[sanitizers] [SystemZ] Add __tls_get_offset interceptor.

[• koriakin]

Summary

s390 is special again - instead of tls_get_addr, it has tls_get_offset
with special calling conventions: the result is TP relative, and
the argument is GOT-relative. Since we need to get address of the caller's
GOT, which is in %r12, we have to use assembly like glibc does.

Aside of tls_get_offset, glibc also implements a slightly saner
tls_get_addr_internal, which takes a pointer as argument, but still
returns a TP-relative offset. It is used for dlsym() called on TLS
symbols, so we have to intercept it was well. Our __tls_get_offset
is also implemented by delegating to it.

[eugenis]

All these macro definitions and conditionals make __tls_get_addr interceptor confusing.
Maybe just copy the whole thing, then you'll end up with a few duplicated lines of code, but each version would be a lot simpler.

Is it possible to avoid this module-level asm by implementing an honest interceptor for __tls_get_offset, and a tiny chunk of asm that simply returns %r12, or finds the address of GOT some other way (aren't there linker symbols for the start of each section?) ?

[kcc]

+1
As I've said before -- avoid #ifdefs inside the code at all costs.

[• koriakin]

In D19778#419152, @eugenis wrote:

All these macro definitions and conditionals make __tls_get_addr interceptor confusing.
Maybe just copy the whole thing, then you'll end up with a few duplicated lines of code, but each version would be a lot simpler.

OK.

Is it possible to avoid this module-level asm by implementing an honest interceptor for __tls_get_offset, and a tiny chunk of asm that simply returns %r12, or finds the address of GOT some other way (aren't there linker symbols for the start of each section?) ?

Finding the address of GOT would be rather hard (you need the caller's GOT not your own, so the symbol isn't of much use). A tiny chunk of asm in a function would be nice, but by that point the register allocator could've already used %r12 for something else. A naked function would make this code slightly less ugly, but it's not supported by gcc for s390. I really don't see how to avoid that top-level assembly...

[• koriakin]

This version duplicates the code for s390.

[eugenis]

LGTM but please wait for a comment from kcc

[kcc]

LGTM (I did not try to review the actual code under #ifdef)
If we get more of these, please move them into a separate s390-specific file

[• koriakin]

In D19778#419333, @kcc wrote:

LGTM (I did not try to review the actual code under #ifdef)
If we get more of these, please move them into a separate s390-specific file

Thanks, that should be the last s390-specific code change for asan - I've got the testsuite passing already, but still need to land many patches on the llvm side.

compiler-rt/trunk/lib/sanitizer_common/sanitizer_common_interceptors.inc

INTERCEPTOR(__sanitizer_clock_t, times, void *tms) {
   return res;
 }
 #define INIT_TIMES COMMON_INTERCEPT_FUNCTION(times);
 #else
 #define INIT_TIMES
 #endif
 
 #if SANITIZER_INTERCEPT_TLS_GET_ADDR
+#if !SANITIZER_S390
 #define INIT_TLS_GET_ADDR COMMON_INTERCEPT_FUNCTION(__tls_get_addr)
 // If you see any crashes around this functions, there are 2 known issues with
 // it: 1. __tls_get_addr can be called with mis-aligned stack due to:
 // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58066
 // 2. It can be called recursively if sanitizer code uses __tls_get_addr
 // to access thread local variables (it should not happen normally,
 // because sanitizers use initial-exec tls model).
 INTERCEPTOR(void *, __tls_get_addr, void *arg) {
   void *ctx;
   COMMON_INTERCEPTOR_ENTER(ctx, __tls_get_addr, arg);
   void *res = REAL(__tls_get_addr)(arg);
   uptr tls_begin, tls_end;
   COMMON_INTERCEPTOR_GET_TLS_RANGE(&tls_begin, &tls_end);
   DTLS::DTV *dtv = DTLS_on_tls_get_addr(arg, res, tls_begin, tls_end);
   if (dtv) {
     // New DTLS block has been allocated.
     COMMON_INTERCEPTOR_INITIALIZE_RANGE((void *)dtv->beg, dtv->size);
   }
   return res;
 }
+#else // SANITIZER_S390
+// On s390, we have to intercept two functions here:
+// - __tls_get_addr_internal, which is a glibc-internal function that is like
+//   the usual __tls_get_addr, but returns a TP-relative offset instead of
+//   a proper pointer.  It is used by dlsym for TLS symbols.
+// - __tls_get_offset, which is like the above, but also takes a GOT-relative
+//   descriptor offset as an argument instead of a pointer.  GOT address
+//   is passed in r12, so it's necessary to write it in assembly.  This is
+//   the function used by the compiler.
+#define INIT_TLS_GET_ADDR COMMON_INTERCEPT_FUNCTION(__tls_get_addr_internal)
+INTERCEPTOR(uptr, __tls_get_addr_internal, void *arg) {
+  void *ctx;
+  COMMON_INTERCEPTOR_ENTER(ctx, __tls_get_addr_internal, arg);
+  uptr res = REAL(__tls_get_addr_internal)(arg);
+  uptr tp = reinterpret_cast<uptr>(__builtin_thread_pointer());
+  void *ptr = reinterpret_cast<void *>(res + tp);
+  uptr tls_begin, tls_end;
+  COMMON_INTERCEPTOR_GET_TLS_RANGE(&tls_begin, &tls_end);
+  DTLS::DTV *dtv = DTLS_on_tls_get_addr(arg, ptr, tls_begin, tls_end);
+  if (dtv) {
+    // New DTLS block has been allocated.
+    COMMON_INTERCEPTOR_INITIALIZE_RANGE((void *)dtv->beg, dtv->size);
+  }
+  return res;
+}
+// We need a protected symbol aliasing the above, so that we can jump
+// directly to it from the assembly below.
+extern "C" __attribute__((alias("__interceptor___tls_get_addr_internal"),
+                          visibility("protected")))
+uptr __interceptor___tls_get_addr_internal_protected(void *arg);
+// Now carefully intercept __tls_get_offset.
+asm(
+  ".text\n"
+  ".global __tls_get_offset\n"
+  "__tls_get_offset:\n"
+// The __intercept_ version has to exist, so that gen_dynamic_list.py
+// exports our symbol.
+  ".global __interceptor___tls_get_offset\n"
+  "__interceptor___tls_get_offset:\n"
+#ifdef __s390x__
+  "la %r2, 0(%r2,%r12)\n"
+  "jg __interceptor___tls_get_addr_internal_protected\n"
+#else
+  "basr %r3,0\n"
+  "0: la %r2,0(%r2,%r12)\n"
+  "l %r4,1f-0b(%r3)\n"
+  "b 0(%r4,%r3)\n"
+  "1: .long __interceptor___tls_get_addr_internal_protected - 0b\n"
+#endif
+  ".type __tls_get_offset, @function\n"
+  ".size __tls_get_offset, .-__tls_get_offset\n"
+);
+#endif // SANITIZER_S390
 #else
 #define INIT_TLS_GET_ADDR
 #endif
 
 #if SANITIZER_INTERCEPT_LISTXATTR
 INTERCEPTOR(SSIZE_T, listxattr, const char *path, char *list, SIZE_T size) {
   void *ctx;
   COMMON_INTERCEPTOR_ENTER(ctx, listxattr, path, list, size);