Emit 'cameraAdded' and 'cameraRemoved' from CameraManager to enable
hotplug and hot-unplug support in application like QCam.
To avoid use-after-free race between the CameraManager and the
application, emit the 'cameraRemoved' with the shared_ptr version
of <Camera *>. This requires to change the function signature of
CameraManager::removeCamera() API.
Also, until now, CameraManager::Private::addCamera() transfers the
entire ownership of camera shared_ptr to CameraManager using
std::move(). This patch changes the signature of Private::addCamera to
accept pass-by-value camera parameter. It is done to make it clear from
the caller point of view that the pointer within the caller will still
be valid after this function returns. With this change in, we can emit
the camera pointer via 'cameraAdded' signal without hitting a segfault.
Signed-off-by: Umang Jain <email@uajain.com>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Emit a signal whenever new MediaDevices are added to the
DeviceEnumerator. This will allow CameraManager to be notified
about the new devices and it can re-emumerate all the devices
currently present on the system.
Device enumeration by the CameraManger is an expensive operation hence,
we want one signal emission per 'x' milliseconds to notify multiple
devices additions as a single batch, by the DeviceEnumerator.
Add a \todo to investigate the support for that.
Signed-off-by: Umang Jain <email@uajain.com>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
This commit introduces no functional changes.
Split pipelines creation code into a separate function,
so that the function can be re-used for upcoming hotplug
functionality in subsequent commits.
Also, fixup correct tag for \todo.
Signed-off-by: Umang Jain <email@uajain.com>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
The pipes_ vector was initially used to store pipeline handlers
instances with the CameraManager when it cannot be referenced from
anywhere else. It was used to retrieve cameras and deleting pipeline
handlers when stopping the camera manager.
In f3695e9b09 ("libcamera: camera_manager: Register cameras with the
camera manager"), cameras started to get registered directly with camera
manager and in 5b02e03199 ("libcamera: camera: Associate cameras with
their pipeline handler") pipeline handlers started to get stored in a
std::shared_ptr<> with each camera starting to hold a strong reference
to its associated pipeline-handler. At this point, both the camera
manager and the camera held a strong reference to the pipeline handler.
Since the additional reference held by the camera manager gets released
only on cleanup(), this lurking reference held on pipeline handler did
not allow it to get destroyed even when cameras instances have been
destroyed. This situation of having a pipeline handler instance around
without having a camera may lead to problems (one of them explained
below) especially when the camera manager is still running.
It was noticed that, there was a dangling driver directory issue (tested
for UVC camera - in /sys/bus/usb/drivers/uvcvideo) on 'unbind' → 'bind'
operation while the CameraManager is running. The directories were still
kept around even after 'unbind' because of the lurking reference of
pipeline handler holding onto them. That reference would clear if and
only if the CameraManager is stopped and then only directories were
getting removed in the above stated path.
Rather than writing a fix to release the pipeline handlers' reference
from camera manager on camera disconnection, it is decided to eliminate
the pipes_ vector from CameraManager moving forwards. There is no
point in holding a reference to it from camera manager's point-of-view
at this stage. It also helps us to fix the issue as explained above.
Now that the pipeline handler instances are referenced via cameras only,
it can happen that the destruction of last the camera instance may
result in destruction of the pipeline handler itself. Such a possibility
exists in PipelineHandler::disconnect(), where the pipeline handler
itself can get destroyed while removing the camera. This is acceptable
as long as we make sure that there is no access of pipeline handler's
members later on in the code path. Address this situation and also add a
detailed comment about it.
Signed-off-by: Umang Jain <email@uajain.com>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
As the ELF parsing code uses non-const pointers to the ELF mapping, we
have to map the module in private read-write mode. This causes issues
with valgrind, due to the IPA manager mapping the module in shared
read-only mode and valgrind having trouble loading debugging symbols
later at dlopen time due to conflicting mappings.
This is likely a bug in valgrind (reported as [1]), but we can easily
work around it by using shared read-only mappings only. As such a
mapping shouldn't be less efficient than private read-write mappings,
switch the mapping type. This requires modifying the ELF parsing
functions to operate on const memory, which is a good idea anyway as
they're not supposed to modify the ELF file.
[1] https://bugs.kde.org/show_bug.cgi?id=422601
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Paul Elder <paul.elder@ideasonboard.com>
The logic of the code guarantees that the PipelineHandler pointer passed
to the RkISP1Frames constructor is an instance of PipelineHandlerRkISP1.
We can thus use static_cast<> instead of dynamic_cast<>.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
The V4L2 compatibility layer uses devnum to match video device nodes to
libcamera Cameras. Some pipeline handlers don't report a devnum for
their camera, which prevents the V4L2 compatibility layer from matching
video device nodes to these cameras. To fix this, we first allow the
camera manager to map multiple devnums to a camera. Next, we walk the
media device and entity list and tell the camera manager to map every
one of these devnums that is a video capture node to the camera.
Since we decided that all video capture nodes that belong to a camera
can be opened via the V4L2 compatibility layer to map to that camera, it
would cause confusion for users if some pipeline handlers decided that
only specific device nodes would map to the camera. To prevent this
confusion, remove the ability for pipeline handlers to declare their own
devnum-to-camera mapping. The only pipeline handler that declares the
devnum mapping is the UVC pipeline handler, so remove the devnum there.
We considered walking the media entity list and taking the devnum from
just the one with the default flag set, but we found that some drivers
(eg. vimc) don't set this flag for any entity. Instead, we take all the
video capture nodes (entities with the sink pad flag set).
Signed-off-by: Paul Elder <paul.elder@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
As the only usage of IPAManager::instance() is by the pipeline handlers
to call IPAManager::createIPA(), remove the former and make the latter
static. Update the pipeline handlers and tests accordingly.
Signed-off-by: Paul Elder <paul.elder@ideasonboard.com>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
If any ipa_context instances are destroyed after the IPAManager is
destroyed, then a segfault will occur, since the modules have been
unloaded by the IPAManager and the context function pointers have been
freed.
Fix this by making the lifetime of the IPAManager explicit, and make the
CameraManager construct and deconstruct (automatically, via a unique
pointer) the IPAManager.
Also update the IPA interface test to do the construction and
deconstruction of the IPAManager, as it does not use the CameraManager.
Signed-off-by: Paul Elder <paul.elder@ideasonboard.com>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Given how the elfSection() function uses the sub-expression
(idx * eHdr->e_shentsize)
it has effectively two (16 bits, unsigned) operands.
The sub-expression is promoted to type int (32 bits, signed) for
multiplication and then added to eHdr->e_shoff, which is uint32_t on
32-bit platforms and uint64_t on 64-bit platforms. Since eHdr->e_shoff
is unsigned, the integer conversion rules dictate that the other signed
operand (i.e. the result of aforementioned sub-expression) will be
converted to unsigned type too. This causes sign-extension for both of
the above operands to match eHdr->e_shoff's type and should be avoided.
The solution is to explicitly cast one of the operands of the
sub-expression with unsigned int type. Hence, the other operand will be
integer promoted and the resultant will also be of unsigned int type,
not requiring to bother about a sign-extension.
Reported-by: Coverity CID=280008
Reported-by: Coverity CID=280009
Reported-by: Coverity CID=280010
Signed-off-by: Umang Jain <email@uajain.com>
Reviewed-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
The libcamera source files are named after class names, using
snake_case. pixelformats.h and pixelformats.cpp don't comply with that
rule. Fix them.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
The DRM(BGRA8888)/V4L2(ARGB8888) format is not supportable by the current
configurations of VIMC.
Remove it from the list of supported configurations.
Signed-off-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Fail and return the Camera::configure() operation if any
of the stream turns out to be a nullptr even after the
PipelineHandler handler seems to have configured the config
successfully. This prevents a null-dereference below in the
loop.
Reported-by: Coverity CID=279069
Signed-off-by: Umang Jain <email@uajain.com>
Reviewed-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Signed-off-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
The reset function in staggered write was using the wrong index when
looking for the last updated camera parameters. This would cause
possibly stale exposure values to be written to the camera on a
mode switch for captures.
Signed-off-by: Naushir Patuck <naush@raspberrypi.com>
Signed-off-by: David Plowman <david.plowman@raspberrypi.com>
Reviewed-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
Signed-off-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
We don't use full stops at the end of \return directives in Doxygen
documentation. Drop it.
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Signed-off-by: Jacopo Mondi <jacopo@jmondi.org>
Extend the SimpleConverter to support scaling, with reporting of the
minimum and maximum output sizes supported for a given input size.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Report the stride when configuring the camera. The stride is retrieved
from the capture device first, and overridden by the converter if used.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
The RPiCameraData class destructor tries to stop its ipa_ instance
without making sure it has been initialized.
If the RPiCameraData gets destroyed before its ipa_ member is
initialized, for example if the sensor initialization fails during the
match() function, a nullptr dereference segfault is triggered preventing
a graceful library teardown.
Fix this by checking for ipa_ to be initialized before stopping it.
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Signed-off-by: Jacopo Mondi <jacopo@jmondi.org>
Update the properties parsing routine in the CameraSensor class to use
the newly defined V4L2 control V4L2_CID_CAMERA_ORIENTATION in place of
the downstream V4L2_CID_CAMERA_SENSOR_LOCATION which has now been
removed.
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Signed-off-by: Jacopo Mondi <jacopo@jmondi.org>
The FileDescriptor class, when constructed from a numerical file
descriptor, duplicates the file descriptor and takes ownership of the
copy. The caller has to close the original file descriptor manually if
needed. This is inefficient as the dup() and close() calls could be
avoided, but can also lead to resource leakage, as recently shown by
commit 353fc4c223 ("libcamera: v4l2_videodevice: Fix dangling file
descriptor").
In an attempt to solve this problem, implement move semantics for the
FileDescriptor constructor. The constructor taking a numerical file
descriptor is split in two variants:
- A "fd copy" constructor that takes a const lvalue reference to a
numerical file descriptor and duplicates it (corresponding to the
current behaviour).
- A "fd move" constructor that takes a rvalue reference to a numerical
file descriptor and takes ownership of it.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Naushir Patuck <naush@raspberrypi.com>
The FileDescriptor constructor used in V4L2VideoDevice::exportDmabufFd()
creates a duplicate of the fd to store in the object. The original
fd returned by the VIDIOC_EXPBUF ioctl was never closed, and left
dangling. This would cause out of memory conditions if the camera stream
was repeatedly started and stopped.
This change closes the original fd explicitly, fixing the leak.
Signed-off-by: Naushir Patuck <naush@raspberrypi.com>
Reviewed-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
The list of public, IPA and internal header files are stored in three
meson variables, named libcamera_api, libcamera_ipa_api and
libcamera_headers respectively. The lack of uniformity is a bit
confusing. Fix it by renaming those variables to
libcamera_public_headers, libcamera_ipa_headers and
libcamera_internal_headers.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Acked-by: Jacopo Mondi <jacopo@jmondi.org>
Acked-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
The IPA headers are installed into $prefix/include/libcamera/ipa/, but
are located in the source tree in include/ipa/. This requires files
within libcamera to include them with
#include <ipa/foo.h>
while a third party IPA would need to use
#include <libcamera/ipa/foo.h>
Not only is this inconsistent, it can create issues later if IPA headers
need to include each other, as the first form of include directive
wouldn't be valid once the headers are installed.
Fix the problem by moving the IPA headers to include/libcamera/ipa/.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Acked-by: Jacopo Mondi <jacopo@jmondi.org>
The libcamera internal headers are located in src/libcamera/include/.
The directory is added to the compiler headers search path with a meson
include_directories() directive, and internal headers are included with
(e.g. for the internal semaphore.h header)
#include "semaphore.h"
All was well, until libcxx decided to implement the C++20
synchronization library. The __threading_support header gained a
#include <semaphore.h>
to include the pthread's semaphore support. As include_directories()
adds src/libcamera/include/ to the compiler search path with -I, the
internal semaphore.h is included instead of the pthread version.
Needless to say, the compiler isn't happy.
Three options have been considered to fix this issue:
- Use -iquote instead of -I. The -iquote option instructs gcc to only
consider the header search path for headers included with the ""
version. Meson unfortunately doesn't support this option.
- Rename the internal semaphore.h header. This was deemed to be the
beginning of a long whack-a-mole game, where namespace clashes with
system libraries would appear over time (possibly dependent on
particular system configurations) and would need to be constantly
fixed.
- Move the internal headers to another directory to create a unique
namespace through path components. This causes lots of churn in all
the existing source files through the all project.
The first option would be best, but isn't available to us due to missing
support in meson. Even if -iquote support was added, we would need to
fix the problem before a new version of meson containing the required
support would be released.
The third option is thus the only practical solution available. Bite the
bullet, and do it, moving headers to include/libcamera/internal/.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Acked-by: Jacopo Mondi <jacopo@jmondi.org>
The StaggeredCtrl class, part of the Raspberry Pi pipeline handler, is
part of libcamera. Move it to the libcamera namespace to simplify usage
of libcamera APIs.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Naushir Patuck <naush@raspberrypi.com>
The StaggeredCtrl class has large functions, move them to a .cpp file
instead of inlining them all to reduce the binary size.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Naushir Patuck <naush@raspberrypi.com>
In an attempt to clarify the license terms of all files in the libcamera
project, the build system files deserve particular attention. While they
describe how the binaries are created, they are not themselves
transformed into any part of binary distributions of the software, and
thus don't influence the copyright on the binary packages. They are
however subject to copyright, and thus influence the distribution terms
of the source packages.
Most of the meson.build files would not meet the threshold of
originality criteria required for copyright protection. Some of the more
complex meson.build files may be eligible for copyright protection. To
avoid any ambiguity and uncertainty, state our intent to not assert
copyrights on the build system files by putting them in the public
domain with the CC0-1.0 license.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
Acked-by: Giulio Benetti <giulio.benetti@micronovasrl.com>
Acked-by: Jacopo Mondi <jacopo@jmondi.org>
Acked-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
Acked-by: Naushir Patuck <naush@raspberrypi.com>
Acked-by: Nicolas Dufresne <nicolas.dufresne@collabora.com>
Acked-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Acked-by: Paul Elder <paul.elder@ideasonboard.com>
Acked-by: Show Liu <show.liu@linaro.org>
The proxy worker is licensed under the GPL. It is compiled as a binary
separate from libcamera.so, and it is our understanding and intent that
the GPL license doesn't propagate to libcamera.so. However, as the
worker is executed by libcamera.so, the GPL license may cause concerns
in this context, regardless of whether the concerns are valid or not.
This uncertainty could be addressed by a combination of a legal review
and an explicit intent clarification from the copyright holders. A
simpler option is to relicense the code under the LGPL.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Acked-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
Acked-by: Jacopo Mondi <jacopo@jmondi.org>
Acked-by: Paul Elder <paul.elder@ideasonboard.com>
The simple pipeline handler rejects devices that have multiple capture
video nodes. There's no real reason to do so, a more dynamic approach is
possible as the pipeline handler already locates the video device by
walking the media graph.
Rework the match sequence by skipping any check on the video nodes, and
create the V4L2VideoDevice for the media entity at the end of the
pipeline when initializing the camera data. The V4L2VideoDevice
instances are managed by the pipeline handler itself, to avoid creating
separate instances in the camera data if multiple sensors are routed to
the same video device.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Andrey Konovalov <andrey.konovalov@linaro.org>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Add support for an optional format converter, supported by the
SimpleConverter class. If a converter is available for the pipeline, it
will be used to expose additional pixel formats.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
This new pipeline handler aims at supporting any simple device without
requiring any device-specific code. Simple devices are currently defined
as a graph made of one or multiple camera sensors and a single video
node, with each sensor connected to the video node through a linear
pipeline.
The simple pipeline handler will automatically parse the media graph,
enumerate sensors, build supported stream configurations, and configure
the pipeline, without any device-specific knowledge. It doesn't support
configuration of any processing in the pipeline at the moment, but may
be extended to support simple processing such as format conversion or
scaling in the future.
The only device-specific information in the pipeline handler is the list
of supported drivers, required for device matching. We may be able to
remove this in the future by matching with the simple pipeline handler
as a last resort option, after all other pipeline handlers have been
tried.
Signed-off-by: Martijn Braam <martijn@brixit.nl>
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Andrey Konovalov <andrey.konovalov@linaro.org>
The CameraSensor class assumes that camera sensors support the exact
same list of sizes of all media bus codes. While allowing a simpler API,
this assumption is incorrect and is blocking usage of some camera
sensors.
Relaxing the constraint is possible without changes to the CameraSensor
API syntax, but requires changing its semantics. The sizes() function
now returns the list of all sizes for all media bus codes, and the
getFormat() function now searches in all supported media bus codes. The
former is likely not the most useful option for pipeline handlers, but
the sizes() function is currently unused. Designing a better API will
require inspecting current and expected future use cases in pipeline
handlers to determine proper heuristics.
While at it, fix a small typo in an unrelated comment.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
While most camera sensors are exposed to userspace as a single subdev,
this is not always the case. A notable exception is the smiapp driver,
which exposes two or three subdevs. The subdev at the output of the
camera sensor will thus have more than one pad. Prepare for this by
removing the single-pad restriction, and locating the source pad
dynamically.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
The CameraSensor::logPrefix() function accesses the entity from the
subdev. Remove the level of indirection as we stored a pointer to the
entity internally.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
