1 files changed, 431 insertions, 168 deletions

diff --git a/fs/bcachefs/recovery_passes.c b/fs/bcachefs/recovery_passes.c
index 22f72bb5b853..dabb29b08ad0 100644
--- a/fs/bcachefs/recovery_passes.c
+++ b/fs/bcachefs/recovery_passes.c
@@ -28,6 +28,145 @@ const char * const bch2_recovery_passes[] = {
 	NULL
 };
 
+static const u8 passes_to_stable_map[] = {
+#define x(n, id, ...)	[BCH_RECOVERY_PASS_##n] = BCH_RECOVERY_PASS_STABLE_##n,
+	BCH_RECOVERY_PASSES()
+#undef x
+};
+
+static const u8 passes_from_stable_map[] = {
+#define x(n, id, ...)	[BCH_RECOVERY_PASS_STABLE_##n] = BCH_RECOVERY_PASS_##n,
+	BCH_RECOVERY_PASSES()
+#undef x
+};
+
+static enum bch_recovery_pass_stable bch2_recovery_pass_to_stable(enum bch_recovery_pass pass)
+{
+	return passes_to_stable_map[pass];
+}
+
+u64 bch2_recovery_passes_to_stable(u64 v)
+{
+	u64 ret = 0;
+	for (unsigned i = 0; i < ARRAY_SIZE(passes_to_stable_map); i++)
+		if (v & BIT_ULL(i))
+			ret |= BIT_ULL(passes_to_stable_map[i]);
+	return ret;
+}
+
+static enum bch_recovery_pass bch2_recovery_pass_from_stable(enum bch_recovery_pass_stable pass)
+{
+	return pass < ARRAY_SIZE(passes_from_stable_map)
+		? passes_from_stable_map[pass]
+		: 0;
+}
+
+u64 bch2_recovery_passes_from_stable(u64 v)
+{
+	u64 ret = 0;
+	for (unsigned i = 0; i < ARRAY_SIZE(passes_from_stable_map); i++)
+		if (v & BIT_ULL(i))
+			ret |= BIT_ULL(passes_from_stable_map[i]);
+	return ret;
+}
+
+static int bch2_sb_recovery_passes_validate(struct bch_sb *sb, struct bch_sb_field *f,
+					    enum bch_validate_flags flags, struct printbuf *err)
+{
+	return 0;
+}
+
+static void bch2_sb_recovery_passes_to_text(struct printbuf *out,
+					    struct bch_sb *sb,
+					    struct bch_sb_field *f)
+{
+	struct bch_sb_field_recovery_passes *r =
+		field_to_type(f, recovery_passes);
+	unsigned nr = recovery_passes_nr_entries(r);
+
+	if (out->nr_tabstops < 1)
+		printbuf_tabstop_push(out, 32);
+	if (out->nr_tabstops < 2)
+		printbuf_tabstop_push(out, 16);
+
+	prt_printf(out, "Pass\tLast run\tLast runtime\n");
+
+	for (struct recovery_pass_entry *i = r->start; i < r->start + nr; i++) {
+		if (!i->last_run)
+			continue;
+
+		unsigned idx = i - r->start;
+
+		prt_printf(out, "%s\t", bch2_recovery_passes[bch2_recovery_pass_from_stable(idx)]);
+
+		bch2_prt_datetime(out, le64_to_cpu(i->last_run));
+		prt_tab(out);
+
+		bch2_pr_time_units(out, le32_to_cpu(i->last_runtime) * NSEC_PER_SEC);
+		prt_newline(out);
+	}
+}
+
+static void bch2_sb_recovery_pass_complete(struct bch_fs *c,
+					   enum bch_recovery_pass pass,
+					   s64 start_time)
+{
+	enum bch_recovery_pass_stable stable = bch2_recovery_pass_to_stable(pass);
+	s64 end_time = ktime_get_real_seconds();
+
+	mutex_lock(&c->sb_lock);
+	struct bch_sb_field_ext *ext = bch2_sb_field_get(c->disk_sb.sb, ext);
+	__clear_bit_le64(stable, ext->recovery_passes_required);
+
+	struct bch_sb_field_recovery_passes *r =
+		bch2_sb_field_get(c->disk_sb.sb, recovery_passes);
+
+	if (stable >= recovery_passes_nr_entries(r)) {
+		unsigned u64s = struct_size(r, start, stable + 1) / sizeof(u64);
+
+		r = bch2_sb_field_resize(&c->disk_sb, recovery_passes, u64s);
+		if (!r) {
+			bch_err(c, "error creating recovery_passes sb section");
+			goto out;
+		}
+	}
+
+	r->start[stable].last_run	= cpu_to_le64(end_time);
+	r->start[stable].last_runtime	= cpu_to_le32(max(0, end_time - start_time));
+out:
+	bch2_write_super(c);
+	mutex_unlock(&c->sb_lock);
+}
+
+static bool bch2_recovery_pass_want_ratelimit(struct bch_fs *c, enum bch_recovery_pass pass)
+{
+	enum bch_recovery_pass_stable stable = bch2_recovery_pass_to_stable(pass);
+	bool ret = false;
+
+	lockdep_assert_held(&c->sb_lock);
+
+	struct bch_sb_field_recovery_passes *r =
+		bch2_sb_field_get(c->disk_sb.sb, recovery_passes);
+
+	if (stable < recovery_passes_nr_entries(r)) {
+		struct recovery_pass_entry *i = r->start + stable;
+
+		/*
+		 * Ratelimit if the last runtime was more than 1% of the time
+		 * since we last ran
+		 */
+		ret = (u64) le32_to_cpu(i->last_runtime) * 100 >
+			ktime_get_real_seconds() - le64_to_cpu(i->last_run);
+	}
+
+	return ret;
+}
+
+const struct bch_sb_field_ops bch_sb_field_ops_recovery_passes = {
+	.validate	= bch2_sb_recovery_passes_validate,
+	.to_text	= bch2_sb_recovery_passes_to_text
+};
+
 /* Fake recovery pass, so that scan_for_btree_nodes isn't 0: */
 static int bch2_recovery_pass_empty(struct bch_fs *c)
 {
@@ -47,11 +186,36 @@ static int bch2_set_may_go_rw(struct bch_fs *c)
 
 	set_bit(BCH_FS_may_go_rw, &c->flags);
 
-	if (keys->nr || !c->opts.read_only || c->opts.fsck || !c->sb.clean || c->opts.recovery_passes)
+	if (keys->nr ||
+	    !c->opts.read_only ||
+	    !c->sb.clean ||
+	    c->opts.recovery_passes ||
+	    (c->opts.fsck && !(c->sb.features & BIT_ULL(BCH_FEATURE_no_alloc_info)))) {
+		if (c->sb.features & BIT_ULL(BCH_FEATURE_no_alloc_info)) {
+			bch_info(c, "mounting a filesystem with no alloc info read-write; will recreate");
+			bch2_reconstruct_alloc(c);
+		}
+
 		return bch2_fs_read_write_early(c);
+	}
 	return 0;
 }
 
+/*
+ * Make sure root inode is readable while we're still in recovery and can rewind
+ * for repair:
+ */
+static int bch2_lookup_root_inode(struct bch_fs *c)
+{
+	subvol_inum inum = BCACHEFS_ROOT_SUBVOL_INUM;
+	struct bch_inode_unpacked inode_u;
+	struct bch_subvolume subvol;
+
+	return bch2_trans_do(c,
+		bch2_subvolume_get(trans, inum.subvol, true, &subvol) ?:
+		bch2_inode_find_by_inum_trans(trans, inum, &inode_u));
+}
+
 struct recovery_pass_fn {
 	int		(*fn)(struct bch_fs *);
 	unsigned	when;
@@ -63,252 +227,351 @@ static struct recovery_pass_fn recovery_pass_fns[] = {
 #undef x
 };
 
-static const u8 passes_to_stable_map[] = {
-#define x(n, id, ...)	[BCH_RECOVERY_PASS_##n] = BCH_RECOVERY_PASS_STABLE_##n,
-	BCH_RECOVERY_PASSES()
-#undef x
-};
+static u64 bch2_recovery_passes_match(unsigned flags)
+{
+	u64 ret = 0;
 
-static enum bch_recovery_pass_stable bch2_recovery_pass_to_stable(enum bch_recovery_pass pass)
+	for (unsigned i = 0; i < ARRAY_SIZE(recovery_pass_fns); i++)
+		if (recovery_pass_fns[i].when & flags)
+			ret |= BIT_ULL(i);
+	return ret;
+}
+
+u64 bch2_fsck_recovery_passes(void)
 {
-	return passes_to_stable_map[pass];
+	return bch2_recovery_passes_match(PASS_FSCK);
 }
 
-u64 bch2_recovery_passes_to_stable(u64 v)
+static void bch2_run_async_recovery_passes(struct bch_fs *c)
 {
-	u64 ret = 0;
-	for (unsigned i = 0; i < ARRAY_SIZE(passes_to_stable_map); i++)
-		if (v & BIT_ULL(i))
-			ret |= BIT_ULL(passes_to_stable_map[i]);
-	return ret;
+	if (!down_trylock(&c->recovery.run_lock))
+		return;
+
+	if (!enumerated_ref_tryget(&c->writes, BCH_WRITE_REF_async_recovery_passes))
+		goto unlock;
+
+	if (queue_work(system_long_wq, &c->recovery.work))
+		return;
+
+	enumerated_ref_put(&c->writes, BCH_WRITE_REF_async_recovery_passes);
+unlock:
+	up(&c->recovery.run_lock);
 }
 
-u64 bch2_recovery_passes_from_stable(u64 v)
+static bool recovery_pass_needs_set(struct bch_fs *c,
+				    enum bch_recovery_pass pass,
+				    enum bch_run_recovery_pass_flags *flags)
 {
-	static const u8 map[] = {
-#define x(n, id, ...)	[BCH_RECOVERY_PASS_STABLE_##n] = BCH_RECOVERY_PASS_##n,
-	BCH_RECOVERY_PASSES()
-#undef x
-	};
+	struct bch_fs_recovery *r = &c->recovery;
+	bool in_recovery = test_bit(BCH_FS_in_recovery, &c->flags);
+	bool persistent = !in_recovery || !(*flags & RUN_RECOVERY_PASS_nopersistent);
 
-	u64 ret = 0;
-	for (unsigned i = 0; i < ARRAY_SIZE(map); i++)
-		if (v & BIT_ULL(i))
-			ret |= BIT_ULL(map[i]);
-	return ret;
+	if ((*flags & RUN_RECOVERY_PASS_ratelimit) &&
+	    !bch2_recovery_pass_want_ratelimit(c, pass))
+		*flags &= ~RUN_RECOVERY_PASS_ratelimit;
+
+	/*
+	 * If RUN_RECOVERY_PASS_nopersistent is set, we don't want to do
+	 * anything if the pass has already run: these mean we need a prior pass
+	 * to run before we continue to repair, we don't expect that pass to fix
+	 * the damage we encountered.
+	 *
+	 * Otherwise, we run run_explicit_recovery_pass when we find damage, so
+	 * it should run again even if it's already run:
+	 */
+
+	if (persistent
+	    ? !(c->sb.recovery_passes_required & BIT_ULL(pass))
+	    : !((r->passes_to_run|r->passes_complete) & BIT_ULL(pass)))
+		return true;
+
+	if (!(*flags & RUN_RECOVERY_PASS_ratelimit) &&
+	    (r->passes_ratelimiting & BIT_ULL(pass)))
+		return true;
+
+	return false;
 }
 
 /*
  * For when we need to rewind recovery passes and run a pass we skipped:
  */
-static int __bch2_run_explicit_recovery_pass(struct bch_fs *c,
-					     enum bch_recovery_pass pass)
+int __bch2_run_explicit_recovery_pass(struct bch_fs *c,
+				      struct printbuf *out,
+				      enum bch_recovery_pass pass,
+				      enum bch_run_recovery_pass_flags flags)
 {
-	if (c->curr_recovery_pass == ARRAY_SIZE(recovery_pass_fns))
-		return -BCH_ERR_not_in_recovery;
+	struct bch_fs_recovery *r = &c->recovery;
+	int ret = 0;
 
-	if (c->recovery_passes_complete & BIT_ULL(pass))
-		return 0;
+	lockdep_assert_held(&c->sb_lock);
 
-	bool print = !(c->opts.recovery_passes & BIT_ULL(pass));
+	bch2_printbuf_make_room(out, 1024);
+	out->atomic++;
 
-	if (pass < BCH_RECOVERY_PASS_set_may_go_rw &&
-	    c->curr_recovery_pass >= BCH_RECOVERY_PASS_set_may_go_rw) {
-		if (print)
-			bch_info(c, "need recovery pass %s (%u), but already rw",
-				 bch2_recovery_passes[pass], pass);
-		return -BCH_ERR_cannot_rewind_recovery;
-	}
+	unsigned long lockflags;
+	spin_lock_irqsave(&r->lock, lockflags);
 
-	if (print)
-		bch_info(c, "running explicit recovery pass %s (%u), currently at %s (%u)",
-			 bch2_recovery_passes[pass], pass,
-			 bch2_recovery_passes[c->curr_recovery_pass], c->curr_recovery_pass);
+	if (!recovery_pass_needs_set(c, pass, &flags))
+		goto out;
 
-	c->opts.recovery_passes |= BIT_ULL(pass);
+	bool in_recovery = test_bit(BCH_FS_in_recovery, &c->flags);
+	bool rewind = in_recovery && r->curr_pass > pass;
+	bool ratelimit = flags & RUN_RECOVERY_PASS_ratelimit;
 
-	if (c->curr_recovery_pass > pass) {
-		c->next_recovery_pass = pass;
-		c->recovery_passes_complete &= (1ULL << pass) >> 1;
-		return -BCH_ERR_restart_recovery;
-	} else {
-		return 0;
+	if (!(in_recovery && (flags & RUN_RECOVERY_PASS_nopersistent))) {
+		struct bch_sb_field_ext *ext = bch2_sb_field_get(c->disk_sb.sb, ext);
+		__set_bit_le64(bch2_recovery_pass_to_stable(pass), ext->recovery_passes_required);
 	}
-}
-
-int bch2_run_explicit_recovery_pass(struct bch_fs *c,
-				    enum bch_recovery_pass pass)
-{
-	unsigned long flags;
-	spin_lock_irqsave(&c->recovery_pass_lock, flags);
-	int ret = __bch2_run_explicit_recovery_pass(c, pass);
-	spin_unlock_irqrestore(&c->recovery_pass_lock, flags);
-	return ret;
-}
 
-int bch2_run_explicit_recovery_pass_persistent_locked(struct bch_fs *c,
-					       enum bch_recovery_pass pass)
-{
-	lockdep_assert_held(&c->sb_lock);
+	if (pass < BCH_RECOVERY_PASS_set_may_go_rw &&
+	    (!in_recovery || r->curr_pass >= BCH_RECOVERY_PASS_set_may_go_rw)) {
+		prt_printf(out, "need recovery pass %s (%u), but already rw\n",
+			   bch2_recovery_passes[pass], pass);
+		ret = -BCH_ERR_cannot_rewind_recovery;
+		goto out;
+	}
 
-	struct bch_sb_field_ext *ext = bch2_sb_field_get(c->disk_sb.sb, ext);
-	__set_bit_le64(bch2_recovery_pass_to_stable(pass), ext->recovery_passes_required);
+	if (ratelimit)
+		r->passes_ratelimiting |= BIT_ULL(pass);
+	else
+		r->passes_ratelimiting &= ~BIT_ULL(pass);
 
-	return bch2_run_explicit_recovery_pass(c, pass);
-}
+	if (in_recovery && !ratelimit) {
+		prt_printf(out, "running recovery pass %s (%u), currently at %s (%u)%s\n",
+			   bch2_recovery_passes[pass], pass,
+			   bch2_recovery_passes[r->curr_pass], r->curr_pass,
+			   rewind ? " - rewinding" : "");
 
-int bch2_run_explicit_recovery_pass_persistent(struct bch_fs *c,
-					       enum bch_recovery_pass pass)
-{
-	enum bch_recovery_pass_stable s = bch2_recovery_pass_to_stable(pass);
+		r->passes_to_run |= BIT_ULL(pass);
 
-	mutex_lock(&c->sb_lock);
-	struct bch_sb_field_ext *ext = bch2_sb_field_get(c->disk_sb.sb, ext);
+		if (rewind) {
+			r->next_pass = pass;
+			r->passes_complete &= (1ULL << pass) >> 1;
+			ret = -BCH_ERR_restart_recovery;
+		}
+	} else {
+		prt_printf(out, "scheduling recovery pass %s (%u)%s\n",
+			   bch2_recovery_passes[pass], pass,
+			   ratelimit ? " - ratelimiting" : "");
 
-	if (!test_bit_le64(s, ext->recovery_passes_required)) {
-		__set_bit_le64(s, ext->recovery_passes_required);
-		bch2_write_super(c);
+		struct recovery_pass_fn *p = recovery_pass_fns + pass;
+		if (p->when & PASS_ONLINE)
+			bch2_run_async_recovery_passes(c);
 	}
-	mutex_unlock(&c->sb_lock);
-
-	return bch2_run_explicit_recovery_pass(c, pass);
+out:
+	spin_unlock_irqrestore(&r->lock, lockflags);
+	--out->atomic;
+	return ret;
 }
 
-static void bch2_clear_recovery_pass_required(struct bch_fs *c,
-					      enum bch_recovery_pass pass)
+int bch2_run_explicit_recovery_pass(struct bch_fs *c,
+				    struct printbuf *out,
+				    enum bch_recovery_pass pass,
+				    enum bch_run_recovery_pass_flags flags)
 {
-	enum bch_recovery_pass_stable s = bch2_recovery_pass_to_stable(pass);
+	int ret = 0;
 
-	mutex_lock(&c->sb_lock);
-	struct bch_sb_field_ext *ext = bch2_sb_field_get(c->disk_sb.sb, ext);
+	scoped_guard(mutex, &c->sb_lock) {
+		if (!recovery_pass_needs_set(c, pass, &flags))
+			return 0;
 
-	if (test_bit_le64(s, ext->recovery_passes_required)) {
-		__clear_bit_le64(s, ext->recovery_passes_required);
+		ret = __bch2_run_explicit_recovery_pass(c, out, pass, flags);
 		bch2_write_super(c);
 	}
-	mutex_unlock(&c->sb_lock);
-}
 
-u64 bch2_fsck_recovery_passes(void)
-{
-	u64 ret = 0;
-
-	for (unsigned i = 0; i < ARRAY_SIZE(recovery_pass_fns); i++)
-		if (recovery_pass_fns[i].when & PASS_FSCK)
-			ret |= BIT_ULL(i);
 	return ret;
 }
 
-static bool should_run_recovery_pass(struct bch_fs *c, enum bch_recovery_pass pass)
+int bch2_run_print_explicit_recovery_pass(struct bch_fs *c, enum bch_recovery_pass pass)
 {
-	struct recovery_pass_fn *p = recovery_pass_fns + pass;
+	enum bch_run_recovery_pass_flags flags = RUN_RECOVERY_PASS_nopersistent;
 
-	if (c->opts.recovery_passes_exclude & BIT_ULL(pass))
-		return false;
-	if (c->opts.recovery_passes & BIT_ULL(pass))
-		return true;
-	if ((p->when & PASS_FSCK) && c->opts.fsck)
-		return true;
-	if ((p->when & PASS_UNCLEAN) && !c->sb.clean)
-		return true;
-	if (p->when & PASS_ALWAYS)
-		return true;
-	return false;
+	if (!recovery_pass_needs_set(c, pass, &flags))
+		return 0;
+
+	struct printbuf buf = PRINTBUF;
+	bch2_log_msg_start(c, &buf);
+
+	mutex_lock(&c->sb_lock);
+	int ret = __bch2_run_explicit_recovery_pass(c, &buf, pass,
+						RUN_RECOVERY_PASS_nopersistent);
+	mutex_unlock(&c->sb_lock);
+
+	bch2_print_str(c, KERN_NOTICE, buf.buf);
+	printbuf_exit(&buf);
+	return ret;
 }
 
 static int bch2_run_recovery_pass(struct bch_fs *c, enum bch_recovery_pass pass)
 {
+	struct bch_fs_recovery *r = &c->recovery;
 	struct recovery_pass_fn *p = recovery_pass_fns + pass;
-	int ret;
 
 	if (!(p->when & PASS_SILENT))
 		bch2_print(c, KERN_INFO bch2_log_msg(c, "%s..."),
 			   bch2_recovery_passes[pass]);
-	ret = p->fn(c);
-	if (ret)
+
+	s64 start_time = ktime_get_real_seconds();
+	int ret = p->fn(c);
+
+	r->passes_to_run &= ~BIT_ULL(pass);
+
+	if (ret) {
+		r->passes_failing |= BIT_ULL(pass);
 		return ret;
+	}
+
+	r->passes_failing = 0;
+
+	if (!test_bit(BCH_FS_error, &c->flags))
+		bch2_sb_recovery_pass_complete(c, pass, start_time);
+
 	if (!(p->when & PASS_SILENT))
 		bch2_print(c, KERN_CONT " done\n");
 
 	return 0;
 }
 
-int bch2_run_online_recovery_passes(struct bch_fs *c)
+static int __bch2_run_recovery_passes(struct bch_fs *c, u64 orig_passes_to_run,
+				      bool online)
 {
-	for (unsigned i = 0; i < ARRAY_SIZE(recovery_pass_fns); i++) {
-		struct recovery_pass_fn *p = recovery_pass_fns + i;
-
-		if (!(p->when & PASS_ONLINE))
-			continue;
+	struct bch_fs_recovery *r = &c->recovery;
+	int ret = 0;
 
-		int ret = bch2_run_recovery_pass(c, i);
-		if (bch2_err_matches(ret, BCH_ERR_restart_recovery)) {
-			i = c->curr_recovery_pass;
-			continue;
-		}
-		if (ret)
-			return ret;
-	}
+	spin_lock_irq(&r->lock);
 
-	return 0;
-}
+	if (online)
+		orig_passes_to_run &= bch2_recovery_passes_match(PASS_ONLINE);
 
-int bch2_run_recovery_passes(struct bch_fs *c)
-{
-	int ret = 0;
+	if (c->sb.features & BIT_ULL(BCH_FEATURE_no_alloc_info))
+		orig_passes_to_run &= ~bch2_recovery_passes_match(PASS_ALLOC);
 
 	/*
-	 * We can't allow set_may_go_rw to be excluded; that would cause us to
-	 * use the journal replay keys for updates where it's not expected.
+	 * A failed recovery pass will be retried after another pass succeeds -
+	 * but not this iteration.
+	 *
+	 * This is because some passes depend on repair done by other passes: we
+	 * may want to retry, but we don't want to loop on failing passes.
 	 */
-	c->opts.recovery_passes_exclude &= ~BCH_RECOVERY_PASS_set_may_go_rw;
 
-	spin_lock_irq(&c->recovery_pass_lock);
+	orig_passes_to_run &= ~r->passes_failing;
 
-	while (c->curr_recovery_pass < ARRAY_SIZE(recovery_pass_fns) && !ret) {
-		unsigned prev_done = c->recovery_pass_done;
-		unsigned pass = c->curr_recovery_pass;
+	r->passes_to_run = orig_passes_to_run;
 
-		c->next_recovery_pass = pass + 1;
+	while (r->passes_to_run) {
+		unsigned prev_done = r->pass_done;
+		unsigned pass = __ffs64(r->passes_to_run);
+		r->curr_pass = pass;
+		r->next_pass = r->curr_pass + 1;
+		r->passes_to_run &= ~BIT_ULL(pass);
 
-		if (c->opts.recovery_pass_last &&
-		    c->curr_recovery_pass > c->opts.recovery_pass_last)
-			break;
+		spin_unlock_irq(&r->lock);
 
-		if (should_run_recovery_pass(c, pass)) {
-			spin_unlock_irq(&c->recovery_pass_lock);
-			ret =   bch2_run_recovery_pass(c, pass) ?:
-				bch2_journal_flush(&c->journal);
-
-			if (!ret && !test_bit(BCH_FS_error, &c->flags))
-				bch2_clear_recovery_pass_required(c, pass);
-			spin_lock_irq(&c->recovery_pass_lock);
-
-			if (c->next_recovery_pass < c->curr_recovery_pass) {
-				/*
-				 * bch2_run_explicit_recovery_pass() was called: we
-				 * can't always catch -BCH_ERR_restart_recovery because
-				 * it may have been called from another thread (btree
-				 * node read completion)
-				 */
-				ret = 0;
-				c->recovery_passes_complete &= ~(~0ULL << c->curr_recovery_pass);
-			} else {
-				c->recovery_passes_complete |= BIT_ULL(pass);
-				c->recovery_pass_done = max(c->recovery_pass_done, pass);
-			}
+		int ret2 = bch2_run_recovery_pass(c, pass) ?:
+			bch2_journal_flush(&c->journal);
+
+		spin_lock_irq(&r->lock);
+
+		if (r->next_pass < r->curr_pass) {
+			/* Rewind: */
+			r->passes_to_run |= orig_passes_to_run & (~0ULL << r->next_pass);
+		} else if (!ret2) {
+			r->pass_done = max(r->pass_done, pass);
+			r->passes_complete |= BIT_ULL(pass);
+		} else {
+			ret = ret2;
 		}
 
-		c->curr_recovery_pass = c->next_recovery_pass;
+		if (ret && !online)
+			break;
 
 		if (prev_done <= BCH_RECOVERY_PASS_check_snapshots &&
-		    c->recovery_pass_done > BCH_RECOVERY_PASS_check_snapshots) {
+		    r->pass_done > BCH_RECOVERY_PASS_check_snapshots) {
 			bch2_copygc_wakeup(c);
 			bch2_rebalance_wakeup(c);
 		}
 	}
 
-	spin_unlock_irq(&c->recovery_pass_lock);
+	clear_bit(BCH_FS_in_recovery, &c->flags);
+	spin_unlock_irq(&r->lock);
 
 	return ret;
 }
+
+static void bch2_async_recovery_passes_work(struct work_struct *work)
+{
+	struct bch_fs *c = container_of(work, struct bch_fs, recovery.work);
+	struct bch_fs_recovery *r = &c->recovery;
+
+	__bch2_run_recovery_passes(c,
+		c->sb.recovery_passes_required & ~r->passes_ratelimiting,
+		true);
+
+	up(&r->run_lock);
+	enumerated_ref_put(&c->writes, BCH_WRITE_REF_async_recovery_passes);
+}
+
+int bch2_run_online_recovery_passes(struct bch_fs *c, u64 passes)
+{
+	return __bch2_run_recovery_passes(c, c->sb.recovery_passes_required|passes, true);
+}
+
+int bch2_run_recovery_passes(struct bch_fs *c, enum bch_recovery_pass from)
+{
+	u64 passes =
+		bch2_recovery_passes_match(PASS_ALWAYS) |
+		(!c->sb.clean ? bch2_recovery_passes_match(PASS_UNCLEAN) : 0) |
+		(c->opts.fsck ? bch2_recovery_passes_match(PASS_FSCK) : 0) |
+		c->opts.recovery_passes |
+		c->sb.recovery_passes_required;
+
+	if (c->opts.recovery_pass_last)
+		passes &= BIT_ULL(c->opts.recovery_pass_last + 1) - 1;
+
+	/*
+	 * We can't allow set_may_go_rw to be excluded; that would cause us to
+	 * use the journal replay keys for updates where it's not expected.
+	 */
+	c->opts.recovery_passes_exclude &= ~BCH_RECOVERY_PASS_set_may_go_rw;
+	passes &= ~c->opts.recovery_passes_exclude;
+
+	passes &= ~(BIT_ULL(from) - 1);
+
+	down(&c->recovery.run_lock);
+	int ret = __bch2_run_recovery_passes(c, passes, false);
+	up(&c->recovery.run_lock);
+
+	return ret;
+}
+
+static void prt_passes(struct printbuf *out, const char *msg, u64 passes)
+{
+	prt_printf(out, "%s:\t", msg);
+	prt_bitflags(out, bch2_recovery_passes, passes);
+	prt_newline(out);
+}
+
+void bch2_recovery_pass_status_to_text(struct printbuf *out, struct bch_fs *c)
+{
+	struct bch_fs_recovery *r = &c->recovery;
+
+	printbuf_tabstop_push(out, 32);
+	prt_passes(out, "Scheduled passes", c->sb.recovery_passes_required);
+	prt_passes(out, "Scheduled online passes", c->sb.recovery_passes_required &
+		   bch2_recovery_passes_match(PASS_ONLINE));
+	prt_passes(out, "Complete passes", r->passes_complete);
+	prt_passes(out, "Failing passes", r->passes_failing);
+
+	if (r->curr_pass) {
+		prt_printf(out, "Current pass:\t%s\n", bch2_recovery_passes[r->curr_pass]);
+		prt_passes(out, "Current passes", r->passes_to_run);
+	}
+}
+
+void bch2_fs_recovery_passes_init(struct bch_fs *c)
+{
+	spin_lock_init(&c->recovery.lock);
+	sema_init(&c->recovery.run_lock, 1);
+
+	INIT_WORK(&c->recovery.work, bch2_async_recovery_passes_work);
+}