Created attachment 41098 [details]
A patch that adds an mmap() interceptor to Helgrind.

Trying to debug ThreadSanitizer's (http://code.google.com/p/data-race-test) mmap() interceptor on Mac OS I've came up with the following test:

=============================
#include <sys/types.h>
#include <sys/mman.h>
#include <stdio.h>
#include <errno.h>
int main () {
 int md;
 caddr_t virt_addr;
 md = shm_open ("apple.shm.notification_center", 0, 0);
 virt_addr = mmap(0, 4096, 1, 1, md, 0);
 if (virt_addr == (void*) -1) {
   printf("errno: %d\n", errno);
   perror("perror");
 }
 printf("mmap: %p\n", virt_addr);
 return 0;
}
=============================

When ran natively, mmap succeeds and virt_addr is 0x8000
However, when ran under ThreadSanitizer mmap fails returning -1 and
errno is 22 (EINVAL).

Other Valgrind tools (Nullgrind, Memcheck, DRD) behave normally, because none of them intercepts mmap(). Disabling the mmap() interceptor fixes the problem, but is unacceptable for us.

I've tried adding the WRAP_MMAP macro to Helgrind (see the patch) and managed to reproduce the same behavior:

valgrind -v  --tool=helgrind  ~/src/shm_test/shm
==49272== Helgrind, a thread error detector
...
--49272-- REDIR: 0x200af8 (mmap) redirected to 0x17f42 (mmap)
errno: 22
--49272-- REDIR: 0x1fff20 (strlen) redirected to 0x1905c (strlen)
perror: Invalid argument
mmap: 0xffffffff
==49272== 
==49272== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 6 from 6)

As the mmap wrapper for Helgrind is really strict and consists of nothing but a call to CALL_FN_W_6W, I believe that CALL_FN_W_6W incorrectly passes the parameters to the original function.