On Tue, Jul 07, 2026 at 12:24:21PM -0700, Mina Almasry wrote:
(I'm kinda reviewing this very late here. Some suggestions/comments but feel free to ignore if not useful).
On Wed, Jul 1, 2026 at 12:22 PM Bobby Eshleman bobbyeshleman@gmail.com wrote:
Every devmem dmabuf binding hands the page_pool PAGE_SIZE niovs today. On NICs that consume one descriptor per netmem, this caps a single RX descriptor at PAGE_SIZE and burns CPU on buffer churn.
In this series, we add a bind-time netlink attribute, NETDEV_A_DMABUF_RX_BUF_SIZE, that lets userspace request a larger niov size (power of two >= PAGE_SIZE).
FWIW we may be able to support arbitrary sizes with devmem. Because the genpool supports byte-aligned allocations AFAIR. Also the dma-mapping happens with the dma-buf size, so the actual niov size doesn't matter. The only thing I can think off which may not be flexible to arbitrary sizes is the driver itself. IDK what happens if you ask the driver to dma into a buffer that is frag size 5023 or something like that.
But that is something that can be relaxed in the future.
I think at least for mlx5 there would be some issues, as it splits the memory region into fixed-size strides (256B), so I'd expect it needs to at least be divisible by the stride length. The mlx5 driver seems to guard against this by checking for sz > PAGE_SIZE && is_power_of_2.
Drivers must opt in via queue_mgmt_ops.QCFG_RX_PAGE_SIZE.
nit that probably doesn't matter: ...QCFG_RX_NETMEM_SIZE, or (...NIOV_SIZE). This doesn't actually work with pages, right?
I probably could have worded this in the message more clearly, but this name is not introduced by this series, so we probably can't get away with changing it.
If you decide to extend to arbrary sizes, I would add to the queue_mgmt ops supports_netmem_size(size_t size) function, and let the driver enforce "it has to be power of 2" if it needs to. AFAICT core doesn't need to.
Selftests use udmabuf, but udmabuf sgtables were previously hardcoded to PAGE_SIZE. This series modifies udmabuf to respect folio sizes in its exported sgtable. The result is that when backing udmabuf with MFD_HUGETLB 2MB pages, the sgtable is populated with 2MB entries, allowing devmem's gen_pool to carve out large (eg. 64K) niovs.
Measurements
Setup: kperf devmem RX/TX cuda, 4 flows, 64 MB messages, 60s, dctcp, num-rx-queues=4, dmabuf-rx/tx-size-mb=2048, 10 runs per niov size, mlx5.
niov RX dev Gbps RX flow avg Gbps app sys %
4K 300.63 +/- 53.21 75.16 +/- 13.30 54.15 +/- 10.23 16K 321.35 +/- 28.20 80.34 +/- 7.05 41.05 +/- 8.87 32K 347.63 +/- 2.20 86.91 +/- 0.55 44.54 +/- 3.51 64K 332.11 +/- 14.26 83.03 +/- 3.56 35.47 +/- 3.11RX app sys % drops ~19% from 4K to 64K.
Hard to read the columns for me but seems like good perf data. Did performance become worse from 32K to 64K? I wonder why.
The drop off struck my eye too, but didn't investigate further. Given the wide stdev, it appears to me like the trend is positive but probably not huge. The cpu util deltas, on the other hand, look stronger to me.
I have some devmem performance fixes that are very critical for our production that I haven't gotten around to upstreaming yet. I wonder if I can send them to you for upstream submission. Are you potentially interested?
Definitely interested!
-- Thanks, Mina
Thanks Mina.
Best, Bobby