There's a remotely triggerable memory corruption issue in SwiftShader that's reachable from WebGL, resulting from an integer overflow issue.
In the GPU process there is validation on the sizes passed to texture creation functions to ensure that they shouldn't cause overflow. However, in the Swiftshader code there is a separate rounding up of render-target sizes to the next even size, which allows bypassing this validation.
(Note the additional +4, which is also (unexpected by the chrome gpu process) unsafe but in practice shouldn't cause an issue.)
https://cs.chromium.org/chromium/src/third_party/swiftshader/src/Renderer/Surface.cpp?l=3261
void *Surface::allocateBuffer(int width, int height, int depth, int border, int samples, Format format)
{
// Render targets require 2x2 quads
int width2 = (width + 1) & ~1;
int height2 = (height + 1) & ~1;
// FIXME: Unpacking byte4 to short4 in the sampler currently involves reading 8 bytes,
// and stencil operations also read 8 bytes per four 8-bit stencil values,
// so we have to allocate 4 extra bytes to avoid buffer overruns.
return allocate(size(width2, height2, depth, border, samples, format) + 4);
}
Size calculation takes place here:
https://cs.chromium.org/chromium/src/third_party/swiftshader/src/Renderer/Surface.cpp?l=2646
unsigned int Surface::size(int width, int height, int depth, int border, int samples, Format format)
{
width += 2 * border;
height += 2 * border;
// Dimensions rounded up to multiples of 4, used for compressed formats
int width4 = align(width, 4);
int height4 = align(height, 4);
switch(format)
{
// ... snip ...
default:
return bytes(format) * width * height * depth * samples;
}
}
The maximum value for bytes(format) is 16, with something like GL_RGBA32F, and samples is 1.
We can't cause this value to overflow if we have to provide the texture contents, since allocating a sufficiently large Float32Array will be larger than the renderer memory limits, but we can use glTexStorage3D to trigger the overflow.
We need to meet the following conditions:
1 <= width <= 0x2000
1 <= height <= 0x2000
1 <= depth <= 0x2000
16 * width * height * depth <= 0x100000000ull;
If these conditions are met, and we can also produce values such that:
16 * ((width + 1) & ~1) * ((height + 1) & ~1) * depth >= 0x100000000ull;
Then we'll get an integer overflow during size calculation, and end up with a small buffer for a large texture.
If we use the path glTexSubImage3D to initialize the texture, this will zero out (Chrome's expected size) of the texture (~4gig) in the (260 byte) allocation, which may make exploitation awkward, but especially in a context like the GPU process with multiple threads interacting, it's likely possible to exploit this issue. There may also be alternative paths which avoid the wild memset, but I'm reporting now so that work on a fix can start.
Note, it is possible for an attacker to force use of the Swiftshader backend for WebGL rendering by simply crashing the GPU process a few times (for a platform dependent value of 'few'). The attached PoC uses 4 domains and cycles between them to trigger 3 (hardware accelerated) GPU process crashes due to OOM (on my workstation, at least) which will then be followed by the (software accelerated) GPU process hitting this bug. Mileage may vary with different GPU drivers/OpenGL implementations.
Crashes with the PoC will be fairly random - whatever you'd expect for zeroing out your entire heap...
Thread 1 "chrome" received signal SIGSEGV, Segmentation fault.
0x00007fe697e94551 in egl::Image::loadImageData(egl::Context*, int, int, int, int, int, int, unsigned int, unsigned int, egl::Image::UnpackInfo const&, void const*) ()
from src/out/non-asan/swiftshader/libGLESv2.so
(gdb) bt
#0 0x00007fe697e94551 in egl::Image::loadImageData(egl::Context*, int, int, int, int, int, int, unsigned int, unsigned int, egl::Image::UnpackInfo const&, void const*) ()
at src/out/non-asan/swiftshader/libGLESv2.so
#1 0x0000000000000000 in ()
(gdb) x/10i $pc
=> 0x7fe697e94551 <_ZN3egl5Image13loadImageDataEPNS_7ContextEiiiiiijjRKNS0_10UnpackInfoEPKv+9911>: jmpq *0x28(%rax)
0x7fe697e94554 <_ZN12_GLOBAL__N_113LoadImageDataILNS_8DataTypeE1EEEviiiiiiiiiiPKvPv>: push %rbp
0x7fe697e94555 <_ZN12_GLOBAL__N_113LoadImageDataILNS_8DataTypeE1EEEviiiiiiiiiiPKvPv+1>: push %r15
0x7fe697e94557 <_ZN12_GLOBAL__N_113LoadImageDataILNS_8DataTypeE1EEEviiiiiiiiiiPKvPv+3>: push %r14
0x7fe697e94559 <_ZN12_GLOBAL__N_113LoadImageDataILNS_8DataTypeE1EEEviiiiiiiiiiPKvPv+5>: push %r13
0x7fe697e9455b <_ZN12_GLOBAL__N_113LoadImageDataILNS_8DataTypeE1EEEviiiiiiiiiiPKvPv+7>: push %r12
0x7fe697e9455d <_ZN12_GLOBAL__N_113LoadImageDataILNS_8DataTypeE1EEEviiiiiiiiiiPKvPv+9>: push %rbx
0x7fe697e9455e <_ZN12_GLOBAL__N_113LoadImageDataILNS_8DataTypeE1EEEviiiiiiiiiiPKvPv+10>: sub $0x48,%rsp
0x7fe697e94562 <_ZN12_GLOBAL__N_113LoadImageDataILNS_8DataTypeE1EEEviiiiiiiiiiPKvPv+14>: mov %r8d,0xc(%rsp)
0x7fe697e94567 <_ZN12_GLOBAL__N_113LoadImageDataILNS_8DataTypeE1EEEviiiiiiiiiiPKvPv+19>: test %r9d,%r9d
(gdb) i r
rax 0x0 0
rbx 0x8814 34836
rcx 0x1 1
rdx 0x10 16
rsi 0x30b20f90860 3346332715104
rdi 0x30b2081f500 3346324911360
rbp 0x1406 0x1406
rsp 0x7ffdafd862c8 0x7ffdafd862c8
r8 0xfffffffffffffff0 -16
r9 0x1 1
r10 0x75 117
r11 0x30b2088ee90 3346325368464
r12 0xc2 194
r13 0x2a3 675
r14 0x8814 34836
r15 0x0 0
rip 0x7fe697e94551 0x7fe697e94551 <egl::Image::loadImageData(egl::Context*, int, int, int, int, int, int, unsigned int, unsigned int, egl::Image::UnpackInfo const&, void const*)+9911>
eflags 0x10202 [ IF RF ]
cs 0x33 51
ss 0x2b 43
ds 0x0 0
es 0x0 0
fs 0x0 0
gs 0x0 0
(gdb)
See crash-id d0573792cf03341d for a crash on the current stable branch.
To test using the attached PoC, either run chrome with --disable-gpu to force software rendering, or create 4 aliases to localhost evil0.com, evil1.com, evil2.com and evil3.com in your /etc/hosts file and run ./server.py <port_number> and point your browser to evil0.com:<port_number>.
Proof of Concept:
https://gitlab.com/exploit-database/exploitdb-bin-sploits/-/raw/main/bin-sploits/45059.zip
