Hi Angelo,
Thanks for your patch!
On Thu, Jun 25, 2020 at 1:28 AM Angelo Dureghello
<angelo.dureghello@xxxxxxxxxxx> wrote:
On mcf5441x, mmu enabled, using a small initramfs, the boot was
recently hanging silently in the initial phase, inside
arch/m68k/mm/init.c, m68k_setup_node()
initrd at 0x47d2f000:0x47d82f64
overlap at 1073741889 for chunk 0
overlap at 1073746176 for chunk 0
overlap at 1073746736 for chunk 0
overlap at 1073746737 for chunk 0
overlap at 1073746738 for chunk 0
overlap at 1073746739 for chunk 0
overlap at 1073746740 for chunk 0
From some debugging, at m68k_setup_node(), the 25-bits shift value
(virt_to_node_shift) seems unset, because applied to a wrong
address (address previously set from m68k_fixup).
Tried some fixes, but the more low-risk and hopefully correct seems
to just add a volatile inside __virt_to_node_shift(). This seems to
produce correct addresses for the next fixup to work. Cannot test this
on other mmu ColdFire cpu's so every test is eventually welcome.
Signed-off-by: Angelo Dureghello <angelo.dureghello@xxxxxxxxxxx>
--- a/arch/m68k/include/asm/page_mm.h
+++ b/arch/m68k/include/asm/page_mm.h
@@ -135,7 +135,7 @@ static inline __attribute_const__ int __virt_to_node_shift(void)
{
int shift;
- asm (
+ asm volatile (
"1: moveq #0,%0\n"
m68k_fixup(%c1, 1b)
: "=d" (shift)
Perhaps we should add the volatile to the other asm statements, too?
Andreas?
Which compiler version are you using?
Have you compared the difference in generated code?
With gcc version 8.4.0 (Ubuntu 8.4.0-1ubuntu1~18.04), the above makes
no difference for m68k_setup_node(), and the compiler doesn't even
optimize away the call to __virt_to_node_shift(), which is marked
__attribute_const__.
Gr{oetje,eeting}s,
Geert
--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx
In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds
