Julia Evans

Day 40: Linkers are amazing.

I have a linker bug in my kernel which is kind of infuriating me. So I’ve spent the whole day so far reading the first 11 parts of this excellent 20-part series about how linkers work by Ian Lance Taylor.

Here are some notes on the series. This is gigantic and basically just for my own reference, but summarizes what he talks about in the first half of the series.

I’m not going to describe the basics of what a linker does because I know already. I talk a little about linkers in How to call Rust from assembly, and I found this Beginner’s guide to linkers pretty helpful. Parts [1] and [2] of the essay also discuss “what’s a linker?”.

For context: Right now I have a bug while statically linking a single-threaded ELF file. So I will read about dynamic linking and threading and things other than ELF, but I will mostly ignore it.

1. Basic linker data types: symbols, relocations, and contents

There are symbols, relocations and contents in object files. I knew about symbols and contents already, but relocations are new to me. The contents are the contents of a variable or function. A key insight here is that the linker doesn’t actually care too much about the contents – it’s just concerned with putting the contents in the right place.

Linkers don’t care about your functions and variables! They’re just bytes. :) In the first article he talks about linker speed – a benchmark you could compare a linker to is cat, because it’s just combining the bytes together, and making some replacements along the way.

Relocations were new to me. The story here is that when you have an object file, the assembly code often refer to a symbol like this:

call awesome_function

awesome_function might be undefined – it could be a function defined in a library that we’re planning to link against later.

So after linking, awesome_function will be somewhere new, and we need to put a memory address in call awesome_function! So there’s a relocation table which keeps track of everything that needs to be moved.

To make this a bit more concrete, I looked at man objdump. Turns out you can look at the relocations in an object file by running objdump -r file.o. When you look at the relocations in an ELF file, there are a bunch of scary-looking things like R_386_PC32 and R_386_GOTPC. Here’s a page explaining what those things mean. For example, I got this as an output from objdump -r

00001a76 R_386_GOTPC       _GLOBAL_OFFSET_TABLE_
00001a7f R_386_GOTOFF      _ZN3mem4base18he097c5c5c82e35fah4v0.0E
00001a95 R_386_GOTOFF      _ZN3mem4base18he097c5c5c82e35fah4v0.0E
00001ac4 R_386_PC32        __morestack

I think the last line of that means “At 00001ac4, there’s a reference to __morestack. I’m going to need you to figure out the distance from 00001ac4 to __morestack and add it to the dword at 00001ac4.

Basically R_386_PC32 and friends are different rules that the linker has to follow. Some possible things that relocation rules might do:

  • Put a relative memory address somewhere
  • Put an absolute memory address somewhere
  • Add something to the Global Offset Table (GOT)
  • Add something to the Procedure Lookup Table (PLT)

I don’t know what the GOT/PLT are yet but hopefully as I keep reading I’ll learn. (I did! See part 4)

From Part 2.

2. Object file formats

Apparently there are many different kinds of object file formats (COFF, ELF, PE, a.out, IEEE-695, Mach-O, etc.). They can be either section-based or record-based. I only care about ELF. ELF is section-based. This means that the file is split up into sections.

You can use readelf --sections myfile.o to list the sections in an ELF file, and objdump -t to list the symbol table and which section each symbol belongs to.

Here are the sections and the symbol table for an OCaml object file I found on my machine. You can see that the sections that have most of the symbols in them are .data, .text, and .bss. And there’s a section called .symtab, which I guess is the symbol table! Neat.

.text is the “code” of the program, and .data, .bss, and .rodata contain different kinds of globals.

3. Debugging symbols

It says

The ELF object file format stores debugging information in sections with special names.

It mentions debugging formats like “stabs strings” and “the DWARF debugging format”, but not going down that rabbit hole right now.

File formats and debugging symbols were in Part 3.

4. Shared libraries and position independence

A shared library is a .dll on Windows or a .so file on Linux.

The deal with a shared library is when you create the library, you don’t know what address in memory it’s going to be loaded at. (it depends on what the dynamic linker decides to do). So some calculation (addition!) needs to happen no matter what.

So there are two different strategies to deal with this.

Strategy 1: Make it position dependent. This basically means “write a huge relocation table and let the dynamic linker figure out where everything should go”. Because there is a huge relocation table, this means the library will take longer to load.

Strategy 2: Make it position independent. This means “don’t write a relocation table, but every time you call a function or reference a variable, look up where it should be in a table”. So the library loads more quickly, but runs more slowly. Tradeoffs!

This brings us back to the “Global Offset Table” and “Procedure Lookup Table” from before! So in position independent code, the “Global Offset Table” is where look up our global and static variables, and the “Procedure Lookup Table” is where we look up functions.

The other super important thing here (discussed more in Part 6) is that in position independent code, the assembly instructions always the same, and only the PLT and the GOT need be different. This means that if you have a huge library two different processes can reference the same (read-only) assembly instructions and they only need to have their own copy of the PLT and GOT.

This is super neat! In my object file, I have

00001a76 R_386_GOTPC       _GLOBAL_OFFSET_TABLE_

I guess that means the object file I am looking at is position independent!

Another neat thing about position independent code is that the dynamic linker can do lazy loading – it can wait to put the address of a function in the PLT until it is actually loaded. So if you run a program that links against all of the math library, but only calls sin, it only needs to figure out the address of sin.

All this about shared libraries is in Part 4.

5. What can go wrong with shared libraries

There’s a super interesting discussion of a kind of bug you can have with shared libraries in Part 5.

In C, you can take the address of a function (&f). The natural address for f is its entry in the PLT, because that’s where you jump to when the function is called. But different shared libraries have different PLTs, so you might end up with two different addresses for the function, and if you compared them you would get different answers. Oh no! I did not know that this was even a thing. Cool!

You can go read about the solution in the essay.

All this stuff about dynamic linking is really interesting and complicated, but I’m actually reading this in an effort to solve a static linking problem that I have, so I’m not reading it with 100% attention.

6. ELF symbol visibility and types

I didn’t know that symbols had visibility and types! I thought everything was global and pretty much the same type. I don’t care too much about this right now, though.

Also from Part 5.

7. Linkers are architecture-dependent

How a linker deals with relocations depends on the architecture of the machine the binaries are for! When you’re writing an OS, you need a cross-compiler that targets your target architecture. But you also need a cross-linker! Here’s why.

C code: g = 0.

Corresponding 386 code: movl 0 g.

Corresponding RISC code:

li 1,0 // Set register 1 to 0
lis 9,[email protected] // Load high-adjusted part of g into register 9
stw 1,[email protected](9) // Store register 1 to address in register 9 plus low adjusted part g

Here g is referenced twice! So the linker is going to need to know about how RISC works and the relocation is going to look different than it would on 386.

This was Part 6.

8. Thread Local Storage

Part 7 is an extremely cool explanation of how ELF systems have special support for making threading more efficient. I had no idea this was a thing the linker did. Linkers are crazy.

The idea here is that you can just mark a variable thread-local in C like this:

__thread int i;

and the compiler and linker will do a bunch of stuff so that every thread has its own copy of the variable.

The alternative here (I think) is to use pthread_key_create, pthread_get_specific and pthread_set_specific. I’ve never used pthreads, but this is cool and definitely seems easier. There’s a discussion of when it is more efficient in the essay. (tl;dr: never slower, sometimes faster)

9. ELF segments

More on “oh man ELF is complicated!” Part 8 has a description of all the section and segment types.

Okay, so what is a segment, anyway? We talked about sections before – we said that they were .text, .data, .rodata, etc. – to separate out read-only data from read-write data from program code. Segments are collections of sections.

ELF is an format for object files, and also executables. Apparently the operating system doesn’t use the sections at all to run programs, just the segments.

I’m a bit confused here, so let’s make this more concrete by looking at a simple “Hello World” C program. I couldn’t figure out how to use objdump to look at segments, but my awesome friend Dave suggested using readelf instead. That totally works!

Here’s the the entire output of readelf --segments a.out. Here’s an excerpt:

  Segment Sections...
   01     .interp 
   02     .interp .note.ABI-tag .note.gnu.build-id .gnu.hash .dynsym .dynstr .gnu.version .gnu.version_r .rela.dyn .rela.plt .init .plt .text .fini .rodata .eh_frame_hdr .eh_frame 
   03     .ctors .dtors .jcr .dynamic .got .got.plt .data .bss 
   04     .dynamic 
   05     .note.ABI-tag .note.gnu.build-id 
   06     .eh_frame_hdr 
   08     .ctors .dtors .jcr .dynamic .got 

So here it looks like .text is in a segment with a bunch of stuff, and .data and .bss are in a segment together. There are a bunch of possible segment types. One of the most important ones seems to be LOAD. I think that means “load into memory”. The two segments marked LOAD are segment 02 and 03. This makes sense, because those are the segments with .text and .data in them!

The other segment types in the output of readelf are:

  • INTERP: Which dynamic loader to use (/lib64/ld-linux-x86-64.so.2). I checked and this is an actual file. <3.
  • NOTE: I guess this is a note.
  • GNU_EH_FRAME, GNU_STACK, GNU_RELRO: Some GNU extensions. I don’t really know.
  • DYNAMIC: Some stuff that the dynamic linker needs.
  • PHDR: I quote: “This indicates the address and size of the segment table. This is not too useful in practice as you have to have already found the segment table before you can find this segment.” lulz. What.

The last thing I want to note about this example is that section 02 (the one with .text and .rodata in it) has flags RE, so it can be executed. I’m wondering if you could trick the program into executing something from .rodata, and if that could be bad. Section 03 has flags RW, so it can be written to but not executed.

So the reason that .text and .data aren’t in the same section is that .text has to be read-only and .data needs to be read-write.

Here’s another example of the segments and the symbol table in the same program, but this time statically linked. There are way less segments (the file is less complicated), but the symbol table is much bigger.

Okay I think I’m pretty good with segments. This is cool!

9. Symbol versions

Apparently in an ELF file, you can have two different versions of the same symbol, in case the function signature has changed! This is crazy. I thought that if you had sin, there was only one sin.

Could you abuse this to have polymorphism in object files without doing name mangling? Huh.

The actual use case described for this is providing two versions for stat: LIBC_1.0 and LIBC_2.0 after it changed to support 64-bit file offsets (whatever that means).

From Part 9.

10. Parallel linking

You can do linking in parallel to some extent. Also, once the output file is laid out and its size is determined, you can use mmap and do the I/O in parallel that way. Yay mmap!

From Part 10.

11. Archives

Apparently you can package a whole bunch of object files together into an archive. This is what files with the .a extension are. Neat! I’ve seen those, but didn’t know what they were. They’re created with the ar utility.

I looked for archive files on my computer using locate .a | egrep '.a$' | head, and I found one in /etc/alternatives/libblas.a. It kind of makes sense that BLAS is made of a whole bunch of object files.

I used objdump -t /etc/alternatives/libblas.a to look at is symbol table. In the output, it lists the symbol table for each object file. Here’s just the list of object files. libblas.a is made up of 313 object files.

From Part 11.