Re: [PATCH v3] staging: android: ion: Add implementation of dma_buf_vmap and dma_buf_vunmap

From: Alexey Skidanov
Date: Sun Dec 16 2018 - 01:35:27 EST




On 12/16/18 7:20 AM, Liam Mark wrote:
> On Tue, 6 Feb 2018, Alexey Skidanov wrote:
>
>>
>>
>> On 02/07/2018 01:56 AM, Laura Abbott wrote:
>>> On 01/31/2018 10:10 PM, Alexey Skidanov wrote:
>>>>
>>>> On 01/31/2018 03:00 PM, Greg KH wrote:
>>>>> On Wed, Jan 31, 2018 at 02:03:42PM +0200, Alexey Skidanov wrote:
>>>>>> Any driver may access shared buffers, created by ion, using
>>>>>> dma_buf_vmap and
>>>>>> dma_buf_vunmap dma-buf API that maps/unmaps previosuly allocated
>>>>>> buffers into
>>>>>> the kernel virtual address space. The implementation of these API is
>>>>>> missing in
>>>>>> the current ion implementation.
>>>>>>
>>>>>> Signed-off-by: Alexey Skidanov <alexey.skidanov@xxxxxxxxx>
>>>>>> ---
>>>>>
>>>>> No review from any other Intel developers? :(
>>>> Will add.
>>>>>
>>>>> Anyway, what in-tree driver needs access to these functions?
>>>> I'm not sure that there are the in-tree drivers using these functions
>>>> and ion as> buffer exporter because they are not implemented in ion :)
>>>> But there are some in-tre> drivers using these APIs (gpu drivers) with
>>>> other buffer exporters.
>>>
>>> It's still not clear why you need to implement these APIs.
>> How the importing kernel module may access the content of the buffer? :)
>> With the current ion implementation it's only possible by dma_buf_kmap,
>> mapping one page at a time. For pretty large buffers, it might have some
>> performance impact.
>> (Probably, the page by page mapping is the only way to access large
>> buffers on 32 bit systems, where the vmalloc range is very small. By the
>> way, the current ion dma_map_kmap doesn't really map only 1 page at a
>> time - it uses the result of vmap() that might fail on 32 bit systems.)
>>
>>> Are you planning to use Ion with GPU drivers? I'm especially
>>> interested in this if you have a non-Android use case.
>> Yes, my use case is the non-Android one. But not with GPU drivers.
>>>
>>> Thanks,
>>> Laura
>>
>> Thanks,
>> Alexey
>
> I was wondering if we could re-open the discussion on adding support to
> ION for dma_buf_vmap.
> It seems like the patch was not taken as the reviewers wanted more
> evidence of an upstream use case.
>
> Here would be my upstream usage argument for including dma_buf_vmap
> support in ION.
>
> Currently all calls to ion_dma_buf_begin_cpu_access result in the creation
> of a kernel mapping for the buffer, unfortunately the resulting call to
> alloc_vmap_area can be quite expensive and this has caused a performance
> regression for certain clients when they have moved to the new version of
> ION.
>
> The kernel mapping is not actually needed in ion_dma_buf_begin_cpu_access,
> and generally isn't needed by clients. So if we remove the creation of the
> kernel mapping in ion_dma_buf_begin_cpu_access and only create it when
> needed we can speed up the calls to ion_dma_buf_begin_cpu_access.
>
> An additional benefit of removing the creation of kernel mappings from
> ion_dma_buf_begin_cpu_access is that it makes the ION code more secure.
> Currently a malicious client could call the DMA_BUF_IOCTL_SYNC IOCTL with
> flags DMA_BUF_SYNC_END multiple times to cause the ION buffer kmap_cnt to
> go negative which could lead to undesired behavior.
>
> One disadvantage of the above change is that a kernel mapping is not
> already created when a client calls dma_buf_kmap. So the following
> dma_buf_kmap contract can't be satisfied.
>
> /**
> * dma_buf_kmap - Map a page of the buffer object into kernel address
> space. The
> * same restrictions as for kmap and friends apply.
> * @dmabuf: [in] buffer to map page from.
> * @page_num: [in] page in PAGE_SIZE units to map.
> *
> * This call must always succeed, any necessary preparations that might
> fail
> * need to be done in begin_cpu_access.
> */
>
> But hopefully we can work around this by moving clients to dma_buf_vmap.
I think the problem is with the contract. We can't ensure that the call
is always succeeds regardless the implementation - any mapping might
fail. Probably this is why *all* clients of dma_buf_kmap() check the
return value (so it's safe to return NULL in case of failure).

I would suggest to fix the contract and to keep the dma_buf_kmap()
support in ION.
>
> Based on discussions at LPC here is what was proposed:
> - #1 Add support to ION for dma_buf_vmap and dma_buf_vunmap
> - #2 Move any existing ION clients over from using dma_buf_kmap to
> dma_buf_vmap
> - #3 Deprecate support in ION for dma_buf_kmap?
> - #4 Make the above performance optimization to
> ion_dma_buf_begin_cpu_access to remove the creation of a kernel mapping.
>
> Thoughts?
>
> Liam
>
> Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
> a Linux Foundation Collaborative Project
>

Thanks,
Alexey