Skip to contentSkip to conversation

mm/mempolicy: kernel_migrate_pages: fix race between security checks and suid exec

PATCH In review Oleg Nesterov 2 participants 10 messages May 24 – May 26

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover mm/mempolicy: kernel_migrate_pages: fix race between security checks and suid exec

Hello,

Sorry, I have no idea how to test these changes. Please review. Hopefully sashiko.dev will take a look too ;)

find_mm_struct() has the same problem, will send a patch tomorrow. Unless you tell me this race doesn't need a fix.

Oleg.
---

 mm/mempolicy.c | 53 +++++++++++++++++++++++++----------------------------
 1 file changed, 25 insertions(+), 28 deletions(-)
#1 lore raw

PATCH 1/4 mm/mempolicy: kernel_migrate_pages: simplify the usage of put_task_struct()

kernel_migrate_pages() calls put_task_struct() twice; this complicates the code. Move put_task_struct() to the unified exit path. This also allows us to move the "out_put" label up and avoid another goto.

---
 mm/mempolicy.c | 11 +++--------
 1 file changed, 3 insertions(+), 8 deletions(-)
mm/mempolicy.c +3 -8
index 4e4421b22b59..4326dff16aa6 100644
@@ -1922,25 +1922,20 @@ static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
 		goto out_put;
 
 	mm = get_task_mm(task);
-	put_task_struct(task);
-
 	if (!mm) {
 		err = -EINVAL;
-		goto out;
+		goto out_put;
 	}
 
 	err = do_migrate_pages(mm, old, new,
 		capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
 
 	mmput(mm);
+out_put:
+	put_task_struct(task);
 out:
 	NODEMASK_SCRATCH_FREE(scratch);
-
 	return err;
-
-out_put:
-	put_task_struct(task);
-	goto out;
 }
 
 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
Signature · 1 line
2.52.0
maintainer Reviewed-by on v1 1/4
#2 lore raw

Re: [PATCH 1/4] mm/mempolicy: kernel_migrate_pages: simplify the usage of put_task_struct()

Quoted 5 lines · Oleg Nesterov

kernel_migrate_pages() calls put_task_struct() twice; this complicates the code. Move put_task_struct() to the unified exit path. This also allows us to move the "out_put" label up and avoid another goto.

Quoted 40 lines
---
 mm/mempolicy.c | 11 +++--------
 1 file changed, 3 insertions(+), 8 deletions(-)
mm/mempolicy.c +3 -8
index 4e4421b22b59..4326dff16aa6 100644
@@ -1922,25 +1922,20 @@ static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
 		goto out_put;
 
 	mm = get_task_mm(task);
-	put_task_struct(task);
-
 	if (!mm) {
 		err = -EINVAL;
-		goto out;
+		goto out_put;
 	}
 
 	err = do_migrate_pages(mm, old, new,
 		capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
 
 	mmput(mm);
+out_put:
+	put_task_struct(task);
 out:
 	NODEMASK_SCRATCH_FREE(scratch);
-
 	return err;
-
-out_put:
-	put_task_struct(task);
-	goto out;
 }
 
 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
Signature · 1 line
2.52.0
maintainer
#3 lore raw

Re: [PATCH 1/4] mm/mempolicy: kernel_migrate_pages: simplify the usage of put_task_struct()

In reply to Oleg Nesterov
Quoted 14 lines · Oleg Nesterov

kernel_migrate_pages() calls put_task_struct() twice; this complicates the code. Move put_task_struct() to the unified exit path. This also allows us to move the "out_put" label up and avoid another goto.

---
 mm/mempolicy.c | 11 +++--------
 1 file changed, 3 insertions(+), 8 deletions(-)
mm/mempolicy.c +0 -0
index 4e4421b22b59..4326dff16aa6 100644

mm/mempolicy.c static int kernel_migrate_pages(pid_t pid, unsigned long maxnode, · line 1926

Quoted 5 lines
mm/mempolicy.c +0 -2
@@ -1922,25 +1922,20 @@ static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
 		goto out_put;
 
 	mm = get_task_mm(task);
-	put_task_struct(task);
-

Hm, I was a little trigger-happy on that review, I just realized that this increases the scope in which there is a task reference held (i.e. we're now holding a reference across do_migrate_pages())

I don't *think* that's an issue, but it might we worth looking at the changelog to see if this was done intentionally.

Quoted 26 lines
mm/mempolicy.c +3 -6
 	if (!mm) {
 		err = -EINVAL;
-		goto out;
+		goto out_put;
 	}
 
 	err = do_migrate_pages(mm, old, new,
 		capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
 
 	mmput(mm);
+out_put:
+	put_task_struct(task);
 out:
 	NODEMASK_SCRATCH_FREE(scratch);
-
 	return err;
-
-out_put:
-	put_task_struct(task);
-	goto out;
 }
 
 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
Signature · 1 line
2.52.0
#4 lore raw

Re: [PATCH 1/4] mm/mempolicy: kernel_migrate_pages: simplify the usage of put_task_struct()

Quoted 11 lines · Gregory Price

mm/mempolicy.c static int kernel_migrate_pages(pid_t pid, unsigned long maxnode, · line 1926

Quoted 4 lines · Oleg Nesterov
mm/mempolicy.c +0 -2
 	mm = get_task_mm(task);
-	put_task_struct(task);
-

Hm, I was a little trigger-happy on that review, I just realized that this increases the scope in which there is a task reference held (i.e. we're now holding a reference across do_migrate_pages())

Quoted 2 lines

I don't *think* that's an issue, but it might we worth looking at the changelog to see if this was done intentionally.

I did this to simplify the code and the next changes. We pin this task's memory across do_migrate_pages(), so I think that delaying put_task_struct() doesn't add too much.

But of course, this is not necessary, I can rework if you don't like it.

Thanks for review!

Oleg.

maintainer
#5 lore raw

Re: [PATCH 1/4] mm/mempolicy: kernel_migrate_pages: simplify the usage of put_task_struct()

Quoted 12 lines · Oleg Nesterov
Quoted 11 lines · Gregory Price

mm/mempolicy.c static int kernel_migrate_pages(pid_t pid, unsigned long maxnode, · line 1926

Quoted 4 lines · Oleg Nesterov
mm/mempolicy.c +0 -2
 	mm = get_task_mm(task);
-	put_task_struct(task);
-

Hm, I was a little trigger-happy on that review, I just realized that this increases the scope in which there is a task reference held (i.e. we're now holding a reference across do_migrate_pages())

Quoted 9 lines
Quoted 2 lines

I don't *think* that's an issue, but it might we worth looking at the changelog to see if this was done intentionally.

I did this to simplify the code and the next changes. We pin this task's memory across do_migrate_pages(), so I think that delaying put_task_struct() doesn't add too much.

But of course, this is not necessary, I can rework if you don't like it.

Not necessary, just a curiosity whether holding the reference across the larger scope has any functional difference. Like I said, I don't think that's an issue, but might be worth just checking the history on the function to see if it was done that way on purpose.

Quoted 4 lines

Thanks for review!

Oleg.

#6 lore raw

PATCH 2/4 mm/mempolicy: kernel_migrate_pages: use find_get_task_by_vpid()

In reply to Oleg Nesterov

kernel_migrate_pages() calls ptrace_may_access() under rcu_read_lock() for no reason. This is a leftover from before the commit 313674661925 ("Unify migrate_pages and move_pages access checks"), where rcu_read_lock() was needed to protect __task_cred(task).

So we can drop the RCU lock right after get_task_struct(). Better yet, if pid != 0, we can use find_get_task_by_vpid() which does get_task_struct() itself, and avoid get/put_task_struct() otherwise.

---
 mm/mempolicy.c | 11 +++--------
 1 file changed, 3 insertions(+), 8 deletions(-)
mm/mempolicy.c +3 -8
index 4326dff16aa6..2ec14001e4dc 100644
@@ -1883,15 +1883,11 @@ static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
 	if (err)
 		goto out;
 
-	/* Find the mm_struct */
-	rcu_read_lock();
-	task = pid ? find_task_by_vpid(pid) : current;
+	task = pid ? find_get_task_by_vpid(pid) : current;
 	if (!task) {
-		rcu_read_unlock();
 		err = -ESRCH;
 		goto out;
 	}
-	get_task_struct(task);
 
 	err = -EINVAL;
 
@@ -1900,11 +1896,9 @@ static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
 	 * Use the regular "ptrace_may_access()" checks.
 	 */
 	if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
-		rcu_read_unlock();
 		err = -EPERM;
 		goto out_put;
 	}
-	rcu_read_unlock();
 
 	task_nodes = cpuset_mems_allowed(task);
 	/* Is the user allowed to access the target nodes? */
@@ -1932,7 +1926,8 @@ static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
 
 	mmput(mm);
 out_put:
-	put_task_struct(task);
+	if (pid)
+		put_task_struct(task);
 out:
 	NODEMASK_SCRATCH_FREE(scratch);
 	return err;
Signature · 1 line
2.52.0
#7 lore raw

PATCH 3/4 mm/mempolicy: kernel_migrate_pages: check ptrace_may_access() after nodes_and()

In reply to Oleg Nesterov

Move the ptrace_may_access() check down, after the nodemask validation. This is just a preparation to simplify the next patch.

Note that because of this reordering, if the caller passes an invalid nodemask and also lacks permissions, the syscall will now return -EINVAL instead of -EPERM. Hopefully, nobody relies on the old error precedence.

---
 mm/mempolicy.c | 24 ++++++++++++------------
 1 file changed, 12 insertions(+), 12 deletions(-)
mm/mempolicy.c +12 -12
index 2ec14001e4dc..c09ff9f9aa96 100644
@@ -1889,17 +1889,6 @@ static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
 		goto out;
 	}
 
-	err = -EINVAL;
-
-	/*
-	 * Check if this process has the right to modify the specified process.
-	 * Use the regular "ptrace_may_access()" checks.
-	 */
-	if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
-		err = -EPERM;
-		goto out_put;
-	}
-
 	task_nodes = cpuset_mems_allowed(task);
 	/* Is the user allowed to access the target nodes? */
 	if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
@@ -1908,8 +1897,19 @@ static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
 	}
 
 	task_nodes = cpuset_mems_allowed(current);
-	if (!nodes_and(*new, *new, task_nodes))
+	if (!nodes_and(*new, *new, task_nodes)) {
+		err = -EINVAL;
 		goto out_put;
+	}
+
+	/*
+	 * Check if this process has the right to modify the specified process.
+	 * Use the regular "ptrace_may_access()" checks.
+	 */
+	if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
+		err = -EPERM;
+		goto out_put;
+	}
 
 	err = security_task_movememory(task);
 	if (err)
Signature · 1 line
2.52.0
#8 lore raw

PATCH 4/4 mm/mempolicy: kernel_migrate_pages: fix race between security checks and suid exec

In reply to Oleg Nesterov

The target task can execute a setuid binary between ptrace_may_access() and get_task_mm(). Protect this critical section with exec_update_lock.

Sadly, we don't have DEFINE_LOCK_GUARD_1_COND(rwsem_read, _kill) yet.

---
 mm/mempolicy.c | 13 ++++++++++---
 1 file changed, 10 insertions(+), 3 deletions(-)
mm/mempolicy.c +10 -3
index c09ff9f9aa96..cbb45a876a93 100644
@@ -1902,24 +1902,31 @@ static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
 		goto out_put;
 	}
 
+	err = down_read_killable(&task->signal->exec_update_lock);
+	if (err)
+		goto out_put;
 	/*
 	 * Check if this process has the right to modify the specified process.
 	 * Use the regular "ptrace_may_access()" checks.
 	 */
 	if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
 		err = -EPERM;
-		goto out_put;
+		goto unlock;
 	}
 
 	err = security_task_movememory(task);
 	if (err)
-		goto out_put;
+		goto unlock;
 
 	mm = get_task_mm(task);
 	if (!mm) {
 		err = -EINVAL;
-		goto out_put;
+		goto unlock;
 	}
+unlock:
+	up_read(&task->signal->exec_update_lock);
+	if (err)
+		goto out_put;
 
 	err = do_migrate_pages(mm, old, new,
 		capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
Signature · 1 line
2.52.0
#9 lore raw
In reply to Oleg Nesterov
Quoted 2 lines · Oleg Nesterov

Hopefully sashiko.dev will take a look too ;)

See https://sashiko.dev/#/patchset/ahMt6xyUNnacZU8-%40redhat.com
Thanks Sashiko, let me reply.

As for reply to 3/4:

I honestly think this is not a problem... but a) I will think again, and b) I rely on review from mm experts anyway.

The main motivation for this patch is that we can create a common helper for find_mm_struct() and kernel_migrate_pages(), but it is not strictly necessary.

As for reply to 4/4:

Yes! And I have already mentioned this in 0/4 and in the previous discussions.

Plus (say) proc_coredump_filter_write() which relies on the open-time checks and don't use ptrace_may_access() at all. This looks wrong.

But this need other changes/discussions, so I think we can ignore this comment.

Oleg.