diff --git a/runtime/prof_win.c b/runtime/prof_win.c index afbe799c..36e9c575 100644 --- a/runtime/prof_win.c +++ b/runtime/prof_win.c @@ -9,6 +9,13 @@ // thread's stack and, after // ResumeThread, unwinds the copy with the .pdata tables (RtlVirtualUnwind), so it takes no lock while the thread is // suspended. The unwind stops at the first frame outside the copy; deeper callers are cut. +// TOTK_PROFILE_SLOW= keeps every thread's chains of the last half second in a ring (a thread outside the exe, mostly +// waiting, copies 8 KiB of stack, one in it STK_COPY; a thread that ran no cycle since its previous sample is not +// suspended and repeats its previous chain; a sample that took over 2 ms to stop its thread, which happens while the +// thread is in a long kernel call, is written as `tstall ...` with the +// chain where the thread stopped) and, after a present at least after the previous one, writes +// that interval's samples (plus the 20 ms before) to .f: the profile format, one `stack 1` line per +// sample, and `tstack ...` lines that keep the thread and the time. At most 64 dumps. #include #include #include @@ -25,14 +32,15 @@ #define PROF_SLOTS (1u << 16) // Rip hash table, open addressing; a full table drops new Rips #define PROF_TOP 4000 // more than the 200 asked for: the report sums per function #define PROF_THREADS 512 -#define PROF_DEPTH 24 // frames kept per call chain, the sampled Rip first #define STK_SLOTS (1u << 15) // call-chain hash table, open addressing; a full table drops new chains #define PROF_STACKS 6000 #define STK_COPY (64u << 10) // stack bytes copied per sample #define STK_SLACK (64u << 10) // zeros after the copy: a frame reaching past it reads zeros, and a zero Rip ends the chain typedef struct { uint64_t rip, n; } RipCount; -typedef struct { DWORD tid; HANDLE h; uint64_t n, in_exe; } ThreadSlot; +#define PROF_DEPTH 24 // frames kept per call chain, the sampled Rip first +// cycles, depth, last: TOTK_PROFILE_SLOW only, the thread's cycle count and chain at its previous sample +typedef struct { DWORD tid; HANDLE h; uint64_t n, in_exe, cycles; uint32_t depth; uint64_t last[PROF_DEPTH]; } ThreadSlot; static RipCount table[PROF_SLOTS]; typedef struct { uint64_t hash, n; uint32_t depth; uint64_t rip[PROF_DEPTH]; } StackCount; @@ -54,6 +62,18 @@ static int chains; // call chains are kept: TOTK_PROFILE_RENDER static volatile uint64_t cur_frame; // last presented frame static uint64_t frame_lo, frame_hi; // TOTK_PROFILE_FRAMES=-: sample only while presents lo..hi-1 are pending +#define SLOW_RING (1u << 15) // TOTK_PROFILE_SLOW: about 0.5 s of samples of every thread +#define SLOW_DUMPS 64 +typedef struct { uint64_t t; DWORD tid; uint32_t depth; uint64_t rip[PROF_DEPTH]; } Sample; +static Sample ring[SLOW_RING]; +static uint64_t ring_n; +#define STALL_RING 1024 +static Sample stalls[STALL_RING]; // t: when the stop began; rip[PROF_DEPTH - 1] holds the stop's ns +static uint64_t stall_n; +static uint64_t slow_ns; // TOTK_PROFILE_SLOW in ns, 0 off +static volatile LONG64 dump_lo, dump_hi, dump_frame; // the render thread asks for a window, the sampler writes it +static volatile DWORD replay_tid; // asynchronous replay (vk_backend.c): its waits hold up the frame too + static void count_rip(uint64_t rip) { uint32_t i = (uint32_t)((rip * 0x9E3779B97F4A7C15ull) >> 48); for (uint32_t k = 0; k < 64; k++, i = (i + 1) & (PROF_SLOTS - 1)) { @@ -138,6 +158,54 @@ static void refresh_threads(DWORD self) { n_threads = n; } +static void write_modules(FILE *f) { + HMODULE mods[1024]; + DWORD need = 0; + if (!EnumProcessModules(GetCurrentProcess(), mods, sizeof mods, &need)) return; + for (DWORD i = 0; i < need / sizeof(HMODULE) && i < 1024; i++) { + MODULEINFO mi; + char path[MAX_PATH]; + if (GetModuleInformation(GetCurrentProcess(), mods[i], &mi, sizeof mi) && GetModuleFileNameA(mods[i], path, sizeof path)) + fprintf(f, "module %llx %lx %s\n", (unsigned long long)(uintptr_t)mi.lpBaseOfDll, mi.SizeOfImage, path); + } +} + +// The sampler, between ticks: the samples of the window the render thread asked for. +static void slow_dump(void) { + uint64_t lo = (uint64_t)dump_lo, hi = (uint64_t)dump_hi, frame = (uint64_t)dump_frame; + static int dumps; + char path[1100]; + snprintf(path, sizeof path, "%s.f%llu", out_path, (unsigned long long)frame); + FILE *f = dumps < SLOW_DUMPS ? fopen(path, "w") : NULL; + if (f) { + dumps++; + fprintf(f, "totk-prof 1\nbase %llx\nperiod_us 1000\nframe %llu interval_ms %.1f\nmain %lu render %lu replay %lu\n", + (unsigned long long)(uintptr_t)GetModuleHandleW(NULL), (unsigned long long)frame, (hi - lo) / 1e6 - 20, + main_tid, render_tid, replay_tid); + write_modules(f); + uint64_t first = ring_n > SLOW_RING ? ring_n - SLOW_RING : 0; + for (int pass = 0; pass < 2; pass++) + for (uint64_t k = first; k < ring_n; k++) { + const Sample *s = &ring[k & (SLOW_RING - 1)]; + if (s->t < lo || s->t > hi) continue; + if (pass) fprintf(f, "tstack %lu %.1f", s->tid, (s->t - lo) / 1e6); + else fprintf(f, "stack 1"); + for (uint32_t d = 0; d < s->depth; d++) fprintf(f, " %llx", (unsigned long long)s->rip[d]); + fputc('\n', f); + } + for (uint64_t k = stall_n > STALL_RING ? stall_n - STALL_RING : 0; k < stall_n; k++) { + const Sample *s = &stalls[k & (STALL_RING - 1)]; + if (s->t + s->rip[PROF_DEPTH - 1] < lo || s->t > hi) continue; + fprintf(f, "tstall %lu %.1f %.1f", s->tid, ((double)s->t - (double)lo) / 1e6, s->rip[PROF_DEPTH - 1] / 1e6); + for (uint32_t d = 0; d < s->depth; d++) fprintf(f, " %llx", (unsigned long long)s->rip[d]); + fputc('\n', f); + } + fclose(f); + rt_log("prof: slow frame %llu (%.1f ms) -> %s", (unsigned long long)frame, (hi - lo) / 1e6 - 20, path); + } + InterlockedExchange64(&dump_hi, 0); +} + static DWORD WINAPI sampler_main(void *arg) { (void)arg; DWORD self = GetCurrentThreadId(); @@ -153,27 +221,54 @@ static DWORD WINAPI sampler_main(void *arg) { Sleep(1); } if (tick % 64 == 0) refresh_threads(self); + if (dump_hi) slow_dump(); if (frame_hi && (cur_frame < frame_lo || cur_frame >= frame_hi)) continue; + uint64_t now = slow_ns ? kern_now_ns() : 0; for (int i = 0; i < n_threads; i++) { ThreadSlot *t = &threads[i]; if ((only_render && t->tid != render_tid) || (only_main && t->tid != main_tid)) continue; + uint64_t cyc = 0; + if (slow_ns && t->depth && QueryThreadCycleTime(t->h, &cyc) && cyc == t->cycles) { // still waiting + Sample *s = &ring[ring_n++ & (SLOW_RING - 1)]; + s->t = now, s->tid = t->tid, s->depth = t->depth; + memcpy(s->rip, t->last, t->depth * sizeof *t->last); + continue; + } ctx.ContextFlags = chains ? CONTEXT_CONTROL | CONTEXT_INTEGER : CONTEXT_CONTROL; + uint64_t stop0 = slow_ns ? kern_now_ns() : 0; if (SuspendThread(t->h) == (DWORD)-1) continue; - BOOL ok = GetThreadContext(t->h, &ctx); + BOOL ok = GetThreadContext(t->h, &ctx); // returns once the thread has stopped in user mode + uint64_t stop_ns = slow_ns ? kern_now_ns() - stop0 : 0; size_t n = 0; MEMORY_BASIC_INFORMATION mb; - if (ok && chains && VirtualQuery((void *)ctx.Rsp, &mb, sizeof mb) && mb.State == MEM_COMMIT) { + // a thread stopped between its sub rsp and the stack probe has Rsp in the guard page: touching it would fault + if (ok && chains && VirtualQuery((void *)ctx.Rsp, &mb, sizeof mb) && mb.State == MEM_COMMIT && + !(mb.Protect & (PAGE_GUARD | PAGE_NOACCESS))) { n = (uintptr_t)mb.BaseAddress + mb.RegionSize - ctx.Rsp; // the committed stack above Rsp - if (n > STK_COPY) n = STK_COPY; + size_t cap = slow_ns && (ctx.Rip < exe_lo || ctx.Rip >= exe_hi) ? 8u << 10 : STK_COPY; + if (n > cap) n = cap; memcpy(stk_buf, (void *)ctx.Rsp, n); } ResumeThread(t->h); if (!ok) continue; count_rip(ctx.Rip); - if (n) { - uint64_t chain[PROF_DEPTH]; - uint32_t d = unwind_copy(&ctx, ctx.Rsp, n, chain); - if (d) count_stack(chain, d); + uint64_t chain[PROF_DEPTH] = {ctx.Rip}; + uint32_t d = n ? unwind_copy(&ctx, ctx.Rsp, n, chain) : 0; + if (d) count_stack(chain, d); + if (slow_ns) { + Sample *s = &ring[ring_n++ & (SLOW_RING - 1)]; + s->t = now, s->tid = t->tid, s->depth = d ? d : 1; + memcpy(s->rip, chain, s->depth * sizeof *chain); + if (stop_ns > 2000000) { + Sample *w = &stalls[stall_n++ & (STALL_RING - 1)]; + *w = *s; + w->t = stop0; + if (w->depth > PROF_DEPTH - 1) w->depth = PROF_DEPTH - 1; + w->rip[PROF_DEPTH - 1] = stop_ns; + } + QueryThreadCycleTime(t->h, &t->cycles); // after ResumeThread: the cycles the sample itself cost + t->depth = s->depth; + memcpy(t->last, chain, s->depth * sizeof *chain); } total++; t->n++; @@ -207,16 +302,7 @@ void prof_write(void) { (unsigned long long)total, (unsigned long long)dropped); for (int i = 0; i < n_threads; i++) fprintf(f, "thread %lu %llu %llu\n", threads[i].tid, (unsigned long long)threads[i].n, (unsigned long long)threads[i].in_exe); - HMODULE mods[1024]; - DWORD need = 0; - if (EnumProcessModules(GetCurrentProcess(), mods, sizeof mods, &need)) { - for (DWORD i = 0; i < need / sizeof(HMODULE) && i < 1024; i++) { - MODULEINFO mi; - char path[MAX_PATH]; - if (GetModuleInformation(GetCurrentProcess(), mods[i], &mi, sizeof mi) && GetModuleFileNameA(mods[i], path, sizeof path)) - fprintf(f, "module %llx %lx %s\n", (unsigned long long)(uintptr_t)mi.lpBaseOfDll, mi.SizeOfImage, path); - } - } + write_modules(f); int cnt = 0; for (uint32_t i = 0; i < PROF_SLOTS; i++) if (table[i].rip) table[cnt++] = table[i]; @@ -267,7 +353,9 @@ void prof_start(void) { only_render = cfg_get("TOTK_PROFILE_RENDER") != NULL; only_main = !only_render && cfg_get("TOTK_PROFILE_MAIN") != NULL; const char *ch = cfg_get("TOTK_PROFILE_CHAINS"); - chains = only_render || only_main || (ch && *ch && strcmp(ch, "0")); + const char *sl = cfg_get("TOTK_PROFILE_SLOW"); + slow_ns = sl && atoi(sl) > 0 ? (uint64_t)atoi(sl) * 1000000ull : 0; + chains = only_render || only_main || slow_ns || (ch && *ch && strcmp(ch, "0")); const char *fr = cfg_get("TOTK_PROFILE_FRAMES"); if (fr && sscanf(fr, "%llu-%llu", (unsigned long long *)&frame_lo, (unsigned long long *)&frame_hi) != 2) frame_hi = 0; MODULEINFO mi; @@ -290,8 +378,6 @@ static uint64_t wait_ns[PERF_N], frame_ns[PERF_N], worst_ns[PERF_N]; // window uint64_t perf_begin(void) { return kern_now_ns(); } -static volatile DWORD replay_tid; // asynchronous replay (vk_backend.c): its waits hold up the frame too - void perf_replay_thread(void) { replay_tid = GetCurrentThreadId(); } void perf_main_thread(void) { main_tid = GetCurrentThreadId(); } @@ -448,6 +534,11 @@ void perf_present(uint64_t frame) { cur_frame = frame; if (rt_trace & RT_TRACE_SCHED) kern_sched_report(frame); uint64_t now = kern_now_ns(), d = last ? now - last : 0; + if (slow_ns && last && d >= slow_ns && !dump_hi) { + dump_lo = (LONG64)(now - d - 20000000ull); + dump_frame = (LONG64)frame; + InterlockedExchange64(&dump_hi, (LONG64)now); + } uint64_t rep = ft_frame(frame, now, d); // d is 0 on the first present: it only starts the period if (__atomic_load_n(&overlay, __ATOMIC_RELAXED)) // the overlay's only cost while it is off: this branch perfmon_frame(now, d, frame_ns, rep, __atomic_load_n(&counts[PERF_CNT_TEX_UP], __ATOMIC_RELAXED));