Re: [RFT PATCH 1/4] usb: dwc2: Simplify and fix DMA alignment code

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Hi Doug,

[add Nicolas the new BCM2835 maintainer]

Am 27.02.20 um 23:06 schrieb Doug Anderson:
> Hi,
>
> On Wed, Feb 26, 2020 at 1:04 PM Guenter Roeck <linux@xxxxxxxxxxxx> wrote:
>> The code to align buffers for DMA was first introduced with commit
>> 3bc04e28a030 ("usb: dwc2: host: Get aligned DMA in a more supported way").
>> It was updated with commit 56406e017a88 ("usb: dwc2: Fix DMA alignment
>> to start at allocated boundary") because it did not really align buffers to
>> DMA boundaries but to word offsets. This was then optimized in commit
>> 1e111e885238 ("usb: dwc2: Fix inefficient copy of unaligned buffers")
>> to only copy actual data rather than the whole buffer. Commit 4a4863bf2e79
>> ("usb: dwc2: Fix DMA cache alignment issues") changed this further to add
>> a padding at the end of the buffer to ensure that the old data pointer is
>> not in the same cache line as the buffer.
>>
>> This last commit states "Otherwise, the stored_xfer_buffer gets corrupted
>> for IN URBs on non-cache-coherent systems". However, such corruptions are
>> still observed. This suggests that the commit may have been hiding a
>> problem rather than fixing it. Further analysis shows that this is indeed
>> the case: The code in dwc2_hc_start_transfer() assumes that the transfer
>> size is a multiple of wMaxPacketSize, and rounds up the transfer size
>> communicated to the chip accordingly. Added debug code confirms that
>> the chip does under some circumstances indeed send more data than requested
>> in the urb receive buffer size.
>>
>> On top of that, it turns out that buffers are still not guaranteed to be
>> aligned to dma_get_cache_alignment(), but to DWC2_USB_DMA_ALIGN (4).
>> Further debugging shows that packets aligned to DWC2_USB_DMA_ALIGN
>> but not to dma_get_cache_alignment() are indeed common and work just fine.
>> This suggests that commit 56406e017a88 was not really necessary because
>> even without it packets were already aligned to DWC2_USB_DMA_ALIGN.
>>
>> To simplify the code, move the old data pointer back to the beginning of
>> the new buffer, restoring most of the original commit. Stop aligning the
>> buffer to dma_get_cache_alignment() since it isn't needed and only makes
>> the code more complex. Instead, ensure that the allocated buffer is a
>> multiple of wMaxPacketSize to ensure that the chip does not write beyond
>> the end of the buffer.
> I do like the cleanliness of being able to easily identify the old
> buffer (AKA by putting it first) and I agree that the existing code
> was only really guaranteeing 4-byte alignment and if we truly needed
> more alignment then we'd be allocating a lot more bounce buffers
> (which is pretty expensive).
>
> ...but the argument in commit 56406e017a88 ("usb: dwc2: Fix DMA
> alignment to start at allocated boundary") is still a compelling one.
> Maybe at least put a comment in the code next to the "#define
> DWC2_USB_DMA_ALIGN" saying that we think that this is enough alignment
> for anyone using dwc2's built-in DMA logic?
>
>
>> Cc: Douglas Anderson <dianders@xxxxxxxxxxxx>
>> Cc: Boris Arzur <boris@xxxxxxxxx>
>> Fixes: 56406e017a88 ("usb: dwc2: Fix DMA alignment to start at allocated boundary")
>> Signed-off-by: Guenter Roeck <linux@xxxxxxxxxxxx>
>> ---
>>  drivers/usb/dwc2/hcd.c | 67 ++++++++++++++++++++++--------------------
>>  1 file changed, 35 insertions(+), 32 deletions(-)
> Sorry for such a mess and thank you for all the work tracking down and
> documenting all the problems.  Clearly deep understanding of DMA is
> not something I can claim.  :(
>
> A few points of order first:
> * Although get_maintainer doesn't identify him, it has seemed like
> Felipe Balbi lands most of the dwc2 things.  Probably a good idea to
> CC him.
> * I have historically found Stefan Wahren interested in dwc2 fixes and
> willing to test them on Raspberry Pi w/ various peripherals.

i'm not the BCM2835 maintainer anymore, but will give it a try.

Regards
Stefan




[Index of Archives]     [Linux Media]     [Linux Input]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [Old Linux USB Devel Archive]

  Powered by Linux