 		      c->pd_controllers_update_seconds);
 	sysfs_pd_controller_store(rebalance,	&c->rebalance.pd);
+	sysfs_pd_controller_store(copy_gc,	&c->copygc_pd);
 
 	sysfs_strtoul(promote_whole_extents,	c->promote_whole_extents);
 
@@ -607,6 +606,7 @@ struct attribute *bch2_fs_internal_files[] = {
 	&sysfs_rebalance_enabled,
 	&sysfs_rebalance_work,
 	sysfs_pd_controller_files(rebalance),
+	sysfs_pd_controller_files(copy_gc),
 
 	&sysfs_new_stripes,
 
@@ -882,7 +882,7 @@ static ssize_t show_dev_alloc_debug(struct bch_dev *ca, char *buf)
 		stats.sectors[BCH_DATA_cached],
 		stats.sectors_ec,
 		stats.sectors_fragmented,
-		ca->copygc_threshold,
+		c->copygc_threshold,
 		c->freelist_wait.list.first		? "waiting" : "empty",
 		c->open_buckets_nr_free, OPEN_BUCKETS_COUNT,
 		BTREE_NODE_OPEN_BUCKET_RESERVE,
@@ -949,8 +949,6 @@ SHOW(bch2_dev)
 		return out.pos - buf;
 	}
 
-	sysfs_pd_controller_show(copy_gc, &ca->copygc_pd);
-
 	if (attr == &sysfs_cache_replacement_policy) {
 		bch2_string_opt_to_text(&out,
 					bch2_cache_replacement_policies,
@@ -1004,8 +1002,6 @@ STORE(bch2_dev)
 	struct bch_fs *c = ca->fs;
 	struct bch_member *mi;
 
-	sysfs_pd_controller_store(copy_gc, &ca->copygc_pd);
-
 	if (attr == &sysfs_discard) {
 		bool v = strtoul_or_return(buf);
 
@@ -1090,8 +1086,6 @@ struct attribute *bch2_dev_files[] = {
 	/* debug: */
 	&sysfs_alloc_debug,
 	&sysfs_wake_allocator,
-
-	sysfs_pd_controller_files(copy_gc),
 	NULL
 };
 
diff --git a/include/trace/events/bcachefs.h b/include/trace/events/bcachefs.h
index 8923ebfc578e..ae1ad69d23cd 100644
--- a/include/trace/events/bcachefs.h
+++ b/include/trace/events/bcachefs.h
@@ -470,10 +470,10 @@ TRACE_EVENT(move_data,
 );
 
 TRACE_EVENT(copygc,
-	TP_PROTO(struct bch_dev *ca,
+	TP_PROTO(struct bch_fs *c,
 		 u64 sectors_moved, u64 sectors_not_moved,
 		 u64 buckets_moved, u64 buckets_not_moved),
-	TP_ARGS(ca,
+	TP_ARGS(c,
 		sectors_moved, sectors_not_moved,
 		buckets_moved, buckets_not_moved),
 
@@ -486,7 +486,7 @@ TRACE_EVENT(copygc,
 	),
 
 	TP_fast_assign(
-		memcpy(__entry->uuid, ca->uuid.b, 16);
+		memcpy(__entry->uuid, c->sb.user_uuid.b, 16);
 		__entry->sectors_moved		= sectors_moved;
 		__entry->sectors_not_moved	= sectors_not_moved;
 		__entry->buckets_moved		= buckets_moved;