Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20210607061751.89752-1-sven@svenpeter.dev>
In-Reply-To: <20210607061751.89752-1-sven@svenpeter.dev>
From:   Arnd Bergmann <arnd@arndb.de>
Date:   Mon, 7 Jun 2021 09:25:43 +0200
Message-ID: <CAK8P3a0vbyq-90pUQ6-0Ed=DadR3Pnf0juupLQ70psQSuu_1nw@mail.gmail.com>
Subject: Re: [PATCH v3] usb: dwc3: support 64 bit DMA in platform driver
To:     Sven Peter <sven@svenpeter.dev>
Cc:     USB list <linux-usb@vger.kernel.org>,
        Felipe Balbi <balbi@kernel.org>,
        Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Mon, Jun 7, 2021 at 8:18 AM Sven Peter <sven@svenpeter.dev> wrote:
>
> Currently, the dwc3 platform driver does not explicitly ask for
> a DMA mask. This makes it fall back to the default 32-bit mask which
> breaks the driver on systems that only have RAM starting above the
> first 4G like the Apple M1 SoC.
>
> Fix this by calling dma_set_mask_and_coherent with a 64bit mask.
>
> Signed-off-by: Sven Peter <sven@svenpeter.dev>
> ---
>
> Third time's a charm I hope - this time much simpler :)

I think this is almost good, but there is still one small issue:

> diff --git a/drivers/usb/dwc3/core.c b/drivers/usb/dwc3/core.c
> index b6e53d8212cd..ba4792b6a98f 100644
> --- a/drivers/usb/dwc3/core.c
> +++ b/drivers/usb/dwc3/core.c
> @@ -1545,6 +1545,10 @@ static int dwc3_probe(struct platform_device *pdev)
>
>         dwc3_get_properties(dwc);
>
> +       ret = dma_set_mask_and_coherent(dwc->sysdev, DMA_BIT_MASK(64));
> +       if (ret)
> +               return ret;

This will now  fail on machines with dwc3 connected to a 32-bit bus (or a
bus that is accidentally not annotated as supporting 64-bit) when there is
some memory that is not addressable through that bus.

If dma_set_mask_and_coherent() fails, the platform should just fall back to
32-bit addressing as it did before your change. dma_alloc_*() will do that
implicitly by allocating from ZONE_DMA32, while dma_map_*() fails
on any non-addressable memory, or falls back to swiotlb if that is available.

        Arnd