Golang log.Trace函数代码示例

// processReplica processes a single replica. This should not be
// called externally to the queue. bq.mu.Lock should not be held
// while calling this method.
func (bq *baseQueue) processReplica(repl *Replica, clock *hlc.Clock) error {
	bq.processMu.Lock()
	defer bq.processMu.Unlock()

	// Load the system config.
	cfg, ok := bq.gossip.GetSystemConfig()
	if !ok {
		log.VEventf(1, bq.ctx, "no system config available. skipping")
		return nil
	}

	if bq.requiresSplit(cfg, repl) {
		// Range needs to be split due to zone configs, but queue does
		// not accept unsplit ranges.
		log.VEventf(3, bq.ctx, "%s: split needed; skipping", repl)
		return nil
	}

	sp := repl.store.Tracer().StartSpan(bq.name)
	ctx := opentracing.ContextWithSpan(context.Background(), sp)
	defer sp.Finish()
	log.Tracef(ctx, "processing replica %s", repl)

	// If the queue requires a replica to have the range lease in
	// order to be processed, check whether this replica has range lease
	// and renew or acquire if necessary.
	if bq.needsLease {
		// Create a "fake" get request in order to invoke redirectOnOrAcquireLease.
		if err := repl.redirectOnOrAcquireLease(ctx); err != nil {
			if _, harmless := err.GetDetail().(*roachpb.NotLeaseHolderError); harmless {
				log.VEventf(3, bq.ctx, "%s: not holding lease; skipping", repl)
				return nil
			}
			return errors.Wrapf(err.GoError(), "%s: could not obtain lease", repl)
		}
		log.Trace(ctx, "got range lease")
	}

	log.VEventf(3, bq.ctx, "%s: processing", repl)
	start := timeutil.Now()
	if err := bq.impl.process(ctx, clock.Now(), repl, cfg); err != nil {
		return err
	}
	log.VEventf(2, bq.ctx, "%s: done: %s", repl, timeutil.Since(start))
	log.Trace(ctx, "done")
	return nil
}

// InitSenderForLocalTestCluster initializes a TxnCoordSender that can be used
// with LocalTestCluster.
func InitSenderForLocalTestCluster(
	nodeDesc *roachpb.NodeDescriptor,
	tracer opentracing.Tracer,
	clock *hlc.Clock,
	latency time.Duration,
	stores client.Sender,
	stopper *stop.Stopper,
	gossip *gossip.Gossip,
) client.Sender {
	var rpcSend rpcSendFn = func(_ SendOptions, _ ReplicaSlice,
		args roachpb.BatchRequest, _ *rpc.Context) (*roachpb.BatchResponse, error) {
		if latency > 0 {
			time.Sleep(latency)
		}
		sp := tracer.StartSpan("node")
		defer sp.Finish()
		ctx := opentracing.ContextWithSpan(context.Background(), sp)
		log.Trace(ctx, args.String())
		br, pErr := stores.Send(ctx, args)
		if br == nil {
			br = &roachpb.BatchResponse{}
		}
		if br.Error != nil {
			panic(roachpb.ErrorUnexpectedlySet(stores, br))
		}
		br.Error = pErr
		if pErr != nil {
			log.Trace(ctx, "error: "+pErr.String())
		}
		return br, nil
	}
	retryOpts := GetDefaultDistSenderRetryOptions()
	retryOpts.Closer = stopper.ShouldDrain()
	distSender := NewDistSender(&DistSenderContext{
		Clock: clock,
		RangeDescriptorCacheSize: defaultRangeDescriptorCacheSize,
		RangeLookupMaxRanges:     defaultRangeLookupMaxRanges,
		LeaderCacheSize:          defaultLeaderCacheSize,
		RPCRetryOptions:          &retryOpts,
		nodeDescriptor:           nodeDesc,
		RPCSend:                  rpcSend,                    // defined above
		RangeDescriptorDB:        stores.(RangeDescriptorDB), // for descriptor lookup
	}, gossip)

	return NewTxnCoordSender(distSender, clock, false /* !linearizable */, tracer,
		stopper, NewTxnMetrics(metric.NewRegistry()))
}

// processReplica processes a single replica. This should not be
// called externally to the queue. bq.mu.Lock should not be held
// while calling this method.
func (bq *baseQueue) processReplica(repl *Replica, clock *hlc.Clock) error {
	// Load the system config.
	cfg, ok := bq.gossip.GetSystemConfig()
	if !ok {
		bq.eventLog.VInfof(log.V(1), "no system config available. skipping")
		return nil
	}

	desc := repl.Desc()
	if !bq.acceptsUnsplitRanges && cfg.NeedsSplit(desc.StartKey, desc.EndKey) {
		// Range needs to be split due to zone configs, but queue does
		// not accept unsplit ranges.
		bq.eventLog.VInfof(log.V(3), "%s: split needed; skipping", repl)
		return nil
	}

	sp := repl.store.Tracer().StartSpan(bq.name)
	ctx := opentracing.ContextWithSpan(repl.context(context.Background()), sp)
	log.Trace(ctx, fmt.Sprintf("queue start for range %d", repl.RangeID))
	defer sp.Finish()

	// If the queue requires a replica to have the range leader lease in
	// order to be processed, check whether this replica has leader lease
	// and renew or acquire if necessary.
	if bq.needsLeaderLease {
		// Create a "fake" get request in order to invoke redirectOnOrAcquireLease.
		if err := repl.redirectOnOrAcquireLeaderLease(ctx); err != nil {
			if _, harmless := err.GetDetail().(*roachpb.NotLeaderError); harmless {
				bq.eventLog.VInfof(log.V(3), "%s: not holding lease; skipping", repl)
				return nil
			}
			return errors.Wrapf(err.GoError(), "%s: could not obtain lease", repl)
		}
		log.Trace(ctx, "got range lease")
	}

	bq.eventLog.VInfof(log.V(3), "%s: processing", repl)
	start := timeutil.Now()
	if err := bq.impl.process(ctx, clock.Now(), repl, cfg); err != nil {
		return err
	}
	bq.eventLog.VInfof(log.V(2), "%s: done: %s", repl, timeutil.Since(start))
	log.Trace(ctx, "done")
	return nil
}

// EvictAndReplace instructs the evictionToken to evict the RangeDescriptor it was
// created with from the rangeDescriptorCache. It also allows the user to provide
// new RangeDescriptors to insert into the cache, all atomically. When called without
// arguments, EvictAndReplace will behave the same as Evict.
func (et *evictionToken) EvictAndReplace(ctx context.Context, newDescs ...roachpb.RangeDescriptor) error {
	var err error
	et.doOnce.Do(func() {
		et.doLocker.Lock()
		defer et.doLocker.Unlock()
		err = et.do()
		if err == nil {
			if len(newDescs) > 0 {
				err = et.doReplace(newDescs...)
				log.Trace(ctx, fmt.Sprintf("evicting cached range descriptor with %d replacements",
					len(newDescs)))
			} else {
				log.Trace(ctx, "evicting cached range descriptor")
			}
		}
	})
	return err
}

// process synchronously invokes admin split for each proposed split key.
func (sq *splitQueue) process(
	ctx context.Context,
	now hlc.Timestamp,
	rng *Replica,
	sysCfg config.SystemConfig,
) error {
	// First handle case of splitting due to zone config maps.
	desc := rng.Desc()
	splitKeys := sysCfg.ComputeSplitKeys(desc.StartKey, desc.EndKey)
	if len(splitKeys) > 0 {
		log.Infof("splitting %s at keys %v", rng, splitKeys)
		log.Trace(ctx, fmt.Sprintf("splitting at keys %v", splitKeys))
		for _, splitKey := range splitKeys {
			if err := sq.db.AdminSplit(splitKey.AsRawKey()); err != nil {
				return errors.Errorf("unable to split %s at key %q: %s", rng, splitKey, err)
			}
		}
		return nil
	}

	// Next handle case of splitting due to size.
	zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
	if err != nil {
		return err
	}
	size := rng.GetMVCCStats().Total()
	// FIXME: why is this implementation not the same as the one above?
	if float64(size)/float64(zone.RangeMaxBytes) > 1 {
		log.Infof("splitting %s size=%d max=%d", rng, size, zone.RangeMaxBytes)
		log.Trace(ctx, fmt.Sprintf("splitting size=%d max=%d", size, zone.RangeMaxBytes))
		if _, pErr := client.SendWrappedWith(rng, ctx, roachpb.Header{
			Timestamp: now,
		}, &roachpb.AdminSplitRequest{
			Span: roachpb.Span{Key: desc.StartKey.AsRawKey()},
		}); pErr != nil {
			return pErr.GoError()
		}
	}
	return nil
}

// Send implements the client.Sender interface. The store is looked up from the
// store map if specified by the request; otherwise, the command is being
// executed locally, and the replica is determined via lookup through each
// store's LookupRange method. The latter path is taken only by unit tests.
func (ls *Stores) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
	// If we aren't given a Replica, then a little bending over
	// backwards here. This case applies exclusively to unittests.
	if ba.RangeID == 0 || ba.Replica.StoreID == 0 {
		rs, err := keys.Range(ba)
		if err != nil {
			return nil, roachpb.NewError(err)
		}
		rangeID, repl, err := ls.lookupReplica(rs.Key, rs.EndKey)
		if err != nil {
			return nil, roachpb.NewError(err)
		}
		ba.RangeID = rangeID
		ba.Replica = *repl
	}

	ctx = log.Add(ctx,
		log.RangeID, ba.RangeID)

	store, err := ls.GetStore(ba.Replica.StoreID)
	if err != nil {
		return nil, roachpb.NewError(err)
	}

	if ba.Txn != nil {
		// For calls that read data within a txn, we keep track of timestamps
		// observed from the various participating nodes' HLC clocks. If we have
		// a timestamp on file for this Node which is smaller than MaxTimestamp,
		// we can lower MaxTimestamp accordingly. If MaxTimestamp drops below
		// OrigTimestamp, we effectively can't see uncertainty restarts any
		// more.
		// Note that it's not an issue if MaxTimestamp propagates back out to
		// the client via a returned Transaction update - when updating a Txn
		// from another, the larger MaxTimestamp wins.
		if maxTS, ok := ba.Txn.GetObservedTimestamp(ba.Replica.NodeID); ok && maxTS.Less(ba.Txn.MaxTimestamp) {
			// Copy-on-write to protect others we might be sharing the Txn with.
			shallowTxn := *ba.Txn
			// The uncertainty window is [OrigTimestamp, maxTS), so if that window
			// is empty, there won't be any uncertainty restarts.
			if !ba.Txn.OrigTimestamp.Less(maxTS) {
				log.Trace(ctx, "read has no clock uncertainty")
			}
			shallowTxn.MaxTimestamp.Backward(maxTS)
			ba.Txn = &shallowTxn
		}
	}
	br, pErr := store.Send(ctx, ba)
	if br != nil && br.Error != nil {
		panic(roachpb.ErrorUnexpectedlySet(store, br))
	}
	return br, pErr
}

func (s *senderTransport) SendNext(done chan BatchCall) {
	if s.called {
		panic("called an exhausted transport")
	}
	s.called = true
	sp := s.tracer.StartSpan("node")
	defer sp.Finish()
	ctx := opentracing.ContextWithSpan(context.Background(), sp)
	log.Trace(ctx, s.args.String())
	br, pErr := s.sender.Send(ctx, s.args)
	if br == nil {
		br = &roachpb.BatchResponse{}
	}
	if br.Error != nil {
		panic(roachpb.ErrorUnexpectedlySet(s.sender, br))
	}
	br.Error = pErr
	if pErr != nil {
		log.Trace(ctx, "error: "+pErr.String())
	}
	done <- BatchCall{Reply: br}
}

// cleanupTxnLocked is called when a transaction ends. The transaction record
// is updated and the heartbeat goroutine signaled to clean up the transaction
// gracefully.
func (tc *TxnCoordSender) cleanupTxnLocked(ctx context.Context, txn roachpb.Transaction) {
	log.Trace(ctx, "coordinator stops")
	txnMeta, ok := tc.txns[*txn.ID]
	// The heartbeat might've already removed the record. Or we may have already
	// closed txnEnd but we are racing with the heartbeat cleanup.
	if !ok || txnMeta.txnEnd == nil {
		return
	}

	// The supplied txn may be newer than the one in txnMeta, which is relevant
	// for stats.
	txnMeta.txn = txn
	// Trigger heartbeat shutdown.
	close(txnMeta.txnEnd)
	txnMeta.txnEnd = nil
}

// cleanupTxn is called when a transaction ends. The transaction record is
// updated and the heartbeat goroutine signaled to clean up the transaction
// gracefully.
func (tc *TxnCoordSender) cleanupTxn(ctx context.Context, txn roachpb.Transaction) {
	log.Trace(ctx, "coordinator stops")
	tc.Lock()
	defer tc.Unlock()
	txnMeta, ok := tc.txns[*txn.ID]
	// The heartbeat might've already removed the record.
	if !ok {
		return
	}

	// The supplied txn may be newer than the one in txnMeta, which is relevant
	// for stats.
	txnMeta.txn = txn
	// Trigger heartbeat shutdown.
	close(txnMeta.txnEnd)
	txnMeta.txnEnd = nil
}

func (tc *TxnCoordSender) heartbeat(ctx context.Context, txnID uuid.UUID) bool {
	tc.Lock()
	txnMeta := tc.txns[txnID]
	txn := txnMeta.txn.Clone()
	tc.Unlock()

	// Before we send a heartbeat, determine whether this transaction should be
	// considered abandoned. If so, exit heartbeat. If ctx.Done() is not nil, then
	// it is a cancellable Context and we skip this check and use the ctx lifetime
	// instead of a timeout.
	if ctx.Done() == nil && txnMeta.hasClientAbandonedCoord(tc.clock.PhysicalNow()) {
		tc.clientHasAbandoned(txnID)
		return false
	}

	ba := roachpb.BatchRequest{}
	ba.Txn = &txn

	hb := &roachpb.HeartbeatTxnRequest{
		Now: tc.clock.Now(),
	}
	hb.Key = txn.Key
	ba.Add(hb)

	log.Trace(ctx, "heartbeat")
	_, err := tc.wrapped.Send(ctx, ba)
	// If the transaction is not in pending state, then we can stop
	// the heartbeat. It's either aborted or committed, and we resolve
	// write intents accordingly.
	if err != nil {
		log.Warningf("heartbeat to %s failed: %s", txn, err)
	}
	// TODO(bdarnell): once we have gotten a heartbeat response with
	// Status != PENDING, future heartbeats are useless. However, we
	// need to continue the heartbeatLoop until the client either
	// commits or abandons the transaction. We could save a little
	// pointless work by restructuring this loop to stop sending
	// heartbeats between the time that the transaction is aborted and
	// the client finds out. Furthermore, we could use this information
	// to send TransactionAbortedErrors to the client so it can restart
	// immediately instead of running until its EndTransaction.
	return true
}

// sendSingleRange gathers and rearranges the replicas, and makes an RPC call.
func (ds *DistSender) sendSingleRange(
	ctx context.Context, ba roachpb.BatchRequest, desc *roachpb.RangeDescriptor,
) (*roachpb.BatchResponse, *roachpb.Error) {
	log.Trace(ctx, fmt.Sprintf("sending RPC to [%s, %s)", desc.StartKey, desc.EndKey))

	// Try to send the call.
	replicas := newReplicaSlice(ds.gossip, desc)

	// Rearrange the replicas so that those replicas with long common
	// prefix of attributes end up first. If there's no prefix, this is a
	// no-op.
	order := ds.optimizeReplicaOrder(replicas)

	// If this request needs to go to a leader and we know who that is, move
	// it to the front.
	if !(ba.IsReadOnly() && ba.ReadConsistency == roachpb.INCONSISTENT) {
		if leader := ds.leaderCache.Lookup(roachpb.RangeID(desc.RangeID)); leader.StoreID > 0 {
			if i := replicas.FindReplica(leader.StoreID); i >= 0 {
				replicas.MoveToFront(i)
				order = orderStable
			}
		}
	}

	// TODO(tschottdorf): should serialize the trace here, not higher up.
	br, pErr := ds.sendRPC(ctx, desc.RangeID, replicas, order, ba)
	if pErr != nil {
		return nil, pErr
	}

	// If the reply contains a timestamp, update the local HLC with it.
	if br.Error != nil && br.Error.Now != roachpb.ZeroTimestamp {
		ds.clock.Update(br.Error.Now)
	} else if br.Now != roachpb.ZeroTimestamp {
		ds.clock.Update(br.Now)
	}

	// Untangle the error from the received response.
	pErr = br.Error
	br.Error = nil // scrub the response error
	return br, pErr
}

// Send implements the batch.Sender interface. If the request is part of a
// transaction, the TxnCoordSender adds the transaction to a map of active
// transactions and begins heartbeating it. Every subsequent request for the
// same transaction updates the lastUpdate timestamp to prevent live
// transactions from being considered abandoned and garbage collected.
// Read/write mutating requests have their key or key range added to the
// transaction's interval tree of key ranges for eventual cleanup via resolved
// write intents; they're tagged to an outgoing EndTransaction request, with
// the receiving replica in charge of resolving them.
func (tc *TxnCoordSender) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
	{
		// Start new or pick up active trace and embed its trace metadata into
		// header for use by RPC recipients. From here on, there's always an active
		// Trace, though its overhead is small unless it's sampled.
		sp := opentracing.SpanFromContext(ctx)
		if sp == nil {
			sp = tc.tracer.StartSpan(opTxnCoordSender)
			defer sp.Finish()
			ctx = opentracing.ContextWithSpan(ctx, sp)
		}
		// TODO(tschottdorf): To get rid of the spurious alloc below we need to
		// implement the carrier interface on ba.Header or make Span non-nullable,
		// both of which force all of ba on the Heap. It's already there, so may
		// not be a big deal, but ba should live on the stack. Also not easy to use
		// a buffer pool here since anything that goes into the RPC layer could be
		// used by goroutines we didn't wait for.
		if ba.Header.Trace == nil {
			ba.Header.Trace = &tracing.Span{}
		}
		if err := tc.tracer.Inject(sp, basictracer.Delegator, ba.Trace); err != nil {
			return nil, roachpb.NewError(err)
		}
	}

	startNS := tc.clock.PhysicalNow()

	if ba.Txn != nil {
		// If this request is part of a transaction...
		if err := tc.maybeBeginTxn(&ba); err != nil {
			return nil, roachpb.NewError(err)
		}
		txnID := *ba.Txn.ID
		// Verify that if this Transaction is not read-only, we have it on file.
		// If not, refuse further operations - the transaction was aborted due
		// to a timeout or the client must have issued a write on another
		// coordinator previously.
		if ba.Txn.Writing {
			tc.Lock()
			_, ok := tc.txns[txnID]
			tc.Unlock()
			if !ok {
				pErr := roachpb.NewErrorf("writing transaction timed out, was aborted, " +
					"or ran on multiple coordinators")
				return nil, pErr
			}
		}

		if rArgs, ok := ba.GetArg(roachpb.EndTransaction); ok {
			et := rArgs.(*roachpb.EndTransactionRequest)
			if len(et.Key) != 0 {
				return nil, roachpb.NewErrorf("EndTransaction must not have a Key set")
			}
			et.Key = ba.Txn.Key
			if len(et.IntentSpans) > 0 {
				// TODO(tschottdorf): it may be useful to allow this later.
				// That would be part of a possible plan to allow txns which
				// write on multiple coordinators.
				return nil, roachpb.NewErrorf("client must not pass intents to EndTransaction")
			}
			tc.Lock()
			txnMeta, metaOK := tc.txns[txnID]
			{
				// Populate et.IntentSpans, taking into account both existing
				// writes (if any) and new writes in this batch, and taking
				// care to perform proper deduplication.
				var keys interval.RangeGroup
				if metaOK {
					keys = txnMeta.keys
				} else {
					keys = interval.NewRangeTree()
				}
				ba.IntentSpanIterate(func(key, endKey roachpb.Key) {
					addKeyRange(keys, key, endKey)
				})
				et.IntentSpans = collectIntentSpans(keys)
			}
			tc.Unlock()

			if len(et.IntentSpans) > 0 {
				// All good, proceed.
			} else if !metaOK {
				// If we don't have the transaction, then this must be a retry
				// by the client. We can no longer reconstruct a correct
				// request so we must fail.
				//
				// TODO(bdarnell): if we had a GetTransactionStatus API then
				// we could lookup the transaction and return either nil or
				// TransactionAbortedError instead of this ambivalent error.
				return nil, roachpb.NewErrorf("transaction is already committed or aborted")
			}
//.........这里部分代码省略.........

func (tc *TxnCoordSender) heartbeat(ctx context.Context, txnID uuid.UUID) bool {
	tc.Lock()
	txnMeta := tc.txns[txnID]
	txn := txnMeta.txn.Clone()
	hasAbandoned := txnMeta.hasClientAbandonedCoord(tc.clock.PhysicalNow())
	tc.Unlock()

	if txn.Status != roachpb.PENDING {
		// A previous iteration has already determined that the transaction is
		// already finalized, so we wait for the client to realize that and
		// want to keep our state for the time being (to dish out the right
		// error once it returns).
		return true
	}

	// Before we send a heartbeat, determine whether this transaction should be
	// considered abandoned. If so, exit heartbeat. If ctx.Done() is not nil, then
	// it is a cancellable Context and we skip this check and use the ctx lifetime
	// instead of a timeout.
	if ctx.Done() == nil && hasAbandoned {
		if log.V(1) {
			log.Infof(ctx, "transaction %s abandoned; stopping heartbeat", txnMeta.txn)
		}
		tc.tryAsyncAbort(txnID)
		return false
	}

	ba := roachpb.BatchRequest{}
	ba.Txn = &txn

	hb := &roachpb.HeartbeatTxnRequest{
		Now: tc.clock.Now(),
	}
	hb.Key = txn.Key
	ba.Add(hb)

	log.Trace(ctx, "heartbeat")
	br, pErr := tc.wrapped.Send(ctx, ba)

	// Correctness mandates that when we can't heartbeat the transaction, we
	// make sure the client doesn't keep going. This is particularly relevant
	// in the case of an ABORTED transaction, but if we can't reach the
	// transaction record at all, we're going to have to assume we're aborted
	// as well.
	if pErr != nil {
		log.Warningf(ctx, "heartbeat to %s failed: %s", txn, pErr)
		// We're not going to let the client carry out additional requests, so
		// try to clean up.
		tc.tryAsyncAbort(*txn.ID)
		txn.Status = roachpb.ABORTED
	} else {
		txn.Update(br.Responses[0].GetInner().(*roachpb.HeartbeatTxnResponse).Txn)
	}

	// Give the news to the txn in the txns map. This will update long-running
	// transactions (which may find out that they have to restart in that way),
	// but in particular makes sure that they notice when they've been aborted
	// (in which case we'll give them an error on their next request).
	tc.Lock()
	tc.txns[txnID].txn.Update(&txn)
	tc.Unlock()

	return true
}

// maybePushTransactions tries to push the conflicting transaction(s)
// responsible for the given intents: either move its
// timestamp forward on a read/write conflict, abort it on a
// write/write conflict, or do nothing if the transaction is no longer
// pending.
//
// Returns a slice of intents which can now be resolved, and an error.
// The returned intents should be resolved via intentResolver.resolveIntents.
//
// If skipIfInFlight is true, then no PushTxns will be sent and no
// intents will be returned for any transaction for which there is
// another push in progress. This should only be used by callers who
// are not relying on the side effect of a push (i.e. only
// pushType==PUSH_TOUCH), and who also don't need to synchronize with
// the resolution of those intents (e.g. asynchronous resolutions of
// intents skipped on inconsistent reads).
//
// Callers are involved with
// a) conflict resolution for commands being executed at the Store with the
//    client waiting,
// b) resolving intents encountered during inconsistent operations, and
// c) resolving intents upon EndTransaction which are not local to the given
//    range. This is the only path in which the transaction is going to be
//    in non-pending state and doesn't require a push.
func (ir *intentResolver) maybePushTransactions(ctx context.Context, intents []roachpb.Intent,
	h roachpb.Header, pushType roachpb.PushTxnType, skipIfInFlight bool) (
	[]roachpb.Intent, *roachpb.Error) {

	now := ir.store.Clock().Now()

	partialPusherTxn := h.Txn
	// If there's no pusher, we communicate a priority by sending an empty
	// txn with only the priority set. This is official usage of PushTxn.
	if partialPusherTxn == nil {
		partialPusherTxn = &roachpb.Transaction{
			TxnMeta: roachpb.TxnMeta{
				Priority: roachpb.MakePriority(h.UserPriority),
			},
		}
	}

	log.Trace(ctx, "intent resolution")

	// Split intents into those we need to push and those which are good to
	// resolve.
	ir.mu.Lock()
	// TODO(tschottdorf): can optimize this and use same underlying slice.
	var pushIntents, nonPendingIntents []roachpb.Intent
	var pErr *roachpb.Error
	for _, intent := range intents {
		if intent.Status != roachpb.PENDING {
			// The current intent does not need conflict resolution
			// because the transaction is already finalized.
			// This shouldn't happen as all intents created are in
			// the PENDING status.
			nonPendingIntents = append(nonPendingIntents, intent)
		} else if _, ok := ir.mu.inFlight[*intent.Txn.ID]; ok && skipIfInFlight {
			// Another goroutine is working on this transaction so we can
			// skip it.
			if log.V(1) {
				log.Infof("skipping PushTxn for %s; attempt already in flight", intent.Txn.ID)
			}
			continue
		} else {
			pushIntents = append(pushIntents, intent)
			ir.mu.inFlight[*intent.Txn.ID]++
		}
	}
	ir.mu.Unlock()
	if len(nonPendingIntents) > 0 {
		return nil, roachpb.NewErrorf("unexpected aborted/resolved intents: %s", nonPendingIntents)
	}

	// Attempt to push the transaction(s) which created the conflicting intent(s).
	var pushReqs []roachpb.Request
	for _, intent := range pushIntents {
		pushReqs = append(pushReqs, &roachpb.PushTxnRequest{
			Span: roachpb.Span{
				Key: intent.Txn.Key,
			},
			PusherTxn: *partialPusherTxn,
			PusheeTxn: intent.Txn,
			PushTo:    h.Timestamp,
			// The timestamp is used by PushTxn for figuring out whether the
			// transaction is abandoned. If we used the argument's timestamp
			// here, we would run into busy loops because that timestamp
			// usually stays fixed among retries, so it will never realize
			// that a transaction has timed out. See #877.
			Now:      now,
			PushType: pushType,
		})
	}
	// TODO(kaneda): Set the transaction in the header so that the
	// txn is correctly propagated in an error response.
	b := &client.Batch{}
	b.InternalAddRequest(pushReqs...)
	br, pErr := ir.store.db.RunWithResponse(b)
	ir.mu.Lock()
	for _, intent := range pushIntents {
		ir.mu.inFlight[*intent.Txn.ID]--
//.........这里部分代码省略.........

// resolveIntents resolves the given intents. For those which are
// local to the range, we submit directly to the local Raft instance;
// all non-local intents are resolved asynchronously in a batch. If
// `wait` is true, all operations are carried out synchronously and an
// error is returned. Otherwise, the call returns without error as
// soon as all local resolve commands have been **proposed** (not
// executed). This ensures that if a waiting client retries
// immediately after calling this function, it will not hit the same
// intents again.
func (ir *intentResolver) resolveIntents(ctx context.Context, r *Replica,
	intents []roachpb.Intent, wait bool, poison bool) error {
	// We're doing async stuff below; those need new traces.
	ctx, cleanup := tracing.EnsureContext(ctx, ir.store.Tracer())
	defer cleanup()
	log.Trace(ctx, fmt.Sprintf("resolving intents [wait=%t]", wait))

	var reqsRemote []roachpb.Request
	baLocal := roachpb.BatchRequest{}
	baLocal.Timestamp = ir.store.Clock().Now()
	for i := range intents {
		intent := intents[i] // avoids a race in `i, intent := range ...`
		var resolveArgs roachpb.Request
		var local bool // whether this intent lives on this Range
		{
			if len(intent.EndKey) == 0 {
				resolveArgs = &roachpb.ResolveIntentRequest{
					Span:      intent.Span,
					IntentTxn: intent.Txn,
					Status:    intent.Status,
					Poison:    poison,
				}
				local = r.ContainsKey(intent.Key)
			} else {
				resolveArgs = &roachpb.ResolveIntentRangeRequest{
					Span:      intent.Span,
					IntentTxn: intent.Txn,
					Status:    intent.Status,
					Poison:    poison,
				}
				local = r.ContainsKeyRange(intent.Key, intent.EndKey)
			}
		}

		// If the intent isn't (completely) local, we'll need to send an external request.
		// We'll batch them all up and send at the end.
		if local {
			baLocal.Add(resolveArgs)
		} else {
			reqsRemote = append(reqsRemote, resolveArgs)
		}
	}

	// The local batch goes directly to Raft.
	var wg sync.WaitGroup
	if len(baLocal.Requests) > 0 {
		action := func() error {
			// Trace this under the ID of the intent owner.
			// Create a new span though, since we do not want to pass a span
			// between goroutines or we risk use-after-finish.
			sp := r.store.Tracer().StartSpan("resolve intents")
			defer sp.Finish()
			ctx = opentracing.ContextWithSpan(ctx, sp)
			// Always operate with a timeout when resolving intents: this
			// prevents rare shutdown timeouts in tests.
			ctxWithTimeout, cancel := context.WithTimeout(ctx, base.NetworkTimeout)
			defer cancel()
			_, pErr := r.addWriteCmd(ctxWithTimeout, baLocal, &wg)
			return pErr.GoError()
		}
		wg.Add(1)
		if wait || !r.store.Stopper().RunLimitedAsyncTask(ir.sem, func() {
			if err := action(); err != nil {
				log.Warningf("unable to resolve local intents; %s", err)
			}
		}) {
			// Still run the task when draining. Our caller already has a task and
			// going async here again is merely for performance, but some intents
			// need to be resolved because they might block other tasks. See #1684.
			// Note that handleSkippedIntents has a TODO in case #1684 comes back.
			if err := action(); err != nil {
				return err
			}
		}
	}

	// Resolve all of the intents which aren't local to the Range.
	if len(reqsRemote) > 0 {
		b := &client.Batch{}
		b.InternalAddRequest(reqsRemote...)
		action := func() error {
			// TODO(tschottdorf): no tracing here yet.
			return r.store.DB().Run(b).GoError()
		}
		if wait || !r.store.Stopper().RunLimitedAsyncTask(ir.sem, func() {
			if err := action(); err != nil {
				log.Warningf("unable to resolve external intents: %s", err)
			}
		}) {
			// As with local intents, try async to not keep the caller waiting, but
			// when draining just go ahead and do it synchronously. See #1684.
//.........这里部分代码省略.........