DOST: Ceph in a security critical OpenStack cloud

Ceph in a security critical OpenStack cloud

Danny Al-Gaaf (Deutsche Telekom)Deutsche OpenStack Tage 2015 - Frankfurt

● Ceph and OpenStack● Secure NFV cloud at DT● Attack surface● Proactive countermeasures

○ Setup○ Vulnerability prevention○ Breach mitigation

● Reactive countermeasures○ 0-days, CVEs○ Security support SLA and lifecycle

● Conclusions

Overview

2

Ceph and OpenStack

Ceph Architecture

4

Ceph and OpenStack

5

Secure NFV Cloud @ DT

NFV Cloud @ Deutsche Telekom

● Datacenter design○ BDCs

■ few but classic DCs ■ high SLAs for infrastructure and services■ for private/customer data and services

○ FDCs■ small but many■ near to the customer■ lower SLAs, can fail at any time■ NFVs:

● spread over many FDCs● failures are handled by services and not the infrastructure

● Run telco core services @ OpenStack/KVM/Ceph7

Fundamentals - The CIA Triad

8

CONF

IDEN

TIAL

ITY INTEGRITY

AVAILABILITY

Preventing sensitive data against unauthorized

access

Maintaining consistency, accuracy, and trustworthiness

of data

Protecting systems against disrupting services and

availability of information

High Security Requirements

● Multiple security placement zones (PZ)○ e.g. EHD, DMZ, MZ, SEC, Management○ TelcoWG “Security Segregation” use case

● Separation between PZs required for:○ compute○ networks○ storage

● Protect against many attack vectors● Enforced and reviewed by security department

9

Solutions for storage separation

● Physical separation○ Large number of clusters (>100)○ Large hardware demand (compute and storage)○ High maintenance effort○ Less flexibility

● RADOS pool separation○ Much more flexible○ Efficient use of hardware

● Question:○ Can we get the same security as physical separation?

10

Separation through Placement Zones

● One RADOS pool for each security zone○ Limit access using Ceph capabilities

● OpenStack AZs as PZs○ Cinder

■ Configure one backend/volume type per pool (with own key)■ Need to map between AZs and volume types via policy

○ Glance■ Lacks separation between control and compute/storage layer■ Separate read-only vs management endpoints

○ Manila■ Currently not planned to use in production with CephFS■ May use RBD via NFS

11

Attack Surface

RadosGW attack surface

● S3/Swift○ Network access to gateway

only○ No direct access for

consumer to other Ceph daemons

● Single API attack surface

13

RBD librbd attack surface

● Protection from hypervisor block layer○ transparent for the guest○ No network access or CephX

keys needed at guest level

● Issue:○ hypervisor is software and

therefore not 100% secure…■ breakouts are no mythical creature

■ e.g., Virtunoid, SYSENTER, Venom!14

RBD.ko attack surface

● RBD kernel module○ e.g. used with XEN or on bare

metal○ Requires direct access to Ceph

public network○ Requires CephX keys/secret at

guest level

● Issue:○ no separation between cluster

and guest15

CephFS attack surface

● pure CephFS tears a big hole in hypervisor separation○ Requires direct access to Ceph

public network○ Requires CephX keys/secret at

guest level○ Complete file system visible to

guest■ Separation currently only via POSIX

user/group

16

Host attack surface

● If KVM is compromised, the attacker ...○ has access to neighbor VMs○ has access to local Ceph keys○ has access to Ceph public network and Ceph daemons

● Firewalls, deep packet inspection (DPI), ...○ partly impractical due to used protocols○ implications to performance and cost

● Bottom line: Ceph daemons must resist attack○ C/C++ is harder to secure than e.g. Python○ Homogenous: if one daemon is vulnerable, all in the cluster are!

17

Network attack surface

● Sessions are authenticated○ Attacker cannot impersonate clients or servers○ Attacker cannot mount man-in-the-middle attacks

● Client/cluster sessions are not encrypted○ Sniffer can recover any data read or written

18

Denial of Service

● Attack against:○ Ceph Cluster:

■ Submit many / large / expensive IOs■ Open many connections■ Use flaws to crash Ceph daemons■ Identify non-obvious but expensive features of client/OSD interface

○ Ceph Cluster hosts:■ Crash complete cluster hosts e.g. through flaws in kernel network layer

○ VMs on same host:■ Saturate the network bandwidth of the host

19

Proactive Countermeasures

Deployment and Setup

● Network ○ Always use separated cluster and public networks○ Always separate your control nodes from other networks○ Don’t expose cluster to the open internet○ Encrypt inter-datacenter traffic

● Avoid hyper-converged infrastructure○ Don’t mix

■ compute and storage resources, isolate them!■ OpenStack and Ceph control nodes

○ Scale resources independently○ Risk mitigation if daemons are compromised or DoS’d

21

Deploying RadosGW

● Big and easy target through HTTP(S) protocol

● Small appliance per tenant with○ Separate network ○ SSL terminated proxy forwarding

requests to radosgw○ WAF (mod_security) to filter○ Placed in secure/managed zone○ different type of webserver than

RadosGW● Don’t share buckets/users

between tenants22

Ceph security: CephX

● Monitors are trusted key servers○ Store copies of all entity keys○ Each key has an associated “capability”

■ Plaintext description of what the key user is allowed to do

● What you get○ Mutual authentication of client + server○ Extensible authorization w/ “capabilities”○ Protection from man-in-the-middle, TCP

session hijacking● What you don’t get

○ Secrecy (encryption over the wire)23

Ceph security: CephX take-aways

● Monitors must be secured○ Protect the key database

● Key management is important○ Separate key for each Cinder backend/AZ○ Restrict capabilities associated with each key○ Limit administrators’ power

■ use ‘allow profile admin’ and ‘allow profile readonly’■ restrict role-definer or ‘allow *’ keys

○ Careful key distribution (Ceph and OpenStack nodes)● To do:

○ Thorough CephX code review by security experts○ Audit OpenStack deployment tools’ key distribution○ Improve security documentation24

● Static Code Analysis (SCA)○ Buffer overflows and other code flaws○ Regular Coverity scans

■ 996 fixed, 284 dismissed; 420 outstanding■ defect density 0.97

○ cppcheck○ LLVM: clang/scan-build

● Runtime analysis○ valgrind memcheck

● Plan○ Reduce backlog of low-priority issues (e.g., issues in test code)○ Automated reporting of new SCA issues on pull requests○ Improve code reviewer awareness of security defects

Preventing Breaches - Defects

25

● Pen-testing○ human attempt to subvert security, generally guided by code review

● Fuzz testing○ computer attempt to subvert or crash, by feeding garbage input

● Harden build○ -fpie -fpic○ -stack-protector=strong○ -Wl,-z,relro,-z,now○ -D_FORTIFY_SOURCE=2 -O2 (?)○ Check for performance regression!

Preventing Breaches - Hardening

26

Mitigating Breaches

● Run non-root daemons (WIP: PR #4456)○ Prevent escalating privileges to get root○ Run as ‘ceph’ user and group○ Pending for Infernalis

● MAC○ SELinux / AppArmor ○ Profiles for daemons and tools planned for Infernalis

● Run (some) daemons in VMs or containers○ Monitor and RGW - less resource intensive○ MDS - maybe○ OSD - prefers direct access to hardware

● Separate MON admin network27

Encryption: Data at Rest

● Encryption at application vs cluster level● Some deployment tools support dm-crypt

○ Encrypt raw block device (OSD and journal)○ Allow disks to be safely discarded if key remains secret

● Key management is still very simple○ Encryption key stored on disk via LUKS○ LUKS key stored in /etc/ceph/keys

● Plan○ Petera, a new key escrow project from Red Hat

■ https://github.com/npmccallum/petera○ Alternative: simple key management via monitor (CDS blueprint)

28

● Goal○ Protect data from someone listening in on network○ Protect administrator sessions configuring client keys

● Plan○ Generate per-session keys based on existing tickets○ Selectively encrypt monitor administrator sessions○ alternative: make use of IPSec (performance and management

implications)

Encryption: On Wire

29

● Limit load from client○ Use qemu IO throttling features - set safe upper bound

● To do:○ Limit max open sockets per OSD○ Limit max open sockets per source IP

■ handle on Ceph or in the network layer?○ Throttle operations per-session or per-client (vs just globally)?

Denial of Service attacks

30

CephFS

● No standard virtualization layer (unlike block)○ Filesystem passthrough (9p/virtfs) to host○ Proxy through gateway (NFS?)○ Allow direct access from tenant VM (most unsecure)

● Granularity of access control is harder○ No simple mapping to RADOS objects

● Work in progress○ root_squash (Infernalis blueprint)○ Restrict mount to subtree○ Restrict mount to user

31

Reactive Countermeasures

● Community○ Single point of contact: [email protected]

■ Core development team■ Red Hat, SUSE, Canonical security teams

○ Security related fixes are prioritized and backported○ Releases may be accelerated on ad hoc basis○ Security advisories to [email protected]

● Red Hat Ceph○ Strict SLA on issues raised with Red Hat security team○ Escalation process to Ceph developers○ Red Hat security team drives CVE process○ Hot fixes distributed via Red Hat’s CDN

Reactive Security Process

33

Detecting and Preventing Breaches

● Brute force attacks○ Good logging of any failed authentication○ Monitoring easy via existing tools like e.g. Nagios

● To do:○ Automatic blacklisting IPs/clients after n-failed attempts on Ceph level

(Jewel blueprint)

● Unauthorized injection of keys○ Monitor the audit log

■ trigger alerts for auth events -> monitoring○ Periodic comparison with signed backup of auth database?

34

Conclusions

Summary

● Reactive processes are in place○ [email protected], CVEs, downstream product updates, etc.

● Proactive measures in progress○ Code quality improves (SCA, etc.)○ Unprivileged daemons○ MAC (SELinux, AppArmor)○ Encryption

● Progress defining security best-practices○ Document best practices for security

● Ongoing process

36

Get involved !

● Ceph○ https://ceph.com/community/contribute/ ○ [email protected]○ IRC: OFTC

■ #ceph, ■ #ceph-devel

○ Ceph Developer Summit

● OpenStack○ Telco Working Group

■ #openstack-nfv ○ Cinder, Glance, Manila, ...

37

[email protected]

dalgaaf

linkedin.com/in/dalgaaf

Danny Al-Gaaf Senior Cloud Technologist

IRC

THANK YOU!