06:35 <ayaka__> I met a problem with debugger linux kernel module, but the gdb can't never stop where I want, it always bother me with arm psci error https://paste.centos.org/view/d47269d8

06:35 <ayaka__> the gdb info breakpoints always be pending, even I have add the modules' text addresses

06:59 <PaulFertser> ayaka__: you might want to try "hbreak" to force hardware breakpoints but I have no idea about psci :( also there's the usual advice of disabling idle etc when debugging Linux over JTAG.

07:00 <ayaka__> PaulFertser, psci is a annoying arm firmware running a separate cpu execute level(not kernel, not userspace, but a monitor level)

07:01 <ayaka__> I tried to lx-symbols, which would load kernel modules auto, then I can't resume kernel to run

07:01 <ayaka__> https://paste.centos.org/view/10e2b8ed

07:02 <PaulFertser> This also suggests https://linaro.atlassian.net/wiki/spaces/TCWGPUB/pages/25296346120/Raspberry+Pi+Linux+kernel+debugging+with+OpenOCD nokaslr rodata=off but there're likely more tricks to it.

07:05 <ayaka__> Let me try to disable kaslr and rodata from the building(bootargs is fixed in the bootloader)

07:06 <ayaka__> I just can't figure out even I have not declare a breakpoint, where the failure comes from https://paste.centos.org/view/ae04deca

07:14 <PaulFertser> ayaka__: you can try using "hbreak" instead of "break" to avoid modifying memory.

07:15 <borneoa___> ayaka__: the abort message after GDB connect looks like the debugger failed to connect or halt the target. Is there any other message at OpenOCD starting?

07:17 <borneoa___> ayaka__: there is also a command "monitor gdb_breakpoint_override hard" to force OpenOCD to only use HW breakpoints

07:23 <ayaka__> borneoa___, https://paste.centos.org/view/10e2b8ed

07:24 <ayaka__> PaulFertser, yes, I would try to when I figured out how to disable kaslr

07:25 <borneoa___> ayaka__: can you post the OpenOCD log from OpenOCD start till GDB attach?

07:26 <ayaka__> borneoa___, should I let the gdb loading kernel modules symbols?

07:27 <borneoa___> ayaka__: then you can also increase the log level and dump GDB communication. Add -d to OpenOCD command line ans run "set debug remote 1" in GDB before connecting to OpenOCD

07:27 <ayaka__> of course, I would follow you

07:29 <ayaka__> RANDOMIZE_BASE is n in my kernel config, I think I didn't enable kaslr

07:31 <borneoa___> ayaka__: yes! Load kernel symbols before connecting to OpenOCD otherwise GDB issues crap random read to the target to guess the stack structure and you get endless abort msgs due to memory protection, plus the secure firmware should detect the illegal memory access and hang or reboot the platform

07:31 <ayaka__> paste.centos.org/view/12ec5183

07:36 <ayaka__> it is breakpoint_add_internal() make this problem, for gdb it is Cannot access memory at address 0xffffffc0080da6a0 . Maybe I should figure out a method to disable rodata

07:36 <borneoa___> ayaka__: I see you have set -rtos linux. Have you patched OpenOCD and recompiled it to align it with your kernel?

07:36 <ayaka__> borneoa___, nope, I didn't patch the openocd, I didn't know I should do that

07:37 <borneoa___> ayaka__: there is a compile flag to disable rodata, search it around the command line flag rodata=off

07:38 <borneoa___> ayaka__: I'm going offline for ~30 minutes...

07:39 <borneoa___> ayaka__: eventually drop rtos linux for the moment and eventually use the GDB script in the kernel

07:42 <ayaka__> you mean using gdb with qemu? unfortunately I want to debug it in a real hardware

07:49 <ayaka__> unfortunately, ARCH_HAS_STRICT_KERNEL_RWX in arm64, I think I have to figure out a method to insert the bootargs

08:20 <ayaka__> After I turn off the rodata and decrease the adapter speed, I could finally resume the cpu

08:21 <ayaka__> but still the kernel breakpoint won't work. Besides I must resume kernel quickly after the gdb halt the kernel or kernel would kill that kernel

08:30 <borneoa___> ayaka__: the flag to disable is CONFIG_STRICT_KERNEL_RWX

08:34 <ayaka__> borneoa___, yes, but that flag is not visual when ARCH_HAS_STRICT_KERNEL_RWX so I disable the rodata in dts which applied to the command line

08:34 <ayaka__> but I still can't make the breakpoint work for the kernel

08:35 <borneoa___> ayaka__: no, I'm not suggesting qemu. You have mentioned the command lx-symbols that is a GDB extension provided by a python script distributed in kernel code; it's the GDB awareness script. Linaro has developed it to somehow replace OpenOCD -rtos linux because it's not available in qemu and even in OpenOCD is hard to use it. So, I suggest to use it in place of OpenOCD rtos

08:36 <ayaka__> borneoa___, if I use the vmlinux-gdb.py, I should use hwthread in openocd instead?

08:39 <borneoa___> ayaka__: yes

08:39 <borneoa___> ayaka__: you are right, arm64 doesn't allow that flag.

08:39 <ayaka__> let me try whether it would help

08:42 <borneoa___> ayaka__: to simplify the debug, keep only one core enabled and halt the others. In this way you can skip hwthread and SMP. I mean stop the core from executing the kernel and become like single CPU

08:43 <ayaka__> borneoa___, for the trial stage, that maybe enough. But the problem I need to debug is a about multiple threads

08:47 <borneoa___> ayaka__: just temporarily. Either recompile with no SMP or stop the CPU with echo 0 > /sys/bus/cpu/devices/cpu1/online

08:51 <ayaka__> I tried, the gdb has no problem on halting or resume the cpu now. But the breakpoint still won't work

08:52 <ayaka__> it is quite strange, I tried the watchpoint API from the linux kernel last time, it would work

08:53 <ayaka__> I just can't insert a hardware point? Or maybe the TEE OS(open-tee) here clear the hardware breakpoint ?

09:04 <ayaka__> borneoa___, I think that is the problem, once I do "set breakpoint auto-hw off"

09:04 <ayaka__> it finally work

10:17 <ayaka__> I am not sure about the debug registers that DAP access there, would it apply to all the execute level or just a level ask arm suggested https://developer.arm.com/documentation/102120/0101/AArch64-debug ?

10:51 <borneoa___> ayaka__: that page refers to self-hosted debug, that is what you get with kgdb. OpenOCD works as an external debugger. In the same page in the first paragraph there is the link to external debugger

10:52 <ayaka__> yes, I found that. it looks like the externel debugger doesn't care about the execute level

10:59 <ayaka__> Learning from tcl/target/stm32mp13x.cfg, I have known how to enable the access to AP after the an ap has been finished its init. But I can't create more targets after that. "Error: The 'target create' command must be used before 'init'."

11:08 <borneoa___> ayaka__: yes targets must be created all at the start, but you can flag them as deferred examine. Once you know they are available, you can examine them. Search arp_examine in the scripts, there are few methods to use it