A befunge interpreter implemented in kernel space.
The kernel module allocates a program state for each opened fd. When calling ioctl to start executing, it puts the task function into a workqueue and starts fungez_worker asynchronously.
However, its global structure (stored in in inode structure) contains a pointer of mmap area, which could be modified during execution by calling start ioctl on another fd.
To exploit this bug, basically we need to open 2 fds, one mmap 0x2000 memory area (fd1) and the other 0x1000 bytes memory area (fd2). Then we call start ioctl on fd1 and fd2, these two fungez_worker tasks are added into workqueue. When fungez_worker is working on fd1 execution and not finishes yet, the start ioctl of fd2 replaces the mem_ptr in global structure. Thus when fungez_worker continues working on fd1 execution, it actually works on fd2's memory area, which is 0x1000 bytes long. But the limit mem_size is still 0x2000, then we could cause out-of-bound read and write over the area.
Since 0x2000 and 0x1000 bytes chunks are kmalloced, there are fd pointer when they are freed. So I choose to leverage this pointer to leak kmalloc-4k mem cache address. I searched this slab and found some pci related structure containing pointers to kernel base. Then I corrupted the fd pointer to get arbitrary write.
The pitfall is kmalloced memory is filled with null bytes. When I try to modify modprobe_path variable, it also cleans other data around it, which messes the kernel state. To overcome this, I dump the memory when running without kaslr, and dynamically fix all pointers in this page after leaking kernel base.
Offical repo: https://github.com/o-o-overflow/dc2020q-fungez-public