OS X IOKit kIOMapReadOnly read-only kernel shared memory bypass leading to kernel memory corruption bug in IOAccelContext2

[GoogleCodeExporter]

The IOAccelContext2::clientMemoryForType accepts type value from 0 to 3. The 
code path for type=2 sets the kIOMapReadOnly flag of the IOOptionBits reference 
passed as the second argument by mapClientMemory64.

This flag is presumably supposed to enforce that the userspace mapping of this 
shared memory is read-only, and by default that is the case.

I was auditing more uses of shared memory and noticed that the kernel was 
trusting values in this read-only shared memory and wondered how this was 
enforced so took a look at the code responsible for handling the kIOMapReadOnly 
flag:

kIOMapReadOnly is used here in IOMemoryDescriptor.cpp:

        IOOptionBits createOptions = 0;
        if (!(kIOMapReadOnly & options))
        {
            createOptions |= kIOMemoryReferenceWrite;

As you can see, the flag is only used to prevent kIOMemoryReferenceWrite from 
being ORed onto the createOptions. 

later in IOMemoryDescriptor.cpp:

    // cache mode & vm_prot
    prot = VM_PROT_READ;
    cacheMode = ((_flags & kIOMemoryBufferCacheMask) >> kIOMemoryBufferCacheShift);
    prot |= vmProtForCacheMode(cacheMode);
    // VM system requires write access to change cache mode
    if (kIODefaultCache != cacheMode)                    prot |= VM_PROT_WRITE;
    if (kIODirectionOut != (kIODirectionOutIn & _flags)) prot |= VM_PROT_WRITE;
    if (kIOMemoryReferenceWrite & options)               prot |= VM_PROT_WRITE;

It turns out that kIOMemoryReferenceWrite is only one of the ways to get 
VM_PROT_WRITE set in the eventual protection flags used for the mapping - if we 
can specify a non-default cache mode then the mapping will also be writable, 
even if kIOMapReadOnly was specified.

The 6th argument to IOConnectMapMemory is an IOOptionBits, looking at IOTypes.h 
we can see the flags which we can pass from userspace:

enum {
    kIODefaultCache   = 0,
    kIOInhibitCache   = 1,
    kIOWriteThruCache   = 2,
    kIOCopybackCache    = 3,
    kIOWriteCombineCache  = 4,
    kIOCopybackInnerCache = 5
};

// IOMemory mapping options
enum {
    kIOMapAnywhere    = 0x00000001,

    kIOMapCacheMask   = 0x00000700,
    kIOMapCacheShift    = 8,
    kIOMapDefaultCache    = kIODefaultCache       << kIOMapCacheShift,
    kIOMapInhibitCache    = kIOInhibitCache       << kIOMapCacheShift,
    kIOMapWriteThruCache  = kIOWriteThruCache     << kIOMapCacheShift,
    kIOMapCopybackCache   = kIOCopybackCache      << kIOMapCacheShift,
    kIOMapWriteCombineCache = kIOWriteCombineCache  << kIOMapCacheShift,
    kIOMapCopybackInnerCache  = kIOCopybackInnerCache << kIOMapCacheShift,

    kIOMapUserOptionsMask = 0x00000fff,
...

mapClientMemory64 enforces the kIOMapUserOptionsMask but this still lets us 
specify kIOWriteThruCache. By specifying this non-default cache mode in the 
call to IOConnectMapMemory the read-only mapping is now writeable :)

This means the following code now has a bug :) :

Selector 5 of IOAccelContext2 is finish_fence_event:

__text:00000000000046A0 ; __int64 __fastcall 
IOAccelContext2::finish_fence_event(IOAccelContext2 *__hidden this, unsigned 
int)
__text:00000000000046A0                 public 
__ZN15IOAccelContext218finish_fence_eventEj
__text:00000000000046A0 __ZN15IOAccelContext218finish_fence_eventEj proc near
__text:00000000000046A0                                         ; DATA XREF: 
__const:000000000003C478o
__text:00000000000046A0                 push    rbp
__text:00000000000046A1                 mov     rbp, rsp
__text:00000000000046A4                 push    rbx
__text:00000000000046A5                 push    rax
__text:00000000000046A6                 mov     rbx, rdi
__text:00000000000046A9                 mov     ecx, [rbx+628h]
__text:00000000000046AF                 test    ecx, ecx
__text:00000000000046B1                 mov     eax, 0E00002C2h
__text:00000000000046B6                 jz      short loc_46FE
__text:00000000000046B8                 shr     ecx, 6
__text:00000000000046BB                 cmp     ecx, esi
__text:00000000000046BD                 jb      short loc_46FE
__text:00000000000046BF                 mov     esi, esi
__text:00000000000046C1                 mov     rax, [rbx+518h]
__text:00000000000046C8                 mov     rdi, [rax+360h]
__text:00000000000046CF                 mov     rax, [rdi]
__text:00000000000046D2                 shl     rsi, 6
__text:00000000000046D6                 add     rsi, [rbx+5F8h]      ; +5F8h == 
pointer to kernel mapping of type 2 shared mem
__text:00000000000046DD                 call    qword ptr [rax+1B8h] ; 
IOAccelEventMachineFast2::finishEventUnlocked

this external method takes one scalar argument which is bounds checked then 
added to the pointer to the type 2 shared memory. This pointer into shared 
memory is passed to the virtual IOAccelEventMachineFast2::finishEventUnlocked 
function:

__text:000000000001E580 ; 
IOAccelEventMachineFast2::finishEventUnlocked(IOAccelEvent *)
__text:000000000001E580                 public 
__ZN24IOAccelEventMachineFast219finishEventUnlockedEP12IOAccelEvent
__text:000000000001E580 
__ZN24IOAccelEventMachineFast219finishEventUnlockedEP12IOAccelEvent proc near
__text:000000000001E580                                         ; DATA XREF: 
__const:0000000000042B48o
__text:000000000001E580
__text:000000000001E580 var_50          = qword ptr -50h
__text:000000000001E580 var_48          = qword ptr -48h
__text:000000000001E580 var_40          = qword ptr -40h
__text:000000000001E580 var_38          = qword ptr -38h
__text:000000000001E580 var_30          = qword ptr -30h
__text:000000000001E580
__text:000000000001E580                 push    rbp
__text:000000000001E581                 mov     rbp, rsp
__text:000000000001E584                 push    r15
__text:000000000001E586                 push    r14
__text:000000000001E588                 push    r13
__text:000000000001E58A                 push    r12
__text:000000000001E58C                 push    rbx
__text:000000000001E58D                 sub     rsp, 28h
__text:000000000001E591                 mov     [rbp+var_50], rsi ; pointer to 
shared mem
__text:000000000001E595                 mov     rbx, rdi
__text:000000000001E598                 xor     r14d, r14d
__text:000000000001E59B                 xor     eax, eax
__text:000000000001E59D
__text:000000000001E59D loc_1E59D:                              ; CODE XREF: 
IOAccelEventMachineFast2::finishEventUnlocked(IOAccelEvent *)+EEj
__text:000000000001E59D                 mov     r12, [rsi+r14*8]  ; reading 
qword from shared mem
__text:000000000001E5A1                 cmp     r12d, 0FFFFFFFFh
__text:000000000001E5A5                 jz      loc_1E667
__text:000000000001E5AB                 mov     r13, r12        ; r12d = lower 
32 bits of shared mem value
__text:000000000001E5AB                                         ; r13d = upper 
32 bits of shared mem value
__text:000000000001E5AE                 shr     r13, 20h
__text:000000000001E5B2                 movsxd  rdi, r12d
__text:000000000001E5B5                 lea     rcx, [rdi+rdi*2]   ; rcx 
controlled
__text:000000000001E5B9                 mov     edx, r13d
__text:000000000001E5BC                 sub     edx, [rbx+rcx*4+0C8h] ; 
controlled read
__text:000000000001E5C3                 test    edx, edx
__text:000000000001E5C5                 jle     loc_1E667
__text:000000000001E5CB                 lea     r15, [rbx+rcx*4+0C8h] ; save 
address of previous controlled read
__text:000000000001E5D3                 mov     rcx, [rbx+28h]
__text:000000000001E5D7                 mov     rcx, [rcx+rdi*8]
__text:000000000001E5DB                 mov     edx, [rcx]
__text:000000000001E5DD                 mov     [r15], edx            ; write a 
different value there

As you can see, the shared memory is trusted to only contain valid indexes 
which are used for a series of memory reads and writes with no bounds checking.

This PoC hooks IOConnectMapMemory to set the kIOWriteThruCache flag and then 
trigger the bug.

tested on: MacBookAir5,2 w/ 10.10.1 (14B25)

Original issue reported on code.google.com by ianb...@google.com on 9 Dec 2014 at 4:01

Attachments:

• finish_fence_event.c

[GoogleCodeExporter]

Original comment by ianb...@google.com on 9 Dec 2014 at 4:05

• Changed title: OS X IOKit kIOMapReadOnly read-only kernel shared memory bypass leading to kernel memory corruption bug in IOAccelContext2
• Added labels: ****
• Removed labels: ****

[GoogleCodeExporter]

Original comment by ianb...@google.com on 9 Dec 2014 at 4:07

• Added labels: Reported-2014-Dec-09
• Removed labels: ****

[GoogleCodeExporter]

Original comment by ianb...@google.com on 9 Dec 2014 at 4:07

• Added labels: Id-615684884
• Removed labels: ****

[GoogleCodeExporter]

Original comment by ianb...@google.com on 5 Feb 2015 at 12:14

• Changed state: Fixed
• Added labels: CVE-2014-4495, Fixed-2015-Jan-27
• Removed labels: Restrict-View-Commit

[GoogleCodeExporter]

Apple advisory: http://support.apple.com/en-us/HT204244

Original comment by ianb...@google.com on 5 Feb 2015 at 12:20

• Added labels: ****
• Removed labels: ****

[GoogleCodeExporter]

Apple advisory (iOS): http://support.apple.com/en-us/HT204245

Original comment by ianb...@google.com on 5 Feb 2015 at 12:22

• Added labels: ****
• Removed labels: ****