Skip to content

09-thinlto.md

Latest commit

 

History

History
67 lines (54 loc) · 2.95 KB

09-thinlto.md

File metadata and controls

67 lines (54 loc) · 2.95 KB

Phase 5 — The modern kernel with ThinLTO

A variant of the modern kernel built with Clang ThinLTO on top of everything from Phase 2 (kCFI, hardening, modern features).

What ThinLTO buys

Link-Time Optimization lets the compiler optimize across translation-unit boundaries — inlining, devirtualization and dead-code elimination that a per-file compile can't see, because at -c time each .o is compiled blind to the others. ThinLTO does this with a lightweight per-module summary instead of merging the whole program into one giant module (Full LTO), so it links far faster and uses far less memory while keeping most of the benefit. On the kernel it also tightens kCFI (CFI + LTO let the compiler prove more call targets).

How it was configured

Start from the exact modern .config, then:

cp build-modern/.config build-modern-lto/.config
scripts/config --file build-modern-lto/.config \
    -e LTO_CLANG_THIN \      # Clang ThinLTO
    -d DEBUG_INFO_BTF \      # drop BTF...
    -e DEBUG_INFO_NONE \     # ...and DWARF, to bound the LTO link's memory
    --set-str LOCALVERSION "-modern-lto"
make LLVM=1 O=build-modern-lto olddefconfig
make LLVM=1 O=build-modern-lto -j$(nproc) bzImage

Why drop BTF/debug-info for the LTO build? ThinLTO defers code generation to link time and holds the bitcode for the whole vmlinux in memory at once; piling full DWARF5 + a BTF pahole pass on top of that is what pushes a 15 GiB box toward swap. Dropping debug info kept the peak RSS at ~6 GiB (vs. 7.7 GiB for the non-LTO modern build with BTF). The main modern kernel keeps BTF; this LTO variant trades it for a comfortable link. ThinLTO requires lld (we use LLVM=1), which x86-64 supports (ARCH_SUPPORTS_LTO_CLANG_THIN).

Results

Metric Modern (Phase 2) Modern + ThinLTO (Phase 5)
CONFIG_LTO_CLANG_THIN off y
CONFIG_CFI (kCFI) y y
Debug info / BTF DWARF5 + BTF none (dropped for the link)
Build time 18 m 55 s 21 m 25 s
Peak build RAM 7.7 GiB 6.0 GiB
bzImage 20 MB 18 MB
Boots in QEMU ✅ ✅ rc=0

Boot test (q35, KVM) — artifacts/modern-lto/boot.log

Linux version 7.1.0-modern-lto … clang 18.1.3 … LLD 18.1.3 … #1 SMP PREEMPT_DYNAMIC
SMP alternatives: CFI: Using rehashed retpoline kCFI    <-- kCFI still live under LTO
landlock: Up and running.
Run /init as init process
reboot: Power down                                       <-- clean poweroff

The ThinLTO + kCFI kernel boots cleanly and powers off. The bzImage is slightly smaller than the non-LTO modern build (cross-module dead-code elimination, and no debug info), and kCFI remains active — LTO and CFI compose.

Full LTO (CONFIG_LTO_CLANG_FULL) would optimize harder but merges the whole program into one module at link — much slower and far more memory-hungry. On a 15 GiB box, ThinLTO is the right call; Full LTO would risk swapping.