Re: [RFC v1 00/22] Large rx buffer support for zcrx
From: Stanislav Fomichev <hidden>
Date: 2025-07-28 17:13:37
Also in:
io-uring
On 07/28, Pavel Begunkov wrote:
This series implements large rx buffer support for io_uring/zcrx on top of Jakub's queue configuration changes, but it can also be used by other memory providers. Large rx buffers can be drastically beneficial with high-end hw-gro enabled cards that can coalesce traffic into larger pages, reducing the number of frags traversing the network stack and resuling in larger contiguous chunks of data for the userspace. Benchamrks showed up to ~30% improvement in CPU util. For example, for 200Gbit broadcom NIC, 4K vs 32K buffers, and napi and userspace pinned to the same CPU: 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 And for napi and userspace on different CPUs: packets=10725082 (MB=1227388), rps=198285 (MB/s=22692) CPU %usr %nice %sys %iowait %irq %soft %idle 0 0.10 0.00 0.50 0.00 0.50 74.50 24.40 1 4.51 0.00 44.33 47.22 2.08 1.85 0.00 packets=14026235 (MB=1605175), rps=198388 (MB/s=22703) CPU %usr %nice %sys %iowait %irq %soft %idle 0 0.10 0.00 0.70 0.00 1.00 43.78 54.42 1 1.09 0.00 31.95 62.91 1.42 2.63 0.00 Patch 19 allows to pass queue config from a memory provider. The zcrx changes are contained in a single patch as I already queued most of work making it size agnostic into my zcrx branch. The uAPI is simple and imperative, it'll use the exact value (if) specified by the user. In the future we might extend it to "choose the best size in a given range". The rest (first 20) patches are from Jakub's series implementing per queue configuration. Quoting Jakub: "... The direct motivation for the series is that zero-copy Rx queues would like to use larger Rx buffers. Most modern high-speed NICs support HW-GRO, and can coalesce payloads into pages much larger than than the MTU. Enabling larger buffers globally is a bit precarious as it exposes us to potentially very inefficient memory use. Also allocating large buffers may not be easy or cheap under load. Zero-copy queues service only select traffic and have pre-allocated memory so the concerns don't apply as much. The per-queue config has to address 3 problems: - user API - driver API - memory provider API For user API the main question is whether we expose the config via ethtool or netdev nl. I picked the latter - via queue GET/SET, rather than extending the ethtool RINGS_GET API. I worry slightly that queue GET/SET will turn in a monster like SETLINK. OTOH the only per-queue settings we have in ethtool which are not going via RINGS_SET is IRQ coalescing. My goal for the driver API was to avoid complexity in the drivers. The queue management API has gained two ops, responsible for preparing configuration for a given queue, and validating whether the config is supported. The validating is used both for NIC-wide and per-queue changes. Queue alloc/start ops have a new "config" argument which contains the current config for a given queue (we use queue restart to apply per-queue settings). Outside of queue reset paths drivers can call netdev_queue_config() which returns the config for an arbitrary queue. Long story short I anticipate it to be used during ndo_open. In the core I extended struct netdev_config with per queue settings. All in all this isn't too far from what was there in my "queue API prototype" a few years ago ..."
Supporting big buffers is the right direction, but I have the same feedback: it would be nice to fit a cohesive story for the devmem as well. We should also aim for another use-case where we allocate page pool chunks from the huge page(s), this should push the perf even more. We need some way to express these things from the UAPI point of view. Flipping the rx-buf-len value seems too fragile - there needs to be something to request 32K chunks only for devmem case, not for the (default) CPU memory. And the queues should go back to default 4K pages when the dmabuf is detached from the queue.