README.md

linux-mauri870

My personal Linux kernel, tuned for gaming on modern hardware. Built with the latest LLVM toolchain targeting native compilation, configured with a 1000Hz tick rate, NTSYNC for low-latency synchronization, Transparent Hugepages, full preemption, sched_ext, and NOHZ_FULL tickless support. Patched with BORE scheduling, HDMI 2.1 48Gbps with ALLM(auto low-latency) and VRR(FreeSync) for AMD GPUs, Google's BBRv3 TCP congestion control, increased power limit for AMD GPUs and various desktop-focused tweaks.

• Linux v7.2
• Config
  • PREEMPT_FULL
  • sched_ext
  • TRANSPARENT_HUGEPAGE_MADVISE
  • NTSYNC
  • NOHZ_FULL for optional tickless
  • 1000 Hz tick rate
  • LLVM/Clang + -O3 + ThinLTO + -march=native
• Patches
  • drm: HDMI FRL (HDMI 2.1(48 Gbps), VRR(FreeSync), ALLM (Auto Low Latency Mode), 10-bit 4:4:4 support for AMD GPUs)
  • sched: BORE (Burst-Oriented Response Enhancer) scheduler
  • compiler: LLVM Polly (polyhedral loop optimizer for better cache locality and parallelism)
  • mm: Lazy RSS stat percpu counters (faster fork/exec for single-threaded tasks)
  • compiler: CONFIG_CC_OPTIMIZE_FOR_PERFORMANCE uses -O3
  • x86: Disable Split Lock mitigation
  • net: Google's BBR3 for TCP congestion control
  • time: tick/nohz: Fix wrong NOHZ idle CPU state
  • drm/ttm: VRAM pressure: keep protected workloads in VRAM when GPU memory runs low
  • sched: more aggressive idle load balancing (halved avg_idle threshold)
  • mm: raise default max_map_count to INT_MAX (prevents game crashes from VMA exhaustion)
  • net: increase default socket buffer size by 4x
  • sched: rate-limit sched_yield to once per jiffy (fixes Proton/Wine games that spam yield)
  • mm: place shared libraries below PIE binary in address space (better code/library cache locality)
  • mm: use mmput_async on process exit (avoids blocking mm teardown under memory pressure)
  • locking/rwsem: spin faster on rwsem owner (reduces contention on mixed read/write workloads)
  • time: reduce default timer_slack_ns from 50µs to 50ns (tighter hrtimer coalescing window for nanosleep/select/futex; inherited by all processes from PID 1)
  • sched/wait: do accept() in LIFO order for cache efficiency
  • boot: parallelize ATA and GPU initialization
  • net: increase TCP write buffer limit from 4MB to 16MB
  • drm/amdgpu: raise SMU maximum power limit by 15% (capped by the driver since v6.7).

Build & Install

Compiling the kernel takes around 15-20min on a 9950X3D with the default -j$(nproc).

makepkg -si -f

Recommendations

Optional packages

Install dmemcg-booster and plasma-foreground-booster-dmemcg to make use of the cgroup vram patch.

Kernel Boot Args

amdgpu

amdgpu.ppfeaturemask=0xffffffff amdgpu.ignore_min_pcap=1 to be able to set custom power limit

amdgpu.dcfeaturemask=0x402 to enable HDMI FRL

nohz

For dual-CCD asymmetric CPUs such as 9950X3D/9900X3D, configure the kernel boot args so X3D cores run tickless and offload RCU callbacks and IRQs to CCD1:

C=$(nproc); Q=$((C/4)); H=$((C/2)); echo "nohz_full=1-$((Q-1)),${H}-$((H+Q-1)) rcu_nocbs=1-$((Q-1)),${H}-$((H+Q-1)) irqaffinity=${Q}-$((H-1)),$((H+Q))-$((C-1))"
# cpu 0 can't be tickless
nohz_full=1-7,16-23 rcu_nocbs=1-7,16-23 irqaffinity=8-15,24-31

For a 9950X3D2 I advise to set CCD1 cores as nohz_full and move system tasks to CCD0. Since cpu0 is the BSP it cannot be made tickless, so you might as well isolate CCD1 for gaming instead.