On 2021/8/19 18:26, Jan Kara wrote:

On Thu 19-08-21 14:57:03, Zhang Yi wrote:

quoted

In ext4_inode_blocks_set(), huge_file feature should exist when setting
i_blocks beyond a 32 bit variable could be represented, return EFBIG if
not. This error should never happen in theory since sb->s_maxbytes should
not have allowed this, and we have already init sb->s_maxbytes according
to this feature in ext4_fill_super(). So switch to use WARN_ON_ONCE
instead.

Signed-off-by: Zhang Yi <yi.zhang@huawei.com>
---

One comment below:

quoted

@@ -4918,10 +4918,15 @@ static int ext4_inode_blocks_set(handle_t *handle,
 		raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
 		raw_inode->i_blocks_high = 0;
 		ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
-		return 0;
+		return;
 	}
-	if (!ext4_has_feature_huge_file(sb))
-		return -EFBIG;
+
+	/*
+	 * This should never happen since sb->s_maxbytes should not have
+	 * allowed this, which was set according to the huge_file feature
+	 * in ext4_fill_super().
+	 */
+	WARN_ON_ONCE(!ext4_has_feature_huge_file(sb));

Thinking about this a bit more, this could also happen due to fs
corruption. So we probably need to call ext4_error_inode() here instead of
WARN_ON_ONCE(). Also it will result in properly marking fs as having
errors. But since we hold i_raw_lock at this call site we need to
keep the error bail out from ext4_inode_blocks_set() and in
ext4_do_update_inode() finish updating inode and then call
ext4_error_inode() after dropping i_raw_lock.

Yes, make sense, ext4_error_inode() is more reasonable.

Thanks,
Yi.