Ok, thanks for your explaination, I got it now.
I will document "posted write" info in the top of the file or top of
startup/shutdown function.

And kernel test robot report me build error and warning with gcc
11.2.0 ia64 / powerpc.
I will fix it and send next patch.

Greg KH [off-list ref] 於 2022年2月8日 週二 下午7:31寫道：

On Tue, Feb 08, 2022 at 07:16:52PM +0800, hammer hsieh wrote:

quoted

Jiri Slaby [off-list ref] 於 2022年2月8日 週二 下午2:27寫道：

quoted

Hi,

On 07. 02. 22, 6:58, Hammer Hsieh wrote:

quoted

+static void sunplus_shutdown(struct uart_port *port)
+{
+     unsigned long flags;
+     unsigned int isc;
+
+     spin_lock_irqsave(&port->lock, flags);
+
+     isc = readl(port->membase + SUP_UART_ISC);
+     isc &= ~(SUP_UART_ISC_RXM | SUP_UART_ISC_TXM);

Is this correct? I mean: will the SUP_UART_ISC read contain the control
bits, not only status bits?

I assume reviewers don't like writel(0,xxx).
So I use definition to let the code easy to read.
The purpose is to clear all interrupt.
Bit[3:0] status bit only for read, write 1 or 0 no effect.

quoted

quoted

+     writel(isc, port->membase + SUP_UART_ISC);
+
+     spin_unlock_irqrestore(&port->lock, flags);
+
+     free_irq(port->irq, port);

I am still waiting for explanation why this is safe with respect to
posted writes.

Actually I'm not IC designer, not expert for bus design.
About data incoherence issue between memory bus and peripheral bus.
In case of AXI bus, use non-posted write can avoid data incoherence issue.
What if in case of posted write:
Send a specific command after last write command.
SDCTRL identify specific command, means previous write command done.
Then send interrupt signal to interrupt controller.
And then interrupt controller send done signal to Master.
Master receive done signal, means write command done.
Then issue a interrupt or proceed next write command.

But how does the kernel know when the write is completed?  The kernel
seems to ignore that here entirely, so the write could actually complete
seconds later, which would not be a good thing, right?

Traditionally, we want to ensure that a write() completes, so on some
busses, we have to do a read to ensure that the write made it to the
hardware before we can continue on.  That is not happening here which is
why Jiri keeps bringing it up.  It looks broken to us, and you need to
document it somewhere (in the changelog?  In the top of the file?) as to
why this is not needed.

thanks,

greg k-h