package messaging

import (
	"log"
	"os"
	"strconv"
	"sync"

	mqtt "github.com/eclipse/paho.mqtt.golang"
)

// Concurrency for the MQTT ingest.
//
// paho delivers messages on a single goroutine, so without this every bill is
// committed one after another. A bill is a full Postgres transaction — advisory
// lock, dedup check, stock row locks, availability check, four inserts, commit —
// which realistically costs 10–30ms. Serially that is 30–100 bills a second,
// and a shop-wide backlog draining after an outage would take minutes to land.
//
// ### Why a bounded pool rather than a goroutine per message
//
// paho can be told to call handlers concurrently, but it spawns without limit.
// A storm would then open a database transaction per message, exhaust the
// connection pool, and stall every one of them at once — turning a slow minute
// into a dead one.
//
// A fixed pool behind a bounded queue does the opposite. When the queue fills,
// submitting **blocks**, which is the point: paho stops acknowledging, the
// broker's in-flight window fills, and it stops sending. Backpressure travels
// all the way back to the till, which holds its bills and retries. Slow, but
// nothing is dropped and nothing is lost.
//
// ### Why bills and heartbeats have separate pools
//
// A heartbeat is one Redis write and a bill is a transaction. Sharing a queue
// would let a backlog of bills delay presence, and every till would appear to
// go dark at exactly the moment the system was busiest — the worst possible
// time to be blind to which counters are alive.

// posPool is a fixed set of workers reading a bounded queue.
type posPool struct {
	name string
	jobs chan func()
	wg   sync.WaitGroup
	once sync.Once

	// Guards the transition to closed. A plain `select` over a done-channel and
	// the job channel is not enough: once both are ready Go picks between them
	// at random, and picking the send panics on a closed channel. Held for
	// reading across the whole of submit, so stop cannot close the queue out
	// from under a send already in progress.
	mu     sync.RWMutex
	closed bool
}

func newPosPool(name string, workers, queue int) *posPool {
	p := &posPool{
		name: name,
		jobs: make(chan func(), queue),
	}

	p.wg.Add(workers)
	for i := 0; i < workers; i++ {
		go func() {
			defer p.wg.Done()
			for job := range p.jobs {
				job()
			}
		}()
	}

	log.Printf("pos: %s pool started with %d workers, queue %d", name, workers, queue)
	return p
}

// submit queues work, blocking when the queue is full.
//
// Blocking is deliberate. Dropping would lose a bill outright; the terminal
// would eventually re-send it, but only after its ack timeout, and meanwhile we
// would have thrown away work we had already accepted. Blocking instead pushes
// back through paho to the broker to the till, which is exactly where the
// decision to slow down belongs.
func (p *posPool) submit(job func()) {
	p.mu.RLock()

	if p.closed {
		p.mu.RUnlock()
		// Shutting down. Running it inline still gets the work done and its ack
		// published, rather than discarding a bill that already reached us.
		job()
		return
	}

	// The read lock is held across the send. Blocking here while the queue is
	// full cannot deadlock against stop: the workers only exit once the channel
	// is closed, and that happens under the write lock this send is holding
	// off — so they stay alive and keep draining until this send completes.
	p.jobs <- job
	p.mu.RUnlock()
}

// stop drains the queue and waits for in-flight work.
//
// Every job already accepted runs to completion, so a bill mid-commit still
// gets its ack. Without one the terminal would hold it and send it again on
// restart — harmless, but avoidable.
func (p *posPool) stop() {
	p.once.Do(func() {
		// The write lock waits for every submit already in progress, so the
		// channel is never closed while something is mid-send.
		p.mu.Lock()
		p.closed = true
		close(p.jobs)
		p.mu.Unlock()

		p.wg.Wait()
		log.Printf("pos: %s pool drained", p.name)
	})
}

// posPoolSize reads a worker count from the environment.
//
// The default is deliberately modest. Each worker holds a database transaction
// while it runs, so the useful ceiling is the Postgres connection pool, not the
// CPU — set this above what the database can serve and the workers simply queue
// inside the driver instead, where there is no backpressure to feel.
func posPoolSize(key string, fallback int) int {
	v, err := strconv.Atoi(os.Getenv(key))
	if err != nil || v <= 0 {
		return fallback
	}
	if v > 128 {
		return 128
	}
	return v
}

// posOrderedDelivery reports whether paho should preserve message order.
//
// Left on: paho then delivers on one goroutine, which hands work to the pool
// and blocks when it is full. That single delivery goroutine is what makes
// backpressure reach the broker at all — with concurrent delivery paho would
// keep reading regardless of how far behind the workers were.
const posOrderedDelivery = true

// wrapHandler puts a paho message handler behind a pool.
//
// The payload is copied because paho reuses its buffer once the handler
// returns, and the work now happens after that.
func wrapHandler(pool *posPool, h mqtt.MessageHandler) mqtt.MessageHandler {
	return func(client mqtt.Client, msg mqtt.Message) {
		topic := msg.Topic()
		payload := make([]byte, len(msg.Payload()))
		copy(payload, msg.Payload())

		pool.submit(func() {
			h(client, copiedMessage{topic: topic, payload: payload})
		})
	}
}

// copiedMessage carries a payload that outlives paho's buffer.
type copiedMessage struct {
	topic   string
	payload []byte
}

func (m copiedMessage) Duplicate() bool   { return false }
func (m copiedMessage) Qos() byte         { return 1 }
func (m copiedMessage) Retained() bool    { return false }
func (m copiedMessage) Topic() string     { return m.topic }
func (m copiedMessage) MessageID() uint16 { return 0 }
func (m copiedMessage) Payload() []byte   { return m.payload }
func (m copiedMessage) Ack()              {}