Re: [PATCH 2/2] mm: set PG_dma_pinned on get_user_pages*()

[John Hubbard <jhubbard@nvidia.com>]

On 06/17/2018 12:53 PM, Dan Williams wrote:
> [..]
>> diff --git a/mm/rmap.c b/mm/rmap.c
>> index 6db729dc4c50..37576f0a4645 100644
>> --- a/mm/rmap.c
>> +++ b/mm/rmap.c
>> @@ -1360,6 +1360,8 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
>>                                 flags & TTU_SPLIT_FREEZE, page);
>>         }
>>
>> +       if (PageDmaPinned(page))
>> +               return false;
>>         /*
>>          * We have to assume the worse case ie pmd for invalidation. Note that
>>          * the page can not be free in this function as call of try_to_unmap()
> 
> We have a similiar problem with DAX and the conclusion we came to is
> that it is not acceptable for userspace to arbitrarily block kernel
> actions. The conclusion there was: 'wait' if the DMA is transient, and
> 'revoke' if the DMA is long lived, or otherwise 'block' long-lived DMA
> if a revocation mechanism is not available.
> 

Dan, thanks...can you please say a few words (or point to the code) about the "revoke" part of
this design? And whether you think it could be applied here (instead of the unconditional
appproach I have above)?