1. 1 Ally: OS-Transparent Packet Inspection Using Sequestered Cores Jen-Cheng Huang 1, Matteo Monchiero 2, Yoshio Turner 3, Hsien-Hsin Lee 1 1 Georgia Tech 2 Intel Labs 3 HP Labs

2. 2 Deep Packet Inspection (DPI) Data Center Middle Boxes Intrusion Detection Content Insertion Traffic Classification Internet Deployment of Packet Processing Services

3. 3 Problem Internet Data Center Middle Boxes Local Traffic is growing in importance… But The traffic within the data center is not inspected!

4. 4 Approach “Co-locate” DPI with the server DPI appliance Server Leverage abundant CPU resources Leverage existing management interfaces on servers, e.g. HP iLO Compatible with heterogeneous architecture, e.g. on-chip accelerators

5. 5 Requirements Transparency –Independent to the server’s software stack Efficiency –Low overhead packet interception Isolation –Resistant to attacks

6. 6 ETTM: a scalable fault tolerant network manager. C. Dixon et al. NSDI ‘11 Related Work Transparency Hypervisor Overhead Hypervisor Vulnerability Virtualization Support for DPI deployment Hypervisor DPI VM Guest VM HW SW Virtualized Platform Processors

7. 7 NIC Unprivileged Partition Multi-core processor core Privileged Partition Ally Architecture Software Stack (OS + Applications) Software Stack (DPI Application) core NIC Traffic

8. 8 Outline Introduction & Motivation Architecture –Overview –Multicore Partitioning –Packet interception Evaluation Conclusions

9. 9 Northbridge MMU Service Processor NIC Memory Controller Interconnect IOMMU Interrupt Unit BIOS External Network Main Memory Core Interrupt Controller MMU Core Interrupt Controller MMU Last Level Cache Management Network Baseline Architecture

10. 10 Northbridge MMU Service Processor NIC Memory Controller Interconnect IOMMU Interrupt Unit BIOS External Network Main Memory Core Interrupt Controller MMU Core Interrupt Controller MMU Last Level Cache Management Network Ally Architecture Unprivileged partitionPrivileged partition

11. 11 Outline Introduction & Motivation Architecture –Overview –Multicore Partitioning –Packet interception Evaluation Conclusions

12. 12 Multicore Partitioning NIC Unprivileged Partition Multi-core processor core Privileged Partition Software Stack (OS + Applications) Software Stack (DPI Application) core Invisible

13. 13 Core Sequestration Modify the BIOS to hide privileged core information from the OS  BSP core - the first core that boots  AP cores - the other cores  IPI - Inter-processor interrupts OS retrieves cores information BSP AP Core Info Table Wakeup IPI Update Ally Booting Procedure: AP DPI Engine DPI core waits for IN/OUT packets Initialize …...

14. 14 Memory Protection TLB TLB Miss Handler CR3 Boundary Register Page Table Range Checking Main Memory Privileged partition Unprivileged partition MMU Unprivileged Core Partition the memory into two physically contiguous regions TLB Miss TLB Fill

15. 15 Outline Introduction & Motivation Architecture –Overview –Multicore Partitioning –Packet interception Evaluation Conclusions

16. 16 NIC Unprivileged Partition Multi-core processor core Privileged Partition Packet Interception Software Stack (OS + Applications) Software Stack (DPI Application) core NIC Traffic

17. 17 Packet Interception Virtualization of the Descriptor Queues NIC OS memory Descriptor queues replicated DPI memory Only one copy of the packet buffers Descriptor queues

18. 18 Packet Interception Virtualization of the Descriptor Queues –Device independent, software independent –No copying on packet buffers Processor and NIC communication –Queue manipulation uses Memory Mapped IO (MMIO) accesses –NIC event notification uses Interrupt

19. 19 MMIO redirection MMU MMU detects specific MMIO addresses MMU redirects RW to a reserved region in DPI memory MMU sends IPI to DPI core DPI memory DPI core OS core IPI R/W redirection Load/store

20. 20 Ally Hardware Properties Simple extensions to existing hardware components No impact expected on critical timing paths Compatible with virtualization support (Intel VT- x/EPT, AMD SVM/NPT)

21. 21 Outline Introduction & Motivation Architecture –Overview –Multicore Partitioning –Packet interception Evaluation Conclusions

22. 22 Evaluation Full system emulation QEMU Core sequestration HW changes Real machine prototype Hardware –Intel Core 2 duo 2.66 GHz with 1 Gbit Intel NIC Benchmarks –Netperf –SPECweb Systems –Ally, Linux and Xen

23. 23 System Configurations Queue Virtualization NIC Driver Kernel Netperf/ Specweb Snort DPI core OS coreHW SW NIC Driver Kernel Netperf/ Specweb Snort DPI core OS coreHW SW IP queue Ally Linux

24. 24 System Configurations Hypervisor Dom0 Kernel Netperf/ Specweb Snort DPI core OS coreHW SW Xen DomU Kernel

25. 25 Netperf CPU Usage

26. 26 SPECweb CPU Usage cycles/request * 10 6

27. 27 Outline Introduction & Motivation Architecture –Overview –Multicore Partitioning –Packet interception Evaluation Conclusions

28. 28 Conclusions  Ally: a framework for transparent deployment of packet inspection appliances  Ally uses a set of simple HW/FW extensions enable reliable multicore partitioning and efficient packet inspection  Ally is fully compatible with new virtualization technology as well as heterogeneous architecture

29. 29 Thanks

30. 30 Throughput

31. 31 DPI using Network Processor

32. 32 NIC Unprivileged Partition Multi-core processor core Conventional Architecture Software Stack (OS + Applications) core cores

33. 33 NIC Unprivileged Partition Multi-core processor core Privileged Partition Transmission Path Software Stack (OS + Applications) Software Stack (DPI Application) core

34. 34 NIC Unprivileged Partition Multi-core processor core Privileged Partition Receive Path Software Stack (OS + Applications) Software Stack (DPI Application) core

35. 35 Integrated Northbridge DPI core Local APIC MMU Interface DPI core Local APIC MMU Interface OS core Local APIC MMU Interface OS core Local APIC MMU Interface Platform Controller Hub NIC Memory Controller On chip interconnect Processor IOMMUPCIe ctrl Interrupt Unit BIOS Network Main Memory DMI Ctrl OS core Local APIC MMU Interface Unprivileged partitionPrivileged partition DPI core Local APIC MMU Interface Last Level Cache IOAPIC Managemen t NIC Service Processor Management Network Privileged partition Unprivilege d partition

36. 36 Integrated Northbridge DPI core Local APIC MMU Interface DPI core Local APIC MMU Interface OS core Local APIC MMU Interface OS core Local APIC MMU Interface Platform Controller Hub NIC Memory Controller On chip interconnect Processor IOMMUPCIe ctrl Interrupt Unit BIOS Network Main Memory DMI Ctrl OS core Local APIC MMU Interface Unprivileged partitionPrivileged partition DPI core Local APIC MMU Interface Last Level Cache IOAPIC Managemen t NIC Service Processor Management Network Privileged partition Unprivilege d partition

37. 37 MMU Modification – Memory Protection TLB TLB Miss Handler CR3Special_reg Page Table DPI core boundary register phys_addr > special_reg ? Main Memory Privileged partition Unprivilege d partition

38. 38 Memory Protection Procedure TLB TLB Miss Handler TLB miss Virtual Address CR3Special_reg Page Table DPI core boundary register phys_addr > special_reg ? Main Memory Privileged partition Unprivilege d partition

39. 39 Memory Protection Procedure TLB TLB Miss Handler TLB miss Virtual Address TLB fill CR3Special_reg Page Table DPI core boundary register phys_addr > special_reg ? Main Memory Privileged partition Unprivilege d partition

40. 40 NIC Unprivileged Partition Multi-core processor core Privileged Partition Memory Protection Software Stack (OS + Applications) Software Stack (DPI Application) core Invisible

41. 41 Integrated Northbridge DPI core Local APIC MMU Interface DPI core Local APIC MMU Interface OS core Local APIC MMU Interface OS core Local APIC MMU Interface Platform Controller Hub NIC Memory Controller On chip interconnect Processor IOMMUPCIe ctrl Interrupt Unit BIOS Network Main Memory DMI Ctrl OS core Local APIC MMU Interface Unprivileged partitionPrivileged partition DPI core Local APIC MMU Interface Last Level Cache IOAPIC Management NIC Service Processor Management Network Privileged partition Unprivilege d partition

42. 42 Integrated Northbridge DPI core Local APIC MMU Interface DPI core Local APIC MMU Interface OS core Local APIC MMU Interface OS core Local APIC MMU Interface Platform Controller Hub NIC Memory Controller On chip interconnect Processor IOMMUPCIe ctrl Interrupt Unit BIOS Network Main Memory DMI Ctrl OS core Local APIC MMU Interface Unprivileged partitionPrivileged partition DPI core Local APIC MMU Interface Last Level Cache IOAPIC Managemen t NIC Service Processor Management Network Privileged partition Unprivilege d partition

43. 43 MMU Modification – MMIO Redirection TLB Redirection BitPhysical Page TLB Miss Handler Check uncacheable address map Redirection Table Physical Address Remapped Address

44. 44 MMIO Redirection – TLB Miss On a TLB miss, the TLB miss handler does the page table walk TLB Redirection BitPhysical Page Virtual Address TLB miss TLB Miss Handler Page Table Lookup

45. 45 MMIO Redirection – TLB Miss The TMH checks if the resulting physical address falls in an uncacheable page and hence potentially a MMIO page TLB Redirection BitPhysical Page TLB Miss Handler Physical Address Check uncacheable address map

46. 46 MMIO Redirection – TLB Miss If the page is uncacheable, the TMH looks up the redirection table to check if any address in this page needs to be redirected TLB Redirection BitPhysical Page TLB Miss Handler Check uncacheable address map Redirection Table Physical Address Remapped Address Physical Address

47. 47 MMIO Redirection – TLB Miss If any address in the page needs to be redirected, the TMH sets the redirection bit in addition to fill the TLB TLB Redirection BitPhysical Page TLB Miss Handler Check uncacheable address map TLB fill Redirection Table Physical Address Remapped Address

48. 48 MMIO Redirection – TLB Hit On a TLB hit, if the redirection bit is set, the MMU looks up the Last Level Cache (LLC) used to cache translations in Redirection Table TLB Redirection Bit Physical Page Offset Physical Address Virtual Address LLC Physical Address Remapped Address

49. 49 MMIO Redirection – TLB Hit If a translation is found, the MMU returns the translated address and sends IPI to privileged cores. TLB Redirection Bit Physical Page LLC Physical Address Remapped Address Translated Address Generate IPI Physical Address Hit

50. 50 MMIO Redirection – TLB Hit If the LLC misses, then Redirection Table Lookup is performed TLB Redirection Bit Physical Page LLC Physical Address Remapped Address Redirection Table Lookup Physical Address Miss

51. 51 Interrupt Unit Modification DPI core OS core Interrupt Unit NIC If Source == NIC, Redirect Interrupt

52. 52 When NIC raises an interrupt, The interrupt Unit redirects the interrupt to DPI core Interrupt Redirection DPI core OS core Interrupt Unit NIC If Source == NIC, Redirect Interrupt Interrupt

53. 53 After the NIC interrupt is handled, DPI core sends an IPI to OS core mimicking NIC interrupt Interrupt Redirection DPI core OS core Interrupt Unit NIC If Source == NIC, Redirect Interrupt IPI

54. 54 Summary of Hardware Modifications UnitDescriptionPurpose OS-core MMU Prevent memory accesses to DPI memory from OS- core Protection Redirect MMIO accesses to DPI memory from OS- core and interrupt DPI core Packet Interception IOMMUPrevent non authorized DMA to DPI MemoryProtection IOAPICRedirect NIC interrupts to DPI-corePacket Interception All UnitsProtected configuration registersProtection

55. 55 Functional Evaluation Full system emulation QEMU Validate Hardware and Firmware Changes

56. 56 DPI core Usage

57. 57 SPECweb Cache Misses

58. 58 NIC Unprivileged Partition Multi-core processor core Privileged Partition Memory Protection Software Stack (OS + Applications) Software Stack (DPI Application) core Invisible How? Modified MMU

59. 59 Challenges -Make privileged partition protected and invisible from the unprivileged partition -Core Sequestration -Memory Protection -Intercept packets efficiently -Packet Interception

60. 60 Ally System NIC Linux kernel NIC Traffic Queue Virtualization NIC Driver Other Apps Snort DPI Core

61. 61 Linux System NIC Linux kernel NIC Traffic IP queue NIC Driver Other Apps Snort Core

62. 62 Xen System NIC Linux VM #0 NIC Traffic IP queue Hypervisor Other Apps Snort Core VM #1