[sanitizer] Use AT_EXECFN in ReExec() if available

[bkramer]

Summary

execve("/proc/self/exe") will not work if the binary relies on
$EXEC_ORIGIN in an rpath. Query AT_EXECFN instead, which will give the
same string that the current binary was exec'd with.

[kcc]

Is a test possible here?

[bkramer]

Add test

[bkramer]

In D54113#1287835, @kcc wrote:

Is a test possible here?

I thought it was only possible on PPC, but I can abuse the fact that MSAN re-execs on unlimited stacks. Test included.

[vitalybuka]

test/msan/Linux/reexec_unlimited_stack.cc

17

How this test triggers changed code?

[bkramer]

test/msan/Linux/reexec_unlimited_stack.cc

17

The idea is that the ulimit -s unlimited makes msan re-exec on startup. Currently this overwrites AT_EXECFN with /proc/self/exe. With this change it's preserved.

[bkramer]

• Add a comment to the test.

[MaskRay]

$EXEC_ORIGIN ($ORIGIN without tracing the symlink) does not exist in vanilla glibc https://sourceware.org/ml/libc-alpha/2017-10/msg01109.html

This breaks TSanitizer-x86_64-Test-Nolibc as getauxval is a glibc symbol, which can be fixed by https://reviews.llvm.org/D54160

ld.lld: error: undefined symbol: getauxval
>>> referenced by sanitizer_linux.cc:646 (../projects/compiler-rt/lib/sanitizer_common/sanitizer_linux.cc:646)
>>>               sanitizer_linux.cc.o:(__sanitizer::ReExec()) in archive libRTSanitizerCommon.test.nolibc.x86_64.a

lib/sanitizer_common/sanitizer_linux.cc

#if SANITIZER_NETBSD
   uptr len;
 
   len = sizeof(path);
   if (internal_sysctl(name, ARRAY_SIZE(name), path, &len, NULL, 0) != -1)
     pathname = path;
 #elif SANITIZER_SOLARIS
   pathname = getexecname();
   CHECK_NE(pathname, NULL);
+#elif SANITIZER_USE_GETAUXVAL
+  // Calling execve with /proc/self/exe sets that as $EXEC_ORIGIN. Binaries that
+  // rely on that will fail to load shared libraries. Query AT_EXECFN instead.
+  pathname = reinterpret_cast<const char *>(getauxval(AT_EXECFN));
 #endif
 
   GetArgsAndEnv(&argv, &envp);
   uptr rv = internal_execve(pathname, argv, envp);
   int rverrno;
   CHECK_EQ(internal_iserror(rv, &rverrno), true);
   Printf("execve failed, errno %d\n", rverrno);
   Die();

test/msan/Linux/reexec_unlimited_stack.cc

+// MSAN re-execs on unlimited stacks. We use that to verify ReExec() uses the
+// right path.
+// RUN: %clangxx_msan -O0 %s -o %t && ulimit -s unlimited && %run %t | FileCheck %s
+
+#include <stdio.h>
+
+#if !defined(__GLIBC_PREREQ)
+#define __GLIBC_PREREQ(a, b) 0
+#endif
+
+#if __GLIBC_PREREQ(2, 16)
+#include <sys/auxv.h>
+#endif
+
+int main() {
+#if __GLIBC_PREREQ(2, 16)
+  // Make sure AT_EXECFN didn't get overwritten by re-exec.
+  puts(reinterpret_cast<const char *>(getauxval(AT_EXECFN)));
+#else
+  puts("No getauxval");
+#endif
+  // CHECK-NOT: /proc/self/exe
+}