Re: [PATCH] proc: Track /proc/$pid/attr/ opener mm_struct

Re: [PATCH] proc: Track /proc/$pid/attr/ opener mm_struct

[Kees Cook]

On Mon, Jun 14, 2021 at 06:02:34PM +0800, youling257 wrote:
> I used mainline kernel on android, this patch cause "failed to retrieve pid context" problem.
> 
> 06-14 02:15:51.165  1685  1685 E ServiceManager: SELinux: getpidcon(pid=1682) failed to retrieve pid context.
> 06-14 02:15:51.166  1685  1685 E ServiceManager: add_service('batteryproperties',1) uid=0 - PERMISSION DENIED
> 06-14 02:15:51.166  1682  1682 I ServiceManager: addService() batteryproperties failed (err -1 - no service manager yet?).  Retrying...
> 06-14 02:15:51.197  1685  1685 E ServiceManager: SELinux: getpidcon(pid=1695) failed to retrieve pid context.
> 06-14 02:15:51.197  1685  1685 E ServiceManager: add_service('android.security.keystore',1) uid=1017 - PERMISSION DENIED
> 06-14 02:15:51.198  1695  1695 I ServiceManager: addService() android.security.keystore failed (err -1 - no service manager yet?).  Retrying...
> 06-14 02:15:51.207  1685  1685 E ServiceManager: SELinux: getpidcon(pid=1708) failed to retrieve pid context.
> 06-14 02:15:51.207  1685  1685 E ServiceManager: add_service('android.service.gatekeeper.IGateKeeperService',1) uid=1000 - PERMISSION DENIED
> 06-14 02:15:51.207  1708  1708 I ServiceManager: addService() android.service.gatekeeper.IGateKeeperService failed (err -1 - no service manager yet?).  Retrying...
> 06-14 02:15:51.275  1685  1685 E ServiceManager: SELinux: getpidcon(pid=1693) failed to retrieve pid context.
> 06-14 02:15:51.275  1692  1692 I cameraserver: ServiceManager: 0xf6d309e0
> 06-14 02:15:51.275  1685  1685 E ServiceManager: add_service('drm.drmManager',1) uid=1019 - PERMISSION DENIED
> 06-14 02:15:51.276  1693  1693 I ServiceManager: addService() drm.drmManager failed (err -1 - no service manager yet?).  Retrying...
> 

Argh. Are you able to uncover what userspace is doing here?

So far, my test cases are:

1) self: open, write, close: allowed
2) self: open, clone thread. thread: change privileges, write, close: allowed
3) self: open, give to privileged process. privileged process: write: reject


-- 
Kees Cook

Re: [PATCH] proc: Track /proc/$pid/attr/ opener mm_struct

[Kees Cook]

On Mon, Jun 14, 2021 at 08:32:35AM -0700, Kees Cook wrote:
> On Mon, Jun 14, 2021 at 06:02:34PM +0800, youling257 wrote:
> > I used mainline kernel on android, this patch cause "failed to retrieve pid context" problem.
> > 
> > 06-14 02:15:51.165  1685  1685 E ServiceManager: SELinux: getpidcon(pid=1682) failed to retrieve pid context.

I found getpidcon() in libselinux:
https://github.com/SELinuxProject/selinux/blob/master/libselinux/src/procattr.c#L159

> > 06-14 02:15:51.166  1685  1685 E ServiceManager: add_service('batteryproperties',1) uid=0 - PERMISSION DENIED
> > 06-14 02:15:51.166  1682  1682 I ServiceManager: addService() batteryproperties failed (err -1 - no service manager yet?).  Retrying...
> > 06-14 02:15:51.197  1685  1685 E ServiceManager: SELinux: getpidcon(pid=1695) failed to retrieve pid context.
> > 06-14 02:15:51.197  1685  1685 E ServiceManager: add_service('android.security.keystore',1) uid=1017 - PERMISSION DENIED
> > 06-14 02:15:51.198  1695  1695 I ServiceManager: addService() android.security.keystore failed (err -1 - no service manager yet?).  Retrying...
> > 06-14 02:15:51.207  1685  1685 E ServiceManager: SELinux: getpidcon(pid=1708) failed to retrieve pid context.
> > 06-14 02:15:51.207  1685  1685 E ServiceManager: add_service('android.service.gatekeeper.IGateKeeperService',1) uid=1000 - PERMISSION DENIED
> > 06-14 02:15:51.207  1708  1708 I ServiceManager: addService() android.service.gatekeeper.IGateKeeperService failed (err -1 - no service manager yet?).  Retrying...
> > 06-14 02:15:51.275  1685  1685 E ServiceManager: SELinux: getpidcon(pid=1693) failed to retrieve pid context.
> > 06-14 02:15:51.275  1692  1692 I cameraserver: ServiceManager: 0xf6d309e0
> > 06-14 02:15:51.275  1685  1685 E ServiceManager: add_service('drm.drmManager',1) uid=1019 - PERMISSION DENIED
> > 06-14 02:15:51.276  1693  1693 I ServiceManager: addService() drm.drmManager failed (err -1 - no service manager yet?).  Retrying...
> > 
> 
> Argh. Are you able to uncover what userspace is doing here?

It looks like this is a case of attempting to _read_ the attr file, and
the new opener check was requiring the opener/target relationship pass
the mm_access() checks, which is clearly too strict.

> So far, my test cases are:
> 
> 1) self: open, write, close: allowed
> 2) self: open, clone thread. thread: change privileges, write, close: allowed
> 3) self: open, give to privileged process. privileged process: write: reject

I've now added:

4) self: open privileged process's attr, read, close: allowed

Can folks please test this patch to double-check?


diff --git a/fs/proc/base.c b/fs/proc/base.c
index 7118ebe38fa6..7c55301674e0 100644
--- a/fs/proc/base.c
+++ b/fs/proc/base.c
@@ -2676,7 +2676,14 @@ static int proc_pident_readdir(struct file *file, struct dir_context *ctx,
 #ifdef CONFIG_SECURITY
 static int proc_pid_attr_open(struct inode *inode, struct file *file)
 {
-	return __mem_open(inode, file, PTRACE_MODE_READ_FSCREDS);
+	struct mm_struct *mm = __mem_open(inode, file, PTRACE_MODE_READ_FSCREDS);
+
+	/* Reads do not require mm_struct access. */
+	if (IS_ERR(mm))
+		mm = NULL;
+
+	file->private_data = mm;
+	return 0;
 }
 
 static ssize_t proc_pid_attr_read(struct file * file, char __user * buf,
@@ -2709,7 +2716,7 @@ static ssize_t proc_pid_attr_write(struct file * file, const char __user * buf,
 	int rv;
 
 	/* A task may only write when it was the opener. */
-	if (file->private_data != current->mm)
+	if (!file->private_data || file->private_data != current->mm)
 		return -EPERM;
 
 	rcu_read_lock();


Wheee.

-- 
Kees Cook

Re: [PATCH] proc: Track /proc/$pid/attr/ opener mm_struct

[Linus Torvalds]

On Mon, Jun 14, 2021 at 9:45 AM Kees Cook <keescook@chromium.org> wrote:
>
>         /* A task may only write when it was the opener. */
> -       if (file->private_data != current->mm)
> +       if (!file->private_data || file->private_data != current->mm)

I don't think this is necessary.

If file->private_data is NULL, then the old test for private_data !=
current->mm will still work just fine.

Because if you can fool kernel threads to do the write for you, you
have bigger security issues than that test.

               Linus

Re: [PATCH] proc: Track /proc/$pid/attr/ opener mm_struct

[Kees Cook]

On Tue, Jun 15, 2021 at 02:46:19AM +0800, youling 257 wrote:
> I test this patch cause "init: cannot setexeccon(u:r:ueventd:s0)
> operation not permitted.
> init ctrl_write_limited.

Thanks for testing!

This appears to come from here:
https://github.com/aosp-mirror/platform_system_core/blob/master/init/service.cpp#L242


In setexeccon(), I see (pid=0, attr="exec"):

        fd = openattr(pid, attr, O_RDWR | O_CLOEXEC);
...
                        ret = write(fd, context2, strlen(context2) + 1);
...
        close(fd);


and openattr() is doing:
...
                rc = asprintf(&path, "/proc/thread-self/attr/%s", attr);
                if (rc < 0)
                        return -1;
                fd = open(path, flags | O_CLOEXEC);
...

I'm not sure how the above could fail. (mm_access() always allows
introspection...)

The only way I can understand the check failing is if a process did:

open, exec, write

But setexeccon() is not doing anything between the open and the write...

I will keep looking...

-Kees

-- 
Kees Cook

Re: [PATCH] proc: Track /proc/$pid/attr/ opener mm_struct

[youling 257]

if try to find problem on userspace, i used linux 5.13rc6 on old
android 7 cm14.1, not aosp android 11.
http://git.osdn.net/view?p=android-x86/system-core.git;a=blob;f=init/service.cpp;h=a5334f447fc2fc34453d2f6a37523bedccadc690;hb=refs/heads/cm-14.1-x86#l457

 457         if (!seclabel_.empty()) {
 458             if (setexeccon(seclabel_.c_str()) < 0) {
 459                 ERROR("cannot setexeccon('%s'): %s\n",
 460                       seclabel_.c_str(), strerror(errno));
 461                 _exit(127);
 462             }
 463         }

2021-06-15 6:50 GMT+08:00, Kees Cook <keescook@chromium.org>:
> On Tue, Jun 15, 2021 at 02:46:19AM +0800, youling 257 wrote:
>> I test this patch cause "init: cannot setexeccon(u:r:ueventd:s0)
>> operation not permitted.
>> init ctrl_write_limited.
>
> Thanks for testing!
>
> This appears to come from here:
> https://github.com/aosp-mirror/platform_system_core/blob/master/init/service.cpp#L242
>
>
> In setexeccon(), I see (pid=0, attr="exec"):
>
>         fd = openattr(pid, attr, O_RDWR | O_CLOEXEC);
> ...
>                         ret = write(fd, context2, strlen(context2) + 1);
> ...
>         close(fd);
>
>
> and openattr() is doing:
> ...
>                 rc = asprintf(&path, "/proc/thread-self/attr/%s", attr);
>                 if (rc < 0)
>                         return -1;
>                 fd = open(path, flags | O_CLOEXEC);
> ...
>
> I'm not sure how the above could fail. (mm_access() always allows
> introspection...)
>
> The only way I can understand the check failing is if a process did:
>
> open, exec, write
>
> But setexeccon() is not doing anything between the open and the write...
>
> I will keep looking...
>
> -Kees
>
> --
> Kees Cook
>

Re: [PATCH] proc: Track /proc/$pid/attr/ opener mm_struct

[Linus Torvalds]

On Mon, Jun 14, 2021 at 6:55 PM youling 257 <youling257@gmail.com> wrote:
>
> if try to find problem on userspace, i used linux 5.13rc6 on old
> android 7 cm14.1, not aosp android 11.
> http://git.osdn.net/view?p=android-x86/system-core.git;a=blob;f=init/service.cpp;h=a5334f447fc2fc34453d2f6a37523bedccadc690;hb=refs/heads/cm-14.1-x86#l457
>
>  457         if (!seclabel_.empty()) {
>  458             if (setexeccon(seclabel_.c_str()) < 0) {
>  459                 ERROR("cannot setexeccon('%s'): %s\n",
>  460                       seclabel_.c_str(), strerror(errno));
>  461                 _exit(127);
>  462             }
>  463         }

I have no idea where the cm14.1 libraries are. Does anybody know where
the matching source code for setexeccon() would be?

For me - obviously not on cm14.1 - all "setexeccon()" does is

   n = openat(AT_FDCWD, "/proc/thread-self/attr/exec", O_RDWR|O_CLOEXEC)
   write(n, string, len)
   close(n)

and if that fails, it would seem to indicate that proc_mem_open()
failed. Which would be mm_access() failing. But I don't see how that
can be the case, because mm_access() explicitly allows "mm ==
current->mm" (which the above clearly should be).

youling, can you double-check with the current -git tree? But as far
as I can tell, my minimal patch is exactly the same as Kees' patch
(just smaller and simpler).

Kees, do you see anything?

           Linus

Re: [PATCH] proc: Track /proc/$pid/attr/ opener mm_struct

[Kees Cook]

On Tue, Jun 15, 2021 at 11:19:04AM -0700, Linus Torvalds wrote:
> On Mon, Jun 14, 2021 at 6:55 PM youling 257 <youling257@gmail.com> wrote:
> >
> > if try to find problem on userspace, i used linux 5.13rc6 on old
> > android 7 cm14.1, not aosp android 11.
> > http://git.osdn.net/view?p=android-x86/system-core.git;a=blob;f=init/service.cpp;h=a5334f447fc2fc34453d2f6a37523bedccadc690;hb=refs/heads/cm-14.1-x86#l457
> >
> >  457         if (!seclabel_.empty()) {
> >  458             if (setexeccon(seclabel_.c_str()) < 0) {
> >  459                 ERROR("cannot setexeccon('%s'): %s\n",
> >  460                       seclabel_.c_str(), strerror(errno));
> >  461                 _exit(127);
> >  462             }
> >  463         }
> 
> I have no idea where the cm14.1 libraries are. Does anybody know where
> the matching source code for setexeccon() would be?
> 
> For me - obviously not on cm14.1 - all "setexeccon()" does is
> 
>    n = openat(AT_FDCWD, "/proc/thread-self/attr/exec", O_RDWR|O_CLOEXEC)
>    write(n, string, len)
>    close(n)
> 
> and if that fails, it would seem to indicate that proc_mem_open()
> failed. Which would be mm_access() failing. But I don't see how that
> can be the case, because mm_access() explicitly allows "mm ==
> current->mm" (which the above clearly should be).

Yeah, that was what I saw too.

> youling, can you double-check with the current -git tree? But as far
> as I can tell, my minimal patch is exactly the same as Kees' patch
> (just smaller and simpler).

FWIW, for that patch:

Acked-by: Kees Cook <keescook@chromium.org>
Cc: stable@vger.kernel.org

> 
> Kees, do you see anything?

No, I haven't been able to reproduce the failure. :(

-- 
Kees Cook

Re: [PATCH] proc: Track /proc/$pid/attr/ opener mm_struct

[Greg KH]

On Tue, Jun 15, 2021 at 02:50:39PM -0700, Kees Cook wrote:
> On Tue, Jun 15, 2021 at 11:19:04AM -0700, Linus Torvalds wrote:
> > On Mon, Jun 14, 2021 at 6:55 PM youling 257 <youling257@gmail.com> wrote:
> > >
> > > if try to find problem on userspace, i used linux 5.13rc6 on old
> > > android 7 cm14.1, not aosp android 11.
> > > http://git.osdn.net/view?p=android-x86/system-core.git;a=blob;f=init/service.cpp;h=a5334f447fc2fc34453d2f6a37523bedccadc690;hb=refs/heads/cm-14.1-x86#l457
> > >
> > >  457         if (!seclabel_.empty()) {
> > >  458             if (setexeccon(seclabel_.c_str()) < 0) {
> > >  459                 ERROR("cannot setexeccon('%s'): %s\n",
> > >  460                       seclabel_.c_str(), strerror(errno));
> > >  461                 _exit(127);
> > >  462             }
> > >  463         }
> > 
> > I have no idea where the cm14.1 libraries are. Does anybody know where
> > the matching source code for setexeccon() would be?
> > 
> > For me - obviously not on cm14.1 - all "setexeccon()" does is
> > 
> >    n = openat(AT_FDCWD, "/proc/thread-self/attr/exec", O_RDWR|O_CLOEXEC)
> >    write(n, string, len)
> >    close(n)
> > 
> > and if that fails, it would seem to indicate that proc_mem_open()
> > failed. Which would be mm_access() failing. But I don't see how that
> > can be the case, because mm_access() explicitly allows "mm ==
> > current->mm" (which the above clearly should be).
> 
> Yeah, that was what I saw too.
> 
> > youling, can you double-check with the current -git tree? But as far
> > as I can tell, my minimal patch is exactly the same as Kees' patch
> > (just smaller and simpler).
> 
> FWIW, for that patch:
> 
> Acked-by: Kees Cook <keescook@chromium.org>
> Cc: stable@vger.kernel.org

Thanks, I'll go pick it up now.

greg k-h

Re: [PATCH] proc: Track /proc/$pid/attr/ opener mm_struct

[youling 257]

I test "proc: only require mm_struct for writing" fixed my cm14.1 problem.

2021-06-16 2:19 GMT+08:00, Linus Torvalds <torvalds@linux-foundation.org>:
> On Mon, Jun 14, 2021 at 6:55 PM youling 257 <youling257@gmail.com> wrote:
>>
>> if try to find problem on userspace, i used linux 5.13rc6 on old
>> android 7 cm14.1, not aosp android 11.
>> http://git.osdn.net/view?p=android-x86/system-core.git;a=blob;f=init/service.cpp;h=a5334f447fc2fc34453d2f6a37523bedccadc690;hb=refs/heads/cm-14.1-x86#l457
>>
>>  457         if (!seclabel_.empty()) {
>>  458             if (setexeccon(seclabel_.c_str()) < 0) {
>>  459                 ERROR("cannot setexeccon('%s'): %s\n",
>>  460                       seclabel_.c_str(), strerror(errno));
>>  461                 _exit(127);
>>  462             }
>>  463         }
>
> I have no idea where the cm14.1 libraries are. Does anybody know where
> the matching source code for setexeccon() would be?
>
> For me - obviously not on cm14.1 - all "setexeccon()" does is
>
>    n = openat(AT_FDCWD, "/proc/thread-self/attr/exec", O_RDWR|O_CLOEXEC)
>    write(n, string, len)
>    close(n)
>
> and if that fails, it would seem to indicate that proc_mem_open()
> failed. Which would be mm_access() failing. But I don't see how that
> can be the case, because mm_access() explicitly allows "mm ==
> current->mm" (which the above clearly should be).
>
> youling, can you double-check with the current -git tree? But as far
> as I can tell, my minimal patch is exactly the same as Kees' patch
> (just smaller and simpler).
>
> Kees, do you see anything?
>
>            Linus
>

Re: [PATCH] proc: Track /proc/$pid/attr/ opener mm_struct

[Casey Schaufler]

On 6/14/2021 8:32 AM, Kees Cook wrote:
> On Mon, Jun 14, 2021 at 06:02:34PM +0800, youling257 wrote:
>> I used mainline kernel on android, this patch cause "failed to retrieve pid context" problem.
>>
>> 06-14 02:15:51.165  1685  1685 E ServiceManager: SELinux: getpidcon(pid=1682) failed to retrieve pid context.
>> 06-14 02:15:51.166  1685  1685 E ServiceManager: add_service('batteryproperties',1) uid=0 - PERMISSION DENIED
>> 06-14 02:15:51.166  1682  1682 I ServiceManager: addService() batteryproperties failed (err -1 - no service manager yet?).  Retrying...
>> 06-14 02:15:51.197  1685  1685 E ServiceManager: SELinux: getpidcon(pid=1695) failed to retrieve pid context.
>> 06-14 02:15:51.197  1685  1685 E ServiceManager: add_service('android.security.keystore',1) uid=1017 - PERMISSION DENIED
>> 06-14 02:15:51.198  1695  1695 I ServiceManager: addService() android.security.keystore failed (err -1 - no service manager yet?).  Retrying...
>> 06-14 02:15:51.207  1685  1685 E ServiceManager: SELinux: getpidcon(pid=1708) failed to retrieve pid context.
>> 06-14 02:15:51.207  1685  1685 E ServiceManager: add_service('android.service.gatekeeper.IGateKeeperService',1) uid=1000 - PERMISSION DENIED
>> 06-14 02:15:51.207  1708  1708 I ServiceManager: addService() android.service.gatekeeper.IGateKeeperService failed (err -1 - no service manager yet?).  Retrying...
>> 06-14 02:15:51.275  1685  1685 E ServiceManager: SELinux: getpidcon(pid=1693) failed to retrieve pid context.
>> 06-14 02:15:51.275  1692  1692 I cameraserver: ServiceManager: 0xf6d309e0
>> 06-14 02:15:51.275  1685  1685 E ServiceManager: add_service('drm.drmManager',1) uid=1019 - PERMISSION DENIED
>> 06-14 02:15:51.276  1693  1693 I ServiceManager: addService() drm.drmManager failed (err -1 - no service manager yet?).  Retrying...
>>
> Argh. Are you able to uncover what userspace is doing here?
>
> So far, my test cases are:
>
> 1) self: open, write, close: allowed
> 2) self: open, clone thread. thread: change privileges, write, close: allowed
> 3) self: open, give to privileged process. privileged process: write: reject

I found an issue under Smack where a privileged process opened
/proc/self/attr/smack/current, wrote to it successfully, then tried
to write to it again, which failed because the cred has changed. 
That's not a common use case. The usual case is open, write, close.
If ServiceManager is assuming that it can leave a descriptor open
while manipulations are in progress it could encounter the same kind
of problem.