coreutils
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [PATCH] chcon: avoid a race when partial contexts are specified


From: Pádraig Brady
Subject: Re: [PATCH] chcon: avoid a race when partial contexts are specified
Date: Sun, 1 Nov 2026 13:03:57 +0000
User-agent: Mozilla Thunderbird Beta

On 01/11/2026 01:44, Collin Funk wrote:
Hi Pádraig,
Pádraig Brady <P@draigBrady.com> writes:

On 31/10/2026 03:55, Collin Funk wrote:
     if (specified_context == NULL)
       {
-      int status = (affect_symlink_referent
-                    ? getfileconat (fd, file, &file_context)
-                    : lgetfileconat (fd, file, &file_context));
-
-      if (status < 0 && errno != ENODATA)
+      /* libselinux versions before 3.4 did not support O_PATH.
+         The O_PATH support works by opening the file descriptor under
+         /proc/self/fd.  We could check if fsetfilecon returns EBADF and
+         do the same, but there is likely not much benefit over just
+         using O_RDONLY.  */
+      fd = openat (dirfd, file,
+                   (O_RDONLY | O_NOCTTY | O_NONBLOCK
+                    | (affect_symlink_referent ? 0 : O_NOFOLLOW)));


I dislike the real open for a metadata update.
There may be side effects on devices etc.
or requiring read permissions where previously we did not,
and also symlinks could cause issues.
E.g. the following now fails with ELOOP:

   d='testd'
   mkdir -p $d
   touch $d/file
   ln -s file $d/link
   se_type=$(stat -c %C $d/link | cut -d: -f3)
   src/chcon -h -t $se_type $d/link

This is a tricky one due to the O_PATH limitation you commented on.
Though libselinux 3.4 was 2022, and missing /proc would be very
unusual on selinux systems, so it's probably fine to fall back
to the existing implementation in that case.  Note I would not
fall back on any openat() failure as an attacker could be replacing
a file so you'd want to bail out to avoid that. I.e. only fall back
when fgetfilecon() returns EBADF.
Note we could fail rather than fall back if any partial option is used,
but that would be a bit too draconian for this I think.

Thanks for the input. I think I am entirely in agreement with you. I
just want to share a few things I noticed while investigating this, so
you are aware and can take them into account.

First, I think our implementation would currently fail with /proc
missing. See the current strace output before this patch:

     $ mkdir -p a/b; strace chcon -R -u root a; rm -rf a
     [...]
     openat(AT_FDCWD, "a", 
O_RDONLY|O_NOCTTY|O_NONBLOCK|O_NOFOLLOW|O_CLOEXEC|O_DIRECTORY) = 3
     fstat(3, {st_mode=S_IFDIR|0755, st_size=2, ...}) = 0
     fcntl(3, F_GETFL)                       = 0x38800 (flags 
O_RDONLY|O_NONBLOCK|O_LARGEFILE|O_NOFOLLOW|O_DIRECTORY)
     fcntl(3, F_SETFD, FD_CLOEXEC)           = 0
     fcntl(3, F_DUPFD_CLOEXEC, 3)            = 4
     [...]
     lgetxattr("/proc/self/fd/4/b", "security.selinux", 
"unconfined_u:object_r:gconf_home"..., 255) = 38
     access("/var/run/setrans/.setrans-unix", F_OK) = -1 ENOENT (No such file 
or directory)
     futex(0x7f865971a5a0, FUTEX_WAKE_PRIVATE, 2147483647) = 0
     lsetxattr("/proc/self/fd/4/b", "security.selinux", 
"root:object_r:gconf_home_t:s0", 30, 0) = 0
     [...]

I.e., the file descriptor functions in selinux-at.h use /proc.

If we use my patch, but with s/O_RDONLY/O_PATHSEARCH/g, then here is
what we see:

     $ mkdir -p a/b; strace ./src/chcon -R -u root a; rm -rf a
     [...]
     openat(AT_FDCWD, "a", 
O_RDONLY|O_NOCTTY|O_NONBLOCK|O_NOFOLLOW|O_CLOEXEC|O_DIRECTORY) = 3
     fstat(3, {st_mode=S_IFDIR|0755, st_size=2, ...}) = 0
     fcntl(3, F_GETFL)                       = 0x38800 (flags 
O_RDONLY|O_NONBLOCK|O_LARGEFILE|O_NOFOLLOW|O_DIRECTORY)
     fcntl(3, F_SETFD, FD_CLOEXEC)           = 0
     fcntl(3, F_DUPFD_CLOEXEC, 3)            = 4
     [...]
     openat(4, "b", O_RDONLY|O_NOCTTY|O_NONBLOCK|O_NOFOLLOW|O_PATH) = 3
     fgetxattr(3, "security.selinux", 0x3f7ca9c0, 255) = -1 EBADF (Bad file 
descriptor)
     fcntl(3, F_GETFL)                       = 0x220000 (flags 
O_RDONLY|O_NOFOLLOW|O_PATH)
     getxattr("/proc/self/fd/3", "security.selinux", 
"unconfined_u:object_r:gconf_home"..., 255) = 38
     access("/var/run/setrans/.setrans-unix", F_OK) = -1 ENOENT (No such file 
or directory)
     futex(0x7f332549b5a0, FUTEX_WAKE_PRIVATE, 2147483647) = 0
     fsetxattr(3, "security.selinux", "root:object_r:gconf_home_t:s0", 30, 0) = 
-1 EBADF (Bad file descriptor)
     fcntl(3, F_GETFL)                       = 0x220000 (flags 
O_RDONLY|O_NOFOLLOW|O_PATH)
     setxattr("/proc/self/fd/3", "security.selinux", 
"root:object_r:gconf_home_t:s0", 30, 0) = 0

I.e., when we use O_PATH, libselinux-3.4 and later will get EBADF from
fgetxattr or fsetxattr, and call getxattr or setxattr respectively on
the /proc/self/fd path associated with the file descriptor. My
understanding is that the versions before libselinux-3.4 would just
return EBADF instead of that fallback.

This also means that a portable implementation could just open a file
descriptor with O_PATH, and call getfilecon and setfilecon on the
/proc/self/fd path directly, instead of letting libselinux handle the
fallback for us. WDYT?
Yes good observations.
However this would couple the implementation to /proc unnecessarily I think.
Also isn't the libselinux /proc usage observation only the fast path.
So it seems better to use the more abstract interfaces IMHO.
I've attached code for what I was thinking.

cheers,
Padraig

Attachment: chcon-robust-partial.patch
Description: Text Data


reply via email to

[Prev in Thread] Current Thread [Next in Thread]