| Summary: | glob("/foo/*/") may also match regular & other kind of files, not just directories | ||
|---|---|---|---|
| Product: | glibc | Reporter: | Șpagoveanu <spagoveanu> |
| Component: | glob | Assignee: | Not yet assigned to anyone <unassigned> |
| Status: | RESOLVED FIXED | ||
| Severity: | normal | CC: | gadelat, stephane+sourceware |
| Priority: | P2 | ||
| Version: | unspecified | ||
| Target Milestone: | 2.36 | ||
| Host: | Target: | ||
| Build: | Last reconfirmed: | ||
Note that it was reported at https://unix.stackexchange.com/questions/638955/what-could-be-a-cause-for-getdents-returning-different-results-on-2-systems, there about a "prl_fs" filesystem (likely related to some Parallel virtualisation software). I can reproduce it with glibc 2.33, and also with broken symlinks: $ mkdir -p testdir/dir $ ln -s /no/such/file testdir/broken $ ./glob 'testdir/*/' testdir/broken testdir/dir/ Note that while the issue can be reproduced with ./*/ or b*/, I can't reproduce it with */ $ cd testdir $ ../glob './*/' ./broken ./dir/ $ ../glob 'b*/' broken $ ../glob '*/' dir/ Fixed in 2.36 by commit 7c477b57a3. |
On Linux, glob(3) assumes that a regular file cannot have d_type == DT_UNKNOWN, irrespective of filesystem support or whether the GLOB_ALTDIRFUNC feature was used. This results in a pattern like "/foo/*/" also matching non-directories, which is contrary to what the standard requires ("The <slash> character in a pathname shall be explicitly matched by using one or more <slash> characters in the pattern"). At least on Linux, you CANNOT refer to a regular file as "/path/to/file/". See the simple testcase below, using a minix filesystem, which does not support d_type. You can also check these, about a (fixed) bug in GNU make, caused by its use of GLOB_ALTDIRFUNC: https://www.mail-archive.com/bug-make@gnu.org/msg11073.html https://lists.gnu.org/archive/html/bug-make/2018-06/msg00009.html Once upon a time I had started writing a patch, but I gave up as I had to rewrite the whole logic of glob_in_dir(). FWIW, the glibc documentation claims the GLOB_ONLYDIR is only a "hint", but the implementation itself is assuming more than that. --------------x---------------- # cat >glob.c <<'EOT'; cc -Wall -O2 glob.c -o glob #include <stdio.h> #include <glob.h> int main(int ac, char **av){ int i, j; for(i = 1; i < ac; i++){ glob_t g = {}; if(glob(av[i], 0, 0, &g) == 0) for(j = 0; j < g.gl_pathc; j++) printf("%s\n", g.gl_pathv[j]); } } EOT # truncate -s1G minix.img; mkfs.minix minix.img ... # mkdir -p dir; mount minix.img dir # touch dir/file # ./glob 'dir/*/' dir/file --------------x----------------