[Bug libc/30737] fdopendir() is not robust - returns bogus DIR* instead of flagging an error

["adhemerval.zanella at linaro dot org" <sourceware-bugzilla@sourceware.org>]

https://sourceware.org/bugzilla/show_bug.cgi?id=30737

Adhemerval Zanella <adhemerval.zanella at linaro dot org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |adhemerval.zanella at linaro dot o
                   |                            |rg

--- Comment #3 from Adhemerval Zanella <adhemerval.zanella at linaro dot org> ---
(In reply to Mark Galeck from comment #0)
> First, dislclaimer: I checked this issue is present on the latest Ubuntu
> distribution, which has glibc version 2.35, not the latest stable 2.38.  I
> was able to build 2.38 locally, but I am not able to force Ubuntu to accept
> it for linking to the sample program, or to upgrade the installed glibc 2.35
> to 2.38. 
> 
> If you know how to test the sample program with 2.38, please let me know,
> and I will do so and I will remove this report if it does not reproduce with
> glibc 2.38.
> ---------------------------
> 
> Here is the SSCCE 
> 'foobar.c'
> 
> (which I obtained by necessary modifications to the sample code in man7
> entry for fdopendir() ) :
> 
> 
> #define _GNU_SOURCE
> 
> #include <stdio.h>
> #include <dirent.h>
> #include <fcntl.h>
> #include <stdlib.h>
> #include <string.h>
> 
> int
> main(int argc, char* argv[])
> {
>     DIR* d;
>     struct dirent* dp;
>     int dfd;
> 
>     if ((d = fdopendir(open("./dir", O_RDONLY | O_PATH | O_DIRECTORY))) ==
> NULL) {
>         fprintf(stderr, "Cannot open directory\n");
>         exit(1);
>     }
> 
>     while ((dp = readdir(d)) != NULL) {
>         if (!strcmp(dp->d_name, "."))
>             continue;
>         if (!strcmp(dp->d_name, ".."))
>             continue;
>         printf("d_name %s\n", dp->d_name);
>     }
>     return 0;
> }
> 
> -------------------------
> Here is a script to show the problem:
> 
> $ mkdir dir
> $ touch dir/file
> $ gcc -lc foobar.c
> $ ./a.out
> $
> 
> When called on a descriptor opened with O_PATH , fdopendir() succeeds
> happily, but then, readdir() reports nothing!

But readdir *does* report it, if you set errno as 0 prior calling, you will see
that EBADF is correctly set (from getdents64).

> 
> So the DIR* returned from fdopendir() is useless, but fdopendir() returns
> it, instead of NULL and setting errno to EBADF , as it should INHO
> 

Although POSIX defines fdopendir EBADF for file descriptors not open for
reading, O_PATH is a Linux extension, and standard-wise not failing with it is
still conformant. However, I agree this is QoI to fail if the file descriptor
is opened with O_PATH (it fails early and Linux fdopendir already call fcntl,
so it is a free check).

> 
> If I remove | O_PATH from the above sample code, I get the expected result:
> 
> $gcc -lc foobar.c
> $ ./a.out
> d_name file

-- 
You are receiving this mail because:
You are on the CC list for the bug.