lotus/storage/fpost_sched.go
laser ea2f53745e split FPoStScheduler from Miner
Rationale: Nodes want to handle scheduling of PoSts. Splitting the 
FPoStScheduler from the Miner allows nodes to reuse the Miner w/out 
having to reuse the PoSt scheduler.
2020-02-03 15:48:11 -08:00

175 lines
3.6 KiB
Go

package storage
import (
"context"
"sync"
"go.opencensus.io/trace"
"golang.org/x/xerrors"
"github.com/filecoin-project/go-address"
"github.com/filecoin-project/go-sectorbuilder"
"github.com/filecoin-project/lotus/build"
"github.com/filecoin-project/lotus/chain/store"
"github.com/filecoin-project/lotus/chain/types"
)
const Inactive = 0
const StartConfidence = 4 // TODO: config
type FPoStScheduler struct {
api storageMinerApi
sb sectorbuilder.Interface
actor address.Address
worker address.Address
cur *types.TipSet
// if a post is in progress, this indicates for which ElectionPeriodStart
activeEPS uint64
abort context.CancelFunc
failed uint64 // eps
failLk sync.Mutex
}
func NewFPoStScheduler(api storageMinerApi, sb sectorbuilder.Interface, actor address.Address, worker address.Address) *FPoStScheduler {
return &FPoStScheduler{api: api, sb: sb, actor: actor, worker: worker}
}
func (s *FPoStScheduler) Run(ctx context.Context) {
notifs, err := s.api.ChainNotify(ctx)
if err != nil {
return
}
current := <-notifs
if len(current) != 1 {
panic("expected first notif to have len = 1")
}
if current[0].Type != store.HCCurrent {
panic("expected first notif to tell current ts")
}
if err := s.update(ctx, current[0].Val); err != nil {
panic(err)
}
defer s.abortActivePoSt()
// not fine to panic after this point
for {
select {
case changes, ok := <-notifs:
if !ok {
log.Warn("FPoStScheduler notifs channel closed")
return
}
ctx, span := trace.StartSpan(ctx, "FPoStScheduler.headChange")
var lowest, highest *types.TipSet = s.cur, nil
for _, change := range changes {
if change.Val == nil {
log.Errorf("change.Val was nil")
}
switch change.Type {
case store.HCRevert:
lowest = change.Val
case store.HCApply:
highest = change.Val
}
}
if err := s.revert(ctx, lowest); err != nil {
log.Error("handling head reverts in fallbackPost sched: %+v", err)
}
if err := s.update(ctx, highest); err != nil {
log.Error("handling head updates in fallbackPost sched: %+v", err)
}
span.End()
case <-ctx.Done():
return
}
}
}
func (s *FPoStScheduler) revert(ctx context.Context, newLowest *types.TipSet) error {
if s.cur == newLowest {
return nil
}
s.cur = newLowest
newEPS, _, err := s.shouldFallbackPost(ctx, newLowest)
if err != nil {
return err
}
if newEPS != s.activeEPS {
s.abortActivePoSt()
}
return nil
}
func (s *FPoStScheduler) update(ctx context.Context, new *types.TipSet) error {
if new == nil {
return xerrors.Errorf("no new tipset in FPoStScheduler.update")
}
newEPS, start, err := s.shouldFallbackPost(ctx, new)
if err != nil {
return err
}
s.failLk.Lock()
if s.failed > 0 {
s.failed = 0
s.activeEPS = 0
}
s.failLk.Unlock()
if newEPS == s.activeEPS {
return nil
}
s.abortActivePoSt()
if newEPS != Inactive && start {
s.doPost(ctx, newEPS, new)
}
return nil
}
func (s *FPoStScheduler) abortActivePoSt() {
if s.activeEPS == Inactive {
return // noop
}
if s.abort != nil {
s.abort()
}
log.Warnf("Aborting Fallback PoSt (EPS: %d)", s.activeEPS)
s.activeEPS = Inactive
s.abort = nil
}
func (s *FPoStScheduler) shouldFallbackPost(ctx context.Context, ts *types.TipSet) (uint64, bool, error) {
eps, err := s.api.StateMinerElectionPeriodStart(ctx, s.actor, ts)
if err != nil {
return 0, false, xerrors.Errorf("getting ElectionPeriodStart: %w", err)
}
if ts.Height() >= eps+build.FallbackPoStDelay {
return eps, ts.Height() >= eps+build.FallbackPoStDelay+StartConfidence, nil
}
return 0, false, nil
}