From Zero to main(): Demystifying Firmware Linker Scripts

This is the second post in our Zero to main() series.

Last time, we talked about bootstrapping a C environment on an MCU before invoking our main function. One thing we took for granted was the fact that functions and data end up in the right place in our binary. Today, we’re going to dig into how that happens by learning about memory regions and linker scripts.

You may remember the following things happening auto-magically:

  1. We used variables like &_ebss, &_sdata, …etc. to know where each of our sections was placed in flash and to define where some needed to go in RAM.
  2. A pointer to our ResetHandler was found at address 0x00000004 for the MCU to find.

While these things are true for many projects, they are held together at best by convention, at worst by generations of copy/paste engineering. You’ll find some MCUs have different memory maps, some startup scripts name those variables differently, and some programs have more or less segments.

Since they are not standardized, those things need to be specified somewhere in our project. In the case of projects linked with a Unix-ld-like tool, that somewhere is the linker script.

Once again, we will use our simple “minimal” program, available on Github.

Brief Primer on Linking

Linking is the last stage in compiling a program. It takes a number of compiled object files and merges them into a single program, filling in addresses so that everything is in the right place.

Prior to linking, the compiler will have taken your source files one by one and compiled them into machine code. In the process, it leaves placeholders for addresses as (1) it does not know where the code will end up within the broader structure of the program and (2) it knows nothing about symbols outside of the current file or compilation unit.

The linker takes all of those object files and merges them together along with external dependencies like the C Standard Library into your program. To figure out which bits go where, the linker relies on a linker script - a blueprint for your program. Lastly, all placeholders are replaced by addresses.

We can see this at play in our minimal program. Let’s follow what happens to our main function in minimal.c for example. The compiler builds it into an object file with:

$ arm-none-eabi-gcc -c -o build/objs/a/b/c/minimal.o minimal.c <CFLAGS>

We can dump symbols in minimal.o to look at main within it:

$ arm-none-eabi-nm build/objs/a/b/c/minimal.o
...
00000000 T main
...

As expected, it does not have addresses yet. We then link everything with:

$ arm-none-eabi-gcc <LDFLAGS> build/objs/a/b/c/minimal.o <other object files> -o build/minimal.elf

And dump the symbols in the resulting elf file:

$ arm-none-eabi-nm build/minimal.elf
...
00000294 T main
...

The linker has done its job, and our main symbol has been assigned an address.

The linker often does a bit more than that. For example, it can generate debug information, garbage collect unused sections of code, or run whole-program optimization (also known as Link-Time Optimization, or LTO). For the sake of this conversation, we will not cover these topics.

For more information on the linker, there’s a great thread on Stack Overflow.

Anatomy of a Linker Script

A linker script contains four things:

  • Memory layout: what memory is available where
  • Section definitions: what part of a program should go where
  • Options: commands to specify architecture, entry point, …etc. if needed
  • Symbols: variables to inject into the program at link time

Memory Layout

In order to allocate program space, the linker needs to know how much memory is available, and at what addresses that memory exists. This is what the MEMORY definition in the linker script is for.

The syntax for MEMORY is defined in the binutils docs and is as follow:

MEMORY
  {
    name [(attr)] : ORIGIN = origin, LENGTH = len
    …
  }

Where

  • name is a name you want to use for this region. Names do not carry meaning, so you’re free to use anything you want. You’ll often find “flash”, and “ram” as region names.
  • (attr) are optional attributes for the region, like whether it’s writable (w), readable (r), or executable (x). Flash memory is usually (rx), while ram is rwx. Marking a region as non-writable does not magically make it write protected: these attributes are meant to describe the properties of the memory, not set it.
  • origin is the start address of the memory region.
  • len is the size of the memory region, in bytes.

The memory map for the SAMD21G18 chip we’ve got on our board can be found in its datasheet in table 10-1, reproduced below.

SAMD21G18 Memory Map
Memory Start Address Size
Internal Flash 0x00000000 256 Kbytes
Internal SRAM 0x20000000 32 Kbytes


Transcribed into a MEMORY definition, this gives us:

MEMORY
{
  rom      (rx)  : ORIGIN = 0x00000000, LENGTH = 0x00040000
  ram      (rwx) : ORIGIN = 0x20000000, LENGTH = 0x00008000
}

Section Definitions

Code and data are bucketed into sections, which are contiguous areas of memory. There are no hard rules about how many sections you should have, or what they should be, but you typically want to put symbols in the same section if:

  1. They should be in the same region of memory, or
  2. They need to be initialized together.

In our previous post, we learned about two types of symbols that are initialized in bulk:

  1. Initialized static variables which must be copied from flash
  2. Uninitialized static variables which must be zeroed.

Our linker script concerns itself with two more things:

  1. Code and constant data, which can live in read-only memory (e.g. flash)
  2. Reserved sections of RAM, like a stack or a heap

By convention, we name those sections as follow:

  1. .text for code & constants
  2. .bss for uninitialized data
  3. .stack for our stack
  4. .data for initialized data

The elf spec holds a full list. Your firmware will work just fine if you call them anything else, but your colleagues may be confused and some tools may fail in odd ways. The only constraint is that you may not call your section /DISCARD/, which is a reserved keyword.

First, let’s look at what happens to our symbols if we do not define any of those sections in the linker script.

MEMORY
{
  rom      (rx)  : ORIGIN = 0x00000000, LENGTH = 0x00040000
  ram      (rwx) : ORIGIN = 0x20000000, LENGTH = 0x00008000
}

SECTIONS
{
    /* empty! */
}

The linker is perfectly happy to link our program with this. Probing the resulting elf file with objdump, we see the following:

$ arm-none-eabi-objdump -h build/minimal.elf
build/minimal.elf:     file format elf32-littlearm

SYMBOL TABLE:
no symbols

No symbols! While the linker is able to make assumptions that will allow it to link in symbols with little information, but it at least needs to know either what the entry point should be, or what symbols to put in the text section.

.text Section

Let’s start by adding our .text section. We want that section in ROM. The syntax is simple:

SECTIONS
{
    .text :
    {

    } > rom
}

This defines a section named .text, and adds it to the ROM. We now need to tell the linker what to put in that section. This is accomplished by listing all of the sections from our input object files we want in .text.

To find out what sections are in our object file, we can once again use objdump:

$ arm-none-eabi-objdump -h
build/objs/a/b/c/minimal.o:     file format elf32-littlearm

Sections:
Idx Name          Size      VMA       LMA       File off  Algn
  0 .text         00000000  00000000  00000000  00000034  2**1
                  CONTENTS, ALLOC, LOAD, READONLY, CODE
  1 .data         00000000  00000000  00000000  00000034  2**0
                  CONTENTS, ALLOC, LOAD, DATA
  2 .bss          00000000  00000000  00000000  00000034  2**0
                  ALLOC
  3 .bss.cpu_irq_critical_section_counter 00000004  00000000  00000000  00000034
2**2
                  ALLOC
  4 .bss.cpu_irq_prev_interrupt_state 00000001  00000000  00000000  00000034
2**0
                  ALLOC
  5 .text.system_pinmux_get_group_from_gpio_pin 0000005c  00000000  00000000
00000034  2**2
                  CONTENTS, ALLOC, LOAD, READONLY, CODE
  6 .text.port_get_group_from_gpio_pin 00000020  00000000  00000000  00000090
2**1
                  CONTENTS, ALLOC, LOAD, RELOC, READONLY, CODE
  7 .text.port_get_config_defaults 00000022  00000000  00000000  000000b0  2**1
                  CONTENTS, ALLOC, LOAD, READONLY, CODE
  8 .text.port_pin_set_output_level 0000004e  00000000  00000000  000000d2  2**1
                  CONTENTS, ALLOC, LOAD, RELOC, READONLY, CODE
  9 .text.port_pin_toggle_output_level 00000038  00000000  00000000  00000120
2**1
                  CONTENTS, ALLOC, LOAD, RELOC, READONLY, CODE
 10 .text.set_output 00000040  00000000  00000000  00000158  2**1
                  CONTENTS, ALLOC, LOAD, RELOC, READONLY, CODE
 11 .text.main    0000002c  00000000  00000000  00000198  2**2
                  CONTENTS, ALLOC, LOAD, RELOC, READONLY, CODE

We see that each of our symbol has a section. This is due to the fact that we compiled our firmware with the -ffunction-sections and -fdata-sections flags. Had we not included them, the compiler would have been free to merge several functions into a single text.<some identifier> section.

To put all of our functions in the .text section in our linker script, we use the following syntax: <filename>(<section>), where filename is the name of the input files whose symbols we want to include, and section is the name of the input sections. Since we want all .text... sections in all files, we use the wildcard *:

.text :
{
    KEEP(*(.vector*))
    *(.text*)
} > rom

Note the .vector input section, which contains functions we want to keep at the very start of our .text section. This is so the Reset_Handler is where the MCU expects it to be. We’ll talk more about the vector table in a future post.

Dumping our elf file, we now see all of our functions (but no data)!

$ arm-none-eabi-objdump -t build/minimal.elf

build/minimal.elf:     file format elf32-littlearm

SYMBOL TABLE:
00000000 l    d  .text  00000000 .text
...
00000000 l    df *ABS*  00000000 minimal.c
00000000 l     F .text  0000005c system_pinmux_get_group_from_gpio_pin
0000005c l     F .text  00000020 port_get_group_from_gpio_pin
0000007c l     F .text  00000022 port_get_config_defaults
0000009e l     F .text  0000004e port_pin_set_output_level
000000ec l     F .text  00000038 port_pin_toggle_output_level
00000124 l     F .text  00000040 set_output
00000000 l    df *ABS*  00000000 port.c
00000190 l     F .text  00000028 system_pinmux_get_config_defaults
00000000 l    df *ABS*  00000000 pinmux.c
00000208 l     F .text  0000005c system_pinmux_get_group_from_gpio_pin
00000264 l     F .text  00000110 _system_pinmux_config
00000164 g     F .text  0000002c main
000001b8 g     F .text  0000004e port_pin_set_config
00000374 g     F .text  00000040 system_pinmux_pin_set_config
...

.bss Section

Now, let’s take care of our .bss. Remember, this is the section we put uninitialized static memory in. .bss must be reserved in the memory map, but there is nothing to load, as all variables are initialized to zero. As such, this is what it should look like:

SECTION {
    ...
    .bss (NOLOAD) :
    {
        *(.bss*)
        *(COMMON)
    } > ram
}

You’ll note that the .bss section also includes *(COMMON). This is a special input section where the compiler puts global uninitialized variables that go beyond file scope. int foo; goes there, while static int foo; does not. This allows the linker to merge multiple definitions into one symbol if they have the same name.

We indicate that this section is not loaded with the NOLOAD property. This is the only section property used in modern linker scripts.

.stack Section

We do the same thing for our .stack memory, since it is in RAM and not loaded. As the stack contains no symbols, we must explicitly reserve space for it by indicating its size. We also must align the stack on an 8-byte boundary per ARM Procedure Call Standards (AAPCS).

In order to achieve these goals, we turn to a special variable ., also known as the “location counter”. The location counter tracks the current offset into a given memory region. As sections are added, the location counter increments accordingly. You can force alignment or gaps by setting the location counter forward. You may not set it backwards, and the linker will throw an error if you try.

We set the location counter with the ALIGN function, to align the section, and use simple assignment and arithmetic to set the section size:

STACK_SIZE = 0x2000; /* 8 kB */

SECTION {
    ...
    .stack (NOLOAD) :
    {
        . = ALIGN(8);
        . = . + STACK_SIZE;
        . = ALIGN(8);
    } > ram
    ...
}

Only one more section to go!

.data Section

The .data section contains static variables which have an initial value at boot. You will remember from our previous article that since RAM isn’t persisted while power is off, those sections need to be loaded from flash. At boot, the Reset_Handler copies the data from flash to RAM before the main function is called.

To make this possible, every section in our linker script has two addresses, its load address (LMA) and its virtual address (VMA). In a firmware context, the LMA is where your JTAG loader needs to place the section and the VMA is where the section is found during execution.

You can think of the LMA as the address “at rest” and the VMA the address during execution i.e. when the device is on and the program is running.

The syntax to specify the LMA and VMA is relatively straightforward: every address is two part: AT . In our case it looks like this:

.data :
{
    *(.data*);
} > ram AT > rom  /* "> ram" is the VMA, "> rom" is the LMA */

Note that instead of appending a section to a memory region, you could also explicitly specify an address like so:

.data ORIGIN(ram) /* VMA */ : AT(ORIGIN(rom)) /* LMA */
{
    . = ALIGN(4);
    _sdata = .;
    *(.data*);
    . = ALIGN(4);
    _edata = .;
}

Where ORIGIN(<region>) is a simple way to specify the start of a region. You can enter an address in hex as well.

And we’re done! Here’s our complete linker script with every section:

Complete Linker Script

MEMORY
{
  rom      (rx)  : ORIGIN = 0x00000000, LENGTH = 0x00040000
  ram      (rwx) : ORIGIN = 0x20000000, LENGTH = 0x00008000
}

STACK_SIZE = 0x2000;

/* Section Definitions */
SECTIONS
{
    .text :
    {
        KEEP(*(.vectors .vectors.*))
        *(.text*)
        *(.rodata*)
    } > rom

    /* .bss section which is used for uninitialized data */
    .bss (NOLOAD) :
    {
        *(.bss*)
        *(COMMON)
    } > ram

    .data :
    {
        *(.data*);
    } > ram AT >rom

    /* stack section */
    .stack (NOLOAD):
    {
        . = ALIGN(8);
        . = . + STACK_SIZE;
        . = ALIGN(8);
    } > ram

    _end = . ;
}

You can find the full details on linker script sections syntax in the ld manual.

Variables

In the first post, our ResetHandler relied on seemingly magic variables to know the address of each of our sections of memory. It turns out, those variable came

In order to make section addresses available to code, the linker is able to generate symbols and add them to the program.

You can find the syntax in the linker documentation, it looks exactly like a C assignment: symbol = expression;

Here, we need:

  1. _etext the end of the code in .text section in flash.
  2. _sdata the start of the .data section in RAM
  3. _edata the end of the .data section in RAM
  4. _sbss the start of the .bss section in RAM
  5. _ebss the end of the .bss section in RAM

They are all relatively straightforward: we can assign our symbols to the value of the location counter (.) at the start and at the end of each section definition.

The code is below:

    .text :
    {
        KEEP(*(.vectors .vectors.*))
        *(.text.*)
        *(.rodata.*)
        _etext = .;
    } > rom

    .bss (NOLOAD) :
    {
        _sbss = . ;
        *(.bss .bss.*)
        *(COMMON)
        _ebss = . ;
    } > ram

    .data :
    {
        _sdata = .;
        *(.data*);
        _edata = .;
    } > ram AT >rom

One quirk of these linker-provided symbols: you must use a reference to them, never the variable themselves. For example, the following gets us a pointer to the start of the .data section:

uint8_t *data_byte = &_sdata;

You can read more details about this in the binutils docs.

Closing

I hope this post gave you confidence in writing your own linker scripts.

In my next post, we’ll talk about writing a bootloader to assist with loading and starting your application.

EDIT: Post written! - Writing a Bootloader from Scratch

As with previous posts, code examples are available on Github in the zero to main repository

See anything you'd like to change? Submit a pull request or open an issue on our GitHub

François Baldassari has worked on the embedded software teams at Sun, Pebble, and Oculus. He is currently the CEO of Memfault.