open-nomad/nomad/acl.go
Tim Gross 1cf28996e7 acl: prevent privilege escalation via workload identity
ACL policies can be associated with a job so that the job's Workload Identity
can have expanded access to other policy objects, including other
variables. Policies set on the variables the job automatically has access to
were ignored, but this includes policies with `deny` capabilities.

Additionally, when resolving claims for a workload identity without any attached
policies, the `ResolveClaims` method returned a `nil` ACL object, which is
treated similarly to a management token. While this was safe in Nomad 1.4.x,
when the workload identity token was exposed to the task via the `identity`
block, this allows a user with `submit-job` capabilities to escalate their
privileges.

We originally implemented automatic workload access to Variables as a separate
code path in the Variables RPC endpoint so that we don't have to generate
on-the-fly policies that blow up the ACL policy cache. This is fairly brittle
but also the behavior around wildcard paths in policies different from the rest
of our ACL polices, which is hard to reason about.

Add an `ACLClaim` parameter to the `AllowVariableOperation` method so that we
can push all this logic into the `acl` package and the behavior can be
consistent. This will allow a `deny` policy to override automatic access (and
probably speed up checks of non-automatic variable access).
2023-03-13 11:13:27 -04:00

460 lines
14 KiB
Go

package nomad
import (
"errors"
"fmt"
"net"
"time"
metrics "github.com/armon/go-metrics"
"github.com/hashicorp/nomad/acl"
"github.com/hashicorp/nomad/helper"
"github.com/hashicorp/nomad/nomad/state"
"github.com/hashicorp/nomad/nomad/structs"
)
// Authenticate extracts an AuthenticatedIdentity from the request context or
// provided token and sets the identity on the request. The caller can extract
// an acl.ACL, WorkloadIdentity, or other identifying tokens to use for
// authorization. Keeping these fields independent rather than merging them into
// an ephemeral ACLToken makes the original of the credential clear to RPC
// handlers, who may have different behavior for internal vs external origins.
//
// Note: when called on the follower we'll be making stale queries, so it's
// possible if the follower is behind that the leader will get a different value
// if an ACL token or allocation's WI has just been created.
//
// This method returns errors that are used for testing diagnostics. RPC callers
// should always return ErrPermissionDenied after checking forwarding when one
// of these errors is received.
func (s *Server) Authenticate(ctx *RPCContext, args structs.RequestWithIdentity) error {
// get the user ACLToken or anonymous token
secretID := args.GetAuthToken()
aclToken, err := s.ResolveSecretToken(secretID)
switch {
case err == nil:
// If ACLs are disabled or we have a non-anonymous token, return that.
if aclToken == nil || aclToken != structs.AnonymousACLToken {
args.SetIdentity(&structs.AuthenticatedIdentity{ACLToken: aclToken})
return nil
}
case errors.Is(err, structs.ErrTokenExpired):
return err
case errors.Is(err, structs.ErrTokenInvalid):
// if it's not a UUID it might be an identity claim
claims, err := s.VerifyClaim(secretID)
if err != nil {
// we already know the token wasn't valid for an ACL in the state
// store, so if we get an error at this point we have an invalid
// token and there are no other options but to bail out
return err
}
args.SetIdentity(&structs.AuthenticatedIdentity{Claims: claims})
return nil
case errors.Is(err, structs.ErrTokenNotFound):
// Check if the secret ID is the leader's secret ID, in which case treat
// it as a management token.
leaderAcl := s.getLeaderAcl()
if leaderAcl != "" && secretID == leaderAcl {
aclToken = structs.LeaderACLToken
break
} else {
// Otherwise, see if the secret ID belongs to a node. We should
// reach this point only on first connection.
node, err := s.State().NodeBySecretID(nil, secretID)
if err != nil {
// this is a go-memdb error; shouldn't happen
return fmt.Errorf("could not resolve node secret: %w", err)
}
if node != nil {
args.SetIdentity(&structs.AuthenticatedIdentity{ClientID: node.ID})
return nil
}
}
// we were passed a bogus token so we'll return an error, but we'll also
// want to capture the IP for metrics
remoteIP, err := s.remoteIPFromRPCContext(ctx)
if err != nil {
s.logger.Error("could not determine remote address", "error", err)
}
args.SetIdentity(&structs.AuthenticatedIdentity{RemoteIP: remoteIP})
return structs.ErrPermissionDenied
default: // any other error
return fmt.Errorf("could not resolve user: %w", err)
}
// If there's no context we're in a "static" handler which only happens for
// cases where the leader is making RPCs internally (volumewatcher and
// deploymentwatcher)
if ctx == nil {
args.SetIdentity(&structs.AuthenticatedIdentity{ACLToken: aclToken})
return nil
}
// At this point we either have an anonymous token or an invalid one.
// Previously-connected clients will have a NodeID set on the context, which
// is available for all yamux streams over the same yamux session (and TCP
// connection). This will be a large portion of the RPCs sent, but we can't
// fast-path this at the top of the method, because authenticated HTTP
// requests to the clients will come in over to the same session.
if ctx.NodeID != "" {
args.SetIdentity(&structs.AuthenticatedIdentity{ClientID: ctx.NodeID})
return nil
}
// Unlike clients that provide their Node ID on first connection, server
// RPCs don't include an ID for the server so we identify servers by cert
// and IP address.
identity := &structs.AuthenticatedIdentity{ACLToken: aclToken}
if ctx.TLS {
identity.TLSName = ctx.Certificate().Subject.CommonName
}
remoteIP, err := s.remoteIPFromRPCContext(ctx)
if err != nil {
s.logger.Error(
"could not authenticate RPC request or determine remote address", "error", err)
return err
}
identity.RemoteIP = remoteIP
args.SetIdentity(identity)
return nil
}
func (s *Server) remoteIPFromRPCContext(ctx *RPCContext) (net.IP, error) {
var remoteAddr *net.TCPAddr
var ok bool
if ctx == nil {
return nil, nil
}
if ctx.Session != nil {
remoteAddr, ok = ctx.Session.RemoteAddr().(*net.TCPAddr)
if !ok {
return nil, errors.New("session address was not a TCP address")
}
}
if remoteAddr == nil && ctx.Conn != nil {
remoteAddr, ok = ctx.Conn.RemoteAddr().(*net.TCPAddr)
if !ok {
return nil, errors.New("session address was not a TCP address")
}
}
if remoteAddr != nil {
return remoteAddr.IP, nil
}
return nil, structs.ErrPermissionDenied
}
// ResolveACL is an authentication wrapper which handles resolving both ACL
// tokens and Workload Identities. If both are provided the ACL token is
// preferred, but it is best for the RPC caller to only include the credentials
// for the identity they intend the operation to be performed with.
func (s *Server) ResolveACL(args structs.RequestWithIdentity) (*acl.ACL, error) {
identity := args.GetIdentity()
if !s.config.ACLEnabled || identity == nil {
return nil, nil
}
aclToken := identity.GetACLToken()
if aclToken != nil {
return s.ResolveACLForToken(aclToken)
}
claims := identity.GetClaims()
if claims != nil {
return s.ResolveClaims(claims)
}
return nil, nil
}
// ResolveACLForToken resolves an ACL from a token only. It should be used only
// by Variables endpoints, which have additional implicit policies for their
// claims so we can't wrap them up in ResolveACL.
//
// TODO: figure out a way to the Variables endpoint implicit policies baked into
// their acl.ACL object so that we can avoid using this method.
func (s *Server) ResolveACLForToken(aclToken *structs.ACLToken) (*acl.ACL, error) {
if !s.config.ACLEnabled {
return nil, nil
}
snap, err := s.fsm.State().Snapshot()
if err != nil {
return nil, err
}
return resolveACLFromToken(snap, s.aclCache, aclToken)
}
// ResolveClientOrACL resolves an ACL if the identity has a token or claim, and
// falls back to verifying the client ID if one has been set
func (s *Server) ResolveClientOrACL(args structs.RequestWithIdentity) (*acl.ACL, error) {
identity := args.GetIdentity()
if !s.config.ACLEnabled || identity == nil || identity.ClientID != "" {
return nil, nil
}
aclObj, err := s.ResolveACL(args)
if err != nil {
return nil, err
}
// Returns either the users aclObj, or nil if ACLs are disabled.
return aclObj, nil
}
// ResolveToken is used to translate an ACL Token Secret ID into
// an ACL object, nil if ACLs are disabled, or an error.
func (s *Server) ResolveToken(secretID string) (*acl.ACL, error) {
// Fast-path if ACLs are disabled
if !s.config.ACLEnabled {
return nil, nil
}
defer metrics.MeasureSince([]string{"nomad", "acl", "resolveToken"}, time.Now())
// Check if the secret ID is the leader secret ID, in which case treat it as
// a management token.
if leaderAcl := s.getLeaderAcl(); leaderAcl != "" && secretID == leaderAcl {
return acl.ManagementACL, nil
}
// Snapshot the state
snap, err := s.fsm.State().Snapshot()
if err != nil {
return nil, err
}
// Resolve the ACL
return resolveTokenFromSnapshotCache(snap, s.aclCache, secretID)
}
// VerifyClaim asserts that the token is valid and that the resulting
// allocation ID belongs to a non-terminal allocation
func (s *Server) VerifyClaim(token string) (*structs.IdentityClaims, error) {
claims, err := s.encrypter.VerifyClaim(token)
if err != nil {
return nil, err
}
snap, err := s.fsm.State().Snapshot()
if err != nil {
return nil, err
}
alloc, err := snap.AllocByID(nil, claims.AllocationID)
if err != nil {
return nil, err
}
if alloc == nil || alloc.Job == nil {
return nil, fmt.Errorf("allocation does not exist")
}
// the claims for terminal allocs are always treated as expired
if alloc.TerminalStatus() {
return nil, fmt.Errorf("allocation is terminal")
}
return claims, nil
}
func (s *Server) ResolveClaims(claims *structs.IdentityClaims) (*acl.ACL, error) {
policies, err := s.resolvePoliciesForClaims(claims)
if err != nil {
return nil, err
}
// Compile and cache the ACL object. For many claims this will result in an
// ACL object with no policies, which can be efficiently cached.
aclObj, err := structs.CompileACLObject(s.aclCache, policies)
if err != nil {
return nil, err
}
return aclObj, nil
}
// resolveTokenFromSnapshotCache is used to resolve an ACL object from a
// snapshot of state, using a cache to avoid parsing and ACL construction when
// possible. It is split from resolveToken to simplify testing.
func resolveTokenFromSnapshotCache(snap *state.StateSnapshot, cache *structs.ACLCache[*acl.ACL], secretID string) (*acl.ACL, error) {
// Lookup the ACL Token
var token *structs.ACLToken
var err error
// Handle anonymous requests
if secretID == "" {
token = structs.AnonymousACLToken
} else {
token, err = snap.ACLTokenBySecretID(nil, secretID)
if err != nil {
return nil, err
}
if token == nil {
return nil, structs.ErrTokenNotFound
}
if token.IsExpired(time.Now().UTC()) {
return nil, structs.ErrTokenExpired
}
}
return resolveACLFromToken(snap, cache, token)
}
func resolveACLFromToken(snap *state.StateSnapshot, cache *structs.ACLCache[*acl.ACL], token *structs.ACLToken) (*acl.ACL, error) {
// Check if this is a management token
if token.Type == structs.ACLManagementToken {
return acl.ManagementACL, nil
}
// Store all policies detailed in the token request, this includes the
// named policies and those referenced within the role link.
policies := make([]*structs.ACLPolicy, 0, len(token.Policies)+len(token.Roles))
// Iterate all the token policies and add these to our policy tracking
// array.
for _, policyName := range token.Policies {
policy, err := snap.ACLPolicyByName(nil, policyName)
if err != nil {
return nil, err
}
if policy == nil {
// Ignore policies that don't exist, since they don't grant any
// more privilege.
continue
}
// Add the policy to the tracking array.
policies = append(policies, policy)
}
// Iterate all the token role links, so we can unpack these and identify
// the ACL policies.
for _, roleLink := range token.Roles {
// Any error reading the role means we cannot move forward. We just
// ignore any roles that have been detailed but are not within our
// state.
role, err := snap.GetACLRoleByID(nil, roleLink.ID)
if err != nil {
return nil, err
}
if role == nil {
continue
}
// Unpack the policies held within the ACL role to form a single list
// of ACL policies that this token has available.
for _, policyLink := range role.Policies {
policy, err := snap.ACLPolicyByName(nil, policyLink.Name)
if err != nil {
return nil, err
}
// Ignore policies that don't exist, since they don't grant any
// more privilege.
if policy == nil {
continue
}
// Add the policy to the tracking array.
policies = append(policies, policy)
}
}
// Compile and cache the ACL object
aclObj, err := structs.CompileACLObject(cache, policies)
if err != nil {
return nil, err
}
return aclObj, nil
}
// ResolveSecretToken is used to translate an ACL Token Secret ID into
// an ACLToken object, nil if ACLs are disabled, or an error.
func (s *Server) ResolveSecretToken(secretID string) (*structs.ACLToken, error) {
// TODO(Drew) Look into using ACLObject cache or create a separate cache
// Fast-path if ACLs are disabled
if !s.config.ACLEnabled {
return nil, nil
}
defer metrics.MeasureSince([]string{"nomad", "acl", "resolveSecretToken"}, time.Now())
if secretID == "" {
return structs.AnonymousACLToken, nil
}
if !helper.IsUUID(secretID) {
return nil, structs.ErrTokenInvalid
}
snap, err := s.fsm.State().Snapshot()
if err != nil {
return nil, err
}
// Lookup the ACL Token
token, err := snap.ACLTokenBySecretID(nil, secretID)
if err != nil {
return nil, err
}
if token == nil {
return nil, structs.ErrTokenNotFound
}
if token.IsExpired(time.Now().UTC()) {
return nil, structs.ErrTokenExpired
}
return token, nil
}
func (s *Server) resolvePoliciesForClaims(claims *structs.IdentityClaims) ([]*structs.ACLPolicy, error) {
snap, err := s.fsm.State().Snapshot()
if err != nil {
return nil, err
}
alloc, err := snap.AllocByID(nil, claims.AllocationID)
if err != nil {
return nil, err
}
if alloc == nil || alloc.Job == nil {
return nil, fmt.Errorf("allocation does not exist")
}
// Find any policies attached to the job
iter, err := snap.ACLPolicyByJob(nil, alloc.Namespace, alloc.Job.ID)
if err != nil {
return nil, err
}
policies := []*structs.ACLPolicy{}
for {
raw := iter.Next()
if raw == nil {
break
}
policy := raw.(*structs.ACLPolicy)
if policy.JobACL == nil {
continue
}
switch {
case policy.JobACL.Group == "":
policies = append(policies, policy)
case policy.JobACL.Group != alloc.TaskGroup:
continue // don't bother checking task
case policy.JobACL.Task == "":
policies = append(policies, policy)
case policy.JobACL.Task == claims.TaskName:
policies = append(policies, policy)
}
}
return policies, nil
}