On Tue, Sep 01, 2015 at 02:54:17PM +0300, Mathias Nyman wrote:

On 31.08.2015 21:58, Duc Dang wrote:

quoted

On Thu, Aug 20, 2015 at 12:38 PM, Duc Dang [off-list ref] wrote:

quoted

The xhci platform driver needs to work on systems that
either only support 64-bit DMA or only support 32-bit DMA.
Attempt to set a coherent dma mask for 64-bit DMA, and
attempt again with 32-bit DMA if that fails.

[dhdang: regenerate the patch over 4.2-rc5 and address new comments]
Signed-off-by: Mark Langsdorf <redacted>
Tested-by: Mark Salter <redacted>
Signed-off-by: Duc Dang <redacted>

---
Changes from v6:
        -Add WARN_ON if dma_mask is NULL
        -Use dma_coerce_mask_and_coherent to assign
        dma_mask and coherent_dma_mask

Changes from v5:
        -Change comment
        -Assign dma_mask to coherent_dma_mask if dma_mask is NULL
        to make sure dma_set_mask_and_coherent does not fail prematurely.

Changes from v4:
        -None

Changes from v3:
        -Re-generate the patch over 4.2-rc5
        -No code change.

Changes from v2:
        -None

Changes from v1:
        -Consolidated to use dma_set_mask_and_coherent
        -Got rid of the check against sizeof(dma_addr_t)

 drivers/usb/host/xhci-plat.c | 21 +++++++++++++--------
 1 file changed, 13 insertions(+), 8 deletions(-)

diff --git a/drivers/usb/host/xhci-plat.c b/drivers/usb/host/xhci-plat.c
index 890ad9d..e4c7f9d 100644
--- a/drivers/usb/host/xhci-plat.c
+++ b/drivers/usb/host/xhci-plat.c

@@ -93,14 +93,19 @@ static int xhci_plat_probe(struct platform_device *pdev)
        if (irq < 0)
                return -ENODEV;

-       /* Initialize dma_mask and coherent_dma_mask to 32-bits */
-       ret = dma_set_coherent_mask(&pdev->dev, DMA_BIT_MASK(32));
-       if (ret)
-               return ret;
-       if (!pdev->dev.dma_mask)
-               pdev->dev.dma_mask = &pdev->dev.coherent_dma_mask;
-       else
-               dma_set_mask(&pdev->dev, DMA_BIT_MASK(32));
+       /* Throw a waring if broken platform code didn't initialize dma_mask */
+       WARN_ON(!pdev->dev.dma_mask);
+       /*
+        * Try setting dma_mask and coherent_dma_mask to 64 bits,
+        * then try 32 bits
+        */
+       ret = dma_coerce_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
+       if (ret) {
+               ret = dma_coerce_mask_and_coherent(&pdev->dev,
+                                                  DMA_BIT_MASK(32));
+               if (ret)
+                       return ret;
+       }

This isn't very good.  If dev.dma_mask is already set,
dma_coerce_mask_and_coherent() will always overwrite it.  There's also
no need to call it twice.  This, imho, is much better:

	/* Try to set a 64-bit DMA mask first */
	if (WARN_ON(!pdev->dev.dma_mask)) {
		/* Eek, platform didn't initialise the streaming DMA mask */
		ret = dma_coerce_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
	} else {
		ret = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
	}

	/* If that failed, fall back to a 32-bit DMA mask */
	if (ret) {
		ret = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32));
		if (ret)
			return ret;
	}

since it preserves the dev.dma_mask pointer if it was properly setup

Really, drivers shouldn't be messing around with that pointer - especially
if it's already been correctly setup.  A platform may require separate
streaming and coherent masks, and we should respect that.

(The whole dma_mask being a pointer thing is a left-over from the PCI
layer which has never been cleaned up through fear of breaking something.)

-- 
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.