devcoredump: Fix circular locking dependency with devcd->mutex.

The original code causes a circular locking dependency found by lockdep.

======================================================
WARNING: possible circular locking dependency detected
6.16.0-rc6-lgci-xe-xe-pw-151626v3+ #1 Tainted: G S   U
------------------------------------------------------
xe_fault_inject/5091 is trying to acquire lock:
ffff888156815688 ((work_completion)(&(&devcd->del_wk)->work)){+.+.}-{0:0}, at: __flush_work+0x25d/0x660

but task is already holding lock:

ffff888156815620 (&devcd->mutex){+.+.}-{3:3}, at: dev_coredump_put+0x3f/0xa0
which lock already depends on the new lock.
the existing dependency chain (in reverse order) is:
-> #2 (&devcd->mutex){+.+.}-{3:3}:
       mutex_lock_nested+0x4e/0xc0
       devcd_data_write+0x27/0x90
       sysfs_kf_bin_write+0x80/0xf0
       kernfs_fop_write_iter+0x169/0x220
       vfs_write+0x293/0x560
       ksys_write+0x72/0xf0
       __x64_sys_write+0x19/0x30
       x64_sys_call+0x2bf/0x2660
       do_syscall_64+0x93/0xb60
       entry_SYSCALL_64_after_hwframe+0x76/0x7e
-> #1 (kn->active#236){++++}-{0:0}:
       kernfs_drain+0x1e2/0x200
       __kernfs_remove+0xae/0x400
       kernfs_remove_by_name_ns+0x5d/0xc0
       remove_files+0x54/0x70
       sysfs_remove_group+0x3d/0xa0
       sysfs_remove_groups+0x2e/0x60
       device_remove_attrs+0xc7/0x100
       device_del+0x15d/0x3b0
       devcd_del+0x19/0x30
       process_one_work+0x22b/0x6f0
       worker_thread+0x1e8/0x3d0
       kthread+0x11c/0x250
       ret_from_fork+0x26c/0x2e0
       ret_from_fork_asm+0x1a/0x30
-> #0 ((work_completion)(&(&devcd->del_wk)->work)){+.+.}-{0:0}:
       __lock_acquire+0x1661/0x2860
       lock_acquire+0xc4/0x2f0
       __flush_work+0x27a/0x660
       flush_delayed_work+0x5d/0xa0
       dev_coredump_put+0x63/0xa0
       xe_driver_devcoredump_fini+0x12/0x20 [xe]
       devm_action_release+0x12/0x30
       release_nodes+0x3a/0x120
       devres_release_all+0x8a/0xd0
       device_unbind_cleanup+0x12/0x80
       device_release_driver_internal+0x23a/0x280
       device_driver_detach+0x14/0x20
       unbind_store+0xaf/0xc0
       drv_attr_store+0x21/0x50
       sysfs_kf_write+0x4a/0x80
       kernfs_fop_write_iter+0x169/0x220
       vfs_write+0x293/0x560
       ksys_write+0x72/0xf0
       __x64_sys_write+0x19/0x30
       x64_sys_call+0x2bf/0x2660
       do_syscall_64+0x93/0xb60
       entry_SYSCALL_64_after_hwframe+0x76/0x7e
other info that might help us debug this:
Chain exists of: (work_completion)(&(&devcd->del_wk)->work) --> kn->active#236 --> &devcd->mutex
 Possible unsafe locking scenario:
       CPU0                    CPU1
       ----                    ----
  lock(&devcd->mutex);
                               lock(kn->active#236);
                               lock(&devcd->mutex);
  lock((work_completion)(&(&devcd->del_wk)->work));
 *** DEADLOCK ***
5 locks held by xe_fault_inject/5091:
 #0: ffff8881129f9488 (sb_writers#5){.+.+}-{0:0}, at: ksys_write+0x72/0xf0
 #1: ffff88810c755078 (&of->mutex#2){+.+.}-{3:3}, at: kernfs_fop_write_iter+0x123/0x220
 #2: ffff8881054811a0 (&dev->mutex){....}-{3:3}, at: device_release_driver_internal+0x55/0x280
 #3: ffff888156815620 (&devcd->mutex){+.+.}-{3:3}, at: dev_coredump_put+0x3f/0xa0
 #4: ffffffff8359e020 (rcu_read_lock){....}-{1:2}, at: __flush_work+0x72/0x660
stack backtrace:
CPU: 14 UID: 0 PID: 5091 Comm: xe_fault_inject Tainted: G S   U              6.16.0-rc6-lgci-xe-xe-pw-151626v3+ #1 PREEMPT_{RT,(lazy)}
Tainted: [S]=CPU_OUT_OF_SPEC, [U]=USER
Hardware name: Micro-Star International Co., Ltd. MS-7D25/PRO Z690-A DDR4(MS-7D25), BIOS 1.10 12/13/2021
Call Trace:
 <TASK>
 dump_stack_lvl+0x91/0xf0
 dump_stack+0x10/0x20
 print_circular_bug+0x285/0x360
 check_noncircular+0x135/0x150
 ? register_lock_class+0x48/0x4a0
 __lock_acquire+0x1661/0x2860
 lock_acquire+0xc4/0x2f0
 ? __flush_work+0x25d/0x660
 ? mark_held_locks+0x46/0x90
 ? __flush_work+0x25d/0x660
 __flush_work+0x27a/0x660
 ? __flush_work+0x25d/0x660
 ? trace_hardirqs_on+0x1e/0xd0
 ? __pfx_wq_barrier_func+0x10/0x10
 flush_delayed_work+0x5d/0xa0
 dev_coredump_put+0x63/0xa0
 xe_driver_devcoredump_fini+0x12/0x20 [xe]
 devm_action_release+0x12/0x30
 release_nodes+0x3a/0x120
 devres_release_all+0x8a/0xd0
 device_unbind_cleanup+0x12/0x80
 device_release_driver_internal+0x23a/0x280
 ? bus_find_device+0xa8/0xe0
 device_driver_detach+0x14/0x20
 unbind_store+0xaf/0xc0
 drv_attr_store+0x21/0x50
 sysfs_kf_write+0x4a/0x80
 kernfs_fop_write_iter+0x169/0x220
 vfs_write+0x293/0x560
 ksys_write+0x72/0xf0
 __x64_sys_write+0x19/0x30
 x64_sys_call+0x2bf/0x2660
 do_syscall_64+0x93/0xb60
 ? __f_unlock_pos+0x15/0x20
 ? __x64_sys_getdents64+0x9b/0x130
 ? __pfx_filldir64+0x10/0x10
 ? do_syscall_64+0x1a2/0xb60
 ? clear_bhb_loop+0x30/0x80
 ? clear_bhb_loop+0x30/0x80
 entry_SYSCALL_64_after_hwframe+0x76/0x7e
RIP: 0033:0x76e292edd574
Code: c7 00 16 00 00 00 b8 ff ff ff ff c3 66 2e 0f 1f 84 00 00 00 00 00 f3 0f 1e fa 80 3d d5 ea 0e 00 00 74 13 b8 01 00 00 00 0f 05 <48> 3d 00 f0 ff ff 77 54 c3 0f 1f 00 55 48 89 e5 48 83 ec 20 48 89
RSP: 002b:00007fffe247a828 EFLAGS: 00000202 ORIG_RAX: 0000000000000001
RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 000076e292edd574
RDX: 000000000000000c RSI: 00006267f6306063 RDI: 000000000000000b
RBP: 000000000000000c R08: 000076e292fc4b20 R09: 0000000000000000
R10: 0000000000000000 R11: 0000000000000202 R12: 00006267f6306063
R13: 000000000000000b R14: 00006267e6859c00 R15: 000076e29322a000
 </TASK>
xe 0000:03:00.0: [drm] Xe device coredump has been deleted.

Fixes: 01daccf ("devcoredump : Serialize devcd_del work")
Cc: Mukesh Ojha <[email protected]>
Cc: Greg Kroah-Hartman <[email protected]>
Cc: Johannes Berg <[email protected]>
Cc: Rafael J. Wysocki <[email protected]>
Cc: Danilo Krummrich <[email protected]>
Cc: [email protected]
Cc: [email protected] # v6.1+
Signed-off-by: Maarten Lankhorst <[email protected]>
Cc: Matthew Brost <[email protected]>
Acked-by: Mukesh Ojha <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Greg Kroah-Hartman <[email protected]>

Author: Maarten Lankhorst

drivers/base/devcoredump.c

@@ -23,50 +23,46 @@ struct devcd_entry {
 	void *data;
 	size_t datalen;
 	/*
-	 * Here, mutex is required to serialize the calls to del_wk work between
-	 * user/kernel space which happens when devcd is added with device_add()
-	 * and that sends uevent to user space. User space reads the uevents,
-	 * and calls to devcd_data_write() which try to modify the work which is
-	 * not even initialized/queued from devcoredump.
+	 * There are 2 races for which mutex is required.
 	 *
+	 * The first race is between device creation and userspace writing to
+	 * schedule immediately destruction.
 	 *
+	 * This race is handled by arming the timer before device creation, but
+	 * when device creation fails the timer still exists.
 	 *
-	 *        cpu0(X)                                 cpu1(Y)
+	 * To solve this, hold the mutex during device_add(), and set
+	 * init_completed on success before releasing the mutex.
 	 *
-	 *        dev_coredump() uevent sent to user space
-	 *        device_add()  ======================> user space process Y reads the
-	 *                                              uevents writes to devcd fd
-	 *                                              which results into writes to
+	 * That way the timer will never fire until device_add() is called,
+	 * it will do nothing if init_completed is not set. The timer is also
+	 * cancelled in that case.
 	 *
-	 *                                             devcd_data_write()
-	 *                                               mod_delayed_work()
-	 *                                                 try_to_grab_pending()
-	 *                                                   timer_delete()
-	 *                                                     debug_assert_init()
-	 *       INIT_DELAYED_WORK()
-	 *       schedule_delayed_work()
-	 *
-	 *
-	 * Also, mutex alone would not be enough to avoid scheduling of
-	 * del_wk work after it get flush from a call to devcd_free()
-	 * mentioned as below.
-	 *
-	 *	disabled_store()
-	 *        devcd_free()
-	 *          mutex_lock()             devcd_data_write()
-	 *          flush_delayed_work()
-	 *          mutex_unlock()
-	 *                                   mutex_lock()
-	 *                                   mod_delayed_work()
-	 *                                   mutex_unlock()
-	 * So, delete_work flag is required.
+	 * The second race involves multiple parallel invocations of devcd_free(),
+	 * add a deleted flag so only 1 can call the destructor.
 	 */
 	struct mutex mutex;
-	bool delete_work;
+	bool init_completed, deleted;
 	struct module *owner;
 	ssize_t (*read)(char *buffer, loff_t offset, size_t count,
 			void *data, size_t datalen);
 	void (*free)(void *data);
+	/*
+	 * If nothing interferes and device_add() was returns success,
+	 * del_wk will destroy the device after the timer fires.
+	 *
+	 * Multiple userspace processes can interfere in the working of the timer:
+	 * - Writing to the coredump will reschedule the timer to run immediately,
+	 *   if still armed.
+	 *
+	 *   This is handled by using "if (cancel_delayed_work()) {
+	 *   schedule_delayed_work() }", to prevent re-arming after having
+	 *   been previously fired.
+	 * - Writing to /sys/class/devcoredump/disabled will destroy the
+	 *   coredump synchronously.
+	 *   This is handled by using disable_delayed_work_sync(), and then
+	 *   checking if deleted flag is set with &devcd->mutex held.
+	 */
 	struct delayed_work del_wk;
 	struct device *failing_dev;
 };
@@ -95,14 +91,27 @@ static void devcd_dev_release(struct device *dev)
 	kfree(devcd);
 }
 
+static void __devcd_del(struct devcd_entry *devcd)
+{
+	devcd->deleted = true;
+	device_del(&devcd->devcd_dev);
+	put_device(&devcd->devcd_dev);
+}
+
 static void devcd_del(struct work_struct *wk)
 {
 	struct devcd_entry *devcd;
+	bool init_completed;
 
 	devcd = container_of(wk, struct devcd_entry, del_wk.work);
 
-	device_del(&devcd->devcd_dev);
-	put_device(&devcd->devcd_dev);
+	/* devcd->mutex serializes against dev_coredumpm_timeout */
+	mutex_lock(&devcd->mutex);
+	init_completed = devcd->init_completed;
+	mutex_unlock(&devcd->mutex);
+
+	if (init_completed)
+		__devcd_del(devcd);
 }
 
 static ssize_t devcd_data_read(struct file *filp, struct kobject *kobj,
@@ -122,12 +131,12 @@ static ssize_t devcd_data_write(struct file *filp, struct kobject *kobj,
 	struct device *dev = kobj_to_dev(kobj);
 	struct devcd_entry *devcd = dev_to_devcd(dev);
 
-	mutex_lock(&devcd->mutex);
-	if (!devcd->delete_work) {
-		devcd->delete_work = true;
-		mod_delayed_work(system_wq, &devcd->del_wk, 0);
-	}
-	mutex_unlock(&devcd->mutex);
+	/*
+	 * Although it's tempting to use mod_delayed work here,
+	 * that will cause a reschedule if the timer already fired.
+	 */
+	if (cancel_delayed_work(&devcd->del_wk))
+		schedule_delayed_work(&devcd->del_wk, 0);
 
 	return count;
 }
@@ -151,11 +160,21 @@ static int devcd_free(struct device *dev, void *data)
 {
 	struct devcd_entry *devcd = dev_to_devcd(dev);
 
+	/*
+	 * To prevent a race with devcd_data_write(), disable work and
+	 * complete manually instead.
+	 *
+	 * We cannot rely on the return value of
+	 * disable_delayed_work_sync() here, because it might be in the
+	 * middle of a cancel_delayed_work + schedule_delayed_work pair.
+	 *
+	 * devcd->mutex here guards against multiple parallel invocations
+	 * of devcd_free().
+	 */
+	disable_delayed_work_sync(&devcd->del_wk);
 	mutex_lock(&devcd->mutex);
-	if (!devcd->delete_work)
-		devcd->delete_work = true;
-
-	flush_delayed_work(&devcd->del_wk);
+	if (!devcd->deleted)
+		__devcd_del(devcd);
 	mutex_unlock(&devcd->mutex);
 	return 0;
 }
@@ -179,12 +198,10 @@ static ssize_t disabled_show(const struct class *class, const struct class_attri
  *                                                                 put_device() <- last reference
  *             error = fn(dev, data)                           devcd_dev_release()
  *             devcd_free(dev, data)                           kfree(devcd)
- *             mutex_lock(&devcd->mutex);
  *
  *
  * In the above diagram, it looks like disabled_store() would be racing with parallelly
- * running devcd_del() and result in memory abort while acquiring devcd->mutex which
- * is called after kfree of devcd memory after dropping its last reference with
+ * running devcd_del() and result in memory abort after dropping its last reference with
  * put_device(). However, this will not happens as fn(dev, data) runs
  * with its own reference to device via klist_node so it is not its last reference.
  * so, above situation would not occur.
@@ -374,7 +391,7 @@ void dev_coredumpm_timeout(struct device *dev, struct module *owner,
 	devcd->read = read;
 	devcd->free = free;
 	devcd->failing_dev = get_device(dev);
-	devcd->delete_work = false;
+	devcd->deleted = false;
 
 	mutex_init(&devcd->mutex);
 	device_initialize(&devcd->devcd_dev);
@@ -383,8 +400,14 @@ void dev_coredumpm_timeout(struct device *dev, struct module *owner,
 		     atomic_inc_return(&devcd_count));
 	devcd->devcd_dev.class = &devcd_class;
 
-	mutex_lock(&devcd->mutex);
 	dev_set_uevent_suppress(&devcd->devcd_dev, true);
+
+	/* devcd->mutex prevents devcd_del() completing until init finishes */
+	mutex_lock(&devcd->mutex);
+	devcd->init_completed = false;
+	INIT_DELAYED_WORK(&devcd->del_wk, devcd_del);
+	schedule_delayed_work(&devcd->del_wk, timeout);
+
 	if (device_add(&devcd->devcd_dev))
 		goto put_device;
 
@@ -401,13 +424,20 @@ void dev_coredumpm_timeout(struct device *dev, struct module *owner,
 
 	dev_set_uevent_suppress(&devcd->devcd_dev, false);
 	kobject_uevent(&devcd->devcd_dev.kobj, KOBJ_ADD);
-	INIT_DELAYED_WORK(&devcd->del_wk, devcd_del);
-	schedule_delayed_work(&devcd->del_wk, timeout);
+
+	/*
+	 * Safe to run devcd_del() now that we are done with devcd_dev.
+	 * Alternatively we could have taken a ref on devcd_dev before
+	 * dropping the lock.
+	 */
+	devcd->init_completed = true;
 	mutex_unlock(&devcd->mutex);
 	return;
  put_device:
-	put_device(&devcd->devcd_dev);
 	mutex_unlock(&devcd->mutex);
+	cancel_delayed_work_sync(&devcd->del_wk);
+	put_device(&devcd->devcd_dev);
+
  put_module:
 	module_put(owner);
  free: