On Mon, Mar 1, 2021 at 4:44 PM 谢华伟(此时此刻） [off-list ref] wrote:

quoted

quoted

quoted

What is the downside of using "pause until the I/O completes" versions?

The downside in virtio PMD is a small performance penalty when we use it
to notify backend. CPU executes unnecessary serializing IO instruction.

I check kernel code, io wrapper for in/out doesn't use p version.

This change is a fix/optimisation.
This is a separate topic from adding MMIO support with x86 ioport.
I would split as a separate patch.

Hi David:

Maybe there is confuse? There is no change. The out/in is added. I don't
remove _p on purpose.

Looking at v8 and repeating previous mails:

+#if defined(RTE_ARCH_X86)
...
+static inline void iowrite8(uint8_t val, void *addr)
+{
+    (uint64_t)(uintptr_t)addr >= PIO_MAX ?
+        *(volatile uint8_t *)addr = val :
+        outb(val, (unsigned long)addr); <======
+}

[...]


-#if defined(RTE_ARCH_X86)
-            outb_p(*s, reg); <======
-#else
-            *(volatile uint8_t *)reg = *s;
-#endif
+            iowrite8(*s, (void *)reg);


This almost went unnoticed (thanks Ferruh for spotting).

Do we _need_ this change on outX_p -> outX?

I am not comfortable at touching such low level internal routines that
have been in dpdk since v1.5.0.

If there is a good reason, it has nothing to do with adding MMIO
support and must be split in a separate patch.
If there is no reason, please restore outX_p, since the safest is not to touch.


-- 
David Marchand