bcachefs: Introduce a separate journal watermark for copygc

Since journal reclaim -> btree key cache flushing may require the
allocation of new btree nodes, it has an implicit dependency on copygc
in order to make forward progress - so we should avoid blocking copygc
unless the journal is really close to full.

This introduces watermarks to replace our single MAY_GET_UNRESERVED bit
in the journal, and adds a watermark for copygc and plumbs it through.

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

fs/bcachefs/alloc_background.c

@@ -670,7 +670,6 @@ static int bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca,
 	ret = bch2_trans_do(c, NULL, &commit_seq,
 			    BTREE_INSERT_NOCHECK_RW|
 			    BTREE_INSERT_NOFAIL|
-			    BTREE_INSERT_JOURNAL_RESERVED|
 			    flags,
 			    bucket_invalidate_btree(&trans, ca, b, &u));
 

fs/bcachefs/btree_key_cache.c

@@ -421,7 +421,7 @@ static int btree_key_cache_flush_pos(struct btree_trans *trans,
 				  BTREE_INSERT_NOFAIL|
 				  BTREE_INSERT_USE_RESERVE|
 				  (ck->journal.seq == journal_last_seq(j)
-				   ? BTREE_INSERT_JOURNAL_RESERVED
+				   ? JOURNAL_WATERMARK_reserved
 				   : 0)|
 				  commit_flags);
 	if (ret) {

fs/bcachefs/btree_update.h

@@ -16,12 +16,12 @@ bool bch2_btree_bset_insert_key(struct btree_trans *, struct btree_path *,
 void bch2_btree_add_journal_pin(struct bch_fs *, struct btree *, u64);
 
 enum btree_insert_flags {
-	__BTREE_INSERT_NOFAIL,
+	/* First two bits for journal watermark: */
+	__BTREE_INSERT_NOFAIL = 2,
 	__BTREE_INSERT_NOCHECK_RW,
 	__BTREE_INSERT_LAZY_RW,
 	__BTREE_INSERT_USE_RESERVE,
 	__BTREE_INSERT_JOURNAL_REPLAY,
-	__BTREE_INSERT_JOURNAL_RESERVED,
 	__BTREE_INSERT_JOURNAL_RECLAIM,
 	__BTREE_INSERT_NOWAIT,
 	__BTREE_INSERT_GC_LOCK_HELD,
@@ -41,9 +41,6 @@ enum btree_insert_flags {
 /* Insert is for journal replay - don't get journal reservations: */
 #define BTREE_INSERT_JOURNAL_REPLAY	(1 << __BTREE_INSERT_JOURNAL_REPLAY)
 
-/* Indicates that we have pre-reserved space in the journal: */
-#define BTREE_INSERT_JOURNAL_RESERVED	(1 << __BTREE_INSERT_JOURNAL_RESERVED)
-
 /* Insert is being called from journal reclaim path: */
 #define BTREE_INSERT_JOURNAL_RECLAIM (1 << __BTREE_INSERT_JOURNAL_RECLAIM)
 

fs/bcachefs/btree_update_interior.c

@@ -599,7 +599,7 @@ static void btree_update_nodes_written(struct btree_update *as)
 			      BTREE_INSERT_NOFAIL|
 			      BTREE_INSERT_NOCHECK_RW|
 			      BTREE_INSERT_JOURNAL_RECLAIM|
-			      BTREE_INSERT_JOURNAL_RESERVED,
+			      JOURNAL_WATERMARK_reserved,
 			      btree_update_nodes_written_trans(&trans, as));
 	bch2_trans_exit(&trans);
 
@@ -964,14 +964,11 @@ bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
 		? BCH_DISK_RESERVATION_NOFAIL : 0;
 	unsigned nr_nodes[2];
 	unsigned update_level = level;
-	int journal_flags = 0;
+	int journal_flags = flags & JOURNAL_WATERMARK_MASK;
 	int ret = 0;
 
 	BUG_ON(!path->should_be_locked);
 
-	if (flags & BTREE_INSERT_JOURNAL_RESERVED)
-		journal_flags |= JOURNAL_RES_GET_RESERVED;
-
 	closure_init_stack(&cl);
 retry:
 	nr_nodes[0] = nr_nodes[1] = 0;
@@ -1972,7 +1969,7 @@ static int __bch2_btree_node_update_key(struct btree_trans *trans,
 				BTREE_INSERT_NOCHECK_RW|
 				BTREE_INSERT_USE_RESERVE|
 				BTREE_INSERT_JOURNAL_RECLAIM|
-				BTREE_INSERT_JOURNAL_RESERVED);
+				JOURNAL_WATERMARK_reserved);
 	if (ret)
 		goto err;
 

fs/bcachefs/btree_update_leaf.c

@@ -296,11 +296,10 @@ static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 	struct bch_fs *c = trans->c;
 	int ret;
 
-	if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
-		flags |= JOURNAL_RES_GET_RESERVED;
-
 	ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
-				   trans->journal_u64s, flags);
+				   trans->journal_u64s,
+				   flags|
+				   (trans->flags & JOURNAL_WATERMARK_MASK));
 
 	return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 }
@@ -902,8 +901,7 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans,
 	ret = bch2_journal_preres_get(&c->journal,
 			&trans->journal_preres, trans->journal_preres_u64s,
 			JOURNAL_RES_GET_NONBLOCK|
-			((trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
-			 ? JOURNAL_RES_GET_RESERVED : 0));
+			(trans->flags & JOURNAL_WATERMARK_MASK));
 	if (unlikely(ret == -EAGAIN))
 		ret = bch2_trans_journal_preres_get_cold(trans,
 						trans->journal_preres_u64s, trace_ip);
@@ -988,7 +986,7 @@ int bch2_trans_commit_error(struct btree_trans *trans,
 		bch2_trans_unlock(trans);
 
 		if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
-		    !(trans->flags & BTREE_INSERT_JOURNAL_RESERVED)) {
+		    !(trans->flags & JOURNAL_WATERMARK_reserved)) {
 			trans->restarted = true;
 			ret = -EAGAIN;
 			break;

fs/bcachefs/journal.c

@@ -19,6 +19,18 @@
 #include "super-io.h"
 #include "trace.h"
 
+#define x(n)	#n,
+static const char * const bch2_journal_watermarks[] = {
+	JOURNAL_WATERMARKS()
+	NULL
+};
+
+static const char * const bch2_journal_errors[] = {
+	JOURNAL_ERRORS()
+	NULL
+};
+#undef x
+
 static inline bool journal_seq_unwritten(struct journal *j, u64 seq)
 {
 	return seq > j->seq_ondisk;
@@ -207,19 +219,19 @@ static int journal_entry_open(struct journal *j)
 	BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
 
 	if (j->blocked)
-		return cur_entry_blocked;
+		return JOURNAL_ERR_blocked;
 
 	if (j->cur_entry_error)
 		return j->cur_entry_error;
 
 	if (bch2_journal_error(j))
-		return cur_entry_insufficient_devices; /* -EROFS */
+		return JOURNAL_ERR_insufficient_devices; /* -EROFS */
 
 	if (!fifo_free(&j->pin))
-		return cur_entry_journal_pin_full;
+		return JOURNAL_ERR_journal_pin_full;
 
 	if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf))
-		return cur_entry_max_in_flight;
+		return JOURNAL_ERR_max_in_flight;
 
 	BUG_ON(!j->cur_entry_sectors);
 
@@ -238,7 +250,7 @@ static int journal_entry_open(struct journal *j)
 	u64s = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
 
 	if (u64s <= 0)
-		return cur_entry_journal_full;
+		return JOURNAL_ERR_journal_full;
 
 	if (fifo_empty(&j->pin) && j->reclaim_thread)
 		wake_up_process(j->reclaim_thread);
@@ -354,13 +366,12 @@ 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)) {
+	if ((flags & JOURNAL_WATERMARK_MASK) < j->watermark) {
 		/*
 		 * Don't want to close current journal entry, just need to
 		 * invoke reclaim:
 		 */
-		ret = cur_entry_journal_full;
+		ret = JOURNAL_ERR_journal_full;
 		goto unlock;
 	}
 
@@ -378,10 +389,10 @@ static int __journal_res_get(struct journal *j, struct journal_res *res,
 	__journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL);
 	ret = journal_entry_open(j);
 
-	if (ret == cur_entry_max_in_flight)
+	if (ret == JOURNAL_ERR_max_in_flight)
 		trace_journal_entry_full(c);
 unlock:
-	if ((ret && ret != cur_entry_insufficient_devices) &&
+	if ((ret && ret != JOURNAL_ERR_insufficient_devices) &&
 	    !j->res_get_blocked_start) {
 		j->res_get_blocked_start = local_clock() ?: 1;
 		trace_journal_full(c);
@@ -393,14 +404,15 @@ static int __journal_res_get(struct journal *j, struct journal_res *res,
 	if (!ret)
 		goto retry;
 
-	if ((ret == cur_entry_journal_full ||
-	     ret == cur_entry_journal_pin_full) &&
+	if ((ret == JOURNAL_ERR_journal_full ||
+	     ret == JOURNAL_ERR_journal_pin_full) &&
 	    !can_discard &&
 	    !nr_unwritten_journal_entries(j) &&
-	    (flags & JOURNAL_RES_GET_RESERVED)) {
+	    (flags & JOURNAL_WATERMARK_MASK) == JOURNAL_WATERMARK_reserved) {
 		struct printbuf buf = PRINTBUF;
 
-		bch_err(c, "Journal stuck! Hava a pre-reservation but journal full");
+		bch_err(c, "Journal stuck! Hava a pre-reservation but journal full (ret %s)",
+			bch2_journal_errors[ret]);
 
 		bch2_journal_debug_to_text(&buf, j);
 		bch_err(c, "%s", buf.buf);
@@ -418,8 +430,8 @@ static int __journal_res_get(struct journal *j, struct journal_res *res,
 	 * Journal is full - can't rely on reclaim from work item due to
 	 * freezing:
 	 */
-	if ((ret == cur_entry_journal_full ||
-	     ret == cur_entry_journal_pin_full) &&
+	if ((ret == JOURNAL_ERR_journal_full ||
+	     ret == JOURNAL_ERR_journal_pin_full) &&
 	    !(flags & JOURNAL_RES_GET_NONBLOCK)) {
 		if (can_discard) {
 			bch2_journal_do_discards(j);
@@ -432,7 +444,7 @@ static int __journal_res_get(struct journal *j, struct journal_res *res,
 		}
 	}
 
-	return ret == cur_entry_insufficient_devices ? -EROFS : -EAGAIN;
+	return ret == JOURNAL_ERR_insufficient_devices ? -EROFS : -EAGAIN;
 }
 
 /*
@@ -1187,13 +1199,14 @@ void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
 	rcu_read_lock();
 	s = READ_ONCE(j->reservations);
 
-	pr_buf(out, "dirty journal entries:\t%llu\n",	fifo_used(&j->pin));
+	pr_buf(out, "dirty journal entries:\t%llu/%llu\n",fifo_used(&j->pin), j->pin.size);
 	pr_buf(out, "seq:\t\t\t%llu\n",			journal_cur_seq(j));
 	pr_buf(out, "seq_ondisk:\t\t%llu\n",		j->seq_ondisk);
 	pr_buf(out, "last_seq:\t\t%llu\n",		journal_last_seq(j));
 	pr_buf(out, "last_seq_ondisk:\t%llu\n",		j->last_seq_ondisk);
 	pr_buf(out, "flushed_seq_ondisk:\t%llu\n",	j->flushed_seq_ondisk);
 	pr_buf(out, "prereserved:\t\t%u/%u\n",		j->prereserved.reserved, j->prereserved.remaining);
+	pr_buf(out, "watermark:\t\t%s\n",		bch2_journal_watermarks[j->watermark]);
 	pr_buf(out, "each entry reserved:\t%u\n",	j->entry_u64s_reserved);
 	pr_buf(out, "nr flush writes:\t%llu\n",		j->nr_flush_writes);
 	pr_buf(out, "nr noflush writes:\t%llu\n",	j->nr_noflush_writes);
@@ -1203,7 +1216,7 @@ void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
 	pr_buf(out, "reclaim runs in:\t%u ms\n",	time_after(j->next_reclaim, now)
 	       ? jiffies_to_msecs(j->next_reclaim - jiffies) : 0);
 	pr_buf(out, "current entry sectors:\t%u\n",	j->cur_entry_sectors);
-	pr_buf(out, "current entry error:\t%u\n",	j->cur_entry_error);
+	pr_buf(out, "current entry error:\t%s\n",	bch2_journal_errors[j->cur_entry_error]);
 	pr_buf(out, "current entry:\t\t");
 
 	switch (s.cur_entry_offset) {

fs/bcachefs/journal.h

@@ -293,9 +293,9 @@ static inline void bch2_journal_res_put(struct journal *j,
 int bch2_journal_res_get_slowpath(struct journal *, struct journal_res *,
 				  unsigned);
 
-#define JOURNAL_RES_GET_NONBLOCK	(1 << 0)
-#define JOURNAL_RES_GET_CHECK		(1 << 1)
-#define JOURNAL_RES_GET_RESERVED	(1 << 2)
+/* First two bits for JOURNAL_WATERMARK: */
+#define JOURNAL_RES_GET_NONBLOCK	(1 << 2)
+#define JOURNAL_RES_GET_CHECK		(1 << 3)
 
 static inline int journal_res_get_fast(struct journal *j,
 				       struct journal_res *res,
@@ -316,8 +316,7 @@ 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))
+		if ((flags & JOURNAL_WATERMARK_MASK) < j->watermark)
 			return 0;
 
 		new.cur_entry_offset += res->u64s;
@@ -370,23 +369,27 @@ static inline int bch2_journal_res_get(struct journal *j, struct journal_res *re
 
 /* journal_preres: */
 
-static inline bool journal_check_may_get_unreserved(struct journal *j)
+static inline void journal_set_watermark(struct journal *j)
 {
 	union journal_preres_state s = READ_ONCE(j->prereserved);
-	bool ret = s.reserved < s.remaining &&
-		fifo_free(&j->pin) > j->pin.size / 4;
+	unsigned watermark = JOURNAL_WATERMARK_any;
 
-	lockdep_assert_held(&j->lock);
+	if (fifo_free(&j->pin) < j->pin.size / 4)
+		watermark = max_t(unsigned, watermark, JOURNAL_WATERMARK_copygc);
+	if (fifo_free(&j->pin) < j->pin.size / 8)
+		watermark = max_t(unsigned, watermark, JOURNAL_WATERMARK_reserved);
 
-	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;
+	if (s.reserved > s.remaining)
+		watermark = max_t(unsigned, watermark, JOURNAL_WATERMARK_copygc);
+	if (!s.remaining)
+		watermark = max_t(unsigned, watermark, JOURNAL_WATERMARK_reserved);
+
+	if (watermark == j->watermark)
+		return;
+
+	swap(watermark, j->watermark);
+	if (watermark > j->watermark)
+		journal_wake(j);
 }
 
 static inline void bch2_journal_preres_put(struct journal *j,
@@ -406,12 +409,8 @@ static inline void bch2_journal_preres_put(struct journal *j,
 		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);
-	}
+	if (s.reserved <= s.remaining && j->watermark)
+		journal_set_watermark(j);
 }
 
 int __bch2_journal_preres_get(struct journal *,
@@ -432,7 +431,7 @@ static inline int bch2_journal_preres_get_fast(struct journal *j,
 		old.v = new.v = v;
 		ret = 0;
 
-		if ((flags & JOURNAL_RES_GET_RESERVED) ||
+		if ((flags & JOURNAL_WATERMARK_reserved) ||
 		    new.reserved + d < new.remaining) {
 			new.reserved += d;
 			ret = 1;

fs/bcachefs/journal_reclaim.c

@@ -195,7 +195,7 @@ void bch2_journal_space_available(struct journal *j)
 	j->can_discard = can_discard;
 
 	if (nr_online < c->opts.metadata_replicas_required) {
-		ret = cur_entry_insufficient_devices;
+		ret = JOURNAL_ERR_insufficient_devices;
 		goto out;
 	}
 
@@ -224,9 +224,9 @@ void bch2_journal_space_available(struct journal *j)
 		bch2_fatal_error(c);
 		spin_lock(&j->lock);
 
-		ret = cur_entry_journal_stuck;
+		ret = JOURNAL_ERR_journal_stuck;
 	} else if (!j->space[journal_space_discarded].next_entry)
-		ret = cur_entry_journal_full;
+		ret = JOURNAL_ERR_journal_full;
 
 	if ((j->space[journal_space_clean_ondisk].next_entry <
 	     j->space[journal_space_clean_ondisk].total) &&
@@ -245,7 +245,7 @@ void bch2_journal_space_available(struct journal *j)
 	j->cur_entry_sectors	= !ret ? j->space[journal_space_discarded].next_entry : 0;
 	j->cur_entry_error	= ret;
 	journal_set_remaining(j, u64s_remaining);
-	journal_check_may_get_unreserved(j);
+	journal_set_watermark(j);
 
 	if (!ret)
 		journal_wake(j);

fs/bcachefs/journal_types.h

@@ -144,16 +144,45 @@ enum journal_space_from {
 enum {
 	JOURNAL_REPLAY_DONE,
 	JOURNAL_STARTED,
-	JOURNAL_MAY_GET_UNRESERVED,
 	JOURNAL_MAY_SKIP_FLUSH,
 };
 
+#define JOURNAL_WATERMARKS()		\
+	x(any)				\
+	x(copygc)			\
+	x(reserved)
+
+enum journal_watermark {
+#define x(n)	JOURNAL_WATERMARK_##n,
+	JOURNAL_WATERMARKS()
+#undef x
+};
+
+#define JOURNAL_WATERMARK_MASK	3
+
+/* Reasons we may fail to get a journal reservation: */
+#define JOURNAL_ERRORS()		\
+	x(ok)				\
+	x(blocked)			\
+	x(max_in_flight)		\
+	x(journal_full)			\
+	x(journal_pin_full)		\
+	x(journal_stuck)		\
+	x(insufficient_devices)
+
+enum journal_errors {
+#define x(n)	JOURNAL_ERR_##n,
+	JOURNAL_ERRORS()
+#undef x
+};
+
 /* Embedded in struct bch_fs */
 struct journal {
 	/* Fastpath stuff up front: */
 	struct {
 
 	union journal_res_state reservations;
+	enum journal_watermark	watermark;
 
 	union journal_preres_state prereserved;
 
@@ -173,15 +202,7 @@ struct journal {
 	 * 0, or -ENOSPC if waiting on journal reclaim, or -EROFS if
 	 * insufficient devices:
 	 */
-	enum {
-		cur_entry_ok,
-		cur_entry_blocked,
-		cur_entry_max_in_flight,
-		cur_entry_journal_full,
-		cur_entry_journal_pin_full,
-		cur_entry_journal_stuck,
-		cur_entry_insufficient_devices,
-	}			cur_entry_error;
+	enum journal_errors	cur_entry_error;
 
 	unsigned		buf_size_want;
 	/*

fs/bcachefs/movinggc.c

@@ -91,7 +91,7 @@ static enum data_cmd copygc_pred(struct bch_fs *c, void *arg,
 			data_opts->target		= io_opts->background_target;
 			data_opts->nr_replicas		= 1;
 			data_opts->btree_insert_flags	= BTREE_INSERT_USE_RESERVE|
-				BTREE_INSERT_JOURNAL_RESERVED;
+				JOURNAL_WATERMARK_copygc;
 			data_opts->rewrite_dev		= p.ptr.dev;
 
 			if (p.has_ec)

fs/bcachefs/recovery.c

@@ -562,8 +562,9 @@ static int bch2_journal_replay(struct bch_fs *c)
 		ret = bch2_trans_do(c, NULL, NULL,
 				    BTREE_INSERT_LAZY_RW|
 				    BTREE_INSERT_NOFAIL|
-				    BTREE_INSERT_JOURNAL_RESERVED|
-				    (!k->allocated ? BTREE_INSERT_JOURNAL_REPLAY : 0),
+				    (!k->allocated
+				     ? BTREE_INSERT_JOURNAL_REPLAY|JOURNAL_WATERMARK_reserved
+				     : 0),
 			     bch2_journal_replay_key(&trans, k));
 		if (ret) {
 			bch_err(c, "journal replay: error %d while replaying key at btree %s level %u",