On Wed, 2021-11-17 at 14:14 +0200, Jarkko Nikula wrote:

Hi

On 11/16/21 3:58 PM, Matthias Schiffer wrote:

quoted

I just noticed that m_can_pci is completely broken on 5.15.2, while
it's working fine on 5.14.y.

Hmm.. so that may explain why I once saw candump received just zeroes on 
v5.15-rc something but earlier kernels were ok. What's odd then next 
time v5.15-rc was ok so went blaming sun spots instead of bisecting.

quoted

I assume something simliar to [1] will be necessary in m_can_pci as
well, however I'm not really familiar with the driver. There is no
"mram_base" in m_can_plat_pci, only "base". Is using "base" with
iowrite32/ioread32 + manual increment the correct solution here?


[1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=99d173fbe8944861a00ebd1c73817a1260d21e60

If your test case after 5.15 reliably fails are you able to bisect or 
check does the regression originate from the same commit?

Jarkko

The Fixes tag of 99d173fbe894 ("can: m_can: fix iomap_read_fifo() and
iomap_write_fifo()") is off AFAICT, the actual breakage happened with
812270e5445b ("can: m_can: Batch FIFO writes during CAN transmit") and
1aa6772f64b4 ("can: m_can: Batch FIFO reads during CAN receive");
reverting these two patches fixes the regression.

I just sent a patch for m_can_pci that applies the same fix that was
done in m_can_platform in 99d173fbe894, and verified that this makes
CAN communication work on Elkhart Lake with Linux 5.15.y together with
my other patches.

Matthias