[HSSkyBoy]

Follow-up to #3881 / #3883.

Problem

block_modules is merged into the kernel's module_blacklist in kernelsu_init, so a blocked module fails to load with EPERM. First stage init loads the vendor_boot modules listed in modules.load with LoadListedModules(strict) and calls LOG(FATAL) << "Failed to load kernel modules" on any failure, unless the module is in modules.blocklist:

• system/core/init/first_stage_init.cpp: LoadKernelModules() → m.LoadListedModules(!want_console), then LOG(FATAL) if it returns false
• libmodprobe/libmodprobe.cpp: LoadListedModules() only tolerates a failure when IsBlocklisted(module)

So blocking any module that first stage init loads bootloops the device, regardless of whether the system needs it. This is likely what happened on the Redmi K60 Ultra with mkp (#3883), and it applies equally to vendor_boot vr / vklp.

Fix

• kernel: install the blacklist when /system/bin/init second_stage is executed instead of in kernelsu_init. Second stage loaders (vendor_dlkm modprobe, init insmod) only log a failed module.
• ksuinit: before handing over to first stage init, append the names from block_modules= in /ksu_config to modules.blocklist in /lib/modules and its subdirectories (first stage init picks one of them by kernel release). The ramdisk is still writable at this point. modprobe then skips these modules without failing.
  • modprobe also skips every module that depends on a blocklisted one, so a name is left out (and logged to kmsg) if a module that would load otherwise depends on it. Unit tests cover the dependency handling.
• manager: only pre-fill known modules that no other module uses (the "used by" column of /proc/modules), and reword the warning.

boot-patch-v2 and built-in kernels have no ksuinit. There only second stage loads are blocked, and first stage is never at risk.

Testing

• Kernel (all DDK KMIs), ksud, ksuinit (aarch64 + riscv64), clippy and rustfmt pass in CI on the SukiSU-Ultra port (feat: block kernel modules at boot without first stage bootloop (#3881, #3883) SukiSU-Ultra/SukiSU-Ultra#986, merged); the code is identical apart from package paths.
• Manager compileDebugKotlin passes on the SukiSU-Ultra port.
• The ksuinit unit tests (cargo test -p ksuinit) were not run (no Linux toolchain available).
• Not tested on a device.

[Ylarod]

Thanks for working on this and for the dependency-aware planning, it's a careful design.

We've decided not to fix the first stage issue in KernelSU itself: the ksuinit approach doesn't cover boot-patch-v2 / built-in kernels, and it adds a fair amount of code for a niche feature. Instead, #3920 removes the option from the Manager and keeps block_modules as a developer-only feature (ksud boot-patch --block-module). Developers who use it are responsible for picking modules that are safe for their device, and for first stage modules they can edit modules.blocklist in vendor_boot directly.

Closing in favor of #3920.