bcachefs: Persist 64 bit io clocks

Originally, bcachefs - going back to bcache - stored, for each bucket, a
16 bit counter corresponding to how long it had been since the bucket
was read from. But, this required periodically rescaling counters on
every bucket to avoid wraparound. That wasn't an issue in bcache, where
we'd perodically rewrite the per bucket metadata all at once, but in
bcachefs we're trying to avoid having to walk every single bucket.

This patch switches to persisting 64 bit io clocks, corresponding to the
64 bit bucket timestaps introduced in the previous patch with
KEY_TYPE_alloc_v2.

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

19 files changed, 141 insertions, 310 deletions

diff --git a/fs/bcachefs/alloc_background.c b/fs/bcachefs/alloc_background.c
index 9a670bb2ccfb..bba83011b18b 100644
--- a/fs/bcachefs/alloc_background.c
+++ b/fs/bcachefs/alloc_background.c
@@ -31,8 +31,6 @@ static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
 #undef x
 };
 
-static void bch2_recalc_oldest_io(struct bch_fs *, struct bch_dev *, int);
-
 /* Ratelimiting/PD controllers */
 
 static void pd_controllers_update(struct work_struct *work)
@@ -340,9 +338,7 @@ static int bch2_alloc_read_fn(struct bch_fs *c, enum btree_id id,
 
 int bch2_alloc_read(struct bch_fs *c, struct journal_keys *journal_keys)
 {
-	struct bch_dev *ca;
-	unsigned i;
-	int ret = 0;
+	int ret;
 
 	down_read(&c->gc_lock);
 	ret = bch2_btree_and_journal_walk(c, journal_keys, BTREE_ID_ALLOC,
@@ -358,22 +354,6 @@ int bch2_alloc_read(struct bch_fs *c, struct journal_keys *journal_keys)
 	bch2_dev_usage_from_buckets(c);
 	percpu_up_write(&c->mark_lock);
 
-	mutex_lock(&c->bucket_clock[READ].lock);
-	for_each_member_device(ca, c, i) {
-		down_read(&ca->bucket_lock);
-		bch2_recalc_oldest_io(c, ca, READ);
-		up_read(&ca->bucket_lock);
-	}
-	mutex_unlock(&c->bucket_clock[READ].lock);
-
-	mutex_lock(&c->bucket_clock[WRITE].lock);
-	for_each_member_device(ca, c, i) {
-		down_read(&ca->bucket_lock);
-		bch2_recalc_oldest_io(c, ca, WRITE);
-		up_read(&ca->bucket_lock);
-	}
-	mutex_unlock(&c->bucket_clock[WRITE].lock);
-
 	return 0;
 }
 
@@ -460,114 +440,6 @@ err:
 
 /* Bucket IO clocks: */
 
-static void bch2_recalc_oldest_io(struct bch_fs *c, struct bch_dev *ca, int rw)
-{
-	struct bucket_clock *clock = &c->bucket_clock[rw];
-	struct bucket_array *buckets = bucket_array(ca);
-	struct bucket *g;
-	u16 max_last_io = 0;
-	unsigned i;
-
-	lockdep_assert_held(&c->bucket_clock[rw].lock);
-
-	/* Recalculate max_last_io for this device: */
-	for_each_bucket(g, buckets)
-		max_last_io = max(max_last_io, bucket_last_io(c, g, rw));
-
-	ca->max_last_bucket_io[rw] = max_last_io;
-
-	/* Recalculate global max_last_io: */
-	max_last_io = 0;
-
-	for_each_member_device(ca, c, i)
-		max_last_io = max(max_last_io, ca->max_last_bucket_io[rw]);
-
-	clock->max_last_io = max_last_io;
-}
-
-static void bch2_rescale_bucket_io_times(struct bch_fs *c, int rw)
-{
-	struct bucket_clock *clock = &c->bucket_clock[rw];
-	struct bucket_array *buckets;
-	struct bch_dev *ca;
-	struct bucket *g;
-	unsigned i;
-
-	trace_rescale_prios(c);
-
-	for_each_member_device(ca, c, i) {
-		down_read(&ca->bucket_lock);
-		buckets = bucket_array(ca);
-
-		for_each_bucket(g, buckets)
-			g->io_time[rw] = clock->hand -
-			bucket_last_io(c, g, rw) / 2;
-
-		bch2_recalc_oldest_io(c, ca, rw);
-
-		up_read(&ca->bucket_lock);
-	}
-}
-
-static inline u64 bucket_clock_freq(u64 capacity)
-{
-	return max(capacity >> 10, 2028ULL);
-}
-
-static void bch2_inc_clock_hand(struct io_timer *timer)
-{
-	struct bucket_clock *clock = container_of(timer,
-						struct bucket_clock, rescale);
-	struct bch_fs *c = container_of(clock,
-					struct bch_fs, bucket_clock[clock->rw]);
-	struct bch_dev *ca;
-	u64 capacity;
-	unsigned i;
-
-	mutex_lock(&clock->lock);
-
-	/* if clock cannot be advanced more, rescale prio */
-	if (clock->max_last_io >= U16_MAX - 2)
-		bch2_rescale_bucket_io_times(c, clock->rw);
-
-	BUG_ON(clock->max_last_io >= U16_MAX - 2);
-
-	for_each_member_device(ca, c, i)
-		ca->max_last_bucket_io[clock->rw]++;
-	clock->max_last_io++;
-	clock->hand++;
-
-	mutex_unlock(&clock->lock);
-
-	capacity = READ_ONCE(c->capacity);
-
-	if (!capacity)
-		return;
-
-	/*
-	 * we only increment when 0.1% of the filesystem capacity has been read
-	 * or written too, this determines if it's time
-	 *
-	 * XXX: we shouldn't really be going off of the capacity of devices in
-	 * RW mode (that will be 0 when we're RO, yet we can still service
-	 * reads)
-	 */
-	timer->expire += bucket_clock_freq(capacity);
-
-	bch2_io_timer_add(&c->io_clock[clock->rw], timer);
-}
-
-static void bch2_bucket_clock_init(struct bch_fs *c, int rw)
-{
-	struct bucket_clock *clock = &c->bucket_clock[rw];
-
-	clock->hand		= 1;
-	clock->rw		= rw;
-	clock->rescale.fn	= bch2_inc_clock_hand;
-	clock->rescale.expire	= bucket_clock_freq(c->capacity);
-	mutex_init(&clock->lock);
-}
-
 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
 			      size_t bucket_nr, int rw)
 {
@@ -577,7 +449,7 @@ int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
 	struct bucket *g;
 	struct bkey_alloc_buf *a;
 	struct bkey_alloc_unpacked u;
-	u64 *time;
+	u64 *time, now;
 	int ret = 0;
 
 	iter = bch2_trans_get_iter(trans, BTREE_ID_ALLOC, POS(dev, bucket_nr),
@@ -599,10 +471,11 @@ int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
 	percpu_up_read(&c->mark_lock);
 
 	time = rw == READ ? &u.read_time : &u.write_time;
-	if (*time == c->bucket_clock[rw].hand)
+	now = atomic64_read(&c->io_clock[rw].now);
+	if (*time == now)
 		goto out;
 
-	*time = c->bucket_clock[rw].hand;
+	*time = now;
 
 	bch2_alloc_pack(c, a, u);
 	ret   = bch2_trans_update(trans, iter, &a->k, 0) ?:
@@ -674,23 +547,22 @@ static int wait_buckets_available(struct bch_fs *c, struct bch_dev *ca)
 	return ret;
 }
 
-static bool bch2_can_invalidate_bucket(struct bch_dev *ca,
-				       size_t bucket,
-				       struct bucket_mark mark)
+static bool bch2_can_invalidate_bucket(struct bch_dev *ca, size_t b,
+				       struct bucket_mark m)
 {
 	u8 gc_gen;
 
-	if (!is_available_bucket(mark))
+	if (!is_available_bucket(m))
 		return false;
 
-	if (mark.owned_by_allocator)
+	if (m.owned_by_allocator)
 		return false;
 
 	if (ca->buckets_nouse &&
-	    test_bit(bucket, ca->buckets_nouse))
+	    test_bit(b, ca->buckets_nouse))
 		return false;
 
-	gc_gen = bucket_gc_gen(ca, bucket);
+	gc_gen = bucket_gc_gen(bucket(ca, b));
 
 	if (gc_gen >= BUCKET_GC_GEN_MAX / 2)
 		ca->inc_gen_needs_gc++;
@@ -704,43 +576,33 @@ static bool bch2_can_invalidate_bucket(struct bch_dev *ca,
 /*
  * Determines what order we're going to reuse buckets, smallest bucket_key()
  * first.
- *
- *
- * - We take into account the read prio of the bucket, which gives us an
- *   indication of how hot the data is -- we scale the prio so that the prio
- *   farthest from the clock is worth 1/8th of the closest.
- *
- * - The number of sectors of cached data in the bucket, which gives us an
- *   indication of the cost in cache misses this eviction will cause.
- *
- * - If hotness * sectors used compares equal, we pick the bucket with the
- *   smallest bucket_gc_gen() - since incrementing the same bucket's generation
- *   number repeatedly forces us to run mark and sweep gc to avoid generation
- *   number wraparound.
  */
 
-static unsigned long bucket_sort_key(struct bch_fs *c, struct bch_dev *ca,
-				     size_t b, struct bucket_mark m)
+static unsigned bucket_sort_key(struct bucket *g, struct bucket_mark m,
+				u64 now, u64 last_seq_ondisk)
 {
-	unsigned last_io = bucket_last_io(c, bucket(ca, b), READ);
-	unsigned max_last_io = ca->max_last_bucket_io[READ];
-
-	/*
-	 * Time since last read, scaled to [0, 8) where larger value indicates
-	 * more recently read data:
-	 */
-	unsigned long hotness = (max_last_io - last_io) * 7 / max_last_io;
-
-	/* How much we want to keep the data in this bucket: */
-	unsigned long data_wantness =
-		(hotness + 1) * bucket_sectors_used(m);
+	unsigned used = bucket_sectors_used(m);
 
-	unsigned long needs_journal_commit =
-		bucket_needs_journal_commit(m, c->journal.last_seq_ondisk);
+	if (used) {
+		/*
+		 * Prefer to keep buckets that have been read more recently, and
+		 * buckets that have more data in them:
+		 */
+		u64 last_read = max_t(s64, 0, now - g->io_time[READ]);
+		u32 last_read_scaled = max_t(u64, U32_MAX, div_u64(last_read, used));
 
-	return  (data_wantness << 9) |
-		(needs_journal_commit << 8) |
-		(bucket_gc_gen(ca, b) / 16);
+		return -last_read_scaled;
+	} else {
+		/*
+		 * Prefer to use buckets with smaller gc_gen so that we don't
+		 * have to walk the btree and recalculate oldest_gen - but shift
+		 * off the low bits so that buckets will still have equal sort
+		 * keys when there's only a small difference, so that we can
+		 * keep sequential buckets together:
+		 */
+		return  (bucket_needs_journal_commit(m, last_seq_ondisk) << 4)|
+			(bucket_gc_gen(g) >> 4);
+	}
 }
 
 static inline int bucket_alloc_cmp(alloc_heap *h,
@@ -763,16 +625,15 @@ static void find_reclaimable_buckets_lru(struct bch_fs *c, struct bch_dev *ca)
 {
 	struct bucket_array *buckets;
 	struct alloc_heap_entry e = { 0 };
+	u64 now, last_seq_ondisk;
 	size_t b, i, nr = 0;
 
-	ca->alloc_heap.used = 0;
-
-	mutex_lock(&c->bucket_clock[READ].lock);
 	down_read(&ca->bucket_lock);
 
 	buckets = bucket_array(ca);
-
-	bch2_recalc_oldest_io(c, ca, READ);
+	ca->alloc_heap.used = 0;
+	now = atomic64_read(&c->io_clock[READ].now);
+	last_seq_ondisk = c->journal.last_seq_ondisk;
 
 	/*
 	 * Find buckets with lowest read priority, by building a maxheap sorted
@@ -780,8 +641,9 @@ static void find_reclaimable_buckets_lru(struct bch_fs *c, struct bch_dev *ca)
 	 * all buckets have been visited.
 	 */
 	for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++) {
-		struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
-		unsigned long key = bucket_sort_key(c, ca, b, m);
+		struct bucket *g = &buckets->b[b];
+		struct bucket_mark m = READ_ONCE(g->mark);
+		unsigned key = bucket_sort_key(g, m, now, last_seq_ondisk);
 
 		if (!bch2_can_invalidate_bucket(ca, b, m))
 			continue;
@@ -816,7 +678,6 @@ static void find_reclaimable_buckets_lru(struct bch_fs *c, struct bch_dev *ca)
 	}
 
 	up_read(&ca->bucket_lock);
-	mutex_unlock(&c->bucket_clock[READ].lock);
 }
 
 static void find_reclaimable_buckets_fifo(struct bch_fs *c, struct bch_dev *ca)
@@ -1031,8 +892,8 @@ retry:
 	u.data_type	= 0;
 	u.dirty_sectors	= 0;
 	u.cached_sectors = 0;
-	u.read_time	= c->bucket_clock[READ].hand;
-	u.write_time	= c->bucket_clock[WRITE].hand;
+	u.read_time	= atomic64_read(&c->io_clock[READ].now);
+	u.write_time	= atomic64_read(&c->io_clock[WRITE].now);
 
 	bch2_alloc_pack(c, &a, u);
 	bch2_trans_update(trans, iter, &a.k,
@@ -1542,8 +1403,6 @@ int bch2_dev_allocator_start(struct bch_dev *ca)
 void bch2_fs_allocator_background_init(struct bch_fs *c)
 {
 	spin_lock_init(&c->freelist_lock);
-	bch2_bucket_clock_init(c, READ);
-	bch2_bucket_clock_init(c, WRITE);
 
 	c->pd_controllers_update_seconds = 5;
 	INIT_DELAYED_WORK(&c->pd_controllers_update, pd_controllers_update);
diff --git a/fs/bcachefs/alloc_types.h b/fs/bcachefs/alloc_types.h
index 1abfff5290bc..be164d6108bb 100644
--- a/fs/bcachefs/alloc_types.h
+++ b/fs/bcachefs/alloc_types.h
@@ -10,30 +10,6 @@
 
 struct ec_bucket_buf;
 
-/* There's two of these clocks, one for reads and one for writes: */
-struct bucket_clock {
-	/*
-	 * "now" in (read/write) IO time - incremented whenever we do X amount
-	 * of reads or writes.
-	 *
-	 * Goes with the bucket read/write prios: when we read or write to a
-	 * bucket we reset the bucket's prio to the current hand; thus hand -
-	 * prio = time since bucket was last read/written.
-	 *
-	 * The units are some amount (bytes/sectors) of data read/written, and
-	 * the units can change on the fly if we need to rescale to fit
-	 * everything in a u16 - your only guarantee is that the units are
-	 * consistent.
-	 */
-	u16			hand;
-	u16			max_last_io;
-
-	int			rw;
-
-	struct io_timer		rescale;
-	struct mutex		lock;
-};
-
 enum alloc_reserve {
 	RESERVE_BTREE_MOVINGGC	= -2,
 	RESERVE_BTREE		= -1,
diff --git a/fs/bcachefs/bcachefs.h b/fs/bcachefs/bcachefs.h
index bd675b88b354..763cac0efa0c 100644
--- a/fs/bcachefs/bcachefs.h
+++ b/fs/bcachefs/bcachefs.h
@@ -451,9 +451,6 @@ struct bch_dev {
 
 	size_t			fifo_last_bucket;
 
-	/* last calculated minimum prio */
-	u16			max_last_bucket_io[2];
-
 	size_t			inc_gen_needs_gc;
 	size_t			inc_gen_really_needs_gc;
 
@@ -693,14 +690,6 @@ struct bch_fs {
 	struct mutex		usage_scratch_lock;
 	struct bch_fs_usage_online *usage_scratch;
 
-	/*
-	 * When we invalidate buckets, we use both the priority and the amount
-	 * of good data to determine which buckets to reuse first - to weight
-	 * those together consistently we keep track of the smallest nonzero
-	 * priority of any bucket.
-	 */
-	struct bucket_clock	bucket_clock[2];
-
 	struct io_clock		io_clock[2];
 
 	/* JOURNAL SEQ BLACKLIST */
diff --git a/fs/bcachefs/bcachefs_format.h b/fs/bcachefs/bcachefs_format.h
index b6c7e57b6bcd..5dab5bfd228a 100644
--- a/fs/bcachefs/bcachefs_format.h
+++ b/fs/bcachefs/bcachefs_format.h
@@ -1143,8 +1143,8 @@ struct bch_sb_field_clean {
 	struct bch_sb_field	field;
 
 	__le32			flags;
-	__le16			read_clock;
-	__le16			write_clock;
+	__le16			_read_clock; /* no longer used */
+	__le16			_write_clock;
 	__le64			journal_seq;
 
 	union {
@@ -1511,7 +1511,8 @@ static inline __u64 __bset_magic(struct bch_sb *sb)
 	x(blacklist,		3)		\
 	x(blacklist_v2,		4)		\
 	x(usage,		5)		\
-	x(data_usage,		6)
+	x(data_usage,		6)		\
+	x(clock,		7)
 
 enum {
 #define x(f, nr)	BCH_JSET_ENTRY_##f	= nr,
@@ -1559,6 +1560,13 @@ struct jset_entry_data_usage {
 	struct bch_replicas_entry r;
 } __attribute__((packed));
 
+struct jset_entry_clock {
+	struct jset_entry	entry;
+	__u8			rw;
+	__u8			pad[7];
+	__le64			time;
+} __attribute__((packed));
+
 /*
  * On disk format for a journal entry:
  * seq is monotonically increasing; every journal entry has its own unique
@@ -1581,8 +1589,8 @@ struct jset {
 
 	__u8			encrypted_start[0];
 
-	__le16			read_clock;
-	__le16			write_clock;
+	__le16			_read_clock; /* no longer used */
+	__le16			_write_clock;
 
 	/* Sequence number of oldest dirty journal entry */
 	__le64			last_seq;
diff --git a/fs/bcachefs/btree_gc.c b/fs/bcachefs/btree_gc.c
index 9e123736a125..5ea9bae09d59 100644
--- a/fs/bcachefs/btree_gc.c
+++ b/fs/bcachefs/btree_gc.c
@@ -1489,7 +1489,7 @@ static int bch2_gc_thread(void *arg)
 {
 	struct bch_fs *c = arg;
 	struct io_clock *clock = &c->io_clock[WRITE];
-	unsigned long last = atomic_long_read(&clock->now);
+	unsigned long last = atomic64_read(&clock->now);
 	unsigned last_kick = atomic_read(&c->kick_gc);
 	int ret;
 
@@ -1510,7 +1510,7 @@ static int bch2_gc_thread(void *arg)
 			if (c->btree_gc_periodic) {
 				unsigned long next = last + c->capacity / 16;
 
-				if (atomic_long_read(&clock->now) >= next)
+				if (atomic64_read(&clock->now) >= next)
 					break;
 
 				bch2_io_clock_schedule_timeout(clock, next);
@@ -1522,7 +1522,7 @@ static int bch2_gc_thread(void *arg)
 		}
 		__set_current_state(TASK_RUNNING);
 
-		last = atomic_long_read(&clock->now);
+		last = atomic64_read(&clock->now);
 		last_kick = atomic_read(&c->kick_gc);
 
 		/*
diff --git a/fs/bcachefs/buckets.h b/fs/bcachefs/buckets.h
index 4103ea7e769a..50989d286190 100644
--- a/fs/bcachefs/buckets.h
+++ b/fs/bcachefs/buckets.h
@@ -58,20 +58,13 @@ static inline struct bucket *bucket(struct bch_dev *ca, size_t b)
 	return __bucket(ca, b, false);
 }
 
-static inline u16 bucket_last_io(struct bch_fs *c, struct bucket *g, int rw)
-{
-	return c->bucket_clock[rw].hand - g->io_time[rw];
-}
-
 /*
  * bucket_gc_gen() returns the difference between the bucket's current gen and
  * the oldest gen of any pointer into that bucket in the btree.
  */
 
-static inline u8 bucket_gc_gen(struct bch_dev *ca, size_t b)
+static inline u8 bucket_gc_gen(struct bucket *g)
 {
-	struct bucket *g = bucket(ca, b);
-
 	return g->mark.gen - g->oldest_gen;
 }
 
diff --git a/fs/bcachefs/buckets_types.h b/fs/bcachefs/buckets_types.h
index 99ab9f48ba9d..b6ea67506cc2 100644
--- a/fs/bcachefs/buckets_types.h
+++ b/fs/bcachefs/buckets_types.h
@@ -37,7 +37,7 @@ struct bucket {
 		const struct bucket_mark mark;
 	};
 
-	u16				io_time[2];
+	u64				io_time[2];
 	u8				oldest_gen;
 	u8				gc_gen;
 	unsigned			gen_valid:1;
diff --git a/fs/bcachefs/clock.c b/fs/bcachefs/clock.c
index 869ba1887757..da91c95e3ffc 100644
--- a/fs/bcachefs/clock.c
+++ b/fs/bcachefs/clock.c
@@ -19,7 +19,7 @@ void bch2_io_timer_add(struct io_clock *clock, struct io_timer *timer)
 
 	spin_lock(&clock->timer_lock);
 
-	if (time_after_eq((unsigned long) atomic_long_read(&clock->now),
+	if (time_after_eq((unsigned long) atomic64_read(&clock->now),
 			  timer->expire)) {
 		spin_unlock(&clock->timer_lock);
 		timer->fn(timer);
@@ -146,7 +146,7 @@ static struct io_timer *get_expired_timer(struct io_clock *clock,
 void __bch2_increment_clock(struct io_clock *clock, unsigned sectors)
 {
 	struct io_timer *timer;
-	unsigned long now = atomic_long_add_return(sectors, &clock->now);
+	unsigned long now = atomic64_add_return(sectors, &clock->now);
 
 	while ((timer = get_expired_timer(clock, now)))
 		timer->fn(timer);
@@ -158,7 +158,7 @@ void bch2_io_timers_to_text(struct printbuf *out, struct io_clock *clock)
 	unsigned i;
 
 	spin_lock(&clock->timer_lock);
-	now = atomic_long_read(&clock->now);
+	now = atomic64_read(&clock->now);
 
 	for (i = 0; i < clock->timers.used; i++)
 		pr_buf(out, "%ps:\t%li\n",
@@ -175,7 +175,7 @@ void bch2_io_clock_exit(struct io_clock *clock)
 
 int bch2_io_clock_init(struct io_clock *clock)
 {
-	atomic_long_set(&clock->now, 0);
+	atomic64_set(&clock->now, 0);
 	spin_lock_init(&clock->timer_lock);
 
 	clock->max_slop = IO_CLOCK_PCPU_SECTORS * num_possible_cpus();
diff --git a/fs/bcachefs/clock_types.h b/fs/bcachefs/clock_types.h
index 92c740a47565..5fae0012d808 100644
--- a/fs/bcachefs/clock_types.h
+++ b/fs/bcachefs/clock_types.h
@@ -26,7 +26,7 @@ struct io_timer {
 typedef HEAP(struct io_timer *)	io_timer_heap;
 
 struct io_clock {
-	atomic_long_t		now;
+	atomic64_t		now;
 	u16 __percpu		*pcpu_buf;
 	unsigned		max_slop;
 
diff --git a/fs/bcachefs/journal.c b/fs/bcachefs/journal.c
index ba37c78c01db..379b9ad2c0f9 100644
--- a/fs/bcachefs/journal.c
+++ b/fs/bcachefs/journal.c
@@ -1123,6 +1123,9 @@ int bch2_fs_journal_init(struct journal *j)
 	j->entry_u64s_reserved +=
 		BTREE_ID_NR * (JSET_KEYS_U64s + BKEY_BTREE_PTR_U64s_MAX);
 
+	j->entry_u64s_reserved +=
+		2 * (sizeof(struct jset_entry_clock) / sizeof(u64));
+
 	atomic64_set(&j->reservations.counter,
 		((union journal_res_state)
 		 { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
diff --git a/fs/bcachefs/journal_io.c b/fs/bcachefs/journal_io.c
index 7e726db77881..a82548983dbd 100644
--- a/fs/bcachefs/journal_io.c
+++ b/fs/bcachefs/journal_io.c
@@ -426,6 +426,32 @@ fsck_err:
 	return ret;
 }
 
+static int journal_entry_validate_clock(struct bch_fs *c,
+					struct jset *jset,
+					struct jset_entry *entry,
+					int write)
+{
+	struct jset_entry_clock *clock =
+		container_of(entry, struct jset_entry_clock, entry);
+	unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
+	int ret = 0;
+
+	if (journal_entry_err_on(bytes != sizeof(*clock),
+				 c, "invalid journal entry clock: bad size")) {
+		journal_entry_null_range(entry, vstruct_next(entry));
+		return ret;
+	}
+
+	if (journal_entry_err_on(clock->rw > 1,
+				 c, "invalid journal entry clock: bad rw")) {
+		journal_entry_null_range(entry, vstruct_next(entry));
+		return ret;
+	}
+
+fsck_err:
+	return ret;
+}
+
 struct jset_entry_ops {
 	int (*validate)(struct bch_fs *, struct jset *,
 			struct jset_entry *, int);
@@ -1361,8 +1387,8 @@ void bch2_journal_write(struct closure *cl)
 
 	end	= bch2_btree_roots_to_journal_entries(c, jset->start, end);
 
-	end	= bch2_journal_super_entries_add_common(c, end,
-						le64_to_cpu(jset->seq));
+	bch2_journal_super_entries_add_common(c, &end,
+				le64_to_cpu(jset->seq));
 	u64s	= (u64 *) end - (u64 *) start;
 	BUG_ON(u64s > j->entry_u64s_reserved);
 
@@ -1371,10 +1397,7 @@ void bch2_journal_write(struct closure *cl)
 
 	journal_write_compact(jset);
 
-	jset->read_clock	= cpu_to_le16(c->bucket_clock[READ].hand);
-	jset->write_clock	= cpu_to_le16(c->bucket_clock[WRITE].hand);
 	jset->magic		= cpu_to_le64(jset_magic(c));
-
 	jset->version		= c->sb.version < bcachefs_metadata_version_new_versioning
 		? cpu_to_le32(BCH_JSET_VERSION_OLD)
 		: cpu_to_le32(c->sb.version);
diff --git a/fs/bcachefs/movinggc.c b/fs/bcachefs/movinggc.c
index 8e6e4cd73886..e2472c19beaf 100644
--- a/fs/bcachefs/movinggc.c
+++ b/fs/bcachefs/movinggc.c
@@ -298,7 +298,7 @@ static int bch2_copygc_thread(void *arg)
 {
 	struct bch_fs *c = arg;
 	struct io_clock *clock = &c->io_clock[WRITE];
-	unsigned long last, wait;
+	u64 last, wait;
 
 	set_freezable();
 
@@ -306,7 +306,7 @@ static int bch2_copygc_thread(void *arg)
 		if (kthread_wait_freezable(c->copy_gc_enabled))
 			break;
 
-		last = atomic_long_read(&clock->now);
+		last = atomic64_read(&clock->now);
 		wait = bch2_copygc_wait_amount(c);
 
 		if (wait > clock->max_slop) {
diff --git a/fs/bcachefs/rebalance.c b/fs/bcachefs/rebalance.c
index f9a12dd797a5..2263ee41c444 100644
--- a/fs/bcachefs/rebalance.c
+++ b/fs/bcachefs/rebalance.c
@@ -169,12 +169,12 @@ static int bch2_rebalance_thread(void *arg)
 	unsigned long start, prev_start;
 	unsigned long prev_run_time, prev_run_cputime;
 	unsigned long cputime, prev_cputime;
-	unsigned long io_start;
+	u64 io_start;
 	long throttle;
 
 	set_freezable();
 
-	io_start	= atomic_long_read(&clock->now);
+	io_start	= atomic64_read(&clock->now);
 	p		= rebalance_work(c);
 	prev_start	= jiffies;
 	prev_cputime	= curr_cputime();
@@ -210,7 +210,7 @@ static int bch2_rebalance_thread(void *arg)
 					(20 - w.dev_most_full_percent),
 					50);
 
-			if (atomic_long_read(&clock->now) + clock->max_slop <
+			if (atomic64_read(&clock->now) + clock->max_slop <
 			    r->throttled_until_iotime) {
 				r->throttled_until_cputime = start + throttle;
 				r->state = REBALANCE_THROTTLED;
@@ -229,7 +229,7 @@ static int bch2_rebalance_thread(void *arg)
 			      max(p.dev_most_full_percent, 1U) /
 			      max(w.dev_most_full_percent, 1U));
 
-		io_start	= atomic_long_read(&clock->now);
+		io_start	= atomic64_read(&clock->now);
 		p		= w;
 		prev_start	= start;
 		prev_cputime	= cputime;
@@ -274,7 +274,7 @@ void bch2_rebalance_work_to_text(struct printbuf *out, struct bch_fs *c)
 	case REBALANCE_THROTTLED:
 		bch2_hprint(&PBUF(h1),
 			    (r->throttled_until_iotime -
-			     atomic_long_read(&c->io_clock[WRITE].now)) << 9);
+			     atomic64_read(&c->io_clock[WRITE].now)) << 9);
 		pr_buf(out, "throttled for %lu sec or %s io\n",
 		       (r->throttled_until_cputime - jiffies) / HZ,
 		       h1);
diff --git a/fs/bcachefs/rebalance_types.h b/fs/bcachefs/rebalance_types.h
index 192c6be20ced..2f62a643c39f 100644
--- a/fs/bcachefs/rebalance_types.h
+++ b/fs/bcachefs/rebalance_types.h
@@ -17,7 +17,7 @@ struct bch_fs_rebalance {
 	atomic64_t		work_unknown_dev;
 
 	enum rebalance_state	state;
-	unsigned long		throttled_until_iotime;
+	u64			throttled_until_iotime;
 	unsigned long		throttled_until_cputime;
 	struct bch_move_stats	move_stats;
 
diff --git a/fs/bcachefs/recovery.c b/fs/bcachefs/recovery.c
index f470e0e233ce..55f7771e11c8 100644
--- a/fs/bcachefs/recovery.c
+++ b/fs/bcachefs/recovery.c
@@ -847,6 +847,12 @@ static int journal_replay_entry_early(struct bch_fs *c,
 				le64_to_cpu(bl_entry->end) + 1);
 		break;
 	}
+	case BCH_JSET_ENTRY_clock: {
+		struct jset_entry_clock *clock =
+			container_of(entry, struct jset_entry_clock, entry);
+
+		atomic64_set(&c->io_clock[clock->rw].now, clock->time);
+	}
 	}
 
 	return ret;
@@ -861,9 +867,6 @@ static int journal_replay_early(struct bch_fs *c,
 	int ret;
 
 	if (clean) {
-		c->bucket_clock[READ].hand = le16_to_cpu(clean->read_clock);
-		c->bucket_clock[WRITE].hand = le16_to_cpu(clean->write_clock);
-
 		for (entry = clean->start;
 		     entry != vstruct_end(&clean->field);
 		     entry = vstruct_next(entry)) {
@@ -876,9 +879,6 @@ static int journal_replay_early(struct bch_fs *c,
 			if (i->ignore)
 				continue;
 
-			c->bucket_clock[READ].hand = le16_to_cpu(i->j.read_clock);
-			c->bucket_clock[WRITE].hand = le16_to_cpu(i->j.write_clock);
-
 			vstruct_for_each(&i->j, entry) {
 				ret = journal_replay_entry_early(c, entry);
 				if (ret)
@@ -942,13 +942,6 @@ static int verify_superblock_clean(struct bch_fs *c,
 		return 0;
 	}
 
-	mustfix_fsck_err_on(j->read_clock != clean->read_clock, c,
-			"superblock read clock %u doesn't match journal %u after clean shutdown",
-			clean->read_clock, j->read_clock);
-	mustfix_fsck_err_on(j->write_clock != clean->write_clock, c,
-			"superblock write clock %u doesn't match journal %u after clean shutdown",
-			clean->write_clock, j->write_clock);
-
 	for (i = 0; i < BTREE_ID_NR; i++) {
 		char buf1[200], buf2[200];
 		struct bkey_i *k1, *k2;
diff --git a/fs/bcachefs/super-io.c b/fs/bcachefs/super-io.c
index 61b947313c88..3b082da934fb 100644
--- a/fs/bcachefs/super-io.c
+++ b/fs/bcachefs/super-io.c
@@ -966,29 +966,25 @@ int bch2_fs_mark_dirty(struct bch_fs *c)
 	return ret;
 }
 
-static void
-entry_init_u64s(struct jset_entry *entry, unsigned u64s)
+static struct jset_entry *jset_entry_init(struct jset_entry **end, size_t size)
 {
-	memset(entry, 0, u64s * sizeof(u64));
+	struct jset_entry *entry = *end;
+	unsigned u64s = DIV_ROUND_UP(size, sizeof(u64));
 
+	memset(entry, 0, u64s * sizeof(u64));
 	/*
 	 * The u64s field counts from the start of data, ignoring the shared
 	 * fields.
 	 */
 	entry->u64s = u64s - 1;
-}
 
-static void
-entry_init_size(struct jset_entry *entry, size_t size)
-{
-	unsigned u64s = DIV_ROUND_UP(size, sizeof(u64));
-	entry_init_u64s(entry, u64s);
+	*end = vstruct_next(*end);
+	return entry;
 }
 
-struct jset_entry *
-bch2_journal_super_entries_add_common(struct bch_fs *c,
-				      struct jset_entry *entry,
-				      u64 journal_seq)
+void bch2_journal_super_entries_add_common(struct bch_fs *c,
+					   struct jset_entry **end,
+					   u64 journal_seq)
 {
 	unsigned i;
 
@@ -1003,59 +999,59 @@ bch2_journal_super_entries_add_common(struct bch_fs *c,
 
 	{
 		struct jset_entry_usage *u =
-			container_of(entry, struct jset_entry_usage, entry);
+			container_of(jset_entry_init(end, sizeof(*u)),
+				     struct jset_entry_usage, entry);
 
-		entry_init_size(entry, sizeof(*u));
 		u->entry.type	= BCH_JSET_ENTRY_usage;
 		u->entry.btree_id = FS_USAGE_INODES;
 		u->v		= cpu_to_le64(c->usage_base->nr_inodes);
-
-		entry = vstruct_next(entry);
 	}
 
 	{
 		struct jset_entry_usage *u =
-			container_of(entry, struct jset_entry_usage, entry);
+			container_of(jset_entry_init(end, sizeof(*u)),
+				     struct jset_entry_usage, entry);
 
-		entry_init_size(entry, sizeof(*u));
 		u->entry.type	= BCH_JSET_ENTRY_usage;
 		u->entry.btree_id = FS_USAGE_KEY_VERSION;
 		u->v		= cpu_to_le64(atomic64_read(&c->key_version));
-
-		entry = vstruct_next(entry);
 	}
 
 	for (i = 0; i < BCH_REPLICAS_MAX; i++) {
 		struct jset_entry_usage *u =
-			container_of(entry, struct jset_entry_usage, entry);
+			container_of(jset_entry_init(end, sizeof(*u)),
+				     struct jset_entry_usage, entry);
 
-		entry_init_size(entry, sizeof(*u));
 		u->entry.type	= BCH_JSET_ENTRY_usage;
 		u->entry.btree_id = FS_USAGE_RESERVED;
 		u->entry.level	= i;
 		u->v		= cpu_to_le64(c->usage_base->persistent_reserved[i]);
-
-		entry = vstruct_next(entry);
 	}
 
 	for (i = 0; i < c->replicas.nr; i++) {
 		struct bch_replicas_entry *e =
 			cpu_replicas_entry(&c->replicas, i);
 		struct jset_entry_data_usage *u =
-			container_of(entry, struct jset_entry_data_usage, entry);
+			container_of(jset_entry_init(end, sizeof(*u) + e->nr_devs),
+				     struct jset_entry_data_usage, entry);
 
-		entry_init_size(entry, sizeof(*u) + e->nr_devs);
 		u->entry.type	= BCH_JSET_ENTRY_data_usage;
 		u->v		= cpu_to_le64(c->usage_base->replicas[i]);
 		unsafe_memcpy(&u->r, e, replicas_entry_bytes(e),
 			      "embedded variable length struct");
-
-		entry = vstruct_next(entry);
 	}
 
 	percpu_up_read(&c->mark_lock);
 
-	return entry;
+	for (i = 0; i < 2; i++) {
+		struct jset_entry_clock *clock =
+			container_of(jset_entry_init(end, sizeof(*clock)),
+				     struct jset_entry_clock, entry);
+
+		clock->entry.type = BCH_JSET_ENTRY_clock;
+		clock->rw	= i;
+		clock->time	= atomic64_read(&c->io_clock[i].now);
+	}
 }
 
 void bch2_fs_mark_clean(struct bch_fs *c)
@@ -1084,15 +1080,13 @@ void bch2_fs_mark_clean(struct bch_fs *c)
 	}
 
 	sb_clean->flags		= 0;
-	sb_clean->read_clock	= cpu_to_le16(c->bucket_clock[READ].hand);
-	sb_clean->write_clock	= cpu_to_le16(c->bucket_clock[WRITE].hand);
 	sb_clean->journal_seq	= cpu_to_le64(journal_cur_seq(&c->journal) - 1);
 
 	/* Trying to catch outstanding bug: */
 	BUG_ON(le64_to_cpu(sb_clean->journal_seq) > S64_MAX);
 
 	entry = sb_clean->start;
-	entry = bch2_journal_super_entries_add_common(c, entry, 0);
+	bch2_journal_super_entries_add_common(c, &entry, 0);
 	entry = bch2_btree_roots_to_journal_entries(c, entry, entry);
 	BUG_ON((void *) entry > vstruct_end(&sb_clean->field));
 
diff --git a/fs/bcachefs/super-io.h b/fs/bcachefs/super-io.h
index 402ae563b3c7..dd8d4ba911f0 100644
--- a/fs/bcachefs/super-io.h
+++ b/fs/bcachefs/super-io.h
@@ -122,9 +122,8 @@ static inline struct bch_member_cpu bch2_mi_to_cpu(struct bch_member *mi)
 
 /* BCH_SB_FIELD_clean: */
 
-struct jset_entry *
-bch2_journal_super_entries_add_common(struct bch_fs *,
-				      struct jset_entry *, u64);
+void bch2_journal_super_entries_add_common(struct bch_fs *,
+					   struct jset_entry **, u64);
 
 void bch2_sb_clean_renumber(struct bch_sb_field_clean *, int);
 
diff --git a/fs/bcachefs/super.c b/fs/bcachefs/super.c
index d451a29b517b..5f5893ab9edf 100644
--- a/fs/bcachefs/super.c
+++ b/fs/bcachefs/super.c
@@ -181,9 +181,6 @@ static void __bch2_fs_read_only(struct bch_fs *c)
 	bch2_copygc_stop(c);
 	bch2_gc_thread_stop(c);
 
-	bch2_io_timer_del(&c->io_clock[READ], &c->bucket_clock[READ].rescale);
-	bch2_io_timer_del(&c->io_clock[WRITE], &c->bucket_clock[WRITE].rescale);
-
 	/*
 	 * Flush journal before stopping allocators, because flushing journal
 	 * blacklist entries involves allocating new btree nodes:
@@ -406,9 +403,6 @@ static int __bch2_fs_read_write(struct bch_fs *c, bool early)
 		bch2_dev_allocator_add(c, ca);
 	bch2_recalc_capacity(c);
 
-	bch2_io_timer_add(&c->io_clock[READ], &c->bucket_clock[READ].rescale);
-	bch2_io_timer_add(&c->io_clock[WRITE], &c->bucket_clock[WRITE].rescale);
-
 	for_each_rw_member(ca, c, i) {
 		ret = bch2_dev_allocator_start(ca);
 		if (ret) {
diff --git a/fs/bcachefs/sysfs.c b/fs/bcachefs/sysfs.c
index 521b6d8d518f..8fdbeaf9df32 100644
--- a/fs/bcachefs/sysfs.c
+++ b/fs/bcachefs/sysfs.c
@@ -705,7 +705,7 @@ static unsigned bucket_last_io_fn(struct bch_fs *c, struct bch_dev *ca,
 {
 	int rw = (private ? 1 : 0);
 
-	return bucket_last_io(c, bucket(ca, b), rw);
+	return atomic64_read(&c->io_clock[rw].now) - bucket(ca, b)->io_time[rw];
 }
 
 static unsigned bucket_sectors_used_fn(struct bch_fs *c, struct bch_dev *ca,
@@ -718,7 +718,7 @@ static unsigned bucket_sectors_used_fn(struct bch_fs *c, struct bch_dev *ca,
 static unsigned bucket_oldest_gen_fn(struct bch_fs *c, struct bch_dev *ca,
 				     size_t b, void *private)
 {
-	return bucket_gc_gen(ca, b);
+	return bucket_gc_gen(bucket(ca, b));
 }
 
 static int unsigned_cmp(const void *_l, const void *_r)