Hi Weifeng, On 5/2/24 6:02 AM, Weifeng Liu wrote: > Greetings, > > This series is intended to remedy a race condition where surface > aggregator module (SAM) which is a serdev driver could fail to probe if > the underlying UART port is not ready to open. In such circumstance, > invoking serdev_device_open() gets errno -ENXIO, leading to failure in > probing of SAM. However, if the probe is retried in a short delay, > serdev_device_open() would work as expected and everything just goes > fine. As a workaround, adding the serial driver (8250_dw) into > initramfs or building it into the kernel image significantly mitigates > the likelihood of encountering this race condition, as in this way the > serial device would be initialized much earlier than probing of SAM. > > However, IMO we should reliably avoid this sort of race condition. A > good way is returning -EPROBE_DEFER when serdev_device_open returns > -ENXIO so that the kernel will be able to retry the probing later. This > is what the first patch tries to do. > > Though this solution might be a good enough solution for this specific > issue, I am wondering why this kind of race condition could ever happen, > i.e., why a serdes device could be not ready yet at the moment the > serdev driver gets called and tries to bind it. And even if this is an > expected behavior how serdev driver works, I do feel it a little bit > weird that we need to identify serdev_device_open() returning -ENXIO as > non-fatal error and thus return -EPROBE_DEFER manually in such case, as > I don't see this sort of behavior in other serdev driver. Maximilian > and Hans suggested discussing the root cause of the race condition here. > I will be grateful if you could provide some reasoning and insights on > this. Ack, I have no objection against the changes and if Maximilian is ok with it I can merge these right away as an interim fix, but I would really like to know why the serdev core / tty code is registering a serdev device for a serial port before it is ready to have serdev_device_open() called on it. To me it seems that the root cause is in somewhere in the 8250_dw code or the serdev core code. Resources sometimes not being ready is sometimes which drivers generally speaking need to handle, but in this case the resource which is not ready is the device the driver is binding to, so it seems that the device is registered too soon. If someone familiar with the serial / serdev code can provide some insight here that would be great. Regards, Hans > > Following is the code path when the issue occurs: > > ssam_serial_hub_probe() > serdev_device_open() > ctrl->ops->open() /* this callback being ttyport_open() */ > tty->ops->open() /* this callback being uart_open() */ > tty_port_open() > port->ops->activate() /* this callback being uart_port_activate() */ > Find bit UPF_DEAD is set in uport->flags and fail with errno -ENXIO. > > I notice that flag UPF_DEAD would be set in serial_core_register_port() > during calling serial_core_add_one_port() but don't have much idea > what's going on under the hood. > > Additionally, add logs to the probe procedure of SAM in the second > patch, hoping it could help provide context to user when something > miserable happens. > > Context of this series is available in [1]. > > Best regards, > Weifeng > > [1] https://github.com/linux-surface/kernel/pull/152 > > Weifeng Liu (2): > platform/surface: aggregator: Defer probing when serdev is not ready > platform/surface: aggregator: Log critical errors during SAM probing > > drivers/platform/surface/aggregator/core.c | 39 ++++++++++++++++++---- > 1 file changed, 32 insertions(+), 7 deletions(-) >
