> On 5 Aug 2026, at 1:59 PM, MUKESH KUMAR PILANIYA <mpilaniy(a)redhat.com&gt; wrote: > > > >> On 4 Aug 2026, at 12:59 PM, HAGIO KAZUHITO(萩尾　一仁) <k-hagio-ab at nec.com> wrote: >> >> On 2026/08/04 15:57, Tao Liu wrote: >>> Previously ORC_REG_SP and ORC_REG_PREV_SP will depend on kernel version >>> for their value, however this is fragile since distributions will >>> backport the upstream patch to a lower kernel version, thus break the >>> version assumption. >>> >>> This patch fixes by checking the value of ORC_REG_PREV_SP, which can be >>> get from orc_fp_entry. ORC_REG_PREV_SP's value can work as the indicator >>> of whether the current kernel have applied the upstream patch 1735858caa4b >>> ("objtool/x86: Reorder ORC register numbering"). >>> >>> Fixes: d0ee428664f9 ("x86_64: Fix "bt" command to use correct ORC register >>> values on Linux 7.1 and later") >>> >>> Signed-off-by: Tao Liu <ltao(a)redhat.com> >> >> thank you for the fix, looks good to me. >> >> Thanks, >> Kazu >> >>> --- >>> v2 -> v1: emit error(WARN) for kernels which doesn't have orc_fp_entry >>> symbol. >>> --- >>> x86_64.c | 30 +++++++++++++++++++++++------- >>> 1 file changed, 23 insertions(+), 7 deletions(-) >>> >>> diff --git a/x86_64.c b/x86_64.c >>> index 55648697baf3..81d20ed419d5 100644 >>> --- a/x86_64.c >>> +++ b/x86_64.c >>> @@ -6661,6 +6661,9 @@ x86_64_ORC_init(void) >>> NULL >>> }; >>> struct ORC_data *orc; >>> + struct gnu_request request, *req; >>> + req = &request; >>> + ulong value = 0; > > declaration after statement so that compiler does not give warning as per C90 standards. > ulong value = 0; > req = &request; > > This will also works. > struct gnu_request request, *req = &request; > ulong value = 0;

> >>> >>> MEMBER_OFFSET_INIT(inactive_task_frame_bp, "inactive_task_frame", "bp"); >>> MEMBER_OFFSET_INIT(inactive_task_frame_ret_addr, "inactive_task_frame", "ret_addr"); >>> @@ -6736,13 +6739,26 @@ x86_64_ORC_init(void) >>> if (orc->has_signal && !orc->has_end) >>> machdep->flags |= ORC_6_4; >>> >>> - /* See kernel commit 1735858caa4b */ >>> - if (THIS_KERNEL_VERSION >= LINUX(7,1,0)) { >>> - ORC_REG_SP = 3; >>> - ORC_REG_PREV_SP = 8; >>> - } else { >>> - ORC_REG_SP = 5; >>> - ORC_REG_PREV_SP = 1; >>> + /* Try get ORC_REG_(PREV)_SP */ >>> + ORC_REG_SP = 5; >>> + ORC_REG_PREV_SP = 1; >>> + >>> + if (kernel_symbol_exists("orc_fp_entry")) { > > If orc_fp_entry doesn't exist in the symbol table (possible on stripped vmlinux without kallsyms, or if a future kernel removes it), the patch silently falls back to old values. > > Also the kernel declares orc_fp_entry as static in arch/x86/kernel/unwind_orc.c. It's generally available in: > - vmlinux with debuginfo (.symtab with STB_LOCAL) > - /proc/kallsyms (with CONFIG_KALLSYMS_ALL=y, the default) > > But it won't be available in fully stripped vmlinux without kallsyms. So it better to implement a fallback logic in else block. > > else if (THIS_KERNEL_VERSION >= LINUX(7,1,0)) { > ORC_REG_SP = 3; > ORC_REG_PREV_SP = 8; > }

> >>> + if (get_symbol_type("orc_fp_entry", "bp_reg", req) == TYPE_CODE_INT) { > > Crash utility already defines kernel_orc_entry in defs.h with the identical packed bitfield layout so why not to use directly that will reduce get_symbol_type + manual bitfield extraction work ? > The bp_reg bitfield is at the same offset in both kernel_orc_entry and kernel_orc_entry_6_4, so either works.

> >>> + get_symbol_data("orc_fp_entry", sizeof(ulong), &value); > > get_symbol_data() calls readmem() with FAULT_ON_ERROR, which aborts crash if the memory page isn't present in the vmcore. In kdump scenarios, not all pages are necessarily captured. The patch should use try_get_symbol_data() instead, which returns FALSE on failure without aborting.

> >>> + /* >>> + * orc_fp_entry.bp_reg = ORC_REG_PREV_SP >>> + * See kernel commit 1735858caa4b. Use ORC_REG_PREV_SP as the >>> + * indicator of the commit. >>> + */ >>> + value >>= req->member_offset; >>> + value &= ((1UL << req->member_length) - 1); >>> + if (value == 8) { >>> + ORC_REG_SP = 3; >>> + ORC_REG_PREV_SP = 8; >>> + } >>> + } else >>> + error(WARNING, "Cannot get orc_fp_entry.bp_reg info"); > > Should not be this is a more user friendly ? > error(WARNING, "Cannot determine ORC register numbering from orc_fp_entry; “ "using legacy values (ORC_REG_SP=5, ORC_REG_PREV_SP=1)\n");

> >>> } >>> >>> machdep->flags |= ORC; > > > Putting it all together. > /* Detect ORC register numbering from orc_fp_entry.bp_reg. > * See kernel commit 1735858caa4b. Note: orc_fp_entry is static, > * so fall back to the version check if the symbol is unavailable. > */ > ORC_REG_SP = 5; > ORC_REG_PREV_SP = 1; > > if (kernel_symbol_exists("orc_fp_entry")) { > kernel_orc_entry fp_entry; > if (try_get_symbol_data("orc_fp_entry", > sizeof(kernel_orc_entry), &fp_entry) && > fp_entry.bp_reg == 8) { > ORC_REG_SP = 3; > ORC_REG_PREV_SP = 8; > } > } else if (THIS_KERNEL_VERSION >= LINUX(7,1,0)) { > ORC_REG_SP = 3; > ORC_REG_PREV_SP = 8; > } > > Thanks > Mukesh Pilaniya