[Valgrind-users] Re: backtraces of leak-checking && dlopened modules (Nicholas Nethercote)

[Peter H S. <smi...@us...>]

I think Nick is on to something here.  You're already keeping a table
of stack traces to identify "duplicate calls".  You know the range of
addresses at munmap time, so you can sift through your table to
find the ones you need to "sourcify" now.  The cost of sifting the
table is linear, and some removal of duplicates has already
occurred.

When you encounter an address in the region you're about to
munmap, create the source string for it and add that address
and its source string to a "special table", and index that table
with a "special address."  Maybe 0xFFFF0001, 0xFFFF0002
could be the "special addresses."  Somewhere that mmap
isn't going to find (I'm not an X86 architect, and I don't play
one on TV :-)

For completeness, when you create the source string, tack
on the original address, and maybe something that helps
with understanding when the text was loaded.  You could
get fancy and refer to the Nth load of the library, and also
log each library load so the poor slobs debugging the code
can match things up.

I think there won't be many pathological cases where the
size of the translated stack address strings is more
than the munmapped file...

Peter