- 22 Nov, 2016 40 commits
-
-
Mauro Carvalho Chehab authored
Despite being exported, there's no prototype for it at the headers, as warned by sparse: Fixes this sparse warning: drivers/media/platform/ti-vpe/vpdma.c:1000:6: warning: no previous prototype for 'vpdma_enable_list_notify_irq' [-Wmissing-prototypes] void vpdma_enable_list_notify_irq(struct vpdma_data *vpdma, int irq_num, ^~~~~~~~~~~~~~~~~~~~~~~~~~~~ Worse than that, it is not even used, as making it static it would produce: drivers/media/platform/ti-vpe/vpdma.c:1000:13: warning: 'vpdma_enable_list_notify_irq' defined but not used [-Wunused-function] static void vpdma_enable_list_notify_irq(struct vpdma_data *vpdma, int irq_num, ^~~~~~~~~~~~~~~~~~~~~~~~~~~~ So, let's just get rid of the dead code. If needed in the future, someone could re-add it. Cc: Benoit Parrot <bparrot@ti.com> Cc: Hans Verkuil <hans.verkuil@cisco.com> Signed-off-by:Mauro Carvalho Chehab <mchehab@s-opensource.com>
-
Hans Verkuil authored
davinci/vpfe_capture.c: In function 'vpfe_probe': davinci/vpfe_capture.c:1992:9: warning: 'ret' may be used uninitialized in this function [-Wmaybe-uninitialized] return ret; ^~~ This is indeed correct, so if the kmalloc fails set ret to -ENOMEM. Signed-off-by:Hans Verkuil <hans.verkuil@cisco.com> Signed-off-by:
-
Benoit Parrot authored
The VPE was restricting the number of plane per buffer based on the fact that if a particular format had color separation it was meant to need 2 planes. However NV12/NV16 are color separate format which are meant to be presented in a single contiguous buffer/plane. It could also be presented in a multi-plane as well if need be. So we must support both modes for more flexibility. The number of plane requested by user space was previously ignored and was therefore always overwritten. The driver now use the requested num plane as hint to calculate needed offset when required. Signed-off-by:
Benoit Parrot <bparrot@ti.com> Signed-off-by:
-
Benoit Parrot authored
Since there might be more then one instance it is better to show the base address when dumping registers to help with debugging. Signed-off-by:
-
Benoit Parrot authored
In preparation to add colorspace conversion support to VIP, we need to turn csc.c into its own kernel module. Signed-off-by:
-
Benoit Parrot authored
Add RAW8 and RAW16 data type to VPDMA. To handle RAW format we are re-using the YUV CBY422 vpdma data type so that we use the vpdma to re-order the incoming bytes, as the VIP parser assumes that the first byte presented on the bus is the MSB of a 2 bytes value. RAW8 handles from 1 to 8 bits. RAW16 handles from 9 to 16 bits. Signed-off-by:
-
Benoit Parrot authored
When scaler is to be used we need to make sure that the input and output frame size do not exceed the maximum frame sizes that the scaler h/w can handle otherwise streaming stall as the scaler cannot proceed. The scaler buffer is limited to 2047 pixels (i.e. 11 bits) when attempting anything larger (2048 for example) the scaler stalls. Realistically in an mem2mem device we can only check for this type of issue when start_streaming is called. We can't do it during the try_fmt/s_fmt because we do not have all of the info needed at that point. So instead when start_streaming is called we need to check that the input and output frames size do not exceed the scaler's capability. The only time larger frame size are allowed is when the input frame szie is the same as the output frame size. Now in the case where we need to fail, start_streaming must return all previously queued buffer back otherwise the vb2 framework will issue kernel WARN messages. In this case we also give an error message. Signed-off-by:
-
Benoit Parrot authored
Since there might be more then one instance it is better to show the base address when dumping registers to help with debugging. Signed-off-by:
-
Benoit Parrot authored
In preparation to add scaler support into VIP we need to turn sc.c into its own kernel module. Add support for multiple SC memory block as VIP contains 2 scaler instances. This is done by passing the resource name to sc_create() and modify the vpe invocation accordingly. Signed-off-by:
-
Benoit Parrot authored
Allow VPE to be able to export DMA buffer. Signed-off-by:
-
Nikhil Devshatwar authored
For deinterlacing operation, VPE hardware uses motion vectors. MV calculated in the previous iteration are used for next interation. Therefore driver allocates two motion vectors in ping-pong fashion. For every transaction, one MV is DMAed in and one is DMAed out. All the outbound DMAs (DMA to memory) use output parameters, but as the motion vectors is generated purely out of input fields, it should use the input parameters for DMA. Fix the add_out_dtd to use source q_data for creating descriptor. If the output size is greater than input stride, without this change, MV DMA may overwrite the buffer causing memory corruption. This CRITICAL fix ensures that the motion vector DMA descriptor is created based on the attributes with which the buffer was allocated. Signed-off-by:
Nikhil Devshatwar <nikhil.nd@ti.com> Signed-off-by:
Ravikumar Kattekola <rk@ti.com> Signed-off-by:
Ravi Babu <ravibabu@ti.com> Signed-off-by:
-
Nikhil Devshatwar authored
VPDMA buffer will be used by CPU as well as by the VPDMA. CPU will write/update the VPDMA descriptors containing data about the video buffer DMA addresses. VPDMA will write the "write descriptor" containing the data about the DMA operation. When mapping/unmapping the buffer, driver has to take care of WriteBack and invalidation of the cache so that all the coherency is maintained from both directions. Use DMA_BIDIRECTIONAL to maintain coherency between CPU and VPDMA. Signed-off-by:
-
Nikhil Devshatwar authored
vpdma_create API is supposed to allocated the struct vpdma_data and return it to the driver. Also, it would call the callback function when the VPDMA firmware is loaded. Typically, VPE driver have following function call: dev->vpdma = vpdma_create(pdev, firmware_load_callback); And the callback implementation would continue the probe further. Also, the dev->vpdma is accessed from the callback implementation. This may lead to race condition between assignment of dev->vpdma and the callback function being triggered. This would lead to kernel crash because of NULL pointer access. Fix this by passing a driver wrapped &vpdma_data instead of allocating inside vpdma_create. Change the vpdma_create prototype accordingly and fix return paths. Also, update the VPE driver to use the updated API and initialize the dev->vpdma before hand so that the race condition is avoided. Signed-off-by: -
Benoit Parrot authored
When stop_streaming is called we need to cleanup the queued vb2 buffers properly. This was not previously being done which caused kernel warning when the application using the resources was killed. Kernel warnings were also generated on successful completion of a de-interlacing case as well as upon aborting a conversion. Make sure every vb2 buffers is properly handled in all cases. Signed-off-by:
-
Nikhil Devshatwar authored
Current scaler library implementation of sc_set_hs_coeffs and sc_set_vs_coeffs tries to return immediately if the calculated coefficient index is already being used. As the same scaler block is going to be used for all the VPE contexts, even if the calculated index is same, the parameters have to be reconfigured for each of the context. Because of this, when multiple contexts use the same coefficients, all other contexts would have zero scaling coefficients. Fix this and also remove the unnecessary hs_index and vs_index fields. Signed-off-by:
-
Benoit Parrot authored
The VPDMA RGB data type definition have been updated to match with Errata i839. But some of the ARGB definition appeared to be wrong in the document also. As they would yield RGBA instead. They have been corrected based on experimentation. Signed-off-by:
-
Benoit Parrot authored
The YUV data type definition below are taken from both the TRM and i839 Errata information. Use the correct data type considering byte reordering of components. Added the 2 missing YUV422 variant. Also since the single use of "C" in the 422 case to mean "Cr" (i.e. V component). It was decided to explicitly label them CR to remove any confusion. Bear in mind that the type label refer to the memory packed order (LSB - MSB). Signed-off-by:
-
Benoit Parrot authored
ti_vpe module currently does not get loaded automatically. Added MODULE_DEVICE_TABLE hint to the driver to assist. Signed-off-by:
Javier Martinez Canillas <javier@osg.samsung.com> Signed-off-by:
-
Nikhil Devshatwar authored
VPDMA block used in ti-vip and ti-vpe modules have support for up to 8 hardware descriptor lists. A descriptor list can be submitted to any of the 8 lists (as long as it's not busy). When multiple clients want to transfer data in parallel, its easier to allocate one list per client and let it use it. This way, the list numbers need not be hard-coded into the driver. Add support for allocating hwlist and maintain them with a priv data. Signed-off-by:
-
Nikhil Devshatwar authored
VPE hardware can generate output in RGB565 or in RGB5551 format. Add these formats in the supported format list for CAPTURE stream. Also, for RGB5551 format, the alpha component is not processed, so the alpha value is taken from the default color. Set the default color to make alpha component full when the dst format is of RGB color space. Signed-off-by:
-
Nikhil Devshatwar authored
vpe_irq checks for the possible interrupt sources and prints the errors for the DEI_ERROR and DS_UV interrupts. But it also post the next descriptor list irrespective of whichever interrupt has occurred. Because of this, driver may release the buffers even before DMA is complete and also schedule next descriptor list. Fix this by _actually_ handling the IRQ only when ListComplete IRQ occurs. Signed-off-by:
-
Nikhil Devshatwar authored
For deinterlacing operation, each operation needs 2 fields in the history. This is achieved by holding three buffers in ctx->src_vbs[0,1,2] (f,f-1,f-2) This is achieved by using the ctx->sequence which gets reset via the s_fmt ioctl. These buffers are dequeued in stream OFF by calling free_vbs() But the corresponding references aren't removed anywhere. When application tries to stream ON and OFF continuously, s_fmt ioctl won't be called and it won't setup the srcdst parameters. Setting source/destination parameters in stream ON ioctl would make sure that the context is re-initialized before it is being used by the driver. Signed-off-by:
-
Nikhil Devshatwar authored
Current driver configures the line mode of the DEI clients from the open function directly. Even if the newly created context is not yet scheduled, it updates some of the VPDMA registers. This causes a problem in multi instance use case where just opening the m2m device second time causes the running job to stall. This happens especially if the source buffers used are NV12. While all other configuration is being written to context specific shadow registers, only line mode configuration is happening directly. As there is no shadow register for line mode configuration, it's better to separate the config_mode setting and line_mode setting. Call the new "set_line_modes" functions only when actually loading the mmrs. This makes sure that no non-running job will write to the registers directly. Signed-off-by:
-
Nikhil Devshatwar authored
VPDMA IRQs are registered for multiple lists When clearing an IRQ for a list interrupt, all the IRQs for the individual lists are to be cleared separately. Signed-off-by:
-
Nikhil Devshatwar authored
Writing to the "VPDMA list attribute" register is considered as a list post. This informs the VPDMA firmware to load the list from the address which should be taken from the "VPDMA list address" register. As these two register writes are dependent, it is important that the two writes happen in atomic manner. This ensures multiple slices (which share same VPDMA) can post lists asynchronously and all of them point to the correct addresses. Slightly modified to implementation for the original patch to use spin_lock instead of mutex as the list post is also called from interrupt context. Signed-off-by:
-
Nikhil Devshatwar authored
Whenever VPDMA processes a data descriptor of a list, it processes it and sets up the channel for the DMA transaction. List manager holds the descriptor in the list until the DMA is complete. If sync_on_channel descriptor, or another descriptor for the same channel is present in the FIFO, list manager keeps them until the current channel is free. When the capture stream is closed suddenly while there are pending descriptors in the FIFO (streamON failed, application killed), it would keep the VPDMA in a busy state. Any further list post would fail with EBUSY. To avoid this, drivers need to stop the current processing list and cleanup all the resources VPDMA has taken and also clear the internal FSM of list manager. The state machine is cleared by issuing channel specific abort descriptor. Therefore, the vpdma_list_cleanup accepts an array of channels for which abort_channel descriptors should be posted. It is driver's responsibility to post for all the channels or the channels which were used in the last context. Signed-off-by:
-
Nikhil Devshatwar authored
Add a helper function to be able to set the maximum VPDMA transfer size to limit potential buffer overrun. Added enums for max_width and max_height fields of the outbound data descriptor. Changed vpdma_add_out_dtd to accept two more arguments for max width and height. Make use of different max width & height sets for different of capture module (i.e. slices). Signed-off-by:
-
Harinarayan Bhatta authored
Free vpdma buffers in vpe_release. Otherwise it was generating random backtrace. Signed-off-by:
Harinarayan Bhatta <harinarayan@ti.com> Signed-off-by:
Somnath Mukherjee <somnath@ti.com> Signed-off-by:
-
Nikhil Devshatwar authored
get_q_data can be called with different values for type e.g. vpe_try_crop calls it with the buffer type which gets passed from user space Framework doesn't check wheather its correct type or not If user space passes wrong type, kernel should not crash. Return NULL when the passed type is invalid. Signed-off-by:
-
Harinarayan Bhatta authored
Increasing max buffer height and width to allow for padded buffers. Signed-off-by:
-
Nikhil Devshatwar authored
The video source can generate the data in the SEQ_TB buffer format. In the case of TI SoC, the IVA_HD can generate the interlaced content in the SEQ_TB buffer format. This is the format where the top and bottom field data can be contained in a single buffer. For example, for NV12, interlaced format, the data in Y buffer will be arranged as Y-top followed by Y-bottom. And likewise for UV plane. Also, queuing one buffer of SEQ_TB is equivalent to queuing two different buffers for top and bottom fields. Driver needs to take care of this when handling source buffer lists. Signed-off-by:
-
Nikhil Devshatwar authored
Current VPE driver does not start the job until all the buffers for a transaction are queued. When running in multiple context, this might increase the processing latency. Alternate solution would be to try to continue the same context as long as buffers for the transaction are ready; else switch the context. This may increase number of context switches but it reduces latency significantly. In this approach, the job_ready always succeeds as long as there are buffers on the CAPTURE and OUTPUT stream. Processing may start immediately as the first 2 iterations don't need extra source buffers. Shift all the source buffers after each iteration and remove the oldest buffer. Also, with this removes the constraint of pre buffering 3 buffers before call to STREAMON in case of de-interlacing. Signed-off-by:
-
Archit Taneja authored
The motion detection block requires 3 fields to create the motion vector data. This means that using the default method the first progressive frame is only generated after 3rd field is consumed. Hence by default for N input field we would generate N - 2 progressive frames. In order to generate N progressive frames from N fields we use the line averaging mode of the de-interlacer for the first 2 fields and then revert back to the preferred Edge Directed Interpolation method (using the motion vector). Thus creating 2 line averaged frames + N - 2 motion based frames for a total of N frames. Signed-off-by:
Archit Taneja <archit@ti.com> Signed-off-by:
-
Benoit Parrot authored
On DRA7 since l3_noc event are being reported it was found that when the write descriptor was being written it was consistently causing bus error events. The write address was improperly programmed. Signed-off-by:
-
Benoit Parrot authored
Add a helper to set the background color during vpdma transfer. This is needed when VPDMA is generating 32 bits RGB format to have the Alpha channel set to an appropriate value. Signed-off-by:
-
Benoit Parrot authored
The VPDMA (Video Port DMA) as found in devices such as DRA7xx is used for both the Video Processing Engine (VPE) and the Video Input Port (VIP). Some devices may have multiple VIP instances each with its own VPDMA engine. Within VIP two slices can use a single VPDMA engine simultaneously. So support for multi instances and multiple clients has been added to VPDMA. Needed modification to the existing helper functions were then reflected to VPE. Multi-clients registers offset have also been added in preparation. Signed-off-by:
-
Benoit Parrot authored
The VPDMA (Video Port DMA) as found in devices such as DRA7xx is used for both the Video Processing Engine (VPE) and the Video Input Port (VIP). In preparation for this we need to turn vpdma into its own kernel module. Signed-off-by:
-
Mauro Carvalho Chehab authored
While checking why we need i386 checking, I noticed that the serial code referred at the driver was moved to another place. Update it to make clear from where such code came from. Signed-off-by: -
Sean Young authored
Signed-off-by:Sean Young <sean@mess.org>
-
Mauro Carvalho Chehab authored
There are a few checkpatch complains here. As we're about to promote this driver out of staging, address them. Signed-off-by:
-
