On 12/2/25 14:44, Paolo Abeni wrote:
On 12/1/25 12:35 AM, Pavel Begunkov wrote:
Note: it's net/ only bits and doesn't include changes, which shoulf be
merged separately and are posted separately. The full branch for
convenience is at [1], and the patch is here:
https://lore.kernel.org/io-uring/7486ab32e99be1f614b3ef8d0e9bc77015b173f7.1764265323.git.asml.sile...@gmail.com
Many modern NICs support configurable receive buffer lengths, and zcrx and
memory providers can use buffers larger than 4K/PAGE_SIZE on x86 to improve
performance. When paired with hw-gro larger rx buffer sizes can drastically
reduce the number of buffers traversing the stack and save a lot of processing
time. It also allows to give to users larger contiguous chunks of data. The
idea was first floated around by Saeed during netdev conf 2024 and was
asked about by a few folks.
Single stream benchmarks showed up to ~30% CPU util improvement.
E.g. comparison for 4K vs 32K buffers using a 200Gbit NIC:
packets=23987040 (MB=2745098), rps=199559 (MB/s=22837)
CPU %usr %nice %sys %iowait %irq %soft %idle
0 1.53 0.00 27.78 2.72 1.31 66.45 0.22
packets=24078368 (MB=2755550), rps=200319 (MB/s=22924)
CPU %usr %nice %sys %iowait %irq %soft %idle
0 0.69 0.00 8.26 31.65 1.83 57.00 0.57
This series adds net infrastructure for memory providers configuring
the size and implements it for bnxt. It's an opt-in feature for drivers,
they should advertise support for the parameter in the qops and must check
if the hardware supports the given size. It's limited to memory providers
as it drastically simplifies implementation. It doesn't affect the fast
path zcrx uAPI, and the sizes is defined in zcrx terms, which allows it
to be flexible and adjusted in the future, see Patch 8 for details.
A liburing example can be found at [2]
full branch:
[1] https://github.com/isilence/linux.git zcrx/large-buffers-v7
Liburing example:
[2] https://github.com/isilence/liburing.git zcrx/rx-buf-len
Dump question, hoping someone could answer in a very short time...
Differently from previous revisions, this is not a PR, just a plain
patch series - that in turn may cause duplicate commits when applied on
different trees.
Is the above intentional? why?
It was based on linus-rc* before and getting merged nice and clean,
now there is a small conflict. In my view, it should either be a
separate pull to Linus that depends on the net+io_uring trees if
Jens would be willing to orchestrate that, or I'll just merge the
leftover io_uring patch for-6.20. In either case, this set shouldn't
get applied to any other tree directly.
--
Pavel Begunkov
