Re: [linux-usb-devel] Re: USB 2.0 ehci failure with large amount of RAM (4GB) on x86_64

[David Brownell]

On Tuesday 09 May 2006 10:37 pm, Nathan Becker wrote:
> I added 1 line to drivers/usb/host/ehci-pci.c which sets the DMA mask,
> and now it seems to work with ehci loaded and with 4 GB of RAM.
> Unfortunately, I don't really understand what I did.  Perhaps you have
> a better idea what this is doing and if it is correct.

Interesting.  My guess is that the IOMMU is helping you out, and the
issue is that the silicon erratum applies to the I/O buffers too, not
just to a subset of the schedule data structures.

Can you confirm that this patch also resolves your issue?  Reboot
from scratch a few times (power off, warm reboot, etc) to be
reasonably sure you're seeing the cases that failed before.

Modify workaround for the EHCI quirk in newer NVidia controllers by
assigning the DMA mask, not just the consistent DMA mask.  It seems
to matter, as if the silicon erratum isn't quite what was described
in the original bug workaround (or there's another erratum).


Index: g26/drivers/usb/host/ehci-pci.c
===================================================================
--- g26.orig/drivers/usb/host/ehci-pci.c	2006-05-12 19:40:22.000000000 -0700
+++ g26/drivers/usb/host/ehci-pci.c	2006-05-12 21:15:31.000000000 -0700
@@ -118,7 +118,15 @@ static int ehci_pci_setup(struct usb_hcd
 			if (pci_set_consistent_dma_mask(pdev,
 						DMA_31BIT_MASK) < 0)
 				ehci_warn(ehci, "can't enable NVidia "
-					"workaround for >2GB RAM\n");
+					"workaround %d for >2GB RAM\n", 1);
+
+			/* some users report problems with 4GB when the IOMMU
+			 * doesn't force lowmem addresses... looks like the
+			 * initial report from NVidia wasn't quite right.
+			 */
+			if (pci_set_dma_mask(pdev, DMA_31BIT_MASK) < 0)
+				ehci_warn(ehci, "can't enable NVidia "
+					"workaround %d for >2GB RAM\n", 2);
 			break;
 		/* Some NForce2 chips have problems with selective suspend;
 		 * fixed in newer silicon.