Re: [dpdk-dev] [RFC] ethdev: introduce DMA memory mapping for external memory

["Burakov, Anatoly" <anatoly.burakov@intel.com>]

On 15-Nov-18 9:46 AM, Shahaf Shuler wrote:
> Wednesday, November 14, 2018 7:06 PM, Burakov, Anatoly:
>> Subject: Re: [RFC] ethdev: introduce DMA memory mapping for external
>> memory
>>
>> On 14-Nov-18 2:53 PM, Shahaf Shuler wrote:
>>> Hi Anatoly,
>>>
>>> Wednesday, November 14, 2018 1:19 PM, Burakov, Anatoly:
>>>> Subject: Re: [RFC] ethdev: introduce DMA memory mapping for external
>>>> memory
>>>>
>>>> Hi Shahaf,
>>>>
>>>> Great to see such effort! Few comments below.
>>>>
>>>> Note: halfway through writing my comments i realized that i am
>>>> starting with an assumption that this API is a replacement for
>>>> current VFIO DMA mapping API's. So, if my comments seem out of left
>>>> field, this is probably why :)
>>>>
>>>> On 04-Nov-18 12:41 PM, Shahaf Shuler wrote:
>>>>> Request for comment on the high level changes present on this patch.
>>>>>
>>>>> The need to use external memory (memory belong to application and
>>>>> not part of the DPDK hugepages) is allready present.
>>>>> Starting from storage apps which prefer to manage their own memory
>>>>> blocks for efficient use of the storage device. Continue with GPU
>>>>> based application which strives to achieve zero copy while
>>>>> processing the packet payload on the GPU core. And finally by
>>>>> vSwitch/vRouter application who just prefer to have a full control
>>>>> over the memory in use
>>>> (e.g. VPP).
>>>>>
>>>>> Recent work[1] in the DPDK enabled the use of external memory,
>>>>> however it mostly focus on VFIO as the only way to map memory.
>>>>> While VFIO is common, there are other vendors which use different
>>>>> ways to map memory (e.g. Mellanox and NXP[2]).
>>>>>
>>>>> The work in this patch moves the DMA mapping to vendor agnostic APIs
>>>>> located under ethdev. The choice in ethdev was because memory map
>>>>> should be associated with a specific port(s). Otherwise the memory
>>>>> is being mapped multiple times to different frameworks and ends up
>>>>> with memory being wasted on redundant translation table in the host
>>>>> or in the
>>>> device.
>>>>
>>>> So, anything other than ethdev (e.g. cryptodev) will not be able to
>>>> map memory for DMA?
>>>
>>> That's is a fair point.
>>>
>>>>
>>>> I have thought about this for some length of time, and i think DMA
>>>> mapping belongs in EAL (more specifically, somewhere at the bus
>>>> layer), rather than at device level.
>>>
>>> I am not sure I agree here. For example take Intel and Mellanox devices.
>> Both are PCI devices, so how will you distinguish which mapping API to use?
>>> Also I still think the mapping should be in device granularity and not
>> bus/system granularity, since it is very typical for a memory to be used for
>> DMA be a specific device.
>>>
>>> Maybe we can say the DMA mapping is a rte_device attribute. It is the
>> parent class for all the DPDK devices.
>>> We need to see w/ vport representors (which all has the same rte_device).
>> On that case I believe the rte_device.map call can register the memory to all
>> of the representors as well (if needed).
>>>
>>> Placing this functionality at device level comes with more work
>>>> to support different device types and puts a burden on device driver
>>>> developers to implement their own mapping functions.
>>>
>>> The mapping function can be shared. For example we can still maintain the
>> vfio mapping scheme as part of eal and have all the related driver to call this
>> function.
>>> The only overhead will be to maintain the function pointer for the dma call.
>>> With this work, instead of the eal layer to guess which type of DMA
>> mapping the devices in the  system needs or alternatively force them all to
>> work w/ VFIO, each driver will select its own function.
>>> The driver is the only one which knows what type of DMA mapping its
>> device needs.
>>>
>>>>
>>>> However, i have no familiarity with how MLX/NXP devices do their DMA
>>>> mapping, so maybe the device-centric approach would be better. We
>>>> could provide "standard" mapping functions at the bus level (such as
>>>> VFIO mapping functions for PCI bus), so that this could would not
>>>> have to be reimplemented in the devices.
>>>
>>> Yes, like I said above, I wasn't intending to re-implement all the mapping
>> function again on each driver. Yet, I believe it should be per device.
>>>
>>>>
>>>> Moreover, i'm not sure how this is going to work for VFIO. If this is
>>>> to be called for each NIC that needs access to the memory, then we'll
>>>> end up with double mappings for any NIC that uses VFIO, unless you
>>>> want each NIC to be in a separate container.
>>>
>>> I am not much familiar w/ VFIO (you are the expert😊).
>>> What will happen if we map the same memory twice (under same
>> container)? The translation on the IOMMU will be doubled? The map will
>> return with error that this memory mapping already exists?
>>
>> The latter. You can't map the same memory twice in the same container.
>> You can't even keep NICs in separate containers because then secondary
>> processes won't work without serious rework.
>>
>> So, all VFIO-mapped things will need to share the mappings.
>>
>> It's not an insurmountable problem, but if we're going to share mapping
>> status for VFIO (i.e. track which area is already mapped), then what's
>> stopping us from doing the same for other DMA mapping mechanisms? I.e.
>> instead of duplicating the mappings in each driver, provide some kind of
>> mechanism for devices to share the DMA mapping caches. Apologies if i'm
>> talking nonsense - i'm completely unfamiliar with how DMA mapping works
>> for MLX/NXP devices :)
> 
> Unfortunately it cannot be done at least w/ Mellanox.
> In Mellanox the kernel driver is the one which maps the memory. The mapping returns a key which identify a memory region which was just registered to the device.
> There is a complete separation between the ports, meaning one port mapping cannot be used by in the other port, even if the key is known.
> 
> The separation is not only in ports, but also in processes (two primary ones, for secondary we have a way to share). If two process work on the same device, the must register the memory independently.

Ah, OK.

So, we're right back to where we started. Right now, external memory 
expects to behave the same way as all other memory - you don't need to 
perform DMA mapping for it.

That said, part of the reason *why* it was done that way was because 
there is no way to trigger VFIO DMA mapping for NXP (or was it MLX?) 
devices. If you look at initial versions of the patchset, the DMA 
mapping was actually done manually. Then, i became convinced that doing 
this automatically is the way to go, both because it erases the 
usability differences as far as memory types are concerned, and because 
it enables whatever services that are subscribing to memory events to 
receive notifications about external memory as well (i.e. consistency).

Given that it's still an experimental API, we can tinker with it all we 
like, so it's not set in stone. However, i would really like to keep the 
current automagic thing, because DMA mapping may not be the only user of 
memory callbacks - they can be used for debug purposes, or for any other 
things.

>>>>> For example, consider a host with Mellanox and Intel devices.
>>>>> Mapping a memory without specifying to which port will end up with
>>>>> IOMMU registration and Verbs (Mellanox DMA map) registration.
>>>>> Another example can be two Mellanox devices on the same host. The
>>>> memory
>>>>> will be mapped for both, even though application will use mempool
>>>>> per device.
>>>>>
>>>>> To use the suggested APIs the application will allocate a memory
>>>>> block and will call rte_eth_dma_map. It will map it to every port
>>>>> that needs DMA access to this memory.
>>>>
>>>> This bit is unclear to me. What do you mean "map it to every port
>>>> that needs DMA access to this memory"? I don't see how this API
>>>> solves the above problem of mapping the same memory to all devices.
>>>> How does a device know which memory it will need? Does the user
>>>> specifically have to call this API for each and every NIC they're using?
>>>
>>> Yes, the user will call this API for every port which needs to have DMA
>> access to this memory.
>>> Remember we are speaking here on external memory the application
>> allocated and wants to use for send/receive.  The device doesn't guess which
>> memory he will need, the user is telling it to him explicitly.
>>>
>>>>
>>>> For DPDK-managed memory, everything will still get mapped to every
>>>> device automatically, correct?
>>>
>>> Yes, even though it is not the case today.
>>
>> What do you mean it is not the case? It is the case today. When external
>> memory chunk is registered at the heap, a mem event callback is triggered
>> just like for regular memory, and this chunk does get mapped to VFIO as well
>> as any other subscribed entity. As i recall, NXP NICs currently are set up to
>> ignore externally allocated memory, but for the general VFIO case,
>> everything is mapped automatically.
>>
>>>
>>> If so, then such a manual approach for
>>>> external memory will be bad for both usability and drop-in
>>>> replacement of internal-to-external memory, because it introduces
>>>> inconsistency between using internal and external memory. From my
>>>> point of view, either we do *everything* manually (i.e. register all
>>>> memory for DMA
>>>> explicitly) and thereby avoid this problem but keep the consistency,
>>>> or we do *everything* automatically and deal with duplication of
>>>> mappings somehow (say, by MLX/NXP drivers sharing their mappings
>>>> through bus interface).
>>>
>>> I understand your point, however I am not sure external and internal
>> memory *must* be consist.
>>> The DPDK-managed memory is part of the DPDK subsystems and the DPDK
>> libs are preparing it for the optimal use of the underlying devices. The
>> external memory is different, it is a proprietary memory the application
>> allocated and the DPDK cannot do anything in advance on it.
>>
>> My view for designing external memory support was that it should behave
>> like regular DPDK memory for all intents and purposes, and be a drop-in
>> replacement, should you choose to use it. I.e. the application should not care
>> whether it uses internal or external memory - it all sits in the same malloc
>> heaps, it all uses the same socket ID mechanisms, etc. - for all intents and
>> purposes, they're one and the same.
>>
>>> Even today there is inconsistency, because if user wants to use external
>> memory it must map it (rte_vfio_dma_map) while he doesn't need to do
>> that for the DPDK-managed memory.
>>
>> Well, now i see where your confusion stems from :) You didn't know about
>> this:
>>
>> https://emea01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fgit.d
>> pdk.org%2Fdpdk%2Ftree%2Flib%2Flibrte_eal%2Fcommon%2Fmalloc_heap.c
>> %23n1169&data=02%7C01%7Cshahafs%40mellanox.com%7Ca1eb06c5a
>> d794e4e725d08d64a537f23%7Ca652971c7d2e4d9ba6a4d149256f461b%7C0%7
>> C0%7C636778119822376211&sdata=duMis6fW2CTmyRqdOAmlZl5cezCfJ
>> aeuoo61QGIBIGk%3D&reserved=0
>>
>> This will trigger all mem event callbacks, including VFIO DMA mapping
>> callback.
> 
> I see, so I am indeed confused 😊.
> On which cases the application should call the existing rte_vfio_dma_map? If the memory is already mapped and the only way to work with it is through the rte_malloc mechanism.

I have to be honest, I didn't consider this question before :D I guess 
there could be cases where using rte_malloc might not be suitable 
because it wastes some memory on malloc elements, i.e. if you want to 
use N pages as memory, you'd have to allocate N+1 pages. If memory is at 
a premium, maybe manual management of it would be better in some cases.

> 
>>
>>>
>>> I guess we can we can add a flag on the device mapping which will say
>> MAP_TO_ALL_DEVICES, to ease the application life in the presence of
>> multiple device in the host.
>>>
>>
>>
>> --
>> Thanks,
>> Anatoly


-- 
Thanks,
Anatoly