Re: [PATCH v4 2/2] glob: implement GLOB_ALTDIRFUNC et al

[Rich Felker <dalias@libc.org>]

On Fri, Sep 06, 2019 at 09:40:22PM +0200, Ismael Luceno wrote:
> diff --git a/include/glob.h b/include/glob.h
> index 0ff70bdfeef2..26ea65cd1097 100644
> --- a/include/glob.h
> +++ b/include/glob.h
> @@ -16,7 +16,16 @@ typedef struct {
>  	char **gl_pathv;
>  	size_t gl_offs;
>  	int __dummy1;
> +
> +#if defined(_BSD_SOURCE) || defined(_GNU_SOURCE)
> +	void (*gl_closedir)(void *);
> +	struct dirent *(*gl_readdir)(void *);
> +	void *(*gl_opendir)(const char *);
> +	int (*gl_lstat)(const char *__restrict, struct stat *__restrict);
> +	int (*gl_stat)(const char *__restrict, struct stat *__restrict);
> +#else
>  	void *__dummy2[5];
> +#endif
>  } glob_t;

I think you need to forward-declare at least struct stat, and possibly
also struct dirent, so that they refer to the same types as the
declarations elsewhere. At least the argument list is a different
scope and declares a *different* struct stat shadowing the file-scope
one if the file-scope one is not already declared.

With the above change and:

> diff --git a/src/regex/glob.c b/src/regex/glob.c
> index 7780e21ee113..4f329f053fe0 100644
> --- a/src/regex/glob.c
> +++ b/src/regex/glob.c
> @@ -1,7 +1,7 @@
>  #define _BSD_SOURCE
           ^^^^^^^^^^^

combined, every use of g->[something] in glob() is now undefined
behavior if the caller was using a standard profile where the
__dummy2-containing version of glob_t above is the one that's used.
That's because it's accessing one struct type via an lvalue of a
different struct type.

> +#include <sys/stat.h>
>  #include <glob.h>
>  #include <fnmatch.h>
> -#include <sys/stat.h>

Unrelated: this change is gratuitous.

> @@ -224,6 +232,10 @@ static int expand_tilde(char **pat, char *buf, size_t *pos)
>  	return 0;
>  }
>  
> +static void wrap_closedir(void *p) { __closedir(p); }
> +static struct dirent *wrap_readdir(void *d) { return __readdir(d); }
> +static void *wrap_opendir(const char *path) { return __opendir(path); }
> +
>  int glob(const char *restrict pat, int flags, int (*errfunc)(const char *path, int err), glob_t *restrict g)
>  {
>  	struct match head = { .next = NULL }, *tail = &head;
> @@ -231,9 +243,24 @@ int glob(const char *restrict pat, int flags, int (*errfunc)(const char *path, i
>  	size_t offs = (flags & GLOB_DOOFFS) ? g->gl_offs : 0;
                                              ^^^^^^^^^^

These are the accesses that become UB. To make them well-defined,
they'd need to be something like G(g,gl_offs), where

#define G(g,m) (*(1?(void*)((char*)(g)+offsetof(glob_t,m)):&((glob_t){0}).m)

is the cleanest abstraction I can think of for the monstrosity needed
to access "something layout-compatible with glob_t, but not glob_t, at
address g".

Imposing this ugliness to avoid UB is part of why I'm hesitant about
this functionality. I think there should be good motivation for it if
we want to go forward with it.

>  	int error = 0;
>  	char buf[PATH_MAX];
> +	struct dirfn dirfn;
>  	
>  	if (!errfunc) errfunc = ignore_err;
>  
> +	if (flags & GLOB_ALTDIRFUNC) {
> +		dirfn.closedir = g->gl_closedir;
> +		dirfn.readdir = g->gl_readdir;
> +		dirfn.opendir = g->gl_opendir;
> +		dirfn.lstat = g->gl_lstat;
> +		dirfn.stat = g->gl_stat;
> +	} else {
> +		dirfn.closedir = wrap_closedir;
> +		dirfn.readdir = wrap_readdir;
> +		dirfn.opendir = wrap_opendir;
> +		dirfn.lstat = lstat;
> +		dirfn.stat = stat;
> +	}
> +
>  	if (!(flags & GLOB_APPEND)) {
>  		g->gl_offs = offs;
>  		g->gl_pathc = 0;
> @@ -249,7 +276,7 @@ int glob(const char *restrict pat, int flags, int (*errfunc)(const char *path, i
>  		if ((flags & (GLOB_TILDE | GLOB_TILDE_CHECK)) && *p == '~')
>  			error = expand_tilde(&s, buf, &pos);
>  		if (!error)
> -			error = do_glob(buf, pos, 0, s, flags, errfunc, &tail);
> +			error = do_glob(buf, pos, 0, s, flags, errfunc, &dirfn, &tail);
>  		free(p);
>  	}

Not necessary for correctness, but if we're adding a context structure
that gets passed down to each level of do_glob, we might as well put
the errfunc in it too, and possibly also flags and &tail in it, so
that we reduce the amount of state at each recursion frame rather than
increasing it.

It's likely this would make more sense to do as a separate patch
though, to keep the diffs clear.

Rich