Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Subject: Re: [PATCH v1] dmaengine: tegra: Use relaxed versions of readl/writel
To:     Jon Hunter <jonathanh@nvidia.com>,
        Laxman Dewangan <ldewangan@nvidia.com>,
        Vinod Koul <vkoul@kernel.org>,
        Thierry Reding <thierry.reding@gmail.com>
Cc:     dmaengine@vger.kernel.org, linux-tegra@vger.kernel.org,
        linux-kernel@vger.kernel.org
References: <20190424231708.21219-1-digetx@gmail.com>
 <d789f1bc-44d1-abf6-a046-7cf835416e8f@nvidia.com>
From:   Dmitry Osipenko <digetx@gmail.com>
Message-ID: <4a315b63-bc71-3c3e-f1ae-8638bcf4033d@gmail.com>
Date:   Fri, 26 Apr 2019 13:45:29 +0300
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:60.0) Gecko/20100101
 Thunderbird/60.6.1
MIME-Version: 1.0
In-Reply-To: <d789f1bc-44d1-abf6-a046-7cf835416e8f@nvidia.com>
Content-Type: text/plain; charset=utf-8
Content-Language: en-US
Content-Transfer-Encoding: 8bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

26.04.2019 12:52, Jon Hunter пишет:
> 
> On 25/04/2019 00:17, Dmitry Osipenko wrote:
>> The readl/writel functions are inserting memory barrier in order to
>> ensure that memory stores are completed. On Tegra20 and Tegra30 this
>> results in L2 cache syncing which isn't a cheapest operation. The
>> tegra20-apb-dma driver doesn't need to synchronize generic memory
>> accesses, hence use the relaxed versions of the functions.
> 
> Do you mean device-io accesses here as this is not generic memory?

Yes. The IOMEM accesses within are always ordered and uncached, while
generic memory accesses are out-of-order and cached.

> Although there may not be any issues with this change, I think I need a
> bit more convincing that we should do this given that we have had it
> this way for sometime and I would not like to see us introduce any
> regressions as this point without being 100% certain we would not.
> Ideally, if I had some good extensive tests I could run to hammer the
> DMA for all configurations with different combinations of channels
> running simultaneously then we could test this, but right now I don't :-(
> 
> Have you ...
> 1. Tested both cyclic and scatter-gather transfers?
> 2. Stress tested simultaneous transfers with various different
>    configurations?
> 3. Quantified the actual performance benefit of this change so we can
>    understand how much of a performance boost this offers?

Actually I found a case where this change causes a problem, I'm seeing
I2C transfer timeout for touchscreen and it breaks the touch input.
Indeed, I haven't tested this patch very well.

And the fix is this:

@@ -1592,6 +1592,8 @@ static int tegra_dma_runtime_suspend(struct device
*dev)
 						  TEGRA_APBDMA_CHAN_WCOUNT);
 	}

+	dsb();
+
 	clk_disable_unprepare(tdma->dma_clk);

 	return 0;


Apparently the problem is that CLK/DMA (PPSB/APB) accesses are
incoherent and CPU disables clock before writes are reaching DMA controller.

I'd say that cyclic and scatter-gather transfers are now tested. I also
made some more testing of simultaneous transfers.

Quantifying performance probably won't be easy to make as the DMA
read/writes are not on any kind of code's hot-path.

Jon, are you still insisting about to drop this patch or you will be
fine with the v2 that will have the dsb() in place?