Post AmpxaL58rZ0ClrGyPI by axboe@fosstodon.org
(DIR) More posts by axboe@fosstodon.org
(DIR) Post #AmptbFpBUmxKwEzQDw by axboe@fosstodon.org
2024-10-09T15:40:31Z
0 likes, 0 repeats
David recently posted the first non-rfc version of the io_uring zerocopy receive work, you can find it here:https://lore.kernel.org/io-uring/ZwW7_cRr_UpbEC-X@LQ3V64L9R2/T/I noticed Broadcom has public fw that supports this now, so really very few (and minor) dependencies left for this work.I ran some testing. Details in the reply below:https://lore.kernel.org/io-uring/2e475d9f-8d39-43f4-adc5-501897c951a8@kernel.dk/tldr - line rate, 59% faster.
(DIR) Post #AmptbGvFPjOsLKpmJU by pro@mu.zaitcev.nu
2024-10-09T16:56:50.071046Z
0 likes, 0 repeats
@axboe Is AF_XDP dead now?
(DIR) Post #AmpueJcZw5nxXK7JdQ by axboe@fosstodon.org
2024-10-09T16:58:28Z
0 likes, 0 repeats
@pro I have no horse in that race.
(DIR) Post #AmpueKaqK20iYEJRZI by pro@mu.zaitcev.nu
2024-10-09T17:08:37.001117Z
0 likes, 0 repeats
@axboe We're looking at migrating away from DPDK to something better supported by kernel drivers. We looked at AF_XDP but it isn't performing and has severe limitations regarding packet alignment that ruin everything. I suppose io_uring is next to try.
(DIR) Post #AmpxaL58rZ0ClrGyPI by axboe@fosstodon.org
2024-10-09T17:28:41Z
0 likes, 0 repeats
@pro I mean, any kind of bypass solution is, in my opinion, the wrong direction. So if it renders AF_XDP less useful, then I won't shed any tears.
(DIR) Post #AmpxaLlKKikKsh0fA0 by pro@mu.zaitcev.nu
2024-10-09T17:41:30.154347Z
0 likes, 0 repeats
@axboe Oh fine, DPDK it is, then. Forever.