1.19. Barebox Bootchooser

In many cases embedded systems are layed out redundantly with multiple kernels and multiple root file systems. The bootchooser framework provides the building blocks to model different use cases without the need to start from scratch over and over again.

The bootchooser works on abstract boot targets, each with a set of properties and implements an algorithm which selects the highest priority target to boot.

1.19.1. Bootchooser Targets

A bootchooser boot target represents one target that barebox can boot. It consists of a set of variables in the global.bootchooser.<targetname> namespace. The following configuration variables are needed to describe a bootchooser boot target:

global.bootchooser.<targetname>.boot
This controls what barebox actually boots for this boot target. This string can contain anything that the boot command understands.
global.bootchooser.<targetname>.default_attempts
The default number of attempts that a boot target shall be tried before skipping it.
global.bootchooser.<targetname>.default_priority
The default priority of a boot target.

Additionally the following run-time variables are needed. Unlike the configuration variables their values are automatically updated by the bootchooser algorithm:

global.bootchooser.<targetname>.priority
The current priority of the boot target. Higher numbers have higher priorities. A priority of 0 means the boot target is disabled and won’t be started.
global.bootchooser.<targetname>.remaining_attempts
The remaining_attempts counter. Only boot targets with a remaining_attempts counter > 0 are started.

The bootchooser algorithm generally only starts boot targets that have a priority > 0 and a remaining_attempts counter > 0.

1.19.2. The Bootchooser Algorithm

The bootchooser algorithm is very simple. It works with two variables per boot target and some optional flags. The variables are the remaining_attempts counter that tells how many times the boot target will be started. The other variable is the priority, the boot target with the highest priority will be used first, a zero priority means the boot target is disabled.

When booting, bootchooser starts the boot target with the highest priority that has a non-zero remaining_attempts counter. With every start of a boot target the remaining_attempts counter of this boot target is decremented by one. This means every boot target’s remaining_attempts counter reaches zero sooner or later and the boot target won’t be booted anymore. To prevent that, the remaining_attempts counter must be reset to its default. There are different flags in the bootchooser which control resetting the remaining_attempts counter, controlled by the global.bootchooser.reset_attempts variable. It holds a list of space separated flags. Possible values are:

  • empty: counters will never be reset
  • power-on: The remaining_attempts counters of all enabled boot targets are reset after a power-on reset ($global.system.reset="POR"). This means after a power cycle all boot targets will be tried again for the configured number of retries.
  • all-zero: The remaining_attempts counters of all enabled boot targets are reset when none of them has any remaining_attempts left.

Additionally the remaining_attempts counter can be reset manually using the bootchoser command. This allows for custom conditions under which a system is marked as good. In case only the booted system itself knows when it is in a good state, the barebox-state tool from the dt-utils package can be used to reset the remaining_attempts counter from the running system.

1.19.3. General Bootchooser Options

In addition to the boot target options described above, bootchooser has some general options not specific to any boot target.

global.bootchooser.disable_on_zero_attempts
Boolean flag. If set to 1, bootchooser disables a boot target (sets priority to 0) whenever the remaining attempts counter reaches 0.
global.bootchooser.default_attempts
The default number of attempts that a boot target shall be tried before skipping it, used when not overwritten with the boot target specific variable of the same name.
global.bootchooser.default_priority
The default priority of a boot target when not overwritten with the target specific variable of the same name.
global.bootchooser.reset_attempts
Already described in Bootchooser Algorithm
global.bootchooser.reset_priorities

A space separated list of events that cause bootchooser to reset the priorities of all boot targets. Possible values:

  • empty: priorities will never be reset
  • all-zero: priorities will be reset when all targets have zero priority
global.bootchooser.retry
If set to 1, bootchooser retries booting until one succeeds or no more valid boot targets exist. Otherwise the boot command will return with an error after the first failed boot target.
global.bootchooser.state_prefix
Variable prefix when bootchooser is used with the state framework as backend for storing run-time data, see below.
global.bootchooser.targets
Space separated list of boot targets that are used. For each entry in the list a corresponding set of global.bootchooser.<targetname>.<variablename> variables must exist.
global.bootchooser.last_chosen
bootchooser sets this to the boot target that was chosen on last boot (index).

1.19.4. Setup Example

We want to set up a redundant machine with two bootable systems within one shared memory, here a NAND type flash memory with a UBI partition. We have a 512 MiB NAND type flash, to be used only for the root filesystem. The devicetree doesn’t define any partition, because we want to run one UBI partition with two volumens for the redundant root filesystems on this flash memory.

nand@0 {
   [...]
};

In order to configure this machine the following steps can be used:

ubiformat /dev/nand0 -y
ubiattach /dev/nand0
ubimkvol /dev/nand0.ubi root_filesystem_1 256MiB
ubimkvol /dev/nand0.ubi root_filesystem_2 0

The last command creates a volume which fills the remaining available space on the NAND type flash memory, which will be most of the time smaller than 256 MiB due to factory bad blocks and lost data blocks for UBI’s management.

After this preparation we can find two devices in /dev:

  • nand0.ubi.root_filesystem_1
  • nand0.ubi.root_filesystem_2

These two devices can now be populated with their filesystem content. In our example here we additionally assume, that these root filesystems contain a Linux kernel with its corresponding devicetree via boot spec (refer to Bootloader Spec for further details).

Either device can be booted with the boot command command, and thus can be used by the bootchooser and we can start to configure the bootchooser variables.

The following example shows how to initialize two boot targets, system1 and system2. Both boot from a UBIFS on nand0, the former has a priority of 21 and boots from the volume root_filesystem_1 whereas the latter has a priority of 20 and boots from the volume root_filesystem_2.

# initialize target 'system1'
nv bootchooser.system1.boot=nand0.ubi.root_filesystem_1
nv bootchooser.system1.default_attempts=3
nv bootchooser.system1.default_priority=21

# initialize target 'system2'
nv bootchooser.system2.boot=nand0.ubi.root_filesystem_2
nv bootchooser.system2.default_attempts=3
nv bootchooser.system2.default_priority=20

# make targets known
nv bootchooser.targets="system1 system2"

# retry until one target succeeds
nv bootchooser.retry=1

# First try bootchooser, when no targets remain boot from network
nv boot.default="bootchooser net"

Note

This example is for testing only, normally the NV variables would be initialized directly by files in the default environment, not with a script.

The run-time values are stored in environment variables as well. Alternatively, they can be stored in a state variable set instead. Refer to using the state framework for further details.

1.19.5. Scenarios

This section describes some scenarios that can be handled by bootchooser. All scenarios assume multiple boot targets that can be booted, where ‘multiple’ is anything higher than one.

1.19.5.1. Scenario 1

  • a system that shall always boot without user interaction
  • staying in the bootloader is not an option.

In this scenario a boot target is started for the configured number of remaining attempts. If it cannot be started successfully, the next boot target is chosen. This repeats until no boot targets are left to start, then all remaining attempts are reset to their defaults and the first boot target is tried again.

1.19.5.1.1. Settings

  • global.bootchooser.reset_attempts="all-zero"
  • global.bootchooser.reset_priorities="all-zero"
  • global.bootchooser.disable_on_zero_attempts=0
  • global.bootchooser.retry=1
  • global.boot.default="bootchooser recovery"
  • Userspace marks as good.

1.19.5.1.2. Deployment

  1. barebox or flash robot fills all boot targets with valid systems.
  2. The all-zero settings will lead to automatically enabling the boot targets, no default settings are needed here.

1.19.5.1.3. Recovery

Recovery will only be called when all boot targets are not startable (That is, no valid kernel found or read failure). Once a boot target is startable (a valid kernel is found and started) bootchooser will never fall through to the recovery boot target.

1.19.5.2. Scenario 2

  • a system with multiple boot targets
  • one recovery system

A boot target that was booted three times without success shall never be booted again (except after update or user interaction).

1.19.5.2.1. Settings

  • global.bootchooser.reset_attempts=""
  • global.bootchooser.reset_priorities=""
  • global.bootchooser.disable_on_zero_attempts=0
  • global.bootchooser.retry=1
  • global.boot.default="bootchooser recovery"
  • Userspace marks as good.

1.19.5.2.2. Deployment

  1. barebox or flash robot fills all boot targets with valid systems.
  2. barebox or flash robot marks boot targets as good or state contains non zero defaults for the remaining_attempts/priorities.

1.19.5.2.3. Recovery

Done by ‘recovery’ boot target which is booted after the bootchooser falls through due to the lack of bootable targets. This boot target can be:

  • a system that will be booted as recovery.
  • a barebox script that will be started.

1.19.5.3. Scenario 3

  • a system with multiple boot targets
  • one recovery system
  • a power cycle shall not be counted as failed boot.

Booting a boot target three times without success disables it.

1.19.5.3.1. Settings

  • global.bootchooser.reset_attempts="power-on"
  • global.bootchooser.reset_priorities=""
  • global.bootchooser.disable_on_zero_attempts=1
  • global.bootchooser.retry=1
  • global.boot.default="bootchooser recovery"
  • Userspace marks as good.

1.19.5.3.2. Deployment

  1. barebox or flash robot fills all boot targets with valid systems.
  2. barebox or flash robot marks boot targets as good.

1.19.5.3.3. Recovery

Done by ‘recovery’ boot target which is booted after the bootchooser falls through due to the lack of bootable targets. This target can be:

  • a system that will be booted as recovery.
  • a barebox script that will be started.

1.19.6. Using the State Framework as Backend for Run-Time Variable Data

Usually the bootchooser modifies its data in global variables which are connected to non volatile variables.

Alternatively the Barebox State Framework can be used for this data, which allows to store this data redundantly in some kind of persistent memory.

In order to let the bootchooser use the state framework for its storage backend, configure the bootchooser.state_prefix variable with the state variable set instance name.

Usually a generic state variable set in the devicetree is defined like this (refer to Barebox state for more details):

some_kind_of_state {
   [...]
};

At barebox run-time this will result in a state variable set instance called some_kind_of_state. You can also store variables unrelated to bootchooser (a serial number, MAC address, …) in it.

Extending this state variable set by information required by the bootchooser is simply done by adding so called ‘boot targets’ and optionally one last_chosen node. It then looks like:

some_kind_of_state {
  [...]
  boot_target_1 {
      [...]
  };
  boot_target_2 {
      [...]
  };
};

It could makes sense to store the result of the last bootchooser operation in the state variable set as well. In order to do so, add a node with the name last_chosen to the state variable set. bootchooser will use it if present. The *state’ variable set definition then looks like:

some_kind_of_state {
  [...]
  boot_target_1 {
      [...]
  };
  boot_target_2 {
      [...]
  };
  last_chosen {
      reg = <offset 0x4>;
      type = "uint32";
  };
};

The boot_target_* names shown above aren’t variables themselves (like the other variables in the state variable set), they are named containers instead, which are used to group variables specific to bootchooser.

A ‘boot target’ container has the following fixed content:

some_boot_target {
       #address-cells = <1>;
       #size-cells = <1>;

       remaining_attempts {
           [...]
           default = <some value>; /* -> read note below */
       };

       priority {
           [...]
           default = <some value>; /* -> read note below */
       };
};

Important

Since each variable in a state variable set requires a reg property, the value of its reg property must be unique, e.g. the offsets must be consecutive from a global point of view, as they describe the storage layout in the backend memory.

So, remaining_attempts and priority are required variable nodes and are used to setup the corresponding run-time environment variables in the global.bootchooser.<targetname> namespace.

Important

It is important to provide a default value for each variable for the case when the state variable set backend memory is uninitialized. This is also true if default values through the bootchooser’s environment variables are defined (e.g. bootchooser.default_attempts, bootchooser.default_priority and their corresponding boot target specific variables). The former ones are forwarded to the bootchooser to make a decision, the latter ones are used by the bootchooser to make a decision the next time.

1.19.6.1. Example

For this example we use the same system and its setup described in setup example. The resulting devicetree content for the state variable set looks like:

system_state {
     [...]
     system1 {
          #address-cells = <1>;
          #size-cells = <1>;
          remaining_attempts@0 {
              reg = <0x0 0x4>;
              type = "uint32";
              default = <3>;
          };
          priority@4 {
              reg = <0x4 0x4>;
              type = "uint32";
              default = <20>;
          };
     };

     system2 {
          #address-cells = <1>;
          #size-cells = <1>;
          remaining_attempts@8 {
              reg = <0x8 0x4>;
              type = "uint32";
              default = <3>;
          };
          priority@c {
              reg = <0xc 0x4>;
              type = "uint32";
              default = <21>;
          };
     };

     last_chosen@10 {
          reg = <0x10 0x4>;
          type = "uint32";
     };
};

Important

While the system1/2 nodes suggest a different namespace inside the state variable set, the actual variable’s reg-properties and their offset part are always relative to the whole state variable set and thus must be consecutive globally.

To make bootchooser use the so called system_state state variable set instead of the NV run-time environment variables, we just set:

global.bootchooser.state_prefix=system_state

Note

Its a good idea to keep the bootchooser.<targetname>.default_priority and bootchooser.<targetname>.default_attempts values in sync with the corresponding default values in the devicetree.

1.19.7. Updating systems

Updating a boot target is the same among the different scenarios. It is assumed that the update is done under a running Linux system which can be one of the regular bootchooser boot targets or a dedicated recovery system. For the regular bootchooser boot targets updating is done like:

  • Disable the inactive (e.g. not used right now) boot target by setting its priority to 0.
  • Update the inactive boot target.
  • Set remaining_attempts of the inactive boot target to nonzero.
  • Enable the inactive boot target by setting its priority to a higher value than any other boot target (including the used one right now).
  • Reboot.
  • If necessary update the now inactive, not yet updated boot target the same way.

One way of updating systems is using RAUC which integrates well with the bootchooser in barebox.