tools: PAGE_SIZE mmap fallback for CONFIG_IO_STRICT_DEVMEM kernels

[widgetii]

Summary

• mem_reg() mmaps a 64 KiB window per call. On CONFIG_IO_STRICT_DEVMEM=y kernels, the kernel rejects any /dev/mem mapping whose range overlaps a driver-claimed page — even for root.
• On Hi3516CV608 the first probe (SCSYSID @ 0x11020EE0) opens a window [0x11020000..0x1102FFFF] that covers PWM @ 0x11029000, so the mmap fails with EPERM and ipctool aborts before printing the chip name.
• On EPERM, retry with a single PAGE_SIZE window aligned to the requested address. The cache-invalidation logic is reworked to compare against the actual loaded window range, so reads of adjacent registers in the same page still hit the cache after a fallback.

Validation

Tested on OpenIPC hi3516cv6xx (kernel 5.10.221 armv7, dual A7) at 10.216.128.115 with stock OpenIPC defconfig (both CONFIG_STRICT_DEVMEM=y and CONFIG_IO_STRICT_DEVMEM=y).

Before:

# ipctool
read_mem_reg mmap error: Operation not permitted (1)
# ipcinfo -s
read_mem_reg mmap error: Operation not permitted (1)

After:

# ipctool
chip:
  vendor: HiSilicon
  model: 3516CV608
board:
  vendor: OpenIPC
  version: 2.6.06.04
...
# ipcinfo -c
hi3516cv608

Three EPERMs remain on register blocks whose pages really are kernel-claimed (0x11010000 CRG/clocks, 0x11060000 I2C, 0x1109xxxx GPIO). Those need the kernel-side companion fix (drop CONFIG_IO_STRICT_DEVMEM from hi3516cv6xx.generic.config — separate firmware PR).

Test plan

• Cross-built for armv7 musl with the in-tree arm-openipc-linux-musleabi.cmake toolchain file
• Deployed and run on hi3516cv6xx — chip detection works, no regression on non-strict kernels (page fallback only triggers on EPERM)
• Smoke-test on a non-strict kernel (e.g. any V4 board) — the success path is unchanged

🤖 Generated with Claude Code