Theory and Practice of Finding Eviction Sets Pepe Vila Boris Kpf - - PowerPoint PPT Presentation

▶
theory and practice of finding eviction sets
SMART_READER_LITE
LIVE PREVIEW

Theory and Practice of Finding Eviction Sets Pepe Vila Boris Kpf - - PowerPoint PPT Presentation

Dec 31, 2022 311 likes •1.07k views

Theory and Practice of Finding Eviction Sets Pepe Vila Boris Kpf Jos F. Morales IMDEA Software Institute Microsoft Research IMDEA Software Institute vwzq.net @cgvwzq github.com/cgvwzq Eviction Sets CACHE Find addresses that collide

slide-1
SLIDE 1

Theory and Practice of Finding Eviction Sets

Pepe Vila Boris Köpf José F. Morales IMDEA Software Institute Microsoft Research IMDEA Software Institute

vwzq.net @cgvwzq github.com/cgvwzq

slide-2
SLIDE 2

Eviction Sets

CACHE

SLICE 0 sets

associativity

SLICE 1 sets

Find addresses that collide in cache: i.e. addresses mapped into the same cache set

associativity

slide-3
SLIDE 3

Eviction Sets

CACHE

SLICE 0 sets

associativity

SLICE 1 sets

Find addresses that collide in cache: i.e. addresses mapped into the same cache set

associativity

slide-4
SLIDE 4

Eviction Sets

CACHE

SLICE 0 sets

associativity

SLICE 1 sets

Find addresses that collide in cache: i.e. addresses mapped into the same cache set

associativity

slide-5
SLIDE 5

Eviction Sets

CACHE

SLICE 0 sets

associativity

SLICE 1 sets

Find addresses that collide in cache: i.e. addresses mapped into the same cache set

associativity

Find associativity many colliding addresses: i.e. an eviction set

slide-6
SLIDE 6

Attacks

Efficient attacks require small eviction sets

slide-7
SLIDE 7

Attacks

Prime+Probe

Efficient attacks require small eviction sets

slide-8
SLIDE 8

Attacks

Prime+Probe Rowhammer

Efficient attacks require small eviction sets

slide-9
SLIDE 9

Attacks

Prime+Probe Rowhammer Spectre

Efficient attacks require small eviction sets

slide-10
SLIDE 10

Problem

PHYSICAL MEMORY CACHE

SLICE 0 sets SLICE 1 sets

associativity associativity

Potentially unknown mapping from physical address to cache set

slide-11
SLIDE 11

Problem

MMU

text heap stack

low high

USER PROCESS PHYSICAL MEMORY CACHE

SLICE 0 sets SLICE 1 sets

Unknown translation from virtual to physical addresses

associativity associativity

slide-12
SLIDE 12

Problem

<script> var foo = new Uint32Array(N); foo[12]; ... </script>

text heap stack

low high

USER PROCESS MMU PHYSICAL MEMORY CACHE

SLICE 0 sets SLICE 1 sets

In some scenarios, even unknown virtual address

associativity associativity

slide-13
SLIDE 13

Problem

<script> var foo = new Uint32Array(N); foo[12]; ... </script>

text heap stack

low high

USER PROCESS MMU PHYSICAL MEMORY CACHE

SLICE 0 sets SLICE 1 sets

Find associativity many elements (e.g. JS array indices) that collide in cache.

associativity associativity

slide-14
SLIDE 14

Contributions

Systematic study of the problem of finding eviction sets

slide-15
SLIDE 15

Contributions

Find eviction sets in O(n) compared to previous O(n2) Systematic study of the problem of finding eviction sets

slide-16
SLIDE 16

Contributions

Systematic study of the problem of finding eviction sets Find eviction sets in O(n) compared to previous O(n2) Reliability and performance evaluation

  • f algorithms in real hardware
slide-17
SLIDE 17

Find a large eviction set for an address V:

  • Pick “enough” addresses at random
  • Timing test:

Finding minimal eviction sets 1

slide-18
SLIDE 18

Find a large eviction set for an address V:

  • Pick “enough” addresses at random
  • Timing test:

Finding minimal eviction sets

Reduce initial large eviction set into its minimal core

2 1

slide-19
SLIDE 19

Find a large eviction set for an address V:

  • Pick “enough” addresses at random
  • Timing test:

Finding minimal eviction sets

Reduce initial large eviction set into its minimal core

2 1

slide-20
SLIDE 20

Baseline algorithm

N S : Start with large enough eviction set S of size N

slide-21
SLIDE 21

Baseline algorithm

N’ S : Pick candidate element C, and Test if remaining set TEST(S\{C}) is still an eviction set

slide-22
SLIDE 22

Baseline algorithm

S : If TEST(S\{C}) = True, discard C N’

slide-23
SLIDE 23

Baseline algorithm

N’ S : and continue with N’=N-1

slide-24
SLIDE 24

Baseline algorithm

N’ S : We repeat this process several times

slide-25
SLIDE 25

Baseline algorithm

N’ S : We repeat this process several times

slide-26
SLIDE 26

Baseline algorithm

N’ S : We repeat this process several times

slide-27
SLIDE 27

Baseline algorithm

N’ S : We repeat this process several times

slide-28
SLIDE 28

Baseline algorithm

N’ S : Until we find an element C such that when removed the remaining set stops being an eviction set: TEST(S\{C}) = False

slide-29
SLIDE 29

Baseline algorithm

N’ S : We learn that C is part of the minimal core

slide-30
SLIDE 30

Baseline algorithm

N’ S : We keep track of it, and insert it again in S

slide-31
SLIDE 31

Baseline algorithm

N’ S : We repeat this process several times

slide-32
SLIDE 32

Baseline algorithm

N’ S : We repeat this process several times

slide-33
SLIDE 33

Baseline algorithm

N’ S : We repeat this process several times

slide-34
SLIDE 34

Baseline algorithm

N’ S : We repeat this process several times

slide-35
SLIDE 35

Baseline algorithm

N’ S : We repeat this process several times

slide-36
SLIDE 36

Baseline algorithm

N’ S : We repeat this process several times

slide-37
SLIDE 37

Baseline algorithm

N’ S : We repeat this process several times

slide-38
SLIDE 38

Baseline algorithm

N’ S : We repeat this process several times

slide-39
SLIDE 39

Baseline algorithm

N’ S : We repeat this process several times

slide-40
SLIDE 40

Baseline algorithm

N’ S : We repeat this process several times

slide-41
SLIDE 41

Baseline algorithm

N’ S : Until we have identified ASSOCIATIVITY many elements representing the eviction set’s core! ASSOCIATIVITY

slide-42
SLIDE 42

Baseline algorithm

S :

O(N2) memory accesses

ASSOCIATIVITY N’

slide-43
SLIDE 43

Threshold Group Testing

Group testing problem by Robert Dorfman (1943)

Blood samples (10 individual tests)

slide-44
SLIDE 44

Threshold Group Testing

Group testing problem by Robert Dorfman (1943)

Blood samples (4 group tests + 3 individual tests )

slide-45
SLIDE 45

Threshold Group Testing

Group testing problem by Robert Dorfman (1943) Generalization by Peter Damaschke (2006):

  • Positive test only if at least “u” defectives
  • Negative test only if at most “l” defectives
  • Random otherwise

Blood samples (4 group tests + 3 individual tests )

slide-46
SLIDE 46

Blood samples (4 group tests + 3 individual tests )

Threshold Group Testing

Group testing problem by Robert Dorfman (1943) Generalization by Peter Damaschke (2006):

  • Positive test if at least “u” defectives
  • Negative test if at most “l” defectives
  • Random answer otherwise

Observation: Our test is a threshold group test!

slide-47
SLIDE 47

Group-testing algorithm

N S : Start with large enough eviction set S of size N

slide-48
SLIDE 48

Group-testing algorithm

N S : Split S in ASSOCIATIVITY+1 subsets

slide-49
SLIDE 49

Group-testing algorithm

N S : In the worst case, there exists a union of ASSOCIATIVITY subsets being an eviction set

slide-50
SLIDE 50

Group-testing algorithm

N S : We can discard N/(ASSOCIATIVITY+1) elements per iteration

slide-51
SLIDE 51

Group-testing algorithm

N’ S : We repeat this process until we have ASSOCIATIVITY many elements

slide-52
SLIDE 52

Group-testing algorithm

We repeat this process until we have ASSOCIATIVITY many elements N’ S :

slide-53
SLIDE 53

Group-testing algorithm

We repeat this process until we have ASSOCIATIVITY many elements N’ S :

slide-54
SLIDE 54

Group-testing algorithm

We repeat this process until we have ASSOCIATIVITY many elements N’ S :

slide-55
SLIDE 55

Group-testing algorithm

We repeat this process until we have ASSOCIATIVITY many elements N’ S :

slide-56
SLIDE 56

Group-testing algorithm

We repeat this process until we have ASSOCIATIVITY many elements N’ S :

slide-57
SLIDE 57

Group-testing algorithm

We repeat this process until we have ASSOCIATIVITY many elements N’ S :

slide-58
SLIDE 58

Group-testing algorithm

We repeat this process until we have ASSOCIATIVITY many elements N’ S :

slide-59
SLIDE 59

Group-testing algorithm

We repeat this process until we have ASSOCIATIVITY many elements N’ S :

slide-60
SLIDE 60

Group-testing algorithm

We repeat this process until we have ASSOCIATIVITY many elements N’ S :

slide-61
SLIDE 61

Group-testing algorithm

We repeat this process until we have ASSOCIATIVITY many elements N’ S :

slide-62
SLIDE 62

Group-testing algorithm

We repeat this process until we have ASSOCIATIVITY many elements N’ S :

slide-63
SLIDE 63

Group-testing algorithm

We repeat this process until we have ASSOCIATIVITY many elements N’ S :

slide-64
SLIDE 64

Group-testing algorithm

We repeat this process until we have ASSOCIATIVITY many elements N’ S :

slide-65
SLIDE 65

Group-testing algorithm

We find our minimal eviction set! ASSOCIATIVITY S :

slide-66
SLIDE 66

Group-testing algorithm

ASSOCIATIVITY S :

O(N) mem accesses

slide-67
SLIDE 67

Tool (C/x86): https://github.com/cgvwzq/evsets O(n) vs. O(n2) advantage shows up in practice! Finding minimal eviction sets is practical without knowledge on any bits

  • f the set index!

Y-right (lines): Average running time for eviction set reduction Y-left (columns): Cost of finding an initial eviction set of certain size X: Eviction set size in number of addresses Experiments on Skylake i5-6500 with 6MB cache (8192 sets x 12 assoc)

timeout

Performance Evaluation

slide-68
SLIDE 68

Robustness Evaluation

Modern replacement policies break our test assumption and introduce errors.

X: Cache set offset (each points aggregates all slices) Y: Average success rate for Green: reduction rate w/o error correcting mechanisms. Yellow: Test rate reliability Experiments on Skylake i5-6500 with 6MB cache (8192 sets x 12 assoc)

slide-69
SLIDE 69

Live demo (Bonus Material!)

Find minimal eviction sets on Chrome with JS and Wasm

Demo running on Chrome 74.0.3729.75 with V8 7.4 - CPU i7-8550U

slide-70
SLIDE 70

Conclusions

Finding minimal eviction sets is a threshold group-testing problem: new insight for research on principled countermeasures Novel linear-time algorithm makes attacks faster and enables them in scenarios previously considered impractical

slide-71
SLIDE 71

Thanks for your attention

Questions?

slide-72
SLIDE 72

Noise on difference machines and cache sets

slide-73
SLIDE 73

run.sh: google-chrome-beta --user-data-dir=/tmp/tmp.u9lo18kaTh

  • -js-flags='--allow-natives-syntax --experimental-wasm-bigint'

http://localhost:8000/ | ./verify_addr.sh

  • -allow-natives-syntax: used for printing found indices to stdout
  • -experimental-wasm-bigint: only for convenience, will have default support soon
  • verify_addr.sh:
  • find chrome’s PID
  • use pmap to find base virtual address for JS buffer
  • read JS indices and add them to virtual address base
  • execute ./virt_to_phys to translate virtual to physical addresses using /proc/pid/pagemap
  • extract slice and cache index set from physical address (uses Intel’s reverse engineered hash function)

Demo verification scripts

slide-74
SLIDE 74

0 55 push rbp 1 4889e5 movq rbp,rsp 4 6a0a push 0xa 6 56 push rsi 7 4883ec10 subq rsp,0x10 b 488b9ea7000000 movq rbx,[rsi+0xa7] 12 6666660f1f840000000000 nop 1d 0f1f00 nop 20 488b96c7000000 movq rdx,[rsi+0xc7] 27 483922 cmpq [rdx],rsp 2a 0f8340000000 jnc <+0x70> 30 8bc0 movl rax,rax 32 49ba0000000001000000 movq r10,0x100000000 3c 4c3bd0 cmpq r10,rax 3f 7320 jnc <+0x61> ... 61 488b0403 movq rax,[rbx+rax*1] 65 4883f800 cmpq rax,0x0 69 75b5 jnz <+0x20> 6b 488be5 movq rsp,rbp 6e 5d pop rbp 6f c3 retl

TurboFan x86_64 output

(func $x (param $ptr i64) (loop $iter (set_local $ptr (i64.load (i32.wrap/i64 (get_local $ptr)))) (br_if $iter (i32.eqz (i64.eqz (get_local $ptr))))) )

traverse.wat sign check stack guard epilogue prologue

(on V8 v7.4.0)

Example of TEST() in Wasm