Load-time relocation of shared libraries

Loading executables

Linux, similarly to other OSes with virtual memory support, loads executables to a fixed memory address. If we examine the ELF header of some random executable, we'll see an Entry point address:

$ readelf -h /usr/bin/uptime
ELF Header:
  Magic:   7f 45 4c 46 01 01 01 00 00 00 00 00 00 00 00 00
  Class:                             ELF32
  [...] some header fields
  Entry point address:               0x8048470
  [...] some header fields

This is placed by the linker to tell the OS where to start executing the executable's code [1]. And indeed if we then load the executable with GDB and examine the address 0x8048470, we'll see the first instructions of the executable's .text segment there.

What this means is that the linker, when linking the executable, can fully resolve all internal symbol references (to functions and data) to fixed and final locations. The linker does some relocations of its own [2], but eventually the output it produces contains no additional relocations.

Or does it? Note that I emphasized the word internal in the previous paragraph. As long as the executable needs no shared libraries [3], it needs no relocations. But if it does use shared libraries (as do the vast majority of Linux applications), symbols taken from these shared libraries need to be relocated, because of how shared libraries are loaded.

Loading shared libraries

Unlike executables, when shared libraries are being built, the linker can't assume a known load address for their code. The reason for this is simple. Each program can use any number of shared libraries, and there's simply no way to know in advance where any given shared library will be loaded in the process's virtual memory. Many solutions were invented for this problem over the years, but in this article I will just focus on the ones currently used by Linux.

But first, let's briefly examine the problem. Here's some sample C code [4] which I compile into a shared library:

int myglob = 42;

int ml_func(int a, int b)
{
    myglob += a;
    return b + myglob;
}

Note how ml_func references myglob a few times. When translated to x86 assembly, this will involve a mov instruction to pull the value of myglob from its location in memory into a register. mov requires an absolute address - so how does the linker know which address to place in it? The answer is - it doesn't. As I mentioned above, shared libraries have no pre-defined load address - it will be decided at runtime.

In Linux, the dynamic loader [5] is a piece of code responsible for preparing programs for running. One of its tasks is to load shared libraries from disk into memory, when the running executable requests them. When a shared library is loaded into memory, it is then adjusted for its newly determined load location. It is the job of the dynamic loader to solve the problem presented in the previous paragraph.

There are two main approaches to solve this problem in Linux ELF shared libraries:

1. Load-time relocation
2. Position independent code (PIC)

Although PIC is the more common and nowadays-recommended solution, in this article I will focus on load-time relocation. Eventually I plan to cover both approaches and write a separate article on PIC, and I think starting with load-time relocation will make PIC easier to explain later. (Update 03.11.2011: the article about PIC was published)

Linking the shared library for load-time relocation

To create a shared library that has to be relocated at load-time, I'll compile it without the -fPIC flag (which would otherwise trigger PIC generation):

gcc -g -c ml_main.c -o ml_mainreloc.o
gcc -shared -o libmlreloc.so ml_mainreloc.o

The first interesting thing to see is the entry point of libmlreloc.so:

$ readelf -h libmlreloc.so
ELF Header:
  Magic:   7f 45 4c 46 01 01 01 00 00 00 00 00 00 00 00 00
  Class:                             ELF32
  [...] some header fields
  Entry point address:               0x3b0
  [...] some header fields

For simplicity, the linker just links the shared object for address 0x0 (the .text section starting at 0x3b0), knowing that the loader will move it anyway. Keep this fact in mind - it will be useful later in the article.

Now let's look at the disassembly of the shared library, focusing on ml_func:

$ objdump -d -Mintel libmlreloc.so

libmlreloc.so:     file format elf32-i386

[...] skipping stuff

0000046c <ml_func>:
 46c: 55                      push   ebp
 46d: 89 e5                   mov    ebp,esp
 46f: a1 00 00 00 00          mov    eax,ds:0x0
 474: 03 45 08                add    eax,DWORD PTR [ebp+0x8]
 477: a3 00 00 00 00          mov    ds:0x0,eax
 47c: a1 00 00 00 00          mov    eax,ds:0x0
 481: 03 45 0c                add    eax,DWORD PTR [ebp+0xc]
 484: 5d                      pop    ebp
 485: c3                      ret

[...] skipping stuff

After the first two instructions which are part of the prologue [6], we see the compiled version of myglob += a [7]. The value of myglob is taken from memory into eax, incremented by a (which is at ebp+0x8) and then placed back into memory.

But wait, the mov takes myglob? Why? It appears that the actual operand of mov is just 0x0 [8]. What gives? This is how relocations work. The linker places some provisional pre-defined value (0x0 in this case) into the instruction stream, and then creates a special relocation entry pointing to this place. Let's examine the relocation entries for this shared library:

$ readelf -r libmlreloc.so

Relocation section '.rel.dyn' at offset 0x2fc contains 7 entries:
 Offset     Info    Type            Sym.Value  Sym. Name
00002008  00000008 R_386_RELATIVE
00000470  00000401 R_386_32          0000200C   myglob
00000478  00000401 R_386_32          0000200C   myglob
0000047d  00000401 R_386_32          0000200C   myglob
[...] skipping stuff

The rel.dyn section of ELF is reserved for dynamic (load-time) relocations, to be consumed by the dynamic loader. There are 3 relocation entries for myglob in the section showed above, since there are 3 references to myglob in the disassembly. Let's decipher the first one.

It says: go to offset 0x470 in this object (shared library), and apply relocation of type R_386_32 to it for symbol myglob. If we consult the ELF spec we see that relocation type R_386_32 means: take the value at the offset specified in the entry, add the address of the symbol to it, and place it back into the offset.

What do we have at offset 0x470 in the object? Recall this instruction from the disassembly of ml_func:

46f:  a1 00 00 00 00          mov    eax,ds:0x0

a1 encodes the mov instruction, so its operand starts at the next address which is 0x470. This is the 0x0 we see in the disassembly. So back to the relocation entry, we now see it says: add the address of myglob to the operand of that mov instruction. In other words it tells the dynamic loader - once you perform actual address assignment, put the real address of myglob into 0x470, thus replacing the operand of mov by the correct symbol value. Neat, huh?

Note also the "Sym. value" column in the relocation section, which contains 0x200C for myglob. This is the offset of myglob in the virtual memory image of the shared library (which, recall, the linker assumes is just loaded at 0x0). This value can also be examined by looking at the symbol table of the library, for example with nm:

$ nm libmlreloc.so
[...] skipping stuff
0000200c D myglob

This output also provides the offset of myglob inside the library. D means the symbol is in the initialized data section (.data).

Load-time relocation in action

To see the load-time relocation in action, I will use our shared library from a simple driver executable. When running this executable, the OS will load the shared library and relocate it appropriately.

Curiously, due to the address space layout randomization feature which is enabled in Linux, relocation is relatively difficult to follow, because every time I run the executable, the libmlreloc.so shared library gets placed in a different virtual memory address [9].

This is a rather weak deterrent, however. There is a way to make sense in it all. But first, let's talk about the segments our shared library consists of:

$ readelf --segments libmlreloc.so

Elf file type is DYN (Shared object file)
Entry point 0x3b0
There are 6 program headers, starting at offset 52

Program Headers:
  Type           Offset   VirtAddr   PhysAddr   FileSiz MemSiz  Flg Align
  LOAD           0x000000 0x00000000 0x00000000 0x004e8 0x004e8 R E 0x1000
  LOAD           0x000f04 0x00001f04 0x00001f04 0x0010c 0x00114 RW  0x1000
  DYNAMIC        0x000f18 0x00001f18 0x00001f18 0x000d0 0x000d0 RW  0x4
  NOTE           0x0000f4 0x000000f4 0x000000f4 0x00024 0x00024 R   0x4
  GNU_STACK      0x000000 0x00000000 0x00000000 0x00000 0x00000 RW  0x4
  GNU_RELRO      0x000f04 0x00001f04 0x00001f04 0x000fc 0x000fc R   0x1

 Section to Segment mapping:
  Segment Sections...
   00     .note.gnu.build-id .hash .gnu.hash .dynsym .dynstr .gnu.version .gnu.version_r .rel.dyn .rel.plt .init .plt .text .fini .eh_frame
   01     .ctors .dtors .jcr .dynamic .got .got.plt .data .bss
   02     .dynamic
   03     .note.gnu.build-id
   04
   05     .ctors .dtors .jcr .dynamic .got

To follow the myglob symbol, we're interested in the second segment listed here. Note a couple of things:

• In the section to segment mapping in the bottom, segment 01 is said to contain the .data section, which is the home of myglob
• The VirtAddr column specifies that the second segment starts at 0x1f04 and has size 0x10c, meaning that it extends until 0x2010 and thus contains myglob which is at 0x200C.

Now let's use a nice tool Linux gives us to examine the load-time linking process - the dl_iterate_phdr function, which allows an application to inquire at runtime which shared libraries it has loaded, and more importantly - take a peek at their program headers.

So I'm going to write the following code into driver.c:

#define _GNU_SOURCE
#include <link.h>
#include <stdlib.h>
#include <stdio.h>


static int header_handler(struct dl_phdr_info* info, size_t size, void* data)
{
    printf("name=%s (%d segments) address=%p\n",
            info->dlpi_name, info->dlpi_phnum, (void*)info->dlpi_addr);
    for (int j = 0; j < info->dlpi_phnum; j++) {
         printf("\t\t header %2d: address=%10p\n", j,
             (void*) (info->dlpi_addr + info->dlpi_phdr[j].p_vaddr));
         printf("\t\t\t type=%u, flags=0x%X\n",
                 info->dlpi_phdr[j].p_type, info->dlpi_phdr[j].p_flags);
    }
    printf("\n");
    return 0;
}


extern int ml_func(int, int);


int main(int argc, const char* argv[])
{
    dl_iterate_phdr(header_handler, NULL);

    int t = ml_func(argc, argc);
    return t;
}

header_handler implements the callback for dl_iterate_phdr. It will get called for all libraries and report their names and load addresses, along with all their segments. It also invokes ml_func, which is taken from the libmlreloc.so shared library.

To compile and link this driver with our shared library, run:

gcc -g -c driver.c -o driver.o
gcc -o driver driver.o -L. -lmlreloc

Running the driver stand-alone we get the information, but for each run the addresses are different. So what I'm going to do is run it under gdb [10], see what it says, and then use gdb to further query the process's memory space:

 $ gdb -q driver
 Reading symbols from driver...done.
 (gdb) b driver.c:31
 Breakpoint 1 at 0x804869e: file driver.c, line 31.
 (gdb) r
 Starting program: driver
 [...] skipping output
 name=./libmlreloc.so (6 segments) address=0x12e000
                header  0: address=  0x12e000
                        type=1, flags=0x5
                header  1: address=  0x12ff04
                        type=1, flags=0x6
                header  2: address=  0x12ff18
                        type=2, flags=0x6
                header  3: address=  0x12e0f4
                        type=4, flags=0x4
                header  4: address=  0x12e000
                        type=1685382481, flags=0x6
                header  5: address=  0x12ff04
                        type=1685382482, flags=0x4

[...] skipping output
 Breakpoint 1, main (argc=1, argv=0xbffff3d4) at driver.c:31
 31    }
 (gdb)

Since driver reports all the libraries it loads (even implicitly, like libc or the dynamic loader itself), the output is lengthy and I will just focus on the report about libmlreloc.so. Note that the 6 segments are the same segments reported by readelf, but this time relocated into their final memory locations.

Let's do some math. The output says libmlreloc.so was placed in virtual address 0x12e000. We're interested in the second segment, which as we've seen in readelf is at ofset 0x1f04. Indeed, we see in the output it was loaded to address 0x12ff04. And since myglob is at offset 0x200c in the file, we'd expect it to now be at address 0x13000c.

So, let's ask GDB:

(gdb) p &myglob
$1 = (int *) 0x13000c

Excellent! But what about the code of ml_func which refers to myglob? Let's ask GDB again:

(gdb) set disassembly-flavor intel
(gdb) disas ml_func
Dump of assembler code for function ml_func:
   0x0012e46c <+0>:   push   ebp
   0x0012e46d <+1>:   mov    ebp,esp
   0x0012e46f <+3>:   mov    eax,ds:0x13000c
   0x0012e474 <+8>:   add    eax,DWORD PTR [ebp+0x8]
   0x0012e477 <+11>:  mov    ds:0x13000c,eax
   0x0012e47c <+16>:  mov    eax,ds:0x13000c
   0x0012e481 <+21>:  add    eax,DWORD PTR [ebp+0xc]
   0x0012e484 <+24>:  pop    ebp
   0x0012e485 <+25>:  ret
End of assembler dump.

As expected, the real address of myglob was placed in all the mov instructions referring to it, just as the relocation entries specified.

Relocating function calls

So far this article demonstrated relocation of data references - using the global variable myglob as an example. Another thing that needs to be relocated is code references - in other words, function calls. This section is a brief guide on how this gets done. The pace is much faster than in the rest of this article, since I can now assume the reader understands what relocation is all about.

Without further ado, let's get to it. I've modified the code of the shared library to be the following:

int myglob = 42;

int ml_util_func(int a)
{
    return a + 1;
}

int ml_func(int a, int b)
{
    int c = b + ml_util_func(a);
    myglob += c;
    return b + myglob;
}

ml_util_func was added and it's being used by ml_func. Here's the disassembly of ml_func in the linked shared library:

000004a7 <ml_func>:
 4a7:   55                      push   ebp
 4a8:   89 e5                   mov    ebp,esp
 4aa:   83 ec 14                sub    esp,0x14
 4ad:   8b 45 08                mov    eax,DWORD PTR [ebp+0x8]
 4b0:   89 04 24                mov    DWORD PTR [esp],eax
 4b3:   e8 fc ff ff ff          call   4b4 <ml_func+0xd>
 4b8:   03 45 0c                add    eax,DWORD PTR [ebp+0xc]
 4bb:   89 45 fc                mov    DWORD PTR [ebp-0x4],eax
 4be:   a1 00 00 00 00          mov    eax,ds:0x0
 4c3:   03 45 fc                add    eax,DWORD PTR [ebp-0x4]
 4c6:   a3 00 00 00 00          mov    ds:0x0,eax
 4cb:   a1 00 00 00 00          mov    eax,ds:0x0
 4d0:   03 45 0c                add    eax,DWORD PTR [ebp+0xc]
 4d3:   c9                      leave
 4d4:   c3                      ret

What's interesting here is the instruction at address 0x4b3 - it's the call to ml_util_func. Let's dissect it:

e8 is the opcode for call. The argument of this call is the offset relative to the next instruction. In the disassembly above, this argument is 0xfffffffc, or simply -4. So the call currently points to itself. This clearly isn't right - but let's not forget about relocation. Here's what the relocation section of the shared library looks like now:

$ readelf -r libmlreloc.so

Relocation section '.rel.dyn' at offset 0x324 contains 8 entries:
 Offset     Info    Type            Sym.Value  Sym. Name
00002008  00000008 R_386_RELATIVE
000004b4  00000502 R_386_PC32        0000049c   ml_util_func
000004bf  00000401 R_386_32          0000200c   myglob
000004c7  00000401 R_386_32          0000200c   myglob
000004cc  00000401 R_386_32          0000200c   myglob
[...] skipping stuff

If we compare it to the previous invocation of readelf -r, we'll notice a new entry added for ml_util_func. This entry points at address 0x4b4 which is the argument of the call instruction, and its type is R_386_PC32. This relocation type is more complicated than R_386_32, but not by much.

It means the following: take the value at the offset specified in the entry, add the address of the symbol to it, subtract the address of the offset itself, and place it back into the word at the offset. Recall that this relocation is done at load-time, when the final load addresses of the symbol and the relocated offset itself are already known. These final addresses participate in the computation.

What does this do? Basically, it's a relative relocation, taking its location into account and thus suitable for arguments of instructions with relative addressing (which the e8 call is). I promise it will become clearer once we get to the real numbers.

I'm now going to build the driver code and run it under GDB again, to see this relocation in action. Here's the GDB session, followed by explanations:

 $ gdb -q driver
 Reading symbols from driver...done.
 (gdb) b driver.c:31
 Breakpoint 1 at 0x804869e: file driver.c, line 31.
 (gdb) r
 Starting program: driver
 [...] skipping output
 name=./libmlreloc.so (6 segments) address=0x12e000
               header  0: address=  0x12e000
                       type=1, flags=0x5
               header  1: address=  0x12ff04
                       type=1, flags=0x6
               header  2: address=  0x12ff18
                       type=2, flags=0x6
               header  3: address=  0x12e0f4
                       type=4, flags=0x4
               header  4: address=  0x12e000
                       type=1685382481, flags=0x6
               header  5: address=  0x12ff04
                       type=1685382482, flags=0x4

[...] skipping output
Breakpoint 1, main (argc=1, argv=0xbffff3d4) at driver.c:31
31    }
(gdb)  set disassembly-flavor intel
(gdb) disas ml_util_func
Dump of assembler code for function ml_util_func:
   0x0012e49c <+0>:   push   ebp
   0x0012e49d <+1>:   mov    ebp,esp
   0x0012e49f <+3>:   mov    eax,DWORD PTR [ebp+0x8]
   0x0012e4a2 <+6>:   add    eax,0x1
   0x0012e4a5 <+9>:   pop    ebp
   0x0012e4a6 <+10>:  ret
End of assembler dump.
(gdb) disas /r ml_func
Dump of assembler code for function ml_func:
   0x0012e4a7 <+0>:    55     push   ebp
   0x0012e4a8 <+1>:    89 e5  mov    ebp,esp
   0x0012e4aa <+3>:    83 ec 14       sub    esp,0x14
   0x0012e4ad <+6>:    8b 45 08       mov    eax,DWORD PTR [ebp+0x8]
   0x0012e4b0 <+9>:    89 04 24       mov    DWORD PTR [esp],eax
   0x0012e4b3 <+12>:   e8 e4 ff ff ff call   0x12e49c <ml_util_func>
   0x0012e4b8 <+17>:   03 45 0c       add    eax,DWORD PTR [ebp+0xc]
   0x0012e4bb <+20>:   89 45 fc       mov    DWORD PTR [ebp-0x4],eax
   0x0012e4be <+23>:   a1 0c 00 13 00 mov    eax,ds:0x13000c
   0x0012e4c3 <+28>:   03 45 fc       add    eax,DWORD PTR [ebp-0x4]
   0x0012e4c6 <+31>:   a3 0c 00 13 00 mov    ds:0x13000c,eax
   0x0012e4cb <+36>:   a1 0c 00 13 00 mov    eax,ds:0x13000c
   0x0012e4d0 <+41>:   03 45 0c       add    eax,DWORD PTR [ebp+0xc]
   0x0012e4d3 <+44>:   c9     leave
   0x0012e4d4 <+45>:   c3     ret
End of assembler dump.
(gdb)

The important parts here are:

1. In the printout from driver we see that the first segment (the code segment) of libmlreloc.so has been mapped to 0x12e000 [11]
2. ml_util_func was loaded to address 0x0012e49c
3. The address of the relocated offset is 0x0012e4b4
4. The call in ml_func to ml_util_func was patched to place 0xffffffe4 in the argument (I disassembled ml_func with the /r flag to show raw hex in addition to disassembly), which is interpreted as the correct offset to ml_util_func.

Obviously we're most interested in how (4) was done. Again, it's time for some math. Interpreting the R_386_PC32 relocation entry mentioned above, we have:

Take the value at the offset specified in the entry (0xfffffffc), add the address of the symbol to it (0x0012e49c), subtract the address of the offset itself (0x0012e4b4), and place it back into the word at the offset. Everything is done assuming 32-bit 2-s complement, of course. The result is 0xffffffe4, as expected.

Extra credit: Why was the call relocation needed?

This is a "bonus" section that discusses some peculiarities of the implementation of shared library loading in Linux. If all you wanted was to understand how relocations are done, you can safely skip it.

When trying to understand the call relocation of ml_util_func, I must admit I scratched my head for some time. Recall that the argument of call is a relative offset. Surely the offset between the call and ml_util_func itself doesn't change when the library is loaded - they both are in the code segment which gets moved as one whole chunk. So why is the relocation needed at all?

Here's a small experiment to try: go back to the code of the shared library, add static to the declaration of ml_util_func. Re-compile and look at the output of readelf -r again.

Done? Anyway, I will reveal the outcome - the relocation is gone! Examine the disassembly of ml_func - there's now a correct offset placed as the argument of call - no relocation required. What's going on?

When tying global symbol references to their actual definitions, the dynamic loader has some rules about the order in which shared libraries are searched. The user can also influence this order by setting the LD_PRELOAD environment variable.

There are too many details to cover here, so if you're really interested you'll have to take a look at the ELF standard, the dynamic loader man page and do some Googling. In short, however, when ml_util_func is global, it may be overridden in the executable or another shared library, so when linking our shared library, the linker can't just assume the offset is known and hard-code it [12]. It makes all references to global symbols relocatable in order to allow the dynamic loader to decide how to resolve them. This is why declaring the function static makes a difference - since it's no longer global or exported, the linker can hard-code its offset in the code.

Extra credit #2: Referencing shared library data from the executable

Again, this is a bonus section that discusses an advanced topic. It can be skipped safely if you're tired of this stuff.

In the example above, myglob was only used internally in the shared library. What happens if we reference it from the program (driver.c)? After all, myglob is a global variable and thus visible externally.

Let's modify driver.c to the following (note I've removed the segment iteration code):

#include <stdio.h>

extern int ml_func(int, int);
extern int myglob;

int main(int argc, const char* argv[])
{
    printf("addr myglob = %p\n", (void*)&myglob);
    int t = ml_func(argc, argc);
    return t;
}

It now prints the address of myglob. The output is:

addr myglob = 0x804a018

Wait, something doesn't compute here. Isn't myglob in the shared library's address space? 0x804xxxx looks like the program's address space. What's going on?

Recall that the program/executable is not relocatable, and thus its data addresses have to bound at link time. Therefore, the linker has to create a copy of the variable in the program's address space, and the dynamic loader will use that as the relocation address. This is similar to the discussion in the previous section - in a sense, myglob in the main program overrides the one in the shared library, and according to the global symbol lookup rules, it's being used instead. If we examine ml_func in GDB, we'll see the correct reference made to myglob:

0x0012e48e <+23>:      a1 18 a0 04 08 mov    eax,ds:0x804a018

This makes sense because a R_386_32 relocation for myglob still exists in libmlreloc.so, and the dynamic loader makes it point to the correct place where myglob now lives.

This is all great, but something is missing. myglob is initialized in the shared library (to 42) - how does this initialization value get to the address space of the program? It turns out there's a special relocation entry that the linker builds into the program (so far we've only been examining relocation entries in the shared library):

$ readelf -r driver

Relocation section '.rel.dyn' at offset 0x3c0 contains 2 entries:
 Offset     Info    Type            Sym.Value  Sym. Name
08049ff0  00000206 R_386_GLOB_DAT    00000000   __gmon_start__
0804a018  00000605 R_386_COPY        0804a018   myglob
[...] skipping stuff

Note the R_386_COPY relocation for myglob. It simply means: copy the value from the symbol's address into this offset. The dynamic loader performs this when it loads the shared library. How does it know how much to copy? The symbol table section contains the size of each symbol; for example the size for myglob in the .symtab section of libmlreloc.so is 4.

I think this is a pretty cool example that shows how the process of executable linking and loading is orchestrated together. The linker puts special instructions in the output for the dynamic loader to consume and execute.

Conclusion

Load-time relocation is one of the methods used in Linux (and other OSes) to resolve internal data and code references in shared libraries when loading them into memory. These days, position independent code (PIC) is a more popular approach, and some modern systems (such as x86-64) no longer support load-time relocation.

Still, I decided to write an article on load-time relocation for two reasons. First, load-time relocation has a couple of advantages over PIC on some systems, especially in terms of performance. Second, load-time relocation is IMHO simpler to understand without prior knowledge, which will make PIC easier to explain in the future. (Update 03.11.2011: the article about PIC was published)

Regardless of the motivation, I hope this article has helped to shed some light on the magic going behind the scenes of linking and loading shared libraries in a modern OS.