Message-ID: <CAEr6+EAEvZiU14H2bGW9iGOwJjshAw5sQk7uneFuyK23YMm-Sw@mail.gmail.com>
Date:   Tue, 18 Oct 2022 15:01:57 +0800
From:   Jeff Xie <xiehuan09@...il.com>
To:     Steven Rostedt <rostedt@...dmis.org>
Cc:     mingo@...hat.com, mhiramat@...nel.org, zanussi@...nel.org,
        linux-kernel@...r.kernel.org, chensong_2000@....cn
Subject: Re: [PATCH v15 0/4] trace: Introduce objtrace trigger to trace the
 kernel object

On Tue, Oct 18, 2022 at 3:40 AM Steven Rostedt <rostedt@...dmis.org> wrote:
>
> On Fri, 19 Aug 2022 11:27:02 +0800
> Jeff Xie <xiehuan09@...il.com> wrote:
>
> So on my way home from London, I tried to update it to use the fetch
> instruction logic of kprobes (which will work), but I also ended up pretty
> much rewriting most of it, which then I realized, I'm going much further
> than I wanted to.

Very excited  to hear it will work. Thank you for your hard work.

> But know I have a much better idea of how this should work, and I'll try to
> go through what I learned here and throughout the patches.
>
>
> > Introduce a method based on function tracer to trace any object and get
> > the value of the object dynamically. the object can be obtained from the
> > dynamic event (kprobe_event/uprobe_event) or the static event(tracepoint).
>
> Note, I can see the value of tracing only the value of the object, but that
> to me would be "objfollow" not "objtrace". That is, to find out what
> functions are calling the object. I guess that may have been your original
> objective, but it was articulated more about knowing how the value changed

Yes, It actually was my original objective -;)  and I also really love
your idea of getting its value.

> and not where the value is going (and the code was written more like that).
> Having a objfollow may be nice, but reading the arguments of a function is
> really a "best attempt" at most, and you can't really trust the arguments
> are what you are tracing.

It's true, you spoke to my heart ;-)

>I would hold off on that until we have a good BTF
> tracing infrastructure in the function tracer.
>
> So I'm going to continue with "objtrace" which is about tracing the
> object's value throughout time.

Thanks.

>
> >
> > Usage:
> > When using the kprobe event, only need to set the objtrace(a new trigger),
> > we can get the value of the object. The object is from the setting of the
> > kprobe event.
> >
> > For example:
> > For the function bio_add_page():
> >
> > int bio_add_page(struct bio *bio, struct page *page,
> >       unsigned int len, unsigned int offset)
> >
> > Firstly, we can set the base of the object, thus the first string "arg1"
> > stands for the value of the first parameter of this function bio_add_gage(),
> >
> > # echo 'p bio_add_page arg1=$arg1' > ./kprobe_events
> >
> > Secondly, we can get the value dynamically based on above object.
> >
> > find the offset of the bi_size in struct bio:
> > $ gdb vmlinux
> > (gdb) p &(((struct bio *)0)->bi_iter.bi_size)
> > $1 = (unsigned int *) 0x28
> >
> > # echo 'objtrace:add:arg1,0x28:u32:1 if comm == "cat"' > ./events/kprobes/ \
> >       p_bio_add_page_0/trigger
> >
> > # cd /sys/kernel/debug/tracing/
> > # echo 'p bio_add_page arg1=$arg1' > ./kprobe_events
> > # echo 'objtrace:add:arg1,0x28:u32:1 if comm == "cat"' > ./events/kprobes/p_bio_add_page_0/trigger
>
> So I was working to change the above to:
>
>   echo 'objtrace:add:+0x28($arg1):u32:1 if comm == "cat"'

It seems easier to use, I like it.

>
>
> >
> > # du -sh /test.txt
> > 12.0K   /test.txt
> >
> > # cat  /test.txt > /dev/null
> > # cat ./trace
> > # tracer: nop
> > #
> > # entries-in-buffer/entries-written: 128/128   #P:4
> > #
> > #                                _-----=> irqs-off/BH-disabled
> > #                               / _----=> need-resched
> > #                              | / _---=> hardirq/softirq
> > #                              || / _--=> preempt-depth
> > #                              ||| / _-=> migrate-disable
> > #                              |||| /     delay
> > #           TASK-PID     CPU#  |||||  TIMESTAMP  FUNCTION
> > #              | |         |   |||||     |         |
> >              cat-117     [002] ...1.     1.602243: __bio_try_merge_page <-bio_add_page object:0xffff88811bee4000 value:0x0
> >              cat-117     [002] ...1.     1.602244: __bio_add_page <-bio_add_page object:0xffff88811bee4000 value:0x0
> >              cat-117     [002] ...2.     1.602244: bio_add_page <-ext4_mpage_readpages object:0xffff88811bee4000 value:0x1000
> >              cat-117     [002] ...1.     1.602245: __bio_try_merge_page <-bio_add_page object:0xffff88811bee4000 value:0x1000
> >              cat-117     [002] ...1.     1.602245: __bio_add_page <-bio_add_page object:0xffff88811bee4000 value:0x1000
>
> Now instead of searching the function arguments for "obj", we should just
> read the obj value at every function and report when it changed, as well as
> the last function it was still the same. The report will be like the
> "func_repeats" entry. And instead we can print when the object changes,
> something like this:
>
>               cat-117     [002] ...2.     1.602245: bio_add_page <-ext4_mpage_readpages object:0xffff88811bee4000 value:0x2000 (last value:0x1000 at [002] __bio_add_page <-bio_add_page ts: 1.602245)

This output is better than I expected.

>
> This will tell us when a value changed.
>
> It would need to store the CPU, ip and pip and timestamp of the last time
> it was the same. And when it changes, it would then report it.
>
> -- Steve

-- 
Thanks,
JeffXie

Powered by blists - more mailing lists