Hi, Thomas,

On 01/08/2021 01:26 AM, Thomas Bogendoerfer wrote:

quoted

quoted

quoted

--- a/arch/mips/kernel/relocate_kernel.S
+++ b/arch/mips/kernel/relocate_kernel.S

@@ -6,6 +6,7 @@

   #include <asm/asm.h>
   #include <asm/asmmacro.h>
+#include <asm/cpu.h>
   #include <asm/regdef.h>
   #include <asm/mipsregs.h>
   #include <asm/stackframe.h>

@@ -133,6 +134,33 @@ LEAF(kexec_smp_wait)
   #else
       sync
   #endif
+
+#ifdef CONFIG_CPU_LOONGSON64

Is there a reason why you can't use the already existing infrastructure
the way cavium-octeon is doing it ? If you can't please explain why
so we can find a way to extend it. But having some sort of poking
loongson registers in generic MIPS code is a non starter.

Thomas.

Unlike the cavium-octeon platform, the Loongson64 platform needs some 
changes. Before the kernel starts, (before entering the kernel_entry), 
each CPU has its own state (the SMP system). For Loongson64, only the 
boot CPU will enter the kernel_entry, and other CPUs will query their 
mailbox value in a loop. This is what the BIOS does for the CPU. Here is 
different from cavium-octeon. All CPUs will enter the kernel_entry on 
cavium-octeon platform. Then the kernel_entry_setup, the co-CPUs will 
enter the query loop. I saw the kernel_entry_setup of other platforms, 
such as ip27, malta, and generic. They are not like cavium-octeon and 
the co-CPUs entering the loop may be earlier than entering kernel_entry. 
So I have reason to guess that most SMP system platform CPUs are similar 
to Loongson64.

relocate_kernel.S is like BIOS doing s omething for the CPU. It allows 
the boot CPU to start from the new kernel_entry and makes the co-CPUs 
enter a loop. The already existing infrastructure may be more suitable 
for non-smp platforms. Although we can do something with 
plat_smp_ops.kexec_nonboot_cpu, more new problems will arise in that 
case. The kexec process actually runs on a copy of relocate_kernel.S, 
which will bring a lot of problems...

Above all just my personal thoughts.

Thanks,
Jinyang