Message ID | 20210606103656.71079-1-sven@svenpeter.dev (mailing list archive) |
---|---|
State | Superseded |
Headers | show |
Series | [v2] usb: dwc3: support 64 bit DMA in platform driver | expand |
On Sun, Jun 6, 2021 at 12:36 PM Sven Peter <sven@svenpeter.dev> wrote: > diff --git a/drivers/usb/dwc3/core.c b/drivers/usb/dwc3/core.c > index b6e53d8212cd..4930541a8984 100644 > --- a/drivers/usb/dwc3/core.c > +++ b/drivers/usb/dwc3/core.c > @@ -1545,6 +1545,21 @@ static int dwc3_probe(struct platform_device *pdev) > > dwc3_get_properties(dwc); > > + /* Try to set 64-bit DMA first */ > + if (!dwc->sysdev->dma_mask) > + /* Platform did not initialize dma_mask */ > + ret = dma_coerce_mask_and_coherent(dwc->sysdev, > + DMA_BIT_MASK(64)); > + else > + ret = dma_set_mask_and_coherent(dwc->sysdev, DMA_BIT_MASK(64)); > + > + /* If seting 64-bit DMA mask fails, fall back to 32-bit DMA mask */ > + if (ret) { > + ret = dma_set_mask_and_coherent(dwc->sysdev, DMA_BIT_MASK(32)); > + if (ret) > + return ret; > + } Please drop the dma_coerce_mask_and_coherent() code path as well: if the device is marked as non-DMA capable in the platform, it's better have it not be usable at all than to assume a particular bus property that may or may not be present on that bus. The 32-bit mask is the default on all buses you might see a dwc3 controller on, so you can drop that as well, and just leave the dma_set_mask_and_coherent(). Arnd
On Sun, Jun 6, 2021, at 17:43, Arnd Bergmann wrote: > On Sun, Jun 6, 2021 at 12:36 PM Sven Peter <sven@svenpeter.dev> wrote: > > diff --git a/drivers/usb/dwc3/core.c b/drivers/usb/dwc3/core.c > > index b6e53d8212cd..4930541a8984 100644 > > --- a/drivers/usb/dwc3/core.c > > +++ b/drivers/usb/dwc3/core.c > > @@ -1545,6 +1545,21 @@ static int dwc3_probe(struct platform_device *pdev) > > > > dwc3_get_properties(dwc); > > > > + /* Try to set 64-bit DMA first */ > > + if (!dwc->sysdev->dma_mask) > > + /* Platform did not initialize dma_mask */ > > + ret = dma_coerce_mask_and_coherent(dwc->sysdev, > > + DMA_BIT_MASK(64)); > > + else > > + ret = dma_set_mask_and_coherent(dwc->sysdev, DMA_BIT_MASK(64)); > > + > > + /* If seting 64-bit DMA mask fails, fall back to 32-bit DMA mask */ > > + if (ret) { > > + ret = dma_set_mask_and_coherent(dwc->sysdev, DMA_BIT_MASK(32)); > > + if (ret) > > + return ret; > > + } > > Please drop the dma_coerce_mask_and_coherent() code path as well: if > the device is marked as non-DMA capable in the platform, it's better have > it not be usable at all than to assume a particular bus property that may > or may not be present on that bus. > > The 32-bit mask is the default on all buses you might see a dwc3 controller > on, so you can drop that as well, and just leave the > dma_set_mask_and_coherent(). > > Arnd > Makes sense, will drop the two things for v3. Thanks, Sven
diff --git a/drivers/usb/dwc3/core.c b/drivers/usb/dwc3/core.c index b6e53d8212cd..4930541a8984 100644 --- a/drivers/usb/dwc3/core.c +++ b/drivers/usb/dwc3/core.c @@ -1545,6 +1545,21 @@ static int dwc3_probe(struct platform_device *pdev) dwc3_get_properties(dwc); + /* Try to set 64-bit DMA first */ + if (!dwc->sysdev->dma_mask) + /* Platform did not initialize dma_mask */ + ret = dma_coerce_mask_and_coherent(dwc->sysdev, + DMA_BIT_MASK(64)); + else + ret = dma_set_mask_and_coherent(dwc->sysdev, DMA_BIT_MASK(64)); + + /* If seting 64-bit DMA mask fails, fall back to 32-bit DMA mask */ + if (ret) { + ret = dma_set_mask_and_coherent(dwc->sysdev, DMA_BIT_MASK(32)); + if (ret) + return ret; + } + dwc->reset = devm_reset_control_array_get_optional_shared(dev); if (IS_ERR(dwc->reset)) return PTR_ERR(dwc->reset);
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 using the same logic already present in xhci-plat.c: First, try to set a coherent dma mask for 64-bit, and then attempt again with a 32-bit mask if this fails. Signed-off-by: Sven Peter <sven@svenpeter.dev> --- I have taken the code directly from the xhci-plat.c driver so I think this change should be fairly low risk. Unfortunately I only have the Apple M1 to test this on but here the driver still works with the iommu enabled which limits the address space to 32 bit. It also enables to use this with the iommu in bypass mode which requires 64 bit addresses. I believe this has been working fine so far since the dwc3 driver only uses a few very small buffers in host mode which might still fit within the first 4G of address space on many devices. The majority of DMA buffers are allocated inside the xhci driver which will already call dma_set_mask_and_coherent. Best, Sven changes from v1: - removed WARN_ON around !dwc->sysdev->dma_mask; pointed out by greg k-h drivers/usb/dwc3/core.c | 15 +++++++++++++++ 1 file changed, 15 insertions(+)