GitLab

Four access plugins need access to the input at the moment. In all
cases, it would misbehave if the input (item) of the main access was
used.

attachment: failing seems like the correct thing. I don't see a use for
attachment from stream_UrlNew().

CDDA uses the input item to append child nodes for individual tracks.
This is wrong in case of stream_UrlNew(), as the input items would be
wrongly appended to the main access item. Like directory, FTP, etc,
CCDA should probably return a playlist in this particular case.

MMSH uses the input item for redirection. This would not work properly
for stream_UrlNew(): it would rewrite the main access URL instead of
the child stream URL. This failing seems actually safer. Handling the
redirection within the MMS plugin would probably be better.

VCDX uses the input item to fill disc-wide meta data from the GET_TITLE
control request. It should use GET_META instead. In any case, it should
not pollute the meta data of the main access.

Bugs have been filed for the last 3 cases. Fortunately, I doubt any of
these access plugins was ever used through stream_UrlNew().

In addition to p_access->psz_location also add p_access->psz_access to the contructed
stream_t *. It will allow a stream_filter to determine the protocol used in case it was
constructed with an access input.

 It fixes segfaults in src/input/stream.c (at least) when demuxers
overflow int64_t seek position, and avoid testing for negative values
everywhere.
 stream_Tell() and stream_Size() still returns signed values as too much code
depend on it.

Signed-off-by: Laurent Aimar <fenrir@videolan.org>

Introduce realloc_or_free() to src/*, and add assert() to mark unhandled ENOMEM error conditions. Allocation shrinking or otherwise handled allocations don't need realloc_or_free.

This piece of code (transparent UTF-16 to UTF-8 conversion) is wrong in
many ways though. This really does NOT belong here, and should be done
in the subsdec plugin instead.

This fixes a regression from [0d1f966a]

Christopher Rath authored Sep 02, 2009 and Rémi Denis-Courmont committed Sep 02, 2009

There is a bug in AStreamSeekStream that is causing all input modules
that support CAN_SEEK to be treated as CAN_FASTSEEK.  This in turn
causes far too many seeks in slower seeking sources which results in
performance problems when playing video.  The fix is a simple change of
CAN_SEEK to CAN_FASTSEEK.
Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>

This try to avoid vlc_object_find() as much as possible.
This is conservative, because where there is no associated parent input, we'll try to find in certain cases the parent input. This will probably be
 removed later on. Because yes, there is not necessarily a parent input for access and demux, especially if created from stream_UrlNew().

Now seeking and skipping behave exactly the same.
It also allows forward seeking in non seekable stream by reading the data.
It closes #2994.
As it is a critical section of code, it must be carefully tested.

Brought to you by M2X. Needless to mention that this is complete 100%
crap. Depending cases, it's not either ignoring errors (and e.g. leaking
memory which is very stupid in OOM situations), or worse, introducing
bugs in places where realloc() was used correctly.

And exactly one day after I WARNED EXPLICITLY AGAINST DOING THIS.

This reverts commit ec5d8611.

It probably closes #2719.

It is up to the access to refuse if it can't.
Anyway, the check was only for one path of a seek command (when using
stream_Read( NULL, size)).
This partially revert fc9f8518, please report any breakage.

It fixes a growing amount of cache as noticed by Sbastien Escudier.

 Seeking on a block boundary that is exactly the end of our already
cached data was incorrect.