| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Date: Wed, 17 Mar 2021 22:25:24 +0000
From: Alexander Lobakin <alobakin@...me>
To: Jesper Dangaard Brouer <brouer@...hat.com>
Cc: Alexander Lobakin <alobakin@...me>,
Mel Gorman <mgorman@...hsingularity.net>,
Andrew Morton <akpm@...ux-foundation.org>,
Chuck Lever <chuck.lever@...cle.com>,
Christoph Hellwig <hch@...radead.org>,
Alexander Duyck <alexander.duyck@...il.com>,
Matthew Wilcox <willy@...radead.org>,
LKML <linux-kernel@...r.kernel.org>,
Linux-Net <netdev@...r.kernel.org>,
Linux-MM <linux-mm@...ck.org>,
Linux-NFS <linux-nfs@...r.kernel.org>
Subject: Re: [PATCH 0/7 v4] Introduce a bulk order-0 page allocator with two in-tree users
From: Jesper Dangaard Brouer <brouer@...hat.com>
Date: Wed, 17 Mar 2021 18:19:43 +0100
> On Wed, 17 Mar 2021 16:52:32 +0000
> Alexander Lobakin <alobakin@...me> wrote:
>
> > From: Jesper Dangaard Brouer <brouer@...hat.com>
> > Date: Wed, 17 Mar 2021 17:38:44 +0100
> >
> > > On Wed, 17 Mar 2021 16:31:07 +0000
> > > Alexander Lobakin <alobakin@...me> wrote:
> > >
> > > > From: Mel Gorman <mgorman@...hsingularity.net>
> > > > Date: Fri, 12 Mar 2021 15:43:24 +0000
> > > >
> > > > Hi there,
> > > >
> > > > > This series is based on top of Matthew Wilcox's series "Rationalise
> > > > > __alloc_pages wrapper" and does not apply to 5.12-rc2. If you want to
> > > > > test and are not using Andrew's tree as a baseline, I suggest using the
> > > > > following git tree
> > > > >
> > > > > git://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.git mm-bulk-rebase-v4r2
> > > >
> > > > I gave this series a go on my setup, it showed a bump of 10 Mbps on
> > > > UDP forwarding, but dropped TCP forwarding by almost 50 Mbps.
> > > >
> > > > (4 core 1.2GHz MIPS32 R2, page size of 16 Kb, Page Pool order-0
> > > > allocations with MTU of 1508 bytes, linear frames via build_skb(),
> > > > GRO + TSO/USO)
> > >
> > > What NIC driver is this?
> >
> > Ah, forgot to mention. It's a WIP driver, not yet mainlined.
> > The NIC itself is basically on-SoC 1G chip.
>
> Hmm, then it is really hard to check if your driver is doing something
> else that could cause this.
>
> Well, can you try to lower the page_pool bulking size, to test the
> theory from Wilcox that we should do smaller bulking to avoid pushing
> cachelines into L2 when walking the LRU list. You might have to go as
> low as bulk=8 (for N-way associative level of L1 cache).
Turned out it suffered from GCC's decisions.
All of the following was taken on GCC 10.2.0 with -O2 in dotconfig.
vmlinux differences between baseline and this series:
(I used your followup instead of the last patch from the tree)
Function old new delta
__rmqueue_pcplist - 2024 +2024
__alloc_pages_bulk - 1456 +1456
__page_pool_alloc_pages_slow 284 600 +316
page_pool_dma_map - 164 +164
get_page_from_freelist 5676 3760 -1916
The uninlining of __rmqueue_pcplist() hurts a lot. It slightly slows
down the "regular" page allocator, but makes __alloc_pages_bulk()
much slower than the per-page (in my case at least) due to calling
this function out from the loop.
One possible solution is to mark __rmqueue_pcplist() and
rmqueue_bulk() as __always_inline. Only both and only with
__always_inline, or GCC will emit rmqueue_bulk.constprop and
make the numbers even poorer.
This nearly doubles the size of bulk allocator, but eliminates
all performance hits.
Function old new delta
__alloc_pages_bulk 1456 3512 +2056
get_page_from_freelist 3760 5744 +1984
find_suitable_fallback.part - 160 +160
min_free_kbytes_sysctl_handler 96 128 +32
find_suitable_fallback 164 28 -136
__rmqueue_pcplist 2024 - -2024
Between baseline and this series with __always_inline hints:
Function old new delta
__alloc_pages_bulk - 3512 +3512
find_suitable_fallback.part - 160 +160
get_page_from_freelist 5676 5744 +68
min_free_kbytes_sysctl_handler 96 128 +32
find_suitable_fallback 164 28 -136
Another suboptimal place I've found is two functions in Page Pool
code which are marked as 'noinline'.
Maybe there's a reason behind this, but removing the annotations
and additionally marking page_pool_dma_map() as inline simplifies
the object code and in fact improves the performance (+15 Mbps on
my setup):
add/remove: 0/3 grow/shrink: 1/0 up/down: 1024/-1096 (-72)
Function old new delta
page_pool_alloc_pages 100 1124 +1024
page_pool_dma_map 164 - -164
page_pool_refill_alloc_cache 332 - -332
__page_pool_alloc_pages_slow 600 - -600
1124 is a normal size for a hotpath function.
These fragmentation and jumps between page_pool_alloc_pages(),
__page_pool_alloc_pages_slow() and page_pool_refill_alloc_cache()
are really excessive and unhealthy for performance, as well as
page_pool_dma_map() uninlined by GCC.
So the best results I got so far were with these additional changes:
- mark __rmqueue_pcplist() as __always_inline;
- mark rmqueue_bulk() as __always_inline;
- drop 'noinline' from page_pool_refill_alloc_cache();
- drop 'noinline' from __page_pool_alloc_pages_slow();
- mark page_pool_dma_map() as inline.
(inlines in C files aren't generally recommended, but well, GCC
is far from perfect)
> In function: __page_pool_alloc_pages_slow() adjust variable:
> const int bulk = PP_ALLOC_CACHE_REFILL;
Regarding bulk size, it makes no sense on my machine. I tried
{ 8, 16, 32, 64 } and they differed by 1-2 Mbps max / standard
deviation.
Most of the bulk operations I've seen usually take the value of
16 as a "golden ratio" though.
>
> --
> Best regards,
> Jesper Dangaard Brouer
> MSc.CS, Principal Kernel Engineer at Red Hat
> LinkedIn: http://www.linkedin.com/in/brouer
Thanks,
Al
Powered by blists - more mailing lists