Merge commit 'main' into 21160-user-activation
[arvados.git] / lib / cloud / ec2 / ec2.go
1 // Copyright (C) The Arvados Authors. All rights reserved.
2 //
3 // SPDX-License-Identifier: AGPL-3.0
4
5 package ec2
6
7 import (
8         "crypto/md5"
9         "crypto/rsa"
10         "crypto/sha1"
11         "crypto/x509"
12         "encoding/base64"
13         "encoding/json"
14         "fmt"
15         "math/big"
16         "strconv"
17         "strings"
18         "sync"
19         "sync/atomic"
20         "time"
21
22         "git.arvados.org/arvados.git/lib/cloud"
23         "git.arvados.org/arvados.git/sdk/go/arvados"
24         "github.com/aws/aws-sdk-go/aws"
25         "github.com/aws/aws-sdk-go/aws/awserr"
26         "github.com/aws/aws-sdk-go/aws/credentials"
27         "github.com/aws/aws-sdk-go/aws/credentials/ec2rolecreds"
28         "github.com/aws/aws-sdk-go/aws/ec2metadata"
29         "github.com/aws/aws-sdk-go/aws/request"
30         "github.com/aws/aws-sdk-go/aws/session"
31         "github.com/aws/aws-sdk-go/service/ec2"
32         "github.com/prometheus/client_golang/prometheus"
33         "github.com/sirupsen/logrus"
34         "golang.org/x/crypto/ssh"
35 )
36
37 // Driver is the ec2 implementation of the cloud.Driver interface.
38 var Driver = cloud.DriverFunc(newEC2InstanceSet)
39
40 const (
41         throttleDelayMin = time.Second
42         throttleDelayMax = time.Minute
43 )
44
45 type ec2InstanceSetConfig struct {
46         AccessKeyID             string
47         SecretAccessKey         string
48         Region                  string
49         SecurityGroupIDs        arvados.StringSet
50         SubnetID                sliceOrSingleString
51         AdminUsername           string
52         EBSVolumeType           string
53         EBSPrice                float64
54         IAMInstanceProfile      string
55         SpotPriceUpdateInterval arvados.Duration
56 }
57
58 type sliceOrSingleString []string
59
60 // UnmarshalJSON unmarshals an array of strings, and also accepts ""
61 // as [], and "foo" as ["foo"].
62 func (ss *sliceOrSingleString) UnmarshalJSON(data []byte) error {
63         if len(data) == 0 {
64                 *ss = nil
65         } else if data[0] == '[' {
66                 var slice []string
67                 err := json.Unmarshal(data, &slice)
68                 if err != nil {
69                         return err
70                 }
71                 if len(slice) == 0 {
72                         *ss = nil
73                 } else {
74                         *ss = slice
75                 }
76         } else {
77                 var str string
78                 err := json.Unmarshal(data, &str)
79                 if err != nil {
80                         return err
81                 }
82                 if str == "" {
83                         *ss = nil
84                 } else {
85                         *ss = []string{str}
86                 }
87         }
88         return nil
89 }
90
91 type ec2Interface interface {
92         DescribeKeyPairs(input *ec2.DescribeKeyPairsInput) (*ec2.DescribeKeyPairsOutput, error)
93         ImportKeyPair(input *ec2.ImportKeyPairInput) (*ec2.ImportKeyPairOutput, error)
94         RunInstances(input *ec2.RunInstancesInput) (*ec2.Reservation, error)
95         DescribeInstances(input *ec2.DescribeInstancesInput) (*ec2.DescribeInstancesOutput, error)
96         DescribeInstanceStatusPages(input *ec2.DescribeInstanceStatusInput, fn func(*ec2.DescribeInstanceStatusOutput, bool) bool) error
97         DescribeSpotPriceHistoryPages(input *ec2.DescribeSpotPriceHistoryInput, fn func(*ec2.DescribeSpotPriceHistoryOutput, bool) bool) error
98         CreateTags(input *ec2.CreateTagsInput) (*ec2.CreateTagsOutput, error)
99         TerminateInstances(input *ec2.TerminateInstancesInput) (*ec2.TerminateInstancesOutput, error)
100 }
101
102 type ec2InstanceSet struct {
103         ec2config              ec2InstanceSetConfig
104         currentSubnetIDIndex   int32
105         instanceSetID          cloud.InstanceSetID
106         logger                 logrus.FieldLogger
107         client                 ec2Interface
108         keysMtx                sync.Mutex
109         keys                   map[string]string
110         throttleDelayCreate    atomic.Value
111         throttleDelayInstances atomic.Value
112
113         prices        map[priceKey][]cloud.InstancePrice
114         pricesLock    sync.Mutex
115         pricesUpdated map[priceKey]time.Time
116
117         mInstances      *prometheus.GaugeVec
118         mInstanceStarts *prometheus.CounterVec
119 }
120
121 func newEC2InstanceSet(config json.RawMessage, instanceSetID cloud.InstanceSetID, _ cloud.SharedResourceTags, logger logrus.FieldLogger, reg *prometheus.Registry) (prv cloud.InstanceSet, err error) {
122         instanceSet := &ec2InstanceSet{
123                 instanceSetID: instanceSetID,
124                 logger:        logger,
125         }
126         err = json.Unmarshal(config, &instanceSet.ec2config)
127         if err != nil {
128                 return nil, err
129         }
130
131         sess, err := session.NewSession()
132         if err != nil {
133                 return nil, err
134         }
135         // First try any static credentials, fall back to an IAM instance profile/role
136         creds := credentials.NewChainCredentials(
137                 []credentials.Provider{
138                         &credentials.StaticProvider{Value: credentials.Value{AccessKeyID: instanceSet.ec2config.AccessKeyID, SecretAccessKey: instanceSet.ec2config.SecretAccessKey}},
139                         &ec2rolecreds.EC2RoleProvider{Client: ec2metadata.New(sess)},
140                 })
141
142         awsConfig := aws.NewConfig().WithCredentials(creds).WithRegion(instanceSet.ec2config.Region)
143         instanceSet.client = ec2.New(session.Must(session.NewSession(awsConfig)))
144         instanceSet.keys = make(map[string]string)
145         if instanceSet.ec2config.EBSVolumeType == "" {
146                 instanceSet.ec2config.EBSVolumeType = "gp2"
147         }
148
149         // Set up metrics
150         instanceSet.mInstances = prometheus.NewGaugeVec(prometheus.GaugeOpts{
151                 Namespace: "arvados",
152                 Subsystem: "dispatchcloud",
153                 Name:      "ec2_instances",
154                 Help:      "Number of instances running",
155         }, []string{"subnet_id"})
156         instanceSet.mInstanceStarts = prometheus.NewCounterVec(prometheus.CounterOpts{
157                 Namespace: "arvados",
158                 Subsystem: "dispatchcloud",
159                 Name:      "ec2_instance_starts_total",
160                 Help:      "Number of attempts to start a new instance",
161         }, []string{"subnet_id", "success"})
162         // Initialize all of the series we'll be reporting.  Otherwise
163         // the {subnet=A, success=0} series doesn't appear in metrics
164         // at all until there's a failure in subnet A.
165         for _, subnet := range instanceSet.ec2config.SubnetID {
166                 instanceSet.mInstanceStarts.WithLabelValues(subnet, "0").Add(0)
167                 instanceSet.mInstanceStarts.WithLabelValues(subnet, "1").Add(0)
168         }
169         if len(instanceSet.ec2config.SubnetID) == 0 {
170                 instanceSet.mInstanceStarts.WithLabelValues("", "0").Add(0)
171                 instanceSet.mInstanceStarts.WithLabelValues("", "1").Add(0)
172         }
173         if reg != nil {
174                 reg.MustRegister(instanceSet.mInstances)
175                 reg.MustRegister(instanceSet.mInstanceStarts)
176         }
177
178         return instanceSet, nil
179 }
180
181 func awsKeyFingerprint(pk ssh.PublicKey) (md5fp string, sha1fp string, err error) {
182         // AWS key fingerprints don't use the usual key fingerprint
183         // you get from ssh-keygen or ssh.FingerprintLegacyMD5()
184         // (you can get that from md5.Sum(pk.Marshal())
185         //
186         // AWS uses the md5 or sha1 of the PKIX DER encoding of the
187         // public key, so calculate those fingerprints here.
188         var rsaPub struct {
189                 Name string
190                 E    *big.Int
191                 N    *big.Int
192         }
193         if err := ssh.Unmarshal(pk.Marshal(), &rsaPub); err != nil {
194                 return "", "", fmt.Errorf("agent: Unmarshal failed to parse public key: %v", err)
195         }
196         rsaPk := rsa.PublicKey{
197                 E: int(rsaPub.E.Int64()),
198                 N: rsaPub.N,
199         }
200         pkix, _ := x509.MarshalPKIXPublicKey(&rsaPk)
201         md5pkix := md5.Sum([]byte(pkix))
202         sha1pkix := sha1.Sum([]byte(pkix))
203         md5fp = ""
204         sha1fp = ""
205         for i := 0; i < len(md5pkix); i++ {
206                 md5fp += fmt.Sprintf(":%02x", md5pkix[i])
207         }
208         for i := 0; i < len(sha1pkix); i++ {
209                 sha1fp += fmt.Sprintf(":%02x", sha1pkix[i])
210         }
211         return md5fp[1:], sha1fp[1:], nil
212 }
213
214 func (instanceSet *ec2InstanceSet) Create(
215         instanceType arvados.InstanceType,
216         imageID cloud.ImageID,
217         newTags cloud.InstanceTags,
218         initCommand cloud.InitCommand,
219         publicKey ssh.PublicKey) (cloud.Instance, error) {
220
221         ec2tags := []*ec2.Tag{}
222         for k, v := range newTags {
223                 ec2tags = append(ec2tags, &ec2.Tag{
224                         Key:   aws.String(k),
225                         Value: aws.String(v),
226                 })
227         }
228
229         var groups []string
230         for sg := range instanceSet.ec2config.SecurityGroupIDs {
231                 groups = append(groups, sg)
232         }
233
234         rii := ec2.RunInstancesInput{
235                 ImageId:      aws.String(string(imageID)),
236                 InstanceType: &instanceType.ProviderType,
237                 MaxCount:     aws.Int64(1),
238                 MinCount:     aws.Int64(1),
239
240                 NetworkInterfaces: []*ec2.InstanceNetworkInterfaceSpecification{
241                         {
242                                 AssociatePublicIpAddress: aws.Bool(false),
243                                 DeleteOnTermination:      aws.Bool(true),
244                                 DeviceIndex:              aws.Int64(0),
245                                 Groups:                   aws.StringSlice(groups),
246                         }},
247                 DisableApiTermination:             aws.Bool(false),
248                 InstanceInitiatedShutdownBehavior: aws.String("terminate"),
249                 TagSpecifications: []*ec2.TagSpecification{
250                         {
251                                 ResourceType: aws.String("instance"),
252                                 Tags:         ec2tags,
253                         }},
254                 UserData: aws.String(base64.StdEncoding.EncodeToString([]byte("#!/bin/sh\n" + initCommand + "\n"))),
255         }
256
257         if publicKey != nil {
258                 keyname, err := instanceSet.getKeyName(publicKey)
259                 if err != nil {
260                         return nil, err
261                 }
262                 rii.KeyName = &keyname
263         }
264
265         if instanceType.AddedScratch > 0 {
266                 rii.BlockDeviceMappings = []*ec2.BlockDeviceMapping{{
267                         DeviceName: aws.String("/dev/xvdt"),
268                         Ebs: &ec2.EbsBlockDevice{
269                                 DeleteOnTermination: aws.Bool(true),
270                                 VolumeSize:          aws.Int64((int64(instanceType.AddedScratch) + (1<<30 - 1)) >> 30),
271                                 VolumeType:          &instanceSet.ec2config.EBSVolumeType,
272                         }}}
273         }
274
275         if instanceType.Preemptible {
276                 rii.InstanceMarketOptions = &ec2.InstanceMarketOptionsRequest{
277                         MarketType: aws.String("spot"),
278                         SpotOptions: &ec2.SpotMarketOptions{
279                                 InstanceInterruptionBehavior: aws.String("terminate"),
280                                 MaxPrice:                     aws.String(fmt.Sprintf("%v", instanceType.Price)),
281                         }}
282         }
283
284         if instanceSet.ec2config.IAMInstanceProfile != "" {
285                 rii.IamInstanceProfile = &ec2.IamInstanceProfileSpecification{
286                         Name: aws.String(instanceSet.ec2config.IAMInstanceProfile),
287                 }
288         }
289
290         var rsv *ec2.Reservation
291         var errToReturn error
292         subnets := instanceSet.ec2config.SubnetID
293         currentSubnetIDIndex := int(atomic.LoadInt32(&instanceSet.currentSubnetIDIndex))
294         for tryOffset := 0; ; tryOffset++ {
295                 tryIndex := 0
296                 trySubnet := ""
297                 if len(subnets) > 0 {
298                         tryIndex = (currentSubnetIDIndex + tryOffset) % len(subnets)
299                         trySubnet = subnets[tryIndex]
300                         rii.NetworkInterfaces[0].SubnetId = aws.String(trySubnet)
301                 }
302                 var err error
303                 rsv, err = instanceSet.client.RunInstances(&rii)
304                 instanceSet.mInstanceStarts.WithLabelValues(trySubnet, boolLabelValue[err == nil]).Add(1)
305                 if !isErrorCapacity(errToReturn) || isErrorCapacity(err) {
306                         // We want to return the last capacity error,
307                         // if any; otherwise the last non-capacity
308                         // error.
309                         errToReturn = err
310                 }
311                 if isErrorSubnetSpecific(err) &&
312                         tryOffset < len(subnets)-1 {
313                         instanceSet.logger.WithError(err).WithField("SubnetID", subnets[tryIndex]).
314                                 Warn("RunInstances failed, trying next subnet")
315                         continue
316                 }
317                 // Succeeded, or exhausted all subnets, or got a
318                 // non-subnet-related error.
319                 //
320                 // We intentionally update currentSubnetIDIndex even
321                 // in the non-retryable-failure case here to avoid a
322                 // situation where successive calls to Create() keep
323                 // returning errors for the same subnet (perhaps
324                 // "subnet full") and never reveal the errors for the
325                 // other configured subnets (perhaps "subnet ID
326                 // invalid").
327                 atomic.StoreInt32(&instanceSet.currentSubnetIDIndex, int32(tryIndex))
328                 break
329         }
330         if rsv == nil || len(rsv.Instances) == 0 {
331                 return nil, wrapError(errToReturn, &instanceSet.throttleDelayCreate)
332         }
333         return &ec2Instance{
334                 provider: instanceSet,
335                 instance: rsv.Instances[0],
336         }, nil
337 }
338
339 func (instanceSet *ec2InstanceSet) getKeyName(publicKey ssh.PublicKey) (string, error) {
340         instanceSet.keysMtx.Lock()
341         defer instanceSet.keysMtx.Unlock()
342         md5keyFingerprint, sha1keyFingerprint, err := awsKeyFingerprint(publicKey)
343         if err != nil {
344                 return "", fmt.Errorf("Could not make key fingerprint: %v", err)
345         }
346         if keyname, ok := instanceSet.keys[md5keyFingerprint]; ok {
347                 return keyname, nil
348         }
349         keyout, err := instanceSet.client.DescribeKeyPairs(&ec2.DescribeKeyPairsInput{
350                 Filters: []*ec2.Filter{{
351                         Name:   aws.String("fingerprint"),
352                         Values: []*string{&md5keyFingerprint, &sha1keyFingerprint},
353                 }},
354         })
355         if err != nil {
356                 return "", fmt.Errorf("Could not search for keypair: %v", err)
357         }
358         if len(keyout.KeyPairs) > 0 {
359                 return *(keyout.KeyPairs[0].KeyName), nil
360         }
361         keyname := "arvados-dispatch-keypair-" + md5keyFingerprint
362         _, err = instanceSet.client.ImportKeyPair(&ec2.ImportKeyPairInput{
363                 KeyName:           &keyname,
364                 PublicKeyMaterial: ssh.MarshalAuthorizedKey(publicKey),
365         })
366         if err != nil {
367                 return "", fmt.Errorf("Could not import keypair: %v", err)
368         }
369         instanceSet.keys[md5keyFingerprint] = keyname
370         return keyname, nil
371 }
372
373 func (instanceSet *ec2InstanceSet) Instances(tags cloud.InstanceTags) (instances []cloud.Instance, err error) {
374         var filters []*ec2.Filter
375         for k, v := range tags {
376                 filters = append(filters, &ec2.Filter{
377                         Name:   aws.String("tag:" + k),
378                         Values: []*string{aws.String(v)},
379                 })
380         }
381         needAZs := false
382         dii := &ec2.DescribeInstancesInput{Filters: filters}
383         for {
384                 dio, err := instanceSet.client.DescribeInstances(dii)
385                 err = wrapError(err, &instanceSet.throttleDelayInstances)
386                 if err != nil {
387                         return nil, err
388                 }
389
390                 for _, rsv := range dio.Reservations {
391                         for _, inst := range rsv.Instances {
392                                 if *inst.State.Name != "shutting-down" && *inst.State.Name != "terminated" {
393                                         instances = append(instances, &ec2Instance{
394                                                 provider: instanceSet,
395                                                 instance: inst,
396                                         })
397                                         if aws.StringValue(inst.InstanceLifecycle) == "spot" {
398                                                 needAZs = true
399                                         }
400                                 }
401                         }
402                 }
403                 if dio.NextToken == nil {
404                         break
405                 }
406                 dii.NextToken = dio.NextToken
407         }
408         if needAZs && instanceSet.ec2config.SpotPriceUpdateInterval > 0 {
409                 az := map[string]string{}
410                 err := instanceSet.client.DescribeInstanceStatusPages(&ec2.DescribeInstanceStatusInput{
411                         IncludeAllInstances: aws.Bool(true),
412                 }, func(page *ec2.DescribeInstanceStatusOutput, lastPage bool) bool {
413                         for _, ent := range page.InstanceStatuses {
414                                 az[*ent.InstanceId] = *ent.AvailabilityZone
415                         }
416                         return true
417                 })
418                 if err != nil {
419                         instanceSet.logger.Warnf("error getting instance statuses: %s", err)
420                 }
421                 for _, inst := range instances {
422                         inst := inst.(*ec2Instance)
423                         inst.availabilityZone = az[*inst.instance.InstanceId]
424                 }
425                 instanceSet.updateSpotPrices(instances)
426         }
427
428         // Count instances in each subnet, and report in metrics.
429         subnetInstances := map[string]int{"": 0}
430         for _, subnet := range instanceSet.ec2config.SubnetID {
431                 subnetInstances[subnet] = 0
432         }
433         for _, inst := range instances {
434                 subnet := inst.(*ec2Instance).instance.SubnetId
435                 if subnet != nil {
436                         subnetInstances[*subnet]++
437                 } else {
438                         subnetInstances[""]++
439                 }
440         }
441         for subnet, count := range subnetInstances {
442                 instanceSet.mInstances.WithLabelValues(subnet).Set(float64(count))
443         }
444
445         return instances, err
446 }
447
448 type priceKey struct {
449         instanceType     string
450         spot             bool
451         availabilityZone string
452 }
453
454 // Refresh recent spot instance pricing data for the given instances,
455 // unless we already have recent pricing data for all relevant types.
456 func (instanceSet *ec2InstanceSet) updateSpotPrices(instances []cloud.Instance) {
457         if len(instances) == 0 {
458                 return
459         }
460
461         instanceSet.pricesLock.Lock()
462         defer instanceSet.pricesLock.Unlock()
463         if instanceSet.prices == nil {
464                 instanceSet.prices = map[priceKey][]cloud.InstancePrice{}
465                 instanceSet.pricesUpdated = map[priceKey]time.Time{}
466         }
467
468         updateTime := time.Now()
469         staleTime := updateTime.Add(-instanceSet.ec2config.SpotPriceUpdateInterval.Duration())
470         needUpdate := false
471         allTypes := map[string]bool{}
472
473         for _, inst := range instances {
474                 ec2inst := inst.(*ec2Instance).instance
475                 if aws.StringValue(ec2inst.InstanceLifecycle) == "spot" {
476                         pk := priceKey{
477                                 instanceType:     *ec2inst.InstanceType,
478                                 spot:             true,
479                                 availabilityZone: inst.(*ec2Instance).availabilityZone,
480                         }
481                         if instanceSet.pricesUpdated[pk].Before(staleTime) {
482                                 needUpdate = true
483                         }
484                         allTypes[*ec2inst.InstanceType] = true
485                 }
486         }
487         if !needUpdate {
488                 return
489         }
490         var typeFilterValues []*string
491         for instanceType := range allTypes {
492                 typeFilterValues = append(typeFilterValues, aws.String(instanceType))
493         }
494         // Get 3x update interval worth of pricing data. (Ideally the
495         // AWS API would tell us "we have shown you all of the price
496         // changes up to time T", but it doesn't, so we'll just ask
497         // for 3 intervals worth of data on each update, de-duplicate
498         // the data points, and not worry too much about occasionally
499         // missing some data points when our lookups fail twice in a
500         // row.
501         dsphi := &ec2.DescribeSpotPriceHistoryInput{
502                 StartTime: aws.Time(updateTime.Add(-3 * instanceSet.ec2config.SpotPriceUpdateInterval.Duration())),
503                 Filters: []*ec2.Filter{
504                         &ec2.Filter{Name: aws.String("instance-type"), Values: typeFilterValues},
505                         &ec2.Filter{Name: aws.String("product-description"), Values: []*string{aws.String("Linux/UNIX")}},
506                 },
507         }
508         err := instanceSet.client.DescribeSpotPriceHistoryPages(dsphi, func(page *ec2.DescribeSpotPriceHistoryOutput, lastPage bool) bool {
509                 for _, ent := range page.SpotPriceHistory {
510                         if ent.InstanceType == nil || ent.SpotPrice == nil || ent.Timestamp == nil {
511                                 // bogus record?
512                                 continue
513                         }
514                         price, err := strconv.ParseFloat(*ent.SpotPrice, 64)
515                         if err != nil {
516                                 // bogus record?
517                                 continue
518                         }
519                         pk := priceKey{
520                                 instanceType:     *ent.InstanceType,
521                                 spot:             true,
522                                 availabilityZone: *ent.AvailabilityZone,
523                         }
524                         instanceSet.prices[pk] = append(instanceSet.prices[pk], cloud.InstancePrice{
525                                 StartTime: *ent.Timestamp,
526                                 Price:     price,
527                         })
528                         instanceSet.pricesUpdated[pk] = updateTime
529                 }
530                 return true
531         })
532         if err != nil {
533                 instanceSet.logger.Warnf("error retrieving spot instance prices: %s", err)
534         }
535
536         expiredTime := updateTime.Add(-64 * instanceSet.ec2config.SpotPriceUpdateInterval.Duration())
537         for pk, last := range instanceSet.pricesUpdated {
538                 if last.Before(expiredTime) {
539                         delete(instanceSet.pricesUpdated, pk)
540                         delete(instanceSet.prices, pk)
541                 }
542         }
543         for pk, prices := range instanceSet.prices {
544                 instanceSet.prices[pk] = cloud.NormalizePriceHistory(prices)
545         }
546 }
547
548 func (instanceSet *ec2InstanceSet) Stop() {
549 }
550
551 type ec2Instance struct {
552         provider         *ec2InstanceSet
553         instance         *ec2.Instance
554         availabilityZone string // sometimes available for spot instances
555 }
556
557 func (inst *ec2Instance) ID() cloud.InstanceID {
558         return cloud.InstanceID(*inst.instance.InstanceId)
559 }
560
561 func (inst *ec2Instance) String() string {
562         return *inst.instance.InstanceId
563 }
564
565 func (inst *ec2Instance) ProviderType() string {
566         return *inst.instance.InstanceType
567 }
568
569 func (inst *ec2Instance) SetTags(newTags cloud.InstanceTags) error {
570         var ec2tags []*ec2.Tag
571         for k, v := range newTags {
572                 ec2tags = append(ec2tags, &ec2.Tag{
573                         Key:   aws.String(k),
574                         Value: aws.String(v),
575                 })
576         }
577
578         _, err := inst.provider.client.CreateTags(&ec2.CreateTagsInput{
579                 Resources: []*string{inst.instance.InstanceId},
580                 Tags:      ec2tags,
581         })
582
583         return err
584 }
585
586 func (inst *ec2Instance) Tags() cloud.InstanceTags {
587         tags := make(map[string]string)
588
589         for _, t := range inst.instance.Tags {
590                 tags[*t.Key] = *t.Value
591         }
592
593         return tags
594 }
595
596 func (inst *ec2Instance) Destroy() error {
597         _, err := inst.provider.client.TerminateInstances(&ec2.TerminateInstancesInput{
598                 InstanceIds: []*string{inst.instance.InstanceId},
599         })
600         return err
601 }
602
603 func (inst *ec2Instance) Address() string {
604         if inst.instance.PrivateIpAddress != nil {
605                 return *inst.instance.PrivateIpAddress
606         }
607         return ""
608 }
609
610 func (inst *ec2Instance) RemoteUser() string {
611         return inst.provider.ec2config.AdminUsername
612 }
613
614 func (inst *ec2Instance) VerifyHostKey(ssh.PublicKey, *ssh.Client) error {
615         return cloud.ErrNotImplemented
616 }
617
618 // PriceHistory returns the price history for this specific instance.
619 //
620 // AWS documentation is elusive about whether the hourly cost of a
621 // given spot instance changes as the current spot price changes for
622 // the corresponding instance type and availability zone. Our
623 // implementation assumes the answer is yes, based on the following
624 // hints.
625 //
626 // https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/spot-requests.html
627 // says: "After your Spot Instance is running, if the Spot price rises
628 // above your maximum price, Amazon EC2 interrupts your Spot
629 // Instance." (This doesn't address what happens when the spot price
630 // rises *without* exceeding your maximum price.)
631 //
632 // https://docs.aws.amazon.com/whitepapers/latest/cost-optimization-leveraging-ec2-spot-instances/how-spot-instances-work.html
633 // says: "You pay the Spot price that's in effect, billed to the
634 // nearest second." (But it's not explicitly stated whether "the price
635 // in effect" changes over time for a given instance.)
636 //
637 // The same page also says, in a discussion about the effect of
638 // specifying a maximum price: "Note that you never pay more than the
639 // Spot price that is in effect when your Spot Instance is running."
640 // (The use of the phrase "is running", as opposed to "was launched",
641 // hints that pricing is dynamic.)
642 func (inst *ec2Instance) PriceHistory(instType arvados.InstanceType) []cloud.InstancePrice {
643         inst.provider.pricesLock.Lock()
644         defer inst.provider.pricesLock.Unlock()
645         // Note updateSpotPrices currently populates
646         // inst.provider.prices only for spot instances, so if
647         // spot==false here, we will return no data.
648         pk := priceKey{
649                 instanceType:     *inst.instance.InstanceType,
650                 spot:             aws.StringValue(inst.instance.InstanceLifecycle) == "spot",
651                 availabilityZone: inst.availabilityZone,
652         }
653         var prices []cloud.InstancePrice
654         for _, price := range inst.provider.prices[pk] {
655                 // ceil(added scratch space in GiB)
656                 gib := (instType.AddedScratch + 1<<30 - 1) >> 30
657                 monthly := inst.provider.ec2config.EBSPrice * float64(gib)
658                 hourly := monthly / 30 / 24
659                 price.Price += hourly
660                 prices = append(prices, price)
661         }
662         return prices
663 }
664
665 type rateLimitError struct {
666         error
667         earliestRetry time.Time
668 }
669
670 func (err rateLimitError) EarliestRetry() time.Time {
671         return err.earliestRetry
672 }
673
674 type capacityError struct {
675         error
676         isInstanceTypeSpecific bool
677 }
678
679 func (er *capacityError) IsCapacityError() bool {
680         return true
681 }
682
683 func (er *capacityError) IsInstanceTypeSpecific() bool {
684         return er.isInstanceTypeSpecific
685 }
686
687 var isCodeQuota = map[string]bool{
688         "InstanceLimitExceeded":             true,
689         "InsufficientAddressCapacity":       true,
690         "InsufficientFreeAddressesInSubnet": true,
691         "InsufficientVolumeCapacity":        true,
692         "MaxSpotInstanceCountExceeded":      true,
693         "VcpuLimitExceeded":                 true,
694 }
695
696 // isErrorQuota returns whether the error indicates we have reached
697 // some usage quota/limit -- i.e., immediately retrying with an equal
698 // or larger instance type will probably not work.
699 //
700 // Returns false if error is nil.
701 func isErrorQuota(err error) bool {
702         if aerr, ok := err.(awserr.Error); ok && aerr != nil {
703                 if _, ok := isCodeQuota[aerr.Code()]; ok {
704                         return true
705                 }
706         }
707         return false
708 }
709
710 // isErrorSubnetSpecific returns true if the problem encountered by
711 // RunInstances might be avoided by trying a different subnet.
712 func isErrorSubnetSpecific(err error) bool {
713         aerr, ok := err.(awserr.Error)
714         if !ok {
715                 return false
716         }
717         code := aerr.Code()
718         return strings.Contains(code, "Subnet") ||
719                 code == "InsufficientInstanceCapacity" ||
720                 code == "InsufficientVolumeCapacity" ||
721                 code == "Unsupported"
722 }
723
724 // isErrorCapacity returns true if the error indicates lack of
725 // capacity (either temporary or permanent) to run a specific instance
726 // type -- i.e., retrying with a different instance type might
727 // succeed.
728 func isErrorCapacity(err error) bool {
729         aerr, ok := err.(awserr.Error)
730         if !ok {
731                 return false
732         }
733         code := aerr.Code()
734         return code == "InsufficientInstanceCapacity" ||
735                 (code == "Unsupported" && strings.Contains(aerr.Message(), "requested instance type"))
736 }
737
738 type ec2QuotaError struct {
739         error
740 }
741
742 func (er *ec2QuotaError) IsQuotaError() bool {
743         return true
744 }
745
746 func wrapError(err error, throttleValue *atomic.Value) error {
747         if request.IsErrorThrottle(err) {
748                 // Back off exponentially until an upstream call
749                 // either succeeds or returns a non-throttle error.
750                 d, _ := throttleValue.Load().(time.Duration)
751                 d = d*3/2 + time.Second
752                 if d < throttleDelayMin {
753                         d = throttleDelayMin
754                 } else if d > throttleDelayMax {
755                         d = throttleDelayMax
756                 }
757                 throttleValue.Store(d)
758                 return rateLimitError{error: err, earliestRetry: time.Now().Add(d)}
759         } else if isErrorQuota(err) {
760                 return &ec2QuotaError{err}
761         } else if isErrorCapacity(err) {
762                 return &capacityError{err, true}
763         } else if err != nil {
764                 throttleValue.Store(time.Duration(0))
765                 return err
766         }
767         throttleValue.Store(time.Duration(0))
768         return nil
769 }
770
771 var boolLabelValue = map[bool]string{false: "0", true: "1"}