Merge branch '19907-oidc-cache-error'
[arvados.git] / lib / controller / localdb / login_oidc.go
index 61dc5c816b35661f39c4a800ab17f1bf55325f06..8a1b8fd82b0d6a6c1c25de742eb61c42b7680f4c 100644 (file)
@@ -14,8 +14,10 @@ import (
        "errors"
        "fmt"
        "io"
+       "net"
        "net/http"
        "net/url"
+       "regexp"
        "strings"
        "sync"
        "text/template"
@@ -28,9 +30,10 @@ import (
        "git.arvados.org/arvados.git/sdk/go/auth"
        "git.arvados.org/arvados.git/sdk/go/ctxlog"
        "git.arvados.org/arvados.git/sdk/go/httpserver"
-       "github.com/coreos/go-oidc"
+       "github.com/coreos/go-oidc/v3/oidc"
        lru "github.com/hashicorp/golang-lru"
        "github.com/jmoiron/sqlx"
+       "github.com/lib/pq"
        "github.com/sirupsen/logrus"
        "golang.org/x/oauth2"
        "google.golang.org/api/option"
@@ -43,6 +46,7 @@ var (
        tokenCacheNegativeTTL = time.Minute * 5
        tokenCacheTTL         = time.Minute * 10
        tokenCacheRaceWindow  = time.Minute
+       pqCodeUniqueViolation = pq.ErrorCode("23505")
 )
 
 type oidcLoginController struct {
@@ -114,6 +118,9 @@ func (ctrl *oidcLoginController) Login(ctx context.Context, opts arvados.LoginOp
                if opts.ReturnTo == "" {
                        return loginError(errors.New("missing return_to parameter"))
                }
+               if err := validateLoginRedirectTarget(ctrl.Parent.cluster, opts.ReturnTo); err != nil {
+                       return loginError(fmt.Errorf("invalid return_to parameter: %s", err))
+               }
                state := ctrl.newOAuth2State([]byte(ctrl.Cluster.SystemRootToken), opts.Remote, opts.ReturnTo)
                var authparams []oauth2.AuthCodeOption
                for k, v := range ctrl.AuthParams {
@@ -177,12 +184,19 @@ func (ctrl *oidcLoginController) getAuthInfo(ctx context.Context, token *oauth2.
        } else if verified, _ := claims[ctrl.EmailVerifiedClaim].(bool); verified || ctrl.EmailVerifiedClaim == "" {
                // Fall back to this info if the People API call
                // (below) doesn't return a primary && verified email.
-               name, _ := claims["name"].(string)
-               if names := strings.Fields(strings.TrimSpace(name)); len(names) > 1 {
-                       ret.FirstName = strings.Join(names[0:len(names)-1], " ")
-                       ret.LastName = names[len(names)-1]
-               } else if len(names) > 0 {
-                       ret.FirstName = names[0]
+               givenName, _ := claims["given_name"].(string)
+               familyName, _ := claims["family_name"].(string)
+               if givenName != "" && familyName != "" {
+                       ret.FirstName = givenName
+                       ret.LastName = familyName
+               } else {
+                       name, _ := claims["name"].(string)
+                       if names := strings.Fields(strings.TrimSpace(name)); len(names) > 1 {
+                               ret.FirstName = strings.Join(names[0:len(names)-1], " ")
+                               ret.LastName = names[len(names)-1]
+                       } else if len(names) > 0 {
+                               ret.FirstName = names[0]
+                       }
                }
                ret.Email, _ = claims[ctrl.EmailClaim].(string)
        }
@@ -381,6 +395,9 @@ func (ta *oidcTokenAuthorizer) WrapCalls(origFunc api.RoutableFunc) api.Routable
        }
 }
 
+// Matches error from oidc UserInfo() when receiving HTTP status 5xx
+var re5xxError = regexp.MustCompile(`^5\d\d `)
+
 // registerToken checks whether tok is a valid OIDC Access Token and,
 // if so, ensures that an api_client_authorizations row exists so that
 // RailsAPI will accept it as an Arvados token.
@@ -401,11 +418,8 @@ func (ta *oidcTokenAuthorizer) registerToken(ctx context.Context, tok string) er
                // cached positive result
                aca := cached.(arvados.APIClientAuthorization)
                var expiring bool
-               if aca.ExpiresAt != "" {
-                       t, err := time.Parse(time.RFC3339Nano, aca.ExpiresAt)
-                       if err != nil {
-                               return fmt.Errorf("error parsing expires_at value: %w", err)
-                       }
+               if !aca.ExpiresAt.IsZero() {
+                       t := aca.ExpiresAt
                        expiring = t.Before(time.Now().Add(time.Minute))
                }
                if !expiring {
@@ -453,6 +467,7 @@ func (ta *oidcTokenAuthorizer) registerToken(ctx context.Context, tok string) er
                return fmt.Errorf("error setting up OpenID Connect provider: %s", err)
        }
        if ok, err := ta.checkAccessTokenScope(ctx, tok); err != nil || !ok {
+               // Note checkAccessTokenScope logs any interesting errors
                ta.cache.Add(tok, time.Now().Add(tokenCacheNegativeTTL))
                return err
        }
@@ -461,6 +476,21 @@ func (ta *oidcTokenAuthorizer) registerToken(ctx context.Context, tok string) er
        }
        userinfo, err := ta.ctrl.provider.UserInfo(ctx, oauth2.StaticTokenSource(oauth2Token))
        if err != nil {
+               if neterr := net.Error(nil); errors.As(err, &neterr) || re5xxError.MatchString(err.Error()) {
+                       // If this token is in fact a valid OIDC
+                       // token, but we failed to validate it here
+                       // because of a network problem or internal
+                       // server error, we error out now with a 5xx
+                       // error, indicating to the client that they
+                       // can try again.  If we didn't error out now,
+                       // the unrecognized token would eventually
+                       // cause a 401 error further down the stack,
+                       // which the caller would interpret as an
+                       // unrecoverable failure.
+                       ctxlog.FromContext(ctx).WithError(err).Debugf("treating OIDC UserInfo lookup error type %T as transient; failing request instead of forwarding token blindly", err)
+                       return err
+               }
+               ctxlog.FromContext(ctx).WithError(err).WithField("HMAC", hmac).Debug("UserInfo failed (not an OIDC token?), caching negative result")
                ta.cache.Add(tok, time.Now().Add(tokenCacheNegativeTTL))
                return nil
        }
@@ -475,7 +505,6 @@ func (ta *oidcTokenAuthorizer) registerToken(ctx context.Context, tok string) er
        // it's expiring.
        exp := time.Now().UTC().Add(tokenCacheTTL + tokenCacheRaceWindow)
 
-       var aca arvados.APIClientAuthorization
        if updating {
                _, err = tx.ExecContext(ctx, `update api_client_authorizations set expires_at=$1 where api_token=$2`, exp, hmac)
                if err != nil {
@@ -483,23 +512,44 @@ func (ta *oidcTokenAuthorizer) registerToken(ctx context.Context, tok string) er
                }
                ctxlog.FromContext(ctx).WithField("HMAC", hmac).Debug("(*oidcTokenAuthorizer)registerToken: updated api_client_authorizations row")
        } else {
-               aca, err = ta.ctrl.Parent.CreateAPIClientAuthorization(ctx, ta.ctrl.Cluster.SystemRootToken, *authinfo)
+               aca, err := ta.ctrl.Parent.CreateAPIClientAuthorization(ctx, ta.ctrl.Cluster.SystemRootToken, *authinfo)
                if err != nil {
                        return err
                }
-               _, err = tx.ExecContext(ctx, `update api_client_authorizations set api_token=$1, expires_at=$2 where uuid=$3`, hmac, exp, aca.UUID)
+               _, err = tx.ExecContext(ctx, `savepoint upd`)
                if err != nil {
+                       return err
+               }
+               _, err = tx.ExecContext(ctx, `update api_client_authorizations set api_token=$1, expires_at=$2 where uuid=$3`, hmac, exp, aca.UUID)
+               if e, ok := err.(*pq.Error); ok && e.Code == pqCodeUniqueViolation {
+                       // unique_violation, given that the above
+                       // query did not find a row with matching
+                       // api_token, means another thread/process
+                       // also received this same token and won the
+                       // race to insert it -- in which case this
+                       // thread doesn't need to update the database.
+                       // Discard the redundant row.
+                       _, err = tx.ExecContext(ctx, `rollback to savepoint upd`)
+                       if err != nil {
+                               return err
+                       }
+                       _, err = tx.ExecContext(ctx, `delete from api_client_authorizations where uuid=$1`, aca.UUID)
+                       if err != nil {
+                               return err
+                       }
+                       ctxlog.FromContext(ctx).WithField("HMAC", hmac).Debug("(*oidcTokenAuthorizer)registerToken: api_client_authorizations row inserted by another thread")
+               } else if err != nil {
+                       ctxlog.FromContext(ctx).Errorf("%#v", err)
                        return fmt.Errorf("error adding OIDC access token to database: %w", err)
+               } else {
+                       ctxlog.FromContext(ctx).WithFields(logrus.Fields{"UUID": aca.UUID, "HMAC": hmac}).Debug("(*oidcTokenAuthorizer)registerToken: inserted api_client_authorizations row")
                }
-               aca.APIToken = hmac
-               ctxlog.FromContext(ctx).WithFields(logrus.Fields{"UUID": aca.UUID, "HMAC": hmac}).Debug("(*oidcTokenAuthorizer)registerToken: inserted api_client_authorizations row")
        }
        err = tx.Commit()
        if err != nil {
                return err
        }
-       aca.ExpiresAt = exp.Format(time.RFC3339Nano)
-       ta.cache.Add(tok, aca)
+       ta.cache.Add(tok, arvados.APIClientAuthorization{ExpiresAt: exp})
        return nil
 }
 
@@ -534,6 +584,6 @@ func (ta *oidcTokenAuthorizer) checkAccessTokenScope(ctx context.Context, tok st
                        return true, nil
                }
        }
-       ctxlog.FromContext(ctx).WithFields(logrus.Fields{"have": claims.Scope, "need": ta.ctrl.AcceptAccessTokenScope}).Infof("unacceptable access token scope")
+       ctxlog.FromContext(ctx).WithFields(logrus.Fields{"have": claims.Scope, "need": ta.ctrl.AcceptAccessTokenScope}).Info("unacceptable access token scope")
        return false, httpserver.ErrorWithStatus(errors.New("unacceptable access token scope"), http.StatusUnauthorized)
 }