boulder/ctpolicy/ctpolicy.go

391 lines
13 KiB
Go

package ctpolicy
import (
"context"
"errors"
"fmt"
"math/rand"
"strings"
"time"
"github.com/letsencrypt/boulder/canceled"
"github.com/letsencrypt/boulder/core"
"github.com/letsencrypt/boulder/ctpolicy/ctconfig"
"github.com/letsencrypt/boulder/ctpolicy/loglist"
berrors "github.com/letsencrypt/boulder/errors"
blog "github.com/letsencrypt/boulder/log"
pubpb "github.com/letsencrypt/boulder/publisher/proto"
"github.com/prometheus/client_golang/prometheus"
)
// CTPolicy is used to hold information about SCTs required from various
// groupings
type CTPolicy struct {
pub pubpb.PublisherClient
// TODO(#5938): Remove groups, informational, and final
groups []ctconfig.CTGroup
informational []ctconfig.LogDescription
final []ctconfig.LogDescription
sctLogs loglist.List
infoLogs loglist.List
finalLogs loglist.List
stagger time.Duration
log blog.Logger
winnerCounter *prometheus.CounterVec
}
// New creates a new CTPolicy struct
func New(
pub pubpb.PublisherClient,
groups []ctconfig.CTGroup,
informational []ctconfig.LogDescription,
sctLogs loglist.List,
infoLogs loglist.List,
finalLogs loglist.List,
stagger time.Duration,
log blog.Logger,
stats prometheus.Registerer,
) *CTPolicy {
var final []ctconfig.LogDescription
for _, group := range groups {
for _, log := range group.Logs {
if log.SubmitFinalCert {
final = append(final, log)
}
}
}
for _, log := range informational {
if log.SubmitFinalCert {
final = append(final, log)
}
}
winnerCounter := prometheus.NewCounterVec(
prometheus.CounterOpts{
Name: "sct_race_winner",
Help: "Counter of logs that win SCT submission races.",
},
[]string{"log", "group"},
)
stats.MustRegister(winnerCounter)
return &CTPolicy{
pub: pub,
groups: groups,
informational: informational,
final: final,
sctLogs: sctLogs,
infoLogs: infoLogs,
finalLogs: finalLogs,
stagger: stagger,
log: log,
winnerCounter: winnerCounter,
}
}
type result struct {
sct []byte
log string
err error
}
// race submits an SCT to each log in a group and waits for the first response back,
// once it has the first SCT it cancels all of the other submissions and returns.
// It allows up to len(group)-1 of the submissions to fail as we only care about
// getting a single SCT.
// TODO(#5938): Remove this when it becomes dead code.
func (ctp *CTPolicy) race(ctx context.Context, cert core.CertDER, group ctconfig.CTGroup, expiration time.Time) ([]byte, error) {
results := make(chan result, len(group.Logs))
isPrecert := true
// Randomize the order in which we send requests to the logs in a group
// so we maximize the distribution of logs we get SCTs from.
for i, logNum := range rand.Perm(len(group.Logs)) {
ld := group.Logs[logNum]
go func(i int, ld ctconfig.LogDescription) {
// Each submission waits a bit longer than the previous one, to give the
// previous log a chance to reply. If the context is already done by the
// time we get here, don't bother submitting. That generally means the
// context was canceled because another log returned a success already.
time.Sleep(time.Duration(i) * group.Stagger.Duration)
if ctx.Err() != nil {
return
}
uri, key, err := ld.Info(expiration)
if err != nil {
ctp.log.Errf("unable to get log info: %s", err)
return
}
sct, err := ctp.pub.SubmitToSingleCTWithResult(ctx, &pubpb.Request{
LogURL: uri,
LogPublicKey: key,
Der: cert,
Precert: isPrecert,
})
if err != nil {
// Only log the error if it is not a result of the context being canceled
if !canceled.Is(err) {
ctp.log.Warningf("ct submission to %q failed: %s", uri, err)
}
results <- result{err: err}
return
}
results <- result{sct: sct.Sct, log: uri}
}(i, ld)
}
for i := 0; i < len(group.Logs); i++ {
select {
case <-ctx.Done():
ctp.winnerCounter.With(prometheus.Labels{"log": "timeout", "group": group.Name}).Inc()
return nil, ctx.Err()
case res := <-results:
if res.sct != nil {
ctp.winnerCounter.With(prometheus.Labels{"log": res.log, "group": group.Name}).Inc()
// Return the very first SCT we get back. Returning triggers
// the defer'd context cancellation method.
return res.sct, nil
}
// We will continue waiting for an SCT until we've seen the same number
// of errors as there are logs in the group as we may still get a SCT
// back from another log.
}
}
ctp.winnerCounter.With(prometheus.Labels{"log": "all_failed", "group": group.Name}).Inc()
return nil, errors.New("all submissions failed")
}
// GetSCTs attempts to retrieve two SCTs from the configured log groups and
// returns the set of SCTs to the caller.
func (ctp *CTPolicy) GetSCTs(ctx context.Context, cert core.CertDER, expiration time.Time) (core.SCTDERs, error) {
if len(ctp.sctLogs) != 0 {
return ctp.getOperatorSCTs(ctx, cert, expiration)
}
return ctp.getGoogleSCTs(ctx, cert, expiration)
}
// getGoogleSCTs retrieves exactly one SCT from each of the configured log
// groups. It expects that there are exactly 2 such groups, and that one of
// those groups contains only logs operated by Google. As such, it enforces
// Google's *old* CT Policy, which required that certs have two SCTs, one of
// which was from a Google log.
// DEPRECATED: Google no longer enforces the "one Google, one non-Google" log
// policy. Use getOperatorSCTs instead.
// TODO(#5938): Remove this after the configured groups have been rearranged.
func (ctp *CTPolicy) getGoogleSCTs(ctx context.Context, cert core.CertDER, expiration time.Time) (core.SCTDERs, error) {
results := make(chan result, len(ctp.groups))
subCtx, cancel := context.WithCancel(ctx)
defer cancel()
for i, g := range ctp.groups {
go func(i int, g ctconfig.CTGroup) {
sct, err := ctp.race(subCtx, cert, g, expiration)
// Only one of these will be non-nil
if err != nil {
results <- result{err: berrors.MissingSCTsError("CT log group %q: %s", g.Name, err)}
}
results <- result{sct: sct}
}(i, g)
}
go ctp.submitPrecertInformational(cert, expiration)
var ret core.SCTDERs
for i := 0; i < len(ctp.groups); i++ {
res := <-results
// If any one group fails to get a SCT then we fail out immediately
// cancel any other in progress work as we can't continue
if res.err != nil {
// Returning triggers the defer'd context cancellation method
return nil, res.err
}
ret = append(ret, res.sct)
}
return ret, nil
}
// getOperatorSCTs retrieves exactly two SCTs from the total collection of
// configured log groups, with at most one SCT coming from each group. It
// expects that all logs run by a single operator (e.g. Google) are in the same
// group, to guarantee that SCTs from logs in different groups do not end up
// coming from the same operator. As such, it enforces Google's current CT
// Policy, which requires that certs have two SCTs from logs run by different
// operators.
// TODO(#5938): Inline this into GetSCTs when getGoogleSCTs is removed.
func (ctp *CTPolicy) getOperatorSCTs(ctx context.Context, cert core.CertDER, expiration time.Time) (core.SCTDERs, error) {
// We'll cancel this sub-context when we have the two SCTs we need, to cause
// any other ongoing submission attempts to quit.
subCtx, cancel := context.WithCancel(ctx)
defer cancel()
// This closure will be called in parallel once for each operator group.
getOne := func(i int, g string) ([]byte, error) {
// Sleep a little bit to stagger our requests to the later groups. Use `i-1`
// to compute the stagger duration so that the first two groups (indices 0
// and 1) get negative or zero (i.e. instant) sleep durations. If the
// context gets cancelled (most likely because two logs from other operator
// groups returned SCTs already) before the sleep is complete, quit instead.
select {
case <-subCtx.Done():
return nil, subCtx.Err()
case <-time.After(time.Duration(i-1) * ctp.stagger):
}
// Pick a random log from among those in the group. In practice, very few
// operator groups have more than one log, so this loses little flexibility.
uri, key, err := ctp.sctLogs.PickOne(g, expiration)
if err != nil {
return nil, fmt.Errorf("unable to get log info: %w", err)
}
sct, err := ctp.pub.SubmitToSingleCTWithResult(ctx, &pubpb.Request{
LogURL: uri,
LogPublicKey: key,
Der: cert,
Precert: true,
})
if err != nil {
return nil, fmt.Errorf("ct submission to %q (%q) failed: %w", g, uri, err)
}
return sct.Sct, nil
}
// Ensure that this channel has a buffer equal to the number of goroutines
// we're kicking off, so that they're all guaranteed to be able to write to
// it and exit without blocking and leaking.
results := make(chan result, len(ctp.sctLogs))
// Kick off a collection of goroutines to try to submit the precert to each
// log operator group. Randomize the order of the groups so that we're not
// always trying to submit to the same two operators.
for i, group := range ctp.sctLogs.Permute() {
go func(i int, g string) {
sctDER, err := getOne(i, g)
results <- result{sct: sctDER, err: err}
}(i, group)
}
go ctp.submitPrecertInformational(cert, expiration)
// Finally, collect SCTs and/or errors from our results channel.
scts := make(core.SCTDERs, 0)
errs := make([]string, 0)
for i := 0; i < len(ctp.sctLogs); i++ {
select {
case <-ctx.Done():
// We timed out (the calling function returned and canceled our context)
// before getting two SCTs.
return nil, berrors.MissingSCTsError("failed to get 2 SCTs before ctx finished: %s", ctx.Err())
case res := <-results:
if res.err != nil {
errs = append(errs, res.err.Error())
continue
}
scts = append(scts, res.sct)
if len(scts) >= 2 {
return scts, nil
}
}
}
// If we made it to the end of that loop, that means we never got two SCTs
// to return. Error out instead.
return nil, berrors.MissingSCTsError("failed to get 2 SCTs, got error(s): %s", strings.Join(errs, "; "))
}
// submitAllBestEffort submits the given certificate or precertificate to every
// log ("informational" for precerts, "final" for certs) configured in the policy.
// It neither waits for these submission to complete, nor tracks their success.
func (ctp *CTPolicy) submitAllBestEffort(blob core.CertDER, precert bool, expiry time.Time) {
logs := ctp.finalLogs
if precert {
logs = ctp.infoLogs
}
for _, group := range logs {
for _, log := range group {
if log.StartInclusive.After(expiry) || log.EndExclusive.Equal(expiry) || log.EndExclusive.Before(expiry) {
continue
}
go func(log loglist.Log) {
_, err := ctp.pub.SubmitToSingleCTWithResult(
context.Background(),
&pubpb.Request{
LogURL: log.Url,
LogPublicKey: log.Key,
Der: blob,
Precert: precert,
},
)
if err != nil {
ctp.log.Warningf("ct submission of cert to log %q failed: %s", log.Url, err)
}
}(log)
}
}
}
// submitPrecertInformational submits precertificates to any configured
// "informational" logs, but does not care about success or returned SCTs.
func (ctp *CTPolicy) submitPrecertInformational(cert core.CertDER, expiration time.Time) {
if len(ctp.sctLogs) != 0 {
ctp.submitAllBestEffort(cert, true, expiration)
return
}
// TODO(#5938): Remove this when it becomes dead code.
for _, log := range ctp.informational {
go func(l ctconfig.LogDescription) {
// We use a context.Background() here instead of a context from the parent
// because these submissions are running in a goroutine and we don't want
// them to be cancelled when the caller of CTPolicy.GetSCTs returns and
// cancels its RPC context.
uri, key, err := l.Info(expiration)
if err != nil {
ctp.log.Errf("unable to get log info: %s", err)
return
}
_, err = ctp.pub.SubmitToSingleCTWithResult(context.Background(), &pubpb.Request{
LogURL: uri,
LogPublicKey: key,
Der: cert,
Precert: true,
})
if err != nil {
ctp.log.Warningf("ct submission to informational log %q failed: %s", uri, err)
}
}(log)
}
}
// SubmitFinalCert submits finalized certificates created from precertificates
// to any configured "final" logs, but does not care about success.
func (ctp *CTPolicy) SubmitFinalCert(cert core.CertDER, expiration time.Time) {
if len(ctp.sctLogs) != 0 {
ctp.submitAllBestEffort(cert, false, expiration)
return
}
// TODO(#5938): Remove this when it becomes dead code.
for _, log := range ctp.final {
go func(l ctconfig.LogDescription) {
uri, key, err := l.Info(expiration)
if err != nil {
ctp.log.Errf("unable to get log info: %s", err)
return
}
_, err = ctp.pub.SubmitToSingleCTWithResult(context.Background(), &pubpb.Request{
LogURL: uri,
LogPublicKey: key,
Der: cert,
Precert: false,
})
if err != nil {
ctp.log.Warningf("ct submission of final cert to log %q failed: %s", uri, err)
}
}(log)
}
}