Advisory TFMV-2 =============== +----------------+-------------------------------------------------------------+ | Title | Invoking Secure functions from handler mode may cause TF-M | | | IPC model to behave unexpectedly. | +================+=============================================================+ | CVE ID | CVE-2021-27562 | +----------------+-------------------------------------------------------------+ | Date | Mar 5, 2021 | +----------------+-------------------------------------------------------------+ | Versions | Affected All versions up to and including TF-M v1.2 | | Affected | | +----------------+-------------------------------------------------------------+ | Configurations | IPC Model on Armv8-M | +----------------+-------------------------------------------------------------+ | Impact | Most likely a crash in secure world or reset whole system, | | | with a very low likelihood of overwriting some memory | | | contents. | +----------------+-------------------------------------------------------------+ | Fix Version | commit `e212ea1637a66255b44d0e7c19ebe9786ab56ccb`_ | +----------------+-------------------------------------------------------------+ | Credit | Øyvind Rønningstad, | | | Senior SW Engineer from Nordic Semiconductor | +----------------+-------------------------------------------------------------+ Background ---------- When a non-secure caller calls the secure function, the execution switches the security state via Secure Gateway (SG), then arrived at the TF-M secure entry if the SG is successful. After SG, TF-M invokes SVC to SPM at first because SPM handles requests from SPE and NSPE in a unified SVC handler. The TF-M SVC handler code relies on the 'SPSEL' bit in 'EXC_RETURN' to get the caller stack pointer: .. code-block:: asm ; Armv8m mainline as the example mrs r2, msp ; r2 = msp; tst lr, #4 ; EXC_RETURN.SPSEL == ? ite eq ; condition preparation moveq r0, r2 ; if (EXC_RETURN.SPSEL == 0) r0 = msp; movne r0, psp ; if (EXC_RETURN.SPSEL == 1) r0 = psp; mov r1, lr ; r1 = EXC_RETURN; bl tfm_core_svc_handler ; r0 = sp(context), r1 = EXC_RETURN, r2 = msp If NSPE calls the secure function in 'Handler mode', the TF-M secure entry runs in 'Handler mode' before invoking SVC because PE mode does not change during the transition from non-secure state to secure state via the successful SG. Main stack pointer (MSP) is always used in 'Handler mode', which is irrelevant to the value of SPSEL. However, the code above selects secure process stack pointer (PSP_S, S suffix here indicates the security state) for further SVC handling based on EXC_RETURN.SPSEL, hence the SVC handler accesses the wrong stack for handling the request in the NSPE caller in 'Handler mode' case. To prevent this vulnerability, TF-M secure SVC handler should fetch the right stack pointer register by checking both the PE mode and SPSEL bit. The handler should also prevent callers in non-secure `Handler mode` from calling TF-M SVC functionalities. The following section analysis impact of the IPC model. Library model is not vulnerable to this attack because it checks the PE mode in the TF-M secure entry. Impact ------ When the vulnerability is happening, PSP_S is the incorrect stack pointer fetched, the impact is decided by the content PSP_S is pointing to. The SVC handler gets the SVC number by referencing the memory at the 2 bytes subtracted position of member 'Return Address' in the PSP_S pointing context, and uses the caller saved registers in the context as parameters for the subsequent functions indicated by the SVC number. The PSP_S may point to the bottom (higher address) of secure thread stack if there is no ongoing secure function call, or it points to a preempted context caused by non-secure preempting secure execution. When PSP_S is pointing to the stack bottom when this issue happens, the caller context is pointing to the underflowed stack of which the top 2 words are stack seal, and the rest of the context is unpredictable. These underflowed content are ZERO initialized for most platforms, hence when fetching the SVC Number at memory address 'Return Address - 2', a 'MemFault' or 'HardFault' would happen. Even if a valid SVC number is returned, the stack seal values are part of the parameters for the subsequent functions that have parameters, which cannot pass the validation for the parameters. When PSP_S is pointing to a preempted context, the preempted place can be the TF-M secure entry area or internal thread space. If the preemption happens when the execution is in the TF-M secure entry area, the PSP_S will point to an NS preemption stack frame and hence will fail the context size validation through this entry into TF-M SVC Handler. In the other scenario that TF-M internal thread is preempted, there is a very remote chance that the exception stack frame will contain a context with a valid SVC Number with valid parameters. As a summary, the impacts of the vulnerability depend on whether an SVC number could be fetched from the incorrect stack and what operation the fetched SVC number corresponds to: - A memory access fault when fetching the SVC number from memory. This fault halts the whole system. A reset can recover the system back to normal. - The SVC number corresponds to the initialization process in TF-M. SPM initialization is called for the second time. This re-initialization process will fail and halt the whole system. A reset can recover the system back to normal. - The SVC number corresponds to the boot data fetching function. If the unpredictable data in stack triggers boot data fetching and passes the parameter validation under a very rare case, boot data is copied into an unpredictable memory address. One note here, the TF-M default boot data has no sensitive information. - The SVC number corresponds to the log output function. If the unpredictable data in stack triggers log output under a very rare case, secure data at unpredictable address might be printed out through the log interface. - The SVC number corresponds to other valid numbers other than above. The system resets because of parameter validation fail. Conclusion ---------- Overall, from the analysis of upstream TF-M implementation, the severity of this vulnerability is Medium, given that a software crash or system reset happen most of the time. The probability for causing secure data leakage via log output or tampering of secure memory with boot data is extremely low as the TF-M internal thread exception stack frame contents have to pass all the validations performed by SPM. .. _e212ea1637a66255b44d0e7c19ebe9786ab56ccb: https://git.trustedfirmware.org/TF-M/trusted-firmware-m.git/commit/?id=e212ea1637a66255b44d0e7c19ebe9786ab56ccb -------------- *Copyright (c) 2021, Arm Limited. All rights reserved.*