why trace.py can not give full userstack with kernel stack

[yangguosdxl]

./trace-bpfcc -UK __alloc_pages -L $(pidof myhi)
I trace my demo program, it output like this:

PID     TID     COMM            FUNC             
132657  132657  myhi            __alloc_pages    
        b'__alloc_pages+0x1 [kernel]'
        b'vma_alloc_folio+0x9c [kernel]'
        b'do_anonymous_page+0xf5 [kernel]'
        b'handle_pte_fault+0x250 [kernel]'
        b'__handle_mm_fault+0x41f [kernel]'
        b'handle_mm_fault+0xba [kernel]'
        b'do_user_addr_fault+0x1be [kernel]'
        b'exc_page_fault+0x81 [kernel]'
        b'asm_exc_page_fault+0x27 [kernel]'
        b'_int_malloc+0xf00 [libc.so.6]'

the user stack not show all stack. the full may like below:

malloc [libc.so.6]
hello [libhi.so]
main [myhi]

anyone know why there is not have full user stack?

thanks

demo.zip

[chenhengqi]

Kernel relies on stack pointer to unwind the stack, and likely, your libc is compiled with omit frame pointer, so the stack is imcomplete.

There is an attempt to using DWARF to unwind the stack, see https://www.polarsignals.com/blog/posts/2022/11/29/dwarf-based-stack-walking-using-ebpf/.

[yangguosdxl]

Kernel relies on stack pointer to unwind the stack, and likely, your libc is compiled with omit frame pointer, so the stack is imcomplete.

There is an attempt to using DWARF to unwind the stack, see https://www.polarsignals.com/blog/posts/2022/11/29/dwarf-based-stack-walking-using-ebpf/.

how do you know my libc is compiled with omit frame pointer?

[chenhengqi]

I don't know, I just guess, most Linux distributions do that.

https://fedoraproject.org/wiki/Changes/fno-omit-frame-pointer

[Bojun-Seo]

These PR may help you to unwind(or guess) stack without recompiling libc.so with -fno-omit-frame-pointer.
#4463 #4679