[U-Boot] usb: gadget: Fix data aborts during USB ethernet boot

On 09/04/2013 04:38 PM, Marek Vasut wrote:
> Dear Joel Fernandes,
> 
>> On 08/27/2013 08:52 AM, Marek Vasut wrote:
>>> Dear Joel Fernandes,
>>>
>>>> As seen on GCC 4.6 Linaro compiler, control_req buffer is not aligned
>>>> on 4 byte boundaray causing data aborts in eth_setup -> conf_buf
>>>> during dhcp boot over usb_ether. Fix the issue my aligning control_req
>>>> buffer to 4-byte boundary.
>>>>
>>>> Tested on am335x_evm platform (beaglebone).
>>>> Applies on 2013.10-rc1 branch.
>>>>
>>>> Cc: Tom Rini <trini@ti.com>
>>>> Signed-off-by: Joel Fernandes <joelf@ti.com>
>>>
>>> Please keep me in the CC next time.
>>
>> Ok.
>>
>>>> ---
>>>>
>>>>  drivers/usb/gadget/ether.c |    2 +-
>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/usb/gadget/ether.c b/drivers/usb/gadget/ether.c
>>>> index 579893c..251d7b2 100644
>>>> --- a/drivers/usb/gadget/ether.c
>>>> +++ b/drivers/usb/gadget/ether.c
>>>> @@ -849,7 +849,7 @@ static struct usb_gadget_strings	stringtab = {
>>>>
>>>>  };
>>>>  
>>>>  /*=====================================================================
>>>>  ===
>>>>
>>>> ====*/ -static u8 control_req[USB_BUFSIZ];
>>>> +static u8 control_req[USB_BUFSIZ] __attribute__ ((aligned(4)));
>>>
>>> Please make this cacheline aligned, so we get rid of bounce buffering of
>>> the requests on stupid hardware.
>>>
>>>>  #if defined(CONFIG_USB_ETH_CDC) || defined(CONFIG_USB_ETH_RNDIS)
>>>>  static u8 status_req[STATUS_BYTECOUNT] __attribute__ ((aligned(4)));
>>>
>>> This could also use fixing, but the STATUS_BYTECOUNT would need to be
>>> up-aligned as well then. Some DEFINE_CACHE_ALIGN_BUFFER might help in
>>> both cases here.
>>
>> Ok, how about __aligned(CONFIG_SYS_CACHELINE_SIZE) instead of defining
>> a new DEFINE_CACHE_ALIGN_BUFFER macro?
> 
> THis is already defined and used throughout the USB code, check include/common.h 
> IIRC.
> 
>> Can you explain what you meant by STATUS_BYTECOUNT up-aligned and why its
>> needed to be up-aligned?
> 
> it's 16 bytes now, no? Up-align it to cacheline size.

Ok, that makes sense. Thanks. Hope below patch is OK, but I have to test it more
once I am near a setup and can send out a proper patch later today.

If its ok can you explain technical reason for need to up-align array size to
cacheline size, and why aligning only buffer location is not enough?

Also I saw the following comment before the macro definition:
* Usage of this macro shall be avoided or used with extreme care!

Regards,

-Joel

Dear Joel Fernandes,

> On 09/04/2013 04:38 PM, Marek Vasut wrote:
> > Dear Joel Fernandes,
> > 
> >> On 08/27/2013 08:52 AM, Marek Vasut wrote:
> >>> Dear Joel Fernandes,
> >>> 
> >>>> As seen on GCC 4.6 Linaro compiler, control_req buffer is not aligned
> >>>> on 4 byte boundaray causing data aborts in eth_setup -> conf_buf
> >>>> during dhcp boot over usb_ether. Fix the issue my aligning control_req
> >>>> buffer to 4-byte boundary.
> >>>> 
> >>>> Tested on am335x_evm platform (beaglebone).
> >>>> Applies on 2013.10-rc1 branch.
> >>>> 
> >>>> Cc: Tom Rini <trini@ti.com>
> >>>> Signed-off-by: Joel Fernandes <joelf@ti.com>
> >>> 
> >>> Please keep me in the CC next time.
> >> 
> >> Ok.
> >> 
> >>>> ---
> >>>> 
> >>>>  drivers/usb/gadget/ether.c |    2 +-
> >>>>  1 file changed, 1 insertion(+), 1 deletion(-)
> >>>> 
> >>>> diff --git a/drivers/usb/gadget/ether.c b/drivers/usb/gadget/ether.c
> >>>> index 579893c..251d7b2 100644
> >>>> --- a/drivers/usb/gadget/ether.c
> >>>> +++ b/drivers/usb/gadget/ether.c
> >>>> @@ -849,7 +849,7 @@ static struct usb_gadget_strings	stringtab = {
> >>>> 
> >>>>  };
> >>>>  
> >>>>  /*===================================================================
> >>>>  == ===
> >>>> 
> >>>> ====*/ -static u8 control_req[USB_BUFSIZ];
> >>>> +static u8 control_req[USB_BUFSIZ] __attribute__ ((aligned(4)));
> >>> 
> >>> Please make this cacheline aligned, so we get rid of bounce buffering
> >>> of the requests on stupid hardware.
> >>> 
> >>>>  #if defined(CONFIG_USB_ETH_CDC) || defined(CONFIG_USB_ETH_RNDIS)
> >>>>  static u8 status_req[STATUS_BYTECOUNT] __attribute__ ((aligned(4)));
> >>> 
> >>> This could also use fixing, but the STATUS_BYTECOUNT would need to be
> >>> up-aligned as well then. Some DEFINE_CACHE_ALIGN_BUFFER might help in
> >>> both cases here.
> >> 
> >> Ok, how about __aligned(CONFIG_SYS_CACHELINE_SIZE) instead of defining
> >> a new DEFINE_CACHE_ALIGN_BUFFER macro?
> > 
> > THis is already defined and used throughout the USB code, check
> > include/common.h IIRC.
> > 
> >> Can you explain what you meant by STATUS_BYTECOUNT up-aligned and why
> >> its needed to be up-aligned?
> > 
> > it's 16 bytes now, no? Up-align it to cacheline size.
> 
> Ok, that makes sense. Thanks. Hope below patch is OK, but I have to test it
> more once I am near a setup and can send out a proper patch later today.
> 
> If its ok can you explain technical reason for need to up-align array size
> to cacheline size, and why aligning only buffer location is not enough?
> 
> Also I saw the following comment before the macro definition:
> * Usage of this macro shall be avoided or used with extreme care!
> 
> Regards,
> 
> -Joel
> 
> diff --git a/drivers/usb/gadget/ether.c b/drivers/usb/gadget/ether.c
> index 579893c..700d5fb 100644
> --- a/drivers/usb/gadget/ether.c
> +++ b/drivers/usb/gadget/ether.c
> @@ -849,9 +849,10 @@ static struct usb_gadget_strings	stringtab = {
>  };
> 
>  /*========================================================================
> ====*/ -static u8 control_req[USB_BUFSIZ];
> +DEFINE_CACHE_ALIGN_BUFFER(u8, control_req, USB_BUFSIZ);
> +
>  #if defined(CONFIG_USB_ETH_CDC) || defined(CONFIG_USB_ETH_RNDIS)
> -static u8 status_req[STATUS_BYTECOUNT] __attribute__ ((aligned(4)));
> +DEFINE_CACHE_ALIGN_BUFFER(u8, status_req, STATUS_BYTECOUNT);
>  #endif

Yeah, I think so. Repost V2 and we should be cool here.

Best regards,
Marek Vasut