mkdir -p implemented in C

mkdir_p.c

#include <stdlib.h>
#include <string.h>
#include <sys/stat.h>   /* mkdir(2) */
#include <errno.h>

/* Make a directory; already existing dir okay */
static int maybe_mkdir(const char* path, mode_t mode)
{
    struct stat st;
    errno = 0;

    /* Try to make the directory */
    if (mkdir(path, mode) == 0)
        return 0;

    /* If it fails for any reason but EEXIST, fail */
    if (errno != EEXIST)
        return -1;

    /* Check if the existing path is a directory */
    if (stat(path, &st) != 0)
        return -1;

    /* If not, fail with ENOTDIR */
    if (!S_ISDIR(st.st_mode)) {
        errno = ENOTDIR;
        return -1;
    }

    errno = 0;
    return 0;
}

int mkdir_p(const char *path)
{
    /* Adapted from http://stackoverflow.com/a/2336245/119527 */
    char *_path = NULL;
    char *p; 
    int result = -1;
    mode_t mode = 0777;

    errno = 0;

    /* Copy string so it's mutable */
    _path = strdup(path);
    if (_path == NULL)
        goto out;

    /* Iterate the string */
    for (p = _path + 1; *p; p++) {
        if (*p == '/') {
            /* Temporarily truncate */
            *p = '\0';

            if (maybe_mkdir(_path, mode) != 0)
                goto out;

            *p = '/';
        }
    }   

    if (maybe_mkdir(_path, mode) != 0)
        goto out;

    result = 0;

out:
    free(_path);
    return result;
}

mkdir_p.h

#ifndef MKDIR_P_H
#define MKDIR_P_H

int mkdir_p(const char *path);

#endif /* MKDIR_P_H */

SConstruct

env = Environment(
    CCFLAGS = ['-Wall', '-Werror'],
)
env.Program('mkdir_p_test', ['mkdir_p.c', 'test.c'])

test.c

#include <stdio.h>
#include "mkdir_p.h"

int main(int argc, char **argv)
{
    const char *path;
    int rc;

    if (argc < 2) {
        fprintf(stderr, "Missing argument: path\n");
        return 1;
    }

    path = argv[1];

    rc = mkdir_p(path);

    fprintf(stderr, "mkdir_p(\"%s\") returned %d: %m\n", path, rc);

    return (rc == 0) ? 0 : 2;
}

Your check to continue if an error is just because of an existing path component, could also include checking that that component is a directory.

PATH_MAX is provided by <linux/limits.h>, not <limits.h>.

Yes. It's true.

In <limits.h>, we may have some thing like POSIX_PATH_MAX (=256) which is imported from <bits/posix1_lim.h> when we turn USE_POSIX on.

In <linux/limits.h>, PATH_MAX = 4096.

Never use PATH_MAX.
Never.
It's never the truth even if you try to get it programmatically from sys_config.
Always use malloc/free.

if (mkdir(_path, S_IRWXU) != 0) {
    if (errno != EEXIST)
        return -1;
}

Must check whether _path is really a directory if errno is EEXIST, see my fork:

if (errno == EEXIST) {
  struct stat st;
  ret = stat(path, &st);
  if (ret != 0) {
    return ret;
  } else if (S_ISDIR(st.st_mode)) {
    return 0;
  } else {
    return -1;
  }
}

@llothar @tangxinfa Thanks everyone for the feedback. I made a few improvements:

1. Switch from fixed array to dynamic allocation to avoid PATH_MAX problems
2. Check that existing paths are really directories, and return ENOTDIR if they are not
3. Don't leak EEXIST on success
4. Change default mode from 0700 to 0777 to match command line mkdir -p behavior (will be affected by umask)
5. Added a test app

BTW mkdir_p( "" ) accesses beyond allocated memory since it assumes the path length is non-zero. Seems it should return ENOENT in this case, as mkdir does for an empty string.

As for optimization, would it make sense to just try maybe_mkdir in the beginning, for the probably-common case where the parent directories already exist?

Thanks for the useful functions and careful implementation.

This doesn't make sense:

if (_path == NULL)
        goto out;

Because it executes free(_path); which is undefined for NULL.

@SamuelMarks

This doesn't make sense:

if (_path == NULL)
        goto out;

Because it executes free(_path); which is undefined for NULL.

Incorrect. From free(3p):

If ptr is a null pointer, no action shall occur.

https://stackoverflow.com/a/2360462 oh looks like my best-practices is a bit out-of-date…