Re: [PATCH net-next v7 2/2] page_pool: Track DMA-mapped pages and unmap them when destroying the pool

["Toke Høiland-Jørgensen" <toke@redhat.com>]

Yunsheng Lin <linyunsheng@huawei.com> writes:

> On 2025/4/5 20:50, Toke Høiland-Jørgensen wrote:
>> Alexander Lobakin <aleksander.lobakin@intel.com> writes:
>> 
>>> From: Alexander Lobakin <aleksander.lobakin@intel.com>
>>> Date: Fri, 4 Apr 2025 17:55:43 +0200
>>>
>>>> From: Toke Høiland-Jørgensen <toke@redhat.com>
>>>> Date: Fri, 04 Apr 2025 12:18:36 +0200
>>>>
>>>>> When enabling DMA mapping in page_pool, pages are kept DMA mapped until
>>>>> they are released from the pool, to avoid the overhead of re-mapping the
>>>>> pages every time they are used. This causes resource leaks and/or
>>>>> crashes when there are pages still outstanding while the device is torn
>>>>> down, because page_pool will attempt an unmap through a non-existent DMA
>>>>> device on the subsequent page return.
>>>>
>>>> [...]
>>>>
>>>>> -#define PP_MAGIC_MASK ~0x3UL
>>>>> +#define PP_MAGIC_MASK ~(PP_DMA_INDEX_MASK | 0x3UL)
>>>>>  
>>>>>  /**
>>>>>   * struct page_pool_params - page pool parameters
>>>>> @@ -173,10 +212,10 @@ struct page_pool {
>>>>>  	int cpuid;
>>>>>  	u32 pages_state_hold_cnt;
>>>>>  
>>>>> -	bool has_init_callback:1;	/* slow::init_callback is set */
>>>>> +	bool dma_sync;			/* Perform DMA sync for device */
>>>>
>>>> Yunsheng said this change to a full bool is redundant in the v6 thread
>>>> ¯\_(ツ)_/¯
>> 
>> AFAIU, the comment was that the second READ_ONCE() when reading the
>> field was redundant, because of the rcu_read_lock(). Which may be the
>> case, but I think keeping it makes the intent of the code clearer. And
>> in any case, it has nothing to do with changing the type of the field...
>
> For changing the type of the field part, there are only two outcomes here
> when using bit field here:
> 1. The reading returns a correct value.
> 2. The reading returns a incorrect value.
>
> So the question seems to be what would possibly go wrong when the reading
> return an incorrect value when there is an additional reading under the rcu
> read lock and there is a rcu sync after clearing pool->dma_sync? Considering
> we only need to ensure there is no dma sync API called after rcu sync.

Okay, so your argument is basically that the barrier in rcu_read_lock()
should prevent the compiler from coalescing the two reads of the
pp->dma_sync field in page_pool_dma_sync_for_device()? And that
READ/WRITE_ONCE() are not needed for the same reason?

> And it seems data_race() can be used to mark the above reading so that KCSAN
> will not complain.

Where would you suggest to add those? Not sure such annotations would
improve readability relative to the current use of READ/WRITE_ONCE()?
The latter is more clear in communicating intent, I would say...

> IOW, changing the type of the field part isn't that necessary as my
> understanding.

Since changing the field doesn't change the size of the structure, I
would be inclined to keep the change for readability reasons, cf the
above.

-Toke