On Fri, May 21, 2021 at 07:15:21PM +0800, Zheyu Ma wrote:

On Fri, May 21, 2021 at 2:51 PM Greg KH [off-list ref] wrote:

quoted

On Fri, May 21, 2021 at 06:08:43AM +0000, Zheyu Ma wrote:

quoted

In 'rp2_probe', the driver registers 'rp2_uart_interrupt' then calls
'rp2_fw_cb' through 'request_firmware_nowait'. In 'rp2_fw_cb', if the
firmware don't exists, function just return without initializing ports
of 'rp2_card'. But now the interrupt handler function has been
registered, and when an interrupt comes, 'rp2_uart_interrupt' may access
those ports then causing NULL pointer dereference or other bugs.

Because the driver does some initialization work in 'rp2_fw_cb', in
order to make the driver ready to handle interrupts, 'request_firmware'
should be used instead of asynchronous 'request_firmware_nowait'.

You just now slowed down the probe function.  Are you _sure_ this is ok?

Sorry, I'm not an expert in the field, but from my point of view, the
previous function 'rp2_fw_cb' does some initialization work that is
not suitable for asynchronous execution. Because after these initial
work, the driver can work normally (including preparing to handle
interrupts).

quoted

Do you have this hardware to test this?  If so, what is the init time
before and after this change?

To be honest, I don't have real hardware, I tested it with QEMU. I
made a total of 5 attempts. Before this change, the average boot time
required by kernel is 6.382s, the time required for insmoding this
module is 0.139s; After this change, the average boot time required by
kernel is 6.426s, the time required for insmoding this module is
0.160s. This change really slowed down the probe function.

Ok, at least you've tested it :)

I'll take a slower init over a broken init any day.  I'll go queue this
up, thanks.

greg k-h