Kernel version independence: umbrella tracker for all legacy platforms

[widgetii]

Overview

Tracking kernel bump feasibility across all HiSilicon platforms that use binary blob modules. The approach differs based on whether blobs go through OSAL (recompilable abstraction layer) or call kernel APIs directly.

Platform Comparison

Platform	Gen	CPU	Kernel	OSAL	Direct dangerous APIs	Bump ceiling	Effort
hi3516cv500	V3.5	Cortex-A7	4.9.37	Yes	0	Any	Just recompile OSAL
hi3516cv300	V3	ARM926EJ-S	3.18.20	Yes	0	Any	Just recompile OSAL
hi3519v101	V3A	Cortex-A17+A7	3.18.20	Yes	2 (3 blobs)	6.6+	2 shim functions
hi3516av100	V2A	Cortex-A7	4.9.37	No	4 (13 blobs)	6.6+	~100 lines of shims
hi3516cv200	V2	ARM926EJ-S	4.9.37	No	4 (~10 blobs)	5.x–6.x	~100 lines of shims
hi3516cv100	V1	ARM926EJ-S	3.0.8	No	18 (30 blobs)	3.0–3.9	Impractical beyond 3.10

Per-Platform Issues

Platform	Issue	Key finding
hi3516cv500	N/A	Zero direct calls. OSAL handles everything via kernel_compat.h
hi3516cv300	#60	Zero direct calls. Same as CV500 — just recompile OSAL
hi3519v101	#57	register_sysctl_table (2 blobs) + init_timer_key (1 blob) bypass OSAL
hi3516av100	#50	do_gettimeofday (10), init_timer_key (3), register_sysctl_table (2)
hi3516cv200	#46	Same 4 APIs as AV100, same shim approach
hi3516cv100	#48	Blocked by create_proc_entry (5 blobs), __arm_ioremap (23 blobs)

Module Migration Status — Complete

All platforms have open_ modules wired into firmware, tested on real hardware, and shipping via official OpenIPC releases.

Platform	openhisilicon PRs	firmware PRs	Hardware tested	Sensor
hi3516cv500	(pre-existing)	(pre-existing)	—	—
hi3516cv300	#59	#2014, #2016	hi3516cv300	IMX291
hi3519v101	— (no code changes)	#2010	hi3516av200	IMX385
hi3516av100	#49	#2006	hi3516dv100	AR0237
hi3516cv200	#45	#2004, #2008, #2009	hi3518ev200	JXF22
hi3516cv100	#47, #52	#2005, #2008	hi3516cv100	IMX323

Dangerous API Summary

Direct kernel calls in blobs (not via OSAL)

API	Removed in	CV100	CV200	AV100	3519v101	CV300	CV500
create_proc_entry()	3.10	5	—	—	—	—	—
__arm_ioremap()	~3.13	23	—	—	—	—	—
arm926_flush_kern_*	ARMv5	1	—	—	—	—	—
do_mmap_pgoff()	3.x→4.x	1	—	—	—	—	—
filp_open()	restricted 5.10+	1	—	—	—	—	—
strlcpy()	6.8	1	—	—	—	—	1
do_gettimeofday()	5.0	8	2	10	—	—	—
init_timer_key()	4.15	6	similar	3	1	—	—
register_sysctl_table()	6.6	1	1	2	2	—	—
rtc_time_to_tm()	5.6	1	—	—	—	—	—

OSAL advantage (CV500, CV300, hi3519v101)

These platforms' blobs call osal_* wrapper functions. Since OSAL is source-compiled with kernel_compat.h force-included, all API changes are handled at compile time. CV300 and CV500 have zero blobs bypassing OSAL; hi3519v101 has only 3.

Recommended Priority

1. CV500 + CV300 — Already work on any kernel. Just build and test.
2. hi3519v101 — 2 shim functions for 6.6+. Near-zero effort.
3. AV100 — 3 shim functions for 5.x–6.6. Best non-OSAL candidate.
4. CV200 — Same shims as AV100, but ARM926 limits long-term upstream support.
5. CV100 — Locked to kernel 3.0–3.9 unless blobs are replaced with source.

Implementation Approach

OSAL platforms (CV500, CV300): Just recompile OSAL against new kernel — kernel_compat.h handles all API changes.

OSAL platforms with exceptions (hi3519v101): Recompile OSAL + add 2 shim exports for the 3 blobs that bypass OSAL.

Direct-call platforms (AV100, CV200): Create compat_shims.ko loaded before blob modules, exporting old API names forwarding to new implementations.

CV100: Not feasible beyond 3.9 due to create_proc_entry struct layout changes in 23+ blobs.

[widgetii]

hi3516cv500 — kernel 7.0 bring-up complete (hardware verified)

Updating this for the CV500 row. The "Just recompile OSAL" summary was
correct in principle but understated the actual work: there were real
bugs in OSAL/init wrappers, the firmware DTS, and the OpenIPC kernel
clock driver that all had to be fixed before video came up on real
hardware. Documenting them here so the next platform doesn't rediscover
the same traps.

Hardware

• Board: hi3516av300_imx415
• Sensor: Sony IMX415 (3840×2160, 4K2K_10BIT_30FPS, MIPI, control via i2c0 @ 0x1A)
• Kernel: openipc/linux upstream-patches branch (Linux 7.0)
• Toolchain: glibc, arm-openipc-linux-gnueabi, GCC 13

What's working end-to-end now

• Boot to login on UART
• SMP both cores
• Full 32 MiB SPI NOR (no addressing clamp needed; 4-byte addressing on
  MX25L256 works fine on upstream hisi-sfc)
• Ethernet @ 100 Mbps full duplex via upstream hisi_femac + mdio-hisi-femac
• Persistent overlay (jffs2 on mtdblock4 + overlayfs)
• All 32 OSDRV open_*.ko modules load cleanly (no Unknown symbol,
  no interworking errors, no NULL-deref)
• IMX415 sensor auto-detected, MIPI/ISP/VI/VPSS/VENC pipeline initialised
• HiSilicon SDK fully started, RTSP server started on port 554
• HTTP/HTTPS web UI reachable
• 4K JPEG snapshot via HTTP works — verified end-to-end with a real
  3840×2160 capture from the camera

Bugs found and fixed during bring-up

Listed in the order they blocked the next thing:

1. Kernel hung silently after clk: Disabling unused clocks
   Cause: minimal DTSI used a fake fixed-clock for uart0's
   apb_pclk instead of the real CRG gate HI3516DV300_UART0_CLK.
   clk_disable_unused() then turned the gate off → console UART died.
   Fix: openipc/linux PR #34
   (use real CRG clock; same PR also re-enabled mdio/femac/i2c which
   we had wrongly disabled chasing what looked like a femac probe hang
   but was actually this same bug masquerading as one).

2. MDIO probe -EINVAL → no eth0
   Upstream mdio-hisi-femac requires hisilicon,phy-reset-delays-us
   when an external-phy reset control is declared (vendor 4.9 driver
   doesn't enforce). Fix: same PR feat: add hi3516cv500 SoC family support #34, board DTS adds the delays +
   phy-handle + phy-mode = "mii".

3. Unknown symbol HI_LOG / cmpi_register_module / ... cascade
   Cause: CONFIG_THUMB2_KERNEL=y made our open_osal.ko Thumb-2,
   but the vendor blobs (obj/hi3516cv500/hi_*.o) are ARM mode with
   R_ARM_CALL relocations. Module loader rejected open_base.ko
   with unsupported interworking call (ARM -> Thumb), breaking every
   downstream module's symbol resolution.
   Fix: openipc/firmware PR #2048
   disables THUMB2_KERNEL (matching hi3516ev300_neo).

4. Sensor unreachable on i2c — i2cdetect empty
   Cause: sys_config.ko's pinmux() never ran because the
   hisilicon,sys_config DT node didn't exist in the modernised
   minimal DTSI. Without it, i2c0 SDA/SCL pins stayed at 0x400
   (default GPIO), MCLK wasn't routed.
   Fix: openipc/linux PR #35
   adds sys_config plus all 14 vendor media nodes (mipi, vi, isp,
   vpss, vgs, vo, tde, gdc, jpegd, vedu, venc, scd, aiao, ive) so the
   blob platform_drivers actually probe.

5. Kernel Oops at hil_mmb_alloc+0x4c/0x6c after a few modules
   Cause: media_mem_init sees g_setup_allocator=\"hisi\" (the default)
   but only the CMA path is implemented in this tree. It returns -EINVAL,
   yet the OSAL module stays loaded (because the module's init is
   osal_module_platform_driver(...) — registering the driver always
   succeeds; the failure is only in the platform probe). g_the_allocator
   stays zero-initialised; downstream modules call hil_mmb_alloc →
   NULL function pointer deref → kernel panic.
   Fix: openhisilicon PR #81
   defaults g_setup_allocator to \"cma\" and falls back gracefully
   (with pr_warn) when something else is requested.

PRs that landed for the CV500 7.0 work

• openipc/linux #34 — DTS UART clock + ethernet PHY
• openipc/linux #35 — DTS media subsystem + sys_config + i2c buses
• openipc/firmware #2023 — hi3516av300_neo variant scaffold
• openipc/firmware #2048 — THUMB2_KERNEL=n + CI matrix
• openipc/openhisilicon #61 — kernel 5.0–7.0 OSAL/init compat shims
• openipc/openhisilicon #81 — CMA allocator default

Lessons for the next OSAL platform (cv300, hi3519v101)

• Don't mark mdio/femac/i2c/sys_config as status=\"disabled\" to
  "work around" probe hangs. A boot hang that looks like a probe
  hang in subsystem X is very often actually clk_disable_unused
  killing the UART after X probed successfully, with the next print
  never making it out. Verify the suspect peripheral with devmem
  before disabling its DTS node.
• Never use a fake fixed-clock for a peripheral that has a real
  CRG gate. The pl011 driver is happy to claim a no-op fixed clock,
  but clk_disable_unused() will still turn the real gate off.
• Don't enable CONFIG_THUMB2_KERNEL=y on a build that loads vendor
  ARM-mode blobs. Mark the seed config with the same # CONFIG_THUMB2_KERNEL is not set
  that hi3516ev300_neo uses.
• The OSAL module load "succeeds" even when the platform probe
  fails. If the probe-time init can't set up g_the_allocator
  function pointers, defer the EXPORT_SYMBOL exposure or have the
  exported wrappers null-check before dispatch. (osal/cv500: default to "cma" allocator and gracefully fall back #81 picks the second
  option; ideally also do the first.)

🤖 Generated with Claude Code