About `bio->bi_iter.bi_size` at write endio time

About `bio->bi_iter.bi_size` at write endio time

[Qu Wenruo]

Hi,

Recently I got a strange case where for write bio, at its endio time, I 
got bio->bi_iter.bi_size == 0, but bio_for_each_segment_all() reports we 
still have some bv_len.

And obviously, when the bio get submitted, its bi_size is not 0.

This is especially common for REQ_OP_ZONE_APPEND bio, but I also get 
rare bi_size == 0 at endio time, for REQ_OP_WRITE too.

So I guess bi_size at endio time is no longer reliable due to bio 
merging/splitting?

Thus the only correct way to get how large a bio really is, is through 
bio_for_each_segment_all()?

Thanks,
Qu

Re: About `bio->bi_iter.bi_size` at write endio time

[Ming Lei]

On Tue, Jun 15, 2021 at 03:22:59PM +0800, Qu Wenruo wrote:
> Hi,
> 
> Recently I got a strange case where for write bio, at its endio time, I got
> bio->bi_iter.bi_size == 0, but bio_for_each_segment_all() reports we still
> have some bv_len.

It is normal to observe zero .bi_size in ->bi_end_io(), see req_bio_endio().
Meantime bio_for_each_segment_all() covers all pages added to the bio via
bio_add_page(), which is invariant after the bio is submitted.

> 
> And obviously, when the bio get submitted, its bi_size is not 0.
> 
> This is especially common for REQ_OP_ZONE_APPEND bio, but I also get rare
> bi_size == 0 at endio time, for REQ_OP_WRITE too.

It shouldn't be rare.

> 
> So I guess bi_size at endio time is no longer reliable due to bio
> merging/splitting?

No, ->bi_size should be zero in .bi_end_io() if this bio is completed
successfully no matter if the bio is splitted or not.

> 
> Thus the only correct way to get how large a bio really is, is through
> bio_for_each_segment_all()?

Yeah, if you mean to get the bio's real size in ->bi_end_io().


Thanks,
Ming