[Intel-gfx] [Linaro-mm-sig] Re: [PATCH 02/19] dma-buf-map: Add helper to initialize second map

Thu Jan 27 11:44:21 UTC 2022

Am 27.01.22 um 12:16 schrieb Daniel Vetter:
> On Thu, Jan 27, 2022 at 11:21:20AM +0100, Christian König wrote:
>> Am 27.01.22 um 11:00 schrieb Daniel Vetter:
>>> On Thu, Jan 27, 2022 at 01:33:32AM -0800, Lucas De Marchi wrote:
>>>> On Thu, Jan 27, 2022 at 09:57:25AM +0100, Daniel Vetter wrote:
>>>>> On Thu, Jan 27, 2022 at 09:02:54AM +0100, Christian König wrote:
>>>>>> Am 27.01.22 um 08:57 schrieb Lucas De Marchi:
>>>>>>> On Thu, Jan 27, 2022 at 08:27:11AM +0100, Christian König wrote:
>>>>>>>> Am 26.01.22 um 21:36 schrieb Lucas De Marchi:
>>>>>>>>> When dma_buf_map struct is passed around, it's useful to be able to
>>>>>>>>> initialize a second map that takes care of reading/writing to an offset
>>>>>>>>> of the original map.
>>>>>>>>>
>>>>>>>>> Add a helper that copies the struct and add the offset to the proper
>>>>>>>>> address.
>>>>>>>> Well what you propose here can lead to all kind of problems and is
>>>>>>>> rather bad design as far as I can see.
>>>>>>>>
>>>>>>>> The struct dma_buf_map is only to be filled in by the exporter and
>>>>>>>> should not be modified in this way by the importer.
>>>>>>> humn... not sure if I was  clear. There is no importer and exporter here.
>>>>>> Yeah, and exactly that's what I'm pointing out as problem here.
>>>>>>
>>>>>> You are using the inter driver framework for something internal to the
>>>>>> driver. That is an absolutely clear NAK!
>>>>>>
>>>>>> We could discuss that, but you guys are just sending around patches to do
>>>>>> this without any consensus that this is a good idea.
>>>>> Uh I suggested this, also we're already using dma_buf_map all over the
>>>>> place as a convenient abstraction. So imo that's all fine, it should allow
>>>>> drivers to simplify some code where on igpu it's in normal kernel memory
>>>>> and on dgpu it's behind some pci bar.
>>>>>
>>>>> Maybe we should have a better name for that struct (and maybe also a
>>>>> better place), but way back when we discussed that bikeshed I didn't come
>>>>> up with anything better really.
>>>> I suggest iosys_map since it abstracts access to IO and system memory.
>>>>
>>>>>>> There is a role delegation on filling out and reading a buffer when
>>>>>>> that buffer represents a struct layout.
>>>>>>>
>>>>>>> struct bla {
>>>>>>>       int a;
>>>>>>>       int b;
>>>>>>>       int c;
>>>>>>>       struct foo foo;
>>>>>>>       struct bar bar;
>>>>>>>       int d;
>>>>>>> }
>>>>>>>
>>>>>>>
>>>>>>> This implementation allows you to have:
>>>>>>>
>>>>>>>       fill_foo(struct dma_buf_map *bla_map) { ... }
>>>>>>>       fill_bar(struct dma_buf_map *bla_map) { ... }
>>>>>>>
>>>>>>> and the first thing these do is to make sure the map it's pointing to
>>>>>>> is relative to the struct it's supposed to write/read. Otherwise you're
>>>>>>> suggesting everything to be relative to struct bla, or to do the same
>>>>>>> I'm doing it, but IMO more prone to error:
>>>>>>>
>>>>>>>       struct dma_buf_map map = *bla_map;
>>>>>>>       dma_buf_map_incr(map, offsetof(...));
>>>>> Wrt the issue at hand I think the above is perfectly fine code. The idea
>>>>> with dma_buf_map is really that it's just a special pointer, so writing
>>>>> the code exactly as pointer code feels best. Unfortunately you cannot make
>>>>> them typesafe (because of C), so the code sometimes looks a bit ugly.
>>>>> Otherwise we could do stuff like container_of and all that with
>>>>> typechecking in the macros.
>>>> I had exactly this code above, but after writting quite a few patches
>>>> using it, particularly with functions that have to write to 2 maps (see
>>>> patch 6 for example), it felt much better to have something to
>>>> initialize correctly from the start
>>>>
>>>> 	struct dma_buf_map other_map = *bla_map;
>>>> 	/* poor Lucas forgetting dma_buf_map_incr(map, offsetof(...)); */
>>>>
>>>> is error prone and hard to debug since you will be reading/writting
>>>> from/to another location rather than exploding
>>>>
>>>> While with the construct below
>>>>
>>>> 	other_map;
>>>> 	...
>>>> 	other_map = INITIALIZER()
>>>>
>>>> I can rely on the compiler complaining about uninitialized var. And
>>>> in most of the cases I can just have this single line in the beggining of the
>>>> function when the offset is constant:
>>>>
>>>> 	struct dma_buf_map other_map = INITIALIZER(bla_map, offsetof(..));
>>> Hm yeah that's a good point that this allows us to rely on the compiler to
>>> check for uninitialized variables.
>>>
>>> Maybe include the above (with editing, but keeping the examples) in the
>>> kerneldoc to explain why/how to use this? With that the concept at least
>>> has my
>>>
>>> Acked-by: Daniel Vetter <daniel.vetter at ffwll.ch>
>>>
>>> I'll leave it up to you & Christian to find a prettier color choice for
>>> the naming bikeshed.
>> There is one major issue remaining with this and that is dma_buf_vunmap():
>>
>> void dma_buf_vunmap(struct dma_buf *dmabuf, struct dma_buf_map *map);
>>
>> Here we expect the original pointer as returned by dma_buf_map(), otherwise
>> we vunmap() the wrong area!
>>
>> For all TTM based driver this doesn't matter since we keep the vmap base
>> separately in the BO anyway (IIRC), but we had at least one case where this
>> made boom last year.
> Yeah but isn't that the same if it's just a void *?
>
> If you pass the wrong pointer to an unmap function and not exactly what
> you go from the map function, then things go boom. This is like
> complaining that the following code wont work
>
> 	u32 *stuff
>
> 	stuff = kmap_local(some_page);
> 	*stuff++ = 0;
> 	*stuff = 1;
> 	kunmap_locak(stuff);
>
> It's just ... don't do that :-) Also since we pass dma_buf_map by value
> and not by pointer anywhere, the risk of this happening is pretty low
> since you tend to work on a copy. Same with void * pointers really.
>
> Now if people start to pass around struct dma_buf_map * as pointers for
> anything else than out parameters, then we're screwed. But that's like
> passing around void ** for lolz, which is just wrong (except when it's an
> out parameter or actually an array of pointers ofc).
>
> Or I really don't get your concern and you mean something else?

No that's pretty much it. It's just that we hide the pointer inside a 
structure and it is absolutely not obvious to a driver dev that you 
can't do:

dma_buf_vmap(.., &map);
dma_buf_map_inr(&map, x);
dma_buf_vunmap(.., &map);

As bare minimum I strongly suggest that we add some WARN_ONs to the 
framework to check that the pointer given to dma_buf_vunmap() is at 
least page aligned.

Christian.

> -Daniel
>
>
>> Christian.
>>
>>> -Daniel
>>>
>>>> Lucas De Marchi
>>>>
>>>>> -Daniel
>>>>>
>>>>>>> IMO this construct is worse because at a point in time in the function
>>>>>>> the map was pointing to the wrong thing the function was supposed to
>>>>>>> read/write.
>>>>>>>
>>>>>>> It's also useful when the function has double duty, updating a global
>>>>>>> part of the struct and a table inside it (see example in patch 6)
>>>>>>>
>>>>>>> thanks
>>>>>>> Lucas De Marchi
>>>>> -- 
>>>>> Daniel Vetter
>>>>> Software Engineer, Intel Corporation
>>>>> https://nam11.safelinks.protection.outlook.com/?url=http%3A%2F%2Fblog.ffwll.ch%2F&data=04%7C01%7Cchristian.koenig%40amd.com%7C0654a16ea3444271d7c308d9e17bd35d%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637788744226808874%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=Q6soluBglaZLhLszdapaWuUVsqMq5qvJOKiJjO%2B9BTg%3D&reserved=0