illumos gate

In investigating why statemaps were slower on illumos than other platforms, it was clear that we were spending an inordinate amount of time in strndup. Take the following program:

#include <stdio.h>
#include <unistd.h>
#include <stdlib.h>
#include <string.h>

/*
 * Made global to prevent the compiler from just optimizing the call away.
 */
char *s;

int
main()
{
    size_t size = 10 * 1024 * 1024;
    int iter = 1000000;
    char *c = malloc(size);

    memset(c, '!', size - 1);
    c[size - 1] = '\0';

    while (iter--) {
        s = strndup(c, 3);
        free(s);
    }
}

On my Mac laptop (2.7 GHz Core i7):

$ time ./strndup 

real    0m0.084s
user    0m0.076s
sys    0m0.005s

On a 1U running SmartOS (3.5 GHz Haswell):

# time ./strndup

real    8m47.107s
user    8m47.051s
sys     0m0.140s

The problem is that strndup needless calls strlcpy to do the actual copying – which will always return (and therefore determine) the length of the source string even when it exceeds the buffer length. This is needless; this should be a bcopy with an explicit NUL byte termination instead.

Running with the fix:

[root@4e773eea-7527-ee67-90ba-de62e536c367 ~]# time ./strndup

real    0m0.051s
user    0m0.045s
sys     0m0.005s

I have also verified that the statemap code (which was the impetus for this) now runs with the expected performance.