===================
Reliable Stacktrace
===================

This document outlines basic information about reliable stacktracing.

.. Table of Contents:

.. contents:: :local:

1. Introduction
===============

The kernel livepatch consistency model relies on accurately identifying which
functions may have live state and therefore may not be safe to patch. One way
to identify which functions are live is to use a stacktrace.

Existing stacktrace code may not always give an accurate picture of all
functions with live state, and best-effort approaches which can be helpful for
debugging are unsound for livepatching. Livepatching depends on architectures
to provide a *reliable* stacktrace which ensures it never omits any live
functions from a trace.


2. Requirements
===============

Architectures must implement one of the reliable stacktrace functions.
Architectures using CONFIG_ARCH_STACKWALK must implement
'arch_stack_walk_reliable', and other architectures must implement
'save_stack_trace_tsk_reliable'.

Principally, the reliable stacktrace function must ensure that either:

* The trace includes all functions that the task may be returned to, and the
  return code is zero to indicate that the trace is reliable.

* The return code is non-zero to indicate that the trace is not reliable.

.. note::
   In some cases it is legitimate to omit specific functions from the trace,
   but all other functions must be reported. These cases are described in
   futher detail below.

Secondly, the reliable stacktrace function must be robust to cases where
the stack or other unwind state is corrupt or otherwise unreliable. The
function should attempt to detect such cases and return a non-zero error
code, and should not get stuck in an infinite loop or access memory in
an unsafe way.  Specific cases are described in further detail below.


3. Compile-time analysis
========================

To ensure that kernel code can be correctly unwound in all cases,
architectures may need to verify that code has been compiled in a manner
expected by the unwinder. For example, an unwinder may expect that
functions manipulate the stack pointer in a limited way, or that all
functions use specific prologue and epilogue sequences. Architectures
with such requirements should verify the kernel compilation using
objtool.

In some cases, an unwinder may require metadata to correctly unwind.
Where necessary, this metadata should be generated at build time using
objtool.


4. Considerations
=================

The unwinding process varies across architectures, their respective procedure
call standards, and kernel configurations. This section describes common
details that architectures should consider.

4.1 Identifying successful termination
--------------------------------------

Unwinding may terminate early for a number of reasons, including:

* Stack or frame pointer corruption.

* Missing unwind support for an uncommon scenario, or a bug in the unwinder.

* Dynamically generated code (e.g. eBPF) or foreign code (e.g. EFI runtime
  services) not following the conventions expected by the unwinder.

To ensure that this does not result in functions being omitted from the trace,
even if not caught by other checks, it is strongly recommended that
architectures verify that a stacktrace ends at an expected location, e.g.

* Within a specific function that is an entry point to the kernel.

* At a specific location on a stack expected for a kernel entry point.

* On a specific stack expected for a kernel entry point (e.g. if the
  architecture has separate task and IRQ stacks).

4.2 Identifying unwindable code
-------------------------------

Unwinding typically relies on code following specific conventions (e.g.
manipulating a frame pointer), but there can be code which may not follow these
conventions and may require special handling in the unwinder, e.g.

* Exception vectors and entry assembly.

* Procedure Linkage Table (PLT) entries and veneer functions.

* Trampoline assembly (e.g. ftrace, kprobes).

* Dynamically generated code (e.g. eBPF, optprobe trampolines).

* Foreign code (e.g. EFI runtime services).

To ensure that such cases do not result in functions being omitted from a
trace, it is strongly recommended that architectures positively identify code
which is known to be reliable to unwind from, and reject unwinding from all
other code.

Kernel code including modules and eBPF can be distinguished from foreign code
using '__kernel_text_address()'. Checking for this also helps to detect stack
corruption.

There are several ways an architecture may identify kernel code which is deemed
unreliable to unwind from, e.g.

* Placing such code into special linker sections, and rejecting unwinding from
  any code in these sections.

* Identifying specific portions of code using bounds information.

4.3 Unwinding across interrupts and exceptions
----------------------------------------------

At function call boundaries the stack and other unwind state is expected to be
in a consistent state suitable for reliable unwinding, but this may not be the
case part-way through a function. For example, during a function prologue or
epilogue a frame pointer may be transiently invalid, or during the function
body the return address may be held in an arbitrary general purpose register.
For some architectures this may change at runtime as a result of dynamic
instrumentation.

If an interrupt or other exception is taken while the stack or other unwind
state is in an inconsistent state, it may not be possible to reliably unwind,
and it may not be possible to identify whether such unwinding will be reliable.
See below for examples.

Architectures which cannot identify when it is reliable to unwind such cases
(or where it is never reliable) must reject unwinding across exception
boundaries. Note that it may be reliable to unwind across certain
exceptions (e.g. IRQ) but unreliable to unwind across other exceptions
(e.g. NMI).

Architectures which can identify when it is reliable to unwind such cases (or
have no such cases) should attempt to unwind across exception boundaries, as
doing so can prevent unnecessarily stalling livepatch consistency checks and
permits livepatch transitions to complete more quickly.

4.4 Rewriting of return addresses
---------------------------------

Some trampolines temporarily modify the return address of a function in order
to intercept when that function returns with a return trampoline, e.g.

* An ftrace trampoline may modify the return address so that function graph
  tracing can intercept returns.

* A kprobes (or optprobes) trampoline may modify the return address so that
  kretprobes can intercept returns.

When this happens, the original return address will not be in its usual
location. For trampolines which are not subject to live patching, where an
unwinder can reliably determine the original return address and no unwind state
is altered by the trampoline, the unwinder may report the original return
address in place of the trampoline and report this as reliable. Otherwise, an
unwinder must report these cases as unreliable.

Special care is required when identifying the original return address, as this
information is not in a consistent location for the duration of the entry
trampoline or return trampoline. For example, considering the x86_64
'return_to_handler' return trampoline:

.. code-block:: none

   SYM_CODE_START(return_to_handler)
           UNWIND_HINT_EMPTY
           subq  $24, %rsp

           /* Save the return values */
           movq %rax, (%rsp)
           movq %rdx, 8(%rsp)
           movq %rbp, %rdi

           call ftrace_return_to_handler

           movq %rax, %rdi
           movq 8(%rsp), %rdx
           movq (%rsp), %rax
           addq $24, %rsp
           JMP_NOSPEC rdi
   SYM_CODE_END(return_to_handler)

While the traced function runs its return address on the stack points to
the start of return_to_handler, and the original return address is stored in
the task's cur_ret_stack. During this time the unwinder can find the return
address using ftrace_graph_ret_addr().

When the traced function returns to return_to_handler, there is no longer a
return address on the stack, though the original return address is still stored
in the task's cur_ret_stack. Within ftrace_return_to_handler(), the original
return address is removed from cur_ret_stack and is transiently moved
arbitrarily by the compiler before being returned in rax. The return_to_handler
trampoline moves this into rdi before jumping to it.

Architectures might not always be able to unwind such sequences, such as when
ftrace_return_to_handler() has removed the address from cur_ret_stack, and the
location of the return address cannot be reliably determined.

It is recommended that architectures unwind cases where return_to_handler has
not yet been returned to, but architectures are not required to unwind from the
middle of return_to_handler and can report this as unreliable. Architectures
are not required to unwind from other trampolines which modify the return
address.

4.5 Obscuring of return addresses
---------------------------------

Some trampolines do not rewrite the return address in order to intercept
returns, but do transiently clobber the return address or other unwind state.

For example, the x86_64 implementation of optprobes patches the probed function
with a JMP instruction which targets the associated optprobe trampoline. When
the probe is hit, the CPU will branch to the optprobe trampoline, and the
address of the probed function is not held in any register or on the stack.

Similarly, the arm64 implementation of DYNAMIC_FTRACE_WITH_REGS patches traced
functions with the following:

.. code-block:: none

   MOV X9, X30
   BL <trampoline>

The MOV saves the link register (X30) into X9 to preserve the return address
before the BL clobbers the link register and branches to the trampoline. At the
start of the trampoline, the address of the traced function is in X9 rather
than the link register as would usually be the case.

Architectures must either ensure that unwinders either reliably unwind
such cases, or report the unwinding as unreliable.

4.6 Link register unreliability
-------------------------------

On some other architectures, 'call' instructions place the return address into a
link register, and 'return' instructions consume the return address from the
link register without modifying the register. On these architectures software
must save the return address to the stack prior to making a function call. Over
the duration of a function call, the return address may be held in the link
register alone, on the stack alone, or in both locations.

Unwinders typically assume the link register is always live, but this
assumption can lead to unreliable stack traces. For example, consider the
following arm64 assembly for a simple function:

.. code-block:: none

   function:
           STP X29, X30, [SP, -16]!
           MOV X29, SP
           BL <other_function>
           LDP X29, X30, [SP], #16
           RET

At entry to the function, the link register (x30) points to the caller, and the
frame pointer (X29) points to the caller's frame including the caller's return
address. The first two instructions create a new stackframe and update the
frame pointer, and at this point the link register and the frame pointer both
describe this function's return address. A trace at this point may describe
this function twice, and if the function return is being traced, the unwinder
may consume two entries from the fgraph return stack rather than one entry.

The BL invokes 'other_function' with the link register pointing to this
function's LDR and the frame pointer pointing to this function's stackframe.
When 'other_function' returns, the link register is left pointing at the BL,
and so a trace at this point could result in 'function' appearing twice in the
backtrace.

Similarly, a function may deliberately clobber the LR, e.g.

.. code-block:: none

   caller:
           STP X29, X30, [SP, -16]!
           MOV X29, SP
           ADR LR, <callee>
           BLR LR
           LDP X29, X30, [SP], #16
           RET

The ADR places the address of 'callee' into the LR, before the BLR branches to
this address. If a trace is made immediately after the ADR, 'callee' will
appear to be the parent of 'caller', rather than the child.

Due to cases such as the above, it may only be possible to reliably consume a
link register value at a function call boundary. Architectures where this is
the case must reject unwinding across exception boundaries unless they can
reliably identify when the LR or stack value should be used (e.g. using
metadata generated by objtool).