Re: Softirq error with mcp251xfd driver

On Wed, Mar 10, 2021 at 10:22:54PM +0100, Daniel Glöckner wrote:

On Wed, Mar 10, 2021 at 08:13:51AM +0100, Marc Kleine-Budde wrote:

quoted

On 10.03.2021 07:46:26, Daniel Glöckner wrote:

quoted

the mcp251xfd driver uses a threaded irq handler to queue skbs with the
can_rx_offload_* helpers. I get the following error on every packet until
the rate limit kicks in:

NOHZ tick-stop error: Non-RCU local softirq work is pending, handler
#08!!!

That's a known problem. But I had no time to investigate it.

quoted

Adding local_bh_disable/local_bh_enable around the can_rx_offload_* calls
gets rid of the error, but is that the correct way to fix this?
Internally the can_rx_offload code uses spin_lock_irqsave to safely
manipulate its queue.

The problem is not the queue handling inside of rx_offload, but the call
to napi_schedule(). This boils down to raising a soft IRQ (the NAPI)
from the threaded IRQ handler of the mcp251xfd driver.

The local_bh_enable() "fixes" the problem running the softirq if needed.

https://elixir.bootlin.com/linux/v5.11/source/kernel/softirq.c#L1913

I'm not sure how to properly fix the problem, yet.

If I understand correctly, the point of using can_rx_offload_* in the
mcp251xfd driver is that it sorts the rx, tx, and error frames according
to their timestamp. In that case calling local_bh_enable after each packet
is not correct because there will never be more than one packet in the
queue. We want to call local_bh_disable + can_rx_offload_schedule +
local_bh_enable only at the end of mcp251xfd_irq after intf_pending
indicated that there are no more packets inside the chip. How about adding
a flag to struct can_rx_offload that suppresses the automatic calls to
can_rx_offload_schedule?

If there is the risk that under high load we will never exit the loop in
mcp251xfd_irq or if can_rx_offload_napi_poll might run again while we add
more packets to the queue, a more complex scheme is needed. We could
extend can_rx_offload_napi_poll to process only packets with a timestamp
below a certain value. That value has to be read from the TBC register
before we read the INT register. Then the three functions can be run after
each iteration to empty the queue. We need to update that timestamp limit
one more time when we finally exit the loop to process those packets that
have arrived after the reading of the TBC register when the INT register
still had bits set. Using the timestamp of the tail of the queue is
probably the easiest way to set the final limit.

Or we leave can_rx_offload unchanged and keep two additional lists of skbs
inside the mcp251xfd driver: One for the packets that arrived before the
timestamp read from TBC and one for the packets that arrived later. At the
end of an iteration we call local_bh_disable, enqueue all packets from the
first list with can_rx_offload_queue_sorted, and the ask the softirq to
process them by calling local_bh_enable. Afterwards we move everything
from the second list to the first list and do the next iteration.

The drawback is that we can't use can_rx_offload_get_echo_skb.

Best regards,

  Daniel

-- 
Dipl.-Math. Daniel Glöckner, emlix GmbH, http://www.emlix.com
Fon +49 551 30664-0, Fax +49 551 30664-11,
Gothaer Platz 3, 37083 Göttingen, Germany
Sitz der Gesellschaft: Göttingen, Amtsgericht Göttingen HR B 3160
Geschäftsführung: Heike Jordan, Dr. Uwe Kracke
Ust-IdNr.: DE 205 198 055

emlix - your embedded linux partner