PM / QoS: Avoid possible deadlock related to sysfs access
authorRafael J. Wysocki <rafael.j.wysocki@intel.com>
Mon, 1 Apr 2013 23:25:24 +0000 (01:25 +0200)
committerRafael J. Wysocki <rafael.j.wysocki@intel.com>
Mon, 1 Apr 2013 23:25:24 +0000 (01:25 +0200)
Commit b81ea1b (PM / QoS: Fix concurrency issues and memory leaks in
device PM QoS) put calls to pm_qos_sysfs_add_latency(),
pm_qos_sysfs_add_flags(), pm_qos_sysfs_remove_latency(), and
pm_qos_sysfs_remove_flags() under dev_pm_qos_mtx, which was a
mistake, because it may lead to deadlocks in some situations.
For example, if pm_qos_remote_wakeup_store() is run in parallel
with dev_pm_qos_constraints_destroy(), they may deadlock in the
following way:

 ======================================================
 [ INFO: possible circular locking dependency detected ]
 3.9.0-rc4-next-20130328-sasha-00014-g91a3267 #319 Tainted: G        W
 -------------------------------------------------------
 trinity-child6/12371 is trying to acquire lock:
  (s_active#54){++++.+}, at: [<ffffffff81301631>] sysfs_addrm_finish+0x31/0x60

 but task is already holding lock:
  (dev_pm_qos_mtx){+.+.+.}, at: [<ffffffff81f07cc3>] dev_pm_qos_constraints_destroy+0x23/0x250

 which lock already depends on the new lock.

 the existing dependency chain (in reverse order) is:

 -> #1 (dev_pm_qos_mtx){+.+.+.}:
        [<ffffffff811811da>] lock_acquire+0x1aa/0x240
        [<ffffffff83dab809>] __mutex_lock_common+0x59/0x5e0
        [<ffffffff83dabebf>] mutex_lock_nested+0x3f/0x50
        [<ffffffff81f07f2f>] dev_pm_qos_update_flags+0x3f/0xc0
        [<ffffffff81f05f4f>] pm_qos_remote_wakeup_store+0x3f/0x70
        [<ffffffff81efbb43>] dev_attr_store+0x13/0x20
        [<ffffffff812ffdaa>] sysfs_write_file+0xfa/0x150
        [<ffffffff8127f2c1>] __kernel_write+0x81/0x150
        [<ffffffff812afc2d>] write_pipe_buf+0x4d/0x80
        [<ffffffff812af57c>] splice_from_pipe_feed+0x7c/0x120
        [<ffffffff812afa25>] __splice_from_pipe+0x45/0x80
        [<ffffffff812b14fc>] splice_from_pipe+0x4c/0x70
        [<ffffffff812b1538>] default_file_splice_write+0x18/0x30
        [<ffffffff812afae3>] do_splice_from+0x83/0xb0
        [<ffffffff812afb2e>] direct_splice_actor+0x1e/0x20
        [<ffffffff812b0277>] splice_direct_to_actor+0xe7/0x200
        [<ffffffff812b15bc>] do_splice_direct+0x4c/0x70
        [<ffffffff8127eda9>] do_sendfile+0x169/0x300
        [<ffffffff8127ff94>] SyS_sendfile64+0x64/0xb0
        [<ffffffff83db7d18>] tracesys+0xe1/0xe6

 -> #0 (s_active#54){++++.+}:
        [<ffffffff811800cf>] __lock_acquire+0x15bf/0x1e50
        [<ffffffff811811da>] lock_acquire+0x1aa/0x240
        [<ffffffff81300aa2>] sysfs_deactivate+0x122/0x1a0
        [<ffffffff81301631>] sysfs_addrm_finish+0x31/0x60
        [<ffffffff812ff77f>] sysfs_hash_and_remove+0x7f/0xb0
        [<ffffffff813035a1>] sysfs_unmerge_group+0x51/0x70
        [<ffffffff81f068f4>] pm_qos_sysfs_remove_flags+0x14/0x20
        [<ffffffff81f07490>] __dev_pm_qos_hide_flags+0x30/0x70
        [<ffffffff81f07cd5>] dev_pm_qos_constraints_destroy+0x35/0x250
        [<ffffffff81f06931>] dpm_sysfs_remove+0x11/0x50
        [<ffffffff81efcf6f>] device_del+0x3f/0x1b0
        [<ffffffff81efd128>] device_unregister+0x48/0x60
        [<ffffffff82d4083c>] usb_hub_remove_port_device+0x1c/0x20
        [<ffffffff82d2a9cd>] hub_disconnect+0xdd/0x160
        [<ffffffff82d36ab7>] usb_unbind_interface+0x67/0x170
        [<ffffffff81f001a7>] __device_release_driver+0x87/0xe0
        [<ffffffff81f00559>] device_release_driver+0x29/0x40
        [<ffffffff81effc58>] bus_remove_device+0x148/0x160
        [<ffffffff81efd07f>] device_del+0x14f/0x1b0
        [<ffffffff82d344f9>] usb_disable_device+0xf9/0x280
        [<ffffffff82d34ff8>] usb_set_configuration+0x268/0x840
        [<ffffffff82d3a7fc>] usb_remove_store+0x4c/0x80
        [<ffffffff81efbb43>] dev_attr_store+0x13/0x20
        [<ffffffff812ffdaa>] sysfs_write_file+0xfa/0x150
        [<ffffffff8127f71d>] do_loop_readv_writev+0x4d/0x90
        [<ffffffff8127f999>] do_readv_writev+0xf9/0x1e0
        [<ffffffff8127faba>] vfs_writev+0x3a/0x60
        [<ffffffff8127fc60>] SyS_writev+0x50/0xd0
        [<ffffffff83db7d18>] tracesys+0xe1/0xe6

 other info that might help us debug this:

  Possible unsafe locking scenario:

        CPU0                    CPU1
        ----                    ----
   lock(dev_pm_qos_mtx);
                                lock(s_active#54);
                                lock(dev_pm_qos_mtx);
   lock(s_active#54);

  *** DEADLOCK ***

To avoid that, remove the calls to functions mentioned above from
under dev_pm_qos_mtx and introduce a separate lock to prevent races
between functions that add or remove device PM QoS sysfs attributes
from happening.

Reported-by: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
drivers/base/power/qos.c

index 5f74587..71671c4 100644 (file)
@@ -46,6 +46,7 @@
 #include "power.h"
 
 static DEFINE_MUTEX(dev_pm_qos_mtx);
+static DEFINE_MUTEX(dev_pm_qos_sysfs_mtx);
 
 static BLOCKING_NOTIFIER_HEAD(dev_pm_notifiers);
 
@@ -216,12 +217,17 @@ void dev_pm_qos_constraints_destroy(struct device *dev)
        struct pm_qos_constraints *c;
        struct pm_qos_flags *f;
 
-       mutex_lock(&dev_pm_qos_mtx);
+       mutex_lock(&dev_pm_qos_sysfs_mtx);
 
        /*
         * If the device's PM QoS resume latency limit or PM QoS flags have been
         * exposed to user space, they have to be hidden at this point.
         */
+       pm_qos_sysfs_remove_latency(dev);
+       pm_qos_sysfs_remove_flags(dev);
+
+       mutex_lock(&dev_pm_qos_mtx);
+
        __dev_pm_qos_hide_latency_limit(dev);
        __dev_pm_qos_hide_flags(dev);
 
@@ -254,6 +260,8 @@ void dev_pm_qos_constraints_destroy(struct device *dev)
 
  out:
        mutex_unlock(&dev_pm_qos_mtx);
+
+       mutex_unlock(&dev_pm_qos_sysfs_mtx);
 }
 
 /**
@@ -558,6 +566,14 @@ static void __dev_pm_qos_drop_user_request(struct device *dev,
        kfree(req);
 }
 
+static void dev_pm_qos_drop_user_request(struct device *dev,
+                                        enum dev_pm_qos_req_type type)
+{
+       mutex_lock(&dev_pm_qos_mtx);
+       __dev_pm_qos_drop_user_request(dev, type);
+       mutex_unlock(&dev_pm_qos_mtx);
+}
+
 /**
  * dev_pm_qos_expose_latency_limit - Expose PM QoS latency limit to user space.
  * @dev: Device whose PM QoS latency limit is to be exposed to user space.
@@ -581,6 +597,8 @@ int dev_pm_qos_expose_latency_limit(struct device *dev, s32 value)
                return ret;
        }
 
+       mutex_lock(&dev_pm_qos_sysfs_mtx);
+
        mutex_lock(&dev_pm_qos_mtx);
 
        if (IS_ERR_OR_NULL(dev->power.qos))
@@ -591,26 +609,27 @@ int dev_pm_qos_expose_latency_limit(struct device *dev, s32 value)
        if (ret < 0) {
                __dev_pm_qos_remove_request(req);
                kfree(req);
+               mutex_unlock(&dev_pm_qos_mtx);
                goto out;
        }
-
        dev->power.qos->latency_req = req;
+
+       mutex_unlock(&dev_pm_qos_mtx);
+
        ret = pm_qos_sysfs_add_latency(dev);
        if (ret)
-               __dev_pm_qos_drop_user_request(dev, DEV_PM_QOS_LATENCY);
+               dev_pm_qos_drop_user_request(dev, DEV_PM_QOS_LATENCY);
 
  out:
-       mutex_unlock(&dev_pm_qos_mtx);
+       mutex_unlock(&dev_pm_qos_sysfs_mtx);
        return ret;
 }
 EXPORT_SYMBOL_GPL(dev_pm_qos_expose_latency_limit);
 
 static void __dev_pm_qos_hide_latency_limit(struct device *dev)
 {
-       if (!IS_ERR_OR_NULL(dev->power.qos) && dev->power.qos->latency_req) {
-               pm_qos_sysfs_remove_latency(dev);
+       if (!IS_ERR_OR_NULL(dev->power.qos) && dev->power.qos->latency_req)
                __dev_pm_qos_drop_user_request(dev, DEV_PM_QOS_LATENCY);
-       }
 }
 
 /**
@@ -619,9 +638,15 @@ static void __dev_pm_qos_hide_latency_limit(struct device *dev)
  */
 void dev_pm_qos_hide_latency_limit(struct device *dev)
 {
+       mutex_lock(&dev_pm_qos_sysfs_mtx);
+
+       pm_qos_sysfs_remove_latency(dev);
+
        mutex_lock(&dev_pm_qos_mtx);
        __dev_pm_qos_hide_latency_limit(dev);
        mutex_unlock(&dev_pm_qos_mtx);
+
+       mutex_unlock(&dev_pm_qos_sysfs_mtx);
 }
 EXPORT_SYMBOL_GPL(dev_pm_qos_hide_latency_limit);
 
@@ -649,6 +674,8 @@ int dev_pm_qos_expose_flags(struct device *dev, s32 val)
        }
 
        pm_runtime_get_sync(dev);
+       mutex_lock(&dev_pm_qos_sysfs_mtx);
+
        mutex_lock(&dev_pm_qos_mtx);
 
        if (IS_ERR_OR_NULL(dev->power.qos))
@@ -659,16 +686,19 @@ int dev_pm_qos_expose_flags(struct device *dev, s32 val)
        if (ret < 0) {
                __dev_pm_qos_remove_request(req);
                kfree(req);
+               mutex_unlock(&dev_pm_qos_mtx);
                goto out;
        }
-
        dev->power.qos->flags_req = req;
+
+       mutex_unlock(&dev_pm_qos_mtx);
+
        ret = pm_qos_sysfs_add_flags(dev);
        if (ret)
-               __dev_pm_qos_drop_user_request(dev, DEV_PM_QOS_FLAGS);
+               dev_pm_qos_drop_user_request(dev, DEV_PM_QOS_FLAGS);
 
  out:
-       mutex_unlock(&dev_pm_qos_mtx);
+       mutex_unlock(&dev_pm_qos_sysfs_mtx);
        pm_runtime_put(dev);
        return ret;
 }
@@ -676,10 +706,8 @@ EXPORT_SYMBOL_GPL(dev_pm_qos_expose_flags);
 
 static void __dev_pm_qos_hide_flags(struct device *dev)
 {
-       if (!IS_ERR_OR_NULL(dev->power.qos) && dev->power.qos->flags_req) {
-               pm_qos_sysfs_remove_flags(dev);
+       if (!IS_ERR_OR_NULL(dev->power.qos) && dev->power.qos->flags_req)
                __dev_pm_qos_drop_user_request(dev, DEV_PM_QOS_FLAGS);
-       }
 }
 
 /**
@@ -689,9 +717,15 @@ static void __dev_pm_qos_hide_flags(struct device *dev)
 void dev_pm_qos_hide_flags(struct device *dev)
 {
        pm_runtime_get_sync(dev);
+       mutex_lock(&dev_pm_qos_sysfs_mtx);
+
+       pm_qos_sysfs_remove_flags(dev);
+
        mutex_lock(&dev_pm_qos_mtx);
        __dev_pm_qos_hide_flags(dev);
        mutex_unlock(&dev_pm_qos_mtx);
+
+       mutex_unlock(&dev_pm_qos_sysfs_mtx);
        pm_runtime_put(dev);
 }
 EXPORT_SYMBOL_GPL(dev_pm_qos_hide_flags);