Re: [PATCH 1/2] iommu/dma: Respect __GFP_DMA and __GFP_DMA32 in incoming GFP flags

[Tomasz Figa]

On Tue, Jun 27, 2017 at 9:26 PM, Robin Murphy <robin.murphy@xxxxxxx> wrote:
> On 27/06/17 12:17, Tomasz Figa wrote:
>> On Tue, Jun 27, 2017 at 8:01 PM, Robin Murphy <robin.murphy@xxxxxxx> wrote:
>>> On 27/06/17 08:28, Tomasz Figa wrote:
>>>> Current implementation of __iommu_dma_alloc_pages() keeps adding
>>>> __GFP_HIGHMEM to GFP flags regardless of whether other zone flags are
>>>> already included in the incoming flags. If __GFP_DMA or __GFP_DMA32 is
>>>> set at the same time as __GFP_HIGHMEM, the allocation fails due to
>>>> invalid zone flag combination.
>>>>
>>>> Fix this by checking for __GFP_DMA and __GFP_DMA32 in incoming GFP flags
>>>> and adding __GFP_HIGHMEM only if they are not present.
>>>
>>> Wouldn't it make more sense to strip off the ZONE_DMA* related flags,
>>> since the whole point here is that we don't care where the pages come from?
>>
>> I guess for my use case it wouldn't break things, as I strip them in
>> my DMA mapping implementation right now (+/- one minor detail below).
>>
>> However I recall existing IOMMUs that could only use pages from within
>> the 32-bit address space (e.g. Tegra X1).
>
> In general, iommu-dma can't really support IOMMUs which can't reach the
> entirety of kernel memory - there's no easy way to determine what such a
> limit is if it exists, nor necessarily enforce it, and either way the
> streaming API callbacks are pretty much dead in the water.

Right. Especially with various user pointer sources it doesn't sound
very realistic to enforce that all memory comes from __GFP_DMA(32)
zone.

On the other hand, support for user pointer is optional in subsystems
such as V4L2 and drivers for such disabled hardware could opt out.
Then at least dma_alloc can be made working, giving some level of
usability, IMHO higher than no IOMMU and contiguous memory allocated
from CMA.

>
>> Also the IOMMU I'm working
>> on is a part of a PCI device and it might actually prefer 32-bit
>> addressable memory as well (to avoid DAC addressing; I still need to
>> evaluate this). With this said, maybe it could actually make sense to
>> leave the choice to the DMA mapping implementation?
>
> I think you're right - we're just not in a position to make any decision
> at this level, so we probably do have to do this for robustness. I would
> like to fix the longstanding dodgy comment, though, to clarify that
> "IOMMU can map any pages" is only an assumption, and particularly one
> which is invalidated by the presence of GFP_DMA flags.

Just to make sure, should I resent with the commit updated? If so,
what would be your preference on the wording?

Best regards,
Tomasz

>
> Robin.
>
>>
>> Best regards,
>> Tomasz
>>
>>>
>>> Robin.
>>>
>>>> Signed-off-by: Tomasz Figa <tfiga@xxxxxxxxxxxx>
>>>> ---
>>>>  drivers/iommu/dma-iommu.c | 10 ++++++++--
>>>>  1 file changed, 8 insertions(+), 2 deletions(-)
>>>>
>>>> diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
>>>> index 9d1cebe7f6cb..29965a092a69 100644
>>>> --- a/drivers/iommu/dma-iommu.c
>>>> +++ b/drivers/iommu/dma-iommu.c
>>>> @@ -445,8 +445,14 @@ static struct page **__iommu_dma_alloc_pages(unsigned int count,
>>>>       if (!pages)
>>>>               return NULL;
>>>>
>>>> -     /* IOMMU can map any pages, so himem can also be used here */
>>>> -     gfp |= __GFP_NOWARN | __GFP_HIGHMEM;
>>>> +     /*
>>>> +      * IOMMU can map any pages, so himem can also be used here,
>>>> +      * unless another DMA zone is explicitly requested.
>>>> +      */
>>>> +     if (!(gfp & (__GFP_DMA | __GFP_DMA32)))
>>>> +             gfp |= __GFP_HIGHMEM;
>>>> +
>>>> +     gfp |= __GFP_NOWARN;
>>>>
>>>>       while (count) {
>>>>               struct page *page = NULL;
>>>>
>>>
>