On 10/16/21 2:40 AM, Jakub Kicinski wrote:

On Fri, 15 Oct 2021 19:10:27 +0200 Alvin Šipraga wrote:

quoted

v2 -> v3:
   - collect Florian's Reviewed-by
   - move IRQ setup earlier in probe per Florian's suggestion
   - follow Jakub's suggestion and use the standard ethtool stats API

Thanks a lot for doing this. The code LGTM, the only thing that stands
out is the use of spin_locks(). I couldn't quickly parse out what bus
this device hangs off, if it's MMIO and registers can be read without
sleeping you could potentially get rid of the delayed work to read
stats, but I think you need to switch to
regmap_read_poll_timeout_atomic() because regmap_read_poll_timeout()
itself can sleep.

It's an "SMI" bus bit-banged with GPIO - details are in 
realtek-smi-core.c - a Realtek peculiarity. Initially I thought this was 
implementation was sleeping, but actually it's using ndelay() which is 
OK in atomic context if I'm not mistaken? So I guess you're right, I can 
get rid of the delayed work. :)

If the register access sleeps (I2C, SPI, MDIO etc) you need to switch
from a spin lock to a mutex.

Either way CONFIG_DEBUG_ATOMIC_SLEEP is your friend.

Thanks, I'll test with this and follow up. I see a kernel test robot 
warning so I will be sending v4 anyway.