decoding of multi-byte nops in dtrace

[Fabian Keil]

Andriy Gapon <avg at FreeBSD.org> wrote:

> on 10/06/2012 20:56 Fabian Keil said the following:
> > Andriy Gapon <avg at FreeBSD.org> wrote:
> > 
> >> on 10/06/2012 16:27 Fabian Keil said the following:
> >>> Andriy Gapon <avg at FreeBSD.org> wrote:
> >>>
> >>>> It seems that the $subj is missing :-)
> >>>> In my environment that causes many functions to not have fbt return probe,
> >>>> because function body decoding fails before 'ret' is found.
> >>>>
> >>>> Here is my attempt at fixing the problem:
> >>>> http://people.freebsd.org/~avg/fbt-nop.patch
> >>>> Reviews and suggestions are welcome.
> >>>
> >>> The patch seems to reduce the number of missing
> >>> fbt return probes by about 50% for me.
> >>>
> >>> Without the patch:
> >>>
> >>> fk at r500 /usr/src $sudo dtrace -ln fbt::: | grep -c entry
> >>> 23395
> >>> fk at r500 /usr/src $sudo dtrace -ln fbt::: | grep -c return
> >>> 16739
> >>>
> >>> With the patch (and updated kernel sources):
> >>>
> >>> fk at r500 ~ $sudo dtrace -ln fbt::: | grep -c entry
> >>> 23409
> >>> fk at r500 ~ $sudo dtrace -ln fbt::: | grep -c return
> >>> 19879
> >>
> >> Interesting observations, thank you.
> >> Do you use -O2 or higher optimization for kernel/modules build?
> > 
> > Yes, I kept the default -O2.
> > 
> >> I use only -O1.
> > 
> > With -O1 (and your patch) I get:
> > 
> > fk at r500 ~ $sudo dtrace -ln fbt::: | grep -c entry
> > 23421
> > fk at r500 ~ $sudo dtrace -ln fbt::: | grep -c return
> > 22621
> > 
> >> Here are some stats from my system:
> >> $ dtrace -ln fbt::: | fgrep -c entry
> >> 16876
> >> $ dtrace -ln fbt::: | fgrep -c return
> >> 16729
> >>
> >> So, 147 functions without return probe.
> >> >From a quick look at them they all seem to really never return.  Either they are
> >> noreturn type such panic, or functions that always call the functions of the
> >> first type, or functions with endless loops in them such as top level functions
> >> of many system threads.
> > 
> > I looked at a couple of the functions that still lack return
> > probes and the ones I looked at don't seem to belong into these
> > categories.
> > 
> > For example I get no return probes for g_eli_crypto_decrypt()
> > and g_eli_crypto_encrypt(). Both return the return code of
> > g_eli_crypto_cipher() for which I get a return probe.
> 
> I don't have GELI in kernel, but it looks like an instance of well-known tail
> call optimization issue.  Although I assumed that GCC wouldn't apply it at -O1.
> Perhaps you use a module that was built with -O2.

That was it. I missed that COPTFLAGS aren't applied to the modules.
After recompiling geom_eli manually with CFLAGS=-O1 the return
codes show up as expected

How did you set your -O1?

Fabian
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 196 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-hackers/attachments/20120610/2b277271/signature.pgp