bcachefs: Add a pre-reserve mechanism for the journal

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

fs/bcachefs/journal.c

@@ -343,6 +343,16 @@ static int __journal_res_get(struct journal *j, struct journal_res *res,
 		return 0;
 	}
 
+	if (!(flags & JOURNAL_RES_GET_RESERVED) &&
+	    !test_bit(JOURNAL_MAY_GET_UNRESERVED, &j->flags)) {
+		/*
+		 * Don't want to close current journal entry, just need to
+		 * invoke reclaim:
+		 */
+		ret = -ENOSPC;
+		goto unlock;
+	}
+
 	/*
 	 * If we couldn't get a reservation because the current buf filled up,
 	 * and we had room for a bigger entry on disk, signal that we want to
@@ -366,7 +376,7 @@ static int __journal_res_get(struct journal *j, struct journal_res *res,
 	} else {
 		ret = journal_entry_open(j);
 	}
-
+unlock:
 	if ((ret == -EAGAIN || ret == -ENOSPC) &&
 	    !j->res_get_blocked_start)
 		j->res_get_blocked_start = local_clock() ?: 1;
@@ -378,6 +388,8 @@ static int __journal_res_get(struct journal *j, struct journal_res *res,
 		goto retry;
 
 	if (ret == -ENOSPC) {
+		BUG_ON(!can_discard && (flags & JOURNAL_RES_GET_RESERVED));
+
 		/*
 		 * Journal is full - can't rely on reclaim from work item due to
 		 * freezing:
@@ -423,6 +435,32 @@ int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res,
 	return ret;
 }
 
+/* journal_preres: */
+
+static bool journal_preres_available(struct journal *j,
+				     struct journal_preres *res,
+				     unsigned new_u64s)
+{
+	bool ret = bch2_journal_preres_get_fast(j, res, new_u64s);
+
+	if (!ret)
+		bch2_journal_reclaim_work(&j->reclaim_work.work);
+
+	return ret;
+}
+
+int __bch2_journal_preres_get(struct journal *j,
+			      struct journal_preres *res,
+			      unsigned new_u64s)
+{
+	int ret;
+
+	closure_wait_event(&j->preres_wait,
+		   (ret = bch2_journal_error(j)) ||
+		   journal_preres_available(j, res, new_u64s));
+	return ret;
+}
+
 /* journal_entry_res: */
 
 void bch2_journal_entry_res_resize(struct journal *j,
@@ -1110,11 +1148,16 @@ ssize_t bch2_journal_print_debug(struct journal *j, char *buf)
 	       "seq:\t\t\t%llu\n"
 	       "last_seq:\t\t%llu\n"
 	       "last_seq_ondisk:\t%llu\n"
+	       "prereserved:\t\t%u/%u\n"
+	       "current entry sectors:\t%u\n"
 	       "current entry:\t\t",
 	       fifo_used(&j->pin),
 	       journal_cur_seq(j),
 	       journal_last_seq(j),
-	       j->last_seq_ondisk);
+	       j->last_seq_ondisk,
+	       j->prereserved.reserved,
+	       j->prereserved.remaining,
+	       j->cur_entry_sectors);
 
 	switch (s.cur_entry_offset) {
 	case JOURNAL_ENTRY_ERROR_VAL:
@@ -1136,8 +1179,9 @@ ssize_t bch2_journal_print_debug(struct journal *j, char *buf)
 	       journal_state_count(s, s.idx));
 
 	if (s.prev_buf_unwritten)
-		pr_buf(&out, "yes, ref %u\n",
-		       journal_state_count(s, !s.idx));
+		pr_buf(&out, "yes, ref %u sectors %u\n",
+		       journal_state_count(s, !s.idx),
+		       journal_prev_buf(j)->sectors);
 	else
 		pr_buf(&out, "no\n");
 

fs/bcachefs/journal.h

@@ -119,6 +119,7 @@ static inline void journal_wake(struct journal *j)
 {
 	wake_up(&j->wait);
 	closure_wake_up(&j->async_wait);
+	closure_wake_up(&j->preres_wait);
 }
 
 static inline struct journal_buf *journal_cur_buf(struct journal *j)
@@ -274,6 +275,7 @@ int bch2_journal_res_get_slowpath(struct journal *, struct journal_res *,
 
 #define JOURNAL_RES_GET_NONBLOCK	(1 << 0)
 #define JOURNAL_RES_GET_CHECK		(1 << 1)
+#define JOURNAL_RES_GET_RESERVED	(1 << 2)
 
 static inline int journal_res_get_fast(struct journal *j,
 				       struct journal_res *res,
@@ -294,6 +296,10 @@ static inline int journal_res_get_fast(struct journal *j,
 
 		EBUG_ON(!journal_state_count(new, new.idx));
 
+		if (!(flags & JOURNAL_RES_GET_RESERVED) &&
+		    !test_bit(JOURNAL_MAY_GET_UNRESERVED, &j->flags))
+			return 0;
+
 		if (flags & JOURNAL_RES_GET_CHECK)
 			return 1;
 
@@ -333,6 +339,89 @@ static inline int bch2_journal_res_get(struct journal *j, struct journal_res *re
 	return 0;
 }
 
+/* journal_preres: */
+
+static inline bool journal_check_may_get_unreserved(struct journal *j)
+{
+	union journal_preres_state s = READ_ONCE(j->prereserved);
+	bool ret = s.reserved <= s.remaining &&
+		fifo_free(&j->pin) > 8;
+
+	lockdep_assert_held(&j->lock);
+
+	if (ret != test_bit(JOURNAL_MAY_GET_UNRESERVED, &j->flags)) {
+		if (ret) {
+			set_bit(JOURNAL_MAY_GET_UNRESERVED, &j->flags);
+			journal_wake(j);
+		} else {
+			clear_bit(JOURNAL_MAY_GET_UNRESERVED, &j->flags);
+		}
+	}
+	return ret;
+}
+
+static inline void bch2_journal_preres_put(struct journal *j,
+					   struct journal_preres *res)
+{
+	union journal_preres_state s = { .reserved = res->u64s };
+
+	if (!res->u64s)
+		return;
+
+	s.v = atomic64_sub_return(s.v, &j->prereserved.counter);
+	res->u64s = 0;
+	closure_wake_up(&j->preres_wait);
+
+	if (s.reserved <= s.remaining &&
+	    !test_bit(JOURNAL_MAY_GET_UNRESERVED, &j->flags)) {
+		spin_lock(&j->lock);
+		journal_check_may_get_unreserved(j);
+		spin_unlock(&j->lock);
+	}
+}
+
+int __bch2_journal_preres_get(struct journal *,
+			struct journal_preres *, unsigned);
+
+static inline int bch2_journal_preres_get_fast(struct journal *j,
+					       struct journal_preres *res,
+					       unsigned new_u64s)
+{
+	int d = new_u64s - res->u64s;
+	union journal_preres_state old, new;
+	u64 v = atomic64_read(&j->prereserved.counter);
+
+	do {
+		old.v = new.v = v;
+
+		new.reserved += d;
+
+		if (new.reserved > new.remaining)
+			return 0;
+	} while ((v = atomic64_cmpxchg(&j->prereserved.counter,
+				       old.v, new.v)) != old.v);
+
+	res->u64s += d;
+	return 1;
+}
+
+static inline int bch2_journal_preres_get(struct journal *j,
+					  struct journal_preres *res,
+					  unsigned new_u64s,
+					  unsigned flags)
+{
+	if (new_u64s <= res->u64s)
+		return 0;
+
+	if (bch2_journal_preres_get_fast(j, res, new_u64s))
+		return 0;
+
+	if (flags & JOURNAL_RES_GET_NONBLOCK)
+		return -EAGAIN;
+
+	return __bch2_journal_preres_get(j, res, new_u64s);
+}
+
 /* journal_entry_res: */
 
 void bch2_journal_entry_res_resize(struct journal *,

fs/bcachefs/journal_io.c

@@ -974,6 +974,12 @@ static int journal_write_alloc(struct journal *j, struct journal_buf *w,
 					journal_space_discarded)) {
 			ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
 			ja->sectors_free = ca->mi.bucket_size;
+
+			/*
+			 * ja->bucket_seq[ja->cur_idx] must always have
+			 * something sensible:
+			 */
+			ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
 		}
 	}
 

fs/bcachefs/journal_reclaim.c

@@ -49,6 +49,18 @@ unsigned bch2_journal_dev_buckets_available(struct journal *j,
 	return available;
 }
 
+static void journal_set_remaining(struct journal *j, unsigned u64s_remaining)
+{
+	union journal_preres_state old, new;
+	u64 v = atomic64_read(&j->prereserved.counter);
+
+	do {
+		old.v = new.v = v;
+		new.remaining = u64s_remaining;
+	} while ((v = atomic64_cmpxchg(&j->prereserved.counter,
+				       old.v, new.v)) != old.v);
+}
+
 static struct journal_space {
 	unsigned	next_entry;
 	unsigned	remaining;
@@ -124,6 +136,7 @@ void bch2_journal_space_available(struct journal *j)
 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
 	struct bch_dev *ca;
 	struct journal_space discarded, clean_ondisk, clean;
+	unsigned overhead, u64s_remaining = 0;
 	unsigned max_entry_size	 = min(j->buf[0].buf_size >> 9,
 				       j->buf[1].buf_size >> 9);
 	unsigned i, nr_online = 0, nr_devs_want;
@@ -176,9 +189,17 @@ void bch2_journal_space_available(struct journal *j)
 
 	if (!discarded.next_entry)
 		ret = -ENOSPC;
+
+	overhead = DIV_ROUND_UP(clean.remaining, max_entry_size) *
+		journal_entry_overhead(j);
+	u64s_remaining = clean.remaining << 6;
+	u64s_remaining = max_t(int, 0, u64s_remaining - overhead);
+	u64s_remaining /= 4;
 out:
 	j->cur_entry_sectors	= !ret ? discarded.next_entry : 0;
 	j->cur_entry_error	= ret;
+	journal_set_remaining(j, u64s_remaining);
+	journal_check_may_get_unreserved(j);
 
 	if (!ret)
 		journal_wake(j);
@@ -454,7 +475,7 @@ void bch2_journal_reclaim(struct journal *j)
 {
 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
 	struct bch_dev *ca;
-	unsigned iter, bucket_to_flush, min_nr = 0;
+	unsigned iter, min_nr = 0;
 	u64 seq_to_flush = 0;
 
 	lockdep_assert_held(&j->reclaim_lock);
@@ -465,13 +486,22 @@ void bch2_journal_reclaim(struct journal *j)
 
 	for_each_rw_member(ca, c, iter) {
 		struct journal_device *ja = &ca->journal;
+		unsigned nr_buckets, bucket_to_flush;
 
 		if (!ja->nr)
 			continue;
 
-
 		/* Try to keep the journal at most half full: */
-		bucket_to_flush = (ja->cur_idx + (ja->nr >> 1)) % ja->nr;
+		nr_buckets = ja->nr / 2;
+
+		/* And include pre-reservations: */
+		nr_buckets += DIV_ROUND_UP(j->prereserved.reserved,
+					   (ca->mi.bucket_size << 6) -
+					   journal_entry_overhead(j));
+
+		nr_buckets = min(nr_buckets, ja->nr);
+
+		bucket_to_flush = (ja->cur_idx + nr_buckets) % ja->nr;
 		seq_to_flush = max_t(u64, seq_to_flush,
 				     ja->bucket_seq[bucket_to_flush]);
 	}
@@ -490,6 +520,9 @@ void bch2_journal_reclaim(struct journal *j)
 		       msecs_to_jiffies(j->reclaim_delay_ms)))
 		min_nr = 1;
 
+	if (j->prereserved.reserved * 2 > j->prereserved.remaining)
+		min_nr = 1;
+
 	journal_flush_pins(j, seq_to_flush, min_nr);
 
 	if (!test_bit(BCH_FS_RO, &c->flags))

fs/bcachefs/journal_types.h

@@ -80,6 +80,14 @@ struct journal_res {
 	u64			seq;
 };
 
+/*
+ * For reserving space in the journal prior to getting a reservation on a
+ * particular journal entry:
+ */
+struct journal_preres {
+	unsigned		u64s;
+};
+
 union journal_res_state {
 	struct {
 		atomic64_t	counter;
@@ -98,6 +106,21 @@ union journal_res_state {
 	};
 };
 
+union journal_preres_state {
+	struct {
+		atomic64_t	counter;
+	};
+
+	struct {
+		u64		v;
+	};
+
+	struct {
+		u32		reserved;
+		u32		remaining;
+	};
+};
+
 /* bytes: */
 #define JOURNAL_ENTRY_SIZE_MIN		(64U << 10) /* 64k */
 #define JOURNAL_ENTRY_SIZE_MAX		(4U  << 20) /* 4M */
@@ -122,6 +145,7 @@ enum {
 	JOURNAL_STARTED,
 	JOURNAL_NEED_WRITE,
 	JOURNAL_NOT_EMPTY,
+	JOURNAL_MAY_GET_UNRESERVED,
 };
 
 /* Embedded in struct bch_fs */
@@ -142,6 +166,8 @@ struct journal {
 	 */
 	int			cur_entry_error;
 
+	union journal_preres_state prereserved;
+
 	/* Reserved space in journal entry to be used just prior to write */
 	unsigned		entry_u64s_reserved;
 
@@ -161,6 +187,7 @@ struct journal {
 	/* Used when waiting because the journal was full */
 	wait_queue_head_t	wait;
 	struct closure_waitlist	async_wait;
+	struct closure_waitlist	preres_wait;
 
 	struct closure		io;
 	struct delayed_work	write_work;