Sugar: Secure GPU Acceleration in Web Browsers

Zhihao Yao, Zongheng Ma, Yingtong Liu, Ardalan Amiri Sani, Aparna Chandramowlishwaran

Trustworthy Systems Lab, UC Irvine

1

WebGL was released in 2011

Source: https://www.google.com/map2

WebGL is popular

WebGL adoption rate by top 100 websites

3

47.0%53.0%

WebGL is popular

Browser support rate (48.8 million visitors)

Source: http://webglstats.com (2017)

Does not support

4

4.0%

96.0%

https://www.apple.com/macos/sierra/

http://dlmf.nist.gov

https://www.google.com/map

https://eyes.nasa.gov/curiosity/ 5

WebGL recap

6

First, a quick recap on OpenGLNative app

GL libs

Kernel mode GPU driver

GPU hardware

7

user space

kernel space

kernel space

hardware

Native app

GL libs

Native app

GL libs

Kernel mode GPU driver

GPU hardware

function call

First, a quick recap on OpenGL

8

user space

kernel space

kernel space

hardware

First, a quick recap on OpenGL

Native app

GL libs

Native app

GL libs

Kernel mode GPU driver

GPU hardware

syscall

9

user space

kernel space

kernel space

hardware

Use the same design for WebGL?

Web app

GL libs

Kernel mode GPU driver

GPU hardware

10

user space

kernel space

kernel space

hardware

Buggy

Compromised

Malicious

Web app

GL libs

Kernel mode GPU driver

GPU hardware

11

Web apps are not trusted

user space

kernel space

kernel space

hardware

Buggy

Compromised

Malicious

Web app

GL libs

Kernel mode GPU driver

GPU hardware

12

GPU driver is buggy

Buggy

Compromised

Malicioususer space

kernel space

kernel space

hardware

web app

GL libs

Kernel driver is compromised

Web app

GL libs

Kernel mode GPU driver

GPU hardware

13

user space

kernel space

kernel space

hardware

Buggy

Compromised

Malicious

Current WebGL design

14

Kernel mode GPU driver

Browser

GPU hardware

GPU Process

ChecksGL libs

Web appWeb appWeb app

user space

kernel space

kernel space

hardware

Current WebGL design

15

Kernel mode GPU driver

Browser

Web appWeb appWeb app

GPU hardware

GPU Process

ChecksGL libs

IPC

Browser user space

kernel space

kernel space

hardware

Security checks in GPU Process

16

Kernel mode GPU driver

Browser

GPU hardware

GPU Process

ChecksGL libs

Web appWeb appWeb app

user space

kernel space

kernel space

hardware

TCB of current WebGL Design

17

Kernel mode GPU driver

Browser

GPU hardware

GPU Process

ChecksGL libs

Web appWeb appWeb app 158,000 LoC (GPU Process)457,000 LoC(GL libraries)

123,000 LoC(GPU driver)

18

Kernel mode GPU driver

Browser

GPU hardware

GPU Process

ChecksGL libs

Web appWeb appWeb app

CVE-2014-1556CVE-2015-7179CVE-2013-2874CVE-2017-5031CVE-2014-1502

Vulnerabilities in GPU process

Kernel driver is compromised

19

CVE-2011-2601*Chrome 153469Chrome 483877*CVE-2011-2367CVE-2011-3653

Kernel mode GPU driver

Browser

GPU hardware

GPU Process

ChecksGL libs

Web appWeb appWeb app

*Not yet fixed

Vulnerability examplesCVE-2014-1556 execute arbitrary codeCVE-2015-7179 execute arbitrary codeCVE-2013-2874 read browser UICVE-2017-5031 read GPU process memoryCVE-2014-1502 use of cross-origin contentsChrome Issue 593680 browser hangChrome Issue 83841 leak system usernameCVE-2011-2601* system UI freezeChrome issue 153469 kernel panicChrome issue 483877* system UI freezeCVE-2011-2367 read of GPU memoryCVE-2011-3653 read of GPU memoryCVE-2014-3173 read of GPU memory

20 *Not yet fixed

Our WebGL vulnerability study

21

https://trusslab.github.io/sugar/webgl_bugs

High performance

Knownvulnerabilities

Zero day vulnerabilities

System UI freeze

Current WebGL design

22

CVE-2014-3173, read of GPU graphics memory

We type some private notes in terminal:

23

CVE-2014-3173, read of GPU graphics memory

24

Overview of Sugar

Key idea:

• Use GPU virtualization to give an untrusted web app a separate vGPU

25

Intel GPU virtualization

26

• We build a prototype on Intel GPU virtualization

● Intel GPU virtualization is available since the 4th generation Core processors [1]

[1] https://www.usenix.org/conference/atc14/technical-sessions/presentation/tian Photo credit: https://www.intel.com/pressroom/archive/releases/2008/20081117comp_sm.htm

27

GPU GPU

vGPU 2 vGPU 1

28

Sugar’s design

Kernel mode GPU driver

Browser

vGPU

GPU hardware

Web app

GL libsvGPU driver

GPU Process

GL libs

29

user space

kernel space

hardware

Sugar’s design

Kernel mode GPU driver

Browser

vGPU

GPU hardware

Web app

GL libsvGPU driver

GPU Process

GL libs

function call

30

user space

kernel space

hardware

Sugar’s design

Kernel mode GPU driver

Browser

vGPU

GPU hardware

Web app

GL libsvGPU driver

GPU Process

GL libsfunction call

31

user space

kernel space

hardware

Sugar’s design

Kernel mode GPU driver

Browser

vGPU

GPU hardware

Web app

GL libsvGPU driver

GPU Process

GL libs

32

user space

kernel space

hardware

Sugar’s design

virtual graphics plane

primary graphics plane

Kernel mode GPU driver

Browser

vGPU

GPU hardware

Web app

GL libsvGPU driver

GPU Process

GL libs

33

Why is Sugar secure?

34

Web app process is untrusted

Kernel mode GPU driver

Browser

vGPU

GPU hardware

Web app

GL libsvGPU driver

GPU Process

GL libs

35

user space

kernel space

hardware

Web app process is sandboxed

Kernel mode GPU driver

Browser

vGPU

GPU hardware

Web app

GL libsvGPU driver

GPU Process

GL libs

36

user space

kernel space

hardware

vGPU is isolated

Kernel mode GPU driver

Browser

vGPU

GPU hardware

Web app

GL libsvGPU driver

GPU Process

GL libs

37

user space

kernel space

hardware

Sugar’s TCB is small

Kernel mode GPU driver

Browser

vGPU

GPU hardware

Web app

GL libsvGPU driver

GPU Process

GL libs

38

34,400 LoC(GPU virtualization)

user space

kernel space

hardware

Vulnerability examplesCVE-2014-1556 execute arbitrary codeCVE-2015-7179 execute arbitrary codeCVE-2013-2874 read browser UICVE-2017-5031 read GPU process memoryCVE-2014-1502 use of cross-origin contentsChrome Issue 593680 browser hangChrome Issue 83841 leak system usernameCVE-2011-2601* system UI freezeChrome issue 153469 kernel panicChrome issue 483877* system UI freezeCVE-2011-2367 read of GPU memoryCVE-2011-3653 read of GPU memoryCVE-2014-3173 read of GPU memory

39 *Not yet fixed

Limitation of this Sugar designIntel vGPU hang will cause a real GPU hang

40

Dual-GPU Sugar

Key idea: Use two GPUs to fully isolate the virtual graphics plane and the primary graphics plane.

● Solves system UI freeze

● Provides better performance isolation

41

Dual-GPU Sugar’s design

Kernel mode GPU 1 driver

GPU 1 hardware

Browser

vGPU

GPU 2 hardware

Kernel mode GPU 2 driver

Web app

GL libsvGPU driver

GPU process

GL libs

42

user space

kernel space

hardware

Photo credit: https://www.amd.com/zh-tw/products/graphics/desktop/6000/6990

Many computers have two GPUs

apple.com/macbook-pro

dell.com/Inspiron15

store.hp.com/envy43

Source: https://newsroom.intel.com/news/8th-gen-intel-core-radeon-rx-vega-m-graphics

Intel’s 8th Generation Core Processors with Radeon RX Vega M Graphics

44

Sugar’s implementation

45

WebKit / Blink

WebGL frontend

WebGL backend

GL libs

vGPU driver

WebGL in web app process

GPU Process

WebGL backend

GL libs

Reuse most of GPU process code

46

Ported from GPU process

vGPU driver as a library

WebKit / Blink

WebGL frontend

WebGL backend

GL libs

vGPU driver

We modify GL libs to issue function calls instead of syscalls

47

function call

Register: trap and emulate

Kernel mode GPU driver

Browser

vGPU

GPU hardware

Web app

GL libsvGPU driver

GPU Process

GL libsMapped registers

48

user space

kernel space

hardware

Register: trap and emulate

Kernel mode GPU driver

Browser

vGPU

GPU hardware

Web app

GL libsvGPU driver

GPU Process

GL libsMapped registers

GPU virtualization layer will emulate

49

user space

kernel space

hardware

Interrupt: deliver as signal

Kernel mode GPU driver

Browser

vGPU

GPU hardware

Web app

GL libsvGPU driver

GPU Process

GL libs

Interrupt

50

user space

kernel space

hardware

Interrupt: deliver as signal

Kernel mode GPU driver

Browser

vGPU

GPU hardware

Web app

GL libsvGPU driver

GPU Process

GL libs

Interrupt

51

The virtualization layer delivers as a signal

user space

kernel space

hardware

Interrupt: deliver as signal

Kernel mode GPU driver

Browser

vGPU

GPU hardware

Web app

GL libsvGPU driver

GPU Process

GL libs

Interrupt

Signal

52

user space

kernel space

hardware

DMA overview

Main memoryGPU

53

DMA

DMA overview

Main memoryvGPU

54

DMAPagetable

Evaluations

55

Sugar’s performance is goodunder the same WebGL benchmarks that Chrome uses

56

Sugar’s performance is goodunder the same WebGL benchmarks that Chrome uses

57

60 FPS

Sugar’s CPU overhead is lowSugar is better than CPU rendering by 375% on average

58

Summary

● Sugar leverages modern GPU virtualization solutions to isolate WebGL

● Sugar addresses this by repurposing Intel vGPU driver to a library

Thank you!

Sugar is open source: https://trusslab.github.io/sugar59