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>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

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 @@ static int bch2_allocator_thread(void *arg)
 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);
 

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++) {

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];
 

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++)

fs/bcachefs/buckets_types.h

@@ -121,6 +121,7 @@ struct disk_reservation {
 };
 
 struct copygc_heap_entry {
+	u8			dev;
 	u8			gen;
 	u32			sectors;
 	u64			offset;

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 */

fs/bcachefs/super.c

@@ -181,10 +181,7 @@ 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);
 
 	/*
@@ -364,8 +361,6 @@ 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);
@@ -374,13 +369,10 @@ 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);
@@ -504,6 +496,7 @@ 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);
@@ -652,6 +645,7 @@ 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);
@@ -1076,10 +1070,6 @@ 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]);
@@ -1318,8 +1308,6 @@ 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:
 	 */
@@ -1340,9 +1328,6 @@ 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;
 }
 

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
 };
 

fs/bcachefs/trace.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;