gstreamer/markdown/design/MT-refcounting.md
Tim-Philipp Müller 09a884c85c design: refcounting: fix typo
Fixes #26.
2018-11-25 15:13:05 +00:00

16 KiB

Conventions for thread a safe API

The GStreamer API is designed to be thread safe. This means that API functions can be called from multiple threads at the same time. GStreamer internally uses threads to perform the data passing and various asynchronous services such as the clock can also use threads.

This design decision has implications for the usage of the API and the objects which this document explains.

MT safety techniques

Several design patterns are used to guarantee object consistency in GStreamer. This is an overview of the methods used in various GStreamer subsystems.

Refcounting:

All shared objects have a refcount associated with them. Each reference obtained to the object should increase the refcount and each reference lost should decrease the refcount.

The refcounting is used to make sure that when another thread destroys the object, the ones which still hold a reference to the object do not read from invalid memory when accessing the object.

Refcounting is also used to ensure that mutable data structures are only modified when they are owned by the calling code.

It is a requirement that when two threads have a handle on an object, the refcount must be more than one. This means that when one thread passes an object to another thread it must increase the refcount. This requirement makes sure that one thread cannot suddenly dispose the object making the other thread crash when it tries to access the pointer to invalid memory.

Shared data structures and writability:

All objects have a refcount associated with them. Each reference obtained to the object should increase the refcount and each reference lost should decrease the refcount.

Each thread having a refcount to the object can safely read from the object. but modifications made to the object should be preceded with a _get_writable() function call. This function will check the refcount of the object and if the object is referenced by more than one instance, a copy is made of the object that is then by definition only referenced from the calling thread. This new copy is then modifiable without being visible to other refcount holders.

This technique is used for information objects that, once created, never change their values. The lifetime of these objects is generally short, the objects are usually simple and cheap to copy/create.

The advantage of this method is that no reader/writers locks are needed. all threads can concurrently read but writes happen locally on a new copy. In most cases _get_writable() can avoid a real copy because the calling method is the only one holding a reference, which makes read/write very cheap.

The drawback is that sometimes 1 needless copy can be done. This would happen when N threads call _get_writable() at the same time, all seeing that N references are held on the object. In this case 1 copy too many will be done. This is not a problem in any practical situation because the copy operation is fast.

Mutable substructures:

Special techniques are necessary to ensure the consistency of compound shared objects. As mentioned above, shared objects need to have a reference count of 1 if they are to be modified. Implicit in this assumption is that all parts of the shared object belong only to the object. For example, a GstStructure in one GstCaps object should not belong to any other GstCaps object. This condition suggests a parent-child relationship: structures can only be added to parent object if they do not already have a parent object.

In addition, these substructures must not be modified while more than one code segment has a reference on the parent object. For example, if the user creates a GstStructure, adds it to a GstCaps, and the GstCaps is then referenced by other code segments, the GstStructure should then become immutable, so that changes to that data structure do not affect other parts of the code. This means that the child is only mutable when the parent's reference count is 1, as well as when the child structure has no parent.

The general solution to this problem is to include a field in child structures pointing to the parent's atomic reference count. When set to NULL, this indicates that the child has no parent. Otherwise, procedures that modify the child structure must check if the parent's refcount is 1, and otherwise must cause an error to be signaled.

Note that this is an internal implementation detail; application or plugin code that calls _get_writable() on an object is guaranteed to receive an object of refcount 1, which must then be writable. The only trick is that a pointer to a child structure of an object is only valid while the calling code has a reference on the parent object, because the parent is the owner of the child.

Object locking:

For objects that contain state information and generally have a longer lifetime, object locking is used to update the information contained in the object.

All readers and writers acquire the lock before accessing the object. Only one thread is allowed access the protected structures at a time.

Object locking is used for all objects extending from GstObject such as GstElement, GstPad.

Object locking can be done with recursive locks or regular mutexes. Object locks in GStreamer are implemented with mutexes which cause deadlocks when locked recursively from the same thread. This is done because regular mutexes are cheaper.

Atomic operations

Atomic operations are operations that are performed as one consistent operation even when executed by multiple threads. They do however not use the conventional aproach of using mutexes to protect the critical section but rely on CPU features and instructions.

The advantages are mostly speed related since there are no heavyweight locks involved. Most of these instructions also do not cause a context switch in case of concurrent access but use a retry mechanism or spinlocking.

Disadvantages are that each of these instructions usually cause a cache flush on multi-CPU machines when two processors perform concurrent access.

Atomic operations are generally used for refcounting and for the allocation of small fixed size objects in a memchunk. They can also be used to implement a lockfree list or stack.

Compare and swap

As part of the atomic operations, compare-and-swap (CAS) can be used to access or update a single property or pointer in an object without having to take a lock.

This technique is currently not used in GStreamer but might be added in the future in performance critical places.

Objects

Locking involved:

  • atomic operations for refcounting
  • object locking

All objects should have a lock associated with them. This lock is used to keep internal consistency when multiple threads call API function on the object.

For objects that extend the GStreamer base object class this lock can be obtained with the macros GST_OBJECT_LOCK() and GST_OBJECT_UNLOCK(). For other object that do not extend from the base GstObject class these macros can be different.

refcounting

All new objects created have the FLOATING flag set. This means that the object is not owned or managed yet by anybody other than the one holding a reference to the object. The object in this state has a reference count of 1.

Various object methods can take ownership of another object, this means that after calling a method on object A with an object B as an argument, the object B is made sole property of object A. This means that after the method call you are not allowed to access the object anymore unless you keep an extra reference to the object. An example of such a method is the _bin_add() method. As soon as this function is called in a Bin, the element passed as an argument is owned by the bin and you are not allowed to access it anymore without taking a _ref() before adding it to the bin. The reason being that after the _bin_add() call disposing the bin also destroys the element.

Taking ownership of an object happens through the process of "sinking" the object. the _sink() method on an object will decrease the refcount of the object if the FLOATING flag is set. The act of taking ownership of an object is then performed as a _ref() followed by a _sink() call on the object.

The float/sink process is very useful when initializing elements that will then be placed under control of a parent. The floating ref keeps the object alive until it is parented, and once the object is parented you can forget about it.

also see relations

parent-child relations

One can create parent-child relationships with the _object_set_parent() method. This method refs and sinks the object and assigns its parent property to that of the managing parent.

The child is said to have a weak link to the parent since the refcount of the parent is not increased in this process. This means that if the parent is disposed it has to unset itself as the parent of the object before disposing itself, else the child object holds a parent pointer to invalid memory.

The responsibilities for an object that sinks other objects are summarised as:

  • taking ownership of the object

    • call _object_set_parent() to set itself as the object parent, this call will _ref() and _sink() the object.
    • keep reference to object in a datastructure such as a list or array.
  • on dispose

    • call _object_unparent() to reset the parent property and unref the object.
    • remove the object from the list.

also see relations

Properties

Most objects also expose state information with public properties in the object. Two types of properties might exist: accessible with or without holding the object lock. All properties should only be accessed with their corresponding macros. The public object properties are marked in the .h files with /< public >/. The public properties that require a lock to be held are marked with /*< public >*/ /* with <lock_type> */, where <lock_type> can be LOCK or STATE_LOCK or any other lock to mark the type(s) of lock to be held.

Example:

in GstPad there is a public property direction. It can be found in the section marked as public and requiring the LOCK to be held. There exists also a macro to access the property.

struct _GstRealPad {
  ...
  /*< public >*/ /* with LOCK */
  ...
  GstPadDirection                direction;
  ...
};

#define GST_RPAD_DIRECTION(pad)      (GST_REAL_PAD_CAST(pad)->direction)

Accessing the property is therefore allowed with the following code example:

GST_OBJECT_LOCK (pad);
direction = GST_RPAD_DIRECTION (pad);
GST_OBJECT_UNLOCK (pad);

Property lifetime

All properties requiring a lock can change after releasing the associated lock. This means that as long as you hold the lock, the state of the object regarding the locked properties is consistent with the information obtained. As soon as the lock is released, any values acquired from the properties might not be valid anymore and can as best be described as a snapshot of the state when the lock was held.

This means that all properties that require access beyond the scope of the critial section should be copied or refcounted before releasing the lock.

Most object provide a _get_<property>() method to get a copy or refcounted instance of the property value. The caller should not wory about any locks but should unref/free the object after usage.

Example:

the following example correctly gets the peer pad of an element. It is required to increase the refcount of the peer pad because as soon as the lock is released, the peer could be unreffed and disposed, making the pointer obtained in the critical section point to invalid memory.

GST_OBJECT_LOCK (pad);
peer = GST_RPAD_PEER (pad);
if (peer)
gst_object_ref (GST_OBJECT (peer));
GST_OBJECT_UNLOCK (pad);
... use peer ...

if (peer)
  gst_object_unref (GST_OBJECT (peer));

Note that after releasing the lock the peer might not actually be the peer anymore of the pad. If you need to be sure it is, you need to extend the critical section to include the operations on the peer.

The following code is equivalent to the above but with using the functions to access object properties.

peer = gst_pad_get_peer (pad);
if (peer) {
  ... use peer ...

  gst_object_unref (GST_OBJECT (peer));
}

Example:

Accessing the name of an object makes a copy of the name. The caller of the function should g_free() the name after usage.

GST_OBJECT_LOCK (object)
name = g_strdup (GST_OBJECT_NAME (object));
GST_OBJECT_UNLOCK (object)
... use name ...

g_free (name);

or:

name = gst_object_get_name (object);

... use name ...

g_free (name);

Accessor methods

For aplications it is encouraged to use the public methods of the object. Most useful operations can be performed with the methods so it is seldom required to access the public fields manually.

All accessor methods that return an object should increase the refcount of the returned object. The caller should _unref() the object after usage. Each method should state this refcounting policy in the documentation.

Accessing lists

If the object property is a list, concurrent list iteration is needed to get the contents of the list. GStreamer uses the cookie mechanism to mark the last update of a list. The list and the cookie are protected by the same lock. Each update to a list requires the following actions:

  • acquire lock
  • update list
  • update cookie
  • release lock

Updating the cookie is usually done by incrementing its value by one. Since cookies use guint32 its wraparound is for all practical reasons is not a problem.

Iterating a list can safely be done by surrounding the list iteration with a lock/unlock of the lock.

In some cases it is not a good idea to hold the lock for a long time while iterating the list. The state change code for a bin in GStreamer, for example, has to iterate over each element and perform a blocking call on each of them potentially causing infinite bin locking. In this case the cookie can be used to iterate a list.

Example:

The following algorithm iterates a list and reverses the updates in the case a concurrent update was done to the list while iterating. The idea is that whenever we reacquire the lock, we check for updates to the cookie to decide if we are still iterating the right list.

GST_OBJECT_LOCK (lock);
/* grab list and cookie */
cookie = object->list_cookie;
list = object->list;
while (list) {
  GstObject *item = GST_OBJECT (list->data);
  /* need to ref the item before releasing the lock */
  gst_object_ref (item);
  GST_OBJECT_UNLOCK (lock);

  ... use/change item here...

  /* release item here */
  gst_object_unref (item);

  GST_OBJECT_LOCK (lock);
  if (cookie != object->list_cookie) {
    /* handle rollback caused by concurrent modification
     * of the list here */

    ...rollback changes to items...

    /* grab new cookie and list */
    cookie = object->list_cookie;
    list = object->list;
  }
  else {
    list = g_list_next (list);
  }
}
GST_OBJECT_UNLOCK (lock);

GstIterator

GstIterator provides an easier way of retrieving elements in a concurrent list. The following code example is equivalent to the previous example.

Example:

it = _get_iterator(object);
while (!done) {
    switch (gst_iterator_next (it, &item)) {
    case GST_ITERATOR_OK:

        ... use/change item here...

        /* release item here */
        gst_object_unref (item);
    break;
    case GST_ITERATOR_RESYNC:
        /* handle rollback caused by concurrent modification
    * of the list here */

    ...rollback changes to items...

    /* resync iterator to start again */
    gst_iterator_resync (it);
    break;
    case GST_ITERATOR_DONE:
    done = TRUE;
    break;
    }
}
gst_iterator_free (it);