diff mbox series

[4.19] mm: fix exec activate_mm vs TLB shootdown and lazy tlb switching race

Message ID 20201104011406.598487-1-mpe@ellerman.id.au
State New
Headers show
Series [4.19] mm: fix exec activate_mm vs TLB shootdown and lazy tlb switching race | expand

Commit Message

Michael Ellerman Nov. 4, 2020, 1:14 a.m. UTC
From: Nicholas Piggin <npiggin@gmail.com>

commit d53c3dfb23c45f7d4f910c3a3ca84bf0a99c6143 upstream.

Reading and modifying current->mm and current->active_mm and switching
mm should be done with irqs off, to prevent races seeing an intermediate
state.

This is similar to commit 38cf307c1f20 ("mm: fix kthread_use_mm() vs TLB
invalidate"). At exec-time when the new mm is activated, the old one
should usually be single-threaded and no longer used, unless something
else is holding an mm_users reference (which may be possible).

Absent other mm_users, there is also a race with preemption and lazy tlb
switching. Consider the kernel_execve case where the current thread is
using a lazy tlb active mm:

  call_usermodehelper()
    kernel_execve()
      old_mm = current->mm;
      active_mm = current->active_mm;
      *** preempt *** -------------------->  schedule()
                                               prev->active_mm = NULL;
                                               mmdrop(prev active_mm);
                                             ...
                      <--------------------  schedule()
      current->mm = mm;
      current->active_mm = mm;
      if (!old_mm)
          mmdrop(active_mm);

If we switch back to the kernel thread from a different mm, there is a
double free of the old active_mm, and a missing free of the new one.

Closing this race only requires interrupts to be disabled while ->mm
and ->active_mm are being switched, but the TLB problem requires also
holding interrupts off over activate_mm. Unfortunately not all archs
can do that yet, e.g., arm defers the switch if irqs are disabled and
expects finish_arch_post_lock_switch() to be called to complete the
flush; um takes a blocking lock in activate_mm().

So as a first step, disable interrupts across the mm/active_mm updates
to close the lazy tlb preempt race, and provide an arch option to
extend that to activate_mm which allows architectures doing IPI based
TLB shootdowns to close the second race.

This is a bit ugly, but in the interest of fixing the bug and backporting
before all architectures are converted this is a compromise.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
[mpe: Manual backport to 4.19 due to membarrier_exec_mmap(mm) changes]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20200914045219.3736466-2-npiggin@gmail.com
---
 arch/Kconfig |  7 +++++++
 fs/exec.c    | 15 ++++++++++++++-
 2 files changed, 21 insertions(+), 1 deletion(-)

Comments

Greg Kroah-Hartman Nov. 4, 2020, 9:05 a.m. UTC | #1
On Wed, Nov 04, 2020 at 12:14:06PM +1100, Michael Ellerman wrote:
> From: Nicholas Piggin <npiggin@gmail.com>

> 

> commit d53c3dfb23c45f7d4f910c3a3ca84bf0a99c6143 upstream.

> 

> Reading and modifying current->mm and current->active_mm and switching

> mm should be done with irqs off, to prevent races seeing an intermediate

> state.

> 

> This is similar to commit 38cf307c1f20 ("mm: fix kthread_use_mm() vs TLB

> invalidate"). At exec-time when the new mm is activated, the old one

> should usually be single-threaded and no longer used, unless something

> else is holding an mm_users reference (which may be possible).

> 

> Absent other mm_users, there is also a race with preemption and lazy tlb

> switching. Consider the kernel_execve case where the current thread is

> using a lazy tlb active mm:

> 

>   call_usermodehelper()

>     kernel_execve()

>       old_mm = current->mm;

>       active_mm = current->active_mm;

>       *** preempt *** -------------------->  schedule()

>                                                prev->active_mm = NULL;

>                                                mmdrop(prev active_mm);

>                                              ...

>                       <--------------------  schedule()

>       current->mm = mm;

>       current->active_mm = mm;

>       if (!old_mm)

>           mmdrop(active_mm);

> 

> If we switch back to the kernel thread from a different mm, there is a

> double free of the old active_mm, and a missing free of the new one.

> 

> Closing this race only requires interrupts to be disabled while ->mm

> and ->active_mm are being switched, but the TLB problem requires also

> holding interrupts off over activate_mm. Unfortunately not all archs

> can do that yet, e.g., arm defers the switch if irqs are disabled and

> expects finish_arch_post_lock_switch() to be called to complete the

> flush; um takes a blocking lock in activate_mm().

> 

> So as a first step, disable interrupts across the mm/active_mm updates

> to close the lazy tlb preempt race, and provide an arch option to

> extend that to activate_mm which allows architectures doing IPI based

> TLB shootdowns to close the second race.

> 

> This is a bit ugly, but in the interest of fixing the bug and backporting

> before all architectures are converted this is a compromise.

> 

> Signed-off-by: Nicholas Piggin <npiggin@gmail.com>

> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>

> [mpe: Manual backport to 4.19 due to membarrier_exec_mmap(mm) changes]

> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

> Link: https://lore.kernel.org/r/20200914045219.3736466-2-npiggin@gmail.com

> ---

>  arch/Kconfig |  7 +++++++

>  fs/exec.c    | 15 ++++++++++++++-

>  2 files changed, 21 insertions(+), 1 deletion(-)


Now queued up, thanks!

greg k-h
diff mbox series

Patch

diff --git a/arch/Kconfig b/arch/Kconfig
index a336548487e6..e3a030f7a722 100644
--- a/arch/Kconfig
+++ b/arch/Kconfig
@@ -366,6 +366,13 @@  config HAVE_RCU_TABLE_FREE
 config HAVE_RCU_TABLE_INVALIDATE
 	bool
 
+config ARCH_WANT_IRQS_OFF_ACTIVATE_MM
+	bool
+	help
+	  Temporary select until all architectures can be converted to have
+	  irqs disabled over activate_mm. Architectures that do IPI based TLB
+	  shootdowns should enable this.
+
 config ARCH_HAVE_NMI_SAFE_CMPXCHG
 	bool
 
diff --git a/fs/exec.c b/fs/exec.c
index cece8c14f377..52788644c4af 100644
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -1028,10 +1028,23 @@  static int exec_mmap(struct mm_struct *mm)
 		}
 	}
 	task_lock(tsk);
+
+	local_irq_disable();
 	active_mm = tsk->active_mm;
-	tsk->mm = mm;
 	tsk->active_mm = mm;
+	tsk->mm = mm;
+	/*
+	 * This prevents preemption while active_mm is being loaded and
+	 * it and mm are being updated, which could cause problems for
+	 * lazy tlb mm refcounting when these are updated by context
+	 * switches. Not all architectures can handle irqs off over
+	 * activate_mm yet.
+	 */
+	if (!IS_ENABLED(CONFIG_ARCH_WANT_IRQS_OFF_ACTIVATE_MM))
+		local_irq_enable();
 	activate_mm(active_mm, mm);
+	if (IS_ENABLED(CONFIG_ARCH_WANT_IRQS_OFF_ACTIVATE_MM))
+		local_irq_enable();
 	tsk->mm->vmacache_seqnum = 0;
 	vmacache_flush(tsk);
 	task_unlock(tsk);