[LSF/MM/BPF TOPIC] dmabuf backed read/write

[LSF/MM/BPF TOPIC] dmabuf backed read/write

[Pavel Begunkov]

Good day everyone,

dma-buf is a powerful abstraction for managing buffers and DMA mappings,
and there is growing interest in extending it to the read/write path to
enable device-to-device transfers without bouncing data through system
memory. I was encouraged to submit it to LSF/MM/BPF as that might be
useful to mull over details and what capabilities and features people
may need.

The proposal consists of two parts. The first is a small in-kernel
framework that allows a dma-buf to be registered against a given file
and returns an object representing a DMA mapping. The actual mapping
creation is delegated to the target subsystem (e.g. NVMe). This
abstraction centralises request accounting, mapping management, dynamic
recreation, etc. The resulting mapping object is passed through the I/O
stack via a new iov_iter type.

As for the user API, a dma-buf is installed as an io_uring registered
buffer for a specific file. Once registered, the buffer can be used by
read / write io_uring requests as normal. io_uring will enforce that the
buffer is only used with "compatible files", which is for now restricted
to the target registration file, but will be expanded in the future.
Notably, io_uring is a consumer of the framework rather than a
dependency, and the infrastructure can be reused.

It took a couple of iterations on the list to get it to the current
design, v2 of the series can be looked up at [1], which implements the
infrastructure and initial wiring for NVMe. It slightly diverges from
the description above, as some of the framework bits are block specific,
and I'll be working on refining that and simplifying some of the
interfaces for v3. A good chunk of block handling is based on prior work
from Keith that was pre DMA mapping buffers [2].

Tushar was helping and mention he got good numbers for P2P transfers
compared to bouncing it via RAM. Anuj, Kanchan and Nitesh also
previously reported encouraging results for system memory backed
dma-buf for optimising IOMMU overhead, quoting Anuj:

- STRICT: before = 570 KIOPS, after = 5.01 MIOPS
- LAZY: before = 1.93 MIOPS, after = 5.01 MIOPS
- PASSTHROUGH: before = 5.01 MIOPS, after = 5.01 MIOPS

[1] https://lore.kernel.org/io-uring/cover.1763725387.git.asml.silence@gmail.com/
[2] https://lore.kernel.org/io-uring/20220805162444.3985535-1-kbusch@fb.com/
-- 
Pavel Begunkov

Re: [LSF/MM/BPF TOPIC] dmabuf backed read/write

[Keith Busch]

On Tue, Feb 03, 2026 at 02:29:55PM +0000, Pavel Begunkov wrote:
> Good day everyone,
> 
> dma-buf is a powerful abstraction for managing buffers and DMA mappings,
> and there is growing interest in extending it to the read/write path to
> enable device-to-device transfers without bouncing data through system
> memory. I was encouraged to submit it to LSF/MM/BPF as that might be
> useful to mull over details and what capabilities and features people
> may need.
> 
> The proposal consists of two parts. The first is a small in-kernel
> framework that allows a dma-buf to be registered against a given file
> and returns an object representing a DMA mapping. The actual mapping
> creation is delegated to the target subsystem (e.g. NVMe). This
> abstraction centralises request accounting, mapping management, dynamic
> recreation, etc. The resulting mapping object is passed through the I/O
> stack via a new iov_iter type.
> 
> As for the user API, a dma-buf is installed as an io_uring registered
> buffer for a specific file. Once registered, the buffer can be used by
> read / write io_uring requests as normal. io_uring will enforce that the
> buffer is only used with "compatible files", which is for now restricted
> to the target registration file, but will be expanded in the future.
> Notably, io_uring is a consumer of the framework rather than a
> dependency, and the infrastructure can be reused.
> 
> It took a couple of iterations on the list to get it to the current
> design, v2 of the series can be looked up at [1], which implements the
> infrastructure and initial wiring for NVMe. It slightly diverges from
> the description above, as some of the framework bits are block specific,
> and I'll be working on refining that and simplifying some of the
> interfaces for v3. A good chunk of block handling is based on prior work
> from Keith that was pre DMA mapping buffers [2].
> 
> Tushar was helping and mention he got good numbers for P2P transfers
> compared to bouncing it via RAM. Anuj, Kanchan and Nitesh also
> previously reported encouraging results for system memory backed
> dma-buf for optimising IOMMU overhead, quoting Anuj:
> 
> - STRICT: before = 570 KIOPS, after = 5.01 MIOPS
> - LAZY: before = 1.93 MIOPS, after = 5.01 MIOPS
> - PASSTHROUGH: before = 5.01 MIOPS, after = 5.01 MIOPS

Thanks for submitting the topic. The performance wins look great, but
I'm a little surpised passthrough didn't show any difference. We're
still skipping a bit of transformations with the dmabuf compared to not
having it, so maybe it's just a matter of crafting the right benchmark
to show the benefit.

Anyway, I look forward to the next version of this feature. I promise to
have more cycles to review and test the v3.

Re: [LSF/MM/BPF TOPIC] dmabuf backed read/write

[Anuj Gupta/Anuj Gupta]

On 2/3/2026 11:37 PM, Keith Busch wrote:
> Thanks for submitting the topic. The performance wins look great, but
> I'm a little surpised passthrough didn't show any difference. We're
> still skipping a bit of transformations with the dmabuf compared to not
> having it, so maybe it's just a matter of crafting the right benchmark
> to show the benefit.
> 

Those numbers were from a drive that saturates at ~5M IOPS, 
sopassthrough didn’t have much headroom. I did a quick run with two such
drives and saw a small improvement (~2–3%): ~5.97 MIOPS -> ~6.13 MIOPS,
but I’ll try tweaking the kernel config a bit to see if there’s more
headroom.

+1 on the topic - I'm interested in attending the discussion and
reviewing/testing v3 when it lands.

Thanks,
Anuj

Re: [LSF/MM/BPF TOPIC] dmabuf backed read/write

[Pavel Begunkov]

On 2/3/26 18:07, Keith Busch wrote:
> On Tue, Feb 03, 2026 at 02:29:55PM +0000, Pavel Begunkov wrote:
>> Good day everyone,
>>
...
>> Tushar was helping and mention he got good numbers for P2P transfers
>> compared to bouncing it via RAM. Anuj, Kanchan and Nitesh also
>> previously reported encouraging results for system memory backed
>> dma-buf for optimising IOMMU overhead, quoting Anuj:
>>
>> - STRICT: before = 570 KIOPS, after = 5.01 MIOPS
>> - LAZY: before = 1.93 MIOPS, after = 5.01 MIOPS
>> - PASSTHROUGH: before = 5.01 MIOPS, after = 5.01 MIOPS
> 
> Thanks for submitting the topic. The performance wins look great, but
> I'm a little surpised passthrough didn't show any difference. We're
> still skipping a bit of transformations with the dmabuf compared to not
> having it, so maybe it's just a matter of crafting the right benchmark
> to show the benefit.

My first thought was that hardware couldn't push more and would
be great to have idle numbers, but Anuj already demystified it.

> Anyway, I look forward to the next version of this feature. I promise to
> have more cycles to review and test the v3.

Thanks! And in general, IMHO at this point waiting for next
version would be more time efficient for reviewers.

-- 
Pavel Begunkov

Re: [LSF/MM/BPF TOPIC] dmabuf backed read/write

[Nitesh Shetty]

[-- Attachment #1: Type: text/plain, Size: 1934 bytes --]

On 03/02/26 02:29PM, Pavel Begunkov wrote:
>Good day everyone,
>
>dma-buf is a powerful abstraction for managing buffers and DMA mappings,
>and there is growing interest in extending it to the read/write path to
>enable device-to-device transfers without bouncing data through system
>memory. I was encouraged to submit it to LSF/MM/BPF as that might be
>useful to mull over details and what capabilities and features people
>may need.
>
>The proposal consists of two parts. The first is a small in-kernel
>framework that allows a dma-buf to be registered against a given file
>and returns an object representing a DMA mapping. The actual mapping
>creation is delegated to the target subsystem (e.g. NVMe). This
>abstraction centralises request accounting, mapping management, dynamic
>recreation, etc. The resulting mapping object is passed through the I/O
>stack via a new iov_iter type.
>
>As for the user API, a dma-buf is installed as an io_uring registered
>buffer for a specific file. Once registered, the buffer can be used by
>read / write io_uring requests as normal. io_uring will enforce that the
>buffer is only used with "compatible files", which is for now restricted
>to the target registration file, but will be expanded in the future.
>Notably, io_uring is a consumer of the framework rather than a
>dependency, and the infrastructure can be reused.
>
We have been following the series, its interesting from couple of angles,
- IOPS wise we see a major improvement especially for IOMMU
- Series provides a way to do p2pdma to accelerator memory

Here are few topics which I am looking into specifically,
- Right now the series uses a PRP list. We need a good way to keep the
   sg_table info around and decide on‑the‑fly whether to expose the buffer
   as a PRP list or an SG list, depending on the I/O size.
- Possibility of futher optimization for new type of iov iter to reduce
   per IO cost

Thanks,
Nitesh

[-- Attachment #2: Type: text/plain, Size: 0 bytes --]