| Message ID | 20230301160315.1022488-1-aleksander.lobakin@intel.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | xdp: recycle Page Pool backed skbs built from XDP frames | expand |
On 01/03/2023 17.03, Alexander Lobakin wrote: > Yeah, I still remember that "Who needs cpumap nowadays" (c), but anyway. > > __xdp_build_skb_from_frame() missed the moment when the networking stack > became able to recycle skb pages backed by a Page Pool. This was making ^^^^^^^^^ When talking about page_pool, can we write "page_pool" instead of capitalized "Page Pool", please. I looked through the git log, and here we all used "page_pool". > e.g. cpumap redirect even less effective than simple %XDP_PASS. veth was > also affected in some scenarios. Thanks for working on closing this gap :-) > A lot of drivers use skb_mark_for_recycle() already, it's been almost > two years and seems like there are no issues in using it in the generic > code too. {__,}xdp_release_frame() can be then removed as it losts its > last user. > Page Pool becomes then zero-alloc (or almost) in the abovementioned > cases, too. Other memory type models (who needs them at this point) > have no changes. > > Some numbers on 1 Xeon Platinum core bombed with 27 Mpps of 64-byte > IPv6 UDP: What NIC driver? > > Plain %XDP_PASS on baseline, Page Pool driver: > > src cpu Rx drops dst cpu Rx > 2.1 Mpps N/A 2.1 Mpps > > cpumap redirect (w/o leaving its node) on baseline: > > 6.8 Mpps 5.0 Mpps 1.8 Mpps > > cpumap redirect with skb PP recycling: > > 7.9 Mpps 5.7 Mpps 2.2 Mpps +22% > It is of cause awesome, that cpumap SKBs are faster than normal SKB path. I do wonder where the +22% number comes from? > Alexander Lobakin (2): > xdp: recycle Page Pool backed skbs built from XDP frames > xdp: remove unused {__,}xdp_release_frame() > > include/net/xdp.h | 29 ----------------------------- > net/core/xdp.c | 19 ++----------------- > 2 files changed, 2 insertions(+), 46 deletions(-) >
From: Jesper Dangaard Brouer <jbrouer@redhat.com> Date: Fri, 3 Mar 2023 11:39:06 +0100 > > On 01/03/2023 17.03, Alexander Lobakin wrote: >> Yeah, I still remember that "Who needs cpumap nowadays" (c), but anyway. >> >> __xdp_build_skb_from_frame() missed the moment when the networking stack >> became able to recycle skb pages backed by a Page Pool. This was making > ^^^^^^^^^ > When talking about page_pool, can we write "page_pool" instead of > capitalized "Page Pool", please. I looked through the git log, and here > we all used "page_pool". Ah okay, no prob :D Yeah, that's probably more correct. "Page Pool" is the name of the API, while page_pool is an entity we create via page_pool_create(). > >> e.g. cpumap redirect even less effective than simple %XDP_PASS. veth was >> also affected in some scenarios. > > Thanks for working on closing this gap :-) > >> A lot of drivers use skb_mark_for_recycle() already, it's been almost >> two years and seems like there are no issues in using it in the generic >> code too. {__,}xdp_release_frame() can be then removed as it losts its >> last user. >> Page Pool becomes then zero-alloc (or almost) in the abovementioned >> cases, too. Other memory type models (who needs them at this point) >> have no changes. >> >> Some numbers on 1 Xeon Platinum core bombed with 27 Mpps of 64-byte >> IPv6 UDP: > > What NIC driver? IAVF with XDP, the series adding XDP support will be sent in a couple weeks, WIP can be found on my open GH[0]. > >> >> Plain %XDP_PASS on baseline, Page Pool driver: >> >> src cpu Rx drops dst cpu Rx >> 2.1 Mpps N/A 2.1 Mpps >> >> cpumap redirect (w/o leaving its node) on baseline: >> >> 6.8 Mpps 5.0 Mpps 1.8 Mpps >> >> cpumap redirect with skb PP recycling: >> >> 7.9 Mpps 5.7 Mpps 2.2 Mpps +22% >> > > It is of cause awesome, that cpumap SKBs are faster than normal SKB path. That's the point of cpumap redirect, right? You separate NAPI poll / IRQ handling from the skb networking stack traveling to a different CPU, including page freeing (or recycling). That takes a lot of load from the source CPU. 0.1 Mpps is not the highest difference I got, cpumap redirect can boost up to 0.5 Mpps IIRC. > I do wonder where the +22% number comes from? (2.2 - 1.8) / 1.8 * 100%. I compare baseline cpumap redirect before/after here :) > >> Alexander Lobakin (2): >> xdp: recycle Page Pool backed skbs built from XDP frames >> xdp: remove unused {__,}xdp_release_frame() >> >> include/net/xdp.h | 29 ----------------------------- >> net/core/xdp.c | 19 ++----------------- >> 2 files changed, 2 insertions(+), 46 deletions(-) >> > There's a build failure on non-PP systems due to skb_mark_for_recycle() being declared only when CONFIG_PAGE_POOL is set. I'll spin v2 in a bit. [0] https://github.com/alobakin/linux/commits/iavf-xdp Thanks, Olek
Yeah, I still remember that "Who needs cpumap nowadays" (c), but anyway. __xdp_build_skb_from_frame() missed the moment when the networking stack became able to recycle skb pages backed by a Page Pool. This was making e.g. cpumap redirect even less effective than simple %XDP_PASS. veth was also affected in some scenarios. A lot of drivers use skb_mark_for_recycle() already, it's been almost two years and seems like there are no issues in using it in the generic code too. {__,}xdp_release_frame() can be then removed as it losts its last user. Page Pool becomes then zero-alloc (or almost) in the abovementioned cases, too. Other memory type models (who needs them at this point) have no changes. Some numbers on 1 Xeon Platinum core bombed with 27 Mpps of 64-byte IPv6 UDP: Plain %XDP_PASS on baseline, Page Pool driver: src cpu Rx drops dst cpu Rx 2.1 Mpps N/A 2.1 Mpps cpumap redirect (w/o leaving its node) on baseline: 6.8 Mpps 5.0 Mpps 1.8 Mpps cpumap redirect with skb PP recycling: 7.9 Mpps 5.7 Mpps 2.2 Mpps +22% Alexander Lobakin (2): xdp: recycle Page Pool backed skbs built from XDP frames xdp: remove unused {__,}xdp_release_frame() include/net/xdp.h | 29 ----------------------------- net/core/xdp.c | 19 ++----------------- 2 files changed, 2 insertions(+), 46 deletions(-)