devcoredump : Serialize devcd_del work
authorMukesh Ojha <quic_mojha@quicinc.com>
Tue, 13 Sep 2022 12:50:24 +0000 (18:20 +0530)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sat, 24 Sep 2022 12:01:40 +0000 (14:01 +0200)
In following scenario(diagram), when one thread X running dev_coredumpm()
adds devcd device to the framework which sends uevent notification to
userspace and another thread Y reads this uevent and call to
devcd_data_write() which eventually try to delete the queued timer that
is not initialized/queued yet.

So, debug object reports some warning and in the meantime, timer is
initialized and queued from X path. and from Y path, it gets reinitialized
again and timer->entry.pprev=NULL and try_to_grab_pending() stucks.

To fix this, introduce mutex and a boolean flag to serialize the behaviour.

  cpu0(X)                 cpu1(Y)

    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

                                         devcd_data_write()
                                           mod_delayed_work()
                                             try_to_grab_pending()
                                               del_timer()
                                                 debug_assert_init()
   INIT_DELAYED_WORK()
   schedule_delayed_work()
                                                   debug_object_fixup()
                                                     timer_fixup_assert_init()
                                                       timer_setup()
                                                         do_init_timer()
                                                       /*
                                                        Above call reinitializes
                                                        the timer to
                                                        timer->entry.pprev=NULL
                                                        and this will be checked
                                                        later in timer_pending() call.
                                                       */
                                                 timer_pending()
                                                  !hlist_unhashed_lockless(&timer->entry)
                                                    !h->pprev
                                                /*
                                                  del_timer() checks h->pprev and finds
                                                  it to be NULL due to which
                                                  try_to_grab_pending() stucks.
                                                */

Link: https://lore.kernel.org/lkml/2e1f81e2-428c-f11f-ce92-eb11048cb271@quicinc.com/
Signed-off-by: Mukesh Ojha <quic_mojha@quicinc.com>
Link: https://lore.kernel.org/r/1663073424-13663-1-git-send-email-quic_mojha@quicinc.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/base/devcoredump.c

index f4d794d..1c06781 100644 (file)
@@ -25,6 +25,47 @@ struct devcd_entry {
        struct device devcd_dev;
        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.
+        *
+        *
+        *
+        *        cpu0(X)                                 cpu1(Y)
+        *
+        *        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
+        *
+        *                                             devcd_data_write()
+        *                                               mod_delayed_work()
+        *                                                 try_to_grab_pending()
+        *                                                   del_timer()
+        *                                                     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.
+        */
+       struct mutex mutex;
+       bool delete_work;
        struct module *owner;
        ssize_t (*read)(char *buffer, loff_t offset, size_t count,
                        void *data, size_t datalen);
@@ -84,7 +125,12 @@ 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);
 
-       mod_delayed_work(system_wq, &devcd->del_wk, 0);
+       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);
 
        return count;
 }
@@ -112,7 +158,12 @@ static int devcd_free(struct device *dev, void *data)
 {
        struct devcd_entry *devcd = dev_to_devcd(dev);
 
+       mutex_lock(&devcd->mutex);
+       if (!devcd->delete_work)
+               devcd->delete_work = true;
+
        flush_delayed_work(&devcd->del_wk);
+       mutex_unlock(&devcd->mutex);
        return 0;
 }
 
@@ -122,6 +173,30 @@ static ssize_t disabled_show(struct class *class, struct class_attribute *attr,
        return sysfs_emit(buf, "%d\n", devcd_disabled);
 }
 
+/*
+ *
+ *     disabled_store()                                        worker()
+ *      class_for_each_device(&devcd_class,
+ *             NULL, NULL, devcd_free)
+ *         ...
+ *         ...
+ *        while ((dev = class_dev_iter_next(&iter))
+ *                                                             devcd_del()
+ *                                                               device_del()
+ *                                                                 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 parallely
+ * 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
+ * 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.
+ */
+
 static ssize_t disabled_store(struct class *class, struct class_attribute *attr,
                              const char *buf, size_t count)
 {
@@ -278,13 +353,16 @@ void dev_coredumpm(struct device *dev, struct module *owner,
        devcd->read = read;
        devcd->free = free;
        devcd->failing_dev = get_device(dev);
+       devcd->delete_work = false;
 
+       mutex_init(&devcd->mutex);
        device_initialize(&devcd->devcd_dev);
 
        dev_set_name(&devcd->devcd_dev, "devcd%d",
                     atomic_inc_return(&devcd_count));
        devcd->devcd_dev.class = &devcd_class;
 
+       mutex_lock(&devcd->mutex);
        if (device_add(&devcd->devcd_dev))
                goto put_device;
 
@@ -301,10 +379,11 @@ void dev_coredumpm(struct device *dev, struct module *owner,
 
        INIT_DELAYED_WORK(&devcd->del_wk, devcd_del);
        schedule_delayed_work(&devcd->del_wk, DEVCD_TIMEOUT);
-
+       mutex_unlock(&devcd->mutex);
        return;
  put_device:
        put_device(&devcd->devcd_dev);
+       mutex_unlock(&devcd->mutex);
  put_module:
        module_put(owner);
  free: