- 24 Sep, 2022 40 commits
-
-
Hans de Goede authored
Drop an unnecessary first_streamoff check from atomisp_streamoff(), above the check there is a: if (!first_streamoff) goto stop_sensor; Code block which will jump over the code with the test, so the test is only executed when first_streamoff is true and therefor the test is not necessary. Reviewed-by:
Andy Shevchenko <andriy.shevchenko@linux.intel.com> Signed-off-by:
Hans de Goede <hdegoede@redhat.com> Signed-off-by:
Mauro Carvalho Chehab <mchehab@kernel.org>
-
Hans de Goede authored
The __atomisp_reqbufs(), __atomisp_streamoff() are 1:1 wrappers for the non __ prefixed functions now, drop these wrappers. The atomisp_s_fmt_cap() wrapper is almost a 1:1 wrapper for atomisp_set_fmt() adjust the latter to have the right function prototype and drop the wrapper. Reviewed-by:
-
Hans de Goede authored
Set video_dev.lock to point to isp->mutex so that the core does the locking surroundig ioctls for us and drop all the now no longer necessary (and conflicting) locking from the ioctl handling code. Reviewed-by:
-
Hans de Goede authored
Now that __atomisp_streamoff() no longer drops isp->mutex to cancel the watchdog timer, the streamoff_mutex is no longer necessary to avoid multiple streamoffs racing with each other. Reviewed-by:
-
Hans de Goede authored
During the first __atomisp_streamoff() call on an asd with only one pipe streaming asd->streaming would get set twice: asd->streaming = ATOMISP_DEVICE_STREAMING_STOPPING; asd->streaming = ATOMISP_DEVICE_STREAMING_DISABLED; Rework the code a bit so that it gets set to the correct value right away instead of doing this in 2 steps. Reviewed-by:
-
Hans de Goede authored
Both callers of __atomisp_css_recover() check atomisp_streaming_count() first, move the check into __atomisp_css_recover(). And __atomisp_css_recover() already calls lockdep_assert_held(&isp->mutex), so drop that from atomisp_css_flush(). Reviewed-by:
-
Hans de Goede authored
The watchdog timer code to recover from the ISP getting stuck has several major issues: 1. There is no way to do fault injection and normally the ISP does not get stuck, so is it is impossible to test it. 2. It in essence just stops all streams, resets the ISP and then brings everything back up. Userspace can easily do this itself by using a timeout on dqbuf and then closing (which causes a poweroff) + re-opening the device. Doing this in userspace (if it ever turns out to be necessary) greatly simplifies the kernel code and in general will be a more robust solution. Even just a quick look at the code finds several more issues: 3. The need to sync-cancel the timers + work on streamoff requires isp->mutex to be dropped halfway during the ioctl opening all sorts of races. 4. The atomisp code supports setting up 2 pipelines, streaming from two sensors at the same time. But there is only a single wdt_work and stopping one of the 2 streams will cancel the timers + work, stopping the wdt even though the other stream might still be running. 5. In case atomisp_css_flush() the sync cancel is done while keeping isp->mutex locked, causing a deadlock when racing with wdt_work which also takes isp->mutex. 6. Even though the watchdog is purely a software/driver thing which just checkes that new frames keep coming in, there are 2 completely different implementations for the ISP2400/ISP2401 which is not necessary at all. So all in all I believe that it is better to just remove the current watchdog implementation. Fixing all the issues with the current implementation will be so much work, that if it turns out that we do need something like this then doing a clean re-implementation from scratch will be better anyways. wdt_work was also (ab)used to reset the ISP after the firmware signalled an fw-assert error through the irq, add a new assert_recover_work to replace this. Reviewed-by:
-
Hans de Goede authored
Several of the ioctl handlers all do the same checks (isp->fatal_error and asd->streaming errors) add an atomisp_pipe_check() helper for this. Note this changes the vidioc_s_fmt_vid_cap and vidioc_s_input handlers to now reject calls made while asd->streaming==STOPPING. This fixes a possible race where one thread can make this ioctls while vidioc_streamoff is running from another thread and it has temporarily released isp->mutex to kill the watchdog timers / work. Reviewed-by:
-
Hans de Goede authored
At probe time isp_subdev_init_entities() sets pipe->asd to a non NULL value for all four (preview/vf/capture/capture_video) pipes by calling atomisp_init_subdev_pipe() for all 4 pipes. So it can never be NULL. Remove the redundant NULL checks. Reviewed-by:
-
Hans de Goede authored
For reading / writing the asd->streaming enum the following rules should be followed: 1. Writers of streaming must hold both isp->mutex and isp->lock. 2. Readers of streaming need to hold only one of the two locks. Not all writers where properly taking both locks this fixes this. In the case of the readers, many readers depend on their caller to hold isp->mutex, add asserts for this And in the case of atomisp_css_get_dis_stat() it is called with isp->mutex held, so there is no need to take the spinlock just for reading the streaming value. Reviewed-by:
-
Hans de Goede authored
The spin-lock embedded in struct atomisp_sub_device is not used anywhere, remove it. Reviewed-by:
-
Hans de Goede authored
There is no reason for atomisp to use a rt_mutex instead of a normal mutex, so switch over to a normal mutex. All the changes in this patch are just s/rt_mutex/mutex/. This is a preparation patch for switching the ioctl locking over to using the video_dev.lock member so that the v4l2-core takes care of the locking. Signed-off-by:
-
Dan Carpenter authored
The "height" and "width" values come from the user so the "height * width" multiplication can overflow. Link: https://lore.kernel.org/r/YxBBCRnm3mmvaiuR@kili Fixes: a49d2536 ("staging/atomisp: Add support for the Intel IPU v2") Signed-off-by:
Dan Carpenter <dan.carpenter@oracle.com> Reviewed-by:
Andy Shevchenko <andriy.shevchenko@intel.com> Signed-off-by:
-
Hans de Goede authored
v4l2_fh_open() can only fail with -ENOMEM and as a generic rule drivers do not log their own errors for -ENOMEM since the kernel will already have complained loudly about this before the -ENOMEM return. Remove the unnecessary error logging from atomisp_open(). Reviewed-by:
-
Hans de Goede authored
When atomisp_open() fails then it must call v4l2_fh_release() to undo the results of v4l2_fh_open(). Reviewed-by:
-
Hans de Goede authored
Now that the registering of the /dev/* video / subdev nodes has been moved to the end of atomisp_pci_probe() the workaround with the loading mutex to delay opens until init is done is no longer necessary. Reviewed-by:
-
Hans de Goede authored
Register /dev/* nodes at the end of atomisp_pci_probe(), this is a prerequisite for dropping the loading mutex + ready flag kludge for delaying open() calls on the /dev/* nodes . Reviewed-by:
-
Hans de Goede authored
Split subdev and video-node registration into 2 steps, this is a preparation step for moving video-node registration to the end of probe() so that the loading() mutex can be removed. Reviewed-by:
-
Hans de Goede authored
atomisp_css_set_cont_prev_start_time() is a no-op, remove it. Reviewed-by:
-
Hans de Goede authored
Remove some more code which is no longer referenced after the removal of the ATOMISP_ACC_* custom ioctls. Reviewed-by:
-
Hans de Goede authored
The ACC /dev/video# device node uses a struct video_device embedded in an atomisp_acc_pipe struct instead of in an atomisp_video_pipe struct. Yet it uses the same file-ops and ioctl-ops even though it does not have a videobuf queue, which makes e.g. the mmap fop nonsense. Worse the only file-ops / ioctls which differentiate between the 2 types and correctly do container_of on the right type are the open/release fops and the vidioc_default handler. The mmap and poll fops and *all* other ioctl handlers unconditionally do container_of on the passed in struct video_device blindly assuming they are dealing with the one embedded in the atomisp_video_pipe struct. This makes it trivial for userspace to cause all sort of undefined behavior by calling mmap, poll or the other ioctls on the ACC device node! Presumably the use of the ACC device node was to allow making the special ioctls to load custom programs while the other /dev/video# nodes were already open, since the /dev/video# nodes can currently all be opened only once (which needs to be fixed). commit 4bbca788 ("media: atomisp: remove private acceleration ioctls") has removed the custom ATOMISP_ACC_* ioctls, so there no longer is any reason to keep the ACC device node. As explained above its presence can easily cause the kernel to crash, so remove the ACC device node and the code for handling it. Reviewed-by:
-
Hans de Goede authored
After the file-injection support removal the file_input flag is always false. Remove the flag and replace any code checking it with the code-path for when it is false. Reviewed-by:
-
Hans de Goede authored
After the file-injection support removal the outq videobuf queue is no longer used, remove it. Reviewed-by:
-
Hans de Goede authored
After the file-injection support removal, atomisp_video_pipe->type never is V4L2_BUF_TYPE_VIDEO_OUTPUT anymore, so the V4L2_BUF_TYPE_VIDEO_OUTPUT support path in atomisp_video_init() is never hit and this path is the only user of atomisp_file_fops and atomisp_file_ioctl_ops. Remove atomisp_file_fops and atomisp_file_ioctl_ops and all of the functions which are only referenced by these ops structs. Reviewed-by:
-
Hans de Goede authored
The file-injection support of the atomisp driver has not been tested and is not necessary for camera support, remove it. Note the main reason for removing this is because it depends on the videobuf (version 1) outq and we want to remove or replace all videobuf usage in the driver. Reviewed-by:
-
Hans de Goede authored
atomisp_subdev_register_entities() had V4L2_CAP_VIDEO_CAPTURE / V4L2_CAP_VIDEO_OUT swapped. Or-ing in V4L2_CAP_VIDEO_OUT for the nodes which allow capturing from the camera and or-ing in V4L2_CAP_VIDEO_CAPTURE for the file-injection node (mem2mem use of the ISP). Things happen to still work for the capture device-nodes because the "shared" caps also included V4L2_CAP_VIDEO_CAPTURE, so those shared nodes advertised V4L2_CAP_VIDEO_CAPTURE | V4L2_CAP_VIDEO_OUT. Fix things so that only the correct caps are advertised. Reviewed-by:
-
Hans de Goede authored
The atomisp code needs USERPTR pointers to be page aligned, otherwise bad things (scribbling over other parts of the process' RAM) happen. Add a check to ensure this and exit VIDIOC_QBUF calls with unaligned pointers with -EINVAL. Reviewed-by:
Andy Shevchenko <andy.shevchenko@gmail.com> Signed-off-by:
-
Hans de Goede authored
alloc_user_pages() is only ever called on qbuf for USERPTR buffers which always hits the get_user_pages_fast() path, so the pin_user_pages() path can be removed. Getting the vma then also is no longer necessary since that is only done to determine which path to use. And this also removes the only users of the mem_type struct hmm_bo member, so remove that as well. Reviewed-by:
-
Hans de Goede authored
Rewrite free_private_pages() using pages_array helper funcs. Reviewed-by:
-
Hans de Goede authored
Rewrite alloc_private_pages() using pages_array helper funcs. Note alloc_pages_bulk_array() skips non NULL pages, so switch the allocating of the pages pointer array to kcalloc to ensure the pages are initially all set to NULL. Reviewed-by:
-
Hans de Goede authored
Further simplify alloc_private_pages(). Reviewed-by:
-
Hans de Goede authored
Since lack_mem starts initialized to true, alloc_private_pages() will always set order to HMM_MIN_ORDER aka 0 / will always alloc 1 page at a time. So all the magic to decrease order if allocs fail is not necessary and can be removed. Reviewed-by:
-
Hans de Goede authored
atomisp_try_fmt() gives results with padding included. So when userspace asks for e.g. 1600x1200 then we should pass 1616x1216 to atomisp_try_fmt() this will then get adjusted back to 1600x1200 before returning it to userspace by the atomisp_adjust_fmt() call at the end of atomisp_try_fmt(). This fixes the resolution list in camorama showing resolutions like e.g. 1584x1184 instead of 1600x1200. Reviewed-by:
-
Hans de Goede authored
atomisp_try_fmt() calls the sensor's try_fmt handler but it does not copy the result back to the passed in v4l2_pix_format under some circumstances. Potentially returning an unsupported resolution to userspace, which VIDIOC_TRY_FMT is not supposed to do. atomisp_set_fmt() also uses atomisp_try_fmt() and relies on this wrong behavior. The VIDIOC_TRY_FMT call passes NULL for the res_overflow argument where as the atomisp_set_fmt() call passes non NULL. Use the res_overflow argument to differentiate between the 2 callers and always propagate the sensors result in the VIDIOC_TRY_FMT case. This fixes the resolution list in camorama showing resolutions like e.g. 1584x1184 instead of 1600x1200. Reviewed-by:
-
Hans de Goede authored
Add info about sensors v4l2_get_subdev_hostdata() use, to notes.txt. Reviewed-by:
-
Hans de Goede authored
Exit with an error on any i2c-write errors, rather then only exiting with an error when ov2680_get_intg_factor() fails. Reviewed-by:
-
Hans de Goede authored
On ov2680_set_fmt() calls with format->which == V4L2_SUBDEV_FORMAT_TRY, ov2680_set_fmt() does not talk to the sensor, so there is no need to lock the dev->input_lock mutex in this case. Reviewed-by:
-
Hans de Goede authored
On sets actually store the set (closest) format inside ov2680_device.dev, so that it also properly gets returned by get_fmt. This fixes the following problem: 1. App does an VIDIOC_SET_FMT 640x480, calling ov2680_set_fmt() 2. Internal buffers (atomisp_create_pipes_stream()) get allocated at 640x480 size by atomisp_set_fmt() 3. ov2680_get_fmt() gets called later on and returns 1600x1200 since ov2680_device.dev was not updated. So things get configured to stream at 1600x1200, but the internal buffers created during atomisp_create_pipes_stream() do not get updated in size 4. streaming starts, internal buffers overflow and the entire machine freezes eventually due to memory being corrupted Reviewed-by:
-
Andy Shevchenko authored
The acpi_evaluate_dsm_typed() provides a way to check the type of the object evaluated by _DSM call. Use it instead of open coded variant. Link: https://lore.kernel.org/r/20220730155905.90091-1-andriy.shevchenko@linux.intel.comSigned-off-by:
Sakari Ailus <sakari.ailus@linux.intel.com> Tested-by:
-
Krzysztof Kozlowski authored
Convert the Samsung Exynos SoC G-Scaler bindings to DT schema. Changes done during conversion: 1. A typical (already used) properties like clocks, iommus and power-domains. 2. Require clocks, because they are essential for the block to operate. 3. Describe the differences in clocks between the Exynos5250/5420 and the Exynos5433 G-Scalers. This includes the fifth Exynos5433 clock "gsd" (GSCL Smart Deck) which was added to the DTS, but not to the bindings and Linux driver. Similarly to Exynos5433 DECON change [1], the clock should be used. [1] https://lore.kernel.org/all/6270db2d-667d-8d6f-9289-be92da486c25@samsung.com/Signed-off-by:
Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org> Reviewed-by:
Rob Herring <robh@kernel.org> Signed-off-by:
-
