undefined instruction d5380001 (arm64 mrs emulation)

On Mon, Oct 02, 2017 at 03:11:18PM +0100, James Morse wrote:

Hi Matwey,

On 02/10/17 12:24, Dave Martin wrote:

quoted

On Fri, Sep 29, 2017 at 10:23:54PM +0300, Matwey V. Kornilov wrote:

quoted

I am running 4.13.3 on rockchip 3328 platform(aarch64) with glibc 2.26
and see the following at booting:

[   11.152061] modprobe[93]: undefined instruction: pc=0000ffff8ca48ff4
[   11.152707] Code: d503201f 8a180320 92750001 365ffc20 (d5380001)
[   11.154347] modprobe[94]: undefined instruction: pc=0000ffff94243ff4
[   11.154991] Code: d503201f 8a180320 92750001 365ffc20 (d5380001)
[   11.157070] modprobe[97]: undefined instruction: pc=0000ffff839a0ff4
[   11.157715] Code: d503201f 8a180320 92750001 365ffc20 (d5380001)
[   11.159265] modprobe[98]: undefined instruction: pc=0000ffffb0591ff4
[   11.159908] Code: d503201f 8a180320 92750001 365ffc20 (d5380001)

As far as I understand d5380001 should be emulated in cpufeature.c but
it is not. What could be wrong here?

The whole sequence is

   0:   d503201f        nop
   4:   8a180320        and     x0, x25, x24
   8:   92750001        and     x1, x0, #0x800
   c:   365ffc20        tbz     w0, #11, 0xffffffffffffff90
  10:*  d5380001        mrs     x1, midr_el1            <-- trapping instruction

This looks the same as:
https://bugzilla.redhat.com/show_bug.cgi?id=1496209

[...]

quoted

What should happen here is that the do_undefinstr() in
arch/arm64/kernel/traps.c should call registered undef hooks until it
finds one that accepts the faulting instruction.

So, either the cpufeatures undef hook is not getting called, or it is
failing the instruction somewhere, possibly in
cpufeatures.c:emulate_id_reg() or emulate_sys_reg().


Can you add some trace to those functions to see what's happening?

I couldn't reproduce this with linux-stable's v4.13.3 defconfig on Seattle or Juno.

What distribution are you running? Could you also try [0] to see if this is
something specific to your version of modprobe?

It is worth noting that we register the MRS instruction handler as late_init call.
Now the question is how late that could be. Given that we are hitting it with
modprobe, which could be used for requesting modules from initrd. Also which explains
why it we can't reproduce it by simple testcases, after it was registered.

Now the question is, how early do we want to push this. Since it doesn't depend really
on any other subsystem, we could move it as early as "early". Or for keeping it in
line with other "arch" specific init calls, we could simply make it arch_initcall.

Matwey,

Please could you check if the following patch fixes the issue for you:

Cheers
Suzuki

----8>----

arm64: Enable MRS emulation early enough in the boot sequence
   
Make sure the MRS emulation is enabled early enough that the
early userspace applications (e.g, those run from initrd) could
run without any trouble.
 
Signed-off-by: Suzuki K Poulose <suzuki.poulose@arm.com>

diff --git a/arch/arm64/kernel/cpufeature.c b/arch/arm64/kernel/cpufeature.c
index 9f9e0064c8c1..048f5469531f 100644
--- a/arch/arm64/kernel/cpufeature.c
+++ b/arch/arm64/kernel/cpufeature.c

@@ -1294,4 +1294,4 @@ static int __init enable_mrs_emulation(void)
 	return 0;
 }
 
-late_initcall(enable_mrs_emulation);
+arch_initcall(enable_mrs_emulation);
---

Thanks,

James

[0] works for me:
---------------------%<---------------------
#include <stdio.h>
#include <sys/auxv.h>

#ifndef HWCAP_CPUID
#define HWCAP_CPUID (1 << 11)
#endif

int main(int argc, char **argv)
{
        register unsigned int midr asm ("r1") = 0;
        unsigned long hwcaps = getauxval(AT_HWCAP);

        if (!(hwcaps & HWCAP_CPUID)) {
                fprintf(stderr, "mrs emulation not supported\n");
                return 1;
        }

        asm("mrs %0, midr_el1" : "=r"(midr));

        fprintf(stderr, "mrs x1, midr_el1; x1=0x%x\n", midr);

        return 0;
}
---------------------%<---------------------