[PATCH] mm-add-vfree_atomic-fix

[Andrey Ryabinin]

DEBUG_PREEMPT complains about using this_cpu_ptr() in preemptible:
	BUG: using smp_processor_id() in preemptible [00000000] code: iperf-300s-cs-l/277
	caller is debug_smp_processor_id+0x17/0x19
	CPU: 1 PID: 277 Comm: iperf-300s-cs-l Not tainted 4.9.0-rc8-00140-gcc639db #2
	 ffffc900003f3cf0 ffffffff8123ae6f 0000000000000001 ffffffff818181da
	 ffffc900003f3d20 ffffffff81252f41 0000000000012de0 00000000fffffdff
	 ffff880009328f40 ffff88000592c400 ffffc900003f3d30 ffffffff81252f6a
	Call Trace:
	 [<ffffffff8123ae6f>] dump_stack+0x9a/0xd0
	 [<ffffffff81252f41>] check_preemption_disabled+0xdd/0xef
	 [<ffffffff81252f6a>] debug_smp_processor_id+0x17/0x19
	 [<ffffffff811796df>] __vfree_deferred+0x16/0x4c
	 [<ffffffff8117b584>] vfree_atomic+0x22/0x24
	 [<ffffffff81094f5d>] free_thread_stack+0xc2/0x106
	 [<ffffffff810951be>] put_task_stack+0x4c/0x62
	 [<ffffffff81095f81>] copy_process+0x7e0/0x16e8
	 [<ffffffff8109702d>] _do_fork+0xbb/0x2d3
	 [<ffffffff810465e8>] ? __do_page_fault+0x2e1/0x384
	 [<ffffffff8112633f>] ? trace_hardirqs_off_caller+0x12/0x24
	 [<ffffffff810972cb>] SyS_clone+0x19/0x1b
	 [<ffffffff81003800>] do_syscall_64+0x143/0x173
	 [<ffffffff81507289>] entry_SYSCALL64_slow_path+0x25/0x25

Use raw_cpu_ptr() instead of this_cpu_ptr() to hide this warning.
It's fine because llist_add() implementation is lock-less, so it works even
if we adding to the list of some other cpu. schedule_work() is also preempt-safe.

Reported-by: kernel test robot <ying.huang@xxxxxxxxxxxxxxx>
Signed-off-by: Andrey Ryabinin <aryabinin@xxxxxxxxxxxxx>
---
 mm/vmalloc.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 43f0608..d8813963 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -1498,7 +1498,14 @@ static void __vunmap(const void *addr, int deallocate_pages)
 
 static inline void __vfree_deferred(const void *addr)
 {
-	struct vfree_deferred *p = this_cpu_ptr(&vfree_deferred);
+	/*
+	 * Use raw_cpu_ptr() because this can be called from preemptible
+	 * context. Preemption is absolutely fine here, because llist_add()
+	 * implementation is lockless, so it works even if we adding to list
+	 * of the other cpu.
+	 * schedule_work() should be fine with this too.
+	 */
+	struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
 
 	if (llist_add((struct llist_node *)addr, &p->list))
 		schedule_work(&p->wq);
-- 
2.7.3