Alexandre Janniaux authored 2 years ago and Steve Lhomme committed 2 years ago

Audio latency callback is using the fifo lock, which cannot be released
until the audio output is destroyed. In general, outputs can use the
fifo, and we need to release/detach them from the input decoder before
freeing the lock.

François Cartegnie authored 2 years ago and Steve Lhomme committed 2 years ago

1111 becomes 1111000064 due to imprecision

Johannes Kauffmann authored 2 years ago and Steve Lhomme committed 2 years ago

When sout_StreamDelete is called from sout_StreamNew, p_stream->psz_name
is set to NULL, so don't pass it to msg_Dbg.

Only the base object should load all the modules descriptors.

There's no other object that can handle the configuration loading/saving.

We don't need to inherit the options, they can only be in the libvlc_int_t
object. var_GetNonEmptyString() is equivalent to var_InheritString() in that
case.

Steve Lhomme authored 2 years ago and Felix Paul Kühne committed 2 years ago

After removing the WSAEWOULDBLOCK check from the Windows implementation, now
that it's done in vlc_accept(). Both implementation are strictly the same.

Steve Lhomme authored 2 years ago and Felix Paul Kühne committed 2 years ago

It done in vlc_accept_i11e() but can be done in all cases.

Steve Lhomme authored 2 years ago and Jean-Baptiste Kempf committed 2 years ago

If a user never wants colors (s)he should not have to write --no-color at each
VLC call.

But don't use the dynamic plugin parameters when they should not exist.

Before it was only called when HAVE_DYNAMIC_PLUGINS is defined because there
were only parameters for that, but now there is more (on Windows for now).

The value of the command line parameter will need to be known early, when we
call vlc_threads_setup().

The value of the command line parameter will need to be known early, when we
call vlc_threads_setup().

It is the default on 3.0 and might be useful when QueryPerformanceCounter() is
not accurate or too slow to call.

* https://docs.microsoft.com/en-US/troubleshoot/windows-server/performance/programs-queryperformancecounter-function-perform-poorly
* https://www.virtualdub.org/blog2/entry_106.html
* https://chromium.googlesource.com/chromium/src/base/+/refs/heads/main/time/time_win.cc

Partially reverts fdbbae70. Now we link
statically with winmm. It should be on all Win7+ machines.

We also don't force a period value smaller than wPeriodMin.

Their granularity are too small to provide a smooth playback.

It's slower than the "perf" clock source and doesn't provide any extra
precision. We don't need the UTC reference it provides.

We keep it as the default source until the user/default preferences are applied
since it provides an good granularity in a time where things may happen fast
(libvlc startup). The 10/16 millisecond granularity of the other sources are
not very useful there.

This also saves a call to vlc_tick_now() on timeout.

DWORD is an unsigned long so it can use up to ULONG_MAX.

DWORD is an unsigned long so it can use up to ULONG_MAX.

First we convert the delay to int64_t (same width/sign as a vlc_tick_t) in
milliseconds from vlc_tick_t/time_t. Then we check it fits in a DWORD, aka an
unsigned long. If it doesn't fit we clip the value to ULONG_MAX.

We either return 0 or ETIMEDOUT from the loop.

Steve Lhomme authored 2 years ago and Jean-Baptiste Kempf committed 2 years ago

Even in Win7 builds we can use it if it's available on the system.

Niklas Haas authored 2 years ago and Jean-Baptiste Kempf committed 2 years ago

This was taking more locks than necessary, causing a possible deadlock,
because window events are not supposed to lock the window lock.

The display lock is more than sufficient to guard access to this field.

Fixes: 354f3aae

Steve Lhomme authored 2 years ago and Rémi Denis-Courmont committed 2 years ago

Similar to dec06aec for UAP builds.

Steve Lhomme authored 2 years ago and Rémi Denis-Courmont committed 2 years ago

We can tell the end size faster than asprintf and replace the forward slash
while filling the output string. (it is safe to replace in a UTF-8 string).

Thomas Guillem authored 2 years ago and Hugo Beauzée-Luyssen committed 2 years ago

If NULL, it continue to use default directories or the
"input-record-path" variable.

Fixes #27204

And then simplify the code locally.

Now there's only one current_thread_ctx TLS lookup for the whole duration of
vlc_tick_wait().

The killed variable is only check if it's a VLC thread and killable is true.

Not sure why the Win8+ code differs from the rest, even the memory order
loading of th->killed is different.

This is what the documentation [1] says we should use.
Kernel32.dll is just redirecting to api-ms-win-core-processthreads-l1-1-3.dll
but kernelbase.dll is not a redirection.

[1] https://docs.microsoft.com/en-us/windows/win32/api/processthreadsapi/nf-processthreadsapi-setthreaddescription

None of this code block works if th is NULL (external thread), as there is no
"killed" variable.

No need for any conversion to int internally.

Steve Lhomme authored 2 years ago and Hugo Beauzée-Luyssen committed 2 years ago

It might provide better documentation to the user. And we don't need to
reinvent the wheel to handle a default value in SelectClockSource().

Steve Lhomme authored 2 years ago and Hugo Beauzée-Luyssen committed 2 years ago

Since vlc_threads_setup is only called once with the proper libvlc_int_t, we
don't need a mechanism to check the first call, nor a special value to check
if it's the first call or not.

Steve Lhomme authored 2 years ago and Hugo Beauzée-Luyssen committed 2 years ago

Now that we always call it with a proper libvlc_int_t.

Steve Lhomme authored 2 years ago and Hugo Beauzée-Luyssen committed 2 years ago

It relies on clk.perf.freq.QuadPart which is uninitialized and unknown until
SelectClockSource() is called. So it calls vlc_threads_setup() each time until
vlc_threads_setup() is called with the libvlc instance.

Except vlc_threads_setup(NULL) will crash on the
var_InheritBool(vlc, "high-priority") call. It's broken since
d0c4c5a9

Usable candidates at startup are:
* mdate_interrupt which is a counter in 100 ns, not a wall clock
* mdate_tick which is a counter in 1 ms, not a wall clock
* mdate_perf_100ns which is a clock in 100 ns but we don't know if the system
uses that resolution
* mdate_wall which is a clock in 100 ns resolution (slowest)

mdate_wall gives the better results and should provide a safe transition with
mdate_perf/mdate_perf_100ns values.

Steve Lhomme authored 2 years ago and Hugo Beauzée-Luyssen committed 2 years ago

QueryUnbiasedInterruptTime() provides ticks in 100 ns resolution, so we can use
VLC_TICK_FROM_MSFTIME().