[RFC RESEND 00/16] Split IOMMU DMA mapping operation to two steps

[Zhu Yanjun]

On 10.06.24 17:12, Zeng, Oak wrote:
> Hi Jason, Leon,
>
> I come back to this thread to ask a question. Per the discussion in another thread, I have integrated the new dma-mapping API (the first 6 patches of this series) to DRM subsystem. The new API seems fit pretty good to our purpose, better than scatter-gather dma-mapping. So we want to continue work with you to adopt this new API.
>
> Did you test the new API in RDMA subsystem? Or this RFC series was just some untested codes sending out to get people's design feedback? Do you have refined version for us to try? I ask because we are seeing some issues but not sure whether it is caused by the new API. We are debugging but it would be good to also ask at the same time.

Hi, Zeng

I have tested this patch series. And a patch about NVMe will cause some 
call trace. But if you revert this patch about NVMe, the whole patches 
can work well. You can develop your patches based on this patch series.

It seems that "some agreements can not be reached" about NVMe. So NVMe 
patch can not work well. I do not delve into this NVMe patch.

Zhu Yanjun

>
> Cc Himal/Krishna who are also working/testing the new API.
>
> Thanks,
> Oak
>
>> -----Original Message-----
>> From: Jason Gunthorpe <jgg at ziepe.ca>
>> Sent: Friday, May 3, 2024 12:43 PM
>> To: Zeng, Oak <oak.zeng at intel.com>
>> Cc: leon at kernel.org; Christoph Hellwig <hch at lst.de>; Robin Murphy
>> <robin.murphy at arm.com>; Marek Szyprowski
>> <m.szyprowski at samsung.com>; Joerg Roedel <joro at 8bytes.org>; Will
>> Deacon <will at kernel.org>; Chaitanya Kulkarni <chaitanyak at nvidia.com>;
>> Brost, Matthew <matthew.brost at intel.com>; Hellstrom, Thomas
>> <thomas.hellstrom at intel.com>; Jonathan Corbet <corbet at lwn.net>; Jens
>> Axboe <axboe at kernel.dk>; Keith Busch <kbusch at kernel.org>; Sagi
>> Grimberg <sagi at grimberg.me>; Yishai Hadas <yishaih at nvidia.com>;
>> Shameer Kolothum <shameerali.kolothum.thodi at huawei.com>; Tian, Kevin
>> <kevin.tian at intel.com>; Alex Williamson <alex.williamson at redhat.com>;
>> Jérôme Glisse <jglisse at redhat.com>; Andrew Morton <akpm at linux-
>> foundation.org>; linux-doc at vger.kernel.org; linux-kernel at vger.kernel.org;
>> linux-block at vger.kernel.org; linux-rdma at vger.kernel.org;
>> iommu at lists.linux.dev; linux-nvme at lists.infradead.org;
>> kvm at vger.kernel.org; linux-mm at kvack.org; Bart Van Assche
>> <bvanassche at acm.org>; Damien Le Moal
>> <damien.lemoal at opensource.wdc.com>; Amir Goldstein
>> <amir73il at gmail.com>; josef at toxicpanda.com; Martin K. Petersen
>> <martin.petersen at oracle.com>; daniel at iogearbox.net; Williams, Dan J
>> <dan.j.williams at intel.com>; jack at suse.com; Leon Romanovsky
>> <leonro at nvidia.com>; Zhu Yanjun <zyjzyj2000 at gmail.com>
>> Subject: Re: [RFC RESEND 00/16] Split IOMMU DMA mapping operation to
>> two steps
>>
>> On Thu, May 02, 2024 at 11:32:55PM +0000, Zeng, Oak wrote:
>>
>>>> Instead of teaching DMA to know these specific datatypes, let's separate
>>>> existing DMA mapping routine to two steps and give an option to
>> advanced
>>>> callers (subsystems) perform all calculations internally in advance and
>>>> map pages later when it is needed.
>>> I looked into how this scheme can be applied to DRM subsystem and GPU
>> drivers.
>>> I figured RDMA can apply this scheme because RDMA can calculate the
>>> iova size. Per my limited knowledge of rdma, user can register a
>>> memory region (the reg_user_mr vfunc) and memory region's sized is
>>> used to pre-allocate iova space. And in the RDMA use case, it seems
>>> the user registered region can be very big, e.g., 512MiB or even GiB
>> In RDMA the iova would be linked to the SVA granual we discussed
>> previously.
>>
>>> In GPU driver, we have a few use cases where we need dma-mapping. Just
>> name two:
>>> 1) userptr: it is user malloc'ed/mmap'ed memory and registers to gpu
>>> (in Intel's driver it is through a vm_bind api, similar to mmap). A
>>> userptr can be of any random size, depending on user malloc
>>> size. Today we use dma-map-sg for this use case. The down side of
>>> our approach is, during userptr invalidation, even if user only
>>> munmap partially of an userptr, we invalidate the whole userptr from
>>> gpu page table, because there is no way for us to partially
>>> dma-unmap the whole sg list. I think we can try your new API in this
>>> case. The main benefit of the new approach is the partial munmap
>>> case.
>> Yes, this is one of the main things it will improve.
>>
>>> We will have to pre-allocate iova for each userptr, and we have many
>>> userptrs of random size... So we might be not as efficient as RDMA
>>> case where I assume user register a few big memory regions.
>> You are already doing this. dma_map_sg() does exactly the same IOVA
>> allocation under the covers.
>>
>>> 2) system allocator: it is malloc'ed/mmap'ed memory be used for GPU
>>> program directly, without any other extra driver API call. We call
>>> this use case system allocator.
>>> For system allocator, driver have no knowledge of which virtual
>>> address range is valid in advance. So when GPU access a
>>> malloc'ed/mmap'ed address, we have a page fault. We then look up a
>>> CPU vma which contains the fault address. I guess we can use the CPU
>>> vma size to allocate the iova space of the same size?
>> No. You'd follow what we discussed in the other thread.
>>
>> If you do a full SVA then you'd split your MM space into granuals and
>> when a fault hits a granual you'd allocate the IOVA for the whole
>> granual. RDMA ODP is using a 512M granual currently.
>>
>> If you are doing sub ranges then you'd probably allocate the IOVA for
>> the well defined sub range (assuming the typical use case isn't huge)
>>
>>> But there will be a true difficulty to apply your scheme to this use
>>> case. It is related to the STICKY flag. As I understand it, the
>>> sticky flag is designed for driver to mark "this page/pfn has been
>>> populated, no need to re-populate again", roughly...Unlike userptr
>>> and RDMA use cases where the backing store of a buffer is always in
>>> system memory, in the system allocator use case, the backing store
>>> can be changing b/t system memory and GPU's device private
>>> memory. Even worse, we have to assume the data migration b/t system
>>> and GPU is dynamic. When data is migrated to GPU, we don't need
>>> dma-map. And when migration happens to a pfn with STICKY flag, we
>>> still need to repopulate this pfn. So you can see, it is not easy to
>>> apply this scheme to this use case. At least I can't see an obvious
>>> way.
>> You are already doing this today, you are keeping the sg list around
>> until you unmap it.
>>
>> Instead of keeping the sg list you'd keep a much smaller datastructure
>> per-granual. The sticky bit is simply a convient way for ODP to manage
>> the smaller data structure, you don't have to use it.
>>
>> But you do need to keep track of what pages in the granual have been
>> DMA mapped - sg list was doing this before. This could be a simple
>> bitmap array matching the granual size.
>>
>> Looking (far) forward we may be able to have a "replace" API that
>> allows installing a new page unconditionally regardless of what is
>> already there.
>>
>> Jason

-- 
Best Regards,
Yanjun.Zhu