Re: [PATCH rdma-next v5 00/15] RDMA: Introduce generic buffer descriptor infrastructure for umem

[Jiri Pirko <jiri@resnulli.us>]

Mon, May 18, 2026 at 03:30:54PM +0200, leon@kernel.org wrote:
>On Sun, May 17, 2026 at 08:29:51AM +0200, Jiri Pirko wrote:
>> From: Jiri Pirko <jiri@nvidia.com>
>> 
>> This patchset introduces a generic buffer descriptor infrastructure
>> for passing memory buffers (dma-buf or user VA) to uverbs commands,
>> and wires it up for CQ and QP creation in the uverbs core, efa, mlx5,
>> bnxt_re and mlx4 drivers.
>> 
>> Instead of adding ad-hoc per-buffer UAPI attributes for each new buffer
>> type, each command declares dedicated UVERBS_ATTR_UMEM attributes that
>> carry one buffer descriptor each. Each descriptor specifies a buffer
>> type, covering both VA and dma-buf. A consumption check ensures
>> userspace and driver agree on which attributes are used.
>> 
>> The patchset:
>> 1-2,4. Plumbing: rename ib_umem_get() to ib_umem_get_va() and re-route
>>    it through the new central ib_umem_get(); no behaviour change.
>> 3. Introduces the core buffer descriptor infrastructure and UAPI.
>> 5. Factors out CQ buffer umem processing into a helper.
>> 6. Adds the CQ buffer UMEM attribute and driver wrappers.
>> 7-10. Converts efa, mlx5, bnxt_re and mlx4 to use the new CQ helpers,
>>    with drivers taking umem ownership.
>> 11. Removes the legacy umem field from struct ib_cq, now that all
>>    drivers use the new helpers.
>> 12. Adds optional whole-QP, RQ and SQ UMEM attributes to QP creation.
>> 13. Converts mlx5 QP creation to use the new attributes.
>> 14-15. Adds mlx5 driver-namespace UMEM attributes for CQ and QP
>>    doorbell records.
>> 
>> ---
>> Based on top of: https://lore.kernel.org/all/0-v1-045258567bd6+9fe-ib_uverbs_support_ko_jgg@nvidia.com/
>
>The overall approach looks sane to me.
>
>I have a few minor comments, mostly related to how I expect the
>ib_umem_get_*() helpers to be used:
>
>1. We should hide the ib_umem_get() function and export only
>   ib_umem_get_va(), ib_umem_get_attr(), and
>   ib_umem_get_attr_or_va(). I do not want to see drivers relying on
>   ib_umem_get() again.
>
>2. Pass a struct uverbs_attr_bundle *attrs instead of struct ib_udata
>   *udata to ib_umem_get().
>
>3. Consider simplifying ib_umem_get() further by passing in the desc.
>   I am not fully convinced it helps, but it could allow handling
>   va_fallback outside of ib_umem_get().

Okay, addressing all of that in v6.

Thanks!