We are waiting for a client connection, channel is already there
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Uri Lublin <uril@redhat.com>
This changes tests/pki/server-cert.pem and tests/pki/ca-cert.pem to have
2048 bits. These certificates were generated using the
instructions on https://www.spice-space.org/spice-user-manual.html
The -subj args were omitted, and the defaults suggested by openssl used.
The -days parameter was changed to -days 10950, the bits to 2048.
This fixes https://gitlab.freedesktop.org/spice/spice/issues/27.
Some distros are starting to use stricter settings for their openssl
configuration, which forbids 1024 bit keys, and causes test suite
failures.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
Now that we have a refcounted RedSurfaceCmd, we can store the command
itself in DisplayChannel rather than copying QXLReleaseInfoExt. This
will let us move the release of the QXL guest resources in red-parse-qxl
in the next commit.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Currently, RedWorker is using stack-allocated variables for RedSurfaceCmd.
Surface commands are rare enough that we can dynamically allocate them
instead, and make the API in red-parse-qxl.h consistent with how other
QXL commands are handled.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Currently, RedUpdateCmd are allocated on the stack, and then
initialized/uninitialized with red_{get,put}_update_cmd
This makes the API inconsistent with what is being done for RedDrawable,
RedCursor and RedMessage. QXLUpdateCmd are not occurring very often,
we can dynamically allocate them instead, and get a consistent API.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Currently, RedMessage are allocated on the stack, and then
initialized/uninitialized with red_{get,put}_message
This makes the API inconsistent with what is being done for RedDrawable
and RedCursor. Since QXLMessage is just a (mostly unused/unsecure) debugging tool,
we can dynamically allocate it instead, and get a consistent API.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Currently, the cursor channel is allocating RedCursorCmd instances itself, and then
calling into red-parse-qxl.h to initialize it, and doing something
similar when releasing the data. This commit moves this common code to
red-parse-qxl.[ch]
The ref/unref are not strictly needed, red_cursor_cmd_free() would
currently be enough, but this makes the API consistent with
red_drawable_{new,ref,unref}.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
At the moment, we'll unconditionally release the guest QXL resources in
red_put_drawable() even if red_get_drawable() failed and did not
initialize drawable->release_info_ext properly.
This commit only sets RedDrawable::qxl once the guest resource have been
successfully retrieved, and only free the guest QXL resources when
RedDrawable::qxl is set.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Rather than needing to call red_drawable_new() and then initialize it
with red_get_drawable(), we can improve slightly red_drawable new so
that red_drawable_{new,ref,unref} is all which is used by code out of
red-parse-qxl.c.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
RedDrawable really is a RedDrawCmd which is parsed by red-parse-qxl.h
Moreover, red_drawable_ref() is already defined inline in
red-parse-qxl.h, and red_drawable_unref() is declared there too even if
its code is still in red-worker.c
This commit moves them close to the other functions creating/unref'ing
QXL commands parsed by red-parse-qxl.h.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Just a style change, on more recent GLib would print a more
friendly error report.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Uri Lublin <uril@redhat.com>
Acked-by: Victor Toso <victortoso@redhat.com>
Follow all other char devices implementation (spicevmc, agent,
stream-device) and set the char device state when
connected/disconnected. This allows qemu to discard writes, optimize a
bit the source polling, and will trigger HUP events.
See related qemu "char/spice: discard write() if backend is
disconnected".
Note: sif->state() should probably be handled at the char-device
level. I am not sure what the smartcard channel really brings over
plain spicevmc...
Signed-off-by: Marc-André Lureau <marcandre.lureau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Without this it's not obvious that a compression setting took effect.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Is possible that port 5913 is already in use as tests that uses
test_new will attempt to use ports from 5912 to 5921 so use a port
not in that range.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Eduardo Lima (Etrunko) <etrunko@redhat.com>
test_new function attempts to detect attempts to listen to tcp ports
already in listening state detecting some messages during
spice_server_init. However the check is wrong (broken in recent
34a44d3e94 "test-display-base: Avoid spurious errors due to listen
failures") and incomplete (missing message).
To better test this conditions put some of the ports in listening
state (like with a "nc -l 5912 & nc -l 5913 &" command) and run
tests in parallel (like with a "make check -j" command).
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Eduardo Lima (Etrunko) <etrunko@redhat.com>
To set up a listening socket usually you call in sequence:
- socket;
- bind;
- listen.
If you try to bind() to a port when another socket is already
listening on that port, the bind() will fail.
However, it is possible that the bind() may succeed and the listen()
will fail, as demonstrated in the following sequence:
- socket() create socket 1;
- bind() to port N on socket 1;
- socket() create socket 2;
- bind() to port N on socket 2;
- listen() on socket 1;
- listen() on socket 2 <-- failure.
When running tests (especially multiple tests running in parallel), it
may sometimes happen that there are other tests already listening on
the port that we are trying to use. In this case, we want to ignore
this error and simply try to listen on a different port. We already
attempted to handle this scenario, but we were only ignoring bind()
errors and not listen() errors. So in the scenario mentioned above,
the listen() error was causing the entire test to fail instead of
allowing us to try to listen on another port.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Jonathon Jongsma <jjongsma@redhat.com>
Just cosmetic changes, the static function had underscores to
distinguish from the exported one which was recently renamed.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Lukáš Hrázký <lhrazky@redhat.com>
Adds a function to create a write buffer for sending a message to
vdagent from the server to prevent code duplication.
Signed-off-by: Lukáš Hrázký <lhrazky@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Instead of having a single red_char_device_write_buffer_get function to
get both client and server buffers and decide by testing client == NULL,
have separate function for a client and for a server. The situation
should always be clear (you're either on the client or on the server
side) and you shouldn't need to parametrize that.
For the server case, add a use_token parameter instead of a separate
red_char_device_write_buffer_get_server_no_token function, as you may
want to parametrize that.
Signed-off-by: Lukáš Hrázký <lhrazky@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Avoid time adjustment issues.
For instance ticket validity can change when daylight time changes.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Lukáš Hrázký <lhrazky@redhat.com>
Do not convert RedStreamSslStatus enumeration type back to int.
This allows compilers to perform some more type safe checks.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
This is a thin wrapper over g_get_monotonic_time_ms, and is called only
once, so we can call directly g_get_monotonic_time_ms instead.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Having a single QXL interface is not enough, there can be other (e.g.
streaming) display channels that make the tablet unusable. Add a check for the
number of display channels also being equal to 1. We still need the check for
QXL interaces, because the tablet only works with QXL.
Signed-off-by: Lukáš Hrázký <lhrazky@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
fix below compile error:
format '%d' expects argument of type 'int', but argument 6 has
type 'long unsigned int' [-Werror=format=]
spice compile failed on 32bit system, since upstream commit
9541cd2fe change %ld to %PRIdPTR, %PRIdPTR is %d, but argument
strm.total_out is uLong.
Signed-off-by: Changqing Li <changqing.li@windriver.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
In a comparison with current autotools build system, meson/ninja
provides a huge improvement in build speed, while keeping the same
functionalities currently available and being considered more user
friendly.
The new system coexists within the same repository with the current one,
so we can do more extensive testing of its functionality before deciding
if the old system can be removed, or for some reason, has to stay for
good.
- Meson: https://mesonbuild.com
This is the equivalent of autogen/configure step in autotools. It
generates the files that will be used by ninja to actually build the
source code.
The project has received lots of traction recently, with many GNOME
projects willing to move to this new build system. The following wiki
page has more details of the status of the many projects being ported:
https://wiki.gnome.org/Initiatives/GnomeGoals/MesonPorting
Meson has a python-like syntax, easy to read, and the documentation
on the project is very complete, with a dedicated page on how to port
from autotools, explaining how most common use cases can be
implemented using meson.
http://mesonbuild.com/Porting-from-autotools.html
Other important sources of information:
http://mesonbuild.com/howtox.htmlhttp://mesonbuild.com/Syntax.htmlhttp://mesonbuild.com/Reference-manual.html
- Ninja: https://ninja-build.org
Ninja is the equivalent of make in an autotools setup, which actually
builds the source code. It has being used by large and complex
projects such as Google Chrome, Android and LLVM. There is not much to
say about ninja (other than it is much faster than make) because we
won't interact directly with it as much, as meson does the middle man
job here. The reasoning for creating ninja in the first place is
explained on the following post:
http://neugierig.org/software/chromium/notes/2011/02/ninja.html
Also its manual provides more in-depth information about the design
principles:
https://ninja-build.org/manual.html
- Basic workflow:
Meson package is available for most if not all distros, so, taking
Fedora as an example, we only need to run:
# dnf -y install meson ninja-build.
With Meson, building in-tree is not possible at all, so we need to
pass a directory as argument to meson where we want the build to be
done. This has the advantage of creating builds with different options
under the same parent directory, e.g.:
$ meson ./build --prefix=/usr
$ meson ./build-extra -Dextra-checks=true -Dalignment-checks=true
After configuration is done, we call ninja to actually do the build.
$ ninja -C ./build
$ ninja -C ./build install
Ninja defaults to parallel builds, and this can be changed with the -j
flag.
$ ninja -j 10 -C ./build
- Hacking:
* meson.build: Mandatory for the project root and usually found under
each directory you want something to be built.
* meson_options.txt: Options that can interfere with the result of the
build.
Signed-off-by: Eduardo Lima (Etrunko) <etrunko@redhat.com>
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Victor Toso <victortoso@redhat.com>
This commit needs an unreleased version of spice-protocol.
The revert is temporary in order to get the spice-server 0.14.1 release
out.
This reverts commit 9f5859c3ba.
The structure usage was removed from commit
2ba69f9f88
("libspice: add surface 0 support").
They were never used by Qemu.
Signed-off-by: Uri Lublin <uril@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Fixes test-stream-device after adding a log warning about an invalid
message received on the stream device, glib tests fail on unexpected
warning messages.
Signed-off-by: Lukáš Hrázký <lhrazky@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
Following commit fcaf8d1a1, build from tarballs/make distcheck is broken
as spice-server-enums.h is not regenerated when building from tarballs,
and we don't have a -I$(top_srcdir) in our build flags, just
-I$(srcdir). This commit changes #include <server/spice-server-enums.h>
to #include <spice-server-enums.h> which avoids the problem fixed by
commit fcaf8d1a1 without breaking make distcheck.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
When using #include "spice-server-enums.h", it will be looked up first
in the directory containing the file being build, which is going to be
$srcdir when dcc.c includes it. However, spice-server-enums.h is a
generated file, so it will be in $builddir, not in $srcdir. This most of
the time won't be causing any problems, except when you happen to have
an invalid spice-server-enums.h in $srcdir and you are doing an
out-of-tree build.
Using #include <spice-server-enums.h> instead allows the correct
spice-server-enums.h file to be looked up.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Lukáš Hrázký <lhrazky@redhat.com>
The reason for this commit is that Meson expects all submodules to be
placed in this subdirectory, and since autotools build is more flexible
in this case, we make some small adjustments to configure.ac and
Makefile.am files to accommodate for this change.
Signed-off-by: Eduardo Lima (Etrunko) <etrunko@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Rather than using
#ifdef COMPRESS_DEBUG
spice_info(...);
#endif
we can #define COMPRESS_DEBUG to spice_debug() or to do nothing for a
slight readability improvement. This opportunity is used to replace
these spice_debug() calls with g_debug(). The "do nothing" macro is a bit
convoluted to ensure that we will have a compile-time check for our
g_debug args.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
There was a small regression introduced in get_compression_for_bitmap()
by f401eb07f dcc: Rewrite dcc_image_compress.
If SPICE_IMAGE_COMPRESSION_AUTO_GLZ is specified, and the bitmap has a
stride which is bigger than its width (ie it has padding), then
get_compression_for_bitmap() will return SPICE_IMAGE_COMPRESSION_OFF
while in that case, we used to use QUIC for compression.
This happens because that function in the AUTO_GLZ case first checks if
QUIC should be used, if not, it decides to use GLZ, but then decides it
can't because of the stride, so falls back to OFF, while it used to
fall back to QUIC.
This commit only slightly reworks a preexisting if (!can_lz_compress())
check so that it's unconditional rather than depending on the previous
checks having been unsuccessful.
This issue could be observed by using a spice-html5 without support for
uncompressed bitmaps with end-of-line padding by simply starting a f28
VM and connecting to it/moving the mouse cursor in it.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Check that data sent to device are collapsed in a single message.
The StreamChannel object is mocked in the test.
This checks that commit dcc3f995d9
("stream-device: handle_data: send whole message") is doing the
right thing.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
If guest sent an empty data message this was not parsed correctly.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
Leak detectors did not manage to find leaks, possibly as double list
have all elements likely with a pointer to them.
The reference from the agent is necessary for inserting it into
the list.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
CentOS 7 compiler generate this warning:
glz-encoder-dict.c: In function 'glz_dictionary_pre_encode':
glz-encoder-dict.c:516:30: error: 'prev_seg_id' may be used uninitialized in this function [-Werror=maybe-uninitialized]
dict->window.segs[prev_seg_id].next = seg_id;
^
glz-encoder-dict.c:492:22: note: 'prev_seg_id' was declared here
uint32_t seg_id, prev_seg_id;
^
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
G_PID_FORMAT was only added in glib 2.50.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Eduardo Lima (Etrunko) <etrunko@redhat.com>
SPICE expects to have each frame in a single message.
So far the stream-device did not do that.
That works fine for H264 streams but not for MJPEG.
The client handles by itself MJPEG streams, and not via
gstreamer, and is assuming that a message contains the
whole frame. Since it currently not, using spice-streaming-agent
with MJPEG plugin, confuses the client which burns CPU
till it fails and keeps complaining:
"GSpice-CRITICAL **: 15:53:36.984: need more input data"
This patch fixes that, by reading the whole message from the
device (the streaming agent) and sending it over to the client.
Signed-off-by: Uri Lublin <uril@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
It's currently using a GPLv2+ header, which was probably a mistake given
the project overall license. It was created by a Red Hat employee, and
only modified by Red Hat employees since then, so the (c) Red Hat is
correct, and there are no other copyright holders to contact.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
self_bitmap flag is used for some complex drawing not possible
by QXL_DRAW_COPY commands. Having this flag set causes
spice-server do draw part of the screen, copy that part on new
allocated image and reduce network optimisations with no visual
changes.
Some drivers (like Windows 10 DOD) set this flag by mistake for
this command so reset it.
More details follow.
The self_bitmap flag is used for some drawing command requiring to mix
the frame buffer with some other image. For this specific
QXL_DRAW_COPY command self_bitmap is used by spice-server code during
cachine/sending (the reason for the cache is to cache images sent to
client so the relationship between the two parts of the code).
However the self_bitmap_image (an image created in spice-server if
this flags is set) is used only if src_bitmap of SpiceCopy structure
(the structure used to store the QXL_DRAW_COPY command inside
spice-server) is NULL. But in red_get_copy_ptr (red-parse-qxl.c, the
function that parse the QXL_DRAW_COPY command form the QXL device)
not having a src_bitmap is considered an error so the
self_bitmap_image won't be used.
Why this flag affects network performance?
When spice-server see this flag it update the frame buffer according
to the pending commands (commands to be sent or still to be drawn on
frame buffer). spice-server maintain a tree of commands used to reduce
rendering and command to send. More or less if a command is covering
other commands (for instance filling the entire screen with a single
color) the pending commands can be removed from the queue and not sent
to the client. However when an update of the frame buffer is requested
spice-server update the frame buffer removing the commands from the
tree but not from the client queue.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
Pointers to memory allocated in user space are never NULL.
The only exception can be if you explicitly map memory at zero.
There is however no reasons for such requirement and this practise
was also removed from Linux due to security reasons.
This API looks copied from a kernel environment where valid virtual
addresses can be NULL.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
In case we pass something like "spice:mjpeg$%*" the last part is
ignore making the string parse correctly.
A single pair should end by either string terminator or pair terminator.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
These functions are in the standard C library, not well known
but quite useful for parsing strings.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Jonathon Jongsma <jjongsma@redhat.com>
Now warnings are printed through g_warning which causes the test to
fail. We need to use g_test_expect_message() to prevent that failure.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
It was only used twice, for what looks like adhoc debugging. This commit
removes it, similarly to what was done for some spice_printerr() calls.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
The remaining occurrences of spice_printerr() are warnings when
something unexpected happens, they can be replaced with g_warning() so
that users of spice-server can redirect them with
g_log_set_default_handler().
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Depending on the context, we want to output a warning or just a debug
log.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
These calls seem to have been added for debugging for a very specific
purpose. At the very least, they should have been using g_debug() rather
than spice_printerr(). This commit removes these.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
With last changes are just used once and are straight forward.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Jonathon Jongsma <jjongsma@redhat.com>
Do not extract all components and compare one by one, can be easily
compared together.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Jonathon Jongsma <jjongsma@redhat.com>
The macros for both depth are the same, reuse the definition.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Jonathon Jongsma <jjongsma@redhat.com>
When a new record channel is added, the code relies on a snd_send() call
in record_channel_client_constructed() to send RECORD_START to the
client. However, at this point, snd_send() is non-functional because
the red_channel_client_pipe_add() call it makes is a no-op because
prepare_pipe_add() makes a connection check through
red_channel_client_is_connected() queueing the item. This connection
check returns FALSE at ::constructed() time as the channel client will
only become connected towards the end of
red_channel_client_initable_init() which runs after the object
instantiation is complete.
This causes a bug where starting recording and then
disconnecting/reconnecting the client does not successfully reenable
recording. This is a regression introduced by commit d8dc09
'sound: Convert SndChannelClient to RedChannelClient'
This commit solves this issue by making PlaybackChannelClient and
RecordChannelClient implement GInitable, and move the code interacting
with the client in their _initable_init() function, as at this point the
objects will be able to send data.
https://bugzilla.redhat.com/show_bug.cgi?id=1549132
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
The result of this function is always cast to a pointer, there
is no reason to return an integer.
This API looks copied from a kernel environment where virtual
addresses can have different sizes compare to pointers.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
Define JpegEncoderContext as an abstract structure.
This allows to reduce casts.
Also remove some alignment warnings on some architecture like mips.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
On some systems you need to call g_spawn_close_pid after
spawning a process to avoid leaks (currently Windows).
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
In some environment pthread.h is not defined but its definitions
are used in some headers.
Actually happens using MingW.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
Just style change. Invert the if to exit earlier.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe de Dinechin <dinechin@redhat.com>
These constants are meant to be used in format string for size_t
types. Use them for portability.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Snir Sheriber <ssheribe@redhat.com>
It was probably meant to be used as a "user_data" argument for the
various callbacks, but turns out not to be used.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Client callbacks in sound channels do not use registered
data so don't pass a valid pointer making clear from
source that the parameter is not used.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
Without an explicit call to SSL_CTX_set_ecdh_auto(reds->ctx, 1), OpenSSL
1.0 (still used by el7) would not use ECDH ciphers (this is now
automatic with OpenSSL 1.1.0). This commit adds this missing call. It's
based on a suggestion from David Jasa
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
https://bugzilla.redhat.com/show_bug.cgi?id=1566597
Most of pipe items use this name for the base field.
This also allows to use SPICE_UPCAST macros instead of a long
SPICE_CONTAINEROF.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
Although capabilities inside link message are handled as arrays
of 4 bytes unsigned integers we don't need capabilities to be
aligned to 4 bytes just to call g_memdup so use a pointer to
uint8_t instead.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
This array is just used locally in red_channel_client_handle_outgoing
so declare it there.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
Removes debug messages that are logged on every draw, spamming the log
excessively when debugging.
Signed-off-by: Lukáš Hrázký <lhrazky@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
From Gitlab CI:
=17955== 16 bytes in 1 blocks are possibly lost in loss record 725 of 2,079
==17955== at 0x4C2DBAB: malloc (vg_replace_malloc.c:299)
==17955== by 0x4011D17: tls_get_addr_tail.isra.0 (in /usr/lib64/ld-2.27.so)
==17955== by 0x4017997: __tls_get_addr (in /usr/lib64/ld-2.27.so)
==17955== by 0xEE4534B: gnutls_rnd (in /usr/lib64/libgnutls.so.30.20.2)
==17955== by 0xEE1F254: ??? (in /usr/lib64/libgnutls.so.30.20.2)
==17955== by 0xEE1F947: ??? (in /usr/lib64/libgnutls.so.30.20.2)
==17955== by 0xEE231B5: ??? (in /usr/lib64/libgnutls.so.30.20.2)
==17955== by 0xEE24D67: gnutls_handshake (in /usr/lib64/libgnutls.so.30.20.2)
==17955== by 0xEBD4FEA: ??? (in /usr/lib64/gio/modules/libgiognutls.so)
==17955== by 0x7463936: g_task_thread_pool_thread (gtask.c:1331)
==17955== by 0x7A3E932: g_thread_pool_thread_proxy (gthreadpool.c:307)
==17955== by 0x7A3DF29: g_thread_proxy (gthread.c:784)
==17955== by 0x8284563: start_thread (in /usr/lib64/libpthread-2.27.so)
==17955== by 0x859631E: clone (in /usr/lib64/libc-2.27.so)
==17955==
==17955== 32 bytes in 1 blocks are possibly lost in loss record 1,234 of 2,079
==17955== at 0x4C2DBAB: malloc (vg_replace_malloc.c:299)
==17955== by 0x4011D17: tls_get_addr_tail.isra.0 (in /usr/lib64/ld-2.27.so)
==17955== by 0x4017997: __tls_get_addr (in /usr/lib64/ld-2.27.so)
==17955== by 0xCAA5173: __cxa_get_globals (in /usr/lib64/libstdc++.so.6.0.25)
==17955== by 0xCAA6186: __cxa_throw (in /usr/lib64/libstdc++.so.6.0.25)
==17955== by 0xC601457: ??? (in /usr/lib64/libproxy.so.1.0.0)
==17955== by 0xC5F6BB6: ??? (in /usr/lib64/libproxy.so.1.0.0)
==17955== by 0xC5F7089: ??? (in /usr/lib64/libproxy.so.1.0.0)
==17955== by 0xC5F7470: px_proxy_factory_get_proxies (in /usr/lib64/libproxy.so.1.0.0)
==17955== by 0xC3E64E3: ??? (in /usr/lib64/gio/modules/libgiolibproxy.so)
==17955== by 0x7463936: g_task_thread_pool_thread (gtask.c:1331)
==17955== by 0x7A3E932: g_thread_pool_thread_proxy (gthreadpool.c:307)
==17955== by 0x7A3DF29: g_thread_proxy (gthread.c:784)
==17955== by 0x8284563: start_thread (in /usr/lib64/libpthread-2.27.so)
==17955== by 0x859631E: clone (in /usr/lib64/libc-2.27.so)
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
encode_32 already deals with endian, don't swap twice.
Tested with a ppc64 server machine and a x64 client.
This looks the reverse of a previous patch (59c6c82) supposed to fix big
endian machine. encode_32 has been always:
static inline void encode_32(Encoder *encoder, unsigned int word)
{
encode(encoder, (uint8_t)(word >> 24));
encode(encoder, (uint8_t)(word >> 16) & 0x0000ff);
encode(encoder, (uint8_t)(word >> 8) & 0x0000ff);
encode(encoder, (uint8_t)(word & 0x0000ff));
}
while encode basically is similar to a putc on a FILE stream so is writing
numbers from host endian to big endian order.
The "main" endian (the one more tested since ever) is host/guest being
little endian. So if you call encode_32 with a 0x01020304 you get 4 bytes
in the order 1, 2, 3, 4.
Before and after 59c6c82 LZ_MAGIC was defined as:
#define LZ_MAGIC (*(uint32_t *)"LZ ")
so on little endian this was 0x4c, 0x5a, 0x20, 0x20 that is 0x20205a4c
which written through encode_32 become 0x20, 0x20, 0x5a, 0x4c so we can say
that at the end on the network we must have 0x20, 0x20, 0x5a, 0x4c.
On big endian however LZ_MAGIC got the value 0x4c5a2020 which written
through encode_32 get 0x4c, 0x5a, 0x20, 0x20 which is the opposite
expected. So patch 59c6c82 reverted the order having again 0x20, 0x20,
0x5a, 0x4c on the network.
However commit 5a7e587 (spice-common), in an attempt to avoid double
swapping on LZ, changed LZ_MAGIC to
#define LZ_MAGIC 0x20205a4c
breaking endianness again for GLZ code.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
It's only called once, and when it's called, we will have dereferenced
worker->display_channel a few lines before in
display_channel_set_monitors_config_to_primary(), so this cannot be
NULL. The 'if (worker->display_channel)' check can thus be removed, so
display_is_connected() becomes just red_channel_is_connected().
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
There's already a 'display' variable equal to worker->display_channel
which is not consistently used. This commit also adds a new 'channel'
local variable to limit the number of upcasts to RedChannel.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
This is reported by GCC 8.0.1 (Fedora 28).
Instead of doing a possible invalid cast destroy and create the
queue again.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Jonathon Jongsma <jjongsma@redhat.com>
Avoid casting function pointer with different argument providing
a proper utility instead.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Jonathon Jongsma <jjongsma@redhat.com>
With GCC 8.0.1 (Fedora 28), cast to different function pointer
can lead to warnings, like:
event-loop.c: In function ‘watch_update_mask’:
event-loop.c:146:42: error: cast between incompatible function types from ‘gboolean (*)(GIOChannel *, GIOCondition, void *)’ {aka ‘int (*)(struct _GIOChannel *, enum <anonymous>, void *)’} to ‘gboolean (*)(void *)’ {aka ‘int (*)(void *)’} [-Werror=cast-function-type]
g_source_set_callback(watch->source, (GSourceFunc)watch_func, watch, NULL);
^
cc1: all warnings being treated as errors
As g_source_set_callback expect a function pointer which type
changes based on the type of source (so is expected) silent
the possible warning.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Jonathon Jongsma <jjongsma@redhat.com>
When GLZ code attempts to send a 64 bit integer the 8 top bit of
the lower (32 bits) part of the number are stripped due to a bug.
This was discovered by Zhongqiang Huang <useprxf@gmail.com>
Reported-by: Zhongqiang Huang <useprxf@gmail.com>
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Jonathon Jongsma <jjongsma@redhat.com>
Although not necessary for a single monitor DisplayChannel implementation
this make the DisplayChannels more coherent from the client
point of view.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Jonathon Jongsma <jjongsma@redhat.com>
SpiceWhiteness/SpiceBlackness/SpiceInvers are 3 typedef for the same
type, no need to have 3 identical red_put_xxx/red_get_xxx methods.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
SpiceBlend is a typedef to SpiceCopy, and red_put_blend() and
red_put_copy() are identical, so we can add a #define red_put_blend
red_put_copy similar to the one we already have for red_get_blend.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
In order to use the new TCP_CORK feature, disable auto flush.
Depending on channel implementation and purpose of the channel enabling
blindly for all channels could cause performance issues, specifically if
flush is not done at the right time.
CommonGraphicsChannel channels were tested to make sure is not that case.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
Cork is a system interface implemented by Linux and some *BSD systems to
tell the system that other data are expected to be written to a socket.
This allows the system to reduce network fragmentation waiting for network
packets to be complete.
Using some replay capture and some instrumentation resulted in a
bandwith reduction of 11% and a packet reduction of 56%.
The tests was done using replay utility so results could be a bit different
from real cases as:
- replay goes as fast as it can, for instance packets could
be merged by the kernel decreasing packet numbers and a bit
byte spent (this actually make the following improves worse);
- there are fewer channels (no much cursor, sound, etc).
The following tests shows count packet and total bytes from server to
client using a real network. I used a direct cable connection using 1gb
connection and 2 laptops.
cork: 537 1582240
cork: 681 1823754
cork: 524 1583287
cork: 538 1582350
no cork: 1329 1834630
no cork: 1290 1829094
no cork: 1289 1830164
no cork: 1317 1833589
no cork: 1320 1835705
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
The writing to network was always immediate.
Every write in the stream causes a write to the OS.
This can have some penalty if you don't write large data as network
packets can be more fragmented or you encrypt data in smaller chunks
(when data are encrypted some padding is added then data is split in
multiple of encryption block which is usually the size of encryption
key and this is done for every write).
Define an interface to allow higher levels code to tell low level when
data should be sent to remote or when can wait more data.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
The name is more consistent with red_marshall_cursor_init.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
There's an implicit API/ABI contract between QEMU and SPICE that SPICE
will keep the guest QXL resources alive as long as QEMU can hold a
pointer to them. This implicit contract was broken in 1c6e7cf7 "Release
cursor as soon as possible", causing crashes at migration time.
While the proper fix would be in QEMU so that spice-server does not need
to have that kind of knowledge regarding QEMU internal implementation,
this commit reverts to the pre-1c6e7cf7 behaviour to avoid a regression
while QEMU is being fixed.
This version of the fix is based on a suggestion from Frediano Ziglio.
https://bugzilla.redhat.com/show_bug.cgi?id=1540919
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Code can have problems reading empty messages, check we can
handle it.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe de Dinechin <dinechin@redhat.com>
Handle capabilities from guest device.
Send capability to the guest when device is opened.
Currently there's no capabilities set on the message sent.
On the tests we need to discard the capability message before
reading the error.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe de Dinechin <dinechin@redhat.com>
This function will be reused to initialise different message
headers.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe de Dinechin <dinechin@redhat.com>
Currently, red-parse-qxl.c uses g_malloc+memcpy to duplicate the cursor
data when it could use g_memdup() instead. red-stream-device.c does the
same thing but uses spice_memdup(). This commit makes use of g_memdup()
in both cases so that this memory is consistently allocated through
glib.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
The name 'com.redhat.stream.0' is too generic and in no way denotes it
belongs to SPICE. It is preferred to have the project's domain in the
name and Red Hat doesn't own the project. Rename it to
org.spice-space.stream.0.
Signed-off-by: Lukáš Hrázký <lhrazky@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
Reuse option from common code.
Also reuse spice_extra_checks constant instead of using the preprocessor
macro directly.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Eduardo Lima (Etrunko) <etrunko@redhat.com>
The timeout is too short when the test run under Valgrind
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
In order to avoid confusion with file named stream-device.h, from
spice-protocol.
Signed-off-by: Eduardo Lima (Etrunko) <etrunko@redhat.com>
Acked-by: Frediano Ziglio <fziglio@redhat.com>
These factor a bit of common code, and more importantly, help with
freeing all event loop related data at the end of each test.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
This test case will be testing the external spice-server API to
configure the address/port it's listening on. For now it sets up a
listening server, spawns a thread which is going to connect to that
port, and check it gets the REDQ magic upon connection. It will be
extended to test for Unix sockets, TLS sockets, ...
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
There is currently a debug printf which is always shown when a mainloop
event is triggered. This is unlikely to be useful unless one is
debugging the event loop code.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Currently if we fail to set up the watch waiting for accept() to be
called on the socket, we still keep the network socket(s) open even if we
are not going to be able to use it. This commit makes sure it's closed a
set to -1 when such a failure occurs rather than having a half
initialized spice-server instance.
Signed-off-by: Christophe Fergeau <cfergeau@redhat.com>
Fix clang warning:
red-record-qxl.c:893:13: error: variable 'fd_in' is used uninitialized whenever 'if' condition is false [-Werror,-Wsometimes-uninitialized]
if (ret)
This is technically impossible but is not on a hot path.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
Due to ticket expiration, it is possible that the streaming channels for
the client are created after the ticket expires. Currently, streaming
channels are created dynamically when the guest starts streaming to the
server, which can happen at any time (for instance if you decide to start
the graphic server manually).
If the ticket has expired before the streaming channel is created,
authentication will fail and the client will not be able to connect.
To avoid this, create the channels when the first main channel connection
is made. This ensures that client will connect to all streaming channels.
This could be considered a temporary solution. There may be other
situations where it would be useful to connect new channels after the
ticket has expired, but enabling this behavior would require protocol
changes and a careful analysis of security implications.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
Move public declaration (stream_device_connect) from char-device.h
to a new stream-device.h.
Add type declaration for StreamDevice.
This allows to use the type outside the implementation file and makes it
easier to extend the interface without changing char-device.h header.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
This is a preparation for meson build, which has built-in support for
generating enums, but requires the template files to be renamed. It uses
the basename of template files to generate the output, and in this case
it would be the same file for both '.c' and '.h'. Ideally meson would
let us specify the name of the output files, but this is not the case.
Without renaming, the following error happens:
Meson encountered an error in file server/meson.build, line 30, column 0:
Tried to create target "spice-server-enums.tmpl", but a target of that
name already exists.
Reference: http://mesonbuild.com/Gnome-module.html#gnomemkenums
Note that by the time of this commit, the documentation is not accurate
and does not mention the fact that output files will get the base name
of the template files if they are specified, I submitted a pull request
to meson fixing this detail in docs:
https://github.com/mesonbuild/meson/pull/3191
Signed-off-by: Eduardo Lima (Etrunko) <etrunko@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
Test case for the issue fixed by previous commit.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
If data message is followed by another message, it's theoretically
possible that device loses the sync with the guest.
The actual Qemu and streaming agent implementation avoids it, but better to
make sure this can't happen in the server code too.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
Previous patch causes a bug in Qemu if the patch
46764fe09ca2e0f15c0981a672c166ed8cf57e72 ("virtio-serial: fix segfault
on disconnect") is not included in that version of Qemu (patch present in
version 2.10.0).
This crash happens when device is closed during a write operation.
For SPICE character device, spice_server_char_device_wakeup is called
to write data which handles both read and write pending operations.
As we want to close the device but we can't do it inside the handler
without causing a crash, this commit schedules a timer that will close the
guest device outside this callback.
The proper solution would be to patch Qemu but making sure of this is not
so easy, hence this workaround in spice-server.
Code is marked with some comments to remember to remove this
hack in a safe future.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
Once the device is an error state, we don't want the guest to keep
reading/writing to it, especially as this could put the device in an
inconsistent state. This commit disables the device when an error occurs to
prevent further unintended use of the device by the guest.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
Due to the way Qemu handle the device, when an error occurs we must consume
all pending data inside the callback which reads data from the device.
If we don't flush this data, the next time spice-server tries to read from
the device (after the guest closes/reopens it), we'll be getting stale
data. This can happen because we cannot prevent the guest from writing to
the device even after it got in an error state.
This needs to be done within this callback, as QEMU returns 0 if you call
SpiceCharDeviceInterface::read() outside of it. QEMU invokes this callback
through a call to spice_server_char_device_wakeup.
On the test now we must test that we receive an error from the device.
Previously we checked that last part of the data was not read. Now
potentially all data are read, so we need another way to check the device
detected the error.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
Test all batched (send together) messages are handled correctly
and device is not stuck.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
Qemu does not trigger a new data read if we don't read all data in
the buffer.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Acked-by: Christophe Fergeau <cfergeau@redhat.com>
