(In reply to Martin Milata from comment #0)
> Description of problem:
> Resolving code address to symbol, e.g. when running eu-stack, seems to fail
> when MiniDebugInfo is used on PPC64. I always got
> 0000003b.plt_call.SOMETHING@@GLIBC_2.3 as the symbol name instead.

Hmmmm. Did you try installing the debuginfo for the program to make sure it was the minisymtab? What does the backtrace look like if you add the debuginfo packages with debuginfo-install coreutils?

If that does work correctly it might be that the minisymtab table is not correctly setup on ppc64. Could you show how it looks with eu-readelf --elf-section --symbols /usr/bin/sleep?

Could you compare it, or add, the full symbol table from the debug file with
eu-readelf --symbols /usr/lib/debug/usr/bin/sleep.debug?

(In reply to Mark Wielaard from comment #2)
> Hmmmm. Did you try installing the debuginfo for the program to make sure it
> was the minisymtab? What does the backtrace look like if you add the
> debuginfo packages with debuginfo-install coreutils?

With debuginfo, the output seems to be alright:

ppc$ yum install coreutils-debuginfo
ppc$ eu-stack -e `which sleep` --core core.*
PID 31401 - core
TID 31401:
#0  0x00003fffa7a68808 __nanosleep
#1  0x0000000010004f30 rpl_nanosleep
#2  0x0000000010004ca4 xnanosleep
#3  0x00000000100017e4 main
#4  0x00003fffa79c44cc generic_start_main.isra.0
eu-stack: dwfl_thread_getframes tid 31401 at 0x3fffa79c44cb in libc.so.6: No DWARF information found

> If that does work correctly it might be that the minisymtab table is not
> correctly setup on ppc64. Could you show how it looks with eu-readelf
> --elf-section --symbols /usr/bin/sleep?
> 
> Could you compare it, or add, the full symbol table from the debug file with
> eu-readelf --symbols /usr/lib/debug/usr/bin/sleep.debug?

Please see the attachment for output of these commands.

Created attachment 849911 [details]
Symbol tables from both minidebuginfo and separate debuginfo

Thanks that is very interesting. So the issue is that the .gnu_debugdata section embedded in the executable doesn't actually contain any function symbols at all.

For example you would expect the xnanosleep one to be there. And it does contain non-STT_FUNC symbols like these 00000000.plt_call. ones. This suggests that the .gnu_debugdata was wrongly generated.

What we could do to make output a little less confusing is refuse to show names of non-STT_FUNC symbols that we do find and that we conclude is "closest" to the address. But that doesn't really solve your issue, since then we just print the address without any name.

So lets reassign this to rpm first so they can figure out what is going wrong with the .gnu_debugdata minisymtab generation on ppc64 in the first place.

I looked on an actual ppc64 install and found the issue.
rpm's find-debuginfo.sh does the following to find the function symbols:

add_minidebug()
{
  local debuginfo="$1"
  local binary="$2"

  local dynsyms=`mktemp`
  local funcsyms=`mktemp`
  local keep_symbols=`mktemp`
  local mini_debuginfo=`mktemp`

  # Extract the dynamic symbols from the main binary, there is no need to also have these
  # in the normal symbol table
  nm -D "$binary" --format=posix --defined-only | awk '{ print $1 }' | sort > "$dynsyms"
  # Extract all the text (i.e. function) symbols from the debuginfo 
  nm "$debuginfo" --format=posix --defined-only | awk '{ if ($2 == "T" || $2 == "t") print $1 }' | sort > "$funcsyms"
  # Keep all the function symbols not already in the dynamic symbol table
  comm -13 "$dynsyms" "$funcsyms" > "$keep_symbols"
  # Copy the full debuginfo, keeping only a minumal set of symbols and removing some unnecessary sections
  objcopy -S --remove-section .gdb_index --remove-section .comment --keep-symbols="$keep_symbols" "$debuginfo" "$mini_debuginfo" &> /dev/null
  #Inject the compressed data into the .gnu_debugdata section of the original binary
  xz "$mini_debuginfo"
  mini_debuginfo="${mini_debuginfo}.xz"
  objcopy --add-section .gnu_debugdata="$mini_debuginfo" "$binary"
  rm -f "$dynsyms" "$funcsyms" "$keep_symbols" "$mini_debuginfo"
}

Where things seem to go wrong is at this step:
  # Extract all the text (i.e. function) symbols from the debuginfo 
  nm "$debuginfo" --format=posix --defined-only | awk '{ if ($2 == "T" || $2 == "t") print $1 }' | sort > "$funcsyms"

If we take sleep as example then it will not find any real function symbols because:

nm /usr/lib/debug/usr/bin/sleep.debug --format=posix --defined-only | grep xnanosleep
xnanosleep B 00000000100204b0 00000000000000cc

Note that it says 'B' ("The symbol is in the uninitialized data section (known as BSS)."), not 'T' ("The symbol is in the text (code) section.").

This is because binutils nm decides it does see a STT_FUNC symbol, but that the value points into the .odp section, which is has type NOBITS in the .debug file. I guess that is one valid interpretation but not a very useful one in this case.

One workaround would be to use elfutils eu-nm instead which does say the type is "T":

eu-nm /usr/lib/debug/usr/bin/sleep.debug --format=posix --defined-only | grep xnanosleep
xnanosleep T 00000000100204b0 00000000000000cc

Another would be to file against binutils nm (assuming the eu-nm interpretation of the symbol type is the correct one.

GDB needed a ppc64 fix which apparently has not made it into find-debuginfo.sh:

https://sourceware.org/git/?p=binutils-gdb.git;a=commitdiff;h=e58fcc15916bdb8be2b318a685430a597824a606

Just the fix uses ELF symbol type "D" while the ELF symbol type is "B" in your example above.

O, another solution might be to use --format-sysv which does print the symbol type explicitly (without any interpretation):

nm /usr/lib/debug/usr/bin/sleep.debug --format=sysv --defined-only | grep xnanosleep
xnanosleep          |00000000100204b0|   B  |              FUNC|00000000000000cc|     |.opd

Making the add_minidebug script line match field 4 against FUNC:
nm "$debuginfo" --format=sysv --defined-only | awk -F \| '{ if ($4 ~ "FUNC") print $1 }'

(In reply to Jan Kratochvil from comment #7)
> GDB needed a ppc64 fix which apparently has not made it into
> find-debuginfo.sh:
> 
> https://sourceware.org/git/?p=binutils-gdb.git;a=commitdiff;
> h=e58fcc15916bdb8be2b318a685430a597824a606
> 
> Just the fix uses ELF symbol type "D" while the ELF symbol type is "B" in
> your example above.

The reason for the difference "D" ("The symbol is in the initialized data section.") and "B" ("The symbol is in the uninitialized data section (known as BSS).") might be because the gbd testsuite is matching against the main elf file (where .odp is initialized data) and the rpm script is matching against the debug file (where .odp is NOBITS).

I think using --format=sysv style might be preferable to match the ELF symbol type directly instead of using the somewhat ambiguous symbol type char used in --format=posix style in binutils nm.

Just post a patch, according to the past experience rpm maintainers left /usr/lib/rpm/find-debuginfo.sh maintenance to the Tools people.

(In reply to Jan Kratochvil from comment #10)
> Just post a patch, according to the past experience rpm maintainers left
> /usr/lib/rpm/find-debuginfo.sh maintenance to the Tools people.

I couldn't find the add_minidebug () code upstream, so I am not sure where it is maintained by who. But the proposed patch is simple (against my installed version of find-debuginfo.sh):

--- find-debuginfo.sh.orig	2014-01-14 18:20:52.969317796 +0100
+++ find-debuginfo.sh.new	2014-01-14 18:23:40.572867474 +0100
@@ -149,7 +149,10 @@
   # in the normal symbol table
   nm -D "$binary" --format=posix --defined-only | awk '{ print $1 }' | sort > "$dynsyms"
   # Extract all the text (i.e. function) symbols from the debuginfo 
-  nm "$debuginfo" --format=posix --defined-only | awk '{ if ($2 == "T" || $2 == "t") print $1 }' | sort > "$funcsyms"
+  # Use format sysv to make sure we can match against the actual ELF FUNC
+  # symbol type. The binutils nm posix format symbol type chars are
+  # ambigous for architectures that might use function descriptors.
+  nm "$debuginfo" --format=sysv --defined-only | awk -F \| '{ if ($4 ~ "FUNC") print $1 }' | sort > "$funcsyms"
   # Keep all the function symbols not already in the dynamic symbol table
   comm -13 "$dynsyms" "$funcsyms" > "$keep_symbols"
   # Copy the full debuginfo, keeping only a minumal set of symbols and removing some unnecessary sections

This of course needs some testing to make sure it actually works correctly on ppc64 and doesn't change the .gnu_debugdata section on other architectures.

(In reply to Mark Wielaard from comment #11)
> I couldn't find the add_minidebug () code upstream, so I am not sure where
> it is maintained by who. But the proposed patch is simple (against my
> installed version of find-debuginfo.sh):

AFAIK the code lives as a patch in dist-git.

Fixed in rpm-4.11.1-11.el7 (by the patch from comment #11)

The patch looks good to me.

This request was resolved in Red Hat Enterprise Linux 7.0.

Contact your manager or support representative in case you have further questions about the request.