Re: [PATCH 06/18] wifi: iwlwifi: keep the TSO and workaround pages mapped

[Ben Greear <greearb@candelatech.com>]

On 7/19/24 00:22, Berg, Benjamin wrote:
> Hi,
> 
> have you applied all patches? It'll be slightly slower until you apply
> the last patch in the series (and break completely unless you have the
> bugfix obviously).

I applied those three in my email below, on top of everything else that seemed
pertintent to iwlwifi/mac80211, but always possible I mis-applied something
or missed another patch.

I think I'll just leave those out and test again when i rebase on 6.11.
If you do start getting other reports of slowdowns, however...I think these
will be the suspect patches.

Thanks,
Ben

> 
> A slowdown to 80MBit/s seems a bit like we are getting packet loss and
> the AMSDU is simply not send out. I suppose, that could happen if some
> of the DMA mapping is failing, but I really do not have a good idea
> right now.
> 
> I do think that these patches have gone through quite some verification
> and were working well for us.
> 
> 
> As for the value of the patchset, the thing to keep in mind is that we
> send out large SKBs that are spanning multiple pages. Before the patch,
> we would do mapping as below, where h1, … and m1, … are each mappings
> that only span a small part of a page (h_n is the header for the nth
> packet while m_n is the corresponding message data):
> 
> TSO page: |- h1 -|- h2 -|- h3 -|- h4 -| …
> SKB page 1: |- m1     -|- m2      -|
> SKB page 2: |- m2 -|-  m3  -|- m4 -|
> SKB page 3: |- m4 -|- …
> 
> With the patches applied, each of the pages will be mapped in a single
> chunk. This brings down the number of DMA mappings and also avoids
> partial mappings which can very be particularly expensive depending on
> the IOMMU configuration.
> 
> Only the last patch in the series removes the old mappings (obviously
> you also need the later bugfix commit). So you might see a slight
> slowdown if the patch "wifi: iwlwifi: use already mapped data when
> TXing an AMSDU" is not applied.
> 
> But not a slowdown to 80MBit/s …
> 
> Benjamin
> 
> On Thu, 2024-07-18 at 12:20 -0700, Ben Greear wrote:
>> On 7/11/24 09:09, Ben Greear wrote:
>>> On 7/10/24 23:15, Berg, Benjamin wrote:
>>>> Hi Ben,
>>>>
>>>> yes, you need to apply:
>>>>
>>>> commit 003eae5a28c6c9d50290a4ac9b955be912f24c9f
>>>> Author: Benjamin Berg <benjamin.berg@intel.com>
>>>> Date:   Tue Jul 9 14:31:49 2024 +0200
>>>>
>>>>       wifi: iwlwifi: correctly reference TSO page information
>>>>
>>>>
>>>> I had not fully tested the last revision and the error slipped
>>>> unfortunately.
>>>
>>> Hello Benjamin,
>>>
>>> Sorry I did not notice that patch on the mailing list on my own.  I
>>> re-applied
>>> the 6/18 and 7/18 patches, and the fix you mention above, and
>>> system appears stable.
>>
>> Hello,
>>
>> We found another regression in our patched 6.10-ish kernel.  Before I
>> apply these 3
>> patches:
>>
>> wifi: iwlwifi: keep the TSO and workaround pages mapped
>> wifi: iwlwifi: use already mapped data when TXing an AMSDU
>> wifi: iwlwifi: correctly reference TSO page information
>>
>> Then I see around 4Gbps TCP upload on my test rig.  After this, it
>> runs very poorly,
>> perhaps bouncing up to high speed for a second or two, but mostly
>> averaging 80Mbps
>> or so after it runs for a bit.
>>
>> What are these patches trying to solve, and are you able to see good
>> TCP upload performance
>> with these patches applied?
>>
>> Thanks,
>> Ben
>>
> 
> Intel Deutschland GmbH
> Registered Address: Am Campeon 10, 85579 Neubiberg, Germany
> Tel: +49 89 99 8853-0, www.intel.de
> Managing Directors: Sean Fennelly, Jeffrey Schneiderman, Tiffany Doon Silva
> Chairperson of the Supervisory Board: Nicole Lau
> Registered Office: Munich
> Commercial Register: Amtsgericht Muenchen HRB 186928

-- 
Ben Greear <greearb@candelatech.com>
Candela Technologies Inc  http://www.candelatech.com