sched/core: Fix illegal RCU from offline CPUs
authorPeter Zijlstra <peterz@infradead.org>
Wed, 1 Apr 2020 21:40:33 +0000 (17:40 -0400)
committerPeter Zijlstra <peterz@infradead.org>
Thu, 30 Apr 2020 18:14:41 +0000 (20:14 +0200)
In the CPU-offline process, it calls mmdrop() after idle entry and the
subsequent call to cpuhp_report_idle_dead(). Once execution passes the
call to rcu_report_dead(), RCU is ignoring the CPU, which results in
lockdep complaining when mmdrop() uses RCU from either memcg or
debugobjects below.

Fix it by cleaning up the active_mm state from BP instead. Every arch
which has CONFIG_HOTPLUG_CPU should have already called idle_task_exit()
from AP. The only exception is parisc because it switches them to
&init_mm unconditionally (see smp_boot_one_cpu() and smp_cpu_init()),
but the patch will still work there because it calls mmgrab(&init_mm) in
smp_cpu_init() and then should call mmdrop(&init_mm) in finish_cpu().

  WARNING: suspicious RCU usage
  -----------------------------
  kernel/workqueue.c:710 RCU or wq_pool_mutex should be held!

  other info that might help us debug this:

  RCU used illegally from offline CPU!
  Call Trace:
   dump_stack+0xf4/0x164 (unreliable)
   lockdep_rcu_suspicious+0x140/0x164
   get_work_pool+0x110/0x150
   __queue_work+0x1bc/0xca0
   queue_work_on+0x114/0x120
   css_release+0x9c/0xc0
   percpu_ref_put_many+0x204/0x230
   free_pcp_prepare+0x264/0x570
   free_unref_page+0x38/0xf0
   __mmdrop+0x21c/0x2c0
   idle_task_exit+0x170/0x1b0
   pnv_smp_cpu_kill_self+0x38/0x2e0
   cpu_die+0x48/0x64
   arch_cpu_idle_dead+0x30/0x50
   do_idle+0x2f4/0x470
   cpu_startup_entry+0x38/0x40
   start_secondary+0x7a8/0xa80
   start_secondary_resume+0x10/0x14

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Qian Cai <cai@lca.pw>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Michael Ellerman <mpe@ellerman.id.au> (powerpc)
Link: https://lkml.kernel.org/r/20200401214033.8448-1-cai@lca.pw
arch/powerpc/platforms/powernv/smp.c
include/linux/sched/mm.h
kernel/cpu.c
kernel/sched/core.c

index 13e2516..b2ba3e9 100644 (file)
@@ -167,7 +167,6 @@ static void pnv_smp_cpu_kill_self(void)
        /* Standard hot unplug procedure */
 
        idle_task_exit();
-       current->active_mm = NULL; /* for sanity */
        cpu = smp_processor_id();
        DBG("CPU%d offline\n", cpu);
        generic_set_cpu_dead(cpu);
index c49257a..a132d87 100644 (file)
@@ -49,6 +49,8 @@ static inline void mmdrop(struct mm_struct *mm)
                __mmdrop(mm);
 }
 
+void mmdrop(struct mm_struct *mm);
+
 /*
  * This has to be called after a get_task_mm()/mmget_not_zero()
  * followed by taking the mmap_sem for writing before modifying the
index 2371292..244d305 100644 (file)
@@ -3,6 +3,7 @@
  *
  * This code is licenced under the GPL.
  */
+#include <linux/sched/mm.h>
 #include <linux/proc_fs.h>
 #include <linux/smp.h>
 #include <linux/init.h>
@@ -564,6 +565,21 @@ static int bringup_cpu(unsigned int cpu)
        return bringup_wait_for_ap(cpu);
 }
 
+static int finish_cpu(unsigned int cpu)
+{
+       struct task_struct *idle = idle_thread_get(cpu);
+       struct mm_struct *mm = idle->active_mm;
+
+       /*
+        * idle_task_exit() will have switched to &init_mm, now
+        * clean up any remaining active_mm state.
+        */
+       if (mm != &init_mm)
+               idle->active_mm = &init_mm;
+       mmdrop(mm);
+       return 0;
+}
+
 /*
  * Hotplug state machine related functions
  */
@@ -1549,7 +1565,7 @@ static struct cpuhp_step cpuhp_hp_states[] = {
        [CPUHP_BRINGUP_CPU] = {
                .name                   = "cpu:bringup",
                .startup.single         = bringup_cpu,
-               .teardown.single        = NULL,
+               .teardown.single        = finish_cpu,
                .cant_stop              = true,
        },
        /* Final state before CPU kills itself */
index 2e6ba9e..f6ae262 100644 (file)
@@ -6197,13 +6197,14 @@ void idle_task_exit(void)
        struct mm_struct *mm = current->active_mm;
 
        BUG_ON(cpu_online(smp_processor_id()));
+       BUG_ON(current != this_rq()->idle);
 
        if (mm != &init_mm) {
                switch_mm(mm, &init_mm, current);
-               current->active_mm = &init_mm;
                finish_arch_post_lock_switch();
        }
-       mmdrop(mm);
+
+       /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
 }
 
 /*