// Multi-threaded file-cache/VAD fault regression test.
// Several threads map the same file mapping and fault/read pages concurrently.

#include <stdint.h>
#include <stddef.h>
#include <string.h>
#include <ntlite/syscall.h>

#define THREAD_COUNT 6u
#define ITERATIONS   16u
#define TEST_SIZE    32768u

static uint8_t gPattern[TEST_SIZE];

typedef struct THREAD_CTX {
    uint32_t Index;
    uint32_t Mapping;
    volatile uint32_t Status;
    volatile uint32_t IterationsDone;
} THREAD_CTX;

static THREAD_CTX gThreads[THREAD_COUNT];

static void print(const char* s) {
    SYS_WRITE_CONSOLE_REQUEST req = { SysSvFromCstr(s) };
    SysWriteConsole(&req);
}

static void print_u64(uint64_t v) {
    char buf[32];
    int i = 0;
    if (v == 0) { print("0"); return; }
    while (v && i < (int)sizeof(buf)) {
        buf[i++] = (char)('0' + (v % 10));
        v /= 10;
    }
    while (i--) {
        char c[2] = { buf[i], 0 };
        print(c);
    }
}

static void exit_now(int code) {
    while (true) SysExitProcess(code);
}

static void fail(const char* msg) {
    print("pcache_multithread: ");
    print(msg);
    print("\n");
    exit_now(1);
}

static void fail_status(const char* what, uint32_t status) {
    print("pcache_multithread: ");
    print(what);
    print(" failed status=");
    print_u64(status);
    print("\n");
    exit_now(1);
}

static uint32_t query_cache(SYS_FILE_CACHE_STATS* out) {
    memset(out, 0, sizeof(*out));
    return SysQueryKernelStats(SYS_KERNEL_STATS_FILE_CACHE, out, sizeof(*out), 0);
}

static void delete_path(const char* path) {
    SYS_PATH_REQUEST req = { SysSvFromCstr(path) };
    (void)SysDeleteFile(&req);
}

static uint32_t create_temp(uint32_t* fileOut, const char** pathOut) {
    static const char* paths[] = {
        "pcache_multithread.tmp",
        "/pcache_multithread.tmp",
        "1:/pcache_multithread.tmp",
        "0:/pcache_multithread.tmp",
    };

    for (size_t i = 0; i < sizeof(paths) / sizeof(paths[0]); ++i) {
        uint32_t h = 0;
        delete_path(paths[i]);
        SYS_OPEN_FILE_REQUEST req = SysMakeOpenFileRequest(SysSvFromCstr(paths[i]),
                                                               SYS_FILE_ACCESS_READ | SYS_FILE_ACCESS_WRITE | SYS_FILE_ACCESS_QUERY,
                                                               SYS_FILE_CREATE_NEW,
                                                               SYS_FILE_FLAG_NONE,
                                                               &h);
        uint32_t s = SysOpenFile(&req);
        if (s == NTLSTATUS_SUCCESS) {
            *fileOut = h;
            if (pathOut) *pathOut = paths[i];
            return NTLSTATUS_SUCCESS;
        }
    }

    return NTLSTATUS_NOT_FOUND;
}

static void fill_pattern(void) {
    for (size_t i = 0; i < TEST_SIZE; ++i)
        gPattern[i] = (uint8_t)(0x5Au ^ (uint8_t)(i * 13u) ^ (uint8_t)(i >> 3));
}

static uint32_t validate_mapping(const uint8_t* base) {
    for (size_t i = 0; i < TEST_SIZE; ++i) {
        if (base[i] != gPattern[i])
            return 0x10000000u | (uint32_t)(i & 0x0FFFFFFFu);
    }
    return NTLSTATUS_SUCCESS;
}

static void thread_main(uintptr_t arg) {
    THREAD_CTX* ctx = (THREAD_CTX*)arg;
    if (!ctx) SysExitThread(1);

    for (uint32_t iter = 0; iter < ITERATIONS; ++iter) {
        void* base = 0;
        uint32_t s = SysMapFile(ctx->Mapping, 0, TEST_SIZE,
                                SYS_MAP_ACCESS_READ, SYS_MAP_FLAG_NONE, &base);
        if (s != NTLSTATUS_SUCCESS || !base) {
            ctx->Status = s ? s : NTLSTATUS_INVALID_PARAMETER;
            SysExitThread(1);
        }

        volatile uint8_t* touch = (volatile uint8_t*)base;
        for (size_t off = 0; off < TEST_SIZE; off += 4096)
            (void)touch[off];
        (void)touch[TEST_SIZE - 1];

        s = validate_mapping((const uint8_t*)base);
        uint32_t fs = SysFreeVm(base);
        if (s != NTLSTATUS_SUCCESS) {
            ctx->Status = s;
            SysExitThread(1);
        }
        if (fs != NTLSTATUS_SUCCESS) {
            ctx->Status = fs;
            SysExitThread(1);
        }
        ctx->IterationsDone++;
    }

    ctx->Status = NTLSTATUS_SUCCESS;
    SysExitThread(0);
}

int main(void) {
    print("pcache_multithread: start\n");
    fill_pattern();

    uint32_t file = 0;
    const char* path = 0;
    uint32_t s = create_temp(&file, &path);
    if (s != NTLSTATUS_SUCCESS) fail_status("create temp", s);

    size_t written = 0;
    SYS_FILE_IO_REQUEST writeReq = SysMakeFileIoRequest(file, gPattern, sizeof(gPattern), &written);
    s = SysWriteFile(&writeReq);
    if (s != NTLSTATUS_SUCCESS) fail_status("write pattern", s);
    if (written != sizeof(gPattern)) fail("short write");

    uint32_t mapping = 0;
    s = SysCreateFileMapping(file, SYS_VM_READ, SYS_MAP_FLAG_NONE, TEST_SIZE, &mapping);
    if (s != NTLSTATUS_SUCCESS) fail_status("create mapping", s);

    SYS_FILE_CACHE_STATS before;
    SYS_FILE_CACHE_STATS after;
    if (query_cache(&before) != NTLSTATUS_SUCCESS) fail("initial stats query failed");

    uint32_t handles[THREAD_COUNT];
    uint32_t tids[THREAD_COUNT];
    memset(handles, 0, sizeof(handles));
    memset(tids, 0, sizeof(tids));
    memset(gThreads, 0, sizeof(gThreads));

    for (uint32_t i = 0; i < THREAD_COUNT; ++i) {
        gThreads[i].Index = i;
        gThreads[i].Mapping = mapping;
        gThreads[i].Status = NTLSTATUS_PENDING;
        SYS_CREATE_THREAD_REQUEST req = {
            (uintptr_t)thread_main,
            (uintptr_t)&gThreads[i],
            1,
            0,
            &handles[i],
            &tids[i]
        };
        s = SysCreateThread(&req);
        if (s != NTLSTATUS_SUCCESS) fail_status("create thread", s);
    }

    for (uint32_t i = 0; i < THREAD_COUNT; ++i) {
        s = SysWait(handles[i], SYS_WAIT_FLAG_NONE);
        if (s != NTLSTATUS_SUCCESS) fail_status("wait thread", s);
        SysCloseHandle(handles[i]);
    }

    for (uint32_t i = 0; i < THREAD_COUNT; ++i) {
        if (gThreads[i].Status != NTLSTATUS_SUCCESS) {
            print("pcache_multithread: worker ");
            print_u64(i);
            print(" status=");
            print_u64(gThreads[i].Status);
            print(" iterations=");
            print_u64(gThreads[i].IterationsDone);
            print("\n");
            exit_now(1);
        }
        if (gThreads[i].IterationsDone != ITERATIONS) fail("worker iteration count mismatch");
    }

    if (query_cache(&after) != NTLSTATUS_SUCCESS) fail("final stats query failed");
    if (after.Loads <= before.Loads && after.Hits <= before.Hits)
        fail("concurrent mmap did not touch cache pages");
    if (after.Hits <= before.Hits) fail("concurrent mmap did not reuse cache pages");

    SysCloseHandle(mapping);
    SysCloseHandle(file);
    delete_path(path);

    print("pcache_multithread: loads+=");
    print_u64(after.Loads - before.Loads);
    print(" hits+=");
    print_u64(after.Hits - before.Hits);
    print(" races+=");
    print_u64(after.RaceDiscards - before.RaceDiscards);
    print("\n");
    print("pcache_multithread: PASS\n");
    exit_now(0);
    return 0;
}