On Thu, 2024-11-14 at 15:58 +0000, Alan Maguire wrote: > Eduard noticed [1] intermittent segmentation faults triggered by caching > done internally in dwarf_getlocation(s). A binary tree of location > information is cached in the CU, and if multiple threads access it > concurrently we can get segmentation faults. It is possible to > compile elfutils with experimental thread-safe support, but safer for > now to add locking to pahole. > > No additional overhead in pahole encoding was observed > as a result of these changes. > > Reported-by: Eduard Zingerman <eddyz87@xxxxxxxxx> > Suggested-by: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx> > Suggested-by: Jiri Olsa <olsajiri@xxxxxxxxx> > Signed-off-by: Alan Maguire <alan.maguire@xxxxxxxxxx> > > [1] https://lore.kernel.org/dwarves/080794545d8eb3df3d6eba90ac621111ab7171f5.camel@xxxxxxxxx/ > --- Looking at the libdw code, both dwarf_getlocations() and dwarf_getlocation() might end up calling __libdw_intern_expression(), where race condition occurs. So I think that this lock positioning should be safe. In theory, the locking could be pushed down, to only occur around __dwarf_getlocations / dwarf_getlocation, but that would complicate the code a bit. Acked-by: Eduard Zingerman <eddyz87@xxxxxxxxx> [...]
