[RyanTheTide]

The camera firmware writes each frame's exposure into the metadata buffer it hands back with the frame. The driver maps that buffer but never reads it, so userspace cannot tell how hard the sensor is working. The camera denoise filter (aurora-silicon/camera-denoise#1) needs that to turn on at the same gain levels as macOS.

Change

• The metadata buffer opens with a table of section offsets. On H17 firmware (T8140, T6040) the fourth section starts with the integration time in microseconds and the total sensor gain in 1/256 steps (256 is unity).
• Read it, through the table and bounded by the buffer size, when a metadata buffer comes back, and report the last frame's values as read-only, volatile V4L2_CID_GAIN and V4L2_CID_EXPOSURE_ABSOLUTE (100 us units).
• Only hardware whose match data sets meta_exposure (T8140 and T6040) gets the controls. No parameters or new knobs.

How the layout was found (Linux-side measurement only): kprobes on ipc_bt_handle with the module's BTF dumped the metadata buffers of the running fleet12 driver, read-only. On a fixed scene, auto exposure starts at 8.3 ms and gain 260, raises the integration time to the frame period (33.3 ms at 30 fps, 66.6 ms at 15 fps) and then climbs the gain field 260, 372, 990, 1758, 2489, 3253 until it settles. Settled values were 3253 to 4168 on the dark J700 scene and 600 to 888 on the brighter J616s one. Both machines have the same section table, with the section at 0x1870.

Testing

• KUnit (arm64 QEMU, apple-isp*): 31/31, including the new isp_test_meta_exposure (table lookup, out-of-bounds section, unwritten section, hardware without the report).
• defconfig + asahi.config, LLVM, W=1 KCFLAGS=-Werror on the ISP objects: 0 warnings.
• checkpatch --strict: clean.
• Not yet booted. I did not install kernels on the test machines, so the controls themselves need a fleet boot. Check with v4l2-ctl -d /dev/video1 -C gain,exposure_time_absolute while streaming. The values should match what the kprobe dump showed.

🤖 Generated with Claude Code

[RyanTheTide]

Tested in fleet12d (aurora-wip cd217c2 + #661 #633 #689 #683) on the J616s, the M1 Pro and the Neo, with no regression: kernel -p3 0 and aurora-diag check --expect 0 fail on each.

• media: apple: isp: report the frame gain and exposure #689: while streaming, gain and exposure_time_absolute read 643 to 689 and 16.7 ms on the J616s, 1229 to 1440 and 32.5 ms on the Neo.
• drm/asahi: present the GPU driver as Aurora #683: the DRM name is aurora 1.0.0 (Aurora GPU for Apple AGX), there is no asahi.ko, Vulkan is Borealis (driverID 35), and GL is native on the J616s and M1 Pro and Zink on the Neo, with Mesa 26.3.0-14.

🤖 Generated with Claude Code