[REVIEW PATCH v2] ptrace: reintroduce usage of subjective credentials in ptrace_has_cap()

[REVIEW PATCH v2] ptrace: reintroduce usage of subjective credentials in ptrace_has_cap()

[Christian Brauner]

Commit 69f594a38967 ("ptrace: do not audit capability check when outputing /proc/pid/stat")
introduced the ability to opt out of audit messages for accesses to
various proc files since they are not violations of policy.
While doing so it somehow switched the check from ns_capable() to
has_ns_capability{_noaudit}(). That means it switched from checking the
subjective credentials of the task to using the objective credentials. I
couldn't find the original lkml thread and so I don't know why this switch
was done. But it seems wrong since ptrace_has_cap() is currently only used
in ptrace_may_access(). And it's used to check whether the calling task
(subject) has the CAP_SYS_PTRACE capability in the provided user namespace
to operate on the target task (object). According to the cred.h comments
this would mean the subjective credentials of the calling task need to be
used.
This switches it to use security_capable() because we only call
ptrace_has_cap() in ptrace_may_access() and in there we already have a
stable reference to the calling tasks creds under cred_guard_mutex so
there's no need to go through another series of dereferences and rcu
locking done in ns_capable{_noaudit}().

As one example where this might be particularly problematic, Jann pointed
out that in combination with the upcoming IORING_OP_OPENAT feature, this
bug might allow unprivileged users to bypass the capability checks while
asynchronously opening files like /proc/*/mem, because the capability
checks for this would be performed against kernel credentials.

Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Eric Paris <eparis@redhat.com>
Cc: stable@vger.kernel.org
Reviewed-by: Serge Hallyn <serge@hallyn.com>
Reviewed-by: Jann Horn <jannh@google.com>
Fixes: 69f594a38967 ("ptrace: do not audit capability check when outputing /proc/pid/stat")
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
---
 kernel/ptrace.c | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/kernel/ptrace.c b/kernel/ptrace.c
index cb9ddcc08119..d146133e97f1 100644
--- a/kernel/ptrace.c
+++ b/kernel/ptrace.c
@@ -264,12 +264,13 @@ static int ptrace_check_attach(struct task_struct *child, bool ignore_state)
 	return ret;
 }
 
-static int ptrace_has_cap(struct user_namespace *ns, unsigned int mode)
+static int ptrace_has_cap(const struct cred *cred, struct user_namespace *ns,
+			  unsigned int mode)
 {
 	if (mode & PTRACE_MODE_NOAUDIT)
-		return has_ns_capability_noaudit(current, ns, CAP_SYS_PTRACE);
+		return security_capable(cred, ns, CAP_SYS_PTRACE, CAP_OPT_NOAUDIT);
 	else
-		return has_ns_capability(current, ns, CAP_SYS_PTRACE);
+		return security_capable(cred, ns, CAP_SYS_PTRACE, CAP_OPT_NONE);
 }
 
 /* Returns 0 on success, -errno on denial. */
@@ -321,7 +322,7 @@ static int __ptrace_may_access(struct task_struct *task, unsigned int mode)
 	    gid_eq(caller_gid, tcred->sgid) &&
 	    gid_eq(caller_gid, tcred->gid))
 		goto ok;
-	if (ptrace_has_cap(tcred->user_ns, mode))
+	if (ptrace_has_cap(cred, tcred->user_ns, mode))
 		goto ok;
 	rcu_read_unlock();
 	return -EPERM;
@@ -340,7 +341,7 @@ static int __ptrace_may_access(struct task_struct *task, unsigned int mode)
 	mm = task->mm;
 	if (mm &&
 	    ((get_dumpable(mm) != SUID_DUMP_USER) &&
-	     !ptrace_has_cap(mm->user_ns, mode)))
+	     !ptrace_has_cap(cred, mm->user_ns, mode)))
 	    return -EPERM;
 
 	return security_ptrace_access_check(task, mode);

base-commit: b3a987b0264d3ddbb24293ebff10eddfc472f653
-- 
2.25.0

Re: [REVIEW PATCH v2] ptrace: reintroduce usage of subjective credentials in ptrace_has_cap()

[Kees Cook]

On Thu, Jan 16, 2020 at 11:45:18PM +0100, Christian Brauner wrote:
> Commit 69f594a38967 ("ptrace: do not audit capability check when outputing /proc/pid/stat")
> introduced the ability to opt out of audit messages for accesses to
> various proc files since they are not violations of policy.
> While doing so it somehow switched the check from ns_capable() to
> has_ns_capability{_noaudit}(). That means it switched from checking the
> subjective credentials of the task to using the objective credentials. I
> couldn't find the original lkml thread and so I don't know why this switch
> was done. But it seems wrong since ptrace_has_cap() is currently only used
> in ptrace_may_access(). And it's used to check whether the calling task
> (subject) has the CAP_SYS_PTRACE capability in the provided user namespace
> to operate on the target task (object). According to the cred.h comments
> this would mean the subjective credentials of the calling task need to be
> used.

I don't follow this description. As far as I can see, both the current
code and your patch end up using current's cred, yes? I'm not following
the subjective/objective change mentioned here.

Before:
bool has_ns_capability(struct task_struct *t,
                       struct user_namespace *ns, int cap)
{
        int ret;

        rcu_read_lock();
        ret = security_capable(__task_cred(t), ns, cap, CAP_OPT_NONE);
        rcu_read_unlock();

        return (ret == 0);
}
...
		return has_ns_capability(current, ns, CAP_SYS_PTRACE)

After:
	const struct cred *cred = current_cred(), ...
...
		return security_capable(cred, ns, CAP_SYS_PTRACE, CAP_OPT_NOAUDIT);

The cred passed to security_capable() is the subject before and after.

> This switches it to use security_capable() because we only call
> ptrace_has_cap() in ptrace_may_access() and in there we already have a
> stable reference to the calling tasks creds under cred_guard_mutex so
> there's no need to go through another series of dereferences and rcu
> locking done in ns_capable{_noaudit}().

This makes sense to me -- now there's no possible race on the cred
changing between the two ptrace_has_cap() checks, yes?

However, I'm still trying to see where cred_guard_mutex() comes into
play for callers of ptrace_may_access(). I see it for the object
("task" arg in ptrace_may_access()), but if this is dealing with the cred
on current, it's just the RCU read lock protecting it (which I think is
fine here), but seems confusing in the commit log.

> As one example where this might be particularly problematic, Jann pointed
> out that in combination with the upcoming IORING_OP_OPENAT feature, this
> bug might allow unprivileged users to bypass the capability checks while
> asynchronously opening files like /proc/*/mem, because the capability
> checks for this would be performed against kernel credentials.

As in, winning a race between the two ptrace_has_cap() calls across a
cred transition?

> Cc: Oleg Nesterov <oleg@redhat.com>
> Cc: Eric Paris <eparis@redhat.com>
> Cc: stable@vger.kernel.org
> Reviewed-by: Serge Hallyn <serge@hallyn.com>
> Reviewed-by: Jann Horn <jannh@google.com>
> Fixes: 69f594a38967 ("ptrace: do not audit capability check when outputing /proc/pid/stat")
> Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
> ---
>  kernel/ptrace.c | 11 ++++++-----
>  1 file changed, 6 insertions(+), 5 deletions(-)
> 
> diff --git a/kernel/ptrace.c b/kernel/ptrace.c
> index cb9ddcc08119..d146133e97f1 100644
> --- a/kernel/ptrace.c
> +++ b/kernel/ptrace.c
> @@ -264,12 +264,13 @@ static int ptrace_check_attach(struct task_struct *child, bool ignore_state)
>  	return ret;
>  }
>  
> -static int ptrace_has_cap(struct user_namespace *ns, unsigned int mode)
> +static int ptrace_has_cap(const struct cred *cred, struct user_namespace *ns,
> +			  unsigned int mode)
>  {
>  	if (mode & PTRACE_MODE_NOAUDIT)
> -		return has_ns_capability_noaudit(current, ns, CAP_SYS_PTRACE);
> +		return security_capable(cred, ns, CAP_SYS_PTRACE, CAP_OPT_NOAUDIT);
>  	else
> -		return has_ns_capability(current, ns, CAP_SYS_PTRACE);
> +		return security_capable(cred, ns, CAP_SYS_PTRACE, CAP_OPT_NONE);
>  }

Style nit -- can we just make this a single invocation of
security_capable(), something like:

	return security_capable(cred, ns, CAP_SYS_PTRACE,
				mode & PTRACE_MODE_NOAUDIT
					?  CAP_OPT_NOAUDIT,
					: CAP_OPT_NONE) == 0;

Obviously not required, but the longer if hurts my eyes. ;)

>  
>  /* Returns 0 on success, -errno on denial. */
> @@ -321,7 +322,7 @@ static int __ptrace_may_access(struct task_struct *task, unsigned int mode)
>  	    gid_eq(caller_gid, tcred->sgid) &&
>  	    gid_eq(caller_gid, tcred->gid))
>  		goto ok;
> -	if (ptrace_has_cap(tcred->user_ns, mode))
> +	if (ptrace_has_cap(cred, tcred->user_ns, mode))
>  		goto ok;
>  	rcu_read_unlock();
>  	return -EPERM;
> @@ -340,7 +341,7 @@ static int __ptrace_may_access(struct task_struct *task, unsigned int mode)
>  	mm = task->mm;
>  	if (mm &&
>  	    ((get_dumpable(mm) != SUID_DUMP_USER) &&
> -	     !ptrace_has_cap(mm->user_ns, mode)))
> +	     !ptrace_has_cap(cred, mm->user_ns, mode)))
>  	    return -EPERM;
>  
>  	return security_ptrace_access_check(task, mode);
> 
> base-commit: b3a987b0264d3ddbb24293ebff10eddfc472f653
> -- 
> 2.25.0
> 

So, I think this change looks correct, but I find the commit subject
and log confusing (perhaps because I am dense) and misleading (again,
perhaps because I am dense).

-- 
Kees Cook

Re: [REVIEW PATCH v2] ptrace: reintroduce usage of subjective credentials in ptrace_has_cap()

[Christian Brauner]

On Thu, Jan 16, 2020 at 06:29:26PM -0800, Kees Cook wrote:
> On Thu, Jan 16, 2020 at 11:45:18PM +0100, Christian Brauner wrote:
> > Commit 69f594a38967 ("ptrace: do not audit capability check when outputing /proc/pid/stat")
> > introduced the ability to opt out of audit messages for accesses to
> > various proc files since they are not violations of policy.
> > While doing so it somehow switched the check from ns_capable() to
> > has_ns_capability{_noaudit}(). That means it switched from checking the
> > subjective credentials of the task to using the objective credentials. I
> > couldn't find the original lkml thread and so I don't know why this switch
> > was done. But it seems wrong since ptrace_has_cap() is currently only used
> > in ptrace_may_access(). And it's used to check whether the calling task
> > (subject) has the CAP_SYS_PTRACE capability in the provided user namespace
> > to operate on the target task (object). According to the cred.h comments
> > this would mean the subjective credentials of the calling task need to be
> > used.
> 
> I don't follow this description. As far as I can see, both the current
> code and your patch end up using current's cred, yes? I'm not following
> the subjective/objective change mentioned here.
> 
> Before:
> bool has_ns_capability(struct task_struct *t,
>                        struct user_namespace *ns, int cap)
> {
>         int ret;
> 
>         rcu_read_lock();
>         ret = security_capable(__task_cred(t), ns, cap, CAP_OPT_NONE);

If I'm not mistaken, you're looking at the cuplrit: "__task_cred()":

 /**
 * __task_cred - Access a task's objective credentials
 * @task: The task to query
 *
 * Access the objective credentials of a task.  The caller must hold the RCU
 * readlock.
 *
 * The result of this function should not be passed directly to get_cred();
 * rather get_task_cred() should be used instead.
 */
#define __task_cred(task)	\
	rcu_dereference((task)->real_cred)

>         rcu_read_unlock();
> 
>         return (ret == 0);
> }
> ...
> 		return has_ns_capability(current, ns, CAP_SYS_PTRACE)
> 
> After:
> 	const struct cred *cred = current_cred(), ...
> ...
> 		return security_capable(cred, ns, CAP_SYS_PTRACE, CAP_OPT_NOAUDIT);
> 
> The cred passed to security_capable() is the subject before and after.
> 
> > This switches it to use security_capable() because we only call
> > ptrace_has_cap() in ptrace_may_access() and in there we already have a
> > stable reference to the calling tasks creds under cred_guard_mutex so
> > there's no need to go through another series of dereferences and rcu
> > locking done in ns_capable{_noaudit}().
> 
> This makes sense to me -- now there's no possible race on the cred
> changing between the two ptrace_has_cap() checks, yes?
> 
> However, I'm still trying to see where cred_guard_mutex() comes into
> play for callers of ptrace_may_access(). I see it for the object
> ("task" arg in ptrace_may_access()), but if this is dealing with the cred
> on current, it's just the RCU read lock protecting it (which I think is
> fine here), but seems confusing in the commit log.

Ah, right. I'll drop that from the commit message and place in the rcu
lock.

Christian

Re: [REVIEW PATCH v2] ptrace: reintroduce usage of subjective credentials in ptrace_has_cap()

[Christian Brauner]

On Thu, Jan 16, 2020 at 06:29:26PM -0800, Kees Cook wrote:
> On Thu, Jan 16, 2020 at 11:45:18PM +0100, Christian Brauner wrote:
> > As one example where this might be particularly problematic, Jann pointed
> > out that in combination with the upcoming IORING_OP_OPENAT feature, this
> > bug might allow unprivileged users to bypass the capability checks while
> > asynchronously opening files like /proc/*/mem, because the capability
> > checks for this would be performed against kernel credentials.

To follow up on this part of your mail. No, afaict, it's not
aboutwinning a race. It's way simpler...
When io uring creates a new kernel context it records the subjective
credentials of the caller:

	ctx = io_ring_ctx_alloc(p);
	if (!ctx) {
		if (account_mem)
			io_unaccount_mem(user, ring_pages(p->sq_entries,
								p->cq_entries));
		free_uid(user);
		return -ENOMEM;
	}
	ctx->compat = in_compat_syscall();
	ctx->account_mem = account_mem;
	ctx->user = user;
------> ctx->creds = get_current_cred(); <------

Later on, when it starts to do work it creates a kernel thread:

			ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
							ctx, cpu,
							"io_uring-sq");
		} else {
			ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
							"io_uring-sq");
		}

and registers io_sq_thread as "callback". The callback io_sq_thread()
runs __with kernel creds__. To prevent this from becoming an issue
io_sq_thread() will override the __subjective credentials__ with the
callers credentials:

	old_cred = override_creds(ctx->creds);

But ptrace_has_cap() currently looks at __task_cred(current) aka
__real_cred__. This means once IORING_OP_OPENAT and IORING_OP_OPENAT2
lands in v5.5-rc6 it is more or less trivial for an unprivileged user to
bypass ptrace_may_access().

Christian

Re: [REVIEW PATCH v2] ptrace: reintroduce usage of subjective credentials in ptrace_has_cap()

[Kees Cook]

On Fri, Jan 17, 2020 at 06:16:23AM +0100, Christian Brauner wrote:
> On Thu, Jan 16, 2020 at 06:29:26PM -0800, Kees Cook wrote:
> > On Thu, Jan 16, 2020 at 11:45:18PM +0100, Christian Brauner wrote:
> > > Commit 69f594a38967 ("ptrace: do not audit capability check when outputing /proc/pid/stat")
> > > introduced the ability to opt out of audit messages for accesses to
> > > various proc files since they are not violations of policy.
> > > While doing so it somehow switched the check from ns_capable() to
> > > has_ns_capability{_noaudit}(). That means it switched from checking the
> > > subjective credentials of the task to using the objective credentials. I
> > > couldn't find the original lkml thread and so I don't know why this switch
> > > was done. But it seems wrong since ptrace_has_cap() is currently only used
> > > in ptrace_may_access(). And it's used to check whether the calling task
> > > (subject) has the CAP_SYS_PTRACE capability in the provided user namespace
> > > to operate on the target task (object). According to the cred.h comments
> > > this would mean the subjective credentials of the calling task need to be
> > > used.
> > 
> > I don't follow this description. As far as I can see, both the current
> > code and your patch end up using current's cred, yes? I'm not following
> > the subjective/objective change mentioned here.
> > 
> > Before:
> > bool has_ns_capability(struct task_struct *t,
> >                        struct user_namespace *ns, int cap)
> > {
> >         int ret;
> > 
> >         rcu_read_lock();
> >         ret = security_capable(__task_cred(t), ns, cap, CAP_OPT_NONE);
> 
> If I'm not mistaken, you're looking at the cuplrit: "__task_cred()":
> [...]
> #define __task_cred(task)	\
> 	rcu_dereference((task)->real_cred)

Ah! Yes, thank you. cred vs real_cred. That's what I missed!

> > However, I'm still trying to see where cred_guard_mutex() comes into
> > play for callers of ptrace_may_access(). I see it for the object
> > ("task" arg in ptrace_may_access()), but if this is dealing with the cred
> > on current, it's just the RCU read lock protecting it (which I think is
> > fine here), but seems confusing in the commit log.
> 
> Ah, right. I'll drop that from the commit message and place in the rcu
> lock.

Thanks for the clarification. With that adjusted, please consider it:

Reviewed-by: Kees Cook <keescook@chromium.org>

(I wonder how hard it might be to build some self-tests for this to
catch future glitches...)

-- 
Kees Cook