Re: [PATCH v4 0/6] zcrx huge pages support Vol 1

[Pavel Begunkov <asml.silence@gmail.com>]

On 7/3/25 00:12, David Wei wrote:
> On 2025-07-02 07:29, Pavel Begunkov wrote:
>> Use sgtable as the common format b/w dmabuf and user pages, deduplicate
>> dma address propagation handling, and use it to omptimise dma mappings.
>> It also prepares it for larger size for NIC pages.
>>
>> v4: further restrict kmap fallback length, v3 isn't be generic
>>      enough to cover future uses.
>>
>> v3: truncate kmap'ed length by both pages
>>
>> v2: Don't coalesce into folios, just use sg_alloc_table_from_pages()
>>      for now. Coalescing will return back later.
>>
>>      Improve some fallback copy code. Patch 1, and Patch 6 adding a
>>      helper to work with larger pages, which also allows to get rid
>>      of skb_frag_foreach_page.
>>
>>      Return copy fallback helpers back to pages instead of folios,
>>      the latter wouldn't be correct in all cases.
>>
>> Pavel Begunkov (6):
>>    io_uring/zcrx: always pass page to io_zcrx_copy_chunk
>>    io_uring/zcrx: return error from io_zcrx_map_area_*
>>    io_uring/zcrx: introduce io_populate_area_dma
>>    io_uring/zcrx: allocate sgtable for umem areas
>>    io_uring/zcrx: assert area type in io_zcrx_iov_page
>>    io_uring/zcrx: prepare fallback for larger pages
>>
>>   io_uring/zcrx.c | 241 +++++++++++++++++++++++++-----------------------
>>   io_uring/zcrx.h |   1 +
>>   2 files changed, 128 insertions(+), 114 deletions(-)
>>
> 
> What did you use to test this patch series or will that come with vol 2?

I used my libuirng example with different areas types but didn't
do anything in terms of benchmarking. This series only improves
iommu, and I don't care too much about that.

-- 
Pavel Begunkov