| Message ID | 20240710-arm64-backtrace-v1-0-5a2ba50485dd@linaro.org |
|---|---|
| Headers | show |
| Series | ARM64: add symbol name lookup and print a backtrace on exception | expand |
On Wed, Jul 10, 2024 at 06:26:17PM +0200, Caleb Connolly wrote: > U-Boot already emits frame pointers on ARM64, but lacks the code to > parse them, as well as a mechanism for looking up symbol names at > runtime. > > There was some (seemingly?) leftover code for symbols lookups in > common/kallsyms.c and associated parts in the makefile, however it > appears to be entirely unused and unsupported. It relied on generating > one long string of all symbol addresses and names. > > The approach taken here is instead largely based on the implementation > in the Xen hypervisor, it performs basic compression using non-ASCII > bytes to tokenize repeated string segments which can later be expanded > back out at runtime. > > This is then utilized in the ARM64 interrupt handling routine to dump a > backtrace in the show_regs() debug function. As well as providing a > general purpose unwind_stack() function which can be used for debugging. Just a few minor comments from me in general, and I'm glad to see this. Please submit the next version non-RFC, thanks.
On 12/07/2024 18:48, Tom Rini wrote: > On Wed, Jul 10, 2024 at 06:26:17PM +0200, Caleb Connolly wrote: > >> U-Boot already emits frame pointers on ARM64, but lacks the code to >> parse them, as well as a mechanism for looking up symbol names at >> runtime. >> >> There was some (seemingly?) leftover code for symbols lookups in >> common/kallsyms.c and associated parts in the makefile, however it >> appears to be entirely unused and unsupported. It relied on generating >> one long string of all symbol addresses and names. >> >> The approach taken here is instead largely based on the implementation >> in the Xen hypervisor, it performs basic compression using non-ASCII >> bytes to tokenize repeated string segments which can later be expanded >> back out at runtime. >> >> This is then utilized in the ARM64 interrupt handling routine to dump a >> backtrace in the show_regs() debug function. As well as providing a >> general purpose unwind_stack() function which can be used for debugging. > > Just a few minor comments from me in general, and I'm glad to see this. > Please submit the next version non-RFC, thanks. Thanks for the review. One thing I'm a little unsure about before sending the next version: is it better to print the address in the link register or the address of the function in the backtrace? In the example below I print the function address offset by _start, but that doesn't seem very useful for either interactive debugging or decompilation. Maybe the absolute address in the link register would be more useful. I'm leaning towards the latter, but open to ideas. Backtrace: <0x0000000001bb4c> do_mem_mw+0xbc <0x000000bb729cec> cmd_process+0x130 <0x000000bb7213f8> run_list_real+0x6d0 <0x000000bb721b08> parse_stream_outer+0x14c <0x000000bb722164> parse_file_outer+0x34 <0x000000bb72929c> cli_loop+0x1c <0x000000bb71f1dc> main_loop+0x54 <0x000000bb7229dc> run_main_loop+0x14 <0x000000bb777664> initcall_run_list+0x7c <0x000000bb722bf8> board_init_r+0x34 >
U-Boot already emits frame pointers on ARM64, but lacks the code to parse them, as well as a mechanism for looking up symbol names at runtime. There was some (seemingly?) leftover code for symbols lookups in common/kallsyms.c and associated parts in the makefile, however it appears to be entirely unused and unsupported. It relied on generating one long string of all symbol addresses and names. The approach taken here is instead largely based on the implementation in the Xen hypervisor, it performs basic compression using non-ASCII bytes to tokenize repeated string segments which can later be expanded back out at runtime. This is then utilized in the ARM64 interrupt handling routine to dump a backtrace in the show_regs() debug function. As well as providing a general purpose unwind_stack() function which can be used for debugging. == Relocation == Since U-Boot relocates itself at runtime, and can be built to be position independent in the first place (effectively "relocating" itself when it first starts too), we can't really rely on gd->reloc_off. The approach taken here is to subtract CONFIG_TEXT_BASE from the address of each symbol in the lookup table (while it's being generated), then when decoding we just subtract the address of the _start label. Since this label address is updated to make U-Boot position independent and during relocation, it allows us to avoid re-implementing the relocation state handling stuff in the symbol decoder. == Size == By default this feature is off, and will not effect the size of U-Boot binaries. The generated symbols object file is ~85k with the (fairly hefty) qcom_defconfig, so there is certainly a cost to be taken into account. I hope that this implementation can be later extended for other platforms. However this is currently beyond my (skill, time) capabilities. --- Caleb Connolly (3): drop unused kallsyms support add support for symbol lookups arm64: unwind stack on exception Makefile | 24 +- arch/arm/include/asm/ptrace.h | 4 + arch/arm/lib/interrupts_64.c | 76 +++++ common/Makefile | 1 - common/kallsyms.c | 43 --- common/system_map.c | 8 - include/symbols.h | 19 ++ lib/Kconfig | 8 + lib/Makefile | 1 + lib/symbols.c | 126 ++++++++ tools/Makefile | 3 + tools/symbols.c | 646 ++++++++++++++++++++++++++++++++++++++++++ 12 files changed, 896 insertions(+), 63 deletions(-) --- change-id: 20240710-arm64-backtrace-2926f764dbdc base-commit: 13f9c5668411aa18ef64846d5bc86e9e6be52082 // Caleb (they/them)