mm, hotplug: fix concurrent memory hot-add deadlock

There's a deadlock when concurrently hot-adding memory through the probe
interface and switching a memory block from offline to online.

When hot-adding memory via the probe interface, add_memory() first takes
mem_hotplug_begin() and then device_lock() is later taken when registering
the newly initialized memory block.  This creates a lock dependency of (1)
mem_hotplug.lock (2) dev->mutex.

When switching a memory block from offline to online, dev->mutex is first
grabbed in device_online() when the write(2) transitions an existing
memory block from offline to online, and then online_pages() will take
mem_hotplug_begin().

This creates a lock inversion between mem_hotplug.lock and dev->mutex.
Vitaly reports that this deadlock can happen when kworker handling a probe
event races with systemd-udevd switching a memory block's state.

This patch requires the state transition to take mem_hotplug_begin()
before dev->mutex.  Hot-adding memory via the probe interface creates a
memory block while holding mem_hotplug_begin(), there is no way to take
dev->mutex first in this case.

online_pages() and offline_pages() are only called when transitioning
memory block state.  We now require that mem_hotplug_begin() is taken
before calling them -- this requires exporting the mem_hotplug_begin() and
mem_hotplug_done() to generic code.  In all hot-add and hot-remove cases,
mem_hotplug_begin() is done prior to device_online().  This is all that is
needed to avoid the deadlock.

Signed-off-by: David Rientjes <rientjes@google.com>
Reported-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Tested-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net>
Cc: "K. Y. Srinivasan" <kys@microsoft.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Tang Chen <tangchen@cn.fujitsu.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zhang Zhen <zhenzhang.zhang@huawei.com>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Wang Nan <wangnan0@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
diff --git a/drivers/base/memory.c b/drivers/base/memory.c
index ab5c9a7..2804aed 100644
--- a/drivers/base/memory.c
+++ b/drivers/base/memory.c
@@ -219,6 +219,7 @@
 /*
  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
  * OK to have direct references to sparsemem variables in here.
+ * Must already be protected by mem_hotplug_begin().
  */
 static int
 memory_block_action(unsigned long phys_index, unsigned long action, int online_type)
@@ -286,6 +287,7 @@
 	if (mem->online_type < 0)
 		mem->online_type = MMOP_ONLINE_KEEP;
 
+	/* Already under protection of mem_hotplug_begin() */
 	ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 
 	/* clear online_type */
@@ -328,17 +330,19 @@
 		goto err;
 	}
 
+	/*
+	 * Memory hotplug needs to hold mem_hotplug_begin() for probe to find
+	 * the correct memory block to online before doing device_online(dev),
+	 * which will take dev->mutex.  Take the lock early to prevent an
+	 * inversion, memory_subsys_online() callbacks will be implemented by
+	 * assuming it's already protected.
+	 */
+	mem_hotplug_begin();
+
 	switch (online_type) {
 	case MMOP_ONLINE_KERNEL:
 	case MMOP_ONLINE_MOVABLE:
 	case MMOP_ONLINE_KEEP:
-		/*
-		 * mem->online_type is not protected so there can be a
-		 * race here.  However, when racing online, the first
-		 * will succeed and the second will just return as the
-		 * block will already be online.  The online type
-		 * could be either one, but that is expected.
-		 */
 		mem->online_type = online_type;
 		ret = device_online(&mem->dev);
 		break;
@@ -349,6 +353,7 @@
 		ret = -EINVAL; /* should never happen */
 	}
 
+	mem_hotplug_done();
 err:
 	unlock_device_hotplug();