From: Ilias Apalodimas <ilias.apalodi...@linaro.org>
Date: Tue, 23 Mar 2021 19:01:52 +0200
> On Tue, Mar 23, 2021 at 04:55:31PM +0000, Alexander Lobakin wrote:
> > > > > > >
>
> [...]
>
> > > > > >
> > > > > > Thanks for the testing!
> > > > > > Any chance you can get a perf measurement on this?
> > > > >
> > > > > I guess you mean perf-report (--stdio) output, right?
> > > > >
> > > >
> > > > Yea,
> > > > As hinted below, I am just trying to figure out if on Alexander's
> > > > platform the
> > > > cost of syncing, is bigger that free-allocate. I remember one armv7
> > > > were that
> > > > was the case.
> > > >
> > > > > > Is DMA syncing taking a substantial amount of your cpu usage?
> > > > >
> > > > > (+1 this is an important question)
> >
> > Sure, I'll drop perf tools to my test env and share the results,
> > maybe tomorrow or in a few days.
Oh we-e-e-ell...
Looks like I've been fooled by I-cache misses or smth like that.
That happens sometimes, not only on my machines, and not only on
MIPS if I'm not mistaken.
Sorry for confusing you guys.
I got drastically different numbers after I enabled CONFIG_KALLSYMS +
CONFIG_PERF_EVENTS for perf tools.
The only difference in code is that I rebased onto Mel's
mm-bulk-rebase-v6r4.
(lunar is my WIP NIC driver)
1. 5.12-rc3 baseline:
TCP: 566 Mbps
UDP: 615 Mbps
perf top:
4.44% [lunar] [k] lunar_rx_poll_page_pool
3.56% [kernel] [k] r4k_wait_irqoff
2.89% [kernel] [k] free_unref_page
2.57% [kernel] [k] dma_map_page_attrs
2.32% [kernel] [k] get_page_from_freelist
2.28% [lunar] [k] lunar_start_xmit
1.82% [kernel] [k] __copy_user
1.75% [kernel] [k] dev_gro_receive
1.52% [kernel] [k] cpuidle_enter_state_coupled
1.46% [kernel] [k] tcp_gro_receive
1.35% [kernel] [k] __rmemcpy
1.33% [nf_conntrack] [k] nf_conntrack_tcp_packet
1.30% [kernel] [k] __dev_queue_xmit
1.22% [kernel] [k] pfifo_fast_dequeue
1.17% [kernel] [k] skb_release_data
1.17% [kernel] [k] skb_segment
free_unref_page() and get_page_from_freelist() consume a lot.
2. 5.12-rc3 + Page Pool recycling by Matteo:
TCP: 589 Mbps
UDP: 633 Mbps
perf top:
4.27% [lunar] [k] lunar_rx_poll_page_pool
2.68% [lunar] [k] lunar_start_xmit
2.41% [kernel] [k] dma_map_page_attrs
1.92% [kernel] [k] r4k_wait_irqoff
1.89% [kernel] [k] __copy_user
1.62% [kernel] [k] dev_gro_receive
1.51% [kernel] [k] cpuidle_enter_state_coupled
1.44% [kernel] [k] tcp_gro_receive
1.40% [kernel] [k] __rmemcpy
1.38% [nf_conntrack] [k] nf_conntrack_tcp_packet
1.37% [kernel] [k] free_unref_page
1.35% [kernel] [k] __dev_queue_xmit
1.30% [kernel] [k] skb_segment
1.28% [kernel] [k] get_page_from_freelist
1.27% [kernel] [k] r4k_dma_cache_inv
+20 Mbps increase on both TCP and UDP. free_unref_page() and
get_page_from_freelist() dropped down the list significantly.
3. 5.12-rc3 + Page Pool recycling + PP bulk allocator (Mel & Jesper):
TCP: 596
UDP: 641
perf top:
4.38% [lunar] [k] lunar_rx_poll_page_pool
3.34% [kernel] [k] r4k_wait_irqoff
3.14% [kernel] [k] dma_map_page_attrs
2.49% [lunar] [k] lunar_start_xmit
1.85% [kernel] [k] dev_gro_receive
1.76% [kernel] [k] free_unref_page
1.76% [kernel] [k] __copy_user
1.65% [kernel] [k] inet_gro_receive
1.57% [kernel] [k] tcp_gro_receive
1.48% [kernel] [k] cpuidle_enter_state_coupled
1.43% [nf_conntrack] [k] nf_conntrack_tcp_packet
1.42% [kernel] [k] __rmemcpy
1.25% [kernel] [k] skb_segment
1.21% [kernel] [k] r4k_dma_cache_inv
+10 Mbps on top of recycling.
get_page_from_freelist() is gone.
NAPI polling, CPU idle cycle (r4k_wait_irqoff) and DMA mapping
routine became the top consumers.
4-5. __always_inline for rmqueue_bulk() and __rmqueue_pcplist(),
removing 'noinline' from net/core/page_pool.c etc.
...makes absolutely no sense anymore.
I see Mel took Jesper's patch to make __rmqueue_pcplist() inline into
mm-bulk-rebase-v6r5, not sure if it's really needed now.
So I'm really glad we sorted out the things and I can see the real
performance improvements from both recycling and bulk allocations.
> > From what I know for sure about MIPS and my platform,
> > post-Rx synching (dma_sync_single_for_cpu()) is a no-op, and
> > pre-Rx (dma_sync_single_for_device() etc.) is a bit expensive.
> > I always have sane page_pool->pp.max_len value (smth about 1668
> > for MTU of 1500) to minimize the overhead.
> >
> > By the word, IIRC, all machines shipped with mvpp2 have hardware
> > cache coherency units and don't suffer from sync routines at all.
> > That may be the reason why mvpp2 wins the most from this series.
>
> Yep exactly. It's also the reason why you explicitly have to opt-in using the
> recycling (by marking the skb for it), instead of hiding the feature in the
> page pool internals
>
> Cheers
> /Ilias
>
> >
> > > > > > >
> > > > > > > [0]
> > > > > > > https://lore.kernel.org/netdev/20210323153550.130385-1-aloba...@pm.me
> > > > > > >
> > > > >
> > >
> > > That would be the same as for mvneta:
> > >
> > > Overhead Shared Object Symbol
> > > 24.10% [kernel] [k] __pi___inval_dcache_area
> > > 23.02% [mvneta] [k] mvneta_rx_swbm
> > > 7.19% [kernel] [k] kmem_cache_alloc
> > >
> > > Anyway, I tried to use the recycling *and* napi_build_skb on mvpp2,
> > > and I get lower packet rate than recycling alone.
> > > I don't know why, we should investigate it.
> >
> > mvpp2 driver doesn't use napi_consume_skb() on its Tx completion path.
> > As a result, NAPI percpu caches get refilled only through
> > kmem_cache_alloc_bulk(), and most of skbuff_head recycling
> > doesn't work.
> >
> > > Regards,
> > > --
> > > per aspera ad upstream
> >
> > Oh, I love that one!
> >
> > Al
> >
Thanks,
Al
