gstreamer/docs/random/ds/0.9-planning2

130 lines
6.1 KiB
Text
Raw Normal View History

plans:
- The primary object that applications deal with is a GstPipeline. A
given pipeline is connected to a particular main loop (GMainLoop for
glib, etc.). Calls to gst_ functions for objects owned by that
pipeline must be done from the context of the pipeline's main loop.
Signals fired by elements are marshalled in the pipeline's main
loop.
Notably, this means the gst_ API is not necessarily thread-safe.
However, it is safe to operate on different GstPipelines from
different threads. This makes it possible, for example, for
rhythmbox to play music and gather metadata from different threads
using different pipelines. Likewise, it's also possible to do
both in the same thread.
- The primary method of scheduling an element is through a generic
'iterate()' method. The iterate method explicitly tells the core
what it is waiting for (a specific time, pads to have available
data, etc.), and the core calls the iterate method when these
"triggers" happen. GstElement subclasses will be created to
emulate 0.8-style get/chain/loop methods. Existing elements will
be converted to the new subclasses rather than implement the
iterate method directly, unless there is a compelling reason to
do so. Iterate implementations are expected not to block, ever.
Rationale: This makes it possible to create completely non-blocking
elements.
- Scheduling elements will be done in either a threaded or
non-threaded way. The idle handler that is called by a pipeline's
main loop determines which elements are ready to be iterated
(based on their triggers), and puts them into a ready queue. In
the non-threaded case, the idle handler then calls the iterate()
method on each element in the ready queue. In the threaded case,
additional helper threads (which are completely owned by the
pipeline) are used to call the iterate methods.
Note that in the threaded case, elements may not always be run
in the same thread.
Some elements are much easier to write if they run in the same
thread as the main loop (i.e., elements that are also GUI widgets).
An element flag can be set to make the manager always call the
iterate method in the manager context (i.e., in the main loop
thread). Also, elements like spider need to make core calls
which may not be allowed from other threads.
Rationale: Doing all bookkeeping in a single thread/context makes
the core code _much_ simpler. This bookkeeping takes only a
minimal amount of CPU time, less than 5% of the CPU time in a
rhythmbox pipeline. There is very little benefit to spreading
this over multiple CPUs until the number of CPUs is greater than
~16, and you have _huge_ pipelines. Also, a single-threaded
manager significantly decreases the number of locks necessary
in the core, decreasing lock contention (if any) and making it
easier to understand deadlocks (if any).
- There are essentially two types of objects/structures. One type
includes objects that are derived from GObject, and are passed in
function calls similarly to gtk. The other type includes objects
(structures, really) that are not reference counted and passed
around similar to how GstCaps works in 0.8. That is, functions
that take 'const GstCaps *' do not take ownership of the passed
object, whereas functions that take 'GstCaps *' do. Similar is
true for return values.
- The concept of GstBuffer from 0.8 will be split into two types.
One type will focus solely on holding information pertaining to
ownership of a memory area (call this GstMemBuffer), and the
other type will focus solely in transfering information between
elements (call this GstPipeBuffer). In case you get confused,
GstMemBuffers _are not_ transferred between elements, and
GstPipeBuffers _do not_ own the memory they point to.
In general, GstPipeBuffers point to (and reference) a GstMemBuffer.
GstMemBuffers are GObjects. GstPipeBuffers are structs, like
GstCaps. GstPipeBuffers have timestamps, durations, and flags.
GstMemBuffers contain read/write flags. There are no subbuffers
for either type, because they are not necessary. Essentially,
GstPipeBuffers completely replace the concept of subbuffers.
(I'd like to continue to use the name GstBuffer for GstPipeBuffers,
since its usage is much more common in elements.)
Rationale: Memory regions need an ultimate owner and reference
counting. However, chunks passed around between elements need
to be small and efficient. These goals are non-overlapping and
conflicting, and thus are inappropriate to be combined in the
same class.
- Core objects should have very few (if any) public fields. This
means that accessor macros will all be eliminated and replaced
with accessor functions.
Rationale: This makes it possible to change the core more during
an ABI-stable series.
- Remove pluggable scheduling.
Rationale: We need one good scheduler. Having multiple schedulers
is directly opposed to this goal.
- 0.8-style element states are split up. One state (AppState)
indicates what the application wants the element to be doing,
and is completely under the control of the application. The
other state (ElementState) indicates what the element is actually
doing, and is under control of the element. If the application
wants an element to pause, it sets the AppState to PAUSED, and
the element eventually changes its ElementState to PAUSED (and
fires a signal). If the element has an error or EOS, it sets
its ElementState to SOME_STATE and fires a signal, while the
AppState remains at PLAYING. The actual number and descriptions
of states has not been discussed.
Rationale: It's pretty obvious that we're mixing concepts for
elements states in 0.8.
- getcaps() methods will be replaced by an element_allowed_caps()
field in the pad. The primary reason for this is because
renegotiation only needs to happen when circumstances change.
This is more easily done by a field in GstPad and notification
of peers when this changes.
Somewhere, there's a document I wrote about completely redoing
caps.