[lxc-devel] [PATCH] utils: reimplement/fix mkdir_p()

[Serge Hallyn]

Quoting Richard Weinberger (richard at nod.at):
> Am 17.04.2013 18:00, schrieb Serge Hallyn:
> >Quoting Richard Weinberger (richard at nod.at):
> >>Reimplement mkdir_p() such that it:
> >>  ...handles relativ paths correctly. (currently it crashes)
> >>  ...does not rely on dirname().
> >>  ...is not recursive.
> >>  ...is shorter. ;-)
> >
> >Looks good, thanks.  Yeah I prefer non-recursive.  Three
> >comments though,
> >
> >>Signed-off-by: Richard Weinberger <richard at nod.at>
> >>---
> >>  src/lxc/utils.c | 48 +++++++++++++++++-------------------------------
> >>  1 file changed, 17 insertions(+), 31 deletions(-)
> >>
> >>diff --git a/src/lxc/utils.c b/src/lxc/utils.c
> >>index e07ca7b..9794553 100644
> >>--- a/src/lxc/utils.c
> >>+++ b/src/lxc/utils.c
> >>@@ -95,39 +95,25 @@ extern int get_u16(unsigned short *val, const char *arg, int base)
> >>  	return 0;
> >>  }
> >>
> >>-static int is_all_slashes(char *path)
> >>-{
> >>-	while (*path && *path == '/')
> >>-		path++;
> >>-	if (*path)
> >>-		return 0;
> >>-	return 1;
> >>-}
> >>-
> >>  extern int mkdir_p(char *dir, mode_t mode)
> >>  {
> >>-	int ret;
> >>-	char *d;
> >>-
> >>-	if (is_all_slashes(dir))
> >>-		return 0;
> >>-
> >>-	d = strdup(dir);
> >>-	if (!d)
> >>-		return -1;
> >>-
> >>-	ret = mkdir_p(dirname(d), mode);
> >>-	free(d);
> >>-	if (ret)
> >>-		return -1;
> >>-
> >>-	if (!access(dir, F_OK))
> >>-		return 0;
> >>-
> >>-	if (mkdir(dir, mode)) {
> >>-		SYSERROR("failed to create directory '%s'\n", dir);
> >>-		return -1;
> >>-	}
> >>+	char *tmp = dir;
> >>+	char *orig = dir;
> >>+	char *makeme;
> >>+
> >>+	do {
> >>+		dir = tmp + strspn(tmp, "/");
> >>+		tmp = dir + strcspn(dir, "/");
> >>+		makeme = strndupa(orig, dir - orig);
> >
> >strndupa *can* fail and return NULL.
> 
> How?

I was hoping that smart libc on embedded would return NULL if it sees
there isn't enough space.  I misread the strndupa manpage to say it
returns NULL on failure - I see there is no such promise.

> strndupa() uses alloca(), which allocates memory on the stack.
> *If* it fails we have already overrun the stack and are on a one-way trip
> to lala land. :-)
> 
> >
> >>+		if (*makeme) {
> >>+			if (!access(makeme, F_OK))
> >>+				return 0;
> >
> >did you mean to continue here?
> 
> No, if the access() we have to stop.
             =========

If the access() does what?  access("/var", F_OK) should immediately
return 0, so you're saying mkdir_p("/var/lib/lxc/c1") should immediately
fail?  It's entirely possible I'm not thinking right, but...

> "mkdir -p /foo/bar" also just exists if it is not allowed to access one parent directory.
> 
> >>+			if (mkdir(makeme, mode)) {
> >
> >As people are starting to want to thread, I think it's worth making sure
> >all mkdirs and creats check for errno=-EEXIST even if they've just
> >checked with access.  Certainly with the cgroup setup a race is possible
> >when starting two simultaneous containers first time after boot using
> >the api.
> 
> Yes, you are right. I'll update my patch to take care of this.
> 
> Thanks,
> //richard