Re: [PATCH 2/2] block: remove the bi_inline_vecs variable sized array from struct bio

[Christoph Hellwig <hch@lst.de>]

On Tue, Sep 09, 2025 at 04:40:33PM +0800, Yu Kuai wrote:
> Hi,
>
> 在 2025/09/09 16:16, John Garry 写道:
>>> diff --git a/drivers/md/bcache/movinggc.c b/drivers/md/bcache/movinggc.c
>>> index 4fc80c6d5b31..73918e55bf04 100644
>>> --- a/drivers/md/bcache/movinggc.c
>>> +++ b/drivers/md/bcache/movinggc.c
>>> @@ -145,9 +145,9 @@ static void read_moving(struct cache_set *c)
>>>               continue;
>>>           }
>>> -        io = kzalloc(struct_size(io, bio.bio.bi_inline_vecs,
>>> -                     DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS)),
>>> -                 GFP_KERNEL);
>>> +        io = kzalloc(sizeof(*io) + sizeof(struct bio_vec) *
>>> +                DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS),
>>> +                GFP_KERNEL);
>>
>> this seems a common pattern, so maybe another helper (which could be used 
>> by bio_kmalloc)? I am not advocating it, but just putting the idea out 
>> there... too many helpers makes it messy IMHO
>
> Not sure how to do this, do you mean a marco to pass in the base
> structure type, nr_vecs and the gfp_mask?

It's not a very common pattern, and should be even less common.
If a driver / file system wants to embed a bio into its own structure
but dynamically allocate the bio_vecs, it can simply use kmalloc_array,
or add a bio_vec VLA to its own structure and use struct_size on that.