[RFC] mm: page-fault: overcommit: PPM (Page-Plugger Under Memory)

[MaeeFilho FL]

Hi Ard, EFI Subsystem, Kernel Subsystem, Doc Subsystem, Jgun NVIDIA, Kees,

Actually, the Kernel Linux memory management subsystem introduces
significant latency spikes during single-use, high-velocity volatile
data streams due to speculative overcommit tracking and reactive page
fault allocation handlers.

To eliminate the cycle overhead of the page fault handler (vector 14
exception routines), this RFC proposes a localized, pre-mapping bypass
called Page-Plugger Under Memory (PPM).

Architectural Overview:
The PPM subsystem introduces a non-destructive hardware plug mechanism
within the core page table layer, targeting the upper execution bits
of gfp_t allocations (allocated within the Bit 54 and Bit 55 matrix
limits to avoid legacy subsystem collision).

1. Allocation Handshake (ppm_kalloc):
Using the GFP_PLUG_MEMORY mask, the allocator completely bypasses
passive overcommit verification. It instantly isolates physical frames
on the RAM, maps the page table entries síncronamente, and forces the
Present Bit and Writable Bit to 1 before returning the address block
register to the CPU.

2. Desallocation Cleanup (ppm_kfree):
Triggered via the GFP_FREE_T parameter, the release phase executes an
immediate length truncation to 0. It forces a synchronous TLB
invalidation (via invlpg execution) and flushes the data out of the L3
cache lines using clflushopt instructions before moving the hot clean
pages into a dedicated priority recycling pool.

Performance Target:
- Elimination of page fault latency penalties on performance-critical pathways.
- Deterministic allocation timings with localized hardware cache purges.

This header infrastructure is prototyped inside
<linux/ppm/gfp_types.h> using global tipified configuration masks
(GFP_MASK) to enforce 1-clock execution checks inside the allocation
hot paths.

Feedback and architectural review regarding this memory plug
optimization layer are highly appreciated.

Signed-off-by: João Guilherme da Silva Freitas Lima <joanandalima@xxxxxxxxx>