#pragma once
#include <stdint.h>

#ifdef __cplusplus
extern "C" {
#endif

#define BBR_SCALE  8
#define BBR_UNIT   (1 << BBR_SCALE)
#define BW_SCALE   24
#define BW_UNIT    (1 << BW_SCALE)

#define BBR_STARTUP_PACING_GAIN  (BBR_UNIT * 277 / 100 + 1)
#define BBR_STARTUP_CWND_GAIN    (BBR_UNIT * 2)
#define BBR_DRAIN_GAIN           (BBR_UNIT * 1000 / 2885)
#define BBR_CWND_GAIN            (BBR_UNIT * 2)
#define BBR_PROBE_RTT_CWND_GAIN  (BBR_UNIT * 1 / 2)
#define BBR_CWND_MIN_TARGET      4
#define BBR_FULL_BW_THRESH       (BBR_UNIT * 5 / 4)
#define BBR_FULL_BW_CNT          3
#define BBR_LOSS_THRESH          (BBR_UNIT * 2 / 100)
#define BBR_BETA                 (BBR_UNIT * 30 / 100)
#define BBR_INFLIGHT_HEADROOM    (BBR_UNIT * 15 / 100)
#define BBR_EXTRA_ACKED_GAIN     BBR_UNIT
#define BBR_EXTRA_ACKED_WIN_RTTS 5
#define BBR_ACK_EPOCH_ACKED_RESET_THRESH (1U << 20)
#define BBR_EXTRA_ACKED_MAX_US   (100 * 1000)
#define BBR_FULL_LOSS_CNT        6
#define BBR_BW_PROBE_MAX_ROUNDS  63
#define BBR_BW_PROBE_RAND_ROUNDS 2
#define BBR_BW_PROBE_BASE_US     (2 * 1000000)
#define BBR_BW_PROBE_RAND_US     (1 * 1000000)
#define BBR_PROBE_RTT_WIN_MS     5000
#define BBR_MIN_RTT_WIN_SEC      10
#define BBR_PROBE_RTT_MS         200

enum bbr_mode         { BBR_STARTUP = 0, BBR_DRAIN = 1, BBR_PROBE_BW = 2, BBR_PROBE_RTT = 3 };
enum bbr_bw_probe_cyc { BBR_BW_PROBE_UP = 0, BBR_BW_PROBE_DOWN = 1, BBR_BW_PROBE_CRUISE = 2, BBR_BW_PROBE_REFILL = 3 };
enum bbr_ack_phase    { BBR_ACKS_INIT = 0, BBR_ACKS_REFILLING, BBR_ACKS_PROBE_STARTING, BBR_ACKS_PROBE_FEEDBACK, BBR_ACKS_PROBE_STOPPING };

struct bbr_rate_sample {
    uint32_t delivered;
    uint32_t interval_us;
    uint32_t rtt_us;
    uint32_t acked_sacked;
    uint32_t prior_delivered;
    uint32_t tx_in_flight;
    int      lost;
    int      is_app_limited;
};

typedef void (*bbr_cwnd_update_fn)(void* ctx, uint32_t new_cwnd);

struct bbr {
    uint32_t min_rtt_us;
    uint32_t bw_lo;
    uint32_t bw_hi[2];
    uint32_t inflight_lo;
    uint32_t inflight_hi;
    uint32_t bw_latest;
    uint32_t inflight_latest;
    uint8_t  mode;
    uint8_t  cycle_idx;
    uint16_t pacing_gain;
    uint16_t cwnd_gain;
    uint32_t next_rtt_delivered;
    uint8_t  round_start     : 1;
    uint8_t  full_bw_reached : 1;
    uint8_t  loss_in_round   : 1;
    uint8_t  loss_in_cycle   : 1;
    uint8_t  idle_restart    : 1;
    uint8_t  try_fast_path   : 1;
    uint8_t  has_seen_rtt    : 1;
    uint8_t  initialized     : 1;
    uint32_t loss_round_delivered;
    uint32_t bw_probe_up_cnt;
    uint32_t bw_probe_up_acks;
    uint8_t  bw_probe_up_rounds;
    uint8_t  bw_probe_samples   : 1;
    uint8_t  prev_probe_too_high: 1;
    uint8_t  stopped_risky_probe: 1;
    uint8_t  rounds_since_probe;
    uint32_t probe_wait_us;
    uint64_t cycle_mstamp_tb;
    uint32_t probe_rtt_min_us;
    uint64_t probe_rtt_min_stamp_tb;
    uint64_t probe_rtt_done_stamp_tb;
    uint8_t  probe_rtt_round_done : 1;
    uint16_t extra_acked[2];
    uint32_t ack_epoch_acked;
    uint64_t ack_epoch_mstamp_tb;
    uint8_t  extra_acked_win_idx  : 1;
    uint8_t  extra_acked_win_rtts : 5;
    uint32_t prior_cwnd;
    uint8_t  init_cwnd     : 7;
    uint8_t  full_bw_cnt   : 2;
    uint8_t  ack_phase     : 3;
    uint8_t  loss_events_in_round : 4;
    uint32_t delivered;
    uint32_t full_bw;
    uint64_t min_rtt_stamp_tb;
    uint8_t  loss_round_start : 1;
    uint8_t  full_bw_now      : 1;
    uint8_t  pad_unused       : 6;
    uint8_t  link_id;              // local_link_id for debug logging
    uint64_t now_tb;               // 0 = real get_time_tb(); >0 = test-controlled time
    bbr_cwnd_update_fn on_cwnd_update;
    void* cwnd_update_ctx;
};

void bbr_init(struct bbr* bbr);
void bbr_main(struct bbr* bbr, const struct bbr_rate_sample* rs,
              uint32_t* cwnd_out, uint32_t* pacing_rate_out,
              uint32_t mss, uint32_t inflight_bytes, int is_cwnd_limited);
void bbr_note_loss(struct bbr* bbr);
void bbr_tx_start(struct bbr* bbr);

#ifdef __cplusplus
}
#endif