Re: [PATCH 1/2] objtool/LoongArch: Restrict stack operation instruction

[Jinyang He]

On 2024-07-30 17:49, Tiezhu Yang wrote:
> On 07/30/2024 05:28 PM, Jinyang He wrote:
> > On 2024-07-30 14:19, Tiezhu Yang wrote:
> >
> > > After commit a0f7085f6a63 ("LoongArch: Add RANDOMIZE_KSTACK_OFFSET
> > > support"), the code flow of do_syscall() was changed when compiled
> > > with GCC due to the secondary stack of add_random_kstack_offset(),
> > > something like this:
> > >
> > >    addi.d          $sp, $sp, -32
> > >    st.d            $fp, $sp, 16
> > >    st.d            $ra, $sp, 24
> > >    addi.d          $fp, $sp, 32
> > >    ...
> > >    sub.d           $sp, $sp, $t1
> > Have you checked the ORC info whether is right or tried backtrace which
> > passed do_syscall? The "sub.d $sp, $sp, $t1" has modified the $sp so the
> > $sp cannot express CFA here. This patch just clear the warning but 
> > ignore
> > the validity of ORC info. The wrong ORC info may cause illegally access
> > memory when backtrace.
>
> I did testing many times before submitting, the call trace is
> expected when testing "echo l > /proc/sysrq-trigger".
Make sure the RANDOMIZE_KSTACK_OFFSET is enable. I tested it by
CONFIG_RANDOMIZE_KSTACK_OFFSET_DEFAULT=y on qemu and it did
not show the frame about handle_syscall.

On the other hand,
$ ./tools/objtool/objtool --dump arch/loongarch/kernel/syscall.o
.noinstr.text+0:type:call sp:sp +   0 fp: (und)  ra: (und)  signal:0
.noinstr.text+4:type:call sp:sp +  32 fp: (und)  ra: (und)  signal:0
.noinstr.text+8:type:call sp:sp +  32 fp:prevsp + -16 ra: (und) signal:0
.noinstr.text+c:type:call sp:sp +  32 fp:prevsp + -16 ra:prevsp + -8 
signal:0
.noinstr.text+e4:type:call sp:sp +  32 fp:prevsp + -16 ra: (und) signal:0
.noinstr.text+e8:type:call sp:sp +  32 fp: (und)  ra: (und) signal:0
.noinstr.text+ec:type:call sp:sp +   0 fp: (und)  ra: (und) signal:0
.noinstr.text+f0:type:? sp: (und)  fp: (und)  ra: (und)  signal:0
.noinstr.text+100:type:call sp:sp +  32 fp:prevsp + -16 ra:prevsp + -8 
signal:0
.noinstr.text+118:type:? sp: (und)  fp: (und)  ra: (und)  signal:0
$ objdump -d arch/loongarch/kernel/syscall.o | grep "sub.*sp.*sp"
  70:    0011b463     sub.d           $sp, $sp, $t1
Obviously the ORC info is wrong.
(Assume the backtrace PC is 0x74, how to find the previous frame?)