"John Garry" <john.g.garry@xxxxxxxxxx> writes:
> On 18/02/2025 11:40, Andreas Hindborg wrote:
>> "John Garry" <john.g.garry@xxxxxxxxxx> writes:
>>
>>> On 15/02/2025 10:58, Andreas Hindborg wrote:
>>>> When cloning a bio, the `bio.bi_vcnt` field is not cloned. This is a
>>>> problem if users want to perform bounds checks on the `bio.bi_io_vec`
>>>> field.
>>>
>>> Is this fixing a potential problem? Or fixing a real issue?
>>
>> It is fixing a problem I ran into in rnull, the rust null block
>> implementation. When running with debug assertions enabled, a bound
>> check on `bi_io_vec` fails for split bio, because `bio_vcnt` becomes
>> zero in the cloned bio.
>>
>> I can work around this by not using a slice type to represent
>> `bi_io_vec` in rust, not a big deal.
>>
>> But I am genuinely curious if there is a reason for not setting
>> `bi_vcnt` during a clone.
>
> I think that it came from commit 59d276fe0 (with the addition of
> bio_clone_fast()), where we assume that the cloned bio is not having the
> bio_vec touched and so does not need to know bi_vcnt (or bi_max_vecs).
> And it is inefficient to needlessly set bi_vcnt then.
I see. That is a few days ago. I am quite confident that for modern
hardware and workloads, this assignment will not have any measurable
impact on performance.
Can we add it back?
I understand if you would prefer not to, since it is not strictly
necessary. But in that case, I would suggest patching the documentation
of `struct bio` something like this:
--- a/include/linux/blk_types.h
+++ b/include/linux/blk_types.h
@@ -255,7 +255,8 @@ struct bio {
struct bio_integrity_payload *bi_integrity; /* data integrity */
#endif
- unsigned short bi_vcnt; /* how many bio_vec's */
+ unsigned short bi_vcnt; /* how many bio_vec's. Not valid if this bio is
+ a clone (flagged BIO_CLONED). */
/*
* Everything starting with bi_max_vecs will be preserved by bio_reset()
Best regards,
Andreas Hindborg
