Re: [PATCH] md/raid1: change r1conf->r1bio_pool to a pointer type

[Wang Jinchao]

On 6/18/25 19:41, Wang Jinchao wrote:
> > In raid1_reshape(), newpool is a stack variable.
> > mempool_init() initializes newpool->wait with the stack address.
> > After assigning newpool to conf->r1bio_pool, the wait queue
> > need to be reinitialized, which is not ideal.
> >
> > Change raid1_conf->r1bio_pool to a pointer type and
> > replace mempool_init() with mempool_create() to
> > avoid referencing a stack-based wait queue.


> Can you also switch to kmalloc pool in this patch?

> Thanks,
> Kuai

Hi Kuai,

Comparing mempool_create_kmalloc_pool() and mempool_create(), the former 
requires the pool element size as a parameter, while the latter uses 
r1bio_pool_alloc() to allocate new elements, with the size calculated 
based on poolinfo->raid_disks.
The key point is poolinfo, which is used for both r1bio_pool and r1buf_pool.
If we change from mempool_create() to mempool_create_kmalloc_pool(), we 
would need to introduce a new concept, such as r1bio_pool_size, and 
store it somewhere. In this case, the original conf->poolinfo would lose 
its meaning and become just r1buf_poolinfo.
So I think keeping poolinfo is a better fit for the pool in RAID1.

By the way, I did not receive your email in my Gmail inbox; I found your 
message on lore.org. The last email I received from you was on June 16, 
so I am not sure what the problem is.
I also sent you an email mentioning that not using poolinfo makes 
rollback in raid1_reshape more difficult.
I wonder whether you received it, or maybe I missed your reply.

I am looking forward to your discussion. I want to gain a deeper 
understanding and contribute more to md/raid.

Thanks.