Public bug reported:

SRU Justification:
==================

[Impact]

On a Raspberry Pi 5 running a realtime kernel, RP1 IRQ threads do not
follow smp_affinity. This means that IRQ threads will always remain
running on the CPU they were initially scheduled on, and cannot be
moved. This also means that IRQ threads can't be moved off of isolated
CPUs, or onto isolated CPUs as desired, which may introduce significant
variability in latencies for some workloads.

The upstream bug link may also have relevant context:
https://github.com/raspberrypi/linux/issues/7301

[Fix]
Cherry-pick commits b6e5abe492944af09154b6ba65d8db5c4b87accb and 
19dec222ead344334230f57732f994c77af6bb6f from upstream rpi-7.0.y. These commits 
were introduced in https://github.com/raspberrypi/linux/pull/7302.

[Test Case]

1. Select an RP1 IRQ thread, and check the CPU it's running on

  ps -eo psr,comm | grep irq/108
  1 irq/108-eth%d # IRQ 108 running on CPU 1

2. Compare the thread's smp_affinity and effective_affinity, and note that 
effective_affinity is not set
  cat /proc/irq/108/smp_affinity_list
  0-3

 cat /proc/irq/108/effective_affinity_list

# The line above represents empty output from cat. We expect to also see
0-3 here on a fixed kernel.

3. Update the thread's smp_affinity to attempt to move it onto another CPU
  echo "3" > /proc/irq/108/smp_affinity_list
  cat /proc/irq/108/effective_affinity_list

# Output of cat is once again empty. We expect to see 3 here on a fixed
kernel.

  ps -eo psr,comm | grep irq/108
  1 irq/108-eth%d # IRQ 108 is still running on CPU 1, but we expect CPU 3 here 
on a fixed kernel.

[Regression Potential]
The cherry-picked fixes from upstream should only impact the RP1 driver for the 
Raspberry Pi 5, and as far as I can tell there have been no reported issues 
upstream with these fixes so I believe the regression potential to be low.

** Affects: linux-raspi (Ubuntu)
     Importance: Medium
     Assignee: April Cardenas (apcaus)
         Status: In Progress

** Affects: linux-raspi (Ubuntu Resolute)
     Importance: Medium
     Assignee: April Cardenas (apcaus)
         Status: In Progress

** Affects: linux-raspi (Ubuntu Stonking)
     Importance: Undecided
         Status: Fix Released

** Changed in: linux-raspi (Ubuntu)
    Milestone: resolute-updates => None

** Description changed:

  SRU Justification:
  ==================
  
  [Impact]
  
  On a Raspberry Pi 5 running a realtime kernel, RP1 IRQ threads do not
  follow smp_affinity. This means that IRQ threads will always remain
  running on the CPU they were initially scheduled on, and cannot be
  moved. This also means that IRQ threads can't be moved off of isolated
  CPUs, or onto isolated CPUs as desired, which may introduce significant
  variability in latencies for some workloads.
  
  The upstream bug link may also have relevant context:
  https://github.com/raspberrypi/linux/issues/7301
  
  [Fix]
  Cherry-pick commits b6e5abe492944af09154b6ba65d8db5c4b87accb and 
19dec222ead344334230f57732f994c77af6bb6f from upstream rpi-7.0.y. These commits 
were introduced in https://github.com/raspberrypi/linux/pull/7302.
  
  [Test Case]
  
  1. Select an RP1 IRQ thread, and check the CPU it's running on
  
-   ps -eo psr,comm | grep irq/108
-   1 irq/108-eth%d # IRQ 108 running on CPU 1
+   ps -eo psr,comm | grep irq/108
+   1 irq/108-eth%d # IRQ 108 running on CPU 1
  
  2. Compare the thread's smp_affinity and effective_affinity, and note that 
effective_affinity is not set
-   cat /proc/irq/108/smp_affinity_list
-   0-3
+   cat /proc/irq/108/smp_affinity_list
+   0-3
  
-  cat /proc/irq/108/effective_affinity_list
+  cat /proc/irq/108/effective_affinity_list
  
  # The line above represents empty output from cat. We expect to also see
  0-3 here on a fixed kernel.
  
  3. Update the thread's smp_affinity to attempt to move it onto another CPU
-   echo "3" > /proc/irq/108/smp_affinity_list
-   cat /proc/irq/108/effective_affinity_list 
+   echo "3" > /proc/irq/108/smp_affinity_list
+   cat /proc/irq/108/effective_affinity_list
  
  # Output of cat is once again empty. We expect to see 3 here on a fixed
  kernel.
  
-   ps -eo psr,comm | grep irq/108
-   1 irq/108-eth%d # IRQ 108 is still running on CPU 1, but we expect CPU 3 
here on a fixed kernel.
-   
- 
+   ps -eo psr,comm | grep irq/108
+   1 irq/108-eth%d # IRQ 108 is still running on CPU 1, but we expect CPU 3 
here on a fixed kernel.
  
  [Regression Potential]
  The cherry-picked fixes from upstream should only impact the RP1 driver for 
the Raspberry Pi 5, and as far as I can tell there have been no reported issues 
upstream with these fixes so I believe the regression potential to be low.
- 
- 
- [Other]
- 
- #

** Also affects: linux-raspi (Ubuntu Resolute)
   Importance: Undecided
       Status: New

** Changed in: linux-raspi (Ubuntu Resolute)
       Status: New => In Progress

-- 
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2173779

Title:
  PREEMPT_RT: RP1 IRQ threads do not follow smp_affinity

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/linux-raspi/+bug/2173779/+subscriptions


-- 
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs

Reply via email to