[PATCH v2] serial: port: Don't block system suspend even if bytes are left to xmit

From: Douglas Anderson
Date: Thu May 30 2024 - 11:49:15 EST


Recently, suspend testing on sc7180-trogdor based devices has started
to sometimes fail with messages like this:

port a88000.serial:0.0: PM: calling pm_runtime_force_suspend+0x0/0xf8 @ 28934, parent: a88000.serial:0
port a88000.serial:0.0: PM: dpm_run_callback(): pm_runtime_force_suspend+0x0/0xf8 returns -16
port a88000.serial:0.0: PM: pm_runtime_force_suspend+0x0/0xf8 returned -16 after 33 usecs
port a88000.serial:0.0: PM: failed to suspend: error -16

I could reproduce these problems by logging in via an agetty on the
debug serial port (which was _not_ used for kernel console) and
running:
cat /var/log/messages
..and then (via an SSH session) forcing a few suspend/resume cycles.

Tracing through the code and doing some printf()-based debugging shows
that the -16 (-EBUSY) comes from the recently added
serial_port_runtime_suspend().

The idea of the serial_port_runtime_suspend() function is to prevent
the port from being _runtime_ suspended if it still has bytes left to
transmit. Having bytes left to transmit isn't a reason to block
_system_ suspend, though. If a serdev device in the kernel needs to
block system suspend it should block its own suspend and it can use
serdev_device_wait_until_sent() to ensure bytes are sent.

The DEFINE_RUNTIME_DEV_PM_OPS() used by the serial_port code means
that the system suspend function will be pm_runtime_force_suspend().
In pm_runtime_force_suspend() we can see that before calling the
runtime suspend function we'll call pm_runtime_disable(). This should
be a reliable way to detect that we're called from system suspend and
that we shouldn't look for busyness.

Fixes: 43066e32227e ("serial: port: Don't suspend if the port is still busy")
Signed-off-by: Douglas Anderson <dianders@xxxxxxxxxxxx>
---
In v1 [1] this was part of a 2-patch series. I'm now just sending this
patch on its own since the Qualcomm GENI serial driver has ended up
having a whole pile of problems that are taking a while to unravel.
It makes sense to disconnect the two efforts. The core problem fixed
by this patch and the geni problems never had any dependencies anyway.

[1] https://lore.kernel.org/r/20240523162207.1.I2395e66cf70c6e67d774c56943825c289b9c13e4@changeid/

Changes in v2:
- Fix "regulator" => "regular" in comment.
- Fix "PM Runtime" => "runtime PM" in comment.
- Commit messages says how serdev devices should ensure bytes xfered.

drivers/tty/serial/serial_port.c | 10 ++++++++++
1 file changed, 10 insertions(+)

diff --git a/drivers/tty/serial/serial_port.c b/drivers/tty/serial/serial_port.c
index 91a338d3cb34..93ca94426162 100644
--- a/drivers/tty/serial/serial_port.c
+++ b/drivers/tty/serial/serial_port.c
@@ -64,6 +64,16 @@ static int serial_port_runtime_suspend(struct device *dev)
if (port->flags & UPF_DEAD)
return 0;

+ /*
+ * We only want to check the busyness of the port if runtime PM is
+ * enabled. Specifically runtime PM will be disabled by
+ * pm_runtime_force_suspend() during system suspend and we don't want
+ * to block system suspend even if there is data still left to
+ * transmit. We only want to block regular runtime PM transitions.
+ */
+ if (!pm_runtime_enabled(dev))
+ return 0;
+
uart_port_lock_irqsave(port, &flags);
if (!port_dev->tx_enabled) {
uart_port_unlock_irqrestore(port, flags);
--
2.45.1.288.g0e0cd299f1-goog