Re: [SCM] systemtap: system-wide probe/trace tool branch, master, updated. release-0.9.8-184-gf982c59

Re: [SCM] systemtap: system-wide probe/trace tool branch, master, updated. release-0.9.8-184-gf982c59

[Mark Wielaard]

Hi Josh,

On Tue, 2009-07-21 at 01:22 +0000, jistone@sourceware.org wrote:
> commit 0c16d51256bc77c2b5497b72ec43b7864d2b47a8
> Author: Josh Stone <jistone@redhat.com>
> Date:   Mon Jul 20 16:32:33 2009 -0700
> 
>     Allow dwflpp to take a vector of module names
>     
>     This will be used to load tracepoint modules as a bunch of little query
>     modules into a single dwflpp.
>     
>     * dwflpp.cxx (setup_user): take a vector instead of a single module
>       (dwflpp::dwflpp): form a vector in the singular case, and add a
>       variant that takes and passes through a vector of modules.

This breaks user space mark probing. See make installcheck
RUNTESTFLAGS="sdt.exp exelib.exp" for example.

Most stuff will fail with "semantic error: libdwfl failure
(getshdrstrndx): no error while resolving probe point"
Which comes from dwarf_builder::probe_table::probe_table().
What goes wrong is that probe_table will try to fetch the probe table
(initiated through dwarf_builder::build) from the dw module (initialized
through dw = get_user_dw(sess, module_name)), if (get_param(parameters,
TOK_MARK, mark_name)). But at that point dw->module isn't setup yet.

I am not completely sure how to correct this.
The following hack makes things work as before:

diff --git a/dwflpp.cxx b/dwflpp.cxx
index 0bba3f1..38be86b 100644
--- a/dwflpp.cxx
+++ b/dwflpp.cxx
@@ -418,6 +418,7 @@ dwflpp::setup_user(const vector<string>& modules, bool debug
                                     module_name.c_str(),
                                     -1);
       // XXX: save mod!
+      module = mod;
 
       if (debuginfo_needed)
         dwfl_assert (string("missing process ") +

That works, since it mimics what happened before your patch. And because
there is really only one module name in the vector anyway.

I checked it in so things at least work again for now.
But it seems there must be a better/more elegant solution.

Cheers,

Mark

Re: [SCM] systemtap: system-wide probe/trace tool branch, master, updated. release-0.9.8-184-gf982c59

[Josh Stone]

On 07/21/2009 04:18 AM, Mark Wielaard wrote:
> Hi Josh,
> 
> On Tue, 2009-07-21 at 01:22 +0000, jistone@sourceware.org wrote:
>> commit 0c16d51256bc77c2b5497b72ec43b7864d2b47a8
>> Author: Josh Stone <jistone@redhat.com>
>> Date:   Mon Jul 20 16:32:33 2009 -0700
>>
>>     Allow dwflpp to take a vector of module names
>>     
>>     This will be used to load tracepoint modules as a bunch of little query
>>     modules into a single dwflpp.
>>     
>>     * dwflpp.cxx (setup_user): take a vector instead of a single module
>>       (dwflpp::dwflpp): form a vector in the singular case, and add a
>>       variant that takes and passes through a vector of modules.
> 
> This breaks user space mark probing. See make installcheck
> RUNTESTFLAGS="sdt.exp exelib.exp" for example.
> 
> Most stuff will fail with "semantic error: libdwfl failure
> (getshdrstrndx): no error while resolving probe point"
> Which comes from dwarf_builder::probe_table::probe_table().
> What goes wrong is that probe_table will try to fetch the probe table
> (initiated through dwarf_builder::build) from the dw module (initialized
> through dw = get_user_dw(sess, module_name)), if (get_param(parameters,
> TOK_MARK, mark_name)). But at that point dw->module isn't setup yet.

Clearly I didn't test the userspace stuff. :(  But as always, thanks for
being an excellent watchdog, Mark!

> I am not completely sure how to correct this.
> The following hack makes things work as before:
> 
> diff --git a/dwflpp.cxx b/dwflpp.cxx
> index 0bba3f1..38be86b 100644
> --- a/dwflpp.cxx
> +++ b/dwflpp.cxx
> @@ -418,6 +418,7 @@ dwflpp::setup_user(const vector<string>& modules, bool debug
>                                      module_name.c_str(),
>                                      -1);
>        // XXX: save mod!
> +      module = mod;
>  
>        if (debuginfo_needed)
>          dwfl_assert (string("missing process ") +
> 
> That works, since it mimics what happened before your patch. And because
> there is really only one module name in the vector anyway.
> 
> I checked it in so things at least work again for now.
> But it seems there must be a better/more elegant solution.

I just checked in a rework to make SDT behave more like the other dwarf
queries that we have.  Now it will iterate over whatever modules are
present (all one of them).  Hopefully that's a little more elegant for
you... :)

Josh