Re: [RFC PATCH v2 00/11] Add support to dma_map_sg for P2PDMA

[Robin Murphy <robin.murphy@arm.com>]

On 2021-03-12 16:18, Logan Gunthorpe wrote:
> 
> 
> On 2021-03-12 8:51 a.m., Robin Murphy wrote:
>> On 2021-03-11 23:31, Logan Gunthorpe wrote:
>>> Hi,
>>>
>>> This is a rework of the first half of my RFC for doing P2PDMA in
>>> userspace
>>> with O_DIRECT[1].
>>>
>>> The largest issue with that series was the gross way of flagging P2PDMA
>>> SGL segments. This RFC proposes a different approach, (suggested by
>>> Dan Williams[2]) which uses the third bit in the page_link field of the
>>> SGL.
>>>
>>> This approach is a lot less hacky but comes at the cost of adding a
>>> CONFIG_64BIT dependency to CONFIG_PCI_P2PDMA and using up the last
>>> scarce bit in the page_link. For our purposes, a 64BIT restriction is
>>> acceptable but it's not clear if this is ok for all usecases hoping
>>> to make use of P2PDMA.
>>>
>>> Matthew Wilcox has already suggested (off-list) that this is the wrong
>>> approach, preferring a new dma mapping operation and an SGL
>>> replacement. I
>>> don't disagree that something along those lines would be a better long
>>> term solution, but it involves overcoming a lot of challenges to get
>>> there. Creating a new mapping operation still means adding support to
>>> more
>>> than 25 dma_map_ops implementations (many of which are on obscure
>>> architectures) or creating a redundant path to fallback with dma_map_sg()
>>> for every driver that uses the new operation. This RFC is an approach
>>> that doesn't require overcoming these blocks.
>>
>> I don't really follow that argument - you're only adding support to two
>> implementations with the awkward flag, so why would using a dedicated
>> operation instead be any different? Whatever callers need to do if
>> dma_pci_p2pdma_supported() says no, they could equally do if
>> dma_map_p2p_sg() (or whatever) returns -ENXIO, no?
> 
> The thing is if the dma_map_sg doesn't support P2PDMA then P2PDMA
> transactions cannot be done, but regular transactions can still go
> through as they always did.
> 
> But replacing dma_map_sg() with dma_map_new() affects all operations,
> P2PDMA or otherwise. If dma_map_new() isn't supported it can't simply
> not support P2PDMA; it has to maintain a fallback path to dma_map_sg().

But AFAICS the equivalent fallback path still has to exist either way. 
My impression so far is that callers would end up looking something like 
this:

	if (dma_pci_p2pdma_supported()) {
		if (dma_map_sg(...) < 0)
			//do non-p2p fallback due to p2p failure
	} else {
		//do non-p2p fallback due to lack of support
	}

at which point, simply:

	if (dma_map_sg_p2p(...) < 0)
		//do non-p2p fallback either way

seems entirely reasonable. What am I missing?

Let's not pretend that overloading an existing API means we can start 
feeding P2P pages into any old subsystem/driver without further changes 
- there already *are* at least some that retry ad infinitum if DMA 
mapping fails (the USB layer springs to mind...) and thus wouldn't 
handle the PCI_P2PDMA_MAP_NOT_SUPPORTED case acceptably.

> Given that the inputs and outputs for dma_map_new() will be completely
> different data structures this will be quite a lot of similar paths
> required in the driver. (ie mapping a bvec to the input struct and the
> output struct to hardware requirements) If a bug crops up in the old
> dma_map_sg(), developers might not notice it for some time seeing it
> won't be used on the most popular architectures.

Huh? I'm specifically suggesting a new interface that takes the *same* 
data structure (at least to begin with), but just gives us more 
flexibility in terms of introducing p2p-aware behaviour somewhat more 
safely. Yes, we already know that we ultimately want something better 
than scatterlists for representing things like this and dma-buf imports, 
but that hardly has to happen overnight.

Robin.