Re: [PATCH net-next v11 0/6] introduce page_pool_alloc() related API

[Jakub Kicinski]

On Wed, 18 Oct 2023 19:47:16 +0800 Yunsheng Lin wrote:
> > mention it in the documentation. Plus the kdoc of the function should
> > say that this is just a thin wrapper around other page pool APIs, and
> > it's safe to mix it with other page pool APIs?  
> 
> I am not sure I understand what do 'safe' and 'mix' mean here.
> 
> For 'safe' part, I suppose you mean if there is a va accociated with
> a 'struct page' without calling some API like kmap()? For that, I suppose
> it is safe when the driver is calling page_pool API without the
> __GFP_HIGHMEM flag. Maybe we should mention that in the kdoc and give a
> warning if page_pool_*alloc_va() is called with the __GFP_HIGHMEM flag?

Sounds good. Warning wrapped in #if CONFIG_DEBUG_NET perhaps?

> For the 'mix', I suppose you mean the below:
> 1. Allocate a page with the page_pool_*alloc_va() API and free a page with
>    page_pool_free() API.
> 2. Allocate a page with the page_pool_*alloc() API and free a page with
>    page_pool_free_va() API.
> 
> For 1, it seems it is ok as some virt_to_head_page() and page_address() call
> between va and 'struct page' does not seem to change anything if we have
> enforce page_pool_*alloc_va() to be called without the __GFP_HIGHMEM flag.
> 
> For 2, If the va is returned from page_address() which the allocation API is
> called without __GFP_HIGHMEM flag. If not, the va is from kmap*()? which means
> we may be calling page_pool_free_va() before kunmap*()? Is that possible?

Right, if someone passes kmap()'ed address they are trying quite hard
to break their own driver. Technically possible but I wouldn't worry.

I just mean that in the common case of non-HIGHMEM page, calling
page_pool_free_va() with the address returned by page_address() 
is perfectly legal.