DTrace: stack() does not print kernel module functions for i386

[John Baldwin]

On Sunday, November 09, 2014 1:57:19 pm Rui Paulo wrote:
> On Nov 9, 2014, at 01:36, Konstantin Belousov <kostikbel at gmail.com> wrote:
> > 
> > On Sat, Nov 08, 2014 at 02:06:39PM -0800, Shrikanth Kamath wrote:
> >> I verified this on a FreeBSD 10.0 STABLE, the stack() feature does not
> >> appear to print functions from kernel modules. I believe there is a
> >> glitch in libdtrace in the function "dt_module_update"
> >> (cddl/contrib/opensolaris/lib/libdtrace/common/dt_module.c).
> >> 
> >> The section header address computation which is currently being done
> >> in the function dt_module_update for elf type ET_REL, a similar
> >> computation needs to be done for the ET_DYN maybe. I lack background
> >> on the elf types but for experiment sakes I changed the line
> >> 
> >> @@ -948,7 +948,7 @@ dt_module_update(dtrace_hdl_t *dtp, struct kld_fil
> >> #if defined(__FreeBSD__)
> >>        mapbase = (uintptr_t)k_stat->address;
> >>        gelf_getehdr(dmp->dm_elf, &ehdr);
> >> -       is_elf_obj = (ehdr.e_type == ET_REL);
> >> +      is_elf_obj = (ehdr.e_type == ET_REL) || (ehdr.e_type == ET_DYN);
> >>        if (is_elf_obj) {
> >>                dmp->dm_sec_offsets =
> >>                    malloc(ehdr.e_shnum * sizeof(*dmp->dm_sec_offsets));
> >> 
> >> So from a previous run where I was running a dtrace one liner
> >> %dtrace -n 'fbt:hwpmc:: { stack(); }'
> >> The output without the above change shows a sample as
> >> 
> >> 0  50825 pmc_find_process_descriptor:entry
> >>              0xc3c35bf5                                  <-- Address
> >> not matched to symbol
> >>              kernel`syscall+0x48b
> >>              kernel`0xc0fd2121
> >> 
> >> whereas with the above nit change to include ET_DYN for section offset
> >> adjustment I get the complete stack trace as
> >> 
> >> 0  50825 pmc_find_process_descriptor:entry
> >>              hwpmc.ko`pmc_hook_handler+0x6a5   <--address matched to 
symbol
> >>              kernel`syscall+0x48b
> >>              kernel`0xc0fd2121
> >> 
> >> I believe without the correct section offset adjustment the following
> >> check in the function dtrace_lookup_by_addr fails to match the address
> >> to the correct symbol
> >> 
> >>            if (addr - dmp->dm_text_va < dmp->dm_text_size ||
> >>                    addr - dmp->dm_data_va < dmp->dm_data_size ||
> >>                    addr - dmp->dm_bss_va < dmp->dm_bss_size)
> >> 
> >> because dml->dm_text_va was not setup correctly in dt_module_update.
> >> 
> >> Can somebody help clarify this?
> >> 
> >> Reference: commit revision 210425.
> > 
> > I have no idea about DTrace guts, but can add one note you might find
> > useful.
> > 
> > From the backtace above I can conclude that your platform is i386.
> > A difference between i386 and amd64 is that i386 modules are dso's
> > (ET_DYN), while on amd64 modules are only linked to object files
> > (ET_REL).  The original author of the code probably tested on amd64.
> > 
> > You may want to special case amd64 by #ifdef, and use ET_DYN on other
> > arches.
> 
> I agree with your analysis.  I think this patch should go in with the #ifdef 
__amd64__ for ET_REL.
> 

Why have the #ifdef?  In theory other platforms besides amd64 could use
sys/kern/link_elf_obj.c.  It doesn't hurt to just let the code always accept
both ET_DYN and ET_REL does it?

-- 
John Baldwin