mbox series

[RFC,v2,0/4] Fix force_irqthread + fast triggered edge-type IRQs

Message ID 20240216075948.131372-2-leobras@redhat.com
Headers
Series Fix force_irqthread + fast triggered edge-type IRQs |

Message

Leonardo Bras Soares Passos Feb. 16, 2024, 7:59 a.m. UTC 
  v1 patchset subject was "Fix serial console for PREEMPT_RT".

While dealing with a bug that breaks the serial console from qemu (8250)
after printing a lot of data, I found some issues on this driver on 
PREEMPT_RT kernel due to it enabling force_irqthreads, ending up causing 
the IRQ to be disabled and the serial console to go down.

I found out 8250 driver get an IRQ request for every tx byte, but the
handler is able to deal with sending multiple bytes per "thread wake up".

If the thread don't get to run that often, and end up handling many irqs at
once, it will cause few changes in threads_handled to be noticed, meaning 
only a few runs of note_interrupt will not increment irqs_unhandled, which 
leads to disabling the IRQ. 

For serial8250, to trigger IRQ disabling it's only needed a fast printing 
of ~300kBytes.

In order to fix this, I propose that we subtract the number of threads 
handled since the last note_interrupt from irqs_unhandled:
(irqs_unhandled -= threads_handled - threads_handled_last)

So this way irqs_unhandled actually reflects how many of those irqs 
actually haven't been handled.

This work is divided in:
Patch #1: Simple change in SPURIOUS_DEFERRED bit (31 -> 0), so we can 
acuratelly measure how many IRQs have been handled even when the 
threads_handled & threads_handled_last are close to overflow.

Patch #2: Subtract the diff from irqs_unhandled on note_interrupt.

Patch #3: Implement IRQ_HANDLED_MANY in order to let the handler return how 
many interruptions have been handled in a single run, and increment 
threads_handled accordingly.

Patch #4: Change serial8250 driver to use IRQ_HANDLED_MANY interface, 
so it can inform how many IRQs have been handled, and help set the correct 
number of irqs_unhandled, thus avoiding IRQ disabled for normal usage.

Changes since RFCv1:
- Implemented a way of better counting threaded_irqs handled, instead of 
  only zeroing irqs_unhandled.
- Rebased on top of linux-rt/v6.8-rc4-rt4-rebase, so I don't need to fix 
  the sleep while atomic issue in serial8250 mainline.
- Better description of what we are actually fixing
- Changed patchset title from "Fix serial console for PREEMPT_RT"
- Link: https://lore.kernel.org/all/20240116073701.2356171-1-leobras@redhat.com/

Leonardo Bras (4):
  irq: Move spurious_deferred bit from BIT(31) to BIT(0)
  irq/spurious: Account for multiple handles in note_interrupt
  irq: Introduce IRQ_HANDLED_MANY
  tty/serial8250: Make use of IRQ_HANDLED_MANY interface

 include/linux/irqdesc.h             | 11 +++++-
 include/linux/irqreturn.h           | 23 +++++++++++-
 include/linux/serial_8250.h         |  2 +-
 drivers/tty/serial/8250/8250_core.c | 13 ++++---
 drivers/tty/serial/8250/8250_port.c | 16 +++++----
 kernel/irq/chip.c                   | 10 ++++--
 kernel/irq/handle.c                 |  3 ++
 kernel/irq/manage.c                 |  8 +++--
 kernel/irq/spurious.c               | 55 ++++++++++++++++++-----------
 9 files changed, 102 insertions(+), 39 deletions(-)


base-commit: 63d966ad6fecb66769e56fe2285de1e3b448f2ff