/* bench/stream_sim.c — 边下边播全链离线验证（REPORT §21；NCM_STREAM_PLAY §8 阶段 0/1/2 矩阵）。
 *
 * 链路：假网络线程（可调速率/抖动/断流/断续网）→ 真实 ncm_cache 写者（write(2) 直写 .tmp）
 *   → 真实 stream_src（pread + condvar + fread 语义 read_cb）→ 真实 decoder
 *   （mp3=open_cb SEEK_TO_BYTE|DO_NOT_SCAN / flac=drflac_open 回调）→ 真实 audio_out(SINK_NULL)
 *   + PCM 捕获，与「同文件整首离线解码」逐字节比对（金标准；断流场景比到截断点）。
 *
 * 矩阵：速率 {10MB/s, 400kbps, 160kbps} × {正常播完, seek 已下载区, seek 未下载区等待,
 *   40% 断流, 断续网}；附加 flac 边下边播用例（dr_flac 回调路径）。
 * 断言：跑完无死锁（看门狗）、PCM 逐字节一致（seek 用例除外，断流比前缀）、首开声延迟、
 *   underrun 统计（慢网/断续网 >0）、commit 后缓存文件与源逐字节一致。
 *
 * 编译（宿主，ncm_app 目录下）：
 *   gcc -O2 -Wall -Wextra -DJSMN_HEADER -I. -Icore -I../sdk -I../meow_launcherd \
 *       -Ivendor/jsmn -Ivendor/minimp3 -Ivendor/dr_flac \
 *       bench/stream_sim.c app/play/stream_src.c app/play/decoder.c app/play/audio_out.c \
 *       core/ncm_cache.c -c   # 逐个 -c 后链接，或整体：
 *   gcc -O2 -Wall -Wextra ...同上... -o stream_sim ... app/play/minimp3_impl.o dr_flac_impl.o -lpthread -lm
 *   （minimp3_impl/dr_flac_impl 为 vendor 豁免 TU，单独 -O2 无 -Wall 编译）
 * mipsel：换 $CC_MIPS -static，产物用 qemu-mipsel 跑。
 * 运行：./stream_sim [mp3 路径]   （0 网络；PASS/FAIL 退出码）
 */
#include "../app/ncm_app_internal.h"   /* audio_out 声明 + 全局态（g_st/g_play/g_app_quit） */
#include "../app/play/decoder.h"
#include "../app/play/stream_src.h"
#include "../core/ncm_cache.h"
#include "../core/ncm_common.h"
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <stdarg.h>
#include <unistd.h>
#include <errno.h>
#include <pthread.h>
#include <time.h>
#include <sys/stat.h>

/* ---- 全局态（audio_out.c / ncm_app_internal.h 需要） ---- */
ncm_app_state_t g_st;
ncm_play_run_t  g_play = { .mu = PTHREAD_MUTEX_INITIALIZER };
volatile int    g_app_quit = 0;
volatile int    g_net_ready = 0;
int             g_host_mock = 1;

/* ---- 桩：SDK/核心库其他 TU 符号（本 sim 不链） ---- */
void meow_log(meow_log_level lvl, const char *fmt, ...) { (void)lvl; (void)fmt; }
uint64_t meow_ui_time_ms(void)
{
    struct timespec ts; clock_gettime(CLOCK_MONOTONIC, &ts);
    return (uint64_t)ts.tv_sec * 1000 + (uint64_t)ts.tv_nsec / 1000000;
}
/* audio_out.c 的 AUDIOD 后端符号（本 sim 恒 NCM_AUDIO_SINK=null，不会调到；仅链接用） */
int  meow_audio_open(int rate, int ch, int fmt) { (void)rate; (void)ch; (void)fmt; return -1; }
int  meow_audio_write(const void *pcm, int frames) { (void)pcm; return frames; }
int  meow_audio_pause(int sid, int paused) { (void)sid; (void)paused; return 0; }
int  meow_audio_close(int sid) { (void)sid; return 0; }
int  meow_audio_query(meow_audio_state *out) { (void)out; return -1; }
void meow_audio_on_preempt(meow_audio_preempt_fn cb, void *ctx) { (void)cb; (void)ctx; }
/* ncm_cache.c 依赖（同 cache_evict_sim 口径内联） */
size_t ncm_strlcpy(char *dst, const char *src, size_t sz)
{
    size_t n = 0;
    if (sz) { while (src[n] && n < sz - 1) { dst[n] = src[n]; n++; } dst[n] = 0; }
    while (src[n]) n++;
    return n;
}
int ncm_store_mkdir_p(const char *path)
{
    if (!path || !*path) return -1;
    char tmp[512];
    size_t len = strlen(path);
    if (len >= sizeof tmp) return -1;
    memcpy(tmp, path, len + 1);
    if (tmp[len - 1] == '/') tmp[len - 1] = '\0';
    for (char *p = tmp + 1; *p; p++) {
        if (*p == '/') { *p = '\0'; mkdir(tmp, 0755); *p = '/'; }
    }
    return mkdir(tmp, 0755) == 0 || errno == EEXIST ? 0 : -1;
}

int ncm_store_read_cfg_int(const char *dir, const char *key, int dflt)
{ (void)dir; (void)key; return dflt; }

/* ---- 小工具 ---- */
static double now_s(void)
{
    struct timespec ts; clock_gettime(CLOCK_MONOTONIC, &ts);
    return (double)ts.tv_sec + ts.tv_nsec / 1e9;
}

typedef struct { unsigned char *b; size_t n, cap; } pcm_buf_t;
static void pcm_put(pcm_buf_t *p, const void *data, size_t len)
{
    if (p->n + len > p->cap) {
        size_t nc = p->cap ? p->cap * 2 : (1 << 20);
        while (nc < p->n + len) nc *= 2;
        p->b = (unsigned char *)realloc(p->b, nc);
        if (!p->b) { fprintf(stderr, "oom\n"); exit(2); }
        p->cap = nc;
    }
    memcpy(p->b + p->n, data, len);
    p->n += len;
}

static int g_fail = 0;
#define CHECK(cond, fmt, ...) do { \
    if (cond) printf("    PASS  " fmt "\n", ##__VA_ARGS__); \
    else { printf("    FAIL  " fmt "\n", ##__VA_ARGS__); g_fail = 1; } \
} while (0)

/* ---- 假网络线程：按速率/抖动/断流/断续网 把源文件写进真实 ncm_cache 写者 ---- */
typedef struct {
    const unsigned char *data;
    size_t   size;
    uint64_t rate_Bps;       /* 字节/秒 */
    int      jitter_ms;      /* 每块延迟 ±jitter */
    double   cut_frac;       /* >0: 写到该比例即断流（finish 带错误） */
    int      flaky;          /* 断续网：每 ~6 块额外停 100-400ms */
    stream_src_t        *src;
    ncm_cache_writer_t  *w;
    int      rc;             /* NCM_OK / 负 */
} net_t;

static void *net_thread(void *arg)
{
    net_t *nt = (net_t *)arg;
    const size_t CH = 4096;
    size_t off = 0;
    size_t cut_at = (nt->cut_frac > 0) ? (size_t)(nt->size * nt->cut_frac) : (size_t)-1;
    unsigned rng = 12345;
    int chunk_no = 0;
    while (off < nt->size && off < cut_at) {
        size_t n = nt->size - off;
        if (n > CH) n = CH;
        if (off + n > cut_at) n = cut_at - off;
        /* 节奏：块/速率 + 抖动 */
        double base_ms = nt->rate_Bps ? (double)n * 1000.0 / (double)nt->rate_Bps : 0.0;
        rng = rng * 1103515245u + 12345u;
        double jit = nt->jitter_ms ? ((double)((rng >> 16) % 200) / 100.0 - 1.0) * nt->jitter_ms : 0.0;
        long sleep_ms = (long)(base_ms + jit);
        if (nt->flaky && chunk_no % 6 == 5) {
            rng = rng * 1103515245u + 12345u;
            sleep_ms += 100 + (long)((rng >> 16) % 300);   /* 断续网：额外停滞 */
        }
        if (sleep_ms > 0) usleep((useconds_t)sleep_ms * 1000);
        if (ncm_cache_write(nt->w, nt->data + off, n) != NCM_OK) { nt->rc = NCM_ERR_IO; goto out; }
        stream_src_note_written(nt->src, n);
        off += n;
        chunk_no++;
    }
    nt->rc = (off >= nt->size) ? NCM_OK : NCM_ERR_NET;      /* cut=断流 */
out:
    stream_src_finish(nt->src, nt->rc == NCM_OK ? 0 : nt->rc);
    return NULL;
}

/* ---- 整文件离线解码（金标准 PCM） ---- */
static int decode_offline(const char *path, pcm_buf_t *pcm, int *rate, int *ch, int *bits,
                          int64_t *frames)
{
    ncm_decoder_t *dec = NULL;
    if (ncm_dec_open(&dec, path, rate, ch, bits, NULL) != 0) return -1;
    static unsigned char buf[4096 * 2 * 4];
    int64_t fr, total = 0;
    while ((fr = ncm_dec_read(dec, buf, 4096)) > 0) {
        pcm_put(pcm, buf, (size_t)fr * (size_t)(*ch) * (size_t)(*bits) / 8);
        total += fr;
    }
    ncm_dec_close(dec);
    if (frames) *frames = total;
    return 0;
}

/* ---- 单条用例 ---- */
enum { SC_NORMAL = 0, SC_SEEK_DL, SC_SEEK_WAIT, SC_CUT, SC_FLAKY };

typedef struct {
    pcm_buf_t pcm;
    int64_t   frames;           /* 实际解码输出帧 */
    int64_t   frames_at_seek;   /* seek 时已解码帧（非 seek 用例=0） */
    int64_t   seek_tgt_frames;  /* seek 目标帧（非 seek 用例=0） */
    int       rate, ch, bits;
    double    first_sound_s;    /* 网线程启动 → 首块 PCM */
    int       underruns;
    uint64_t  wait_ms;
    double    wall_s;
    int       open_ok;
} case_result_t;

static int run_case(const char *tag, const unsigned char *data, size_t size,
                    const char *ext, uint64_t rate_Bps, int jitter_ms, double cut_frac, int flaky,
                    int scenario, int64_t golden_total_frames, case_result_t *res)
{
    memset(res, 0, sizeof *res);
    char dir[] = "/tmp/ncm_stream_sim_XXXXXX";
    if (!mkdtemp(dir)) { printf("    FAIL  mkdtemp\n"); return -1; }
    char cdir[600];
    snprintf(cdir, sizeof cdir, "%s/cache", dir);

    ncm_cache_t cache;
    ncm_cache_init(&cache, cdir);

    char key[32];
    static int case_no = 0;
    snprintf(key, sizeof key, "sim%d", case_no++);

    ncm_cache_writer_t *w = NULL;
    if (ncm_cache_open(&cache, key, ext, &w) != NCM_OK) { printf("    FAIL  cache open\n"); return -1; }
    stream_src_t *src = NULL;
    if (stream_src_open_stream(&src, ncm_cache_writer_tmp_path(w)) != 0) {
        printf("    FAIL  stream open\n"); return -1;
    }

    net_t nt = { data, size, rate_Bps, jitter_ms, cut_frac, flaky, src, w, NCM_OK };
    pthread_t nth;
    double t0 = now_s();
    if (pthread_create(&nth, NULL, net_thread, &nt) != 0) { printf("    FAIL  pthread\n"); return -1; }

    /* 预缓冲 ≥128KB（同 player_manager 门槛） */
    double wd_t0 = now_s();
    int wr;
    while ((wr = stream_src_wait(src, 128 * 1024, 100)) == 0) {
        if (now_s() - wd_t0 > 120) { printf("    FAIL  prebuffer watchdog (deadlock?)\n"); g_fail = 1; _exit(1); }
    }
    if (stream_src_failed(src) && stream_src_written(src) < 128 * 1024) {
        /* 首开声前失败（written < 预缓冲门槛）。已下过门槛的断流
         *（如 cut40@10MB/s：finish 与预缓冲满足的竞态）照常放行，播到截断点（F5） */
        printf("    FAIL  download failed before prebuffer\n");
        pthread_join(nth, NULL);
        ncm_cache_abort(w); stream_src_close(src);
        return -1;
    }

    ncm_decoder_t *dec = NULL;
    int rate = 0, ch = 0, bits = 0;
    if (ncm_dec_open_src(&dec, src, &rate, &ch, &bits, NULL) != 0) {
        printf("    FAIL  dec open\n");
        pthread_join(nth, NULL);
        ncm_cache_abort(w); stream_src_close(src);
        return -1;
    }
    res->open_ok = 1;
    res->rate = rate; res->ch = ch; res->bits = bits;
    if (audio_out_open(rate, ch, bits) != 0) { printf("    FAIL  audio open\n"); g_fail = 1; _exit(1); }

    int64_t dur_ms = (golden_total_frames > 0 && rate > 0) ? golden_total_frames * 1000 / rate : 0;
    static unsigned char pcm[4096 * 2 * 4];
    int64_t decoded = 0;            /* 实际解码输出帧（seek 断言口径） */
    int seek_done = 0;
    double first_sound = -1;
    for (;;) {
        if (now_s() - t0 > 150) { printf("    FAIL  case watchdog (deadlock?)\n"); g_fail = 1; _exit(1); }
        int64_t fr = ncm_dec_read(dec, pcm, 4096);
        if (fr == 0) break;
        if (first_sound < 0) first_sound = now_s() - t0;
        if (audio_out_write(pcm, (int)fr) != (int)fr) { printf("    FAIL  audio write\n"); g_fail = 1; _exit(1); }
        pcm_put(&res->pcm, pcm, (size_t)fr * (size_t)ch * (size_t)bits / 8);
        decoded += fr;
        /* 实时步调（模拟 audio_out 背压节拍器，慢网才会真 underrun） */
        if (rate > 0) usleep((useconds_t)(fr * 1000000 / rate));
        /* seek 用例：播 ~1s 后 seek */
        if (!seek_done && dur_ms > 0 && rate > 0 && decoded * 1000 / rate >= 1000) {
            seek_done = 1;
            int64_t tgt = 0;
            if (scenario == SC_SEEK_DL)        tgt = dur_ms / 4;      /* 25%：预缓冲内，必已下载 */
            else if (scenario == SC_SEEK_WAIT) tgt = dur_ms * 9 / 10; /* 90%：慢网未下载区，等待 */
            if (tgt > 0) {
                res->frames_at_seek = decoded;
                res->seek_tgt_frames = tgt * rate / 1000;
                ncm_dec_seek_ms(dec, tgt, (int64_t)size, dur_ms);
            }
        }
    }
    res->first_sound_s = first_sound;
    res->frames = decoded;
    stream_src_stats(src, &res->underruns, &res->wait_ms);
    res->wall_s = now_s() - t0;

    audio_out_close();
    ncm_dec_close(dec);
    pthread_join(nth, NULL);
    stream_src_abort(src);      /* 已完成则无副作用 */

    /* commit/abort + 完整下载的落盘校验（缓存文件与源逐字节一致） */
    if (nt.rc == NCM_OK) {
        if (ncm_cache_commit(w) != 0) { printf("    FAIL  commit\n"); g_fail = 1; }
        else {
            char fp[700];
            if (ncm_cache_path(&cache, key, ext, fp, sizeof fp) == NCM_OK) {
                FILE *f = fopen(fp, "rb");
                if (!f) { printf("    FAIL  committed file missing\n"); g_fail = 1; }
                else {
                    unsigned char *got = (unsigned char *)malloc(size);
                    size_t rn = fread(got, 1, size, f);
                    int extra = fgetc(f);
                    fclose(f);
                    if (rn != size || extra != EOF || memcmp(got, data, size) != 0) {
                        printf("    FAIL  committed bytes mismatch (rn=%zu size=%zu)\n", rn, size);
                        g_fail = 1;
                    }
                    free(got);
                }
            }
        }
    } else {
        ncm_cache_abort(w);
    }
    stream_src_close(src);
    printf("  [%s] wall=%.2fs first_sound=%.2fs frames=%lld underruns=%d wait=%llums\n",
           tag, res->wall_s, res->first_sound_s, (long long)res->frames,
           res->underruns, (unsigned long long)res->wait_ms);
    return 0;
}

int main(int argc, char **argv)
{
    (void)argc;
    const char *mp3 = (argc > 1) ? argv[1] : "vendor/minimp3/vectors/performance/MIPSTest.mp3";
    setenv("NCM_AUDIO_SINK", "null", 1);

    /* 源文件入内存 */
    FILE *f = fopen(mp3, "rb");
    if (!f) { fprintf(stderr, "open %s failed\n", mp3); return 1; }
    fseek(f, 0, SEEK_END); long fsz = ftell(f); fseek(f, 0, SEEK_SET);
    unsigned char *data = (unsigned char *)malloc((size_t)fsz);
    if (!data || fread(data, 1, (size_t)fsz, f) != (size_t)fsz) { fprintf(stderr, "read failed\n"); return 1; }
    fclose(f);
    size_t size = (size_t)fsz;
    printf("src %s: %zu bytes\n", mp3, size);

    /* 金标准：整首离线解码（同样走 cb 路径——ncm_dec_open 内部也是静态 stream_src） */
    pcm_buf_t golden = {0};
    int grate, gch, gbits;
    int64_t gframes;
    if (decode_offline(mp3, &golden, &grate, &gch, &gbits, &gframes) != 0) {
        fprintf(stderr, "golden decode failed\n"); return 1;
    }
    printf("golden: %dHz x%d %dbit frames=%lld pcm=%zu bytes\n",
           grate, gch, gbits, (long long)gframes, golden.n);

    /* 40% 前缀金标准（断流用例比对到截断点） */
    char prefix[] = "/tmp/ncm_stream_prefix_XXXXXX";
    int pfd = mkstemp(prefix);
    if (pfd < 0) return 1;
    size_t cut_sz = size * 2 / 5;
    if (write(pfd, data, cut_sz) != (ssize_t)cut_sz) { fprintf(stderr, "prefix write\n"); return 1; }
    close(pfd);
    pcm_buf_t golden_cut = {0};
    int t1, t2, t3;
    if (decode_offline(prefix, &golden_cut, &t1, &t2, &t3, NULL) != 0) {
        fprintf(stderr, "golden prefix decode failed\n"); return 1;
    }
    unlink(prefix);
    printf("golden-40%%: pcm=%zu bytes\n", golden_cut.n);

    struct { const char *name; uint64_t Bps; } rates[] = {
        { "10MB/s",  10485760u },
        { "400kbps", 50000u },
        { "160kbps", 20000u },
    };
    case_result_t r;
    int flaky_ur_total = 0;   /* 断续网 underrun 跨速率聚合（见 flaky 用例注释） */

    for (int ri = 0; ri < 3; ri++) {
        const char *rn = rates[ri].name;
        uint64_t Bps = rates[ri].Bps;
        char tag[64];

        snprintf(tag, sizeof tag, "normal@%s", rn);
        printf("== CASE %s ==\n", tag);
        if (run_case(tag, data, size, "mp3", Bps, 0, 0, 0, SC_NORMAL, gframes, &r) == 0) {
            CHECK(r.pcm.n == golden.n && memcmp(r.pcm.b, golden.b, golden.n) == 0,
                  "PCM 与整首离线解码逐字节一致 (%zu/%zu)", r.pcm.n, golden.n);
            if (Bps >= 1000000u)
                CHECK(r.first_sound_s >= 0 && r.first_sound_s < 2.0,
                      "首开声 %.2fs < 2s", r.first_sound_s);
            else if (Bps >= 50000u)
                /* 物理下限：预缓冲 128KB / 50KB/s = 2.62s（F3 红线：MINIMP3_IO_SIZE 强制），
                 * <2s@400kbps 与该红线冲突——按可达阈值断言并在 REPORT §21 记录。 */
                CHECK(r.first_sound_s >= 0 && r.first_sound_s < 4.0,
                      "首开声 %.2fs < 4s (128KB/50KBps=2.62s 下限)", r.first_sound_s);
            if (Bps <= 20000u)
                CHECK(r.underruns > 0, "慢网 underrun 统计 >0 (=%d)", r.underruns);
        }

        snprintf(tag, sizeof tag, "seek_dl@%s", rn);
        printf("== CASE %s ==\n", tag);
        if (run_case(tag, data, size, "mp3", Bps, 0, 0, 0, SC_SEEK_DL, gframes, &r) == 0) {
            /* 期望 = seek 前已解码 + 目标之后剩余（字节近似 seek，±10% 容差） */
            int64_t expect = r.frames_at_seek + (gframes - r.seek_tgt_frames);
            double dev = expect > 0 ? (double)(r.frames - expect) / (double)expect : 1;
            if (dev < 0) dev = -dev;
            CHECK(dev < 0.10, "seek 已下载区后播完：decoded=%lld ≈ %lld (dev=%.1f%%)",
                  (long long)r.frames, (long long)expect, dev * 100);
        }

        snprintf(tag, sizeof tag, "seek_wait@%s", rn);
        printf("== CASE %s ==\n", tag);
        if (run_case(tag, data, size, "mp3", Bps, 0, 0, 0, SC_SEEK_WAIT, gframes, &r) == 0) {
            int64_t expect = r.frames_at_seek + (gframes - r.seek_tgt_frames);
            double dev = expect > 0 ? (double)(r.frames - expect) / (double)expect : 1;
            if (dev < 0) dev = -dev;
            CHECK(dev < 0.15, "seek 未下载区等待后播完：decoded=%lld ≈ %lld (dev=%.1f%%)",
                  (long long)r.frames, (long long)expect, dev * 100);
        }

        snprintf(tag, sizeof tag, "cut40@%s", rn);
        printf("== CASE %s ==\n", tag);
        if (run_case(tag, data, size, "mp3", Bps, 0, 0.4, 0, SC_CUT, gframes, &r) == 0) {
            CHECK(r.pcm.n == golden_cut.n && memcmp(r.pcm.b, golden_cut.b, golden_cut.n) == 0,
                  "断流：PCM 与 40%% 前缀离线解码逐字节一致 (%zu/%zu)", r.pcm.n, golden_cut.n);
        }

        snprintf(tag, sizeof tag, "flaky@%s", rn);
        printf("== CASE %s ==\n", tag);
        if (run_case(tag, data, size, "mp3", Bps, 15, 0, 1, SC_FLAKY, gframes, &r) == 0) {
            CHECK(r.pcm.n == golden.n && memcmp(r.pcm.b, golden.b, golden.n) == 0,
                  "断续网：PCM 逐字节一致 (%zu/%zu)", r.pcm.n, golden.n);
            /* 单次断续网是否打空缓冲取决于解码速度（qemu 远慢于宿主→缓冲消耗慢→可能 0 次），
             * 属时序非确定性——不做单用例硬断言，跨速率聚合见循环后。 */
            flaky_ur_total += r.underruns;
        }
    }
    CHECK(flaky_ur_total > 0, "断续网跨速率 underrun 聚合 >0 (=%d)", flaky_ur_total);

    /* flac 边下边播兼容（dr_flac 回调路径接到 stream_src） */
    const char *flac = "out/test-vectors/sweep_44k1_16.flac";
    FILE *ff = fopen(flac, "rb");
    if (!ff) {
        printf("== CASE flac_stream == SKIP（%s 不存在；§19.9 可一键重造）\n", flac);
    } else {
        fseek(ff, 0, SEEK_END); long fz = ftell(ff); fseek(ff, 0, SEEK_SET);
        unsigned char *fdata = (unsigned char *)malloc((size_t)fz);
        if (!fdata || fread(fdata, 1, (size_t)fz, ff) != (size_t)fz) { fprintf(stderr, "flac read\n"); return 1; }
        fclose(ff);
        pcm_buf_t fg = {0};
        int64_t fgframes;
        if (decode_offline(flac, &fg, &t1, &t2, &t3, &fgframes) != 0) { fprintf(stderr, "flac golden\n"); return 1; }
        printf("== CASE flac_stream@400kbps ==\n");
        if (run_case("flac_stream@400kbps", fdata, (size_t)fz, "flac",
                     50000u, 10, 0, 1, SC_NORMAL, fgframes, &r) == 0) {
            CHECK(r.rate == 44100 && r.bits == 16, "flac 参数 44.1k/16bit (=%d/%d)", r.rate, r.bits);
            CHECK(r.pcm.n == fg.n && memcmp(r.pcm.b, fg.b, fg.n) == 0,
                  "flac 边下边播 PCM 逐字节一致 (%zu/%zu)", r.pcm.n, fg.n);
        }
        free(fdata);
        free(fg.b);
    }

    free(golden.b); free(golden_cut.b); free(data);
    printf(g_fail ? "== STREAM SIM: FAIL ==\n" : "== STREAM SIM: ALL PASS ==\n");
    return g_fail;
}