Message-ID: <c4748c8f-ae4b-9bbc-7d20-9c7b7ce433c8@st.com>
Date:   Fri, 2 Dec 2016 14:51:34 +0100
From:   Giuseppe CAVALLARO <peppe.cavallaro@...com>
To:     Pavel Machek <pavel@....cz>
CC:     <alexandre.torgue@...com>, David Miller <davem@...emloft.net>,
        <netdev@...r.kernel.org>, <linux-kernel@...r.kernel.org>,
        Alexandre TORGUE <alexandre.torgue@...com>
Subject: Re: stmmac ethernet in kernel 4.9-rc6: coalescing related pauses.

On 12/2/2016 1:32 PM, Pavel Machek wrote:
> Hi!
>
>>> Well, if you have a workload that sends and receive packets, it tends
>>> to work ok, as you do tx_clean() in stmmac_poll(). My workload is not
>>> like that -- it is "sending packets at 3MB/sec, receiving none". So
>>> the stmmac_tx_timer() is rescheduled and rescheduled and rescheduled,
>>> and then we run out of transmit descriptors, and then 40msec passes,
>>> and then we clean them. Bad.
>>>
>>> And that's why low-res timers do not cut it.
>>
>> in that case, I expect that the tuning of the driver could help you.
>> I mean, by using ethtool, it could be enough to set the IC bit on all
>> the descriptors. You should touch the tx_coal_frames.
>>
>> Then you can use ethtool -S to monitor the status.
>
> Yes, I did something similar. Unfortnunately that meant crash within
> minutes, at least with 4.4 kernel. (If you know what was fixed between
> 4.4 and 4.9, that would be helpful).

4.4 has no GMAC4 support.
Alex, do you remember any patches to fix that?

>> We had experimented this tuning on STB IP where just datagrams
>> had to send externally. To be honest, although we had seen
>> better results w/o any timer, we kept this approach enabled
>> because the timer was fast enough to cover our tests on SH4 boxes.
>
> Please reply to David, and explain how it is supposed to
> work... because right now it does not. 40 msec delays are not
> acceptable in default configuration.

I mean, that on UP and SMP system this schema helped
to improve the performance saving CPU on my side and this has been
tested since a long time (~4 years).
I tested something similar to yours where unidirectional traffic
with limited throughput was needed and I can confirm you that
tuning/removing coalesce parameters this helped. The tuning I decided
to keep in the driver was suitable in several user cases and if now
you have problems or you want to review it I can just confirm that
there are no problems on my side. If you want to simply the logic
around the tx process and remove timer on official driver I can accept
that. I will just ask you uto double check if the throughput and
CPU usage when request max throughput (better if on GiGa setup) has
no regressions.
Otherwise we could start thinking about adaptive schema if feasible.

>>>> In the ring, some descriptors can raise the irq (according to a
>>>> threshold) and set the IC bit. In this path, the NAPI  poll will be
>>>> scheduled.
>>>
>>> Not NAPI poll but stmmac_tx_timer(), right?
>>
>> in the xmit according the the threshold the timer is started or the
>> interrupt is set inside the descriptor.
>> Then stmmac_tx_clean will be always called and, if you see the flow,
>> no irqlock protection is needed!
>
> Agreed that no irqlock protection is needed if we rely on napi and timers.

ok

>
>>>> Concerning the lock protection, we had reviewed long time ago and
>>>> IIRC, no raise condition should be present. Open to review it,
>>>> again!
> ...
>>> There's nothing that protect stmmac_poll() from running concurently
>>> with stmmac_dma_interrupt(), right?
>>
>> This is not necessary.
>
> dma_interrupt accesses shared priv->xstats; variables are of type
> unsigned long (not atomic_t), yet they are accesssed from interrupt
> context and from stmmac_ethtool without any locking. That can result
> in broken statistics AFAICT.

ok we can check this and welcome patches and I'd prefer to
remove xstats from critical part of the code like ISR (that
comes from old story of the driver).

>
> Please take another look. As far as I can tell, you can have two cpus
> at #1 and #2 in the code, at the same time. It looks like napi_... has
> some atomic opertions inside so that looks safe at the first look. But
> I'm not sure if they also include enough memory barriers to make it
> safe...?

Although I have never reproduced related issues on SMP platforms
due to reordering of memory operations but, as said above, welcome
review on this especially if you are seeing problems when manage NAPI.

FYI, the only memory barrier you will see in the driver are about the
OWN_BIT setting till now.

> static void stmmac_dma_interrupt(struct stmmac_priv *priv)
> {
> ...
>         status = priv->hw->dma->dma_interrupt(priv->ioaddr, &priv->xstats);
>         if (likely((status & handle_rx)) || (status & handle_tx)) {
>                 if (likely(napi_schedule_prep(&priv->napi))) {
> #1
>                         stmmac_disable_dma_irq(priv);
>                         __napi_schedule(&priv->napi);
>                 }
>         }
>
>
> static int stmmac_poll(struct napi_struct *napi, int budget)
> {
>         struct stmmac_priv *priv = container_of(napi, struct stmmac_priv, napi);
>         int work_done = 0;
>
>         priv->xstats.napi_poll++;
>         stmmac_tx_clean(priv);
>
>         work_done = stmmac_rx(priv, budget);
> 	if (work_done < budget) {
>                 napi_complete(napi);
> #2
> 	        stmmac_enable_dma_irq(priv);
>         }

hmm, I have to check (and refresh my memory) but the driver
uses the napi_schedule_prep.

Regards

Peppe

>         return work_done;
> }
>
>
> Best regards,
> 									Pavel
>

Powered by blists - more mailing lists