[PATCH] Fix: membarrier: racy access to p->mm in membarrier_global_expedited()

Jann Horn identified a racy access to p->mm in the global expedited
command of the membarrier system call.

The suggested fix is to hold the task_lock() around the accesses to
p->mm and to the mm_struct membarrier_state field to guarantee the
existence of the mm_struct.

Link: https://lore.kernel.org/lkml/CAG48ez2G8ctF8dHS42TF37pThfr3y0RNOOYTmxvACm4u8Yu3cw@mail.gmail.com
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Tested-by: Jann Horn <jannh@google.com>
CC: Jann Horn <jannh@google.com>
CC: Thomas Gleixner <redacted>
CC: Peter Zijlstra (Intel) <peterz@infradead.org>
CC: Ingo Molnar <mingo@kernel.org>
CC: Andrea Parri <parri.andrea@gmail.com>
CC: Andy Lutomirski <luto@kernel.org>
CC: Avi Kivity <redacted>
CC: Benjamin Herrenschmidt <benh@kernel.crashing.org>
CC: Boqun Feng <redacted>
CC: Dave Watson <redacted>
CC: David Sehr <redacted>
CC: H. Peter Anvin <hpa@zytor.com>
CC: Linus Torvalds <torvalds@linux-foundation.org>
CC: Maged Michael <redacted>
CC: Michael Ellerman <mpe@ellerman.id.au>
CC: Paul E. McKenney <redacted>
CC: Paul Mackerras <redacted>
CC: Russell King <linux@armlinux.org.uk>
CC: Will Deacon <redacted>
CC: stable@vger.kernel.org # v4.16+
CC: linux-api@vger.kernel.org
---
 kernel/sched/membarrier.c | 27 +++++++++++++++++++++------
 1 file changed, 21 insertions(+), 6 deletions(-)

diff --git a/kernel/sched/membarrier.c b/kernel/sched/membarrier.c
index 76e0eaf4654e..305fdcc4c5f7 100644
--- a/kernel/sched/membarrier.c
+++ b/kernel/sched/membarrier.c

@@ -81,12 +81,27 @@ static int membarrier_global_expedited(void)
 
 		rcu_read_lock();
 		p = task_rcu_dereference(&cpu_rq(cpu)->curr);
-		if (p && p->mm && (atomic_read(&p->mm->membarrier_state) &
-				   MEMBARRIER_STATE_GLOBAL_EXPEDITED)) {
-			if (!fallback)
-				__cpumask_set_cpu(cpu, tmpmask);
-			else
-				smp_call_function_single(cpu, ipi_mb, NULL, 1);
+		/*
+		 * Skip this CPU if the runqueue's current task is NULL or if
+		 * it is a kernel thread.
+		 */
+		if (p && READ_ONCE(p->mm)) {
+			bool mm_match;
+
+			/*
+			 * Read p->mm and access membarrier_state while holding
+			 * the task lock to ensure existence of mm.
+			 */
+			task_lock(p);
+			mm_match = p->mm && (atomic_read(&p->mm->membarrier_state) &
+					     MEMBARRIER_STATE_GLOBAL_EXPEDITED);
+			task_unlock(p);
+			if (mm_match) {
+				if (!fallback)
+					__cpumask_set_cpu(cpu, tmpmask);
+				else
+					smp_call_function_single(cpu, ipi_mb, NULL, 1);
+			}
 		}
 		rcu_read_unlock();
 	}

-- 
2.17.1

On Mon, Jan 28, 2019 at 05:07:07PM -0500, Mathieu Desnoyers wrote:

Are we guaranteed that this p->mm will be the same as the one loaded via
READ_ONCE() above?  Either way, wouldn't it be better to READ_ONCE() it a
single time and use the same value everywhere?

							Thanx, Paul

+					     MEMBARRIER_STATE_GLOBAL_EXPEDITED);
+			task_unlock(p);
+			if (mm_match) {
+				if (!fallback)
+					__cpumask_set_cpu(cpu, tmpmask);
+				else
+					smp_call_function_single(cpu, ipi_mb, NULL, 1);
+			}
 		}
 		rcu_read_unlock();
 	}
-- 
2.17.1

On Mon, Jan 28, 2019 at 11:39 PM Paul E. McKenney [off-list ref] wrote:

On Mon, Jan 28, 2019 at 05:07:07PM -0500, Mathieu Desnoyers wrote:

quoted

Jann Horn identified a racy access to p->mm in the global expedited
command of the membarrier system call.

The suggested fix is to hold the task_lock() around the accesses to
p->mm and to the mm_struct membarrier_state field to guarantee the
existence of the mm_struct.

Link: https://lore.kernel.org/lkml/CAG48ez2G8ctF8dHS42TF37pThfr3y0RNOOYTmxvACm4u8Yu3cw@mail.gmail.com
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

[...]

quoted

--- a/kernel/sched/membarrier.c
+++ b/kernel/sched/membarrier.c

@@ -81,12 +81,27 @@ static int membarrier_global_expedited(void)

              rcu_read_lock();
              p = task_rcu_dereference(&cpu_rq(cpu)->curr);
-             if (p && p->mm && (atomic_read(&p->mm->membarrier_state) &
-                                MEMBARRIER_STATE_GLOBAL_EXPEDITED)) {
-                     if (!fallback)
-                             __cpumask_set_cpu(cpu, tmpmask);
-                     else
-                             smp_call_function_single(cpu, ipi_mb, NULL, 1);
+             /*
+              * Skip this CPU if the runqueue's current task is NULL or if
+              * it is a kernel thread.
+              */
+             if (p && READ_ONCE(p->mm)) {
+                     bool mm_match;
+
+                     /*
+                      * Read p->mm and access membarrier_state while holding
+                      * the task lock to ensure existence of mm.
+                      */
+                     task_lock(p);
+                     mm_match = p->mm && (atomic_read(&p->mm->membarrier_state) &

Are we guaranteed that this p->mm will be the same as the one loaded via
READ_ONCE() above?

No; the way I read it, that's just an optimization and has no effect
on correctness.

Either way, wouldn't it be better to READ_ONCE() it a
single time and use the same value everywhere?

No; the first READ_ONCE() returns a pointer that you can't access
because it wasn't read under a lock. You can only use it for a NULL
check.

----- On Jan 28, 2019, at 5:39 PM, paulmck paulmck@linux.ibm.com wrote:

On Mon, Jan 28, 2019 at 05:07:07PM -0500, Mathieu Desnoyers wrote:

quoted

Jann Horn identified a racy access to p->mm in the global expedited
command of the membarrier system call.

The suggested fix is to hold the task_lock() around the accesses to
p->mm and to the mm_struct membarrier_state field to guarantee the
existence of the mm_struct.

Link:
https://lore.kernel.org/lkml/CAG48ez2G8ctF8dHS42TF37pThfr3y0RNOOYTmxvACm4u8Yu3cw@mail.gmail.com
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Tested-by: Jann Horn <jannh@google.com>
CC: Jann Horn <jannh@google.com>
CC: Thomas Gleixner <redacted>
CC: Peter Zijlstra (Intel) <peterz@infradead.org>
CC: Ingo Molnar <mingo@kernel.org>
CC: Andrea Parri <parri.andrea@gmail.com>
CC: Andy Lutomirski <luto@kernel.org>
CC: Avi Kivity <redacted>
CC: Benjamin Herrenschmidt <benh@kernel.crashing.org>
CC: Boqun Feng <redacted>
CC: Dave Watson <redacted>
CC: David Sehr <redacted>
CC: H. Peter Anvin <hpa@zytor.com>
CC: Linus Torvalds <torvalds@linux-foundation.org>
CC: Maged Michael <redacted>
CC: Michael Ellerman <mpe@ellerman.id.au>
CC: Paul E. McKenney <redacted>
CC: Paul Mackerras <redacted>
CC: Russell King <linux@armlinux.org.uk>
CC: Will Deacon <redacted>
CC: stable@vger.kernel.org # v4.16+
CC: linux-api@vger.kernel.org
---
 kernel/sched/membarrier.c | 27 +++++++++++++++++++++------
 1 file changed, 21 insertions(+), 6 deletions(-)

diff --git a/kernel/sched/membarrier.c b/kernel/sched/membarrier.c
index 76e0eaf4654e..305fdcc4c5f7 100644
--- a/kernel/sched/membarrier.c
+++ b/kernel/sched/membarrier.c

@@ -81,12 +81,27 @@ static int membarrier_global_expedited(void)

 		rcu_read_lock();
 		p = task_rcu_dereference(&cpu_rq(cpu)->curr);
-		if (p && p->mm && (atomic_read(&p->mm->membarrier_state) &
-				   MEMBARRIER_STATE_GLOBAL_EXPEDITED)) {
-			if (!fallback)
-				__cpumask_set_cpu(cpu, tmpmask);
-			else
-				smp_call_function_single(cpu, ipi_mb, NULL, 1);
+		/*
+		 * Skip this CPU if the runqueue's current task is NULL or if
+		 * it is a kernel thread.
+		 */
+		if (p && READ_ONCE(p->mm)) {
+			bool mm_match;
+
+			/*
+			 * Read p->mm and access membarrier_state while holding
+			 * the task lock to ensure existence of mm.
+			 */
+			task_lock(p);
+			mm_match = p->mm && (atomic_read(&p->mm->membarrier_state) &

Are we guaranteed that this p->mm will be the same as the one loaded via
READ_ONCE() above?  Either way, wouldn't it be better to READ_ONCE() it a
single time and use the same value everywhere?

The first "READ_ONCE()" above is _outside_ of the task_lock() critical section.
Those two accesses _can_ load two different pointers, and this is why we
need to re-read the p->mm pointer within the task_lock() critical section to
ensure existence of the mm_struct that we use.

If we move the READ_ONCE() into the task_lock(), we need to uselessly
take a lock before we can skip kernel threads.

If we lead the READ_ONCE() outside the task_lock(), then p->mm can be updated
between the READ_ONCE() and reference to the mm_struct content within the
task_lock(), which is racy and does not guarantee its existence.

Or am I missing your point ?

Thanks,

Mathieu

						Thanx, Paul

quoted

+					     MEMBARRIER_STATE_GLOBAL_EXPEDITED);
+			task_unlock(p);
+			if (mm_match) {
+				if (!fallback)
+					__cpumask_set_cpu(cpu, tmpmask);
+				else
+					smp_call_function_single(cpu, ipi_mb, NULL, 1);
+			}
 		}
 		rcu_read_unlock();
 	}
--
2.17.1

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

On Mon, Jan 28, 2019 at 11:45:32PM +0100, Jann Horn wrote:

On Mon, Jan 28, 2019 at 11:39 PM Paul E. McKenney [off-list ref] wrote:

quoted

On Mon, Jan 28, 2019 at 05:07:07PM -0500, Mathieu Desnoyers wrote:

quoted

Jann Horn identified a racy access to p->mm in the global expedited
command of the membarrier system call.

The suggested fix is to hold the task_lock() around the accesses to
p->mm and to the mm_struct membarrier_state field to guarantee the
existence of the mm_struct.

Link: https://lore.kernel.org/lkml/CAG48ez2G8ctF8dHS42TF37pThfr3y0RNOOYTmxvACm4u8Yu3cw@mail.gmail.com
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

[...]

quoted

quoted

--- a/kernel/sched/membarrier.c
+++ b/kernel/sched/membarrier.c

@@ -81,12 +81,27 @@ static int membarrier_global_expedited(void)

              rcu_read_lock();
              p = task_rcu_dereference(&cpu_rq(cpu)->curr);
-             if (p && p->mm && (atomic_read(&p->mm->membarrier_state) &
-                                MEMBARRIER_STATE_GLOBAL_EXPEDITED)) {
-                     if (!fallback)
-                             __cpumask_set_cpu(cpu, tmpmask);
-                     else
-                             smp_call_function_single(cpu, ipi_mb, NULL, 1);
+             /*
+              * Skip this CPU if the runqueue's current task is NULL or if
+              * it is a kernel thread.
+              */
+             if (p && READ_ONCE(p->mm)) {
+                     bool mm_match;
+
+                     /*
+                      * Read p->mm and access membarrier_state while holding
+                      * the task lock to ensure existence of mm.
+                      */
+                     task_lock(p);
+                     mm_match = p->mm && (atomic_read(&p->mm->membarrier_state) &

Are we guaranteed that this p->mm will be the same as the one loaded via
READ_ONCE() above?

No; the way I read it, that's just an optimization and has no effect
on correctness.

quoted

Either way, wouldn't it be better to READ_ONCE() it a
single time and use the same value everywhere?

No; the first READ_ONCE() returns a pointer that you can't access
because it wasn't read under a lock. You can only use it for a NULL
check.

Ah, of course!  Thank you both!

							Thanx, Paul