Received-SPF: pass (google.com: domain of linux-wireless-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <ba8e3ae1-70a9-e016-6c44-ae0c52f97e8e@wetzel-home.de>
Date:   Fri, 8 Jul 2022 18:42:32 +0200
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Thunderbird/91.11.0
Subject: Re: [RFC] mac80211: Switch all drivers over to iTXQs
Content-Language: en-US
To:     Johannes Berg <johannes@sipsolutions.net>
Cc:     linux-wireless@vger.kernel.org
References: <20220707143230.21686-1-alexander@wetzel-home.de>
From:   Alexander Wetzel <alexander@wetzel-home.de>
In-Reply-To: <20220707143230.21686-1-alexander@wetzel-home.de>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk

> This patch is working for at least iwldvm. It exposes what seems to
> be an existing iwldvm driver bug, though. Driver and card continues
> to work, making that a minor side show for now. (I started to look into
> that a bit, maybe more about that in another mail later.)
> Initial tests are indicating comparable speeds in my setup for iwldvm,
> maybe trending to be a bit slower.
> 
> For now I'm interested what I may have missed or what should be done
> differently, so we dare to merge it. I'll planning to run some more
> tests and use iwldvm with this patch on my main workstation for some
> days at least.

I followed up on the iwldvm driver issue.
The problem is, that the push API drivers are expecting that tx 
immediately stops after calling ieee80211_stop_queue(). And this patch 
breaks that.

For iwldvm this happens:
- the HW queues are getting full, iwldvm calls ieee80211_stop_queue()

- the (patched) mac80211 stops the queue but is not aborting the running
   wake_tx_queue() tasklet. Which goes on to hand over skbs to the driver

- The driver runs out of the "internal reserve" hw buffers and starts to
   add the new skbs to the overflow queue

- some hw buffers are reclaimed and the driver tries to process the skbs
   in the overflow q

- For that it releases the internal lock for tx, assuming mac80211
   stopped handing over skbs

- the still running wake_tx_queue() function from mac80211 hands over
   the next skb, grapping the slot(s) intended to be used for the skbs in
   the driver overflow queue.

- which breaks an iwlwifi internal rule how sequence numbers are mapped
   to hw queue slots, causing the warning I got.

I'll fix that and send a new patch.

Alexander