Performance comparison: Function pointers in memory-protected global struct

[robertdfrench]

from @nekopsykose

#include <err.h>
#include <unistd.h>
#include <sys/mman.h>
#include <stdio.h>

struct myfuncs {
  unsigned (*multiply)(unsigned, unsigned);
};

static struct myfuncs *page;


unsigned multiply_base(unsigned x, unsigned y) {
  return x * y;
}

unsigned multiply_sse42(unsigned x, unsigned y) {
  return x + y; /* lol */
}

void initfuncs(void) {
  int pagesize = getpagesize();
  page = mmap(NULL, pagesize, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANON, 0, 0);

  if (page == MAP_FAILED)
    err(123, "mmap");
  

  /* magical runtime check */
  if (0) {
    page->multiply = multiply_sse42;
  } else {
    page->multiply = multiply_base;
  }

  if(mprotect(page, pagesize, PROT_READ))
    err(123, "mprotect");


  /* segfaults */
  // page->multiply = multiply_sse42;
}


int main(void) {
  initfuncs();

  unsigned x = page->multiply(4, 4);

  printf("%u\n", x);
}

[robertdfrench]

TEST	LOW	HIGH	AVG
fixed	2.00	2.21	2.092
pointer	3.90	4.59	4.341
struct	3.61	3.88	3.76
ifunc	5.29	5.70	5.514
upfront	7.75	8.04	7.875
always	12.88	13.70	13.448

So the "Struct of pointers in its own page" approach seems slightly faster than the "global static pointer not necessarily in its own page" approach, thought I'm not sure why that should be the case.

[CharaVerKys]

interesting idea, and also naturally provide lib->function() pattern
is it just hardware/os lvl optimizations for protected memory region? well result, if i will ever need it, thx.

(but)
i have no doubt that relying on something from system might be not secure by nature, but i have doubts that ifunc not solve speed problem for real application
is any one tested speed difference for big .so libs? may be GOT will work better than pointer there
i not sure is ifunc speed is ifunc speed, or actually GOT speed, added sleep(10) for resolve function and got real 18.55 | user 8.41
so it should be resolving ptr with gnu instead of just language ptr
i not know enough to say more about it, if it is issue with speed for .so table, why not everyone just use protected ptrs pattern?

not wanted to add another issue to just express this