#pragma once

#include <atomic>
#include <fcntl.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <stdint.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <unistd.h>
#include <vector>
#include <chrono>
#include <string>

#include "spdlog/spdlog.h"


#define MAX_WORKERS 60
#define MAX_DEVICES 16

static inline std::string get_shm_name() {
  const char *env_shm_name = getenv("VLLM_VNPU_SHM_NAME");
  if (env_shm_name) {
    if (env_shm_name[0] != '/') {
      spdlog::error(
          "The shm name specified by VLLM_VNPU_SHM_NAME should start "
          "with '/'");
      exit(-1);
    }
    return std::string(env_shm_name);
  }
  return std::string("/vllm_acl_vnpu_offload_shm");
}

static constexpr uint32_t heartbeat_us = 1000; // microseconds
static constexpr uint32_t heartbeat_check_everyN = 50;
static constexpr uint32_t heartbeat_timeout_us =
    heartbeat_check_everyN * heartbeat_us;

struct alignas(64) WorkerHeartBeat {
  std::atomic<uint64_t> timestamp;
  std::atomic<int32_t> tgid;
  uint8_t _padding[64 - sizeof(std::atomic<uint64_t>) -
                   sizeof(std::atomic<int32_t>)];
};

static inline uint64_t heartbeat_ts_us() {
  return static_cast<uint64_t>(
      std::chrono::duration_cast<std::chrono::microseconds>(
          std::chrono::steady_clock::now().time_since_epoch())
          .count());
}

static inline uint32_t unpack_lock_field(uint64_t gpu_flag) {
  return static_cast<uint32_t>(gpu_flag >> 32);
}

static inline int32_t unpack_tgid_field(uint64_t gpu_flag) {
  return static_cast<int32_t>(gpu_flag & 0xFFFFFFFF);
}

static inline uint64_t pack_locked_tgid(int32_t tgid) {
  return (static_cast<uint64_t>(1) << 32) | static_cast<uint64_t>(tgid);
}

static inline uint64_t pack_unlocked_tgid(int32_t tgid) {
  return static_cast<uint64_t>(tgid);
}

// mmap usually page-aligned
struct alignas(64) ShmHelper {
  struct VramInfo {
    uint64_t total_vmem_size;
    uint64_t shareable_handle;
  };
  VramInfo vram_info[MAX_DEVICES]; // support max 16 NPUs
  // GPU lock flag
  std::atomic<uint64_t> gpu_flag[MAX_DEVICES];
  // uint8_t _padding1[64 - sizeof(std::atomic<uint64_t>)];

  // request
  enum RequestType: uint32_t {
    REQUEST_TYPE_REGISTER_WORKER = 1,
  };
  /* ready:
    * 0: worker store: no request & worker get response
    * 1: worker store: worker preparing request
    * 2: worker store: request ready for listener
    * 3: listener store: listener processed request
  */
  enum ReadyState : uint64_t {
    READY_STATE_NO_REQUEST = 0,
    READY_STATE_PREPARING_REQUEST = 1,
    READY_STATE_REQUEST_READY = 2,
    READY_STATE_REQUEST_PROCESSED = 3
  };
  std::atomic<uint64_t> req_ready;
  // currently only allow one parameter and one response
  struct {
    uint32_t type;
    int32_t tgid;
    uint64_t parameter;
    uint64_t response;
  } request;
  uint8_t _padding2[64 - sizeof(std::atomic<uint64_t>) - sizeof(request)];

  // heart beats
  WorkerHeartBeat heart_beats[MAX_WORKERS];

  void init() {
    memset(vram_info, 0, sizeof(vram_info));
    for (size_t i = 0; i < MAX_DEVICES; ++i) {
      gpu_flag[i].store(0, std::memory_order_release);
    }
    req_ready.store(READY_STATE_NO_REQUEST, std::memory_order_release);
  }

  void set_gpu_info(int gpu_id, uint64_t vmem_size, uint64_t shared_handle) {
    vram_info[gpu_id].total_vmem_size = vmem_size;
    vram_info[gpu_id].shareable_handle = shared_handle;
  }
};

static constexpr size_t SHM_SIZE = (sizeof(ShmHelper) + 4095) & ~4095;