Patchwork [1/3] gdb: Don't fault for 'maint print psymbols' when using an index

login
register
mail settings
Submitter Andrew Burgess
Date Sept. 1, 2019, 4:08 p.m.
Message ID <d83c07adee7c0d29462dd1f64a11345f095609b0.1567353237.git.andrew.burgess@embecosm.com>
Download mbox | patch
Permalink /patch/34366/
State New
Headers show

Comments

Andrew Burgess - Sept. 1, 2019, 4:08 p.m.
I found that these tests:

  make check-gdb RUNTESTFLAGS="--target_board=cc-with-gdb-index gdb.base/maint.exp"
  make check-gdb RUNTESTFLAGS="--target_board=cc-with-debug-names gdb.base/maint.exp"

were causing GDB to segfault.  It turns out that this test runs this
command:

  maint print psymbols -pc main /path/to/some/file

which tries to lookup the partial_symtab for 'main'.  The problem is
that there is no partial_symtab for 'main' as we are using the
.gdb_index or .debug_names instead of partial_symtabs.

What happens is that maintenance_print_symbols calls
find_pc_sect_psymtab, which looks for the partial_symtab in the
objfile's objfile->partial_symtabs->psymtabs_addrmap.

This is a problem because when we are using the indexes
psymtabs_addrmap is reused to hold things other than partial_symtabs,
this can be seen in dwarf2read.c in create_addrmap_from_index and
create_addrmap_from_aranges.  If we then lookup in psymtabs_addrmap we
end up returning a pointer to something that isn't really a
partial_symtab, after which everything goes wrong.

Initially I simply added a check at the start of find_pc_sect_psymtab
that the objfile had some partial_symtabs, like:

  if (objfile->partial_symtabs->psymtabs == NULL)
    return NULL;

Figuring that if there were no partial_symtabs at all then this
function should always return NULL, however, this caused a failure in
the test gdb.python/py-event.exp which I didn't dig into too deeply,
but seems to be that in this tests there are initially no psymtabs,
but the second part of find_pc_sect_psymtab does manage to read some
in from somewhere, with the check I added the test fails as we
returned NULL here and this caused GDB to load in the full symtabs
earlier than was expected.

Instead I chose to guard only the access to psymtabs_addrmap with a
check that the function has some psymtabs.  This allows my original
tests to pass, and the py-event.exp test to pass too.

Now, a good argument can be made that we simply should never call
find_pc_sect_psymtab on an objfile that is using indexes instead of
partial_symtabs.  I did consider this approach, we could easily add an
assert into find_pc_sect_psymtab that if we find a partial_symtab in
psymtabs_addrmap then the psymtabs pointer must be non-null.  The
responsibility would then be on the user of find_pc_sect_psymtab to
ensure that the objfile being checked is suitable.  In the end I
didn't take this approach as the check in find_pc_sect_psymtab is
cheap and this ensures that any future miss-uses of the function will
not cause problems.

I also extended the comment on psymtabs_addrmap to indicate that it
holds more than just partial_symtabs as this was not at all clear from
the original comment, and caused me some confusion when I was
initially debugging this problem.

gdb/ChangeLog:

	* psymtab.c (find_pc_sect_psymtab): Move baseaddr local into more
	inner scope, add check that the objfile has psymtabs before
	checking psymtabs_addrmap.
	* psymtab.h (psymtab_storage) <psymtabs_addrmap>: Extend comment.
---
 gdb/ChangeLog |  7 +++++++
 gdb/psymtab.c | 24 +++++++++++++++++-------
 gdb/psymtab.h |  6 +++++-
 3 files changed, 29 insertions(+), 8 deletions(-)

Patch

diff --git a/gdb/psymtab.c b/gdb/psymtab.c
index 6cc7566580a..bc0f1d17783 100644
--- a/gdb/psymtab.c
+++ b/gdb/psymtab.c
@@ -308,14 +308,24 @@  find_pc_sect_psymtab (struct objfile *objfile, CORE_ADDR pc,
 		      struct obj_section *section,
 		      struct bound_minimal_symbol msymbol)
 {
-  CORE_ADDR baseaddr = ANOFFSET (objfile->section_offsets,
-				 SECT_OFF_TEXT (objfile));
-
-  /* Try just the PSYMTABS_ADDRMAP mapping first as it has better granularity
-     than the later used TEXTLOW/TEXTHIGH one.  */
-
-  if (objfile->partial_symtabs->psymtabs_addrmap != NULL)
+  /* Try just the PSYMTABS_ADDRMAP mapping first as it has better
+     granularity than the later used TEXTLOW/TEXTHIGH one.  However, we need
+     to take care as the PSYMTABS_ADDRMAP can hold things other than partial
+     symtabs in some cases.
+
+     This function should only be called for objfiles that are using partial
+     symtabs, not for objfiles that are using indexes (.gdb_index or
+     .debug_names), however 'maintenance print psymbols' calls this function
+     directly for all objfiles.  If we assume that PSYMTABS_ADDRMAP contains
+     partial symtabs then we will end up returning a pointer to an object
+     that is not a partial_symtab, which doesn't end well.  */
+
+  if (objfile->partial_symtabs->psymtabs != NULL
+      && objfile->partial_symtabs->psymtabs_addrmap != NULL)
     {
+      CORE_ADDR baseaddr = ANOFFSET (objfile->section_offsets,
+				     SECT_OFF_TEXT (objfile));
+
       struct partial_symtab *pst
 	= ((struct partial_symtab *)
 	   addrmap_find (objfile->partial_symtabs->psymtabs_addrmap,
diff --git a/gdb/psymtab.h b/gdb/psymtab.h
index aed686258d4..0ad2b49d9a5 100644
--- a/gdb/psymtab.h
+++ b/gdb/psymtab.h
@@ -109,7 +109,11 @@  public:
   /* Map addresses to the entries of PSYMTABS.  It would be more efficient to
      have a map per the whole process but ADDRMAP cannot selectively remove
      its items during FREE_OBJFILE.  This mapping is already present even for
-     PARTIAL_SYMTABs which still have no corresponding full SYMTABs read.  */
+     PARTIAL_SYMTABs which still have no corresponding full SYMTABs read.
+
+     The DWARF parser reuses this addrmap to store things other than
+     psymtabs in the cases where debug information is being read from, for
+     example, the .debug-names section.  */
 
   struct addrmap *psymtabs_addrmap = nullptr;