On 2024/3/6 1:58, Josh Poimboeuf wrote:
> Adding ARM folks -- see
> https://lkml.kernel.org/lkml/[email protected]
> for the original bug report.
>
> This is an off-by-one bug which is common in unwinders, due to the fact
> that the address on the stack points to the return address rather than
> the call address.
>
> So, for example, when the last instruction of a function is a function
> call (e.g., to a noreturn function), it can cause the unwinder to
> incorrectly try to unwind from the function *after* the callee.
>
> For ORC (x86), we fixed this by decrementing the PC for call frames (but
> not exception frames). I've seen user space unwinders do similar, for
> non-signal frames.
>
> Something like the following might fix your issue (completely untested):
>
Thank you very much. I have verified that your patch can fix my issue.
But I have some little questions.
> diff --git a/arch/arm/include/asm/stacktrace.h
> b/arch/arm/include/asm/stacktrace.h
> index 360f0d2406bf..4891e38cdc1f 100644
> --- a/arch/arm/include/asm/stacktrace.h
> +++ b/arch/arm/include/asm/stacktrace.h
> @@ -21,9 +21,7 @@ struct stackframe {
> struct llist_node *kr_cur;
> struct task_struct *tsk;
> #endif
> -#ifdef CONFIG_UNWINDER_FRAME_POINTER
> bool ex_frame;
> -#endif
> };
>
> static __always_inline
> @@ -37,9 +35,8 @@ void arm_get_current_stackframe(struct pt_regs *regs,
> struct stackframe *frame)
> frame->kr_cur = NULL;
> frame->tsk = current;
> #endif
> -#ifdef CONFIG_UNWINDER_FRAME_POINTER
> - frame->ex_frame = in_entry_text(frame->pc);
> -#endif
> + frame->ex_frame = !!regs;
> +
'regs' must not be NULL, frame->ex_frame will always be TRUE.
So I think we just need to remove CONFIG_UNWINDER_FRAME_POINTER here.
We don't need to change the frame->ex_frame assignment statement.
> diff --git a/arch/arm/kernel/unwind.c b/arch/arm/kernel/unwind.c
> index 9d2192156087..99ded32196af 100644
> --- a/arch/arm/kernel/unwind.c
> +++ b/arch/arm/kernel/unwind.c
> @@ -407,7 +407,7 @@ int unwind_frame(struct stackframe *frame)
> {
> const struct unwind_idx *idx;
> struct unwind_ctrl_block ctrl;
> - unsigned long sp_low;
> + unsigned long sp_low, pc;
>
> /* store the highest address on the stack to avoid crossing it*/
> sp_low = frame->sp;
> @@ -417,19 +417,22 @@ int unwind_frame(struct stackframe *frame)
> pr_debug("%s(pc = %08lx lr = %08lx sp = %08lx)\n", __func__,
> frame->pc, frame->lr, frame->sp);
>
> - idx = unwind_find_idx(frame->pc);
> + pc = frame->ex_frame ? frame->pc : frame->pc - 4;
For details, see the unwind_next_frame function in the unwind_orc.c.
Why subtract 4 here instead of 1?
`pc = frame->ex_frame ? frame->pc : frame->pc - 1`
Is it more appropriate?
> +
> + idx = unwind_find_idx(pc);
> if (!idx) {
> - if (frame->pc && kernel_text_address(frame->pc)) {
> - if (in_module_plt(frame->pc) && frame->pc != frame->lr)
> {
> + if (kernel_text_address(pc)) {
> + if (in_module_plt(pc) && frame->pc != frame->lr) {
> /*
> * Quoting Ard: Veneers only set PC using a
> * PC+immediate LDR, and so they don't affect
> * the state of the stack or the register file
> */
> frame->pc = frame->lr;
> + frame->ex_frame = false;
> return URC_OK;
> }
> - pr_warn("unwind: Index not found %08lx\n", frame->pc);
> + pr_warn("unwind: Index not found %08lx\n", pc);
> }
> return -URC_FAILURE;
> }
> @@ -442,7 +445,7 @@ int unwind_frame(struct stackframe *frame)
> if (idx->insn == 1)
> /* can't unwind */
> return -URC_FAILURE;
> - else if (frame->pc == prel31_to_addr(&idx->addr_offset)) {
> + else if (frame->ex_frame && pc == prel31_to_addr(&idx->addr_offset)) {
> /*
> * Unwinding is tricky when we're halfway through the prologue,
> * since the stack frame that the unwinder expects may not be
> @@ -451,9 +454,10 @@ int unwind_frame(struct stackframe *frame)
> * a function, we are still effectively in the stack frame of
> * the caller, and the unwind info has no relevance yet.
> */
> - if (frame->pc == frame->lr)
> + if (pc == frame->lr)
> return -URC_FAILURE;
> frame->pc = frame->lr;
> + frame->ex_frame = false;
> return URC_OK;
> } else if ((idx->insn & 0x80000000) == 0)
> /* prel31 to the unwind table */
> @@ -515,6 +519,7 @@ int unwind_frame(struct stackframe *frame)
> frame->lr = ctrl.vrs[LR];
> frame->pc = ctrl.vrs[PC];
> frame->lr_addr = ctrl.lr_addr;
> + frame->ex_frame = false;
Why is the value of `frame->ex_frame` directly set to false?
Why is the value not determined based on `frame->pc`?
That is, `frame->ex_frame = in_entry_text(frame->pc)`
>
> return URC_OK;
> }
> @@ -544,6 +549,7 @@ void unwind_backtrace(struct pt_regs *regs, struct
> task_struct *tsk,
> */
> here:
> frame.pc = (unsigned long)&&here;
> + frame.ex_frame = false;
> } else {
> /* task blocked in __switch_to */
> frame.fp = thread_saved_fp(tsk);
> @@ -554,11 +560,12 @@ void unwind_backtrace(struct pt_regs *regs, struct
> task_struct *tsk,
> */
> frame.lr = 0;
> frame.pc = thread_saved_pc(tsk);
> + frame.ex_frame = false;
> }
>
> while (1) {
> int urc;
> - unsigned long where = frame.pc;
> + unsigned long where = frame.ex_frame ? frame.pc : frame.pc - 4;
>
> urc = unwind_frame(&frame);
> if (urc < 0)
> .
>
If I refer to your demo patch and submit a new bugfix patch,
can I mark you as "Co-developed-by" in this new bugfix patch?
