[bpf-next] bpf/test_run: increase Page Pool's ptr_ring size in live frames mode

  Currently, when running xdp-trafficgen, test_run creates page_pools with
the ptr_ring size of %NAPI_POLL_WEIGHT (64).
This might work fine if XDP Tx queues are polled with the budget
limitation. However, we often clear them with no limitation to ensure
maximum free space when sending.
For example, in ice and idpf (upcoming), we use "lazy" cleaning, i.e. we
clean XDP Tx queue only when the free space there is less than 1/4 of
the queue size. Let's take the ring size of 512 just as an example. 3/4
of the ring is 384 and often times, when we're entering the cleaning
function, we have this whole amount ready (or 256 or 192, doesn't
matter).
Then we're calling xdp_return_frame_bulk() and after 64th frame,
page_pool_put_page_bulk() starts returning pages to the page allocator
due to that the ptr_ring is already full. put_page(), alloc_page() et at
starts consuming a ton of CPU time and leading the board of the perf top
output.

Let's not limit ptr_ring to 64 for no real reason and allow more pages
to be recycled. Just don't put anything to page_pool_params::size and
let the Page Pool core pick the default of 1024 entries (I don't believe
there are real use cases to clean more than that amount of descriptors).
After the change, the MM layer disappears from the perf top output and
all pages get recycled to the PP. On my test setup on idpf with the
default ring size (512), this gives +80% of Tx performance with no
visible memory consumption increase.

Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
---
 net/bpf/test_run.c | 1 -
 1 file changed, 1 deletion(-)
  

Alexander Lobakin <aleksander.lobakin@intel.com> writes:

> Currently, when running xdp-trafficgen, test_run creates page_pools with
> the ptr_ring size of %NAPI_POLL_WEIGHT (64).
> This might work fine if XDP Tx queues are polled with the budget
> limitation. However, we often clear them with no limitation to ensure
> maximum free space when sending.
> For example, in ice and idpf (upcoming), we use "lazy" cleaning, i.e. we
> clean XDP Tx queue only when the free space there is less than 1/4 of
> the queue size. Let's take the ring size of 512 just as an example. 3/4
> of the ring is 384 and often times, when we're entering the cleaning
> function, we have this whole amount ready (or 256 or 192, doesn't
> matter).
> Then we're calling xdp_return_frame_bulk() and after 64th frame,
> page_pool_put_page_bulk() starts returning pages to the page allocator
> due to that the ptr_ring is already full. put_page(), alloc_page() et at
> starts consuming a ton of CPU time and leading the board of the perf top
> output.
>
> Let's not limit ptr_ring to 64 for no real reason and allow more pages
> to be recycled. Just don't put anything to page_pool_params::size and
> let the Page Pool core pick the default of 1024 entries (I don't believe
> there are real use cases to clean more than that amount of descriptors).
> After the change, the MM layer disappears from the perf top output and
> all pages get recycled to the PP. On my test setup on idpf with the
> default ring size (512), this gives +80% of Tx performance with no
> visible memory consumption increase.
>
> Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>

Hmm, so my original idea with keeping this low was to avoid having a lot
of large rings lying around if it is used by multiple processes at once.
But we need to move away from the per-syscall allocation anyway, and
with Lorenzo's patches introducing a global system page pool we have an
avenue for that. So in the meantime, I have no objection to this...

Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
  

Toke Høiland-Jørgensen <toke@redhat.com> writes:

> Alexander Lobakin <aleksander.lobakin@intel.com> writes:
>
>> Currently, when running xdp-trafficgen, test_run creates page_pools with
>> the ptr_ring size of %NAPI_POLL_WEIGHT (64).
>> This might work fine if XDP Tx queues are polled with the budget
>> limitation. However, we often clear them with no limitation to ensure
>> maximum free space when sending.
>> For example, in ice and idpf (upcoming), we use "lazy" cleaning, i.e. we
>> clean XDP Tx queue only when the free space there is less than 1/4 of
>> the queue size. Let's take the ring size of 512 just as an example. 3/4
>> of the ring is 384 and often times, when we're entering the cleaning
>> function, we have this whole amount ready (or 256 or 192, doesn't
>> matter).
>> Then we're calling xdp_return_frame_bulk() and after 64th frame,
>> page_pool_put_page_bulk() starts returning pages to the page allocator
>> due to that the ptr_ring is already full. put_page(), alloc_page() et at
>> starts consuming a ton of CPU time and leading the board of the perf top
>> output.
>>
>> Let's not limit ptr_ring to 64 for no real reason and allow more pages
>> to be recycled. Just don't put anything to page_pool_params::size and
>> let the Page Pool core pick the default of 1024 entries (I don't believe
>> there are real use cases to clean more than that amount of descriptors).
>> After the change, the MM layer disappears from the perf top output and
>> all pages get recycled to the PP. On my test setup on idpf with the
>> default ring size (512), this gives +80% of Tx performance with no
>> visible memory consumption increase.
>>
>> Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
>
> Hmm, so my original idea with keeping this low was to avoid having a lot
> of large rings lying around if it is used by multiple processes at once.
> But we need to move away from the per-syscall allocation anyway, and
> with Lorenzo's patches introducing a global system page pool we have an
> avenue for that. So in the meantime, I have no objection to this...
>
> Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>

Actually, since Lorenzo's patches already landed in net-next, let's just
move to using those straight away. I'll send a patch for this tomorrow :)

-Toke
  

From: Toke Høiland-Jørgensen <toke@redhat.com>
Date: Thu, 15 Feb 2024 00:02:27 +0100

> Toke Høiland-Jørgensen <toke@redhat.com> writes:
> 
>> Alexander Lobakin <aleksander.lobakin@intel.com> writes:
>>
>>> Currently, when running xdp-trafficgen, test_run creates page_pools with
>>> the ptr_ring size of %NAPI_POLL_WEIGHT (64).
>>> This might work fine if XDP Tx queues are polled with the budget
>>> limitation. However, we often clear them with no limitation to ensure
>>> maximum free space when sending.
>>> For example, in ice and idpf (upcoming), we use "lazy" cleaning, i.e. we
>>> clean XDP Tx queue only when the free space there is less than 1/4 of
>>> the queue size. Let's take the ring size of 512 just as an example. 3/4
>>> of the ring is 384 and often times, when we're entering the cleaning
>>> function, we have this whole amount ready (or 256 or 192, doesn't
>>> matter).
>>> Then we're calling xdp_return_frame_bulk() and after 64th frame,
>>> page_pool_put_page_bulk() starts returning pages to the page allocator
>>> due to that the ptr_ring is already full. put_page(), alloc_page() et at
>>> starts consuming a ton of CPU time and leading the board of the perf top
>>> output.
>>>
>>> Let's not limit ptr_ring to 64 for no real reason and allow more pages
>>> to be recycled. Just don't put anything to page_pool_params::size and
>>> let the Page Pool core pick the default of 1024 entries (I don't believe
>>> there are real use cases to clean more than that amount of descriptors).
>>> After the change, the MM layer disappears from the perf top output and
>>> all pages get recycled to the PP. On my test setup on idpf with the
>>> default ring size (512), this gives +80% of Tx performance with no
>>> visible memory consumption increase.
>>>
>>> Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
>>
>> Hmm, so my original idea with keeping this low was to avoid having a lot
>> of large rings lying around if it is used by multiple processes at once.
>> But we need to move away from the per-syscall allocation anyway, and
>> with Lorenzo's patches introducing a global system page pool we have an
>> avenue for that. So in the meantime, I have no objection to this...
>>
>> Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
> 
> Actually, since Lorenzo's patches already landed in net-next, let's just
> move to using those straight away. I'll send a patch for this tomorrow :)

Keep in mind that system page_pools do direct recycling based on cpuid
and for now, memory leaks are possible. Pls see my patch[0] for the
details :D

> 
> -Toke
> 

[0]
https://lore.kernel.org/netdev/20240215113905.96817-1-aleksander.lobakin@intel.com

Olek