On Wed, Jun 20, 2012 at 11:15:26PM +0200, Marek Vasut wrote:
Dear Tom Rini,
On Wed, Jun 20, 2012 at 09:00:45PM +0200, Marek Vasut wrote:
Dear Tom Rini,
The USB spec says that 32 bytes is the minimum required alignment. However on some platforms we have a larger minimum requirement for cache coherency. In those cases, use that value rather than the USB spec minimum. We add a cpp check to <usb.h> to define USB_DMA_MINALIGN and make use of it in ehci-hcd.c and musb_core.h. We cannot use MAX() here as we are not allowed to have tests inside of align(...).
Cc: Marek Vasut marex@denx.de Signed-off-by: Tom Rini trini@ti.com
-- Changes in v2:
- Move test to <usb.h>, expand comment.
drivers/usb/host/ehci-hcd.c | 13 +++++++------ drivers/usb/musb/musb_core.h | 2 +- include/usb.h | 10 ++++++++++ 3 files changed, 18 insertions(+), 7 deletions(-)
diff --git a/drivers/usb/host/ehci-hcd.c b/drivers/usb/host/ehci-hcd.c index 04300be..5a86117 100644 --- a/drivers/usb/host/ehci-hcd.c +++ b/drivers/usb/host/ehci-hcd.c @@ -29,12 +29,13 @@
#include "ehci.h"
-int rootdev; -struct ehci_hccr *hccr; /* R/O registers, not need for volatile */ -volatile struct ehci_hcor *hcor; +int rootdev __attribute__((aligned(USB_DMA_MINALIGN))); +/* R/O registers, not need for volatile */ +struct ehci_hccr *hccr __attribute__((aligned(USB_DMA_MINALIGN))); +volatile struct ehci_hcor *hcor __attribute__((aligned(USB_DMA_MINALIGN)));
^ these need to be aligned too?
Yes, these were the first and easy to spot ones in the cache flush messages, I would swear.
hcor and hccr are register locations though ... so it's pretty weird.
OK, now that I look harder at the messages, doing that is only masking a problem. At issue (from u-boot.map): .bss 0x801586c0 0x1c0 drivers/usb/host/libusb_host.o 0x80158860 hcor And a message of: ERROR: v7_dcache_inval_range - stop address is not aligned - 0x9ffbc860 Using 'bdinfo' to see reloc offset gives us hcor, but _end_ not start means it's something within the anonymous part of the bss and padding hcor out just masks the real problem. I'll resubmit a v4 shortly. Thanks!