Re: [RFC v8 3/5] page_pool: fix IOMMU crash when driver has already unbound

[Yunsheng Lin <linyunsheng@huawei.com>]

On 2025/2/7 0:54, Mina Almasry wrote:
> On Tue, Feb 4, 2025 at 6:23 AM Yunsheng Lin <yunshenglin0825@gmail.com> wrote:
>>
>> On 1/28/2025 2:12 PM, Christoph Hellwig wrote:
>>> On Mon, Jan 27, 2025 at 10:57:32AM +0800, Yunsheng Lin wrote:
>>>> Note, the devmem patchset seems to make the bug harder to fix,
>>>> and may make backporting harder too. As there is no actual user
>>>> for the devmem and the fixing for devmem is unclear for now,
>>>> this patch does not consider fixing the case for devmem yet.
>>>
>>> Is there another outstanding patchet?  Or do you mean the existing
>>> devmem code already merged?  If that isn't actually used it should
>>> be removed, but otherwise you need to fix it.
>>
>> The last time I checked, only the code for networking stack supporting
>> the devmem had been merged.
>>
>> The first driver suppporting seems to be bnxt, which seems to be under
>> review:
>> https://lore.kernel.org/all/20241022162359.2713094-1-ap420073@gmail.com/
>>
>> As my understanding, this should work for the devmem too if the devmem
> 
>>From a quick look at this patch, it looks like you're handling
> netmem/net_iovs in the implementation, so this implementation is
> indeed considering netmem. I think the paragraph in the commit message
> that Christoph is responding to should be deleted, because in recent
> iterations you're handling netmem.
> 
>> provide a ops to do the per-netmem dma unmapping
>> It would be good that devmem people can have a look at it and see if
>> this fix works for the specific page_pool mp provider.
>>
> 
> We set pool->dma_map==false for memory providers that do not need
> mapping/unmapping, which you are checking in
> __page_pool_release_page_dma.

In page_pool_item_unmap(), it will return early when pool->mp_priv is
set no matter pool->dma_map is set or not, that was why I had the above
paragraph in the commit message.

The question seems to be about whether net_devmem_unbind_dmabuf() might
be called after driver has already unbound as it seems to be doing the
DMA unmapping operation based on the device of netdev->dev.parent, and
page_pool doesn't seems to have direct calling of net_devmem_unbind_dmabuf()
related API.

I am not able to find who and where is seting the netdev->dev.parent, but
I guess the DMA API is also not allowed to be called on the device of
netdev->dev.parent when the networking driver has unbound?

>