bcachefs: Fix for buffered writes getting -ENOSPC

Buffered writes may have to increase their disk reservation at btree
update time, due to compression and erasure coding being unpredictable:
O_DIRECT writes should be checking for -ENOSPC, but buffered writes have
already been accepted and should not.

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

fs/bcachefs/btree_update_leaf.c

@@ -690,6 +690,7 @@ int bch2_trans_commit_error(struct btree_trans *trans,
 		}
 		break;
 	case BTREE_INSERT_ENOSPC:
+		BUG_ON(flags & BTREE_INSERT_NOFAIL);
 		ret = -ENOSPC;
 		break;
 	case BTREE_INSERT_NEED_MARK_REPLICAS:
@@ -743,6 +744,8 @@ int bch2_trans_commit_error(struct btree_trans *trans,
 		break;
 	}
 
+	BUG_ON(ret == -ENOSPC && (flags & BTREE_INSERT_NOFAIL));
+
 	return ret;
 }
 

fs/bcachefs/fs-io.c

@@ -1910,6 +1910,7 @@ static long bch2_dio_write_loop(struct dio_write *dio)
 		if ((req->ki_flags & IOCB_DSYNC) &&
 		    !c->opts.journal_flush_disabled)
 			dio->op.flags |= BCH_WRITE_FLUSH;
+		dio->op.flags |= BCH_WRITE_CHECK_ENOSPC;
 
 		ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
 						dio->op.opts.data_replicas, 0);
@@ -2725,7 +2726,7 @@ static int __bchfs_fallocate(struct bch_inode_info *inode, int mode,
 
 		ret = bch2_extent_update(&trans, iter, &reservation.k_i,
 				&disk_res, &inode->ei_journal_seq,
-				0, &i_sectors_delta);
+				0, &i_sectors_delta, true);
 		i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
 bkey_err:
 		bch2_quota_reservation_put(c, inode, &quota_res);

fs/bcachefs/io.c

@@ -197,7 +197,7 @@ int bch2_sum_sector_overwrites(struct btree_trans *trans,
 			       struct btree_iter *extent_iter,
 			       struct bkey_i *new,
 			       bool *maybe_extending,
-			       bool *should_check_enospc,
+			       bool *usage_increasing,
 			       s64 *i_sectors_delta,
 			       s64 *disk_sectors_delta)
 {
@@ -209,7 +209,7 @@ int bch2_sum_sector_overwrites(struct btree_trans *trans,
 	int ret = 0;
 
 	*maybe_extending	= true;
-	*should_check_enospc	= false;
+	*usage_increasing	= false;
 	*i_sectors_delta	= 0;
 	*disk_sectors_delta	= 0;
 
@@ -229,10 +229,10 @@ int bch2_sum_sector_overwrites(struct btree_trans *trans,
 			? sectors * bch2_bkey_nr_ptrs_fully_allocated(old)
 			: 0;
 
-		if (!*should_check_enospc &&
+		if (!*usage_increasing &&
 		    (new_replicas > bch2_bkey_replicas(c, old) ||
 		     (!new_compressed && bch2_bkey_sectors_compressed(old))))
-			*should_check_enospc = true;
+			*usage_increasing = true;
 
 		if (bkey_cmp(old.k->p, new->k.p) >= 0) {
 			/*
@@ -267,11 +267,12 @@ int bch2_extent_update(struct btree_trans *trans,
 		       struct disk_reservation *disk_res,
 		       u64 *journal_seq,
 		       u64 new_i_size,
-		       s64 *i_sectors_delta_total)
+		       s64 *i_sectors_delta_total,
+		       bool check_enospc)
 {
 	/* this must live until after bch2_trans_commit(): */
 	struct bkey_inode_buf inode_p;
-	bool extending = false, should_check_enospc;
+	bool extending = false, usage_increasing;
 	s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 	int ret;
 
@@ -281,17 +282,20 @@ int bch2_extent_update(struct btree_trans *trans,
 
 	ret = bch2_sum_sector_overwrites(trans, iter, k,
 			&extending,
-			&should_check_enospc,
+			&usage_increasing,
 			&i_sectors_delta,
 			&disk_sectors_delta);
 	if (ret)
 		return ret;
 
+	if (!usage_increasing)
+		check_enospc = false;
+
 	if (disk_res &&
 	    disk_sectors_delta > (s64) disk_res->sectors) {
 		ret = bch2_disk_reservation_add(trans->c, disk_res,
 					disk_sectors_delta - disk_res->sectors,
-					!should_check_enospc
+					!check_enospc
 					? BCH_DISK_RESERVATION_NOFAIL : 0);
 		if (ret)
 			return ret;
@@ -346,6 +350,7 @@ int bch2_extent_update(struct btree_trans *trans,
 		bch2_trans_commit(trans, disk_res, journal_seq,
 				BTREE_INSERT_NOCHECK_RW|
 				BTREE_INSERT_NOFAIL);
+	BUG_ON(ret == -ENOSPC);
 	if (ret)
 		return ret;
 
@@ -384,7 +389,7 @@ int bch2_fpunch_at(struct btree_trans *trans, struct btree_iter *iter,
 
 		ret = bch2_extent_update(trans, iter, &delete,
 				&disk_res, journal_seq,
-				0, i_sectors_delta);
+				0, i_sectors_delta, false);
 		bch2_disk_reservation_put(c, &disk_res);
 btree_err:
 		if (ret == -EINTR) {
@@ -457,7 +462,8 @@ int bch2_write_index_default(struct bch_write_op *op)
 
 		ret = bch2_extent_update(&trans, iter, sk.k,
 					 &op->res, op_journal_seq(op),
-					 op->new_i_size, &op->i_sectors_delta);
+					 op->new_i_size, &op->i_sectors_delta,
+					 op->flags & BCH_WRITE_CHECK_ENOSPC);
 		if (ret == -EINTR)
 			continue;
 		if (ret)

fs/bcachefs/io.h

@@ -38,11 +38,12 @@ enum bch_write_flags {
 	BCH_WRITE_ONLY_SPECIFIED_DEVS	= (1 << 6),
 	BCH_WRITE_WROTE_DATA_INLINE	= (1 << 7),
 	BCH_WRITE_FROM_INTERNAL		= (1 << 8),
+	BCH_WRITE_CHECK_ENOSPC		= (1 << 9),
 
 	/* Internal: */
-	BCH_WRITE_JOURNAL_SEQ_PTR	= (1 << 9),
+	BCH_WRITE_JOURNAL_SEQ_PTR	= (1 << 10),
-	BCH_WRITE_SKIP_CLOSURE_PUT	= (1 << 10),
+	BCH_WRITE_SKIP_CLOSURE_PUT	= (1 << 11),
-	BCH_WRITE_DONE			= (1 << 11),
+	BCH_WRITE_DONE			= (1 << 12),
 };
 
 static inline u64 *op_journal_seq(struct bch_write_op *op)
@@ -68,7 +69,7 @@ int bch2_sum_sector_overwrites(struct btree_trans *, struct btree_iter *,
 			       struct bkey_i *, bool *, bool *, s64 *, s64 *);
 int bch2_extent_update(struct btree_trans *, struct btree_iter *,
 		       struct bkey_i *, struct disk_reservation *,
-		       u64 *, u64, s64 *);
+		       u64 *, u64, s64 *, bool);
 int bch2_fpunch_at(struct btree_trans *, struct btree_iter *,
 		   struct bpos, u64 *, s64 *);
 int bch2_fpunch(struct bch_fs *c, u64, u64, u64, u64 *, s64 *);

fs/bcachefs/reflink.c

@@ -293,7 +293,8 @@ s64 bch2_remap_range(struct bch_fs *c,
 
 		ret = bch2_extent_update(&trans, dst_iter, new_dst.k,
 					 &disk_res, journal_seq,
-					 new_i_size, i_sectors_delta);
+					 new_i_size, i_sectors_delta,
+					 true);
 		bch2_disk_reservation_put(c, &disk_res);
 		if (ret)
 			continue;