Under 4.1.5 x86-64 kernel, PT_GNU_STACK interferes with mprotect.
With PT_GNU_STACK,

__mprotect (newp->start, newp->len, PROT_READ|PROT_WRITE)

changes the pages to PROT_READ|PROT_WRITE|PROT_EXEC:

[hjl@gnu-6 pr18801]$ cat foo.c
int zoo_1 (void) {
  return 0;
}

void *selector (void) {
  return (void *)&zoo_1;
}

int zoo(void) __attribute__ ((ifunc ("selector")));

int main(void) {
  return zoo ();
}
[hjl@gnu-6 pr18801]$ gcc -O2 foo.c -mcmodel=large -pie 
[hjl@gnu-6 pr18801]$ ./a.out 
Segmentation fault   # <--- This is expected.
[hjl@gnu-6 pr18801]$ gcc -O2 foo.c -mcmodel=large -pie -Wl,-z,execstack
[hjl@gnu-6 pr18801]$ ./a.out # <--- This should segfault.
[hjl@gnu-6 pr18801]$ 

Before mprotect:

555555554000-555555555000 r-xp 00000000 08:11 24735096 ..

After mprotect:

555555554000-555555555000 rwxp 00000000 08:11 24735096
                            ^  It should be off.

IA32 is OK:

[hjl@gnu-6 pr18801]$ gcc -O2 foo.c -m32 -pie
[hjl@gnu-6 pr18801]$ ./a.out 
Segmentation fault
[hjl@gnu-6 pr18801]$ gcc -O2 foo.c -m32 -pie -Wl,-z,execstack
[hjl@gnu-6 pr18801]$ ./a.out 
Segmentation fault
[hjl@gnu-6 pr18801]$