Stale TLB entries after demand-paging page-out let arm64 user threads access reused physical frames

Overview

The arm64 MMU backend invalidates a single virtual address from the TLB with invalidate_tlb_page() in arch/arm64/core/mmu.c, which issued TLBI VAE1{IS} with an operand of virt >> 12. That operand encodes the ASID in bits [63:48], so the invalidation only ever matched ASID 0 (the kernel page tables) and global entries. Memory domains, however, are allocated ASIDs starting at 1 in arch_mem_domain_init(), and every domain-private leaf created by private_map() is mapped with MT_NG, i.e. non-global. A domain's cached translation therefore survived the invalidation.

All callers of invalidate_tlb_page() are demand-paging paths. When the eviction algorithm reclaims a frame, page_frame_prepare_locked() in kernel/mmu.c calls arch_mem_page_out(), which marks the kernel leaf invalid, propagates the invalid descriptor into every domain via sync_domains(), and then flushes only ASID 0. do_page_fault() immediately reuses the very same page frame for the incoming mapping through arch_mem_page_in(). A user thread that had already touched the evicted page — a k_mem_partition added with arch_mem_domain_partition_add(), or its own stack mapped by map_thread_stack() — retains a valid EL0 read/write translation to that physical frame under its own ASID.

Unprivileged code running in a user thread can drive memory pressure to force eviction and then keep accessing the stale address, obtaining read and write access to whatever the kernel subsequently placed in that frame: supervisor data, another memory domain's partition, or another thread's stack. This defeats the isolation that CONFIG_USERSPACE is meant to enforce, allowing information disclosure and memory corruption up to full compromise of the system.

The fix switches the implementation to TLBI VAAE1{IS}, which matches every ASID and both global and non-global entries, so no domain-tagged translation survives a page out, dirty-bit update or access-flag clear. Exploitation requires an arm64 build with both CONFIG_USERSPACE and CONFIG_DEMAND_PAGING enabled, an active backing store, and at least one demand-paged page that is also privately mapped into a user memory domain. The same commit additionally stops globalize_table() from stripping EL0 permission and non-global attributes off domain-private leaves during paging synchronization, which previously caused user threads to lose access to their own partitions and stacks.

Affected

• zephyr >= 4.3.0 < 4.5.0

Remediation

Refer to the linked advisories for vendor-supplied fixes and affected version ranges.