Files
totk-pc-recomp/runtime/host.h
T
Valiantsin DzerakhandClaude Opus 5.5 1fe552ae0a Mirror buffer-bound pools in device-local memory on discrete GPUs
Imported pools put vertices, indices, constants and storage buffers in system memory, so discrete GPUs read them over
PCIe on every draw. A pool now gets a device-local mirror on its first use as a vertex, index, uniform, storage or
indirect buffer; bindings, buffer copies, fills and uniform updates use it at the same offsets. The first access of a
submission syncs the pages written since (write-watch snapshots shared with the texture checks, plus
nvnMemoryPoolFlushMappedRange ranges) in a prologue command buffer before the part; what the GPU writes into a mirror
goes back to guest memory at the end of the submission. TOTK_MEMORY=import|mirror|verify overrides the device class;
verify compares sampled mirror pages with guest memory at the fence. The vkexec probe gains a mirror part.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-25 00:45:56 +03:00

233 lines
18 KiB
C

// Shared interface of the M3 host runtime. Owned by T1; changes by agreement.
//
// Guest addresses are host addresses. Every guest-visible region lies below 2^39 (IPC buffer descriptors carry 39
// address bits). Fixed parts of the guest address map:
// modules MOD_<X>_BASE from layout.h, from 0x70_0000_0000
// kernel region RT_KERNEL_BASE, 256 MiB: TLR, TLS blocks, host objects the guest reads (mem_kernel_alloc)
// host functions RT_HOSTFN_BASE, 1 MiB, never accessible: synthetic addresses of host functions
// The stack, heap and alias regions are placed at start-up (mem_layout).
#pragma once
#include <stdarg.h>
#include <stddef.h>
#include <stdint.h>
#include <string.h>
#include "cpu.h"
#include "dispatch.h"
// ---- logging and results
void rt_log(const char *fmt, ...) __attribute__((format(printf, 1, 2))); // "[host] <msg>\n" to stderr, one write
enum { RT_TRACE_SVC = 1, RT_TRACE_HOSTCALL = 2, RT_TRACE_IPC = 4, RT_TRACE_FS = 8, RT_TRACE_GFX = 16, RT_TRACE_PERF = 32,
RT_TRACE_SCHED = 64, RT_TRACE_AUDIO = 128, RT_TRACE_PRESENT = 256 };
extern unsigned rt_trace; // from TOTK_TRACE=svc,host,ipc,fs,gfx,perf,sched,audio,present (or all)
// prof_win.c: TOTK_PROFILE=<path> sampling profiler; TOTK_TRACE=perf frame timing (waits of the render thread)
// PERF_TEX: texture uploads on the render thread (deswizzle, decode, staging), PERF_ASTC included
enum { PERF_PACE, PERF_FENCE, PERF_DRAIN, PERF_PIPE, PERF_XLATE, PERF_TEX, PERF_ASTC, PERF_N };
void prof_start(void);
void prof_write(void); // idempotent; boot.c calls it before _exit
uint64_t perf_begin(void); // start time of a wait; pair with perf_end
void perf_end(int kind, uint64_t t0); // counts render-thread time only
void perf_replay(uint64_t t0); // replay (NVN -> Vulkan recording) time on any thread, waits inside it included
// Events per 60-frame window: graphics pipelines finished by the background workers, draws skipped while pending,
// first texture uploads and re-uploads after CPU writes, draws skipped while a shader translates in the background,
// draws that waited for their pending pipeline and got it, draws into a target the previous frame did not render that
// waited for their pending shader or pipeline instead of being skipped (vk_exec_draw.c prepare_draw). The rest feed only
// the flight recorder (flight.c): graphics pipelines created on any thread, texture images created, images created at an
// address where a live image of another descriptor already starts, write-backs done and skipped because the CPU
// rewrote the storage after the GPU render (vk_texture.c vk_texture_write_back).
enum { PERF_CNT_PIPE_ASYNC, PERF_CNT_PIPE_SKIP, PERF_CNT_TEX_UP, PERF_CNT_TEX_REUP, PERF_CNT_SHADER_SKIP,
PERF_CNT_PIPE_WAIT, PERF_CNT_ONESHOT_WAIT, PERF_CNT_PIPE_BUILT, PERF_CNT_TEX_NEW, PERF_CNT_TEX_ALIAS,
PERF_CNT_WB, PERF_CNT_WB_SKIP, PERF_CNT_N };
void perf_count(int kind);
void perf_present(uint64_t frame);
void perf_summary_flush(void); // prints the frame-time summary of the running period (exit); the 60 s lines come from perf_present
void perf_gpu_wait(uint64_t t0); // a wait for a GPU fence ended (any thread; the render thread's also counts as PERF_FENCE)
// GPU side of the 60-frame perf line, filled by the graphics back end when it set perf_gpu_stats. gpu_ns: GPU busy time
// since the previous call (timestamp queries, overlaps counted once), have_gpu 0 when the device has no timestamps.
typedef struct {
int have_gpu, have_vram, vram_exact;
uint64_t gpu_ns, gpu_parts, gpu_missed; // busy time, submission parts timed, parts left untimed
uint64_t vram_used, vram_budget;
} PerfGpu;
extern void (*perf_gpu_stats)(PerfGpu *out);
#define RT_RESULT(module, desc) ((uint32_t)(module) | ((uint32_t)(desc) << 9))
void rt_sha256(const void *p, size_t n, uint8_t out[32]); // BCrypt; fatal if the provider fails
// ---- guest address map (fixed part)
#define RT_ADDRESS_LIMIT (1ull << 39)
#define RT_KERNEL_BASE 0x7010000000ull
#define RT_KERNEL_SIZE 0x10000000ull
#define RT_HOSTFN_BASE 0x7020000000ull
#define RT_HOSTFN_SIZE 0x100000ull
// ---- AAPCS64 argument access for host functions (section 2.2)
static inline uint64_t arg_x(const Cpu *c, int i) { return i < 8 ? c->x[i] : LD64(c->sp + 8 * (uint64_t)(i - 8)); }
static inline void *arg_p(const Cpu *c, int i) { return (void *)(uintptr_t)arg_x(c, i); }
static inline float arg_s(const Cpu *c, int i) { uint32_t b = (uint32_t)c->v[i]; float f; memcpy(&f, &b, 4); return f; }
static inline double arg_d(const Cpu *c, int i) { uint64_t b = (uint64_t)c->v[i]; double d; memcpy(&d, &b, 8); return d; }
static inline void ret_x(Cpu *c, uint64_t v) { c->x[0] = v; }
static inline void ret_s(Cpu *c, float f) { uint32_t b; memcpy(&b, &f, 4); c->v[0] = b; }
static inline void ret_d(Cpu *c, double d) { uint64_t b; memcpy(&b, &d, 8); c->v[0] = b; }
// ---- host functions (hostfn.c, T1)
// Every host function is a GuestFn at a synthetic address RT_HOSTFN_BASE + 16*i. It runs on the calling guest thread,
// may clobber x0-x18, v0-v7, v16-v31, must preserve x19-x29, sp, v8-v15, tpidr, and returns to the guest BL site with
// a C return. On entry c->pc holds the function's own synthetic address.
typedef struct { const char *name; GuestFn fn; } RtHostEntry;
#define RT_CAT2(a, b) a##b
#define RT_CAT(a, b) RT_CAT2(a, b)
// File-scope declarations; fn must be declared before. The loader binds them by symbol name.
#define HOST_OVERRIDE(sym, fn) /* replaces sym in every GOT/GLOB_DAT/ABS64 of every module */ \
__attribute__((used, section(".ovr$m"))) static const RtHostEntry RT_CAT(rt_ovr_, __COUNTER__) = {sym, fn}
#define HOST_OBSERVER(sym, fn) /* JUMP_SLOT only; fn calls rt_original(c) */ \
__attribute__((used, section(".obs$m"))) static const RtHostEntry RT_CAT(rt_obs_, __COUNTER__) = {sym, fn}
uint64_t hostfn_register(const char *name, GuestFn fn); // synthetic address; before hostfn_seal only. The same
// (name, fn) pair returns the same address
GuestFn hostfn_override(const char *name); // NULL if none
GuestFn hostfn_observer(const char *name); // NULL if none
void hostfn_add_override(const char *name, GuestFn fn); // an override outside the linker table (probes); before
// ldr_load
const char *hostfn_name(uint64_t addr); // name registered at a synthetic address, NULL if none
void hostfn_seal(void); // build the dispatch table (rt_dispatch_init over every
// linked module); idempotent; registration ends here
uint64_t hostfn_calls(void); // host-function calls so far, all threads
void hostfn_report(void); // call count per host function to stderr, busiest first
int hostfn_table_size(int observers); // linked HOST_OVERRIDE (0) / HOST_OBSERVER (1) entries
const char *hostfn_table_name(int observers, int i); // symbol of linked entry i, NULL past the end
uint64_t rt_original(Cpu *c); // inside an observer: call the guest definition with the
// current x0..x7, return x0. Uses c->pc, so call it before
// anything that changes c->pc (rt_call restores it)
uint64_t rt_call(Cpu *c, uint64_t fn, int n, ...); // host -> guest, section 2.3. n <= 8 arguments, each read
// as uint64_t: cast every argument. Seals the dispatch
// table on first use. Sets x30 = RT_RET_HOST, restores
// x30 and pc, returns x0. Does not touch sp
#define RT_RET_HOST 0x70200ff000ull // x30 sentinel inside the host-function window; a jump to
// it is a named fatal
// ---- guest memory (memory.c, T1)
typedef struct { uint64_t base, size; } RtRange;
typedef struct { RtRange aslr, stack, heap, alias, kernel, hostfn; uint64_t pool; } RtGuestLayout;
typedef struct { uint64_t base, size; uint32_t state, attr, perm, ipc, device, pad; } RtMemoryInfo; // Horizon layout
void mem_init(void); // reserve every region; on a collision re-run the process
// (up to 20 times, TOTK_REEXEC), then fatal
const RtGuestLayout *mem_layout(void);
uint32_t mem_set_heap_size(uint64_t size, uint64_t *addr); // 0x01
uint32_t mem_set_permission(uint64_t a, uint64_t n, uint32_t perm); // 0x02
uint32_t mem_set_attribute(uint64_t a, uint64_t n, uint32_t mask, uint32_t at); // 0x03
uint32_t mem_map(uint64_t dst, uint64_t src, uint64_t n); // 0x04
uint32_t mem_unmap(uint64_t dst, uint64_t src, uint64_t n); // 0x05
uint32_t mem_query(uint64_t a, RtMemoryInfo *out); // 0x06
uint32_t mem_map_physical(uint64_t a, uint64_t n); // 0x2C
uint32_t mem_unmap_physical(uint64_t a, uint64_t n); // 0x2D
int mem_info(uint32_t id, uint64_t *out); // GetInfo memory ids 2-7, 12-17, 21, 22, 28; 0 if not a memory id
void *mem_kernel_alloc(uint64_t size, uint64_t align); // zeroed, guest-visible, in the kernel region; align <= 64 KiB
uint64_t mem_written_pages(uint64_t addr, uint64_t size, uint64_t *pages, uint64_t max);
// GetWriteWatch(RESET): fills pages[] (4 KiB addresses)
// with the pages of [addr, addr+size) written since the
// last call, returns their count (at most max; a page
// past max stays reported next time). The range must lie
// inside the stack/heap block or the alias region
// (write watch); otherwise returns UINT64_MAX
uint64_t mem_written_peek(uint64_t addr, uint64_t size, uint64_t *pages, uint64_t max);
// the same without the reset: the bits stay for the next
// mem_written_pages (verify checks only)
void mem_kernel_free(void *p, uint64_t size); // TLR/TLS reuse after thread exit
// ---- modules (loader.c, T1)
typedef struct {
const char *name; uint64_t base, size, text_end, data_off; // text_end, data_off: offsets from base
uint64_t init, init_array, init_array_size, ctor_flag; // absolute guest addresses; init_array_size in bytes
} RtLoaded;
void ldr_load(const char *exefs_dir); // load, relocate (overrides applied), seal; once
int ldr_count(void);
const RtLoaded *ldr_module(int i); // order: main, subsdk0, sdk
uint64_t ldr_symbol(const char *name); // exported definition, 0 if none
uint64_t ldr_plt_stub(const char *module, const char *name); // PLT stub whose GOT slot holds name, 0 if none
const char *ldr_describe(uint64_t a, char *buf, int n); // "main+0x1234 (symbol+0x10)", "host <name>", or the address
uint8_t *ldr_read_verified(const char *dir, const char *file, size_t *len); // whole file, SHA-256 checked against
// modules.c (fatal on mismatch); free() it
// Written by tools/build_game.py into build/<version>/modules.c.
typedef struct { const char *file, *sha256; } RtFileHash;
extern const RtFileHash rt_exefs_sha256[]; // tools/data/expected-<version>.json "exefs_sha256"
extern const int rt_exefs_sha256_count;
extern const char rt_default_exefs[], rt_default_romfs[], rt_default_profile[];
// ---- kernel (kernel.c, T2)
typedef struct RtThread RtThread;
Cpu *kern_cpu(void); // current guest thread's Cpu, NULL on host threads
uint32_t kern_main_thread(Cpu **cpu); // main thread object and Cpu for the calling host thread
// (main.npdm: 1 MiB stack at the top of the stack region,
// priority 44, core 0); on the main thread: its handle
void kern_main_stack(uint64_t *base, uint64_t *size); // the main thread's stack, size 0 before kern_main_thread;
// svc.c reports it to QueryMemory, memory.c does not know it
void kern_run_on_guest_main(void (*fn)(Cpu *c)); // start the guest main host thread, run fn there, join
int kern_core_away(void); // give the virtual core up around a blocking host wait
void kern_core_back(int *away);
void kern_checkpoint(void); // preemption point, at every SVC and indirect call
// A preempt request is pending for the thread (Cpu.thread): its RtThread starts with the flag (kernel.c).
static inline int kern_preempt_pending(const void *thread) { return __atomic_load_n((const int *)thread, __ATOMIC_RELAXED); }
void kern_sched_report(uint64_t frame); // TOTK_TRACE=sched: per-thread time split, on present
uint32_t kern_handle_new(int kind, void *obj); // kind: KERN_SESSION or KERN_TMEM (ipc.c)
void *kern_handle_get(uint32_t h, int kind);
void kern_handle_close(uint32_t h); // CloseHandle from the host side, any thread; fatal on a
// bad handle
enum { KERN_SESSION = 4, KERN_TMEM = 5 };
uint32_t kern_event_new(void); // writable handle
uint32_t kern_event_readable(uint32_t writable);
void kern_event_signal(uint32_t writable);
void kern_bind_system_event(Cpu *c, uint64_t ev, int autoclear); // guest thread only
void kern_signal_system_event(uint64_t ev); // any thread
int kern_svc(Cpu *c, uint32_t imm); // thread/sync SVCs; 1 if handled
uint64_t kern_now_ns(void);
typedef struct { uint64_t svcs, threads, host_calls; } RtStats;
RtStats rt_stats(void); // host_calls: hostfn_calls()
// ---- IPC (ipc.c, T3)
int ipc_svc(Cpu *c, uint32_t imm); // 0x15, 0x1F, 0x21, 0x22; 1 if handled
void ipc_handle_closed(int kind, void *obj); // CloseHandle on a KERN_SESSION/KERN_TMEM handle
// ---- file system (romfs.c / fs_host.c, T5)
void fs_init(void); // idempotent; paths from rt_config()
// ---- account (ov_system.c, T4)
void account_uid(uint64_t out[2]); // the local user's nn::account::Uid {lo, hi}, persistent in
// <profile>/account.uid; fs_host.c names save data after it
// ---- movies (ov_movie.c)
void movie_tick(Cpu *c); // at every present, on the presenting guest thread: hands
// decoded movie frames over when due, signals their events
// ---- code patches (patches.c) and the frame step (frame_rate.c)
void rt_patches_init(void); // after ldr_load: switches from the environment, the patch
// page committed; logs the sets
int rt_patch_set(const char *name); // index of a set of generated patches.h, -1 if absent
int rt_patch_enabled(int set); // the set is on (0 for -1)
void rt_patch_disable(int set); // before the guest starts: a set needs another that is off
void *rt_patch_page(void); // the page the patched code reads (NULL without the table)
void frame_step_init(void); // after rt_patches_init: the real-time step's first value
int frame_step_realtime(void); // the real-time step is on: presents ignore the interval
void frame_step_present(uint64_t display_ns, int interval); // nvnQueuePresentTexture: the frame's display time and
// present interval (60 Hz ticks)
// ---- boot (boot.c T6, main.c T1)
typedef struct { const char *exefs, *romfs, *profile; int boot; int frames; const char *shot, *gpu; } RtConfig;
// romfs: directory with romfs.bin and index.json;
// frames: --frames N (0 = no limit); shot: --shot
// PATH; gpu: TOTK_GPU
const RtConfig *rt_config(void);
void rt_boot_event(const char *what, Cpu *c, const char *detail); // checkpoint log; "nvn" ends --boot
void rt_boot(void); // T6: steps 9-14 of section 1
// Modes of totk.exe. main.c has run mem_init and parsed the paths; each mode calls ldr_load itself if it needs the
// modules, registers its host functions, then seals (hostfn_seal, or implicitly at the first rt_call). Return the
// process exit code.
int probe_cfg(void); // cfg_probe.c
int probe_loader(void), probe_kernel(void), probe_ipc(void), probe_services(void), probe_fs(void);
int probe_gfx_abi(void); // M4 T1: gfx_abi_probe.c
int probe_vkdevice(void); // M4 T2: vk_device.c
int probe_nvn(void), probe_nvn_tex(void); // M4 T3: nvn_probe.c
int probe_window(void); // M4 T2
int probe_shader(void); // M4 T4
int probe_vkmem(void), probe_vkexec(void); // M4 T5, T6
int probe_movie(void); // movie_probe.c: host VP9 movie::Decoder on one RomFS movie
int vk_pipeline_prepare_shaders(void); // vk_pipeline.c: --prepare-shaders (tools/prepare_shaders.py)
// gfx_stubs.c defines each of these weakly: prints "probe-x: not built yet", returns 2.