The architecture of your embedded system will likely look a lot like this:
+-----------------------------------+
| user-space code |
+-----------------------------------+
| kernel-space code (drivers, etc.) |
+-----------------------------------+
| device tree |
+-----------------------------------+
| boot loader |
+-----------------------------------+
Your user-space program(s) will communicate with kernel-space drivers to access hardware devices, manage scheduling, and so on. The kernel will get the details of which devices are installed in the system and where they are from the device tree and will then pass that information on to the various device drivers.
The boot loader is in charge of preparing the system, enabling clocks, powering up peripherals, and so on. It will then load the Linux kernel and execute it with some set of arguments that you can choose.
In the boot loader, you will initialize the very basics of the processor. This includes powering it on, enabling any peripherals required for boot (such as a serial console for boot loader/kernel output, ethernet for network boot, or writing identifying information to an EEPROM), and then actually booting the kernel from the network, an MMC device, or a some other type of storage device.
For many systems, little will need to be done in the boot loader. For the most part, cloning the u-boot sources, doing a little bit of configuring and patching, and then building it is enough to have a working bootloader. YMMV.
The device tree is arguably the most important (and worst-documented) part of the entire process. Some vendors have or had proprietary file formats for configuring hardware devices for the kernel. Many opted to maintain a fork of the kernel for only marginally-different ARM devices, which resulted in a lot of fragmented information and device support. As a result, Linux has settled on using the device tree format for different boards. In order to make your system work, you will create a device tree for your board, likely based on a reference one for an EVM or developer kit for your chosen processor.
The device tree will tell the kernel which clocks are where, what registers are
used for which devices, what kinds of devices you have, and which drivers should
be used with them. For example, say you have an I2C bus with two slave devices,
mydeviceA and mydeviceB. They will be supported by mydriverX and
mydriverY. Their address are 0x20 and 0x21. You might have something like
this in your device tree:
/* i2c0 will have been setup elsewhere, we will just add devices to it */
&i2c0 {
mydeviceA@20 {
compatible = "mydriverX";
reg = <0x20>;
/* each device may have driver-specific properties or other common i2c
slave properties */
};
mydeviceB@21 {
compatible = "mydriverY";
reg = <0x21>;
};
};
Linux will understand that it should load the two drivers and tell them about
the devices at addresses 0x20 and 0x21 on the bus referred to as i2c0,
which was setup elsewhere. This way, you will be able to simply talk to the
device driver at a high level while the critical timing and IO happens in
kernel-space without context-switching. This will be more efficient than if you
use the user-space I2C driver to communicate with the device directly, and it's
also much easier to maintain your userspace code.
If possible, you should write a kernel device driver for your custom devices (or those that don't already have drivers in the kernel source tree). If you can't or don't want to, then you can just fall back to using user-space I/O drivers for I2C, SPI, GPIO, etc.
Kernel drivers will expose a device somewhere in the sysfs. Read the documentation for your device's driver (usually included with the kernel source, but some companies did not feel the need to write any documentation--looking at you, MAX31790) to learn what the driver exposes in the sysfs for your devices. These things will appear as files, which you can open and read and write, but they will only accept certain inputs and will give you certain outputs. Some files are made for polling rather than reading/writing, and so on. It all depends on the device.
Your user-space program is where you will write the main control logic for your program unless you're a masochist and prefer to write your entire program in the kernel. This doesn't work very well for a lot of things and you shouldn't do it. Your program will be run probably as root (or by another user with sufficient permissions to access the required devices) and will communicate with the device drivers in the kernel to control the system and acquire input. This is the easy part; it's just like any other program, though you may need to deal with real-time scheduling and such.