When CONFIG_FUNCTION_GRAPH_TRACER is selected and the function graph
tracer is in use, unwind_frame() may erroneously associate a traced
function with an incorrect return address. This can happen when starting
an unwind from a pt_regs, or when unwinding across an exception
boundary.
This can be seen when recording with perf while the function graph
tracer is in use. For example:
| # echo function_graph > /sys/kernel/debug/tracing/current_tracer
| # perf record -g -e raw_syscalls:sys_enter:k /bin/true
| # perf report
... reports the callchain erroneously as:
| el0t_64_sync
| el0t_64_sync_handler
| el0_svc_common.constprop.0
| perf_callchain
| get_perf_callchain
| syscall_trace_enter
| syscall_trace_enter
... whereas when the function graph tracer is not in use, it reports:
| el0t_64_sync
| el0t_64_sync_handler
| el0_svc
| do_el0_svc
| el0_svc_common.constprop.0
| syscall_trace_enter
| syscall_trace_enter
The underlying problem is that ftrace_graph_get_ret_stack() takes an
index offset from the most recent entry added to the fgraph return
stack. We start an unwind at offset 0, and increment the offset each
time we encounter a rewritten return address (i.e. when we see
`return_to_handler`). This is broken in two cases:
1) Between creating a pt_regs and starting the unwind, function calls
may place entries on the stack, leaving an arbitrary offset which we
can only determine by performing a full unwind from the caller of the
unwind code (and relying on none of the unwind code being
instrumented).
This can result in erroneous entries being reported in a backtrace
recorded by perf or kfence when the function graph tracer is in use.
Currently show_regs() is unaffected as dump_backtrace() performs an
initial unwind.
2) When unwinding across an exception boundary (whether continuing an
unwind or starting a new unwind from regs), we currently always skip
the LR of the interrupted context. Where this was live and contained
a rewritten address, we won't consume the corresponding fgraph ret
stack entry, leaving subsequent entries off-by-one.
This can result in erroneous entries being reported in a backtrace
performed by any in-kernel unwinder when that backtrace crosses an
exception boundary, with entries after the boundary being reported
incorrectly. This includes perf, kfence, show_regs(), panic(), etc.
To fix this, we need to be able to uniquely identify each rewritten
return address such that we can map this back to the original return
address. We can use HAVE_FUNCTION_GRAPH_RET_ADDR_PTR to associate
each rewritten return address with a unique location on the stack. As
the return address is passed in the LR (and so is not guaranteed a
unique location in memory), we use the FP upon entry to the function
(i.e. the address of the caller's frame record) as the return address
pointer. Any nested call will have a different FP value as the caller
must create its own frame record and update FP to point to this.
Since ftrace_graph_ret_addr() requires the return address with the PAC
stripped, the stripping of the PAC is moved before the fixup of the
rewritten address. As we would unconditionally strip the PAC, moving
this earlier is not harmful, and we can avoid a redundant strip in the
return address fixup code.
I've tested this with the perf case above, the ftrace selftests, and
a number of ad-hoc unwinder tests. The tests all pass, and I have seen
no unexpected behaviour as a result of this change. I've tested with
pointer authentication under QEMU TCG where magic-sysrq+l correctly
recovers the original return addresses.
Note that this doesn't fix the issue of skipping a live LR at an
exception boundary, which is a more general problem and requires more
substantial rework. Were we to consume the LR in all cases this would
result in warnings where the interrupted context's LR contains
`return_to_handler`, but the FP has been altered, e.g.
| func:
| <--- ftrace entry ---> // logs FP & LR, rewrites LR
| STP FP, LR, [SP, #-16]!
| MOV FP, SP
| <--- INTERRUPT --->
... as ftrace_graph_get_ret_stack() fill not find a matching entry,
triggering the WARN_ON_ONCE() in unwind_frame().
Link: https://lore.kernel.org/r/20211025164925.GB2001@C02TD0UTHF1T.local
Link: https://lore.kernel.org/r/20211027132529.30027-1-mark.rutland@arm.com
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Madhavan T. Venkataraman <madvenka@linux.microsoft.com>
Cc: Mark Brown <broonie@kernel.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Will Deacon <will@kernel.org>
Reviewed-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20211029162245.39761-1-mark.rutland@arm.com
Signed-off-by: Will Deacon <will@kernel.org>
c6d3cd32fd