Update bcachefs sources to 09be9a5a5147 bcachefs; fix bch2_dev_get_ioref()

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

.bcachefs_revision

@@ -1 +1 @@
-eab5671b52626036abd5a31e7743c74fb3b59635
+09be9a5a514734ebdd6fd840964815b2b0823c66

libbcachefs/btree/iter.c

@@ -3557,6 +3557,13 @@ static inline struct btree_trans *bch2_trans_alloc(struct bch_fs *c)
 struct btree_trans *__bch2_trans_get(struct bch_fs *c, unsigned fn_idx)
 	__acquires(&c->btree_trans_barrier)
 {
+	/*
+	 * No multithreaded btree access until we've gone RW and are no longer
+	 * modifying the journal keys gap buffer
+	 */
+	EBUG_ON(!test_bit(BCH_FS_may_go_rw, &c->flags) &&
+		current != c->recovery_task);
+
 	struct btree_trans *trans = bch2_trans_alloc(c);
 
 	trans->c		= c;
@@ -3574,6 +3581,7 @@ struct btree_trans *__bch2_trans_get(struct bch_fs *c, unsigned fn_idx)
 
 	*trans_paths_nr(trans->paths) = BTREE_ITER_INITIAL;
 
+	/* Reserve path 0 for our sentinal value */
 	trans->paths_allocated[0] = 1;
 
 	static struct lock_class_key lockdep_key;

libbcachefs/btree/journal_overlay.c

@@ -284,7 +284,9 @@ int bch2_journal_key_insert_take(struct bch_fs *c, enum btree_id id,
 	struct journal_keys *keys = &c->journal_keys;
 	size_t idx = bch2_journal_key_search(keys, id, level, k->k.p);
 
+	BUG_ON(test_bit(BCH_FS_may_go_rw, &c->flags));
 	BUG_ON(test_bit(BCH_FS_rw, &c->flags));
+	BUG_ON(current != c->recovery_task);
 
 	if (idx < keys->size &&
 	    journal_key_cmp(c, &n, &keys->data[idx]) == 0) {
@@ -659,8 +661,10 @@ void __bch2_btree_and_journal_iter_init_node_iter(struct btree_trans *trans,
 
 	if (trans->journal_replay_not_finished) {
 		bch2_journal_iter_init(trans->c, &iter->journal, b->c.btree_id, b->c.level, pos);
-		if (!test_bit(BCH_FS_may_go_rw, &trans->c->flags))
+		if (!test_bit(BCH_FS_may_go_rw, &trans->c->flags)) {
+			BUG_ON(current != trans->c->recovery_task);
 			list_add(&iter->journal.list, &trans->c->journal_iters);
+		}
 	}
 }
 

libbcachefs/data/copygc.c

@@ -475,6 +475,8 @@ static int bch2_copygc_thread(void *arg)
 	 */
 	kthread_wait_freezable(c->recovery.pass_done > BCH_RECOVERY_PASS_check_snapshots ||
 			       kthread_should_stop());
+	if (kthread_should_stop())
+		return 0;
 
 	bch2_move_stats_init(&move_stats, "copygc");
 	bch2_moving_ctxt_init(&ctxt, c, NULL, &move_stats,

libbcachefs/data/read.c

@@ -1043,142 +1043,41 @@ static inline bool can_narrow_crc(struct bch_extent_crc_unpacked n)
 		!crc_is_compressed(n);
 }
 
-int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig,
+static inline struct bch_read_bio *read_extent_rbio_alloc(struct btree_trans *trans,
-		       struct bvec_iter iter, struct bpos read_pos,
+			struct bch_read_bio *orig,
-		       enum btree_id data_btree, struct bkey_s_c k,
+			struct bvec_iter iter, struct bpos read_pos,
-		       unsigned offset_into_extent,
+			enum btree_id data_btree, struct bkey_s_c k,
-		       struct bch_io_failures *failed, unsigned flags, int dev)
+			struct extent_ptr_decoded pick,
+			struct bch_dev *ca,
+			unsigned offset_into_extent,
+			struct bch_io_failures *failed, unsigned flags,
+			bool bounce, bool read_full, bool narrow_crcs)
 {
 	struct bch_fs *c = trans->c;
-	struct extent_ptr_decoded pick;
-	struct bch_read_bio *rbio = NULL;
-	bool bounce = false, read_full = false, narrow_crcs = false;
 	struct bpos data_pos = bkey_start_pos(k.k);
-	struct data_update *u = rbio_data_update(orig);
-	int ret = 0;
 
-	if (bkey_extent_is_inline_data(k.k)) {
+	struct bch_read_bio *rbio = orig->opts.promote_target || have_io_error(failed)
-		unsigned bytes = min_t(unsigned, iter.bi_size,
+		? promote_alloc(trans, iter, k, &pick, flags, orig,
-				       bkey_inline_data_bytes(k.k));
+				&bounce, &read_full, failed)
-
+		: NULL;
-		swap(iter.bi_size, bytes);
-		memcpy_to_bio(&orig->bio, iter, bkey_inline_data_p(k));
-		swap(iter.bi_size, bytes);
-		bio_advance_iter(&orig->bio, &iter, bytes);
-		zero_fill_bio_iter(&orig->bio, iter);
-		this_cpu_add(c->counters.now[BCH_COUNTER_io_read_inline],
-			     bvec_iter_sectors(iter));
-		goto out_read_done;
-	}
-
-	if ((bch2_bkey_extent_flags(k) & BIT_ULL(BCH_EXTENT_FLAG_poisoned)) &&
-	    !orig->data_update) {
-		ret = bch_err_throw(c, extent_poisoned);
-		goto err;
-	}
-retry_pick:
-	ret = bch2_bkey_pick_read_device(c, k, failed, &pick, dev);
-
-	/* hole or reservation - just zero fill: */
-	if (!ret)
-		goto hole;
-
-	if (unlikely(ret < 0)) {
-		if (ret == -BCH_ERR_data_read_csum_err) {
-			int ret2 = maybe_poison_extent(trans, orig, data_btree, k);
-			if (ret2) {
-				ret = ret2;
-				goto err;
-			}
-
-			trace_and_count(c, io_read_fail_and_poison, &orig->bio);
-		}
-
-		if (!(flags & BCH_READ_in_retry)) {
-			CLASS(printbuf, buf)();
-			bch2_read_err_msg_trans(trans, &buf, orig, read_pos);
-			prt_printf(&buf, "%s\n  ", bch2_err_str(ret));
-			bch2_bkey_val_to_text(&buf, c, k);
-			bch_err_ratelimited(c, "%s", buf.buf);
-		}
-		goto err;
-	}
-
-	if (unlikely(bch2_csum_type_is_encryption(pick.crc.csum_type)) &&
-	    !c->chacha20_key_set) {
-		if (!(flags & BCH_READ_in_retry)) {
-			CLASS(printbuf, buf)();
-			bch2_read_err_msg_trans(trans, &buf, orig, read_pos);
-			prt_printf(&buf, "attempting to read encrypted data without encryption key\n  ");
-			bch2_bkey_val_to_text(&buf, c, k);
-
-			bch_err_ratelimited(c, "%s", buf.buf);
-		}
-		ret = bch_err_throw(c, data_read_no_encryption_key);
-		goto err;
-	}
-
-	struct bch_dev *ca = bch2_dev_get_ioref(c, pick.ptr.dev, READ,
-					BCH_DEV_READ_REF_io_read);
 
 	/*
-	 * Stale dirty pointers are treated as IO errors, but @failed isn't
+	 * If it's being moved internally, we don't want to flag it as a cache
-	 * allocated unless we're in the retry path - so if we're not in the
+	 * hit:
-	 * retry path, don't check here, it'll be caught in bch2_read_endio()
-	 * and we'll end up in the retry path:
 	 */
-	if ((flags & BCH_READ_in_retry) &&
+	if (ca && pick.ptr.cached && !orig->data_update)
-	    !pick.ptr.cached &&
+		bch2_bucket_io_time_reset(trans, pick.ptr.dev,
-	    ca &&
+			PTR_BUCKET_NR(ca, &pick.ptr), READ);
-	    unlikely(dev_ptr_stale(ca, &pick.ptr))) {
-		read_from_stale_dirty_pointer(trans, ca, k, pick.ptr);
-		bch2_mark_io_failure(failed, &pick, bch_err_throw(c, data_read_ptr_stale_dirty));
-		propagate_io_error_to_data_update(c, rbio, &pick);
-		enumerated_ref_put(&ca->io_ref[READ], BCH_DEV_READ_REF_io_read);
-		goto retry_pick;
-	}
 
-	if (likely(!u)) {
+	/*
-		if (!(flags & BCH_READ_last_fragment) ||
+	 * Done with btree operations:
-		    bio_flagged(&orig->bio, BIO_CHAIN))
+	 * Unlock the iterator while the btree node's lock is still in cache,
-			flags |= BCH_READ_must_clone;
+	 * before allocating the clone/fragment (if any) and doing the IO:
-
+	 */
-		narrow_crcs = !(flags & BCH_READ_in_retry) && can_narrow_crc(pick.crc);
+	if (!(flags & BCH_READ_in_retry))
-
+		bch2_trans_unlock(trans);
-		if (narrow_crcs && (flags & BCH_READ_user_mapped))
+	else
-			flags |= BCH_READ_must_bounce;
+		bch2_trans_unlock_long(trans);
-
-		EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
-
-		if (crc_is_compressed(pick.crc) ||
-		    (pick.crc.csum_type != BCH_CSUM_none &&
-		     (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
-		      (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
-		       (flags & BCH_READ_user_mapped)) ||
-		      (flags & BCH_READ_must_bounce)))) {
-			read_full = true;
-			bounce = true;
-		}
-	} else {
-		/*
-		 * can happen if we retry, and the extent we were going to read
-		 * has been merged in the meantime:
-		 */
-		if (pick.crc.compressed_size > u->op.wbio.bio.bi_iter.bi_size) {
-			if (ca)
-				enumerated_ref_put(&ca->io_ref[READ],
-					BCH_DEV_READ_REF_io_read);
-			rbio->ret = bch_err_throw(c, data_read_buffer_too_small);
-			goto out_read_done;
-		}
-
-		iter.bi_size	= pick.crc.compressed_size << 9;
-		read_full = true;
-	}
-
-	if (orig->opts.promote_target || have_io_error(failed))
-		rbio = promote_alloc(trans, iter, k, &pick, flags, orig,
-				     &bounce, &read_full, failed);
 
 	if (!read_full) {
 		EBUG_ON(crc_is_compressed(pick.crc));
@@ -1264,38 +1163,227 @@ retry_pick:
 	rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
 	rbio->bio.bi_end_io	= bch2_read_endio;
 
-	async_object_list_add(c, rbio, rbio, &rbio->list_idx);
-
-	if (rbio->bounce)
-		trace_and_count(c, io_read_bounce, &rbio->bio);
-
-	if (!u)
-		this_cpu_add(c->counters.now[BCH_COUNTER_io_read], bio_sectors(&rbio->bio));
-	else
-		this_cpu_add(c->counters.now[BCH_COUNTER_io_move_read], bio_sectors(&rbio->bio));
-	bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
-
-	/*
-	 * If it's being moved internally, we don't want to flag it as a cache
-	 * hit:
-	 */
-	if (ca && pick.ptr.cached && !u)
-		bch2_bucket_io_time_reset(trans, pick.ptr.dev,
-			PTR_BUCKET_NR(ca, &pick.ptr), READ);
-
 	if (!(flags & (BCH_READ_in_retry|BCH_READ_last_fragment))) {
 		bio_inc_remaining(&orig->bio);
 		trace_and_count(c, io_read_split, &orig->bio);
 	}
 
-	/*
+	async_object_list_add(c, rbio, rbio, &rbio->list_idx);
-	 * Unlock the iterator while the btree node's lock is still in
+
-	 * cache, before doing the IO:
+	if (rbio->bounce)
-	 */
+		trace_and_count(c, io_read_bounce, &rbio->bio);
-	if (!(flags & BCH_READ_in_retry))
+
-		bch2_trans_unlock(trans);
+	if (!orig->data_update)
+		this_cpu_add(c->counters.now[BCH_COUNTER_io_read], bio_sectors(&rbio->bio));
 	else
-		bch2_trans_unlock_long(trans);
+		this_cpu_add(c->counters.now[BCH_COUNTER_io_move_read], bio_sectors(&rbio->bio));
+	bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
+	return rbio;
+}
+
+static inline int read_extent_done(struct bch_read_bio *rbio, unsigned flags, int ret)
+{
+	if (flags & BCH_READ_in_retry)
+		return ret;
+
+	if (ret)
+		rbio->ret = ret;
+
+	if (flags & BCH_READ_last_fragment)
+		bch2_rbio_done(rbio);
+	return 0;
+}
+
+static noinline int read_extent_inline(struct bch_fs *c,
+				       struct bch_read_bio *rbio,
+				       struct bvec_iter iter,
+				       struct bkey_s_c k,
+				       unsigned offset_into_extent,
+				       unsigned flags)
+{
+	this_cpu_add(c->counters.now[BCH_COUNTER_io_read_inline], bvec_iter_sectors(iter));
+
+	unsigned bytes = min(iter.bi_size, offset_into_extent << 9);
+	swap(iter.bi_size, bytes);
+	zero_fill_bio_iter(&rbio->bio, iter);
+	swap(iter.bi_size, bytes);
+
+	bio_advance_iter(&rbio->bio, &iter, bytes);
+
+	bytes = min(iter.bi_size, bkey_inline_data_bytes(k.k));
+
+	swap(iter.bi_size, bytes);
+	memcpy_to_bio(&rbio->bio, iter, bkey_inline_data_p(k));
+	swap(iter.bi_size, bytes);
+
+	bio_advance_iter(&rbio->bio, &iter, bytes);
+
+	zero_fill_bio_iter(&rbio->bio, iter);
+
+	return read_extent_done(rbio, flags, 0);
+}
+
+static noinline int read_extent_hole(struct bch_fs *c,
+				     struct bch_read_bio *rbio,
+				     struct bvec_iter iter,
+				     unsigned flags)
+{
+	this_cpu_add(c->counters.now[BCH_COUNTER_io_read_hole],
+		     bvec_iter_sectors(iter));
+	/*
+	 * won't normally happen in the data update (bch2_move_extent()) path,
+	 * but if we retry and the extent we wanted to read no longer exists we
+	 * have to signal that:
+	 */
+	if (rbio->data_update)
+		rbio->ret = bch_err_throw(c, data_read_key_overwritten);
+
+	zero_fill_bio_iter(&rbio->bio, iter);
+
+	return read_extent_done(rbio, flags, 0);
+}
+
+static noinline int read_extent_pick_err(struct btree_trans *trans,
+					 struct bch_read_bio *rbio,
+					 struct bpos read_pos,
+					 enum btree_id data_btree, struct bkey_s_c k,
+					 unsigned flags, int ret)
+{
+	struct bch_fs *c = trans->c;
+
+	if (ret == -BCH_ERR_data_read_csum_err) {
+		/* We can only return errors directly in the retry path */
+		BUG_ON(!(flags & BCH_READ_in_retry));
+
+		try(maybe_poison_extent(trans, rbio, data_btree, k));
+		trace_and_count(c, io_read_fail_and_poison, &rbio->bio);
+	}
+
+	if (!(flags & BCH_READ_in_retry)) {
+		CLASS(printbuf, buf)();
+		bch2_read_err_msg_trans(trans, &buf, rbio, read_pos);
+		prt_printf(&buf, "%s\n  ", bch2_err_str(ret));
+		bch2_bkey_val_to_text(&buf, c, k);
+		bch_err_ratelimited(c, "%s", buf.buf);
+	}
+
+	return read_extent_done(rbio, flags, ret);
+}
+
+static noinline int read_extent_no_encryption_key(struct btree_trans *trans,
+					 struct bch_read_bio *rbio,
+					 struct bpos read_pos,
+					 struct bkey_s_c k,
+					 unsigned flags)
+{
+	struct bch_fs *c = trans->c;
+
+	CLASS(printbuf, buf)();
+	bch2_read_err_msg_trans(trans, &buf, rbio, read_pos);
+	prt_printf(&buf, "attempting to read encrypted data without encryption key\n  ");
+	bch2_bkey_val_to_text(&buf, c, k);
+
+	bch_err_ratelimited(c, "%s", buf.buf);
+
+	return read_extent_done(rbio, flags, bch_err_throw(c, data_read_no_encryption_key));
+}
+
+int __bch2_read_extent(struct btree_trans *trans,
+		       struct bch_read_bio *orig,
+		       struct bvec_iter iter, struct bpos read_pos,
+		       enum btree_id data_btree, struct bkey_s_c k,
+		       unsigned offset_into_extent,
+		       struct bch_io_failures *failed, unsigned flags, int dev)
+{
+	struct bch_fs *c = trans->c;
+	struct extent_ptr_decoded pick;
+	bool bounce = false, read_full = false, narrow_crcs = false;
+	struct data_update *u = rbio_data_update(orig);
+	int ret = 0;
+
+	if (bkey_extent_is_inline_data(k.k))
+		return read_extent_inline(c, orig, iter, k, offset_into_extent, flags);
+
+	if (unlikely((bch2_bkey_extent_flags(k) & BIT_ULL(BCH_EXTENT_FLAG_poisoned))) &&
+	    !orig->data_update)
+		return read_extent_done(orig, flags, bch_err_throw(c, extent_poisoned));
+
+	ret = bch2_bkey_pick_read_device(c, k, failed, &pick, dev);
+
+	/* hole or reservation - just zero fill: */
+	if (unlikely(!ret))
+		return read_extent_hole(c, orig, iter, flags);
+
+	if (unlikely(ret < 0))
+		return read_extent_pick_err(trans, orig, read_pos, data_btree, k, flags, ret);
+
+	if (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
+	    unlikely(!c->chacha20_key_set))
+		return read_extent_no_encryption_key(trans, orig, read_pos, k, flags);
+
+	struct bch_dev *ca = bch2_dev_get_ioref(c, pick.ptr.dev, READ,
+					BCH_DEV_READ_REF_io_read);
+
+	/*
+	 * Stale dirty pointers are treated as IO errors, but @failed isn't
+	 * allocated unless we're in the retry path - so if we're not in the
+	 * retry path, don't check here, it'll be caught in bch2_read_endio()
+	 * and we'll end up in the retry path:
+	 */
+	if (unlikely(flags & BCH_READ_in_retry) &&
+	    !pick.ptr.cached &&
+	    ca &&
+	    unlikely(dev_ptr_stale(ca, &pick.ptr))) {
+		enumerated_ref_put(&ca->io_ref[READ], BCH_DEV_READ_REF_io_read);
+		read_from_stale_dirty_pointer(trans, ca, k, pick.ptr);
+
+		bch2_mark_io_failure(failed, &pick, ret);
+		propagate_io_error_to_data_update(c, orig, &pick);
+
+		return read_extent_done(orig, flags, bch_err_throw(c, data_read_ptr_stale_dirty));
+	}
+
+	if (likely(!u)) {
+		if (!(flags & BCH_READ_last_fragment) ||
+		    bio_flagged(&orig->bio, BIO_CHAIN))
+			flags |= BCH_READ_must_clone;
+
+		narrow_crcs = !(flags & BCH_READ_in_retry) && can_narrow_crc(pick.crc);
+
+		if (narrow_crcs && (flags & BCH_READ_user_mapped))
+			flags |= BCH_READ_must_bounce;
+
+		EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
+
+		if (crc_is_compressed(pick.crc) ||
+		    (pick.crc.csum_type != BCH_CSUM_none &&
+		     (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
+		      (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
+		       (flags & BCH_READ_user_mapped)) ||
+		      (flags & BCH_READ_must_bounce)))) {
+			read_full = true;
+			bounce = true;
+		}
+	} else {
+		/*
+		 * can happen if we retry, and the extent we were going to read
+		 * has been merged in the meantime:
+		 */
+		if (unlikely(pick.crc.compressed_size > u->op.wbio.bio.bi_iter.bi_size)) {
+			if (ca)
+				enumerated_ref_put(&ca->io_ref[READ],
+					BCH_DEV_READ_REF_io_read);
+			return read_extent_done(orig, flags, bch_err_throw(c, data_read_buffer_too_small));
+		}
+
+		iter.bi_size	= pick.crc.compressed_size << 9;
+		read_full = true;
+	}
+
+	struct bch_read_bio *rbio =
+		read_extent_rbio_alloc(trans, orig, iter, read_pos, data_btree, k,
+				       pick, ca, offset_into_extent, failed, flags,
+				       bounce, read_full, narrow_crcs);
 
 	if (likely(!rbio->pick.do_ec_reconstruct)) {
 		if (unlikely(!rbio->have_ioref)) {
@@ -1340,8 +1428,6 @@ out:
 	} else {
 		bch2_trans_unlock(trans);
 
-		int ret;
-
 		rbio->context = RBIO_CONTEXT_UNBOUND;
 		bch2_read_endio(&rbio->bio);
 
@@ -1356,31 +1442,6 @@ out:
 
 		return ret;
 	}
-
-err:
-	if (flags & BCH_READ_in_retry)
-		return ret;
-
-	orig->ret = ret;
-	goto out_read_done;
-
-hole:
-	this_cpu_add(c->counters.now[BCH_COUNTER_io_read_hole],
-		     bvec_iter_sectors(iter));
-	/*
-	 * won't normally happen in the data update (bch2_move_extent()) path,
-	 * but if we retry and the extent we wanted to read no longer exists we
-	 * have to signal that:
-	 */
-	if (u)
-		orig->ret = bch_err_throw(c, data_read_key_overwritten);
-
-	zero_fill_bio_iter(&orig->bio, iter);
-out_read_done:
-	if ((flags & BCH_READ_last_fragment) &&
-	    !(flags & BCH_READ_in_retry))
-		bch2_rbio_done(orig);
-	return 0;
 }
 
 int __bch2_read(struct btree_trans *trans, struct bch_read_bio *rbio,

libbcachefs/data/reconcile.c

@@ -1848,6 +1848,8 @@ static int bch2_reconcile_thread(void *arg)
 	 */
 	kthread_wait_freezable(c->recovery.pass_done > BCH_RECOVERY_PASS_check_snapshots ||
 			       kthread_should_stop());
+	if (kthread_should_stop())
+		return 0;
 
 	struct moving_context ctxt __cleanup(bch2_moving_ctxt_exit);
 	bch2_moving_ctxt_init(&ctxt, c, NULL, &r->work_stats,
@@ -1940,6 +1942,13 @@ void bch2_reconcile_status_to_text(struct printbuf *out, struct bch_fs *c)
 
 void bch2_reconcile_scan_pending_to_text(struct printbuf *out, struct bch_fs *c)
 {
+	/*
+	 * No multithreaded btree access until BCH_FS_may_go_rw and we're no
+	 * longer modifying the journal keys gap buffer:
+	 */
+	if (!test_bit(BCH_FS_may_go_rw, &c->flags))
+		return;
+
 	CLASS(btree_trans, trans)(c);
 	CLASS(btree_iter, iter)(trans, BTREE_ID_reconcile_scan, POS_MIN, 0);
 

libbcachefs/debug/debug.c

@@ -366,6 +366,13 @@ static ssize_t bch2_read_btree(struct file *file, char __user *buf,
 	i->size	= size;
 	i->ret	= 0;
 
+	/*
+	 * No multithreaded btree access until BCH_FS_may_go_rw and we're no
+	 * longer modifying the journal keys gap buffer:
+	 */
+	if (!test_bit(BCH_FS_may_go_rw, &i->c->flags))
+		return 0;
+
 	CLASS(btree_trans, trans)(i->c);
 	return bch2_debugfs_flush_buf(i) ?:
 		for_each_btree_key(trans, iter, i->id, i->from,
@@ -396,6 +403,9 @@ static ssize_t bch2_read_btree_formats(struct file *file, char __user *buf,
 	i->size	= size;
 	i->ret	= 0;
 
+	if (!test_bit(BCH_FS_may_go_rw, &i->c->flags))
+		return 0;
+
 	try(bch2_debugfs_flush_buf(i));
 
 	if (bpos_eq(SPOS_MAX, i->from))
@@ -428,6 +438,9 @@ static ssize_t bch2_read_bfloat_failed(struct file *file, char __user *buf,
 	i->size	= size;
 	i->ret	= 0;
 
+	if (!test_bit(BCH_FS_may_go_rw, &i->c->flags))
+		return 0;
+
 	CLASS(btree_trans, trans)(i->c);
 	return bch2_debugfs_flush_buf(i) ?:
 		for_each_btree_key(trans, iter, i->id, i->from,

libbcachefs/init/fs.c

@@ -1030,6 +1030,7 @@ static int bch2_fs_init(struct bch_fs *c, struct bch_sb *sb,
 	CLASS(printbuf, name)();
 
 	c->stdio = (void *)(unsigned long) opts->stdio;
+	c->recovery_task = current;
 
 	__module_get(THIS_MODULE);
 
@@ -1245,7 +1246,6 @@ static int bch2_fs_init(struct bch_fs *c, struct bch_sb *sb,
 	scoped_guard(mutex, &bch2_fs_list_lock)
 		try(bch2_fs_online(c));
 
-	c->recovery_task = current;
 	return 0;
 }
 

libbcachefs/init/passes.c

@@ -226,30 +226,6 @@ static int bch2_recovery_pass_empty(struct bch_fs *c)
 	return 0;
 }
 
-static int bch2_set_may_go_rw(struct bch_fs *c)
-{
-	struct journal_keys *keys = &c->journal_keys;
-
-	/*
-	 * After we go RW, the journal keys buffer can't be modified (except for
-	 * setting journal_key->overwritten: it will be accessed by multiple
-	 * threads
-	 */
-	move_gap(keys, keys->nr);
-
-	set_bit(BCH_FS_may_go_rw, &c->flags);
-
-	if (go_rw_in_recovery(c)) {
-		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:

libbcachefs/init/recovery.c

@@ -202,6 +202,30 @@ static void zero_out_btree_mem_ptr(struct journal_keys *keys)
 	}
 }
 
+int bch2_set_may_go_rw(struct bch_fs *c)
+{
+	struct journal_keys *keys = &c->journal_keys;
+
+	/*
+	 * After we go RW, the journal keys buffer can't be modified (except for
+	 * setting journal_key->overwritten: it will be accessed by multiple
+	 * threads
+	 */
+	move_gap(keys, keys->nr);
+
+	set_bit(BCH_FS_may_go_rw, &c->flags);
+
+	if (go_rw_in_recovery(c)) {
+		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;
+}
+
 /* journal replay: */
 
 static void replay_now_at(struct journal *j, u64 seq)
@@ -344,13 +368,13 @@ int bch2_journal_replay(struct bch_fs *c)
 	bool immediate_flush = false;
 	int ret = 0;
 
+	BUG_ON(!atomic_read(&keys->ref));
+	BUG_ON(keys->gap != keys->nr);
+
 	if (keys->nr)
 		try(bch2_journal_log_msg(c, "Starting journal replay (%zu keys in entries %llu-%llu)",
 					 keys->nr, start_seq, end_seq));
 
-	BUG_ON(!atomic_read(&keys->ref));
-
-	move_gap(keys, keys->nr);
 	CLASS(btree_trans, trans)(c);
 
 	/*
@@ -960,8 +984,7 @@ int bch2_fs_initialize(struct bch_fs *c)
 	struct journal_start_info journal_start = { .start_seq = 1 };
 	try(bch2_fs_journal_start(&c->journal, journal_start));
 
-	set_bit(BCH_FS_may_go_rw, &c->flags);
+	try(bch2_set_may_go_rw(c));
-	try(bch2_fs_read_write_early(c));
 	try(bch2_journal_replay(c));
 	try(bch2_fs_freespace_init(c));
 	try(bch2_initialize_subvolumes(c));

libbcachefs/init/recovery.h

@@ -5,6 +5,7 @@
 int bch2_btree_lost_data(struct bch_fs *, struct printbuf *, enum btree_id);
 void bch2_reconstruct_alloc(struct bch_fs *);
 
+int bch2_set_may_go_rw(struct bch_fs *);
 int bch2_journal_replay(struct bch_fs *);
 
 int bch2_fs_recovery(struct bch_fs *);

libbcachefs/sb/members.h

@@ -332,8 +332,7 @@ static inline struct bch_dev *bch2_dev_get_ioref(struct bch_fs *c, unsigned dev,
 	if (!ca || !enumerated_ref_tryget(&ca->io_ref[rw], ref_idx))
 		return NULL;
 
-	if (ca->mi.state == BCH_MEMBER_STATE_rw ||
+	if (ca->mi.state == BCH_MEMBER_STATE_rw || rw == READ)
-	    (ca->mi.state == BCH_MEMBER_STATE_ro && rw == READ))
 		return ca;
 
 	enumerated_ref_put(&ca->io_ref[rw], ref_idx);