This avoids having to do the sorting everytime we use typefind
The behaviour of gst_type_find_factory_get_list has subtlely changed
in the sense that the order was previously undefined, whereas now
it returns them sorted by rank and then by name.
Simply use casting macros for accessing the message fields like we do for
buffers and events. Avoids some costly typechecking that does not really buy us
much.
CAPS_IS_ANY and CAPS_IS_EMPTY are the equivalent of their gst_caps_*
counterpart except that they avoid the typechecking and are inlined.
CAPS_IS_EMPTY_SIMPLE only checks for empty caps (without checking if
the caps is ANY).
Maintain a hashtable of the plugin basename. We can then use this
hashtable to speedup the search for an existing plugin and avoid
a whole lot of strcmp calls.
This patch adds gst_caps_set_value() and allows gst_caps_set_simple() to
work on non-simple caps. See the API documentation for the functions
about what they do.
The intention of these changes is to ease working with caps in caps
transform functions. An example for this would be ffmpegcolorspace,
where the caps transform function could be changed to look roughly like
this (pseudocode ahead):
result = gst_caps_copy (template_caps);
value = gst_structure_get_value (gst_caps_get_structure (caps, 0),
"widh");
gst_caps_set_value (result, value);
/* same for height, framerate and par */
return caps;
which is much cleaner and easier to understand than the current code.
https://bugzilla.gnome.org/show_bug.cgi?id=597690
This is slightly more efficient because the compiler can't do tail
recursion here and has to keep all stack frames.
Not that efficiency is that important here but I already had
the iterative version somewhere else and both are easy to read.
This allows the macosx versions to properly error out when fds are closed.
This is only a temporary fix until the pluginloader is switched to not
use GstPoll but GIOChannels.
In the plugin loader subprocess, move stdin and stdout to new fd's
so that plugins printing things during plugin init or (*gasp*)
possibly reading from stdin don't interfere with the data sent to
and from the parent.
When the plugin-scanner load fails (because the helper can't be
spawned), make sure to load the plugin that failed in-process, so
that all plugins do get loaded.
Fix an off-by-one error in the size guard for unpack_element, and
improve various debug statements in the failure paths.
Also, swap some g_new0 to g_malloc0 for the fun of it.
In most places in core and baseclasses we just need the caps to do caps-
intersections. In that case ref'ed caps are enough (no need to copy).
This patch also switches the code to use the new functions.
API: gst_pad_get_caps_refed(), gst_pad_peer_get_caps_refed()
Everyone running an uninstalled git setup is going to wonder about
this failure next time they update, so let's mention the solution
in the error message.
Reflow the code to move error handling to the end of the functions. Initialize
gvalue like we do in the setter. Add a unit-test module with two simple tests
the catche this bug.
We don't need the hold the proxy mutex locked for getting the internal pad and
for linking the new target pad when we retarget. So take the lock a little later
and release it earlier.
Fixes#596366
Guard against a hostile child process that sends bogus data
due to memory corruption by adding a magic number to each packet,
and limit the maximum size of any message to 32MB
When trying to find a function plugin-scanner, include a check on the
version of the binary registry chunks it sends, to make sure it's
what we understand.
Add a simple version check when starting the plugin-scanner so we can
verify we're talking to one that talks the same language.
First try a plugin-scanner in the installed path, then try one via the
GST_PLUGIN_SCANNER env var if that doesn't work.
Update the uninstalled script.
Install the plugin-scanner to the libexec dir
When updating the registry, we don't need to re-read the registry cache
and waste time replacing all our existing, hopefully identical, plugins
and features that we're about to re-scan anyway.
phase 2 - make the plugin loader receive the list of plugins to load and
send back the results asynchronously, so we don't context switch back
and forth so much.
The structure_change message was originally emitted on source pads and
then recently changed to be sink pads. This causes a failure in the
gst-python testsuite. Disable the restriction so that the published
behaviour is still allowed.
Also parse the nano of the version and assume that X.Y.Z-1.1 >= X.Y.Z
With this change we can also check development versions against the version of
the upcomming release.
This is available in newer gcc releases and it should only exist
on platforms that provide some native 128bit integer arithmetic
instructions.
The x86-64 assembly for this is still kept for non-gcc compilers
that don't provide __uint128_t magic.
Post the structure change messages on the sinkpads of the elements. This allows
us to catch unlinked pads earlier without ending up with inconsistent element
degrees.
When we detect a pad unlink in progress, we will not be updating the degree of
the parent element. This can cause false loop detected warnings because the
degree counter is invalid. Handle this case by marking the iterator as 'dirty'
when we detect a pad unlink and avoid emiting the warning in this case. We have
to continue our state change as good as we can, we will eventually resync when
the pad unlink completed.
When an element is added to the bin, only set the index if we have a
cached index, rather than setting a NULL index on elements that might
have a default index object of their own.
elementfactory field is filled in by gst_element_base_class_init,
but it needs some info set on the element's type, so have it
available prior to class structure creation spinning up.
This affects elements that have a well-known/public type (e.g. pipeline)
and can be created by other means than gst_element_factory_make
(which will also fill in the element's factory).
Cache the last index that was set with _set_index() and return this in the
_get_index() call.
Set the cached index on newly added elements.
Fixes#566881
Based on clock implementation by Håvard Graff <havard.graff@tandberg.com>
Try to get the time on windows using the performance counters. These have a much
higher resolution and accuracy than the regular getcurrenttime(). Be careful to
fall back to regular getcurrenttime() or posix clocks when performance counters
are not available.
We can use a shift for scaling the denominator instead of a divide since the
denom is always positive. This avoids having the compiler generate code for the
different rounding rules when scaling negative values.
64bit x86 has native 64x64->128 bit multiply that we can use with some inline
assembler to speed up large multiplications.
Use bsr to find the number of leading zeros more efficiently.
The internal links function is deprecated since some time and
there already were GST_REMOVE_DEPRECATED markers in the source file,
now add them to the header too.
Fixes bug #592209.
Before the signal handler would get the ghostpad passed as second
argument but it could've already been unreffed and destroyed.
This would then lead to crashes and all that.
Now we get the ghostpad from the proxy pad, which we get from the
target pad as it's peer.
Fixes bug #591318.
The new functions are
gst_util_uint64_scale_int_round()
gst_util_uint64_scale_int_ceil()
gst_util_uint64_scale_round()
gst_util_uint64_scale_ceil()
Fixes bug #590919.
Before it returned that [start,stop] is inside the segment and that the
difference between segment_start and start needs to be clipped. If the
clipping is done on a buffer (like in baseaudiosink) this will result
in the data pointer being at a invalid memory position.
Fixes bug #589849.
Check for metadata writability when setting caps on buffer or when copying
metadata flags. Only enable these extra assertions in git versions.
This should help us find bad elements.
Often we don't need the result of the intersection. Add a variant that only
tries to intersect. It can break out earlier and does less GValue copying.
API: gst_caps_can_intersect()
The "album artist" tag is used when the artist of an entire
album differs from the artist of an individual track; for example,
when a "guest artist" appears on an album, or on compilations.
Fixes bug #590430.
GstTaks does not always unref the taskpool it was created from because it
depends on when the pool provided an ID for joining the task.
Rework some code so that we always unref the pool and optionally join when the
pool provided an id.
Fixes#589127
For now, don't show a g_warning() for empty tag strings and NULL
tags with non-git versions; we should wait for the fixes in our
plugin modules to make it into a release before we enable this
unconditionally.
Add a quick return if two types are the same. Change the check for the
intersection function to be the same as the one used in intersect(). The
later tries both directions.
Don't forbid the empty string "" in generic structures, only in taglists.
Properly allow the NULL string by adding special cases for serialising
and deserialising it. prop1=(string)NULL is the NULL string,
prop1=(string)"NULL" is the actual string with the value "NULL"
Add a new macro GST_DEBUG_CATEGORY_GET to get a log category by name. This
allows plugins to use e.g. core categories like PERFORMANCE or CLOCK.
API: GST_DEBUG_CATEGORY_GET
This makes the generated code faster since:
* It won't have to read an undirect value (which will most likely be
outside of the L1/L2 cache)
* We know that value never changes (the compiler has no clue that it doesn't).
It is not too obvious that getting and releasing request pads is not entierly
symetrical regarding to the pad refcount. Add a note about that to the docs.
This might deserve a FIXME-0.11 too.
Just in case someone who clearly can't be deterred by any number of leading
underscores uses this very private but still somewhat documented symbol
directly in their code (*cough* qtdemux *cough*).
Add new method to iterate a bufferlist without having to allocate an iterator.
Add convenience method for getting an item from the list based on the group and
index.
Remove redundant _do_data callback and method.
Update unit-tests and add some more for the new methods.
Add functions to create a new tag list and set tags in one go, which
is nice for use in combination with functions that take ownership of
the taglist, such as gst_event_new_tag() or gst_element_found_tags().
API: add gst_tag_list_new_full()
API: add gst_tag_list_new_full_valist()
Update design doc with step-start docs.
Add eos field to step done message
when stepping in reverse, update the segment time field.
Flush out the current step when we are flushing.
Small micro-optimisation: look up value table for fundamental types
via an array dedicated to fundamental types instead of going through
a hash table lookup. Since there can be only 255 fundamental types,
the table size/efficiency trade-off should be acceptable, esp. since
the most commonly-used types are all fundamental types. The size of
the table could probably be minimised further if needed by allocating
the table dynamically and only expanding it on demand.
GTypes are not ints and as such are not guaranteed to fit into an int
(with the exception of fundamental types), so we really shouldn't put
them into int variables. Even if a rather unlikely obscure corner case,
this has actually been a problem at some point in the past, see commit
99f16655f4.
Add a bunch of vararg getter convenience functions to complement
the vararg setter functions, and a basic unit test. Fixes#534208.
API: gst_structure_get()
API: gst_structure_id_get()
API: gst_structure_get_valist()
API: gst_structure_id_get_valist()
Don't dump hundreds of kB of hexdata into debug logs when converting
taglists containing huge images into a string. Instead, shorten the
buffer data so that the string is still readable and debug logs
stay managable. Can be turned off with GST_DEBUG_OPTIONS=full-tags.
See #584988.
The first time one calls gst_element_factory_make(), gst recreates the plugin
feature and the element factory. As a side effect we ref the class to fill
in detail we already have filled from the registry cache. This patch changes
the behaviour to just update the existing entries. The factory is now attached
to the type and set in gst_element_base_class_init().
Add a GType->GstValueTable hashtable mapping.
Avoid _get_type() multiple times when we can.
Use GSlice for fraction range dynamic memory
Add G_LIKELY when we can
Improve lookup of the value table using the hashtable
Fix some memory leaks shown by the new serialisation/deserialisation unit
test. Split the gst_string_wrap function in gstvalue.c into components and
use them to make gst_string_take_and_wrap, which takes ownership of the
string, avoiding a strdup.
Add some G_LIKELY/UNLIKELY, and clean up some leaks in error paths.
Reword this message a bit to make it clearer what it means, namely that
the state change may have failed for good reasons, but that the element
just failed to post a proper error on the bus. This is not an internal
GStreamer bug, and we really don't need people to flood bugzilla with
bug reports if one such plugin bug ever makes it into the wild.
Add gst_segment_to_position() that converts a running_time to a position in the
segment. A faulty variant of this function is currently used in inputselector
but we'll need it for frame stepping too.
API: GstSegment::gst_segment_to_position()
Use string_warp/unwrap to escape delimiters, otherwise deserialisation fails.
Also move GST_ASCII_IS_STRING to private header to avoid keeping it in sync.
Also use '\0' when terminating a string for better readability.
Update framestep document, we want to pass the flush flag in the step-done
message.
Add flush flag to the gstmessage.
Update examples to use the new step-done message api.
Implement framestep with playback rates < 0.0 too.
Add new STEP event and methods for creating/parsing the event
Update design docs.
Add new STEP_DONE message and method to create/parse.
API: GstEvent::gst_event_new_step()
API: GstEvent::gst_event_parse_step()
API: GstMessage::gst_message_new_step_done()
API: GstMessage::gst_message_parse_step_done()
Fixes minor memory leak in unit tests caused by the recent changes.
Since we're expected to take ownership of the GValue in the structure
field struct here, we need to unset it if we don't use it.
Make gst_tag_list_get_string() return FALSE for NULL strings and
empty strings, and gst_tag_list_get_pointer() return FALSE for
NULL pointers, like we do with dates and buffers.
Fixes#560345.
Also warn if an element or application tries to add a field with an
empty string to a structure (NULL strings are still needed and
allowed though) and do all those checks in the right function.
Fixes#559643.
Use gst_structure_id_empty_new() in combination with GST_QUARK
rather than gst_structure_id_new() when creating message, event,
query and taglist structures. Mostly just because we can.
Use the element START_TIME to keep track of the running time when the pipeline
paused so that it can be used to restore the base_time.
Take the start_time before setting the children to PAUSED so that we can
distribute the start_time to the children.
Add a start_time field and some methods. The start_time will contain the
running_time of when the element last went to paused. This time can be user to
report the position in PAUSED but also to do more correct clipping and
stepping later.
This reverts commit bebfde7502.
This change shouldn't be done in a stable release series as
applications are actually expecting the sender to be an
GstElement. One example is totem.
Unset the running flag after we released the lock for posting the stream-status
message. If we set the running flag to FALSE too early, the join method will
just continue without waiting for the message to be posted, leading to potential
crashes.
When writing a cache chunk fails, we were freeing the node and jump to a final
cleanup which dereferenced a null pointer. Leve freeing the node to the cleanup
code in fail_free_list. (sorry for committing wrong fix before).
When writing a cache chunk fails, we were freeing the node and jump to a final
cleanup which dereferenced a null pointer. Leve freeing the node to the cleanup
code in fail_free_list.
Move MT safety to main description (it does not belong to Return: or Since:
statement). Add a few missing return docs. Downgrade a normal comment froma doc
comment. Fix a doc header to only contain symbol name.
The padding of GstClock is a GstClockTime and not a
gpointer, so adding a pointer requires the padding
size to be changed depending on the pointer size.
Use an union instead.
Fixes bug #582878.
Add a gst_element_lost_state_full() with an extra argument to control
distribution of a new base_time. We will need this for flushing step
operations.
API: GstElement::gst_element_lost_state_full()
dot does not take the head/tail labels into account. For unfixed caps they get
quite large. Double the padding to make it sort of readable in more cases. Also
make normal font bigger and caps-label font smaller to increase our luck.
Possibly fixes GObject class creation/unref race conditions when
creating the last-message string in fakesink for events with
structures that have fields with these enum types.
If deserializing a property fails, check if the value type is a string and if so
attempt to create a bin from the string value. This allows to e.g. specify
audio-sink/video-sink for playbin on gst-launch commandline.
Reuse buffer code for bufferlists. Not sure if this measurably impacts performance
for the simple buffer case, if it does after doing some benchmarks, we can
decouple it later.
Fixes#572285
Buffer lists are a means to manage disjoint buffers as one buffer. It's also
possible to put many of those buffers into a list.
The idea is that when support is added to various elements, we will be able to
more efficiently slice and dice buffers, reduce the amount of memcpy and also
reduce data passing overhead.
The implementation is kept simple on purpose, reusing all of the memory
management features we have for miniobjects and buffers.
Access to the bufferlist object is done with an iterator, which allows for
efficient iteration and modification of the list.
See #572285
When an element lost its state but was busy doing a state change, still post the
async-start message with the base_time reset flag or else we might end up with
an old base_time.
this can happen when a sink is goin async to paused and then a flushing seek is
performed. This would cause the base_time to remain unmodified because the
async-start message was not sent.
Remove the static function set on the TaskPool before _prepare() is called and
allow for assigning a function to a Task when we _push().
Update the examples
Post a stream-status message indicating that a new task was created so that the
application has a chance to change the properties of the task.
Fix unit test to take into account the new ref of the message.
Call the leave_thread callback before we signal the thread performing the _join
so that we can be sure that the listener still has valid info in the callback.