GitLab

As of commit d7402513 ("arm64: smp: IPI_CPU_STOP and
IPI_CPU_CRASH_STOP should try for NMI"), if we've got pseudo-NMI
enabled then we'll use it to stop CPUs at panic time. This is nice,
but it does mean that there's a pretty good chance that we'll end up
stopping a CPU while it holds the port lock for the console
UART. Specifically, I see a CPU get stopped while holding the port
lock nearly 100% of the time on my sc7180-trogdor based Chromebook by
enabling the "buddy" hardlockup detector and then doing:

  sysctl -w kernel.hardlockup_all_cpu_backtrace=1
  sysctl -w kernel.hardlockup_panic=1
  echo HARDLOCKUP > /sys/kernel/debug/provoke-crash/DIRECT

UART drivers are _supposed_ to handle this case OK and this is why
UART drivers check "oops_in_progress" and only do a "trylock" in that
case. However, before we enabled pseudo-NMI to stop CPUs it wasn't a
very well-tested situation.

Now that we're testing the situation a lot, it can be seen that the
Qualcomm GENI UART driver is pretty broken. Specifically, when I run
my test case and look at the console output I just see a bunch of
garbled output like:

  [  201.069084] NMI backtrace[  201.069084] NM[  201.069087] CPU: 6
  PID: 10296 Comm: dnsproxyd Not tainted 6.7.0-06265-gb13e8c0ede12
  #1 01112b9f14923cbd0b[  201.069090] Hardware name: Google Lazor
  ([  201.069092] pstate: 80400009 (Nzcv daif +PAN -UAO -TCO -DI[
  201.069095] pc : smp_call_function_man[  201.069099]

That's obviously not so great. This happens because each call to the
console driver exits after the data has been written to the FIFO but
before it's actually been flushed out of the serial port. When we have
multiple calls into the console one after the other then (if we can't
get the lock) each call tells the UART to throw away any data in the
FIFO that hadn't been transferred yet.

I've posted up a patch to change the arm64 core to avoid this
situation most of the time [1] much like x86 seems to do, but even if
that patch lands the GENI driver should still be fixed.

>From testing, it appears that we can just delete the cancel/abort in
the case where we weren't able to get the UART lock and the output
looks good. It makes sense that we'd be able to do this since that
means we'll just call into __qcom_geni_serial_console_write() and
__qcom_geni_serial_console_write() looks much like
qcom_geni_serial_poll_put_char() but with a loop. However, it seems
safest to poll the FIFO and make sure it's empty before our
transfer. This should reliably make sure that we're not
interrupting/clobbering any existing transfers.

As part of this change, we'll also avoid re-setting up a TX at the end
of the console write function if we weren't able to get the lock,
since accessing "port->tx_remaining" without the lock is not
safe. This is only needed to re-start userspace initiated transfers.

[1] https://lore.kernel.org/r/20231207170251.1.Id4817adef610302554b8aa42b090d57270dc119c@changeid

Signed-off-by: Douglas Anderson <dianders@chromium.org>
Link: https://lore.kernel.org/r/20240112150307.2.Idb1553d1d22123c377f31eacb4486432f6c9ac8d@changeid

Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>