mm/page_alloc.c: fix build_all_zonelist() where percpu_alloc() is wrongly called...
authorKAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Wed, 24 Nov 2010 20:57:09 +0000 (12:57 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 24 Nov 2010 21:50:45 +0000 (06:50 +0900)
During memory hotplug, build_allzonelists() may be called under
stop_machine_run().  In this function, setup_zone_pageset() is called.
But it's bug because it will do page allocation under stop_machine_run().

Here is a report from Alok Kataria.

  BUG: sleeping function called from invalid context at kernel/mutex.c:94
  in_atomic(): 0, irqs_disabled(): 1, pid: 4, name: migration/0
  Pid: 4, comm: migration/0 Not tainted 2.6.35.6-45.fc14.x86_64 #1
  Call Trace:
   [<ffffffff8103d12b>] __might_sleep+0xeb/0xf0
   [<ffffffff81468245>] mutex_lock+0x24/0x50
   [<ffffffff8110eaa6>] pcpu_alloc+0x6d/0x7ee
   [<ffffffff81048888>] ? load_balance+0xbe/0x60e
   [<ffffffff8103a1b3>] ? rt_se_boosted+0x21/0x2f
   [<ffffffff8103e1cf>] ? dequeue_rt_stack+0x18b/0x1ed
   [<ffffffff8110f237>] __alloc_percpu+0x10/0x12
   [<ffffffff81465e22>] setup_zone_pageset+0x38/0xbe
   [<ffffffff810d6d81>] ? build_zonelists_node.clone.58+0x79/0x8c
   [<ffffffff81452539>] __build_all_zonelists+0x419/0x46c
   [<ffffffff8108ef01>] ? cpu_stopper_thread+0xb2/0x198
   [<ffffffff8108f075>] stop_machine_cpu_stop+0x8e/0xc5
   [<ffffffff8108efe7>] ? stop_machine_cpu_stop+0x0/0xc5
   [<ffffffff8108ef57>] cpu_stopper_thread+0x108/0x198
   [<ffffffff81467a37>] ? schedule+0x5b2/0x5cc
   [<ffffffff8108ee4f>] ? cpu_stopper_thread+0x0/0x198
   [<ffffffff81065f29>] kthread+0x7f/0x87
   [<ffffffff8100aae4>] kernel_thread_helper+0x4/0x10
   [<ffffffff81065eaa>] ? kthread+0x0/0x87
   [<ffffffff8100aae0>] ? kernel_thread_helper+0x0/0x10
  Built 5 zonelists in Node order, mobility grouping on.  Total pages: 289456
  Policy zone: Normal

This patch tries to fix the issue by moving setup_zone_pageset() out from
stop_machine_run(). It's obviously not necessary to be called under
stop_machine_run().

[akpm@linux-foundation.org: remove unneeded local]
Reported-by: Alok Kataria <akataria@vmware.com>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Petr Vandrovec <petr@vmware.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/page_alloc.c

index 07a654486f75cfdfe13035f7bf501e2d2dbf0a4a..e4092704c1a9a281985d7fc650e7655b1e00ea09 100644 (file)
@@ -3008,14 +3008,6 @@ static __init_refok int __build_all_zonelists(void *data)
                build_zonelist_cache(pgdat);
        }
 
-#ifdef CONFIG_MEMORY_HOTPLUG
-       /* Setup real pagesets for the new zone */
-       if (data) {
-               struct zone *zone = data;
-               setup_zone_pageset(zone);
-       }
-#endif
-
        /*
         * Initialize the boot_pagesets that are going to be used
         * for bootstrapping processors. The real pagesets for
@@ -3064,7 +3056,11 @@ void build_all_zonelists(void *data)
        } else {
                /* we have to stop all cpus to guarantee there is no user
                   of zonelist */
-               stop_machine(__build_all_zonelists, data, NULL);
+#ifdef CONFIG_MEMORY_HOTPLUG
+               if (data)
+                       setup_zone_pageset((struct zone *)data);
+#endif
+               stop_machine(__build_all_zonelists, NULL, NULL);
                /* cpuset refresh routine should be here */
        }
        vm_total_pages = nr_free_pagecache_pages();