On 8/14/18 10:24 AM, Jens Axboe wrote:
> On 8/14/18 10:22 AM, Linus Torvalds wrote:
>> On Tue, Aug 14, 2018 at 8:24 AM Ilya Dryomov <idryo...@gmail.com> wrote:
>>>
>>> Looks like it's coming from that fsync():
>>>
>>> sys_fsync
>>> do_fsync
>>> vfs_fsync_range
>>> blkdev_fsync
>>> blkdev_issue_flush
>>>
>>> I think we need to teach blkdev_issue_flush() to bail out if the bdev
>>> is read-only, similar to blkdev_issue_discard(), _write_zeroes(), etc.
>>> The question is which error code to use. blkdev_fsync() already skips
>>> over EOPNOTSUPP, so it is a (no-so-good) option. Other blkdev_issue_
>>> functions return EPERM.
>>
>> Oh, just make issue_flush() return EROFS for a read-only device.
>>
>> Or maybe we should even just consider the flush to be a read operation?
>>
>> But I guess the error code gets percolated all the way to user space?
>> The safest option might just be to return 0.
>
> We probably just want to special case a flush for this check. In other
> situations, like resource allocation and issue, we'd want to consider
> it a write.
Ala:
diff --git a/block/blk-core.c b/block/blk-core.c
index 12550340418d..29f8a60965bc 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -2162,7 +2163,9 @@ static inline bool should_fail_request(struct hd_struct
*part,
static inline bool bio_check_ro(struct bio *bio, struct hd_struct *part)
{
- if (part->policy && op_is_write(bio_op(bio))) {
+ const int op = bio_op(bio);
+
+ if (part->policy && (op_is_write(op) && !op_is_flush(op))) {
char b[BDEVNAME_SIZE];
WARN_ONCE(1,
--
Jens Axboe
