● CVE-2024-29745 / Commercial Spyware Zero-Day

CVE-2024-29745: Arm Mali GPU Kernel Memory Race Condition Teardown

How an unlocked page table teardown in drivers/gpu/arm/midgard/mali_kbase_mem.c enabled commercial spyware to bypass Android sandboxes.

💡 Explication en Termes Simples (ELI5)

Imagine a hotel where guests return room keys to the front desk. A dishonest guest hands in their key, and while the clerk is marking the room empty in the computer, the guest's accomplice runs into the room, replaces the door lock with their own, and starts renting out the room privately without the hotel knowing.

Concepts Clés et Termes du Sous-système

Mali kbase Driver: The Linux kernel module interfacing Arm Mali GPU cores with user-space rendering libraries.
Page Table Mapping (mmap): Mapping GPU hardware physical memory pages into user-space process virtual memory addresses.
Race Condition: A timing bug where two execution threads attempt to access and modify shared resources simultaneously without adequate synchronization.
Privilege Escalation: Gaining kernel execution privileges from an unprivileged sandbox application.

Mécanique d'Exécution Étape par Étape

Step 1

1. Multi-Threaded Memory Allocation

The attacker launches two threads mapping GPU memory chunks via mmap().

Step 2

2. Race Unmap Against Access

Thread A invokes munmap() while Thread B simultaneously queues GPU drawing commands on the same region.

Step 3

3. Free Memory Access

The kernel frees the memory pages, but Thread B's GPU commands execute against the freed pages.

Step 4

4. Kernel Control Hijack

The attacker sprays page tables into the slot, rewriting GPU descriptors to access all physical device RAM.

Code Source : Faille Critique vs. Correctif Sécurisé

Fourni en code source de haut niveau lisible (sans assembleur brut ni diff binaire).

FAILLE NON CORRIGÉE

// VULNERABLE: drivers/gpu/arm/midgard/mali_kbase_mem.c
int kbase_mem_free(struct kbase_context *kctx, u64 gpu_addr) {
    struct kbase_va_region *reg = kbase_region_tracker_find(kctx, gpu_addr);
    
    // ROOT CAUSE:
    // Memory region is unmapped and freed WITHOUT holding the context's page lock!
    // Concurrent GPU commands can still dereference 'reg' while it is being destroyed!
    kbase_gpu_vm_lock(kctx);
    kbase_mem_free_region(kctx, reg); // Frees underlying pages
    kbase_gpu_vm_unlock(kctx);
    return 0;
}

CORRECTIF RENFORCÉ

// SECURE: drivers/gpu/arm/midgard/mali_kbase_mem.c patch
int kbase_mem_free(struct kbase_context *kctx, u64 gpu_addr) {
    // 1. Acquire global context lock BEFORE looking up region
    kbase_gpu_vm_lock(kctx);
    
    struct kbase_va_region *reg = kbase_region_tracker_find(kctx, gpu_addr);
    if (!reg) {
        kbase_gpu_vm_unlock(kctx);
        return -EINVAL;
    }
    
    // 2. Wait for all in-flight GPU job chains to complete before deallocating
    kbase_wait_for_in_flight_jobs(kctx, reg);
    
    kbase_mem_free_region(kctx, reg);
    kbase_gpu_vm_unlock(kctx);
    return 0;
}

Checklist d'Ingénierie et de Durcissement Système

Enforce strict lock acquisition (`kbase_gpu_vm_lock`) across all lookup and deallocation code paths.
Synchronously block and drain all active GPU execution queues before unmapping backing physical memory pages.
Use reference counting on memory region objects (`kbase_va_region`) to prevent premature destruction.
Audit multi-threaded device driver ioctl calls for asynchronous lifecycle races.

← Parcourir l'Annuaire de Sécurité Tous les Bulletins de Sécurité →