[symtab/27831] Fix OBJF_MAINLINE assert

This commit fixes a bug mentioned by Florian Weimer during the
libpthread/ld.so load order discussion from 2021.  Florian provided
instructions for reproducing the bug here:

https://sourceware.org/pipermail/gdb-patches/2021-April/177923.html

That particular test does some interesting things involving forks,
threads, and thread local storage.  Fortunately, none of that is
needed to reproduce the problem.

I've made a new test case (which is now found in a separate commit)
contained in the files gdb.base/add-symbol-file-attach.{c,exp}.  The
.c file is fairly simple as is the recipe for reproducing the problem.

After separately starting the test case and noting the process id,
start gdb (w/ no arguments), and do the following to reproduce the
assertion failure - for this run, the process id of the separately
started add-symbol-file-attach process is 4103218:

(gdb) add-symbol-file add-symbol-file-attach
add symbol table from file "add-symbol-file-attach"
(y or n) y
Reading symbols from add-symbol-file-attach...
(gdb) attach 4103218
Attaching to process 4103218
Load new symbol table from "/tmp/add-symbol-file-attach"? (y or n) y
Reading symbols from /tmp/add-symbol-file-attach...
Reading symbols from /lib64/libc.so.6...
(No debugging symbols found in /lib64/libc.so.6)
Reading symbols from /lib64/ld-linux-x86-64.so.2...
(No debugging symbols found in /lib64/ld-linux-x86-64.so.2)
0x00007f502130bf27 in pause () from /lib64/libc.so.6
(gdb) p foo
symtab.c:6417: internal-error: CORE_ADDR get_msymbol_address(objfile*,
  const minimal_symbol*): Assertion `(objf->flags & OBJF_MAINLINE) == 0'
  failed.
A problem internal to GDB has been detected,
further debugging may prove unreliable.

The add-symbol-file command causes the symbols to be loaded without
the SYMFILE_MAINLINE (and hence the OBJFILE_MAINLINE) flags being
set.  This, in turn, causes the "maybe_copied" flag to be set for
the global symbol (named "foo" in the provided test case).

The attach command will cause another objfile to be created, but
it will reuse the symtabs from the objfile created by add-symbol-file,
leading to a situation in which the OBJFILE_MAINLINE flag will be set
for the new (attach-created) objfile, however the "maybe_copied"
flag will still be set for the global symbol.  Had it been loaded
anew, this flag would not be set due to OBJFILE_MAINLINE being set
for the objfile.

At present, minimal_symbol::value_address looks like this:

CORE_ADDR
minimal_symbol::value_address (objfile *objfile) const
{
  if (this->maybe_copied (objfile))
    return get_msymbol_address (objfile, this);
  else
    return (CORE_ADDR (this->unrelocated_address ())
	    + objfile->section_offsets[this->section_index ()]);
}

So, we can now see the problem: When the "maybe_copied" flag is set,
get_msymbol_address() will be called.  However, get_msymbol_address()
assumes that it won't be called with the OBF_MAINLINE flag set for
the objfile in question.  It, in fact, contains an assert() which
makes sure that this is the case:

  gdb_assert ((objf->flags & OBJF_MAINLINE) == 0);

(If this assert is removed, then get_msymbol_address() recurses
infinitely for the case under consideration.)

So, the problem here is that the maybe_copied flag is set for the
symbol AND the OBJF_MAINLINE flag is set for the objfile.  As noted
earlier, this happens due to add-symbol-file being used; this causes
the maybe_copied flag to be set.  Later, when the attach is performed,
OBJF_MAINLINE will be set for that objfile, leading to this
unfortunate situation.

My first cut at a solution involved adjusting the
MSYMBOL_VALUE_ADDRESS macro (which has since been changed to be the
method noted above) to include a test of the OBJFILE_MAINLINE flag.
However, Simon Marchi, in his review of my patch, suggested a better
solution.  Simon observed that the 'maybe_copied' flag is (was, after
this commit) being set/initialized in record_minimal_symbol() using
using the objfile in the context in which the symbol was created.

Simon further observed:

  Today, a single copy is created, as symtabs are shared between
  objfiles.  This means that everything that we store into a symbol
  must be independent of any objfile.  However, the value of the
  maybe_copied field is dependent on the objfile in the context of
  which the symbol was created.  Meaning that when the symbol is
  re-used in the context of another objfile, the maybe_copied value is
  not right in the context of that objfile.

  So I think it means there isn't a single "is this symbol maybe
  copied" value, but instead "is this symbol maybe copied, in the
  context of this given objfile".  And the answer is yes or no,
  depending on whether the objfile is mainline.  So maybe_copied
  should become a method that takes an objfile and returns an answer
  based on that.

Simon's full review can be found here:

  https://sourceware.org/pipermail/gdb-patches/2021-May/178855.html

Simon also provided a patch which implements this suggestion.  The
current patch is mostly his work, though I did make some adjustments
during a rebase in addition to making some changes to account for a
concern from Tom Tromey.

During his review of the v3 series, Tom noted, "The old approach was
specific to ELF, while the new approach will be used by any object
format." Tom further observed, "...it seems like it could result in an
incorrect evaluation in some scenario."  This seemed plausible to me,
so I introduced the flag 'object_format_has_copy_relocs' to struct
objfile.  It is set at the end of elf_symfile_read() in elfread.c.
The minimal_symbol::maybe_copied method tests this new flag, forcing
this method to return false when the flag is not set.  If we find that
other object file formats use the same copy reloc mechanism as ELF,
then 'object_format_has_copy_relocs' should be set for objfiles using
those formats.

Lastly, I'll note that this is a strange use case.  It's far more
common to either let gdb figure out which file to load by itself when
attaching, i.e.

(gdb) attach 4104360
Attaching to process 4104360
Reading symbols from /tmp/add-symbol-file-attach...
Reading symbols from /lib64/libc.so.6...
(No debugging symbols found in /lib64/libc.so.6)
Reading symbols from /lib64/ld-linux-x86-64.so.2...
(No debugging symbols found in /lib64/ld-linux-x86-64.so.2)
0x00007fdb1fc33f27 in pause () from /lib64/libc.so.6
(gdb) p foo
$1 = 42

...or to use the "file" command prior to the attach, like this:

(gdb) file add-symbol-file-attach
Reading symbols from add-symbol-file-attach...
(gdb) attach 4104360
Attaching to program: /tmp/add-symbol-file-attach, process 4104360
Reading symbols from /lib64/libc.so.6...
(No debugging symbols found in /lib64/libc.so.6)
Reading symbols from /lib64/ld-linux-x86-64.so.2...
(No debugging symbols found in /lib64/ld-linux-x86-64.so.2)
0x00007fdb1fc33f27 in pause () from /lib64/libc.so.6

Both of these more common scenarios work perfectly fine; using
"add-symbol-file" to load the program to which you will attach
isn't recommended as a normal use case.  That said, it's bad for
gdb to assert, hence this fix.

Reviewed-by: Simon Marchi <simon.marchi@polymtl.ca>
Co-Authored-by: Simon Marchi <simon.marchi@polymtl.ca>
Approved-by: Tom Tromey <tom@tromey.com>
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=27831

1 files changed, 11 insertions, 0 deletions

diff --git a/gdb/objfiles.h b/gdb/objfiles.h
index 91dda1c..b3b216f 100644
--- a/gdb/objfiles.h
+++ b/gdb/objfiles.h
@@ -871,6 +871,17 @@ public:
      next time.  If an objfile does not have the symbols, it will
      never have them.  */
   bool skip_jit_symbol_lookup = false;
+
+  /* Flag which indicates, when true, that the object format
+     potentially supports copy relocations.  ABIs for some
+     architectures that use ELF have a copy relocation in which the
+     initialization for a global variable defined in a shared object
+     will be copied to memory allocated to the main program during
+     dynamic linking.  Therefore this flag will be set for ELF
+     objfiles.  Other object formats that use the same copy relocation
+     mechanism as ELF should set this flag too.  This flag is used in
+     conjunction with the minimal_symbol::maybe_copied method.  */
+  bool object_format_has_copy_relocs = false;
 };
 
 /* A deleter for objfile.  */