Christophe Leroy wrote:
Le 19/06/2023 à 11:47, Naveen N Rao a écrit :
With ppc64 -mprofile-kernel and ppc32 -pg, profiling instructions to
call into ftrace are emitted right at function entry. The instruction
sequence used is minimal to reduce overhead. Crucially, a stackframe is
not created for the function being traced. This breaks stack unwinding
since the function being traced does not have a stackframe for itself.
As such, it never shows up in the backtrace:
/sys/kernel/debug/tracing # echo 1 > /proc/sys/kernel/stack_tracer_enabled
/sys/kernel/debug/tracing # cat stack_trace
Depth Size Location (17 entries)
----- ---- --------
0) 4144 32 ftrace_call+0x4/0x44
1) 4112 432 get_page_from_freelist+0x26c/0x1ad0
2) 3680 496 __alloc_pages+0x290/0x1280
3) 3184 336 __folio_alloc+0x34/0x90
4) 2848 176 vma_alloc_folio+0xd8/0x540
5) 2672 272 __handle_mm_fault+0x700/0x1cc0
6) 2400 208 handle_mm_fault+0xf0/0x3f0
7) 2192 80 ___do_page_fault+0x3e4/0xbe0
8) 2112 160 do_page_fault+0x30/0xc0
9) 1952 256 data_access_common_virt+0x210/0x220
10) 1696 400 0xc00000000f16b100
11) 1296 384 load_elf_binary+0x804/0x1b80
12) 912 208 bprm_execve+0x2d8/0x7e0
13) 704 64 do_execveat_common+0x1d0/0x2f0
14) 640 160 sys_execve+0x54/0x70
15) 480 64 system_call_exception+0x138/0x350
16) 416 416 system_call_common+0x160/0x2c4
Fix this by having ftrace create a dummy stackframe for the function
being traced. Since this is only relevant when ftrace is active, we nop
out the instruction to store LR in the LR save area in the profiling
instruction sequence on ppc32 (and in ppc64 with older gcc versions).
Instead, in ftrace, we store LR in the LR save area of the previous
stackframe, and create a minimal stackframe to represent the function
being traced. With this, backtraces now capture the function being
traced:
/sys/kernel/debug/tracing # cat stack_trace
Depth Size Location (17 entries)
----- ---- --------
0) 3888 32 _raw_spin_trylock+0x8/0x70
1) 3856 576 get_page_from_freelist+0x26c/0x1ad0
2) 3280 64 __alloc_pages+0x290/0x1280
3) 3216 336 __folio_alloc+0x34/0x90
4) 2880 176 vma_alloc_folio+0xd8/0x540
5) 2704 416 __handle_mm_fault+0x700/0x1cc0
6) 2288 96 handle_mm_fault+0xf0/0x3f0
7) 2192 48 ___do_page_fault+0x3e4/0xbe0
8) 2144 192 do_page_fault+0x30/0xc0
9) 1952 608 data_access_common_virt+0x210/0x220
10) 1344 16 0xc0000000334bbb50
11) 1328 416 load_elf_binary+0x804/0x1b80
12) 912 64 bprm_execve+0x2d8/0x7e0
13) 848 176 do_execveat_common+0x1d0/0x2f0
14) 672 192 sys_execve+0x54/0x70
15) 480 64 system_call_exception+0x138/0x350
16) 416 416 system_call_common+0x160/0x2c4
This results in two additional stores in the ftrace entry code, but
produces reliable backtraces. Note that this change now aligns with
other architectures (arm64, s390, x86).
Signed-off-by: Naveen N Rao <nav...@kernel.org>
Reviewed-by: Christophe Leroy <christophe.le...@csgroup.eu>
Thanks. This patch can be dropped from this series. I posted a minimal
fix separately to aid backport:
http://lore.kernel.org/20230621051349.759567-1-nav...@kernel.org
I will post a separate patch to also nop out the duplicate stores in the
function profiling sequence.
- Naveen