From 0f703069296664eb7c649c837cc8bb936c3ef07f Mon Sep 17 00:00:00 2001 From: "Rafael J. Wysocki" Date: Tue, 2 Apr 2013 01:25:24 +0200 Subject: PM / QoS: Avoid possible deadlock related to sysfs access 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: [] sysfs_addrm_finish+0x31/0x60 but task is already holding lock: (dev_pm_qos_mtx){+.+.+.}, at: [] 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){+.+.+.}: [] lock_acquire+0x1aa/0x240 [] __mutex_lock_common+0x59/0x5e0 [] mutex_lock_nested+0x3f/0x50 [] dev_pm_qos_update_flags+0x3f/0xc0 [] pm_qos_remote_wakeup_store+0x3f/0x70 [] dev_attr_store+0x13/0x20 [] sysfs_write_file+0xfa/0x150 [] __kernel_write+0x81/0x150 [] write_pipe_buf+0x4d/0x80 [] splice_from_pipe_feed+0x7c/0x120 [] __splice_from_pipe+0x45/0x80 [] splice_from_pipe+0x4c/0x70 [] default_file_splice_write+0x18/0x30 [] do_splice_from+0x83/0xb0 [] direct_splice_actor+0x1e/0x20 [] splice_direct_to_actor+0xe7/0x200 [] do_splice_direct+0x4c/0x70 [] do_sendfile+0x169/0x300 [] SyS_sendfile64+0x64/0xb0 [] tracesys+0xe1/0xe6 -> #0 (s_active#54){++++.+}: [] __lock_acquire+0x15bf/0x1e50 [] lock_acquire+0x1aa/0x240 [] sysfs_deactivate+0x122/0x1a0 [] sysfs_addrm_finish+0x31/0x60 [] sysfs_hash_and_remove+0x7f/0xb0 [] sysfs_unmerge_group+0x51/0x70 [] pm_qos_sysfs_remove_flags+0x14/0x20 [] __dev_pm_qos_hide_flags+0x30/0x70 [] dev_pm_qos_constraints_destroy+0x35/0x250 [] dpm_sysfs_remove+0x11/0x50 [] device_del+0x3f/0x1b0 [] device_unregister+0x48/0x60 [] usb_hub_remove_port_device+0x1c/0x20 [] hub_disconnect+0xdd/0x160 [] usb_unbind_interface+0x67/0x170 [] __device_release_driver+0x87/0xe0 [] device_release_driver+0x29/0x40 [] bus_remove_device+0x148/0x160 [] device_del+0x14f/0x1b0 [] usb_disable_device+0xf9/0x280 [] usb_set_configuration+0x268/0x840 [] usb_remove_store+0x4c/0x80 [] dev_attr_store+0x13/0x20 [] sysfs_write_file+0xfa/0x150 [] do_loop_readv_writev+0x4d/0x90 [] do_readv_writev+0xf9/0x1e0 [] vfs_writev+0x3a/0x60 [] SyS_writev+0x50/0xd0 [] 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 Signed-off-by: Rafael J. Wysocki --- drivers/base/power/qos.c | 60 +++++++++++++++++++++++++++++++++++++----------- 1 file changed, 47 insertions(+), 13 deletions(-) (limited to 'drivers/base') diff --git a/drivers/base/power/qos.c b/drivers/base/power/qos.c index 5f74587ef25..71671c42ef4 100644 --- a/drivers/base/power/qos.c +++ b/drivers/base/power/qos.c @@ -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); -- cgit v1.2.3