#pragma once // ensure that a given header file is included only once in a
             // single compilation unit
#include "arena.h"
#include "base.h"
#include "lf_mpmc.h"

#include "arena.c"
#include <stdio.h>

// xxhash include
#define XXH_STATIC_LINKING_ONLY
#include "xxh_x86dispatch.h"

// ----------------------------- Config -------------------------------------
#define FILE_HASHES_TXT "file_hashes.txt"
#define HASH_STRLEN 33 // 128-bit hex (32 chars) + null
#define MAX_PATHLEN 4096
#define READ_BLOCK (KiB(64))

// ----------------------------- Globals ------------------------------------
static atomic_uint_fast64_t g_files_found = 0;
static atomic_uint_fast64_t g_files_hashed = 0;
static atomic_uint_fast64_t g_bytes_processed = 0;
static atomic_int g_scan_done = 0;

// ================== OS-agnostic functions abstraction =====================
// ----------------------------- Timer functions --------------
typedef struct {
  u64 start;
  u64 now;
} HiResTimer;

#if defined(_WIN32) || defined(_WIN64)

static LARGE_INTEGER g_freq;

static void timer_init(void) { QueryPerformanceFrequency(&g_freq); }

static void timer_start(HiResTimer *t) {
  LARGE_INTEGER v;
  QueryPerformanceCounter(&v);
  t->start = v.QuadPart;
}

static double timer_elapsed(HiResTimer *t) {
  LARGE_INTEGER v;
  QueryPerformanceCounter(&v);
  t->now = v.QuadPart;

  return (double)(t->now - t->start) / (double)g_freq.QuadPart;
}

#elif defined(__linux__)

void timer_init(void) {}

void timer_start(HiResTimer *t) {
  struct timespec ts;
  clock_gettime(CLOCK_MONOTONIC, &ts);
  t->start = ts.tv_sec * 1000000000ULL + ts.tv_nsec;
}

double timer_elapsed(HiResTimer *t) {
  struct timespec ts;
  clock_gettime(CLOCK_MONOTONIC, &ts);

  uint64_t now = ts.tv_sec * 1000000000ULL + ts.tv_nsec;

  return (double)(now - t->start) / 1e9;
}

#endif

// ----------------------------- Get HW info --------------
#if defined(_WIN32) || defined(_WIN64)

size_t platform_physical_cores(void) {
  DWORD len = 0;
  GetLogicalProcessorInformation(NULL, &len);

  SYSTEM_LOGICAL_PROCESSOR_INFORMATION buf[len];

  GetLogicalProcessorInformation(buf, &len);
  DWORD count = 0;
  DWORD n = len / sizeof(SYSTEM_LOGICAL_PROCESSOR_INFORMATION);
  for (DWORD i = 0; i < n; i++) {
    if (buf[i].Relationship == RelationProcessorCore)
      count++;
  }
  return count ? count : 1;
}

#elif defined(__linux__)

size_t platform_physical_cores(void) {
  long n = sysconf(_SC_NPROCESSORS_ONLN);
  return n > 0 ? (size_t)n : 1;
}

#endif

const char *get_xxhash_instruction_set(void) {
  int vecID = XXH_featureTest();

  switch (vecID) {
  case XXH_SCALAR:
    return "Scalar (portable C)";
  case XXH_SSE2:
    return "SSE2";
  case XXH_AVX2:
    return "AVX2";
  case XXH_AVX512:
    return "AVX-512";
  default:
    return "Unknown";
  }
}

// -------------------- File IO -------------------

#if defined(_WIN32) || defined(_WIN64)
typedef HANDLE FileHandle;
#define FLAG_SEQUENTIAL_READ FILE_FLAG_SEQUENTIAL_SCAN
#define FLAG_ASYNC_DIRECT_READ (FILE_FLAG_OVERLAPPED | FILE_FLAG_NO_BUFFERING)
#define INVALID_FILE_HANDLE INVALID_HANDLE_VALUE

// File open function
static FileHandle os_file_open(const char *path, DWORD flags) {
  return CreateFileA(path, GENERIC_READ, FILE_SHARE_READ | FILE_SHARE_WRITE,
                     NULL, OPEN_EXISTING, flags, NULL);
}

// File read function
static int os_file_read(FileHandle handle, void *buf, size_t count,
                        uint64_t *bytes_read) {
  DWORD read = 0;
  BOOL result = ReadFile(handle, buf, (DWORD)count, &read, NULL);
  *bytes_read = read;
  return (result && read > 0) ? 0 : -1;
}

// File close function
static void os_file_close(FileHandle handle) { CloseHandle(handle); }

#elif defined(__linux__)
typedef int FileHandle;
#define FLAG_SEQUENTIAL_READ (0)
#define FLAG_ASYNC_DIRECT_READ (O_DIRECT)
#define INVALID_FILE_HANDLE (-1)

// File open function
static FileHandle os_file_open(const char *path, int flags) {
  // Combine your mandatory flags with the user-provided flag
  int fd = open(path, O_RDONLY | O_NOFOLLOW | flags);

  // If sequential was requested, advise the kernel
  if (fd != -1 && (flags == FLAG_SEQUENTIAL_READ)) {
    posix_fadvise(fd, 0, 0, POSIX_FADV_SEQUENTIAL);
  }

  return fd;
}

// File read function
static int os_file_read(FileHandle handle, void *buf, size_t count,
                        uint64_t *bytes_read) {
  ssize_t result = read(handle, buf, count);
  if (result >= 0) {
    *bytes_read = (uint64_t)result;
    return 0;
  }
  *bytes_read = 0;
  return -1;
}

// File close function
static void os_file_close(FileHandle handle) { close(handle); }

#endif

// -------------------- Thread abstraction -------------------
// Threads context
typedef struct {
  u8 num_threads;

  mem_arena *path_arena;
  mem_arena *meta_arena;

  MPMCQueue *dir_queue;
  MPMCQueue *file_queue;
} ScannerContext;

typedef struct {
  mem_arena *arena;
  MPMCQueue *file_queue;
} WorkerContext;

#if defined(_WIN32) || defined(_WIN64)
typedef HANDLE ThreadHandle;
typedef DWORD(WINAPI *ThreadFunc)(void *);
#define THREAD_RETURN DWORD WINAPI
#define THREAD_RETURN_VALUE 0;

typedef struct {
  ThreadHandle handle;
  int valid; // Track if thread was successfully created
} Thread;

// Thread function wrapper to handle different return types
#define THREAD_FUNCTION(name) DWORD WINAPI name(LPVOID arg)

// Thread creation function
static int thread_create(Thread *thread, ThreadFunc func, void *arg) {
  thread->handle =
      CreateThread(NULL, 0, (LPTHREAD_START_ROUTINE)func, arg, 0, NULL);
  return (thread->handle != NULL) ? 0 : -1;
}

// Thread join function
static int thread_join(Thread *thread) {
  return (WaitForSingleObject(thread->handle, INFINITE) == WAIT_OBJECT_0) ? 0
                                                                          : -1;
}

// Thread close/detach function
static void thread_close(Thread *thread) { CloseHandle(thread->handle); }

// Wait for multiple threads
static int thread_wait_multiple(Thread *threads, size_t count) {
  HANDLE handles[64]; // Max 64 threads for Windows
  for (size_t i = 0; i < count; i++) {
    handles[i] = threads[i].handle;
  }
  return (WaitForMultipleObjects((DWORD)count, handles, TRUE, INFINITE) ==
          WAIT_OBJECT_0)
             ? 0
             : -1;
}

#elif defined(__linux__)
typedef pthread_t ThreadHandle;
typedef void *(*ThreadFunc)(void *);
#define THREAD_RETURN void *
#define THREAD_RETURN_VALUE NULL;

typedef struct {
  ThreadHandle handle;
  int valid; // Track if thread was successfully created
} Thread;

// Thread function wrapper to handle different return types
typedef struct {
  void *(*func)(void *);
  void *arg;
} ThreadWrapper;

static void *thread_start_routine(void *arg) {
  ThreadWrapper *wrapper = (ThreadWrapper *)arg;
  void *result = wrapper->func(wrapper->arg);
  free(wrapper);
  return result;
}

// Thread creation function
static int thread_create(Thread *thread, ThreadFunc func, void *arg) {
  int ret = pthread_create(&thread->handle, NULL, func, arg);
  if (ret == 0) {
    thread->valid = 1;
  }
  return ret;
}

// Thread join function
static int thread_join(Thread *thread) {
  int ret = pthread_join(thread->handle, NULL);
  thread->valid = 0;
  return ret;
}

// Thread close/detach function
static void thread_close(Thread *thread) {
  if (thread->valid) {
    pthread_detach(thread->handle);
    thread->valid = 0;
  }
}

// Wait for multiple threads
static int thread_wait_multiple(Thread *threads, size_t count) {
  for (size_t i = 0; i < count; i++) {
    if (thread_join(&threads[i]) != 0) {
      return -1;
    }
  }
  return 0;
}

#endif

// ======================== Get file metadata ========================
// -------------------- Path parsing -------------------
static void normalize_path(char *p) {
  char *src = p;
  char *dst = p;
  int prev_slash = 0;

  while (*src) {
    char c = *src++;

    if (c == '\\' || c == '/') {
      if (!prev_slash) {
        *dst++ = '/';
        prev_slash = 1;
      }
    } else {
      *dst++ = c;
      prev_slash = 0;
    }
  }

  *dst = '\0';
}

static int parse_paths(char *line, char folders[][MAX_PATHLEN],
                       int max_folders) {
  int count = 0;
  char *p = line;

  while (*p && count < max_folders) {

    while (*p && isspace((unsigned char)*p))
      p++;

    if (!*p)
      break;

    char *start;
    char quote = 0;

    if (*p == '"' || *p == '\'') {
      quote = *p++;
      start = p;

      while (*p && *p != quote)
        p++;
    } else {
      start = p;

      while (*p && !isspace((unsigned char)*p))
        p++;
    }

    size_t len = p - start;
    if (len >= MAX_PATHLEN)
      len = MAX_PATHLEN - 1;

    memcpy(folders[count], start, len);
    folders[count][len] = 0;

    normalize_path(folders[count]);

    count++;

    if (quote && *p == quote)
      p++;
  }
  return count;
}

// ----------------------------- File time -------------------------
#if defined(_WIN32) || defined(_WIN64)
static void format_time(uint64_t t, char *out, size_t out_sz) {
  if (t == 0) {
    snprintf(out, out_sz, "N/A");
    return;
  }

  time_t tt = (time_t)t;
  struct tm tm;

  localtime_s(&tm, &tt);

  strftime(out, out_sz, "%Y-%m-%d %H:%M:%S", &tm);
}
// ----------------------------- Convert filetime to epoch --------------
static uint64_t filetime_to_epoch(const FILETIME *ft) {
  ULARGE_INTEGER ull;
  ull.LowPart = ft->dwLowDateTime;
  ull.HighPart = ft->dwHighDateTime;

  // Windows epoch (1601) ￢ﾆﾒ Unix epoch (1970)
  return (ull.QuadPart - 116444736000000000ULL) / 10000000ULL;
}

void platform_get_file_times(const char *path, uint64_t *out_created,
                             uint64_t *out_modified) {
  WIN32_FILE_ATTRIBUTE_DATA fad;
  if (GetFileAttributesExA(path, GetFileExInfoStandard, &fad)) {
    *out_created = filetime_to_epoch(&fad.ftCreationTime);
    *out_modified = filetime_to_epoch(&fad.ftLastWriteTime);
  } else {
    *out_created = 0;
    *out_modified = 0;
  }
}

#elif defined(__linux__)
static void format_time(uint64_t t, char *out, size_t out_sz) {
  if (t == 0) {
    snprintf(out, out_sz, "N/A");
    return;
  }

  time_t tt = (time_t)t;
  struct tm tm;

  localtime_r(&tt, &tm);

  strftime(out, out_sz, "%Y-%m-%d %H:%M:%S", &tm);
}

void platform_get_file_times(const char *path, uint64_t *out_created,
                             uint64_t *out_modified) {
  struct stat st;
  if (stat(path, &st) == 0) {
    *out_created = (uint64_t)st.st_ctime;
    *out_modified = (uint64_t)st.st_mtime;
  } else {
    *out_created = 0;
    *out_modified = 0;
  }
}

#endif

// ----------------------------- File owner ---------------------
#if defined(_WIN32) || defined(_WIN64)
static void get_file_owner(const char *path, char *out, size_t out_sz) {
  PSID sid = NULL;
  PSECURITY_DESCRIPTOR sd = NULL;

  if (GetNamedSecurityInfoA(path, SE_FILE_OBJECT, OWNER_SECURITY_INFORMATION,
                            &sid, NULL, NULL, NULL, &sd) == ERROR_SUCCESS) {

    char name[64], domain[64];
    DWORD name_len = sizeof(name);
    DWORD domain_len = sizeof(domain);
    SID_NAME_USE use;

    if (LookupAccountSidA(NULL, sid, name, &name_len, domain, &domain_len,
                          &use)) {
      snprintf(out, out_sz, "%s\\%s", domain, name);
    } else {
      snprintf(out, out_sz, "UNKNOWN");
    }
  } else {
    snprintf(out, out_sz, "UNKNOWN");
  }

  if (sd)
    LocalFree(sd);
}

void platform_get_file_owner(const char *path, char *out_owner,
                             size_t out_owner_size) {
  get_file_owner(path, out_owner, out_owner_size);
}

#elif defined(__linux__)
static void get_file_owner(uid_t uid, char *out, size_t out_sz) {
  struct passwd *pw = getpwuid(uid);
  if (pw) {
    snprintf(out, out_sz, "%s", pw->pw_name);
  } else {
    snprintf(out, out_sz, "UNKNOWN");
  }
}

void platform_get_file_owner(const char *path, char *out_owner,
                             size_t out_owner_size) {
  struct stat st;
  if (stat(path, &st) == 0) {
    get_file_owner(st.st_uid, out_owner, out_owner_size);
  } else {
    snprintf(out_owner, out_owner_size, "UNKNOWN");
  }
}

#endif

// ----------------------------- Scan helpers -----------------------------
typedef struct FileEntry {
  char *path;

  uint64_t size_bytes;
  uint64_t created_time;  // epoch
  uint64_t modified_time; // epoch seconds
  char owner[128];        // resolved owner name
} FileEntry;

typedef struct {
  char buffer[MAX_PATHLEN];
  char *base_end;     // Points to end of base path
  char *filename_pos; // Points to where filename should be written
  size_t base_len;
} PathBuilder;

static void path_builder_init(PathBuilder *pb, const char *base) {
  pb->base_len = strlen(base);
  memcpy(pb->buffer, base, pb->base_len);
  pb->base_end = pb->buffer + pb->base_len;

#if defined(_WIN32) || defined(_WIN64)
  *pb->base_end = '\\';
#elif defined(__linux__)
  *pb->base_end = '/';
#endif

  // Ensure null termination
  *(pb->base_end + 1) = '\0';
  pb->filename_pos = pb->base_end + 1;
}

static void path_builder_set_filename(PathBuilder *pb, const char *filename,
                                      size_t name_len) {
  memcpy(pb->filename_pos, filename, name_len);
  pb->filename_pos[name_len] = '\0'; // Ensure null termination
}

static char *path_builder_dup_arena(PathBuilder *pb, mem_arena *arena,
                                    bool zero) {
  // Calculate total length including base + separator + filename + null
  // terminator
  size_t total_len =
      (pb->filename_pos - pb->buffer) + strlen(pb->filename_pos) + 1;
  char *dup = arena_push(&arena, total_len, zero);
  memcpy(dup, pb->buffer, total_len);
  return dup;
}

#if defined(_WIN32) || defined(_WIN64)
void scan_folder(const char *base, ScannerContext *ctx) {
  PathBuilder pb;
  path_builder_init(&pb, base);

  char search[MAX_PATHLEN];
  memcpy(search, pb.buffer, pb.base_len + 1); // Copy base + separator
  memcpy(search + pb.base_len + 1, "*", 2);   // Add "*" and null

  WIN32_FIND_DATAA fd;
  HANDLE h = FindFirstFileA(search, &fd);
  if (h == INVALID_HANDLE_VALUE)
    return;

  do {
    // Skip . and ..
    if (fd.cFileName[0] == '.' &&
        (fd.cFileName[1] == 0 ||
         (fd.cFileName[1] == '.' && fd.cFileName[2] == 0)))
      continue;

    if (fd.dwFileAttributes & FILE_ATTRIBUTE_REPARSE_POINT)
      continue;

    size_t name_len = strlen(fd.cFileName);
    path_builder_set_filename(&pb, fd.cFileName, name_len);

    if (fd.dwFileAttributes & FILE_ATTRIBUTE_DIRECTORY) {
      char *dir = path_builder_dup_arena(&pb, ctx->path_arena, false);
      mpmc_push_work(ctx->dir_queue, dir);
    } else {
      atomic_fetch_add(&g_files_found, 1);

      FileEntry *fe = arena_push(&ctx->meta_arena, sizeof(FileEntry), true);

      // Create a temporary copy for normalization to avoid corrupting pb.buffer
      char temp_path[MAX_PATHLEN];
      memcpy(temp_path, pb.buffer,
             (pb.filename_pos - pb.buffer) + name_len + 1);
      normalize_path(temp_path);

      fe->path = arena_push(&ctx->path_arena, strlen(temp_path) + 1, false);
      strcpy(fe->path, temp_path);

      platform_get_file_times(pb.buffer, &fe->created_time, &fe->modified_time);
      platform_get_file_owner(pb.buffer, fe->owner, sizeof(fe->owner));
      fe->size_bytes = ((uint64_t)fd.nFileSizeHigh << 32) | fd.nFileSizeLow;

      mpmc_push(ctx->file_queue, fe);
    }

  } while (FindNextFileA(h, &fd));

  FindClose(h);
}

#elif defined(__linux__)
static int platform_get_file_times_fd(int dir_fd, const char *name,
                                      uint64_t *created, uint64_t *modified) {
  struct stat st;
  if (fstatat(dir_fd, name, &st, 0) == 0) {
    *created = st.st_ctime; // or st.st_birthtime on systems that support it
    *modified = st.st_mtime;
    return 0;
  }
  return -1;
}

static int platform_get_file_owner_fd(int dir_fd, const char *name, char *owner,
                                      size_t owner_size) {
  struct stat st;
  if (fstatat(dir_fd, name, &st, 0) == 0) {
    struct passwd pw;
    struct passwd *result;
    char buffer[4096]; // Sufficiently large buffer for passwd data

    // Reentrant version (thread-safe)
    if (getpwuid_r(st.st_uid, &pw, buffer, sizeof(buffer), &result) == 0 &&
        result != NULL && result->pw_name != NULL) {
      strncpy(owner, result->pw_name, owner_size - 1);
      owner[owner_size - 1] = '\0';
    } else {
      // Fallback to uid
      snprintf(owner, owner_size, "uid:%d", st.st_uid);
    }
    return 0;
  }
  return -1;
}

void scan_folder(const char *base, ScannerContext *ctx) {
  PathBuilder pb;
  path_builder_init(&pb, base);

  int dir_fd = open(base, O_RDONLY | O_DIRECTORY | O_NOFOLLOW);
  if (dir_fd == -1)
    return;

  DIR *dir = fdopendir(dir_fd);
  if (!dir) {
    close(dir_fd);
    return;
  }

  struct dirent *entry;

  while ((entry = readdir(dir)) != NULL) {
    if (entry->d_name[0] == '.' &&
        (entry->d_name[1] == 0 ||
         (entry->d_name[1] == '.' && entry->d_name[2] == 0)))
      continue;

    size_t name_len = strlen(entry->d_name);
    path_builder_set_filename(&pb, entry->d_name, name_len);

    int file_type = DT_UNKNOWN;
#ifdef _DIRENT_HAVE_D_TYPE
    file_type = entry->d_type;
#endif

    // Fast path using d_type
    if (file_type != DT_UNKNOWN) {
      if (file_type == DT_LNK)
        continue; // Skip symlinks

      if (file_type == DT_DIR) {
        char *dir_path = path_builder_dup_arena(&pb, ctx->path_arena, false);
        mpmc_push_work(ctx->dir_queue, dir_path);
        continue;
      }

      if (file_type == DT_REG) {
        atomic_fetch_add(&g_files_found, 1);
        FileEntry *fe = arena_push(&ctx->meta_arena, sizeof(FileEntry), true);

        // Use fstatat for file info
        struct stat st;
        if (fstatat(dir_fd, entry->d_name, &st, 0) == 0) {
          // Convert times using fd variant
          platform_get_file_times_fd(dir_fd, entry->d_name, &fe->created_time,
                                     &fe->modified_time);
          platform_get_file_owner_fd(dir_fd, entry->d_name, fe->owner,
                                     sizeof(fe->owner));
          fe->size_bytes = (uint64_t)st.st_size;

          // Normalize path
          char temp_path[MAX_PATHLEN];
          memcpy(temp_path, pb.buffer,
                 (pb.filename_pos - pb.buffer) + name_len + 1);
          normalize_path(temp_path);

          fe->path = arena_push(&ctx->path_arena, strlen(temp_path) + 1, false);
          strcpy(fe->path, temp_path);

          mpmc_push(ctx->file_queue, fe);
        }
        continue;
      }
    }

    // Fallback for unknown types
    struct stat st;
    if (fstatat(dir_fd, entry->d_name, &st, AT_SYMLINK_NOFOLLOW) == 0) {
      if (S_ISLNK(st.st_mode))
        continue;

      if (S_ISDIR(st.st_mode)) {
        char *dir_path = path_builder_dup_arena(&pb, ctx->path_arena, false);
        mpmc_push_work(ctx->dir_queue, dir_path);
      } else if (S_ISREG(st.st_mode)) {
        atomic_fetch_add(&g_files_found, 1);
        FileEntry *fe = arena_push(&ctx->meta_arena, sizeof(FileEntry), true);

        platform_get_file_times(pb.buffer, &fe->created_time,
                                &fe->modified_time);
        platform_get_file_owner(pb.buffer, fe->owner, sizeof(fe->owner));
        fe->size_bytes = (uint64_t)st.st_size;

        char temp_path[MAX_PATHLEN];
        memcpy(temp_path, pb.buffer,
               (pb.filename_pos - pb.buffer) + name_len + 1);
        normalize_path(temp_path);

        fe->path = arena_push(&ctx->path_arena, strlen(temp_path) + 1, false);
        strcpy(fe->path, temp_path);

        mpmc_push(ctx->file_queue, fe);
      }
    }
  }

  closedir(dir); // Closes dir_fd automatically
}

#endif

// ------------------------- Scan worker --------------------------------
static THREAD_RETURN scan_worker(void *arg) {
  ScannerContext *ctx = (ScannerContext *)arg;

  for (;;) {
    char *dir = mpmc_pop(ctx->dir_queue);
    if (!dir)
      break;

    scan_folder(dir, ctx);

    mpmc_task_done(ctx->dir_queue, ctx->num_threads);
  }

  return THREAD_RETURN_VALUE;
}

// ----------------------------- Hashing helpers -----------------------------
static void xxh3_hash_file_stream(const char *path, char *out_hex,
                                  unsigned char *buf) {
  XXH128_hash_t h;
  XXH3_state_t state;
  XXH3_128bits_reset(&state);

  FileHandle handle = os_file_open(path, FLAG_SEQUENTIAL_READ);
  if (handle == INVALID_FILE_HANDLE) {
    strcpy(out_hex, "ERROR");
    return;
  }

  uint64_t bytes_read;
  while (os_file_read(handle, buf, READ_BLOCK, &bytes_read) == 0 &&
         bytes_read > 0) {
    XXH3_128bits_update(&state, buf, (size_t)bytes_read);
    atomic_fetch_add(&g_bytes_processed, bytes_read);
  }

  os_file_close(handle);

  h = XXH3_128bits_digest(&state);
  snprintf(out_hex, HASH_STRLEN, "%016llx%016llx", (unsigned long long)h.high64,
           (unsigned long long)h.low64);
}

// ------------------------- Hash worker --------------------------------
static THREAD_RETURN hash_worker(void *arg) {
  WorkerContext *ctx = (WorkerContext *)arg;
  unsigned char *buf = (unsigned char *)malloc(READ_BLOCK);

  for (;;) {
    FileEntry *fe = mpmc_pop(ctx->file_queue);
    if (!fe)
      break;

    char hash[HASH_STRLEN];
    xxh3_hash_file_stream(fe->path, hash, buf);

    char created[32], modified[32];
    format_time(fe->created_time, created, sizeof(created));
    format_time(fe->modified_time, modified, sizeof(modified));

    double size_kib = (double)fe->size_bytes / 1024.0;

    char stack_buf[1024];

    int len =
        snprintf(stack_buf, sizeof(stack_buf), "%s\t%s\t%.2f\t%s\t%s\t%s\n",
                 hash, fe->path, size_kib, created, modified, fe->owner);

    char *dst = arena_push(&ctx->arena, len, false);
    memcpy(dst, stack_buf, len);

    atomic_fetch_add(&g_files_hashed, 1);
  }

  free(buf);

  return THREAD_RETURN_VALUE;
}

// ------------------------- Progress display ---------------------------
static THREAD_RETURN progress_thread(void *arg) {
  (void)arg;

  HiResTimer progress_timer;
  timer_start(&progress_timer);

  uint64_t last_bytes = 0;
  double last_time = 0.0;
  double displayed_speed = 0.0;
  const double sample_interval = 0.5;

  // Hide cursor to prevent flickering
  printf("\033[?25l");

  for (;;) {
    uint64_t found = atomic_load(&g_files_found);
    uint64_t hashed = atomic_load(&g_files_hashed);
    uint64_t bytes = atomic_load(&g_bytes_processed);
    int scan_done = atomic_load(&g_scan_done);

    double t = timer_elapsed(&progress_timer);
    double dt = t - last_time;

    if (dt >= sample_interval) {
      uint64_t db = (bytes > last_bytes) ? bytes - last_bytes : 0;
      displayed_speed = (double)db / (1024.0 * 1024.0) / dt;
      last_bytes = bytes;
      last_time = t;
    }

    printf("\r");

    if (!scan_done) {
      printf("\033[1mScanning:\033[0m %llu files | Hashed: %llu | \033[32m%.2f "
             "MB/s\033[0m    ",
             (unsigned long long)found, (unsigned long long)hashed,
             displayed_speed);
    } else {
      double pct = found ? (double)hashed / (double)found : 0.0;
      int barw = 40;
      int filled = (int)(pct * barw);

      printf("[");
      // Print filled part in Green (\033[32m)
      printf("\033[32m");
      for (int i = 0; i < filled; i++)
        putchar('#');
      // Reset color for empty part
      printf("\033[0m");
      for (int i = filled; i < barw; i++)
        putchar('.');

      printf("] %6.2f%% (%llu/%llu) \033[32m%.2f MB/s\033[0m    ", pct * 100.0,
             (unsigned long long)hashed, (unsigned long long)found,
             displayed_speed);
    }

    fflush(stdout);

    if (scan_done && hashed == found)
      break;
    sleep_ms(100);
  }

  // Restore cursor (\033[?25h) and move to next line
  printf("\033[?25h\n");

  return THREAD_RETURN_VALUE;
}

// ======================== Hash worker IO Ring ========================
// -------------------------- Configuration ---------------------------
// #define IORING_BUFFER_SIZE (KiB(32))
#define IORING_BUFFER_SIZE (KiB(256))
#define NUM_BUFFERS_PER_THREAD 32
#define MAX_ACTIVE_FILES 1
#define SUBMIT_TIMEOUT_MS 30000

// Globals
u64 g_ioring_buffer_size = 4096 * 64;
static atomic_uint_fast64_t g_io_ring_fallbacks = 0;

// -------------------------- Data structures ---------------------------

#if defined(_WIN32) || defined(_WIN64)
// Windows I/O Ring types
typedef HIORING AsyncIoRing;
typedef HIORING AsyncIoHandle;
#define INVALID_ASYNC_IO_HANDLE NULL
#define SUBMIT_READ_RETURN_VALUE HRESULT

#elif defined(__linux__)
// Linux io_uring types
typedef struct {
  struct io_uring ring;
  int event_fd;
  struct io_uring_cqe *cqe_cache;
  int cqe_cache_index;
  int cqe_cache_count;
} AsyncIoRingImpl;

typedef AsyncIoRingImpl *AsyncIoRing;
typedef int AsyncIoHandle;
typedef struct iovec IORING_BUFFER_INFO;
#define INVALID_ASYNC_IO_HANDLE (-1)
#define SUBMIT_READ_RETURN_VALUE int

typedef struct {
  int ResultCode;
  uint32_t Information;
  uintptr_t UserData;
} AsyncIoCompletion;

#endif

typedef struct IoBuffer IoBuffer;

typedef struct FileReadContext {
  FileHandle hFile;
  uint64_t file_size;
  int use_incremental_hash;
  union {
    XXH3_state_t hash_state;   // For incremental hash (large files)
    XXH128_hash_t single_hash; // For single-shot hash (small files)
  };
  FileEntry *fe;

  // Completion tracking
  int reads_submitted;
  int reads_completed;

  int active_reads;
  int failed_reads;

  int reads_hashed;
  uint64_t bytes_hashed;

  // For in-order hashing
  uint64_t next_hash_offset;
  uint64_t next_read_offset;

  IoBuffer *head;
  IoBuffer *tail;

} FileReadContext;

// -------------------------- Buffer structure ---------------------------
typedef struct IoBuffer {
  void *data;
  uint64_t offset;
  size_t size;
  size_t bytes_read;
  SUBMIT_READ_RETURN_VALUE result;

  int buffer_id;
  int completed;

  FileReadContext *file;

  struct IoBuffer *next;
} IoBuffer;

// Thread-local I/O Ring context
typedef struct ThreadIoContext {
  AsyncIoRing ring;
  void *completion_event;
  unsigned char *fallback_buffer;
  IoBuffer buffers[NUM_BUFFERS_PER_THREAD];
  int buffer_pool[NUM_BUFFERS_PER_THREAD];
  int free_count;
  int submitting;
  int num_submissions;
  int active_files;

  int use_registered_buffers;
} ThreadIoContext;

typedef struct {
  uint32_t MaxSubmissionQueueSize;
  uint32_t MaxCompletionQueueSize;
  uint32_t MaxVersion;
} AsyncIoCapabilities;

// ----------------------------- Async I/O Abstraction -------------------------
#if defined(_WIN32) || defined(_WIN64)

// Windows I/O Ring functions
static void async_io_query_capabilities(AsyncIoCapabilities *caps) {
  IORING_CAPABILITIES win_caps;
  QueryIoRingCapabilities(&win_caps);
  caps->MaxSubmissionQueueSize = win_caps.MaxSubmissionQueueSize;
  caps->MaxCompletionQueueSize = win_caps.MaxCompletionQueueSize;
  caps->MaxVersion = win_caps.MaxVersion;
}

static void *async_io_create_completion_event(void) {
  return CreateEvent(NULL, FALSE, FALSE, NULL);
}

static void async_io_set_completion_event(AsyncIoRing ring, void *event) {
  SetIoRingCompletionEvent(ring, event);
}

static void async_io_wait_for_completion(ThreadIoContext *ctx) {
  if (ctx->num_submissions > 0) {
    WaitForSingleObject(ctx->completion_event, SUBMIT_TIMEOUT_MS);
    return;
  }
}

static int async_io_create_ring(ThreadIoContext *thread_ctx,
                                uint32_t queue_size) {
  IORING_CREATE_FLAGS flags = {0};
  HRESULT hr = CreateIoRing(IORING_VERSION_3, flags, queue_size, queue_size * 2,
                            &thread_ctx->ring);

  // Create completion event
  thread_ctx->completion_event = async_io_create_completion_event();
  if (thread_ctx->completion_event) {
    async_io_set_completion_event(thread_ctx->ring,
                                  thread_ctx->completion_event);
  }
  return SUCCEEDED(hr) ? 0 : -1;
}

#define USERDATA_REGISTER 1

#define MAKE_BUF_INFO(a, l)                                                    \
  (IORING_BUFFER_INFO) { .Address = (a), .Length = (uint32_t)(l) }

static int async_io_submit(ThreadIoContext *thread_ctx, uint32_t wait_count,
                           uint32_t timeout_ms, uint32_t *submitted) {
  HRESULT hr = SubmitIoRing(thread_ctx->ring, 0, timeout_ms, submitted);
  // HRESULT hr = SubmitIoRing(ring, wait_count, timeout_ms, submitted);

  // The wait_count in windows is not implemented yet, so we wait with a
  // completion event for a single completion
  async_io_wait_for_completion(thread_ctx);

  return SUCCEEDED(hr) ? 0 : -1;
}

static int async_io_register_buffers(ThreadIoContext *thread_ctx,
                                     uint32_t num_buffers,
                                     IORING_BUFFER_INFO *buf_info) {

  HRESULT hr = BuildIoRingRegisterBuffers(
      thread_ctx->ring, NUM_BUFFERS_PER_THREAD, buf_info, USERDATA_REGISTER);
  if (FAILED(hr)) {
    LPSTR messageBuffer = NULL;

    size_t size = FormatMessageA(
        FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM |
            FORMAT_MESSAGE_IGNORE_INSERTS,
        NULL, hr, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT),
        (LPSTR)&messageBuffer, 0, NULL);

    if (size > 0) {
      fprintf(stderr, "Error registering buffers: %s (0x%08X)\n", messageBuffer,
              (unsigned int)hr);
      LocalFree(messageBuffer); // Free the memory allocated by FormatMessage
    } else {
      fprintf(stderr, "Error registering buffers: Unknown HRESULT (0x%08X)\n",
              (unsigned int)hr);
    }
  }
  // Submit registration
  async_io_submit(thread_ctx, 0, 0, NULL);

  return hr;
}

static void async_io_close_event(void *event) { CloseHandle(event); }

static int async_io_close_ring(ThreadIoContext *thread_ctx) {

  if (thread_ctx->completion_event)
    async_io_close_event(thread_ctx->completion_event);
  CloseIoRing(thread_ctx->ring);
  return 0;
}

static SUBMIT_READ_RETURN_VALUE
async_io_build_read(ThreadIoContext *thread_ctx, AsyncIoHandle file_handle,
                    uint32_t buffer_id, size_t size, uint64_t offset,
                    uintptr_t user_data) {
  IORING_HANDLE_REF file_ref = IoRingHandleRefFromHandle(file_handle);
  IORING_BUFFER_REF buffer_ref =
      IoRingBufferRefFromIndexAndOffset(buffer_id, 0);
  HRESULT hr =
      BuildIoRingReadFile(thread_ctx->ring, file_ref, buffer_ref,
                          (uint32_t)size, offset, user_data, IOSQE_FLAGS_NONE);
  return hr;
}

typedef struct {
  HRESULT ResultCode;
  uint32_t Information;
  uintptr_t UserData;
} AsyncIoCompletion;

static int async_io_pop_completion(AsyncIoRing ring, AsyncIoCompletion *cqe) {
  IORING_CQE win_cqe;

  while (1) {
    HRESULT hr = PopIoRingCompletion(ring, &win_cqe);

    if (hr == S_FALSE)
      return 0;

    if (FAILED(hr))
      return -1;

    // Unlike linux, in addition of IO operations, Windows IO Ring produces CQEs
    // (completion queue entries) when doing operations like register buffer or
    // submiting, we filter them here cqe.UserData == USERDATA_REGISTER
    // cqe.ResultCode == S_OK (or error)
    if (win_cqe.UserData == USERDATA_REGISTER)
      continue;

    cqe->ResultCode = win_cqe.ResultCode;
    cqe->Information = win_cqe.Information;
    cqe->UserData = win_cqe.UserData;

    // Check for error and print warning
    if (FAILED(win_cqe.ResultCode)) {
      char error_msg[256];
      FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
                     NULL, win_cqe.ResultCode, 0, error_msg, sizeof(error_msg),
                     NULL);
      fprintf(stderr,
              "WARNING: I/O completion error - Code: 0x%lx, Error: %s\n",
              win_cqe.ResultCode, error_msg);
    }

    return 1;
  }
}

#elif defined(__linux__)
// Linux io_uring functions implementation
static void async_io_query_capabilities(AsyncIoCapabilities *caps) {
  // Get system limits for io_uring
  long max_entries = sysconf(_SC_IOV_MAX);
  if (max_entries <= 0)
    max_entries = 4096;

  caps->MaxSubmissionQueueSize =
      (uint32_t)(max_entries < 4096 ? max_entries : 4096);
  caps->MaxCompletionQueueSize = caps->MaxSubmissionQueueSize * 2;
  caps->MaxVersion = 1;
}

// static int async_io_create_ring(uint32_t queue_size, AsyncIoRing *ring) {
static int async_io_create_ring(ThreadIoContext *thread_ctx,
                                uint32_t queue_size) {
  AsyncIoRingImpl *impl = (AsyncIoRingImpl *)calloc(1, sizeof(AsyncIoRingImpl));
  if (!impl)
    return -1;

  // Initialize io_uring
  struct io_uring_params params = {0};

  params.flags =
      IORING_SETUP_CQSIZE |
      IORING_SETUP_SINGLE_ISSUER | // Thread local io_uring
      IORING_SETUP_DEFER_TASKRUN;  // Do not send interupts when a CQE is ready,
                                   // send them when a wait function is called,
                                   // and groupe them acording to the nbre or
                                   // entries to wait (reduces syscall overhead)

  params.cq_entries = queue_size * 2;

  int ret = io_uring_queue_init_params(queue_size, &impl->ring, &params);
  if (ret < 0) {
    // Fallback to without params
    printf("WARNING: Creating io_uring with default params\n");
    ret = io_uring_queue_init(queue_size, &impl->ring, 0);
    if (ret < 0) {
      free(impl);
      return -1;
    }
  }

  impl->cqe_cache = NULL;
  impl->cqe_cache_index = 0;
  impl->cqe_cache_count = 0;

  thread_ctx->ring = impl;
  return 0;
}

#define MAKE_BUF_INFO(a, l)                                                    \
  (IORING_BUFFER_INFO) { .iov_base = (a), .iov_len = (size_t)(l) }

static int async_io_register_buffers(ThreadIoContext *thread_ctx,
                                     uint32_t num_buffers,
                                     IORING_BUFFER_INFO *buf_info) {
  AsyncIoRingImpl *impl = (AsyncIoRingImpl *)thread_ctx->ring;

  int hr = io_uring_register_buffers(&impl->ring, buf_info, num_buffers);

  if (hr < 0) {
    fprintf(stderr, "Error registering buffers: %s (code: %d)\n", strerror(-hr),
            hr);
    fprintf(stderr, "WARNING: Memlock limit too low buffer size! Fallback to "
                    "unregistred buffers\n");
  }
  return hr == 0 ? 0 : -1;
}

static int async_io_submit(ThreadIoContext *thread_ctx, uint32_t wait_count,
                           uint32_t timeout_ms, uint32_t *submitted) {
  AsyncIoRingImpl *impl = (AsyncIoRingImpl *)thread_ctx->ring;
  if (!impl)
    return -1;

  int ret;

  if (wait_count > 0) {
    ret = io_uring_submit_and_wait(&impl->ring, wait_count);
  } else {
    ret = io_uring_submit(&impl->ring);
  }

  if (ret < 0) {
    fprintf(stderr, "submit error: %s\n", strerror(-ret));
    return -1;
  }

  if (submitted)
    *submitted = (uint32_t)ret;

  return 0;
}

static int async_io_close_ring(ThreadIoContext *thread_ctx) {
  AsyncIoRingImpl *impl = (AsyncIoRingImpl *)thread_ctx->ring;
  if (!impl)
    return -1;

  if (thread_ctx->use_registered_buffers) {
    io_uring_unregister_buffers(&impl->ring);
  }
  close(impl->event_fd);
  io_uring_queue_exit(&impl->ring);
  free(impl);

  return 0;
}

static int async_io_build_read(ThreadIoContext *thread_ctx,
                               AsyncIoHandle file_handle, uint32_t buffer_id,
                               size_t size, uint64_t offset,
                               uintptr_t user_data) {
  AsyncIoRing ring = thread_ctx->ring;
  AsyncIoRingImpl *impl = (AsyncIoRingImpl *)ring;
  if (!impl)
    return -1;

  struct io_uring_sqe *sqe = io_uring_get_sqe(&impl->ring);
  if (!sqe) {
    printf("SQE FULL\n");
    return -1;
  }

  ThreadIoContext *ctx = thread_ctx; // or pass it properly TODO : look here

  void *buf = ctx->buffers[buffer_id].data;

  if (thread_ctx->use_registered_buffers) {
    io_uring_prep_read_fixed(sqe, file_handle, buf, size, offset, buffer_id);
  } else {
    io_uring_prep_read(sqe, file_handle, buf, size, offset);
  }

  io_uring_sqe_set_data64(sqe, user_data);
  return 0;
}

static int async_io_pop_completion(AsyncIoRing ring, AsyncIoCompletion *cqe) {
  AsyncIoRingImpl *impl = (AsyncIoRingImpl *)ring;

  struct io_uring_cqe *cqe_ptr = NULL;

  int ret = io_uring_peek_cqe(&impl->ring, &cqe_ptr);

  if (ret == -EAGAIN) {
    // No CQE available
    return 0;
  }

  if (ret < 0) {
    // Error
    fprintf(stderr, "io_uring_peek_cqe error: %d (%s)\n", ret, strerror(-ret));
    return -1;
  }

  if (!cqe_ptr) {
    return 0;
  }

  int res = cqe_ptr->res;

  if (res >= 0) {
    cqe->ResultCode = 0;
    cqe->Information = (uint32_t)res;
  } else {
    cqe->ResultCode = res;
    cqe->Information = 0;
  }

  cqe->UserData = (uintptr_t)cqe_ptr->user_data;

  io_uring_cqe_seen(&impl->ring, cqe_ptr);

  // Check for error and print warning
  if (res < 0) {
    fprintf(stderr, "WARNING: I/O completion error - Code: %d, Error: %s\n",
            res, strerror(-res));
  }

  return 1;
}

#endif

// OS-agnostic helper macros
#define ASYNC_IO_SUCCEEDED(result) ((result) >= 0)
#define ASYNC_IO_FAILED(result) ((result) < 0)

// ---------------------- FIFO queue operations ---------------------------
typedef struct FileQueue {
  FileReadContext files[MAX_ACTIVE_FILES];
  int head;
  int tail;
  int count;
} FileQueue;

static FileReadContext *fq_push(FileQueue *fq) {
  if (fq->count == MAX_ACTIVE_FILES)
    return NULL;

  FileReadContext *f = &fq->files[fq->tail];
  fq->tail = (fq->tail + 1) % MAX_ACTIVE_FILES;
  fq->count++;
  return f;
}

static FileReadContext *fq_peek(FileQueue *fq) {
  if (fq->count == 0)
    return NULL;
  return &fq->files[fq->head];
}

static void fq_pop(FileQueue *fq) {
  fq->head = (fq->head + 1) % MAX_ACTIVE_FILES;
  fq->count--;
}

static void fq_remove_at(FileQueue *fq, int index) {
  if (fq->count == 0)
    return;

  int remove_idx = (fq->head + index) % MAX_ACTIVE_FILES;

  int last_logical = fq->count - 1;
  int last_idx = (fq->head + last_logical) % MAX_ACTIVE_FILES;

  // Swap with last logical element if needed
  if (index != last_logical) {
    fq->files[remove_idx] = fq->files[last_idx];
  }

  // Just decrease count
  fq->count--;

  // Recompute tail properly
  fq->tail = (fq->head + fq->count) % MAX_ACTIVE_FILES;
}

// ---------------------- Initialize thread context ---------------------------
static ThreadIoContext *io_ring_init_thread(void) {
  ThreadIoContext *thread_ctx =
      (ThreadIoContext *)calloc(1, sizeof(ThreadIoContext));
  if (!thread_ctx)
    return NULL;

  // Query I/O Ring capabilities
  AsyncIoCapabilities caps;
  async_io_query_capabilities(&caps);

  uint32_t queue_size = caps.MaxSubmissionQueueSize;
  if (queue_size > 4096)
    queue_size = 4096; // Cap at 4096 for reasonable memory usage

  // Create I/O Ring
  if (async_io_create_ring(thread_ctx, queue_size) != 0) {
    free(thread_ctx);
    thread_ctx = NULL;
    return NULL;
  }

  // Initialize buffer pool
  thread_ctx->fallback_buffer = (unsigned char *)malloc(READ_BLOCK);

  IORING_BUFFER_INFO buf_info[NUM_BUFFERS_PER_THREAD];

  u64 buf_pool_size = g_ioring_buffer_size * NUM_BUFFERS_PER_THREAD;

  // Reserve and Commit memory for buffers
  void *base_ptr = plat_mem_reserve(buf_pool_size);
  if (base_ptr) {
    if (!plat_mem_commit(base_ptr, buf_pool_size)) {
      plat_mem_release(base_ptr, 0);
      async_io_close_ring(thread_ctx);
      free(thread_ctx);
      thread_ctx = NULL;
      return NULL;
    }
  } else {

    async_io_close_ring(thread_ctx);
    free(thread_ctx);
    thread_ctx = NULL;
    return NULL;
  }

  for (int i = 0; i < NUM_BUFFERS_PER_THREAD; i++) {
    thread_ctx->buffers[i].data = (u8 *)base_ptr + (i * g_ioring_buffer_size);
    thread_ctx->buffer_pool[i] = i;
    thread_ctx->buffers[i].buffer_id = i;

    buf_info[i] =
        MAKE_BUF_INFO(thread_ctx->buffers[i].data, g_ioring_buffer_size);
  }

  thread_ctx->free_count = NUM_BUFFERS_PER_THREAD;

  // Register buffers
  int hr =
      async_io_register_buffers(thread_ctx, NUM_BUFFERS_PER_THREAD, buf_info);

  thread_ctx->use_registered_buffers = (hr == 0);
  thread_ctx->submitting = 1;
  thread_ctx->num_submissions = 0;
  thread_ctx->active_files = 0;

  return thread_ctx;
}

static void io_ring_cleanup_thread(ThreadIoContext *thread_ctx) {
  if (!thread_ctx)
    return;

  if (thread_ctx->ring)
    async_io_close_ring(thread_ctx);

  // Free the buffer pool memory
  if (thread_ctx->buffers[0].data) {
    u64 buf_pool_size = g_ioring_buffer_size * NUM_BUFFERS_PER_THREAD;
    plat_mem_release(thread_ctx->buffers[0].data, buf_pool_size);
  }

  free(thread_ctx);
  thread_ctx = NULL;
}

// -------------------------- Buffer get and return ------------------------
static IoBuffer *get_free_buffer(ThreadIoContext *ctx) {

  if (ctx->free_count == 0) {
    return NULL;
  }

  int idx = ctx->buffer_pool[--ctx->free_count];
  IoBuffer *buf = &ctx->buffers[idx];
  buf->completed = 0;
  buf->bytes_read = 0;
  buf->result = 0;

  return buf;
}

static void return_buffer(ThreadIoContext *ctx, IoBuffer *buf) {
  if (!buf)
    return;

  ctx->buffer_pool[ctx->free_count++] = buf->buffer_id;
}

// -------------------------- Submit async read ---------------------------
static int build_read(ThreadIoContext *thread_ctx, FileReadContext *file_ctx,
                      IoBuffer *buf, uint64_t offset, size_t size) {
  buf->offset = offset;
  buf->size = size;
  buf->file = file_ctx;

  SUBMIT_READ_RETURN_VALUE result =
      async_io_build_read(thread_ctx, file_ctx->hFile, buf->buffer_id, size,
                          offset, (uintptr_t)buf);

  if (ASYNC_IO_SUCCEEDED(result)) {
    file_ctx->active_reads++;
    file_ctx->reads_submitted++;
    thread_ctx->num_submissions++;
  } else {
    buf->completed = 1;
    buf->result = result; // Store the error code
    return_buffer(thread_ctx, buf);
  }
  return result;
}

// ------------ Link completed buffers in an ordered list -------------
static void insert_buffer_ordered(FileReadContext *file, IoBuffer *buf) {
  buf->next = NULL;

  // empty list
  if (!file->head) {
    file->head = file->tail = buf;
    return;
  }

  // insert at head
  if (buf->offset < file->head->offset) {
    buf->next = file->head;
    file->head = buf;
    return;
  }

  // find position
  IoBuffer *cur = file->head;

  while (cur->next && cur->next->offset < buf->offset) {
    cur = cur->next;
  }

  buf->next = cur->next;
  cur->next = buf;

  if (!buf->next)
    file->tail = buf;
}

// -------------------------- Process completions ---------------------------
static void process_completions(ThreadIoContext *thread_ctx, FileQueue *fq) {
  AsyncIoCompletion cqe;

  // Keep processing as long as there are completions available
  while (async_io_pop_completion(thread_ctx->ring, &cqe) == 1) {

    IoBuffer *buf = (IoBuffer *)cqe.UserData;
    FileReadContext *file = buf->file;

    buf->result = cqe.ResultCode;
    buf->bytes_read = cqe.Information;
    buf->completed = 1;

    file->active_reads--;
    file->reads_completed++;
    thread_ctx->num_submissions--;

    if (ASYNC_IO_SUCCEEDED(cqe.ResultCode) && cqe.Information > 0) {

      buf->next = NULL;

      insert_buffer_ordered(file, buf);

    } else {
      file->failed_reads++;
      return_buffer(thread_ctx, buf);
    }
  }
}

// -------------------- File operations -----------------------
static int init_file(FileReadContext *f, FileEntry *fe) {
  memset(f, 0, sizeof(*f));

  f->fe = fe;
  f->file_size = fe->size_bytes;

  // Use the abstracted os_file_open_async for async I/O with no buffering
  f->hFile = os_file_open(fe->path, FLAG_ASYNC_DIRECT_READ);

  if (f->hFile == INVALID_ASYNC_IO_HANDLE) {
    return 0;
  }

  // Determine hash method based on file size
  if (f->file_size > g_ioring_buffer_size) {
    f->use_incremental_hash = true;
    XXH3_128bits_reset(&f->hash_state);
  } else {
    f->use_incremental_hash = false;
  }
  return 1;
}

static void finalize_file(FileReadContext *file, ThreadIoContext *thread_ctx,
                          WorkerContext *worker_ctx) {

  FileEntry *fe = file->fe;

  char hash[HASH_STRLEN];

  if (file->failed_reads == 0 && file->bytes_hashed == file->file_size) {
    if (file->use_incremental_hash) {
      // Large file: digest the accumulated hash state
      XXH128_hash_t h = XXH3_128bits_digest(&file->hash_state);
      snprintf(hash, HASH_STRLEN, "%016llx%016llx",
               (unsigned long long)h.high64, (unsigned long long)h.low64);
    } else {
      // Small file: hash already computed, stored directly in single_hash
      snprintf(hash, HASH_STRLEN, "%016llx%016llx",
               (unsigned long long)file->single_hash.high64,
               (unsigned long long)file->single_hash.low64);
    }
  } else {
    atomic_fetch_add(&g_io_ring_fallbacks, 1);
    xxh3_hash_file_stream(fe->path, hash, thread_ctx->fallback_buffer);
    // DEBUG
    // printf("Fallback for path: %s\n", fe->path);
  }

  char created[32], modified[32];
  format_time(fe->created_time, created, sizeof(created));
  format_time(fe->modified_time, modified, sizeof(modified));

  double size_kib = (double)fe->size_bytes / 1024.0;

  char stack_buf[KiB(4)];
  int len = snprintf(stack_buf, sizeof(stack_buf), "%s\t%s\t%.2f\t%s\t%s\t%s\n",
                     hash, fe->path, size_kib, created, modified, fe->owner);

  char *dst = arena_push(&worker_ctx->arena, len, false);
  memcpy(dst, stack_buf, len);

  atomic_fetch_add(&g_files_hashed, 1);
}

// -------------------- Hash head file -----------------------
static void hash_head_file(ThreadIoContext *thread_ctx, FileQueue *fq,
                           WorkerContext *worker_ctx) {

  FileReadContext *file = fq_peek(fq);
  if (!file)
    return;

  // Keep hashing while the next buffer in sequence is ready at head
  while (file->head && file->head->offset == file->next_hash_offset) {

    IoBuffer *buf = file->head;

    // Consume from head
    file->head = buf->next;
    if (!file->head)
      file->tail = NULL;

    // Process the buffer
    if (ASYNC_IO_SUCCEEDED(buf->result) && buf->bytes_read > 0) {
      // Calculate actual bytes to hash (handle last partial sector)
      size_t bytes_to_hash = buf->bytes_read;

      // If this is the last buffer and we read beyond file size, trim it
      if (buf->offset + buf->bytes_read > file->file_size) {
        bytes_to_hash = file->file_size - buf->offset;
      }

      if (bytes_to_hash > 0) {
        if (file->use_incremental_hash) {
          // Large file: update incremental hash state
          XXH3_128bits_update(&file->hash_state, buf->data, bytes_to_hash);
        } else {
          // Small file: single-shot hash
          file->single_hash = XXH3_128bits(buf->data, bytes_to_hash);
        }

        file->bytes_hashed += bytes_to_hash;
        atomic_fetch_add(&g_bytes_processed, bytes_to_hash);
      }

      file->reads_hashed++;
    } else if (buf->bytes_read == 0 && ASYNC_IO_SUCCEEDED(buf->result)) {
      // Read operation completed with 0 bytes (EOF)
      file->reads_hashed++;
    } else {
      // Read failed
      file->failed_reads++;
      file->reads_hashed++;
    }

    // Move to next offset
    file->next_hash_offset += buf->size;

    // Return buffer to pool
    return_buffer(thread_ctx, buf);
  }

  // Finalize file when all reads are complete
  if (file->active_reads == 0 && file->bytes_hashed >= file->file_size) {
    finalize_file(file, thread_ctx, worker_ctx);
    os_file_close(file->hFile);
    fq_pop(fq);
    thread_ctx->active_files--;
  }
}

static void hash_ready_files(ThreadIoContext *thread_ctx, FileQueue *fq,
                             WorkerContext *worker_ctx) {
  for (int i = 0; i < fq->count;) {

    int idx = (fq->head + i) % MAX_ACTIVE_FILES;
    FileReadContext *file = &fq->files[idx];

    bool progressed = false;

    // ---- HASH READY BUFFERS ----
    while (file->head) {

      IoBuffer *buf = file->head;

      if (buf->offset != file->bytes_hashed)
        break;

      progressed = true;

      file->head = buf->next;
      if (!file->head)
        file->tail = NULL;

      size_t bytes_to_hash = buf->bytes_read;

      if (buf->offset + buf->bytes_read > file->file_size) {
        bytes_to_hash = file->file_size - buf->offset;
      }

      if (bytes_to_hash > 0) {
        if (file->use_incremental_hash) {
          XXH3_128bits_update(&file->hash_state, buf->data, bytes_to_hash);
        } else {
          file->single_hash = XXH3_128bits(buf->data, bytes_to_hash);
        }

        file->bytes_hashed += bytes_to_hash;
        atomic_fetch_add(&g_bytes_processed, bytes_to_hash);
      }

      return_buffer(thread_ctx, buf);
    }

    // ---- FINALIZE ----
    if (file->active_reads == 0 && file->bytes_hashed >= file->file_size) {

      finalize_file(file, thread_ctx, worker_ctx);
      os_file_close(file->hFile);

      fq_remove_at(fq, i);
      thread_ctx->active_files--;

      continue;
    }
    i++;
  }
}

// ------------- Submit pending reads - fill all free buffers -----------------
static void build_pending_reads(ThreadIoContext *thread_ctx, FileQueue *fq,
                                WorkerContext *worker_ctx) {

  MPMCQueue *file_queue = worker_ctx->file_queue;

  // Try to submit reads for the current head file
  FileReadContext *f = fq_peek(fq);

  for (;;) {

    if (f) {
      while (f->next_read_offset < f->file_size) {

        IoBuffer *buf = get_free_buffer(thread_ctx);
        if (!buf)
          return;

        size_t remaining = f->file_size - f->next_read_offset;
        size_t size;

        // Check if this is the last read and the file size is not
        // sector-aligned
        int is_last_read = (remaining <= g_ioring_buffer_size);

        if (remaining >= g_ioring_buffer_size) {
          // Normal full read
          size = g_ioring_buffer_size;
        } else {
          // Last read - handle partial sector
          if (remaining % g_pagesize != 0) {
            size = ALIGN_UP_POW2(remaining, g_pagesize);

          } else {
            size = remaining;
          }
        }

        SUBMIT_READ_RETURN_VALUE hr =
            build_read(thread_ctx, f, buf, f->next_read_offset, size);

        if (ASYNC_IO_FAILED(hr)) {
          return_buffer(thread_ctx, buf);
          f->failed_reads++;
          f->active_reads = 0;
          f->reads_submitted = 0;
          f->next_read_offset = f->file_size;
          break;
        }

        f->next_read_offset += size;
      }
    }

    // Add new file if possible
    if (!thread_ctx->submitting)
      return;

    if (thread_ctx->active_files >= MAX_ACTIVE_FILES)
      return;

    FileEntry *fe = mpmc_pop(file_queue);
    if (!fe) {
      thread_ctx->submitting = 0;
      return;
    }

    FileReadContext *newf = fq_push(fq);

    if (!init_file(newf, fe)) {
      // File can't be opened with NO_BUFFERING, process with fallback
      char hash[HASH_STRLEN];
      finalize_file(newf, thread_ctx, worker_ctx);
      fq_pop(fq);
      continue;
    }

    f = newf;
    thread_ctx->active_files++;
  }
}

// -------------------------- Hash worker I/O Ring ---------------------------
static THREAD_RETURN hash_worker_io_ring(void *arg) {
  WorkerContext *ctx = (WorkerContext *)arg;

  // Init IO ring
  ThreadIoContext *thread_ctx = io_ring_init_thread();
  if (!thread_ctx || !thread_ctx->ring) {
    printf("I/O Ring unavailable, using buffered I/O\n");
    return hash_worker(arg);
  }

  // Initialize pipeline state
  FileQueue fq;
  memset(&fq, 0, sizeof(fq));

  uint32_t submitted = 0;
  uint32_t wait_count;

  // Main pipeline loop
  for (;;) {

    // Submit new reads
    build_pending_reads(thread_ctx, &fq, ctx);

    wait_count = MIN(thread_ctx->num_submissions, NUM_BUFFERS_PER_THREAD - 2);

    submitted = 0;
    // async_io_submit(ring_ctx->ring, 0, 0, &submitted);
    async_io_submit(thread_ctx, wait_count, 0, &submitted);

    // Process completions
    process_completions(thread_ctx, &fq);

    // debug
    // printf("Free buffers: %d, Submissions: %d, Active files: %d\n",
    //        ring_ctx->free_count, ring_ctx->num_submissions,
    //        ring_ctx->active_files);
    // debug end

    // Hash files
    for (int i = 0; i < fq.count; i++) {
      hash_head_file(thread_ctx, &fq, ctx);
    }
    // hash_ready_files(ring_ctx, &fq, ctx);

    // Exit condition
    if (!thread_ctx->submitting && thread_ctx->active_files == 0 &&
        thread_ctx->num_submissions == 0) {
      break;
    }
  }

  io_ring_cleanup_thread(thread_ctx);
  return THREAD_RETURN_VALUE;
}