The idle HTTP interace is down from 30 to 10 wakeups/s (I think the remaining ones are due to HTTPd core).

 - Really yield the input (and release it) on a need-basis in the HTTP handler rather than in a custom thread. This is probably not thread-safe, but it's less unsafe than the previous code anyway.
This makes the HTTP interface thread very much useless, niark niark

LibVLC always creates the (singleton) playlist before the interfaces anyway.
Besides, the HTTP interface callbacks assumes the playlist is yield.
This removes one of two reasons for running the damn HTTP interface thread.

control/media_instance.c: Remove access to p_input_thread->p_internals. (which was a broken hack by the way).

input slave) Untested.
media_instance.c no longer depends on input_internal.h
Clean up a bit es_out pause handling.
Fixed DEMUX_GET_FPS description (double not float)

input_query_e).

Btw input_internal.h is an internal header for src/input/* ONLY.
If input/control/* need it, then it is BROKEN (or there is a missing
functionnality, just ask for it then)
 I see it calling demux2_Control, this will segfault if you are unlucky.
 Accessing directly to p_input->p member is also forbidden.

This would be necessary to rid the networking code of arbitrary timers (but it does not work yet).

Made demux2_IsPathExtension case insensitive.
Used demux2_IsPathExtension/p_demux->b_force when appropriate
Fixed a regression in flac demuxer (segfault when reading meta data)
Added a VLC_UNUSED(x) macro to avoid seeing ugly (void)x around the code.
Improved a bit a few vlc_common.h macro (do {} while(0))

Good news is, no plugins used it anyway.
Bad news is, some parts of libvlc still use which is wrong (i_refcount is protected by the "structure lock", meaning only misc/objects.c functions can use it safely).