(mem hotplug, pcpu_alloc) BUG: sleeping function called from invalid context at kernel/mutex.c:94

Hi,

We have seen following might_sleep warning while hot adding memory...

[  142.339267] BUG: sleeping function called from invalid context at kernel/mutex.c:94
[  142.339276] in_atomic(): 0, irqs_disabled(): 1, pid: 4, name: migration/0
[  142.339283] Pid: 4, comm: migration/0 Not tainted 2.6.35.6-45.fc14.x86_64 #1
[  142.339288] Call Trace:
[  142.339305]  [<ffffffff8103d12b>] __might_sleep+0xeb/0xf0
[  142.339316]  [<ffffffff81468245>] mutex_lock+0x24/0x50
[  142.339326]  [<ffffffff8110eaa6>] pcpu_alloc+0x6d/0x7ee
[  142.339336]  [<ffffffff81048888>] ? load_balance+0xbe/0x60e
[  142.339343]  [<ffffffff8103a1b3>] ? rt_se_boosted+0x21/0x2f
[  142.339349]  [<ffffffff8103e1cf>] ? dequeue_rt_stack+0x18b/0x1ed
[  142.339356]  [<ffffffff8110f237>] __alloc_percpu+0x10/0x12
[  142.339362]  [<ffffffff81465e22>] setup_zone_pageset+0x38/0xbe
[  142.339373]  [<ffffffff810d6d81>] ? build_zonelists_node.clone.58+0x79/0x8c
[  142.339384]  [<ffffffff81452539>] __build_all_zonelists+0x419/0x46c
[  142.339395]  [<ffffffff8108ef01>] ? cpu_stopper_thread+0xb2/0x198
[  142.339401]  [<ffffffff8108f075>] stop_machine_cpu_stop+0x8e/0xc5
[  142.339407]  [<ffffffff8108efe7>] ? stop_machine_cpu_stop+0x0/0xc5
[  142.339414]  [<ffffffff8108ef57>] cpu_stopper_thread+0x108/0x198
[  142.339420]  [<ffffffff81467a37>] ? schedule+0x5b2/0x5cc
[  142.339426]  [<ffffffff8108ee4f>] ? cpu_stopper_thread+0x0/0x198
[  142.339434]  [<ffffffff81065f29>] kthread+0x7f/0x87
[  142.339443]  [<ffffffff8100aae4>] kernel_thread_helper+0x4/0x10
[  142.339449]  [<ffffffff81065eaa>] ? kthread+0x0/0x87
[  142.339455]  [<ffffffff8100aae0>] ? kernel_thread_helper+0x0/0x10
[  142.340099] Built 5 zonelists in Node order, mobility grouping on.  Total pages: 289456
[  142.340108] Policy zone: Normal

This warning was seen on the FC14 kernel, though looking at the current git, the problem seems to exist on mainline too.
The problem is that pcpu_alloc expects that it is called from non-atomic context as a result it grabs the pcpu_alloc_mutex.
In the memory-hotplug case though, we do end up calling pcpu_alloc from atomic context, while all cpus are stopped.

void build_all_zonelists(void *data)
{
set_zonelist_order();

   if (system_state == SYSTEM_BOOTING) {
      __build_all_zonelists(NULL);
      mminit_verify_zonelist();
      cpuset_init_current_mems_allowed();
   } else {
      /* we have to stop all cpus to guarantee there is no user
         of zonelist */
      stop_machine(__build_all_zonelists, data, NULL);   <=========
      /* cpuset refresh routine should be here */
   }

__build_all_zonelists eventually calls pcpu_alloc.

I didn't dive through the history, so am not sure when was this regression introduced, but could have regressed with the new pcpu memory allocator.

--
Alok

Hello,

Meh... the percpu allocator required user context from the beginning. The new allocator didn't change that.

Wouldn't it be possible to prepare hotplug outside of cpu_stop and use stop_machine() only to make it available to the system. In general, it's a very bad idea to allocate memory from inside stop_machine. The whole machine is stopped, after all. In general, it shouldn't be too difficult to add new resource without stop_machine too unlike removing one. Pekka, Christoph, any ideas?

Thanks.

tejun

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom policy in Canada: sign http://dissolvethecrtc.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Fix here. I'm glad if someone test this.
==

At 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.

[  142.339267] BUG: sleeping function called from invalid context at kernel/mutex.c:94
[  142.339276] in_atomic(): 0, irqs_disabled(): 1, pid: 4, name: migration/0
[  142.339283] Pid: 4, comm: migration/0 Not tainted 2.6.35.6-45.fc14.x86_64 #1
[  142.339288] Call Trace:
[  142.339305]  [<ffffffff8103d12b>] __might_sleep+0xeb/0xf0
[  142.339316]  [<ffffffff81468245>] mutex_lock+0x24/0x50
[  142.339326]  [<ffffffff8110eaa6>] pcpu_alloc+0x6d/0x7ee
[  142.339336]  [<ffffffff81048888>] ? load_balance+0xbe/0x60e
[  142.339343]  [<ffffffff8103a1b3>] ? rt_se_boosted+0x21/0x2f
[  142.339349]  [<ffffffff8103e1cf>] ? dequeue_rt_stack+0x18b/0x1ed
[  142.339356]  [<ffffffff8110f237>] __alloc_percpu+0x10/0x12
[  142.339362]  [<ffffffff81465e22>] setup_zone_pageset+0x38/0xbe
[  142.339373]  [<ffffffff810d6d81>] ? build_zonelists_node.clone.58+0x79/0x8c
[  142.339384]  [<ffffffff81452539>] __build_all_zonelists+0x419/0x46c
[  142.339395]  [<ffffffff8108ef01>] ? cpu_stopper_thread+0xb2/0x198
[  142.339401]  [<ffffffff8108f075>] stop_machine_cpu_stop+0x8e/0xc5
[  142.339407]  [<ffffffff8108efe7>] ? stop_machine_cpu_stop+0x0/0xc5
[  142.339414]  [<ffffffff8108ef57>] cpu_stopper_thread+0x108/0x198
[  142.339420]  [<ffffffff81467a37>] ? schedule+0x5b2/0x5cc
[  142.339426]  [<ffffffff8108ee4f>] ? cpu_stopper_thread+0x0/0x198
[  142.339434]  [<ffffffff81065f29>] kthread+0x7f/0x87
[  142.339443]  [<ffffffff8100aae4>] kernel_thread_helper+0x4/0x10
[  142.339449]  [<ffffffff81065eaa>] ? kthread+0x0/0x87
[  142.339455]  [<ffffffff8100aae0>] ? kernel_thread_helper+0x0/0x10
[  142.340099] Built 5 zonelists in Node order, mobility grouping on.  Total pages: 289456
[  142.340108] 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().

---
 mm/page_alloc.c |   16 +++++++---------
 1 file changed, 7 insertions(+), 9 deletions(-)

Index: mmotm-1117/mm/page_alloc.c
===================================================================
@@ -3027,14 +3027,6 @@ static __init_refok int __build_all_zone
 		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

@@ -3083,7 +3075,13 @@ 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) {
+			struct zone *zone = (struct zone *)data;
+			setup_zone_pageset(zone);
+		}
+#endif
+		stop_machine(__build_all_zonelists, NULL, NULL);
 		/* cpuset refresh routine should be here */
 	}
 	vm_total_pages = nr_free_pagecache_pages();