For remote (or non-gl) clients, if a valid gl_draw stream exists,
then we first extract the dmabuf fd associated with the scanout and
share it with the encoder along with other key parameters such as
stride, width and height. Once the encoder finishes creating an
encoded buffer (using the dmabuf fd as input), we then send it
over to the client. And, as soon as the encoder notifies that it
is no longer using the dmabuf fd, we send a gl_draw_done async to
the application.
Cc: Frediano Ziglio <freddy77@gmail.com>
Cc: Gerd Hoffmann <kraxel@redhat.com>
Cc: Marc-André Lureau <marcandre.lureau@redhat.com>
Cc: Dongwon Kim <dongwon.kim@intel.com>
Signed-off-by: Vivek Kasireddy <vivek.kasireddy@intel.com>
Acked-by: Frediano Ziglio <freddy77@gmail.com>
Store pointers to surface object in "surfaces" and allows to
have different surfaces with same ID in memory.
The surface was keep "busy" if there was pending drawing around.
Consider the following case:
1- receive drawing command
2- queue command on DCCs
3- destroy surface
4- send draw
Previously at point 4) you would have to use a surface from
"surfaces" which was destroyed, that is we would have to maintain
the pointer (and canvas) to the surface until reference counter
was 0.
However consider this case:
1- receive drawing command
2- queue command on DCCs
3- destroy surface
4- create surface
5- send draw
What would happen in point 4) ?
We could not change the surface as it will be used by point 5).
To avoid this the code attempts to wait the commands to release the
surface. However this can be an issue, you can't force the clients
to receive pending data if network is slow.
So this patch change this allowing to create surfaces while the old
version will still be used.
This is also more clean from the reference pointer prospective,
as the reference is increased for a specific surface.
Note that now instead of checking for canvas to not be NULL a
simple check for surface pointer is enough.
Signed-off-by: Frediano Ziglio <freddy77@gmail.com>
Acked-by: Victor Toso <victortoso@redhat.com>
Mostly left on dcc-send.cpp.
Other minor too.
The change in BitmapData seems odd but the id for cached image
was not used so the only information left was the surface.
Signed-off-by: Frediano Ziglio <freddy77@gmail.com>
Acked-by: Victor Toso <victortoso@redhat.com>
Instead of computing the value inside the function to then
compute also in the caller return it.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Victor Toso <victortoso@redhat.com>
As we use reference counting is more direct to use direct pointers.
Also this will allow to have a surface in a "released state"
reducing the complexity of code destroying a surface.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Victor Toso <victortoso@redhat.com>
Found with modernize-use-nullptr
NULL in C++ is 0 whereas it is a void pointer in C. Avoids implicit
conversions.
Signed-off-by: Rosen Penev <rosenp@gmail.com>
Acked-by: Frediano Ziglio <freddy77@gmail.com>
The regression was introduced by
commit 22fc6a48e6
red-channel-client: Change GQueue into a std::list
Starts using smart pointers in the queue.
dcc_add_surface_area_image added a new pipe item in front if pos was
NULL. Before the commit red_add_lossless_drawable_dependencies passed
a pointer to the last element in the pipe adding the new item on the
back. The change caused the item to be inserted in front instead of
back. Restore the insert position.
This caused in some condition the pipe to grow uncontrollably (like
using a large notepad window on Windows 7 and moving it around).
Signed-off-by: Frediano Ziglio <freddy77@gmail.com>
Now that items are managed by constructors/destructors avoid some
manual releases.
Signed-off-by: Frediano Ziglio <freddy77@gmail.com>
Acked-by: Julien Ropé <jrope@gmail.com>
This allows to:
- reuse reference counting;
- avoid having to manually call g_free to release item memory;
- assure item is initialized;
- avoids some manual casts.
It will also allows to use smart pointers.
Signed-off-by: Frediano Ziglio <freddy77@gmail.com>
Acked-by: Julien Ropé <jrope@gmail.com>
To avoid memory errors marshaller_unref_pipe_item should be
matched with a previous call to red_pipe_item_ref.
This is correctly done but to reduce the possibility of breaking
this rule move both referencing and unreferencing into a new
RedPipeItem::add_to_marshaller method.
Signed-off-by: Frediano Ziglio <freddy77@gmail.com>
Acked-by: Julien Ropé <jrope@gmail.com>
Allows to simplify a bit marshalling code in both CursorChannel
and DisplayChannel as they use the same marshalling structure.
Signed-off-by: Frediano Ziglio <freddy77@gmail.com>
Acked-by: Julien Ropé <jrope@gmail.com>
RedCacheItem was using an union to reuse cache item memory
as a pipe item to release that specific cache item.
Instead of spreading that structure everywhere move the specific
optimization all in cache-item.tmpl.cpp.
This make also code less cluttered.
Add some comment on specific code.
Signed-off-by: Frediano Ziglio <freddy77@gmail.com>
Acked-by: Julien Ropé <jrope@gmail.com>
The patch seems pretty huge but mainly are mechanical steps:
- remove GObject declarations
- do not inherit from GObject
- add SPICE_CXX_GLIB_ALLOCATOR to avoid using C++ allocators
- CLASS_init and CLASS_constructor code goes into C++ constructor
- CLASS_dispose and CLASS_finalize code goes into C++ destructor
- g_object_new is replaced by new operator
- class members goes into virtual methods
- class parameters became argument to constructor
- use push-visibility.h and pop-visibility.h to limit visibility
- temporary use XXX_CAST for old GObject casts, they will
be replaced
- g_object_get is replaced by accessors
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
The items are send from RedChannelClient so move the callback
to a virtual function in RedChannelClient instead of RedChannel.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Make all RedChannelClient hierarchy a C++ class.
This allows to use virtual methods.
Added a normal contructor instead or properties and g_object_new.
As we remove GObject conversion macros I added a macro XXX_CAST
to create a function to replace the old macro.
They will be removed when more type safety is introduced.
There's a new SPICE_CXX_GLIB_ALLOCATOR macro in red-common.h.
This macro, added to a class define the class allocator allowing
to use, in this case, GLib for allocation. This to avoid C++ library
dependency and to initialize all structure to 0 (not all fields
are manually initialized, will be improved with more encapsulation).
Currently the methods are mainly public, access will be modified
when more encapsulation (all functions in method) are done.
Some classes are now defined in the header, C++ uses access to
limit accessibility but for efficiency and type safety/inline and
other features require types to be defined in the headers.
Some fields were moved from XxxPrivate structure to class, C++
has accessibility.
Many destructors are defined as protected to forbid the use of
stack, this as these objects uses internal reference counting
to have normal pointers. Maybe in the future pointers like
std::shared_ptr could be used instead.
Reference counting is now implemented very easily using atomic
operations.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
