On Mon, Oct 25, 2021 at 05:49:25PM +0100, Mark Rutland wrote: > From f3e66ca75aff3474355839f72d123276028204e1 Mon Sep 17 00:00:00 2001 > From: Mark Rutland <mark.rutland@xxxxxxx> > Date: Mon, 25 Oct 2021 13:23:11 +0100 > Subject: [PATCH] arm64: ftrace: use HAVE_FUNCTION_GRAPH_RET_ADDR_PTR > > When CONFIG_FUNCTION_GRAPH_TRACER is selected, and the function graph: > tracer is in use, unwind_frame() may erroneously asscociate 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. > > 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 `return_to_handler`, which indicates a rewritten > return address. This is broken in two cases: > > * Between creating a pt_regs and starting the unwind, function calls may > place entries on the stack, leaving an abitrary offset which we can > only determine by performing a full unwind from the caller of the > unwind code. While this initial unwind is open-coded in > dump_backtrace(), this is not performed for other unwinders such as > perf_callchain_kernel(). > > * When unwinding across an exception boundary (whether continuing an > unwind or starting a new unwind from regs), we always consume the LR > of the interrupted context, though this may not have been live at the > time of the exception. Where the LR was not live but happened to > contain `return_to_handler`, we'll recover an address from the graph > return stack and increment the current offset, leaving subsequent > entries off-by-one. > > Where the LR was not live and did not contain `return_to_handler`, we > will still report an erroneous address, but subsequent entries will be > unaffected. It turns out I had this backwards, and we currently always *skip* the LR when unwinding across regs, because: * The entry assembly creates a synthetic frame record with the original FP and the ELR_EL1 value (i.e. the PC at the point of the exception), skipping the LR. * In arch_stack_walk() we start the walk from regs->pc, and continue with the frame record, skipping the LR. * In the existing dump_backtrace, we skip until we hit a frame record whose FP value matches the FP in the regs (i.e. the synthetic frame record created by the entry assembly). That'll dump the ELR_EL1 value, then continue to the next frame record, skipping the LR. So case two is bogus, and only case one can happen today. This cleanup shouldn't trigger the WARN_ON_ONCE() in unwind_frame(), and we can fix the missing LR entry in a subsequent cleanup. Thanks, Mark.
