[asan] [SystemZ] Add slop for stack address detection.

[• koriakin]

Summary

On s390, siginfo reports the faulting address with page granularity -
we need to mask off the low bits of sp before comparison.

[eugenis]

LGTM w/ a nit

lib/asan/asan_posix.cc

49

Do you mean that sig.addr is rounded down to the nearest page boundary?
Do you need some extra space below sp (same as +512 in the common case below)? I.e. is it possible that the true fault address is below sp and falls into a different page, and then sig.addr would be ((sig.sp & 0xFFF) - 0x1000) ?

[• koriakin]

lib/asan/asan_posix.cc

49

sig.addr is rounded down to nearest page bounduary (because s390 gives you only the page index on a fault, not the full address). However, there is no redzone - accesses below sp are verboten. I'll try to think of a better wording for the comment before commiting.

[• koriakin]

Changed the wording in the comment. Does it sound clearer now?

[eugenis]

Yes, looks great.

lib/asan/asan_posix.cc

void AsanOnDeadlySignal(int signo, void *siginfo, void *context) {
   int code = (int)((siginfo_t*)siginfo)->si_code;
   // Write the first message using the bullet-proof write.
   if (18 != internal_write(2, "ASAN:DEADLYSIGNAL\n", 18)) Die();
   SignalContext sig = SignalContext::Create(siginfo, context);
 
   // Access at a reasonable offset above SP, or slightly below it (to account
   // for x86_64 or PowerPC redzone, ARM push of multiple registers, etc) is
   // probably a stack overflow.
+#ifdef __s390__
+  // On s390, the fault address in siginfo points to start of the page, not
+  // to the precise word that was accessed.  Mask off the low bits of sp to
+  // take it into account.
+  bool IsStackAccess = sig.addr >= (sig.sp & ~0xFFF) &&
+                       sig.addr < sig.sp + 0xFFFF;
+#else
   bool IsStackAccess = sig.addr + 512 > sig.sp && sig.addr < sig.sp + 0xFFFF;
+#endif
 
 #if __powerpc__
   // Large stack frames can be allocated with e.g.
   //   lis r0,-10000
   //   stdux r1,r1,r0 # store sp to [sp-10000] and update sp by -10000
   // If the store faults then sp will not have been updated, so test above
   // will not work, becase the fault address will be more than just "slightly"
   // below sp.