Hi,
在 2023/12/19 0:04, Song Liu 写道:
On Sun, Dec 17, 2023 at 5:39 PM Yu Kuai <yukuai1@xxxxxxxxxxxxxxx> wrote:
[...]
We only use this for idle or not check, the behavior is OK (I think).
However, this logic is error prone.
On 64-bit systems, there is a 4-byte hole behind sync_io. I think we can
just use it for atomic64_t so that we don't have to worry about overflow.
I'm not sure about this, because other than this ubsan warning, this
overflow doesn't have any impact on functionality to me.
Fixing warnings for zero or low cost is always a good idea. It helps boost
the signal when UBSAN (and other debug features) detects real issues.
If we care about this 'hole', there are lots of holes in gendisk, and
can be avoiled, for example, moving 'sync_io' near to 'node_id'.
The point was not "let's fill the hole", but "we can use atomic64_t
without extra memory cost". In general, I don't think we care too
much about holes in "struct gendisk".
Does this make sense?
Of course, I don't have strong preference on this. Because our syzkaller
reported lots of UBSAN warnings, hence only fix real issues is how we do
it. For upstream, I'm good at fix this warning with zero or low cost.
Thanks,
Kuai
Thanks,
Song
.
