On Tue, Jan 16, 2024 at 05:57:14PM +0100, Hans de Goede wrote: > Hi, > > On 1/16/24 17:54, Sakari Ailus wrote: > > Hi Hans, > > > > On Tue, Jan 16, 2024 at 05:12:50PM +0100, Hans de Goede wrote: > >> Hi, > >> > >> On 1/11/24 07:55, bingbu.cao@xxxxxxxxx wrote: > >>> From: Bingbu Cao <bingbu.cao@xxxxxxxxx> > >>> > >>> This patch series adds a driver for Intel IPU6 input system. > >>> IPU6 is the sixth generation of Imaging Processing Unit, it is a PCI > >>> device which can be found in some Intel Client Platforms. User can use > >>> IPU6 to capture images from MIPI camera sensors. > >>> > >>> IPU6 has its own firmware which exposes ABIs to driver, and communicates > >>> with CSE to do firmware authentication. IPU6 has its MMU hardware, so > >>> the driver sets up a page table to allow IPU6 DMA to access the system > >>> memory. > >>> > >>> IPU6 input system driver uses MC and V4L2 sub-device APIs besides V4L2. > >>> --- > >>> v2 -> v3: > >>> - Add line-based metadata capture support > >>> - Fix header files inclusion issues > >>> - Fix the CSI2 timing calculation > >>> - Fix crash when remove module during streaming > >>> - Remove some unused code > >>> - Use cross-referencing links in documentation > >>> - Update Makefile to use ":=" for objects > >>> - Fix several bugs and coding style issues > >> > >> So I've given this version a try on a Lenovo X1 yoga gen 8 with ov2740 > >> sensor using the ongoing libcamera SoftISP work + this small patch > >> to enable the SoftISP on IPU6 : > >> > >> https://github.com/jwrdegoede/libcamera/commit/3172f3703cf7076390fbf86c3b43e388c2422b31 > >> > >> things work fine when using patch 1-15 + 17 on top of 6.7, note > >> I'm skipping patch 16 ("media: ipu6/isys: support line-based > >> metadata capture support")" here! > >> > >> However when I instead apply the whole series on top of: > >> https://git.linuxtv.org/sailus/media_tree.git/log/?h=metadata > >> > >> Then things stop working, with the following errors > >> (I added extra error logging to figure out in which syscall > >> resetRoutingTable() fails and made libcamera ignore the routing > >> errors): > >> > >> [2:02:04.466310686] [8943] ERROR SimplePipeline simple.cpp:1443 GetRouting() failed -25 > >> [2:02:04.466315975] [8943] ERROR SimplePipeline simple.cpp:1574 Failed to reset routes for /dev/v4l-subdev1: Inappropriate ioctl for device > >> [2:02:04.466366331] [8943] ERROR SimplePipeline simple.cpp:1443 GetRouting() failed -25 > >> [2:02:04.466370025] [8943] ERROR SimplePipeline simple.cpp:1574 Failed to reset routes for /dev/v4l-subdev4: Inappropriate ioctl for device > >> [2:03:32.334708887] [8929] INFO Camera camera.cpp:1183 configuring streams: (0) 1924x1088-BGR888 > >> [2:03:32.335129023] [8943] ERROR SimplePipeline simple.cpp:1205 Unable to configure capture in 1932x1092-BA10 (got 0x0-@...) > >> Failed to configure camera. > >> > >> I was sorta assuming that the new metadata-stream support would > >> be backwards compatible for userspace without support for this, > >> so I think this is a bug ? > > > > That's the intention when it comes to the kernel APIs indeed. > > > > I wonder how the simple pipeline handler works with this, does it try to > > configure both streams all the way from the internal source pad to the > > video node? This will certainly fail without metadata support in the ISYS > > driver. Just guessing the cause though. An extra stream from the source pad > > won't fail pipeline validation. > > > > I should be able to set up a system to test this, too. > > I did include Bingbu's patch to add metadata to the isys driver for my > testing (and I also tried reverting just that patch). I think the issue > might be that libcamera already has some streams awareness based on > an older streams patch-set. Correct. It uses the G_ROUTING and S_ROUTING ioctls, and those have chenged in the latest kernel patch series. The simple pipeline handler doesn't configure routing as such, it gets the default routes with G_ROUTING(TRY) if the device advertises routing support, and calls S_ROUTING(ACTIVE) to reset the routes to the default. It then uses the routes to walk the pipeline, but doesn't change them. > I'll retest when Laurent's branch for this is ready. -- Regards, Laurent Pinchart