Re: BUG: net-next (7.0-rc6 based and later) fails to boot on Jetson Xavier NX

[Linus Torvalds]

On Wed, 8 Apr 2026 at 09:08, Russell King (Oracle)
<linux@xxxxxxxxxxxxxxx> wrote:
>
> The rebase is still progressing, but it's landed on:
>
> c7d812e33f3e dmaengine: xilinx: xilinx_dma: Fix unmasked residue subtraction

Well, that commit looks completely bogus.

The explanation is just garbage: when subtracting two values that may
have random crud in the top bits, it's actually likely *better* to do
the masking *after* the subtraction.

The subtract of bogus upper bits will only affect upper bits. The
carry-chain only works upwards, not downwards.

So the old code that did

                       residue += (cdma_hw->control - cdma_hw->status) &
                                  chan->xdev->max_buffer_len;

would correctly mask out the upper bits, and the result of the
subtraction would be done "modulo mac_buffer_len". Which is rather
reasonable.

The code was changed to

                       residue += (cdma_hw->control &
chan->xdev->max_buffer_len) -
                                  (cdma_hw->status &
chan->xdev->max_buffer_len);

and now it does obviously still mask out the upper bits on each of the
values), but then the subtraction is done "modulo the arithmetic C
type" (which is 'u32')

In particular, if the status bits are bigger than the control bits,
that residue addition will now add a *huge* 32-bit number. It used to
add a number that was limited by the  max_buffer_len mask.

So the "interference from those top bits" stated in the commit message
is simply NOT TRUE. It's just complete rambling garbage.

Instead, the commit purely changes the final modulus of the
subtraction - which has nothing to do with any upper bits, and
everything to do with what kind of answer you want.

I think that commit is just very very wrong. At least the commit
message is wrong. And see above why I think the changed arithmetic is
likely wrong too.

It's very possible that the 'residue' is now a random 32-bit number
with the high bits set, and you get DMA corruption.

That would explain why this happens on Jetson but I haven't seen other reports.

                    Linus