Merge the host cost fixes: the performance counter is scaled without a 128-bit divide, NVN calls are counted per thread, and nn::fs host functions and NVN memory pool setup give the virtual core up; core diagnostics report waits on owners in host calls

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Valiantsin DzerakhandClaude Opus 5.5 committed 2026-10-27 16:46:38 +03:00
commit 6c14dd5937
7 files changed
+77 -20

No files matched your search

+16 -1
View File
@@ -25,6 +25,7 @@ typedef struct {
GuestFn fn;
uint64_t original; // observers: the guest definition rt_original calls
uint64_t calls;
int blocking; // runs with the virtual core given up (hostfn_blocking)
} Slot;
static Slot slots[MAX_HOSTFN];
static int nslots;
@@ -82,7 +83,7 @@ uint64_t hostfn_register(const char *name, GuestFn fn) {
if (nslots == MAX_HOSTFN) rt_fatal("hostfn_register: more than %d host functions", MAX_HOSTFN);
uint64_t a = slot_addr(nslots);
if (a >= RT_RET_HOST) rt_fatal("hostfn_register: host-function window full");
slots[nslots] = (Slot){.name = name, .fn = fn};
slots[nslots] = (Slot){.name = name, .fn = fn, .blocking = !strncmp(name, "_ZN2nn2fs", 9)};
rt_register_host(a, hostfn_entry);
nslots++;
return a;
@@ -120,6 +121,16 @@ __attribute__((noinline)) static void hostfn_tracked(Cpu *c, GuestFn fn) {
if (!n && kern_preempt_pending(t)) kern_host_left(t);
}
// nn::fs: the call waits for the disk and the OS, and on the Switch it is an IPC request that blocks its thread, so
// the thread gives its virtual core up for it. TOTK_CORE_DIAG found the game's loader threads (priority 58) holding
// a core through CloseFile and CreateCacheStorage while job workers and better threads waited 2-6 ms for it.
__attribute__((noinline)) static void hostfn_blocking(Cpu *c, GuestFn fn) {
int away = kern_core_away();
if (kern_host_track && c->thread) hostfn_tracked(c, fn);
else fn(c);
kern_core_back(&away);
}
static void hostfn_entry(Cpu *c) {
uint64_t i = (c->pc - RT_HOSTFN_BASE) >> 4;
if (__builtin_expect((c->pc & 15) || i >= (uint64_t)nslots, 0)) bad_entry(c);
@@ -129,6 +140,10 @@ static void hostfn_entry(Cpu *c) {
char lr[96];
rt_log("host call %s from %s", s->name, ldr_describe(c->x[30] - 4, lr, sizeof lr));
}
if (s->blocking) {
hostfn_blocking(c, s->fn);
return;
}
if (__builtin_expect(kern_host_track && c->thread, 0)) {
hostfn_tracked(c, s->fn);
return;
+17 -7
View File
@@ -104,6 +104,7 @@ struct RtThread {
uint64_t dg_host_end; // as an owner: the return from the host call or SVC it was in at the request
uint64_t dg_pc, dg_lr; // as an owner: guest pc and lr at the checkpoint that gave the core up
int dg_host_kick; // as an owner: in a host call or SVC at the request
uint32_t dg_svc; // the thread's latest SVC (TOTK_CORE_DIAG: what an owner in a host call is in)
struct DgRel { uint64_t kick, host_end, pc, lr; int host, pre, prio; uint32_t h; } dg_rel; // as a waiter: the
// owner that released
};
@@ -627,7 +628,7 @@ static void dg_done(RtThread *t) {
while (b < DG_NB - 1 && w >= dg_edge[b]) b++;
g_dgw.n[top]++;
g_dgw.hist[top][b]++;
if (part[DG_GAME] >= 2000000 && g_dg_lines < DG_LONG_LINES) {
if ((part[DG_GAME] >= 2000000 || part[DG_HOST] >= 2000000) && g_dg_lines < DG_LONG_LINES) {
g_dg_lines++;
char a[128], l[128];
rt_log("core-diag: wait %.2f ms (game %.2f host %.2f): %#x p%d mask %llx; released by %#x p%d%s at %s lr %s",
@@ -639,8 +640,8 @@ static void dg_done(RtThread *t) {
memset(&t->dg_rel, 0, sizeof t->dg_rel);
}
// Once per long DG_GAME wait: where the owner's host thread runs (exe RVA). Holding G; the owner is resumed before
// anything that could take a lock it holds.
// Once per long DG_GAME or DG_HOST wait: where the owner's host thread runs (module and offset), its latest SVC and
// its guest lr. Holding G; the owner is resumed before anything that could take a lock it holds.
static void dg_sample(RtThread *t) {
RtThread *o = t->target;
if (!o || o->cur_core < 0 || !o->tid || g_dg_samples >= DG_SAMPLES) return;
@@ -658,9 +659,17 @@ static void dg_sample(RtThread *t) {
CloseHandle(h);
if (!ok) return;
g_dg_samples++;
uintptr_t base = (uintptr_t)GetModuleHandleW(NULL);
rt_log("core-diag: sample owner %#x p%d depth %d rip exe+%#llx (waiter %#x p%d, %.2f ms)", o->handle, o->prio, depth,
(unsigned long long)(ctx.Rip - base), t->handle, t->prio, (kern_now_ns() - t->dg_since) / 1e6);
HMODULE m = NULL;
wchar_t mw[MAX_PATH] = L"?";
if (GetModuleHandleExW(GET_MODULE_HANDLE_EX_FLAG_FROM_ADDRESS | GET_MODULE_HANDLE_EX_FLAG_UNCHANGED_REFCOUNT,
(LPCWSTR)(uintptr_t)ctx.Rip, &m) && m)
GetModuleFileNameW(m, mw, MAX_PATH);
const wchar_t *mn = wcsrchr(mw, L'\\') ? wcsrchr(mw, L'\\') + 1 : mw;
char lr[128];
rt_log("core-diag: sample owner %#x p%d depth %d rip %ls+%#llx, last svc %#x, guest lr %s (waiter %#x p%d, %s, "
"%.2f ms)", o->handle, o->prio, depth, mn, (unsigned long long)(ctx.Rip - (uintptr_t)m), o->dg_svc,
o->cur ? ldr_describe(o->cur->x[30], lr, sizeof lr) : "-", t->handle, t->prio,
t->dg_cls == DG_HOST ? "host" : "game", (kern_now_ns() - t->dg_since) / 1e6);
}
@@ -875,7 +884,7 @@ static void core_take(RtThread *t) {
while (t->cur_core < 0) {
DWORD ms = INFINITE;
if (g_ap_int && ap_try(t) != AP_NONE) ms = 1; // the owner keeps its core: again in a millisecond
int sample = g_diag > 0 && !sampled && t->dg_cls == DG_GAME;
int sample = g_diag > 0 && !sampled && (t->dg_cls == DG_GAME || t->dg_cls == DG_HOST);
if (sample && ms == INFINITE) ms = 2;
SleepConditionVariableSRW(&t->cv, &G, ms, 0);
if (sample && t->cur_core < 0 && kern_now_ns() - t->dg_since >= 2000000) {
@@ -1426,6 +1435,7 @@ static void cv_hook(void) {
int kern_svc(Cpu *c, uint32_t imm) {
uint64_t *x = c->x;
if (g_diag > 0 && tcur) tcur->dg_svc = imm;
switch (imm) {
case 0x08: { // CreateThread(-> w1 handle; x1 entry, x2 arg, x3 stack top, w4 priority, w5 core)
int prio = (int)x[4], core = (int)x[5];
+11 -8
View File
@@ -27,6 +27,7 @@
#include "cfg.h"
#include "host.h"
#include "platform/platform.h"
#define MAX_MODULES 16
@@ -105,17 +106,19 @@ void rt_udf(Cpu *c, uint64_t pc) {
ldr_describe(c->x[30], lr, sizeof lr));
}
// CNTPCT_EL0: the Switch counter runs at 19.2 MHz.
// CNTPCT_EL0: the Switch counter runs at 19.2 MHz. The scale is set before main, so no thread reads it half written.
static PlatScale tick_scale;
__attribute__((constructor)) static void tick_scale_init(void) {
LARGE_INTEGER f;
QueryPerformanceFrequency(&f);
tick_scale = plat_scale(19200000u, (uint64_t)f.QuadPart);
}
uint64_t rt_ticks(void) {
static uint64_t freq;
LARGE_INTEGER t;
if (!freq) {
LARGE_INTEGER f;
QueryPerformanceFrequency(&f);
freq = (uint64_t)f.QuadPart;
}
QueryPerformanceCounter(&t);
return (uint64_t)((unsigned __int128)(uint64_t)t.QuadPart * 19200000u / freq);
return plat_scaled(tick_scale, (uint64_t)t.QuadPart);
}
// ---- crash reporter
+14 -1
View File
@@ -24,8 +24,15 @@
#define EVENT_MAGIC MAGIC('E', 'V', 'N', 'T')
// ---------- shared helpers (nvn_api.h)
// The call count goes to the shared counter 1024 at a time: the main thread and the job workers all record command
// buffers, and one locked add per call kept the counter's cache line moving between their cores (0.8k samples a
// fight). The count lags by at most 1023 calls per thread.
void nvn_called(uint8_t *seen, const char *name) {
__atomic_fetch_add(&gfx_counters.nvn_calls, 1, __ATOMIC_RELAXED);
static _Thread_local uint32_t calls;
if (++calls == 1024) {
__atomic_fetch_add(&gfx_counters.nvn_calls, calls, __ATOMIC_RELAXED);
calls = 0;
}
if (__atomic_load_n(seen, __ATOMIC_RELAXED) || __atomic_exchange_n(seen, 1, __ATOMIC_RELAXED)) return;
uint64_t n = __atomic_add_fetch(&gfx_counters.nvn_distinct, 1, __ATOMIC_RELAXED);
NVN_LOG("first call %llu: %s", (unsigned long long)n, name);
@@ -260,7 +267,11 @@ NVN_API(nvnMemoryPoolBuilderSetStorage) {
// void nvnMemoryPoolBuilderSetFlags(NVNmemoryPoolBuilder *b, int flags)
NVN_API(nvnMemoryPoolBuilderSetFlags) { ((pool_builder_t *)arg_p(c, 0))->flags = (uint32_t)arg_x(c, 1); }
// NVNboolean nvnMemoryPoolInitialize(NVNmemoryPool *p, const NVNmemoryPoolBuilder *b)
// Pool setup imports, mirrors and maps memory in the back end (the driver, VirtualAlloc): up to 20 ms on the game's
// loader threads, which held a virtual core through it while job workers waited (TOTK_CORE_DIAG). The thread gives
// the core up, as for any blocking host call.
NVN_API(nvnMemoryPoolInitialize) {
NVN_BLOCKING();
pool_t *o = arg_p(c, 0);
const pool_builder_t *s = arg_p(c, 1);
memset(o, 0, sizeof *o);
@@ -290,6 +301,7 @@ NVN_API(nvnMemoryPoolInitialize) {
}
// void nvnMemoryPoolFinalize(NVNmemoryPool *p)
NVN_API(nvnMemoryPoolFinalize) {
NVN_BLOCKING(); // as nvnMemoryPoolInitialize
pool_t *o = arg_p(c, 0);
if (o->magic != NVN_POOL_MAGIC) return;
if (pool_gpu_visible(o)) index_del(o);
@@ -365,6 +377,7 @@ static int check_request(const pool_t *v, const nvn_mapping_t *r, int i) {
}
// NVNboolean nvnMemoryPoolMapVirtual(NVNmemoryPool *p, int n, const NVNmappingRequest *requests)
NVN_API(nvnMemoryPoolMapVirtual) {
NVN_BLOCKING(); // as nvnMemoryPoolInitialize
pool_t *v = arg_p(c, 0);
int n = (int)arg_x(c, 1);
const nvn_mapping_t *r = arg_p(c, 2);
+4 -2
View File
@@ -327,7 +327,8 @@ static void obs_main(Cpu *c) { observe(c, "nnMain"); }
static void obs_address_region(Cpu *c) {
uint64_t *out = arg_p(c, 0), size = arg_x(c, 1);
uint32_t r = (uint32_t)rt_original(c);
rt_log("AllocateAddressRegion size %#llx result %#x address %#llx", (unsigned long long)size, r,
if (rt_trace & RT_TRACE_HOSTCALL) // 20,000 lines a session otherwise, each a write on a game thread
rt_log("AllocateAddressRegion size %#llx result %#x address %#llx", (unsigned long long)size, r,
(unsigned long long)(r || !out ? 0 : *out));
}
@@ -335,7 +336,8 @@ static void obs_address_region(Cpu *c) {
static void obs_memory_pages(Cpu *c) {
uint64_t a = arg_x(c, 0), size = arg_x(c, 1);
uint32_t r = (uint32_t)rt_original(c);
rt_log("AllocateMemoryPages address %#llx size %#llx result %#x", (unsigned long long)a, (unsigned long long)size, r);
if (rt_trace & RT_TRACE_HOSTCALL)
rt_log("AllocateMemoryPages address %#llx size %#llx result %#x", (unsigned long long)a, (unsigned long long)size, r);
}
// ==== nn::account (account.c): one offline local user whose Uid survives restarts in <profile>/account.uid
+10
View File
@@ -23,6 +23,16 @@ uint32_t plat_cpu_count(void); // logical processors
// ---- time
uint64_t plat_now_ns(void); // monotonic, high resolution, arbitrary epoch
// t * num / den for a fixed num / den without a 128-bit division per call (__udivti3 took 5% of the game's running
// samples through plat_now_ns and rt_ticks): t * q + (t * f >> 64) with q = num / den and f = (num % den) * 2^64 / den.
// Monotonic in t and at most one below the exact quotient while t * (num % den) / den fits 64 bits.
typedef struct { uint64_t q, f; } PlatScale;
static inline PlatScale plat_scale(uint64_t num, uint64_t den) {
return (PlatScale){num / den, (uint64_t)(((unsigned __int128)(num % den) << 64) / den)};
}
static inline uint64_t plat_scaled(PlatScale s, uint64_t t) {
return t * s.q + (uint64_t)(((unsigned __int128)t * s.f) >> 64);
}
void plat_sleep_ns(uint64_t ns); // returns no earlier than ns later; overshoot well under 1 ms
// ---- locks (not recursive) and condition variables; zero-initialised storage is a valid unlocked lock and an empty
+5 -1
View File
@@ -78,10 +78,14 @@ static uint64_t qpc_freq(void) {
return freq;
}
static PlatScale qpc_ns; // set before main, so no thread reads it half written
__attribute__((constructor)) static void qpc_ns_init(void) { qpc_ns = plat_scale(1000000000u, qpc_freq()); }
uint64_t plat_now_ns(void) {
LARGE_INTEGER t;
QueryPerformanceCounter(&t);
return (uint64_t)((unsigned __int128)(uint64_t)t.QuadPart * 1000000000u / qpc_freq());
return plat_scaled(qpc_ns, (uint64_t)t.QuadPart);
}
// A high-resolution waitable timer per sleeping thread (Windows 10 1803+; one handle per thread, kept until exit)