On 07 Jun 2017, at 09:53, Marek Vasut marex@denx.de wrote:
On 06/07/2017 09:49 AM, Dr. Philipp Tomsich wrote:
On 07 Jun 2017, at 08:28, Marek Vasut <marex@denx.de mailto:marex@denx.de> wrote:
On 06/06/2017 05:11 PM, Dr. Philipp Tomsich wrote:
On 06 Jun 2017, at 16:35, Marek Vasut <marex@denx.de mailto:marex@denx.de> wrote:
On 06/06/2017 03:42 PM, Philipp Tomsich wrote:
For LP64 build w/ GCC 6.3, the handling of the ep->dma_buf variable when (truncating and) writing into the controller's registers is unsafe and triggers the following compiler warning (thus failing by buildman tests): ../drivers/usb/gadget/dwc2_udc_otg_xfer_dma.c: In function 'setdma_rx': ../drivers/usb/gadget/dwc2_udc_otg_xfer_dma.c:116:9: warning: cast from pointer to integer of different size [-Wpointer-to-int-cast] writel((unsigned int) ep->dma_buf, ®->out_endp[ep_num].doepdma); ^
This change fixes the warning and makes the code a bit more robust by doing the following:
- we check whether ep->dma_buf can be safely truncated to 32 bits
(i.e. whether dma_buf is the same value when masked to 32 bits) and emit an error message if that is not the case
- we cast to a uintptr_t and let the writel macro truncate the
uintptr_t (potentially a 64bit value) to 32bits
Signed-off-by: Philipp Tomsich <philipp.tomsich@theobroma-systems.com mailto:philipp.tomsich@theobroma-systems.com>
Changes in v2:
- (new patch) fix warnings and add some robustness for the truncation
of ep->dma_buf in dwc2-otg for LP64 builds to fix a buildman failure for u-boot-rockchip/master@2b19b2f
drivers/usb/gadget/dwc2_udc_otg_xfer_dma.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)
diff --git a/drivers/usb/gadget/dwc2_udc_otg_xfer_dma.c b/drivers/usb/gadget/dwc2_udc_otg_xfer_dma.c index 0d6d2fb..f5c926c 100644 --- a/drivers/usb/gadget/dwc2_udc_otg_xfer_dma.c +++ b/drivers/usb/gadget/dwc2_udc_otg_xfer_dma.c @@ -113,7 +113,11 @@ static int setdma_rx(struct dwc2_ep *ep, struct dwc2_request *req) invalidate_dcache_range((unsigned long) ep->dma_buf, (unsigned long) ep->dma_buf + ep->len);
-writel((unsigned int) ep->dma_buf, ®->out_endp[ep_num].doepdma); +/* ensure that ep->dma_buf fits into 32 bits */ +if (((uintptr_t)ep->dma_buf & 0xffffffff) != (uintptr_t)ep->dma_buf) +error("ep->dma_buf does not fit into a 32 bits");
You print an error, but still continue with the function ? That'll probably lead to a crash, you should rather abort. In fact, I suspect you can use bounce-buffer to assure the dma buffer is in 32bit space.
This tries to stay close to the current behaviour (we have no hardware to test this, but I wanted to get it out of the way as it caused buildman-failures) and to provide diagnostics for whoever first encounters an issue here.
Note that I would only expect a DMA error, if someone really hit this issue ...
So this patch is just some hypothetical fix ? What triggered creation of this patch ?
Running buildman (w/ a GCC 6.3 configured as aarch-unknown-elf) against unrelated patch-series (for a Rockchip device that has a different USB controller) fails due to the warnings raised from this. Just casting those warnings away (without having an “assertion-like” diagnostic printed) doesn’t sound like the best plan either, though.
Details of the buildman-failure are in the above commit-message.
Ah, hm. In that case, you should really implement the bounce buffer here I think. Letting the transfer continue will lead to really weird failures and/or memory corruption.
Alright. I’ll put it to (the back of) my queue and resubmit. Someone (with hardware) will need to test, once it’s submitted.
Cheers, Philipp.