Re: [PATCH v2 3/9] LoongArch/ftrace: Add dynamic function tracer support

[Qing Zhang]

On 2022/9/1 上午10:59, Jeff Xie wrote:
> On Fri, Aug 26, 2022 at 4:24 PM Qing Zhang <zhangqing@xxxxxxxxxxx> wrote:
> > The compiler has inserted 2 NOPs before the regular function prologue.
> > T series registers are available and safe because of LoongArch psABI.
> >
> > At runtime, replace nop with bl to enable ftrace call and replace bl with
> > nop to disable ftrace call. The bl requires us to save the original RA value,
> > so here it saves RA at t0.
> > details are:
> >
> > | Compiled   |       Disabled         |        Enabled         |
> > +------------+------------------------+------------------------+
> > | nop        | move     t0, ra        | move     t0, ra        |
> > | nop        | nop                    | bl      ftrace_caller  |
> > | func_body  | func_body              | func_body              |
> >
> > The RA value will be recovered by ftrace_regs_entry, and restored into RA
> > before returning to the regular function prologue. When a function is not
> > being traced, the move t0, ra is not harmful.
> >
> > 1) ftrace_make_call, ftrace_make_nop (in kernel/ftrace.c)
> >     The two functions turn each recorded call site of filtered functions
> >     into a call to ftrace_caller or nops.
> >
> > 2) ftracce_update_ftrace_func (in kernel/ftrace.c)
> >     turns the nops at ftrace_call into a call to a generic entry for
> >     function tracers.
> >
> > 3) ftrace_caller (in kernel/mcount-dyn.S)
> >     The entry where each _mcount call sites calls to once they are
> >     filtered to be traced.
[...]
> >
>
> When use the option func_stack_trace for the function tracer, I found a issue:
>
> Steps:
> 1. Enable the function tracer and the option func_stack_trace:
>
> /sys/kernel/tracing # echo blk_update_request > ./set_ftrace_filter
> /sys/kernel/tracing # echo 1 > ./options/func_stack_trace
> /sys/kernel/tracing # echo function > ./current_tracer
>
> 2. Let the blk_update_request() be called.
>
>   # mount /dev/vda /tmp
>
>
> 3. cat ./trace
>            <idle>-0       [000] ..s1.   126.016445: blk_update_request
> <-blk_mq_end_request
>            <idle>-0       [000] ..s1.   126.017937: <stack trace>
>   => blk_mq_end_request
>
> We can see only one stack trace.
>
>
> I found the default unwinder(for loongson3_defconfig) is
> CONFIG_UNWINDER_PROLOGUE, if switch it to CONFIG_UNWINDER_GUESS
> it works well:
[...]
> Maybe the issue happened on the CONFIG_UNWINDER_PROLOGUE,  but I
> haven't dug deep into it ;)

Hi, Jeff

Thanks a lot for your feedback!

I fixed it in v3, which was caused by the ftrace_regs_entry assembly not 
being considered by the prologue analyze method. :)

regards
-Qing

>