[buyemune-sys]

The problem

Two things in what dl_iterate_phdr hands out differ from glibc. Together they break LuaJIT's memprof and sysprof, which name C functions by walking the loaded objects (lj_symtab.c).

1. dlpi_adds and dlpi_subs were always zero. glibc counts every object that ever joined the list, and those that left it: adds=4 at startup, 5 after a dlopen. memprof dumps the libraries added since its previous dump. With zero adds it dumps none, and a profile names no C function. Tarantool's gh-5813-resolving-of-c-symbols test fails all 5 checks.

2. PT_DYNAMIC entries kept their link-time offsets. ld.so rebases the address-valued entries of a writable dynamic section in place, and walkers read them that way: DT_STRTAB, DT_SYMTAB, the hash tables, the relocation tables, DT_PLTGOT, DT_VERSYM. For a stripped library memprof reads the dynamic symbols through them.

   glibc: /usr/lib/x86_64-linux-gnu/libz.so.1  base=0x727c1afe1000 strtab=0x727c1afe2258
   solo:  /usr/lib/x86_64-linux-gnu/libz.so.1  base=0x7dc8f6d3e000 strtab=0x1258
   

   This stayed hidden only because of (1). Once memprof sees the libraries, it dereferences 0x240 and crashes (misclib-memprof-lapi.test.lua).

The fix

Two commits, in this order:

• Leave PT_DYNAMIC addresses absolute after loading, as ld.so does.
  • Right after parseDynamic, and before RELRO seals the section, the loader applies glibc's adjustment to the same tags glibc adjusts, for images whose PT_DYNAMIC is writable (LinkMap::rebaseDynamic, lib/elf_loader.cpp:1705). parseDynamic has already cached everything the loader itself reads.
  • The bridge's own image is no object ld.so would have loaded, and its static-PIE startup left the offsets in place. dl_iterate_phdr reports it with PT_DYNAMIC blanked to PT_NULL (hostHeaders, lib/glibc_shim.cpp:3772), so walkers skip it, as LuaJIT skips the read-only vDSO under glibc.
• Count loaded objects in dl_iterate_phdr's dlpi_adds.
  • The loader counts the images that reach Ready.
  • dl_iterate_phdr adds the host entries it reports itself.
  • Nothing unloads, so dlpi_subs stays zero, which is also what glibc reports before any dlclose.

Test

The conformance battery's dynamicLinking() checks two things:

• no reported image (vDSO aside) holds an offset in DT_STRTAB;
• dlpi_adds covers the images walked.

On main both fail, one per commit. With the fix: ./build test passes.

Under solo, LuaJIT's test suite then fails only where it also fails natively, plus lj-522, which needs LD_PRELOAD.