[compiler-rt] Fix build of Sanitizer-Test_Nolibc after D80648

[arichardson]

Summary

Running ninja check-sanitizer fails for after that patch (commit
058f5f6fd813d1ee1480497394d6fd44e65ec62b) with the following error:

libRTSanitizerCommon.test.nolibc.x86_64.a(sanitizer_posix.cpp.o): In
function `__sanitizer::GetNamedMappingFd(char const*, unsigned long,
int*)':
..../llvm-project/compiler-rt/lib/sanitizer_common/sanitizer_posix.cpp:358:
undefined reference to `fcntl'
clang-12: error: linker command failed with exit code 1 (use -v to see
invocation)

This patch works around the problem by only calling fcntl if O_CLOEXEC
is not defined.

[arichardson]

This is only a workaround and a better solution would probably be an internal_fcntl() wrapper? Or maybe internal_set_cloexec() to avoid having to add a varargs syscall wrapper? However, I don't have time to work on that so I'd like to get the minimal compilation fix committed first since it's breaking ninja check-all for me.

[plopresti]

This is pretty similar to my original patch, actually... But I ended up trying to minimize #ifdefs.

I do have a question, though. fcntl() and FD_CLOEXEC have been mandated by POSIX since the beginning of time (https://pubs.opengroup.org/onlinepubs/009695399/functions/fcntl.html). So how could it be an undefined symbol?

[arichardson]

In D85114#2191044, @plopresti wrote:

This is pretty similar to my original patch, actually... But I ended up trying to minimize #ifdefs.

I do have a question, though. fcntl() and FD_CLOEXEC have been mandated by POSIX since the beginning of time (https://pubs.opengroup.org/onlinepubs/009695399/functions/fcntl.html). So how could it be an undefined symbol?

The nolibc version of the tests builds a sanitizer runtime that does not link against libc and reimplements the C system call wrappers inside the runtime (e.g. internal_open()/internal_unlink()), so unless fcntl is also provided by the compiler-rt runtime it will be undefined. The error only happens when building the sanitizer_nolibc_test_main test since it explicitly compiles with -nostdlib (all other tests end up using fcntl from libc).

[plopresti]

Got it. Was never expecting to run in a non-POSIX-compliant environment.

internal_fcntl() seems like a lot of effort for this one-off need. Thanks
for the fix.

compiler-rt/lib/sanitizer_common/sanitizer_posix.cpp

internal_snprintf(shmname, sizeof(shmname), "/dev/shm/%zu [%s]",
                     internal_getpid(), name);
   int o_cloexec = 0;
 #if defined(O_CLOEXEC)
   o_cloexec = O_CLOEXEC;
 #endif
   int fd = ReserveStandardFds(
       internal_open(shmname, O_RDWR | O_CREAT | O_TRUNC | o_cloexec, S_IRWXU));
   CHECK_GE(fd, 0);
-  if (!o_cloexec) {
-    int res = fcntl(fd, F_SETFD, FD_CLOEXEC);
-    CHECK_EQ(0, res);
-  }
   int res = internal_ftruncate(fd, size);
+#if !defined(O_CLOEXEC)
+  res = fcntl(fd, F_SETFD, FD_CLOEXEC);
+  CHECK_EQ(0, res);
+#endif
   CHECK_EQ(0, res);
   res = internal_unlink(shmname);
   CHECK_EQ(0, res);
   *flags &= ~(MAP_ANON | MAP_ANONYMOUS);
   return fd;
 }
 #else
 int GetNamedMappingFd(const char *name, uptr size, int *flags) {