ARM: tag bits cleared in FAR_EL1

[Peter Collingbourne]

The ARM Architecture Reference Manual provides the following for FAR_EL1:

"For a Data Abort or Watchpoint exception, if address tagging is enabled for the address accessed by the data access that caused the exception, then this field includes the tag."

However, I have found that the tag bits in FAR_EL1 are always clear, even if the tag bits were set in the original access.

I can reproduce the problem on both 4.1.1 and master (6e8a73e911f066527e775e04b98f31ebd19db600).

[Richard Henderson (rth)]

As it happens, I posted some cleanups for this last week:
https://<email address hidden>/

Some of them have been queued to Peter's target-arm.next branch,
but that hasn't made it to master yet.

[Richard Henderson (rth)]

Actually, I take that back: Peter has merged my TBI patch set,
and is included in 6e8a73e911f066.

Do you have a test case?

[Richard Henderson (rth)]

Ho hum, I must have been asleep last night.
Peter only merged 7 of 9 patches. The final 2 were re-posted:
https://<email address hidden>/

which includes the critical change that affects FAR_ELx.

[Peter Collingbourne (pcc-goog)]

With those two patches applied I can no longer reproduce the problem, thanks!

For posterity, this is how I've been reproducing the problem:

1. Build a Linux kernel with this patch applied: https://patchwork.kernel.org/patch/11435077/
2. Run this program under the kernel:

#include <stdint.h>
#include <stdio.h>
#include <signal.h>

void handler(int signo, siginfo_t *siginfo, void *context) {
  uint32_t *begin = (uint32_t *)context;
  uint32_t *end = ((uint32_t *)context) + (sizeof(ucontext_t)/4);
  for (uint32_t *i = begin; i != end; ++i) {
    printf("%08p %08x\n", i, *i);
  }
  _exit(0);
}

int main() {
  struct sigaction sa;
  sa.sa_sigaction = handler;
  sa.sa_flags = SA_SIGINFO;
  sigaction(SIGSEGV, &sa, 0);

  return *(int *)((1ULL << 56) + 0x123456);
}

I would expect this program's output to include something like the following:

0xffffd5869bd0 46415201
0xffffd5869bd4 00000010
0xffffd5869bd8 00123456
0xffffd5869bdc 01000000

But the output that I was seeing with the bad qemu looked like this:

0xffffd5869bd0 46415201
0xffffd5869bd4 00000010
0xffffd5869bd8 00123456
0xffffd5869bdc 00000000

[Richard Henderson (rth)]

Fix now in master.

[Laurent Vivier (laurent-vivier)]

Fixed here:
https://git.qemu.org/?p=qemu.git;a=commitdiff;h=38d931687fa1