On 10/1/2021 3:02 PM, Konstantin Ananyev wrote: > v3 changes: > - Changes in public struct naming (Jerin/Haiyue) > - Split patches > - Update docs > - Shamelessly included Andrew's patch: > > https://patches.dpdk.org/project/dpdk/patch/20210928154856.1015020-1-andrew.rybche...@oktetlabs.ru/ > into these series. > I have to do similar thing here, so decided to avoid duplicated effort. > > The aim of these patch series is to make rte_ethdev core data structures > (rte_eth_dev, rte_eth_dev_data, rte_eth_rxtx_callback, etc.) internal to > DPDK and not visible to the user. > That should allow future possible changes to core ethdev related structures > to be transparent to the user and help to improve ABI/API stability. > Note that current ethdev API is preserved, but it is a formal ABI break. > > The work is based on previous discussions at: > https://www.mail-archive.com/dev@dpdk.org/msg211405.html > https://www.mail-archive.com/dev@dpdk.org/msg216685.html > and consists of the following main points: > 1. Copy public 'fast' function pointers (rx_pkt_burst(), etc.) and > related data pointer from rte_eth_dev into a separate flat array. > We keep it public to still be able to use inline functions for these > 'fast' calls (like rte_eth_rx_burst(), etc.) to avoid/minimize slowdown. > Note that apart from function pointers itself, each element of this > flat array also contains two opaque pointers for each ethdev: > 1) a pointer to an array of internal queue data pointers > 2) points to array of queue callback data pointers. > Note that exposing this extra information allows us to avoid extra > changes inside PMD level, plus should help to avoid possible > performance degradation. > 2. Change implementation of 'fast' inline ethdev functions > (rte_eth_rx_burst(), etc.) to use new public flat array. > While it is an ABI breakage, this change is intended to be transparent > for both users (no changes in user app is required) and PMD developers > (no changes in PMD is required). > One extra note - with new implementation RX/TX callback invocation > will cost one extra function call with this changes. That might cause > some slowdown for code-path with RX/TX callbacks heavily involved. > Hope such trade-off is acceptable for the community. > 3. Move rte_eth_dev, rte_eth_dev_data, rte_eth_rxtx_callback and related > things into internal header: <ethdev_driver.h>. > > That approach was selected to: > - Avoid(/minimize) possible performance losses. > - Minimize required changes inside PMDs. >
Overall +1 to the approach. Only 'metrics' library is failing to build and it also needs to include driver header: diff --git a/lib/metrics/rte_metrics_telemetry.c b/lib/metrics/rte_metrics_telemetry.c index 269f8ef6135c..5be21b2e86c5 100644 --- a/lib/metrics/rte_metrics_telemetry.c +++ b/lib/metrics/rte_metrics_telemetry.c @@ -2,7 +2,7 @@ * Copyright(c) 2020 Intel Corporation */ -#include <rte_ethdev.h> +#include <ethdev_driver.h> #include <rte_string_fns.h> #ifdef RTE_LIB_TELEMETRY #include <telemetry_internal.h> > Performance testing results (ICX 2.0GHz, E810 (ice)): > - testpmd macswap fwd mode, plus > a) no RX/TX callbacks: > no actual slowdown observed > b) bpf-load rx 0 0 JM ./dpdk.org/examples/bpf/t3.o: > ~2% slowdown > - l3fwd: no actual slowdown observed > > Would like to thank Ferruh and Jerrin for reviewing and testing previous > versions of these series. All other interested parties please don't be shy > and provide your feedback. > > Konstantin Ananyev (7): > ethdev: allocate max space for internal queue array > ethdev: change input parameters for rx_queue_count > ethdev: copy ethdev 'fast' API into separate structure > ethdev: make burst functions to use new flat array > ethdev: add API to retrieve multiple ethernet addresses > ethdev: remove legacy Rx descriptor done API > ethdev: hide eth dev related structures > <...>
