[libcamera-devel] [PATCH v11] test: gstreamer: Add test for gstreamer single stream
Laurent Pinchart
laurent.pinchart at ideasonboard.com
Wed Aug 18 13:35:27 CEST 2021
Hi Vedant,
On Wed, Aug 18, 2021 at 04:54:45PM +0530, Vedant Paranjape wrote:
> On Wed, Aug 18, 2021 at 12:04 AM Laurent Pinchart wrote:
> > On Tue, Aug 17, 2021 at 12:24:05PM -0400, Nicolas Dufresne wrote:
> > > Le samedi 14 août 2021 à 02:46 +0300, Laurent Pinchart a écrit :
> > > > From: Vedant Paranjape <vedantparanjape160201 at gmail.com>
> > > >
> > > > This patch adds a test to test if single stream using libcamera's
> > > > gstreamer element works.
> > > >
> > > > We need to work around two distinct issues with ASan when enabled in the
> > > > build:
> > > >
> > > > - glib has a known leak at initialization time. This is covered by the
> > > > suppression file shipped with glib, but it's not clear how to use it
> > > > automatically. For now, disable leak detection to avoid test failures.
> > >
> > > Are Valgrind suppression usable for Asan ? Glib installs it into:
> > >
> > > /usr/share/glib-2.0/valgrind/glib.supp
> > >
> > > For GStreamer suppressions, we don't install them (yet, under discussion). They
> > > are located in each repos. I would enable GStreamer leak tracer though, it does
> > > not have such a high run-time overhead, and will detect any GObject or
> > > GstMiniObject leak, as long as the test calls gst_deinit() at the end (all tests
> > > should).
> >
> > Yes, I think they are usable. I was however unsure whether the path to
> > the supression file was standard, especially considering embedded
> > systems. I thus went for a workaround. It would be nice if there was a
> > standard mechanism to pick suppressions automatically.
> >
> > > > - GStreamer spawns a child process to scan plugins. If GStreamer is
> > > > compiled without ASan (which is likely) but libcamera is, dlopen()ing
> > > > the libcamera plugin will cause an ASan link order verification
> > > > failure. Disable the verification child processes to work around the
> > > > problem. This requires gcc 8 or newer.
> > >
> > > Have you considered simply disabling forks ? See gst_registry_fork_set_enabled()
> >
> > No, as I had no idea that existed :-)
>
> Tested this, it works as expected. Now need to figure out to patch
> gstreamer as it still leaks memory
>
> =================================================================
> ==676137==ERROR: LeakSanitizer: detected memory leaks
>
> Direct leak of 16384 byte(s) in 1 object(s) allocated from:
> #0 0x7f4faf3b8bc8 in malloc
> (/usr/lib/x86_64-linux-gnu/libasan.so.5+0x10dbc8)
> #1 0x7f4fae763e98 in g_malloc
> (/usr/lib/x86_64-linux-gnu/libglib-2.0.so.0+0x57e98)
>
> SUMMARY: AddressSanitizer: 16384 byte(s) leaked in 1 allocation(s).
> ――――――――――――――――――――――――――――――――――――――――
That's a known glib issue, and that's why ASan leak detection is
disabled (in addition to disabling ASan link order verification, which
could be dropped in favour of using gst_registry_fork_set_enabled()).
glib has a valgrind suppression file that could be used by ASan as well,
but I wasn't sure how to integrate it correctly in a way that would work
on all distributions.
> > > Signed-off-by: Vedant Paranjape <vedantparanjape160201 at gmail.com>
> > > > Reviewed-by: Paul Elder <paul.elder at ideasonboard.com>
> > > > Reviewed-by: Kieran Bingham <kieran.bingham at ideasonboard.com>
> > > > Tested-by: Kieran Bingham <kieran.bingham@@ideasonboard.com>
> > > > Reviewed-by: Laurent Pinchart <laurent.pinchart at ideasonboard.com>
> > > > Signed-off-by: Laurent Pinchart <laurent.pinchart at ideasonboard.com>
> > > > ---
> > > > This version incorporates changes coming from my review of v10, and
> > > > fixes for ASan issues. ASan is a bit of a pain with GStreamer, for two
> > > > independent reasons as explained in the commit message. I'd like to find
> > > > a way to use leak suppression files to handle the glib initialization
> > > > leak, but that's a big rabbit hole. The workaround for the second issue
> > > > is acceptable in my opinion.
> > > >
> > > > The biggest trouble with these workarounds is that they don't work with
> > > > gcc version older than 8. As the stable version of the most common Linux
> > > > distributions ship gcc 8 or newer, skipping this test for gcc 7 (which
> > > > is the oldest version that libcamera supports) is also acceptable in my
> > > > opinion.
> > > >
> > > > Changes since v10:
> > > >
> > > > - Disable ASan leak detection
> > > > - Disable ASan link order verification for child processes
> > > > - Include source_path.h
> > > > - Remove unneeded explicit std::string construction
> > > > - Declare variables at usage site
> > > > - Make constants constexpr
> > > > - Add a variable for the message type
> > > > - Blank space fixes
> > > > ---
> > > > .../gstreamer_single_stream_test.cpp | 188 ++++++++++++++++++
> > > > test/gstreamer/meson.build | 19 ++
> > > > test/meson.build | 1 +
> > > > 3 files changed, 208 insertions(+)
> > > > create mode 100644 test/gstreamer/gstreamer_single_stream_test.cpp
> > > > create mode 100644 test/gstreamer/meson.build
> > > >
> > > > diff --git a/test/gstreamer/gstreamer_single_stream_test.cpp b/test/gstreamer/gstreamer_single_stream_test.cpp
> > > > new file mode 100644
> > > > index 000000000000..e26673b3471a
> > > > --- /dev/null
> > > > +++ b/test/gstreamer/gstreamer_single_stream_test.cpp
> > > > @@ -0,0 +1,188 @@
> > > > +/* SPDX-License-Identifier: GPL-2.0-or-later */
> > > > +/*
> > > > + * Copyright (C) 2021, Vedant Paranjape
> > > > + *
> > > > + * ipa_interface_test.cpp - Test the IPA interface
> > > > + */
> > > > +
> > > > +#include <iostream>
> > > > +#include <unistd.h>
> > > > +
> > > > +#include <libcamera/base/utils.h>
> > > > +
> > > > +#include "libcamera/internal/source_paths.h"
> > > > +
> > > > +#include <gst/gst.h>
> > > > +
> > > > +#include "test.h"
> > > > +
> > > > +using namespace std;
> > > > +
> > > > +extern "C" {
> > > > +const char *__asan_default_options()
> > > > +{
> > > > + /*
> > > > + * Disable leak detection due to a known global variable initialization
> > > > + * leak in glib's g_quark_init(). This should ideally be handled by
> > > > + * using a suppression file instead of disabling leak detection.
> > > > + */
> > > > + return "detect_leaks=false";
> > >
> > > Have you consider GST_TRACERS=leaks ? It will heck for refcounted object leaks
> > > on gstreamer/glib side, with very low overhead, and will abort on gst_deinit()
> > > if it found some leaks.
> >
> > Again, no idea it existed :-) It could be enabled in init() below.
> >
> > > > +}
> > > > +}
> > > > +
> > > > +class GstreamerSingleStreamTest : public Test
> > > > +{
> > > > +protected:
> > > > + int init() override
> > > > + {
> > > > + /*
> > > > + * GStreamer spawns a process to run the gst-plugin-scanner
> > > > + * helper. If libcamera is compiled with ASan enabled, and as
> > > > + * GStreamer is most likely not, this will cause the ASan link
> > > > + * order check to fail when gst-plugin-scanner dlopen()s the
> > > > + * plugin as many libraries will have already been loaded by
> > > > + * then. Work around this issue by disabling the link order
> > > > + * check. This will only affect child processes, as ASan is
> > > > + * already loaded for this process by the time this code is
> > > > + * executed, and should thus hopefully be safe.
> > > > + *
> > > > + * This option is not available in gcc older than 8, the only
> > > > + * option in that case is to skip the test.
> > > > + */
> > > > +#if defined(__SANITIZE_ADDRESS__) && !defined(__clang__) && __GNUC__ < 8
> > > > + return TestSkip;
> > > > +#endif
> > > > + setenv("ASAN_OPTIONS", "verify_asan_link_order=0", 1);
> > >
> > > I think gst_registry_fork_set_enabled(FALSE) would be much cleaner, what do you
> > > think ?
> >
> > If it does the job and has no drawback, that's fine with me.
> >
> > > > +
> > > > + /* Initialize GStreamer */
> > > > + GError *errInit = nullptr;
> > > > + if (!gst_init_check(nullptr, nullptr, &errInit)) {
> > > > + g_printerr("Could not initialize GStreamer: %s\n",
> > > > + errInit ? errInit->message : "unknown error");
> > > > + if (errInit)
> > > > + g_error_free(errInit);
> > > > +
> > > > + return TestFail;
> > > > + }
> > > > +
> > > > + /*
> > > > + * Remove the system libcamera plugin, if any, and add the
> > > > + * plugin from the build directory.
> > > > + */
> > > > + GstRegistry *registry = gst_registry_get();
> > > > + GstPlugin *plugin = gst_registry_lookup(registry, "libgstlibcamera.so");
> > > > + if (plugin) {
> > > > + gst_registry_remove_plugin(registry, plugin);
> > > > + gst_object_unref(plugin);
> > > > + }
> > > > +
> > > > + std::string path = libcamera::utils::libcameraBuildPath()
> > > > + + "src/gstreamer";
> > > > + if (!gst_registry_scan_path(registry, path.c_str())) {
> > > > + g_printerr("Failed to add plugin to registry\n");
> > > > + gst_deinit();
> > > > + return TestFail;
> > > > + }
> > > > +
> > > > + /* Create the elements */
> > > > + libcameraSrc_ = gst_element_factory_make("libcamerasrc", "libcamera");
> > > > + convert0_ = gst_element_factory_make("videoconvert", "convert0");
> > > > + sink0_ = gst_element_factory_make("fakesink", "sink0");
> > >
> > > Please use fakevideosink instead.
> >
> > That's what Vedant was doing, and the element wasn't present on Kieran's
> > system. For my own culture, what's the advantage of using fakevideosink
> > over fakesink ?
> >
> > > > +
> > > > + /* Create the empty pipeline_ */
> > > > + pipeline_ = gst_pipeline_new("test-pipeline");
> > > > +
> > > > + if (!pipeline_ || !convert0_ || !sink0_ || !libcameraSrc_) {
> > >
> > > pipeline_ cannot be NULL, glib will abort if malloc failed.
> >
> > Good point.
> >
> > > > + g_printerr("Not all elements could be created. %p.%p.%p.%p\n",
> > > > + pipeline_, convert0_, sink0_, libcameraSrc_);
> > > > + if (pipeline_)
> > > > + gst_object_unref(pipeline_);
> > > > + if (convert0_)
> > > > + gst_object_unref(convert0_);
> > > > + if (sink0_)
> > > > + gst_object_unref(sink0_);
> > > > + if (libcameraSrc_)
> > > > + gst_object_unref(libcameraSrc_);
> > >
> > > Use local g_autoptr() and g_steal_pointer() on success perhaps ? This will
> > > reduce a lot the about of error prone code in failure case. Note that in other
> > > software I've seen, an abort is used instead, as clean exit is just coding
> > > overhead for a test.
> > >
> > > An alternative, add them immediately to the pipeline, they are not owned by your
> > > class anyway.
> > >
> > > > + gst_deinit();
> > > > +
> > > > + return TestFail;
> > > > + }
> > > > +
> > > > + return TestPass;
> > > > + }
> > > > +
> > > > + void cleanup() override
> > > > + {
> > > > + gst_object_unref(pipeline_);
> > > > + gst_deinit();
> > > > + }
> > > > +
> > > > + int run() override
> > > > + {
> > > > + GstStateChangeReturn ret;
> > > > +
> > > > + /* Build the pipeline */
> > > > + gst_bin_add_many(GST_BIN(pipeline_), libcameraSrc_, convert0_, sink0_, NULL);
> > >
> > > This can fail (e.g. duplicate name).
> > >
> > > > + if (gst_element_link_many(libcameraSrc_, convert0_, sink0_, NULL) != TRUE) {
> > > > + g_printerr("Elements could not be linked.\n");
> > > > + return TestFail;
> > > > + }
> > > > +
> > > > + /* Start playing */
> > > > + ret = gst_element_set_state(pipeline_, GST_STATE_PLAYING);
> > > > + if (ret == GST_STATE_CHANGE_FAILURE) {
> > > > + g_printerr("Unable to set the pipeline to the playing state.\n");
> > > > + return TestFail;
> > > > + }
> > > > +
> > > > + /* Wait until error or EOS or timeout after 2 seconds */
> > > > + constexpr GstMessageType msgType =
> > > > + static_cast<GstMessageType>(GST_MESSAGE_ERROR | GST_MESSAGE_EOS);
> > > > + constexpr GstClockTime timeout = 2000000000;
> > >
> > > Perhaps 2 * GST_SECOND.
> >
> > Much better.
> >
> > > > +
> > > > + g_autoptr(GstBus) bus = gst_element_get_bus(pipeline_);
> > > > + g_autoptr(GstMessage) msg = gst_bus_timed_pop_filtered(bus, timeout, msgType);
> > > > +
> > > > + gst_element_set_state(pipeline_, GST_STATE_NULL);
> > > > +
> > > > + /* Parse error message */
> > > > + if (msg == NULL)
> > > > + return TestPass;
> > >
> > > I would like some minimal validation. I would expect that after 2s some frames
> > > got "rendered" properly. You can read the GstStructure property "stats" from
> > > fakevideosink / fakesink, and read the "rendered" field. Make sure this not zero
> > > perhaps ?
> >
> > Very good idea.
> >
> > Vedant, I think there's work for you :-)
> >
> > > > +
> > > > + switch (GST_MESSAGE_TYPE(msg)) {
> > > > + case GST_MESSAGE_ERROR:
> > > > + gstreamer_print_error(msg);
> > > > + break;
> > > > + case GST_MESSAGE_EOS:
> > > > + g_print("End-Of-Stream reached.\n");
> > > > + break;
> > > > + default:
> > > > + g_printerr("Unexpected message received.\n");
> > > > + break;
> > > > + }
> > > > +
> > > > + return TestFail;
> > > > + }
> > > > +
> > > > +private:
> > > > + void gstreamer_print_error(GstMessage *msg)
> > > > + {
> > > > + GError *err;
> > > > + gchar *debug_info;
> > > > +
> > > > + gst_message_parse_error(msg, &err, &debug_info);
> > > > + g_printerr("Error received from element %s: %s\n",
> > > > + GST_OBJECT_NAME(msg->src), err->message);
> > > > + g_printerr("Debugging information: %s\n",
> > > > + debug_info ? debug_info : "none");
> > > > + g_clear_error(&err);
> > > > + g_free(debug_info);
> > > > + }
> > > > +
> > > > + GstElement *pipeline_;
> > > > + GstElement *libcameraSrc_;
> > > > + GstElement *convert0_;
> > > > + GstElement *sink0_;
> > > > +};
> > > > +
> > > > +TEST_REGISTER(GstreamerSingleStreamTest)
> > > > diff --git a/test/gstreamer/meson.build b/test/gstreamer/meson.build
> > > > new file mode 100644
> > > > index 000000000000..b99aa0da0ba3
> > > > --- /dev/null
> > > > +++ b/test/gstreamer/meson.build
> > > > @@ -0,0 +1,19 @@
> > > > +# SPDX-License-Identifier: CC0-1.0
> > > > +
> > > > +if not gst_enabled
> > > > + subdir_done()
> > > > +endif
> > > > +
> > > > +gstreamer_tests = [
> > > > + ['single_stream_test', 'gstreamer_single_stream_test.cpp'],
> > > > +]
> > > > +gstreamer_dep = dependency('gstreamer-1.0', required: true)
> > > > +
> > > > +foreach t : gstreamer_tests
> > > > + exe = executable(t[0], t[1],
> > > > + dependencies : [libcamera_private, gstreamer_dep],
> > > > + link_with : test_libraries,
> > > > + include_directories : test_includes_internal)
> > > > +
> > > > + test(t[0], exe, suite : 'gstreamer', is_parallel : false)
> > > > +endforeach
> > > > diff --git a/test/meson.build b/test/meson.build
> > > > index 3bceb5df586f..d0466f17d7b6 100644
> > > > --- a/test/meson.build
> > > > +++ b/test/meson.build
> > > > @@ -11,6 +11,7 @@ subdir('libtest')
> > > >
> > > > subdir('camera')
> > > > subdir('controls')
> > > > +subdir('gstreamer')
> > > > subdir('ipa')
> > > > subdir('ipc')
> > > > subdir('log')
--
Regards,
Laurent Pinchart
More information about the libcamera-devel
mailing list