[RFC PATCH v3 0/4] arm64: Stack trace reliability checks in the unwinder

Mon May 3 10:36:11 PDT 2021

From: "Madhavan T. Venkataraman" <madvenka at linux.microsoft.com>

There are a number of places in kernel code where the stack trace is not
reliable. Enhance the unwinder to check for those cases and mark the
stack trace as unreliable. Once all of the checks are in place, the unwinder
can provide a reliable stack trace. But before this can be used for livepatch,
some other entity needs to guarantee that the frame pointers are all set up
correctly in kernel functions. objtool is currently being worked on to
address that need.

Return address check
====================

Check the return PC of every stack frame to make sure that it is a valid
kernel text address (and not some generated code, for example). If it is
not a valid kernel text address, mark the stack trace as unreliable.

Assembly functions
==================

There are a number of assembly functions in arm64. Except for a couple of
them, these functions do not have a frame pointer prolog or epilog. Also,
many of them manipulate low-level state such as registers. These functions
are, by definition, unreliable from a stack unwinding perspective. That is,
when these functions occur in a stack trace, the unwinder would not be able
to unwind through them reliably.

Assembly functions are defined as SYM_FUNC_*() functions or SYM_CODE_*()
functions. objtool peforms static analysis of SYM_FUNC functions. It ignores
SYM_CODE functions because they have low level code that is difficult to
analyze. When objtool becomes ready eventually, SYM_FUNC functions will
be analyzed and "fixed" as necessary. So, they are not "interesting" for
the reliable unwinder.

That leaves SYM_CODE functions. It is for the unwinder to deal with these
for reliable stack trace. The unwinder needs to do the following:

	- Recognize SYM_CODE functions in a stack trace.

	- If a particular SYM_CODE function can be unwinded through using
	  some special logic, then do it. E.g., the return trampoline for
	  Function Graph Tracing.

	- Otherwise, mark the stack trace unreliable.

Using text sections to recognize SYM_CODE functions
===================================================

SYM_CODE functions are present in the following text sections:

	(1) .entry.text
	(2) .idmap.text
	(3) .hyp.idmap.text
	(4) .hyp.text
	(5) .hibernate_exit.text
	(6) .entry.tramp.text
	(7) .text
	(8) .init.text

For each of these sections, there are global variables that contain the
starting and ending addresses generated by the linker. So, they can be
recognized easily. Create an array called sym_code_ranges[] to contain
the ranges for sections (1) thru (6).

Check if the return PC falls in any of these sections. If it does, mark
the stack trace unreliable.

Sections (7) and (8)
====================

Sections (7) and (8) are generic sections which also contain tons of other
functions which are actually reliable. The SYM_CODE functions in these
sections must be dealt with differently.

Some of these are "don't care" functions so they can be left in their
current sections. E.g.,

	efi_enter_kernel().
	arm64_relocate_new_kernel()

		These functions enter the kernel. So, they will not occur
		in a stack trace examined by livepatch.

	__kernel_rt_sigreturn()

		This only gets invoked in user context.

	hypervisor vectors

		I don't believe these are interesting to livepatch.

Others need to be moved to a special section. I have created a "catch-all"
section called ".code.text" for this purpose. Add this section to
vmlinux.lds.S and sym_code_ranges[].

Reliable SYM_CODE functions
===========================

The unwinder currently has logic to recognize the return trampoline of the
Function Graph Tracer (return_to_handler()), retrieve the original return
address and use that in the stack trace. So, this is a SYM_CODE function
that the unwinder can actually unwind through.

However, the current logic only addreses stack traces taken while still in
the traced function. When the traced function returns, control is transferred
to return_to_handler(). Any stack trace taken while in the return trampoline
is not handled. This messes up the stack trace as the unwinder has to keep
track of the current index within the return address stack.

There are two options:

	- Either remove this logic altogether. This would work since the
	  unwinder would recognize the section of the trampoline and
	  treat the stack trace as unreliable. So, from a live patch
	  perspective, this is sufficient.

	- Or, fix the logic. I have taken this approach. See the patch
	  for more details. That said, I am open to option 1.

Other cases like kretprobe_trampoline() and ftrace entry code can be
addressed in a similar fashion. But this is outside the scope of this
work. The only reason I fixed the logic for return_to_handler() in the
unwinder is because the logic is already there.

Last stack frame
================

If a SYM_CODE function occurs in the very last frame in the stack trace,
then the stack trace is not considered unreliable. This is because there
is no more unwinding to do. Examples:

	- EL0 exception stack traces end in the top level EL0 exception
	  handlers.

	- All kernel thread stack traces end in ret_from_fork().

Special cases
=============

Some special cases need to be mentioned.

EL1 interrupt and exception handlers are present in .entry.text. So, all
EL1 interrupt and exception stack traces will be considered unreliable.
This the correct behavior as interrupts and exceptions can happen on any
instruction including ones in the frame pointer prolog and epilog. Unless
objtool generates metadata so the unwinder can unwind through these
special cases, such stack traces will be considered unreliable.

A task can get preempted at the end of an interrupt. Stack traces of
preempted tasks will show the interrupt frame in the stack trace and
will be considered unreliable.

Breakpoints are exceptions. So, all stack traces in the break point
handler (including probes) will be considered unreliable.

All of the ftrace trampoline code that gets executed at the beginning
of a traced function falls in ".code.text". All stack traces taken from
tracer functions will be considered unreliable.

The same is true for kretprobe trampolines.

---
Changelog:

v3:
	- Implemented a sym_code_ranges[] array to contains sections bounds
	  for text sections that contain SYM_CODE_*() functions. The unwinder
	  checks each return PC against the sections. If it falls in any of
	  the sections, the stack trace is marked unreliable.

	- Moved SYM_CODE functions from .text and .init.text into a new
	  text section called ".code.text". Added this section to
	  vmlinux.lds.S and sym_code_ranges[].

	- Fixed the logic in the unwinder that handles Function Graph
	  Tracer return trampoline.

	- Removed all the previous code that handles:
		- ftrace entry code for traced function
		- special_functions[] array that lists individual functions
		- kretprobe_trampoline() special case

	- Synced with mainline v5.12-rc8.

v2
	- Removed the terminating entry { 0, 0 } in special_functions[]
	  and replaced it with the idiom { /* sentinel */ }.

	- Change the ftrace trampoline entry ftrace_graph_call in
	  special_functions[] to ftrace_call + 4 and added explanatory
	  comments.

	- Unnested #ifdefs in special_functions[] for FTRACE.

v1
	- Define a bool field in struct stackframe. This will indicate if
	  a stack trace is reliable.

	- Implement a special_functions[] array that will be populated
	  with special functions in which the stack trace is considered
	  unreliable.

	- Using kallsyms_lookup(), get the address ranges for the special
	  functions and record them.

	- Implement an is_reliable_function(pc). This function will check
	  if a given return PC falls in any of the special functions. If
	  it does, the stack trace is unreliable.

	- Implement check_reliability() function that will check if a
	  stack frame is reliable. Call is_reliable_function() from
	  check_reliability().

	- Before a return PC is checked against special_funtions[], it
	  must be validates as a proper kernel text address. Call
	  __kernel_text_address() from check_reliability().

	- Finally, call check_reliability() from unwind_frame() for
	  each stack frame.

	- Add EL1 exception handlers to special_functions[].

		el1_sync();
		el1_irq();
		el1_error();
		el1_sync_invalid();
		el1_irq_invalid();
		el1_fiq_invalid();
		el1_error_invalid();

	- The above functions are currently defined as LOCAL symbols.
	  Make them global so that they can be referenced from the
	  unwinder code.

	- Add FTRACE trampolines to special_functions[]:

		ftrace_graph_call()
		ftrace_graph_caller()
		return_to_handler()

	- Add the kretprobe trampoline to special functions[]:

		kretprobe_trampoline()

Previous versions and discussion
================================

v2: https://lore.kernel.org/linux-arm-kernel/20210405204313.21346-1-madvenka@linux.microsoft.com/
v1: https://lore.kernel.org/linux-arm-kernel/20210330190955.13707-1-madvenka@linux.microsoft.com/

Madhavan T. Venkataraman (4):
  arm64: Introduce stack trace reliability checks in the unwinder
  arm64: Check the return PC against unreliable code sections
  arm64: Handle miscellaneous functions in .text and .init.text
  arm64: Handle funtion graph tracer better in the unwinder

 arch/arm64/include/asm/sections.h             |   1 +
 arch/arm64/include/asm/stacktrace.h           |   7 +
 arch/arm64/kernel/entry-ftrace.S              |   5 +
 arch/arm64/kernel/entry.S                     |   6 +
 arch/arm64/kernel/head.S                      |   3 +-
 arch/arm64/kernel/probes/kprobes_trampoline.S |   2 +
 arch/arm64/kernel/stacktrace.c                | 129 ++++++++++++++++--
 arch/arm64/kernel/vmlinux.lds.S               |   7 +
 8 files changed, 149 insertions(+), 11 deletions(-)

base-commit: bf05bf16c76bb44ab5156223e1e58e26dfe30a88
-- 
2.25.1