Environment Variables

  LD_PRELOAD=<l_so>       colon separated list of libso's to be pre loaded
  LD_DEBUG=<opts>         comma separated list of debug options
          =help           list available options
          =libs           show library search path
          =files          processing of input files
          =symbols        show search path for symbol lookup
          =bindings       show against which definition a symbol is bound

LD_LIBRARY_PATH and dlopen(3)

When dynamically loading a shared library during program runtime with dlopen(3), only the LD_LIBRARY_PATH as it was during program startup is evaluated. Therefore the following is a code smell:

// at startup LD_LIBRARY_PATH=/moose

// Assume /foo/libbar.so
setenv("LD_LIBRARY_PATH", "/foo", true /* overwrite */);

// Will look in /moose and NOT in /foo.
dlopen("libbar.so", RTLD_LAZY);

Libraries specified in LD_PRELOAD are loaded from left-to-right but initialized from right-to-left.

  > ldd ./main
    >> libc.so.6 => /usr/lib/libc.so.6

  > LD_PRELOAD=liba.so:libb.so ./main
      preloaded in this order
      initialized in this order

The preload order determines:

  • the order libraries are inserted into the link map
  • the initialization order for libraries

For the example listed above the resulting link map will look like the following:

  +------+    +------+    +------+    +------+
  | main | -> | liba | -> | libb | -> | libc |
  +------+    +------+    +------+    +------+

This can be seen when running with LD_DEBUG=files:

  > LD_DEBUG=files LD_PRELOAD=liba.so:libb.so ./main
    # load order (-> determines link map)
    >> file=liba.so [0];  generating link map
    >> file=libb.so [0];  generating link map
    >> file=libc.so.6 [0];  generating link map

    # init order
    >> calling init: /usr/lib/libc.so.6
    >> calling init: <path>/libb.so
    >> calling init: <path>/liba.so
    >> initialize program: ./main

To verify the link map order we let ld.so resolve the memcpy(3) libc symbol (used in main) dynamically, while enabling LD_DEBUG=symbols,bindings to see the resolving in action.

  > LD_DEBUG=symbols,bindings LD_PRELOAD=liba.so:libb.so ./main
    >> symbol=memcpy;  lookup in file=./main [0]
    >> symbol=memcpy;  lookup in file=<path>/liba.so [0]
    >> symbol=memcpy;  lookup in file=<path>/libb.so [0]
    >> symbol=memcpy;  lookup in file=/usr/lib/libc.so.6 [0]
    >> binding file ./main [0] to /usr/lib/libc.so.6 [0]: normal symbol `memcpy' [GLIBC_2.14]

Dynamic Linking (x86_64)

Dynamic linking basically works via one indirect jump. It uses a combination of function trampolines (.plt section) and a function pointer table (.got.plt section). On the first call the trampoline sets up some metadata and then jumps to the ld.so runtime resolve function, which in turn patches the table with the correct function pointer.

  .plt ....... procedure linkage table, contains function trampolines, usually
               located in code segment (rx permission)
  .got.plt ... global offset table for .plt, holds the function pointer table

Using radare2 we can analyze this in more detail:

  [0x00401040]> pd 4 @ section..got.plt
              ;-- section..got.plt:
              ;-- .got.plt:    ; [22] -rw- section size 32 named .got.plt
              ;-- _GLOBAL_OFFSET_TABLE_:
         [0]  0x00404000      .qword 0x0000000000403e10 ; section..dynamic
         [1]  0x00404008      .qword 0x0000000000000000
              ; CODE XREF from section..plt @ +0x6
         [2]  0x00404010      .qword 0x0000000000000000
              ;-- reloc.puts:
              ; CODE XREF from sym.imp.puts @ 0x401030
         [3]  0x00404018      .qword 0x0000000000401036 ; RELOC 64 puts

  [0x00401040]> pd 6 @ section..plt
              ;-- section..plt:
              ;-- .plt:       ; [12] -r-x section size 32 named .plt
          ┌─> 0x00401020      ff35e22f0000   push qword [0x00404008]
          ╎   0x00401026      ff25e42f0000   jmp qword [0x00404010]
          ╎   0x0040102c      0f1f4000       nop dword [rax]
  ┌ 6: int sym.imp.puts (const char *s);
  └       ╎   0x00401030      ff25e22f0000   jmp qword [reloc.puts]
          ╎   0x00401036      6800000000     push 0
          └─< 0x0040103b      e9e0ffffff     jmp sym..plt
  • At address 0x00401030 in the .plt section we see the indirect jump for puts using the function pointer in _GLOBAL_OFFSET_TABLE_[3] (GOT).
  • GOT[3] initially points to instruction after the puts trampoline 0x00401036.
  • This pushes the relocation index 0 and then jumps to the first trampoline 0x00401020.
  • The first trampoline jumps to GOT[2] which will be filled at program startup by the ld.so with its resolve function.
  • The ld.so resolve function fixes the relocation referenced by the relocation index pushed by the puts trampoline.
  • The relocation entry at index 0 tells the resolve function which symbol to search for and where to put the function pointer:
      > readelf -r <main>
        >> Relocation section '.rela.plt' at offset 0x4b8 contains 1 entry:
        >>   Offset          Info           Type           Sym. Value    Sym. Name + Addend
        >> 000000404018  000200000007 R_X86_64_JUMP_SLO 0000000000000000 puts@GLIBC_2.2.5 + 0
    As we can see the offset from relocation at index 0 points to GOT[3].