[PATCH v3] sched/mmcid: Synchronize MM teardown with affinity updates

[Weiming Shi]

set_cpus_allowed_common() passes a remote task's p->mm to
mm_update_cpus_allowed() while holding p->pi_lock, but that lock did not
stabilize the pointer against MM replacement. A same-owner
sched_setaffinity() caller can therefore retain the old pointer while exec
replaces and frees it. On architectures which shoot down rather than
refcount lazy TLB MMs, exit_mm() and kthread_unuse_mm() have the same
lifetime gap.

  BUG: KASAN: slab-use-after-free in set_cpus_allowed_common
  Read of size 4 at addr ffff888123155e4c by task poc/5132
  Call Trace:
   set_cpus_allowed_common (kernel/sched/core.c:10835)
   set_cpus_allowed_fair (kernel/sched/fair.c:10046)
   __set_cpus_allowed_ptr_locked (kernel/sched/core.c:3224)
   __set_cpus_allowed_ptr (kernel/sched/core.c:3258)
   __sched_setaffinity (kernel/sched/syscalls.c:1161)
   sched_setaffinity (kernel/sched/syscalls.c:1239)
   __x64_sys_sched_setaffinity (kernel/sched/syscalls.c:1275)
   do_syscall_64 (arch/x86/entry/syscall_64.c:84)
   entry_SYSCALL_64_after_hwframe (arch/x86/entry/entry_64.S:121)

  Freed by task 5127:
   kmem_cache_free (mm/slub.c:6757)
   __mmput (kernel/fork.c:1204)
   mmput (kernel/fork.c:1215)
   setup_new_exec (fs/exec.c:1395)
   load_elf_binary (fs/binfmt_elf.c:1024)
   bprm_execve (fs/exec.c:1865)
   do_execveat_common.isra.0 (fs/exec.c:1963)
   __x64_sys_execveat (fs/exec.c:2047)
   do_syscall_64 (arch/x86/entry/syscall_64.c:84)
   entry_SYSCALL_64_after_hwframe (arch/x86/entry/entry_64.S:121)

  Kernel panic - not syncing: KASAN: panic_on_warn set ...

Affinity paths hold p->pi_lock across the p->mm load and its complete use.
After publishing a replacement or NULL p->mm, take and release that same
lock before dropping the old MM reference. For an updater that is already
using the old pointer, the ordering is:

  affinity updater                     MM teardown
  ----------------                     -----------
  raw_spin_lock(&p->pi_lock)
  mm = p->mm            /* old_mm */
                                       p->mm = new_mm or NULL
                                       raw_spin_lock(&p->pi_lock)
                                         /* waits */
  mm_update_cpus_allowed(mm, mask)
  raw_spin_unlock(&p->pi_lock) /* release */
                                         /* acquire succeeds */
                                       raw_spin_unlock(&p->pi_lock)
                                       ...
                                       mmput/mmdrop(old_mm)

The teardown-side acquire cannot complete until an updater ordered before
it has released p->pi_lock, so old_mm remains alive throughout that
updater's access. If teardown is ordered first instead, its unlock is a
release that orders the preceding p->mm store before a later updater's lock
acquire. The later updater therefore observes new_mm or NULL. An updater
that acquired the lock before the store but has not loaded p->mm yet is
still ordered before the teardown-side acquisition; it may observe either
pointer, but teardown waits for its complete use before dropping old_mm.

All runtime callers of set_cpus_allowed_common() hold p->pi_lock across
this access. The runqueue lock is not part of the p->mm lifetime ordering,
so the grace period only needs p->pi_lock. Place it after publication and
before the old reference drop in exec_mmap(), exit_mm(), and
kthread_unuse_mm(). The last two paths matter on architectures which shoot
down rather than refcount lazy TLB MMs.

This change was developed with assistance from an LLM.

Fixes: 7e019dcc470f ("sched: Improve cache locality of RSEQ concurrency IDs for intermittent workloads")
Cc: stable@xxxxxxxxxxxxxxx
Reported-by: <co+cab19e2ebe4d535a@xxxxxxx>
Signed-off-by: Weiming Shi <bestswngs@xxxxxxxxx>
---
Changes in v3:
- Document the p->pi_lock grace period and acquire/release ordering.
- Use p->pi_lock directly and cover exit_mm() and kthread_unuse_mm().
- Refresh the KASAN trace and provide a standalone reproducer in reply.

Changes in v2:
- Replace the schedule-in deferral with an exec-side runqueue-lock barrier.
- Keep affinity expansions synchronous, including expansions narrowed before
  the next schedule-in.
- Avoid changes to the MM-CID fork, exit, and scheduling state machine.

 fs/exec.c             | 1 +
 include/linux/sched.h | 2 ++
 kernel/exit.c         | 1 +
 kernel/kthread.c      | 1 +
 kernel/sched/core.c   | 9 +++++++++
 5 files changed, 14 insertions(+)

diff --git a/fs/exec.c b/fs/exec.c
index 8d2b14e1d455..3196421cb81d 100644
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -905,6 +905,7 @@ static int exec_mmap(struct linux_binprm *bprm)
 		BUG_ON(active_mm != old_mm);
 		/* Defer teardown to setup_new_exec(), outside the exec locks. */
 		bprm->old_mm = old_mm;
+		sched_mm_cid_mm_barrier(tsk);
 	} else {
 		mmdrop_lazy_tlb(active_mm);
 	}
diff --git a/include/linux/sched.h b/include/linux/sched.h
index 1c93da75226e..f88ab0445be3 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -2358,6 +2358,7 @@ static __always_inline void alloc_tag_restore(struct alloc_tag *tag, struct allo
 #ifdef CONFIG_SCHED_MM_CID
 void sched_mm_cid_before_execve(struct task_struct *t);
 void sched_mm_cid_after_execve(struct task_struct *t);
+void sched_mm_cid_mm_barrier(struct task_struct *t);
 void sched_mm_cid_exit(struct task_struct *t);
 static __always_inline int task_mm_cid(struct task_struct *t)
 {
@@ -2366,6 +2367,7 @@ static __always_inline int task_mm_cid(struct task_struct *t)
 #else
 static inline void sched_mm_cid_before_execve(struct task_struct *t) { }
 static inline void sched_mm_cid_after_execve(struct task_struct *t) { }
+static inline void sched_mm_cid_mm_barrier(struct task_struct *t) { }
 static inline void sched_mm_cid_exit(struct task_struct *t) { }
 static __always_inline int task_mm_cid(struct task_struct *t)
 {
diff --git a/kernel/exit.c b/kernel/exit.c
index d5e7bc2af804..5f9979d6216c 100644
--- a/kernel/exit.c
+++ b/kernel/exit.c
@@ -582,6 +582,7 @@ static void exit_mm(void)
 	local_irq_enable();
 	task_unlock(current);
 	mmap_read_unlock(mm);
+	sched_mm_cid_mm_barrier(current);
 	mm_update_next_owner(mm);
 	mmput(mm);
 	if (test_thread_flag(TIF_MEMDIE))
diff --git a/kernel/kthread.c b/kernel/kthread.c
index fdb8fb07f40b..e7b393cfdcee 100644
--- a/kernel/kthread.c
+++ b/kernel/kthread.c
@@ -1682,6 +1682,7 @@ void kthread_unuse_mm(struct mm_struct *mm)
 	enter_lazy_tlb(mm, tsk);
 	local_irq_enable();
 	task_unlock(tsk);
+	sched_mm_cid_mm_barrier(tsk);
 
 	mmdrop(mm);
 }
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 04b62c16fa8c..d78a174b5405 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -11128,6 +11128,15 @@ void sched_mm_cid_after_execve(struct task_struct *t)
 		sched_mm_cid_fork(t);
 }
 
+/*
+ * Pair with set_cpus_allowed_common()'s p->pi_lock critical section. Call
+ * after publishing p->mm and before dropping the old MM reference.
+ */
+void sched_mm_cid_mm_barrier(struct task_struct *t)
+{
+	guard(raw_spinlock_irq)(&t->pi_lock);
+}
+
 static void mm_cid_work_fn(struct work_struct *work)
 {
 	struct mm_struct *mm = container_of(work, struct mm_struct, mm_cid.work);
-- 
2.56.0