What is an OS bootstrap?

An operating system's bootstrap is the process that happens shortly after you press the power on button, as shown below:

Power-on -> Bios -> Bootloader -> OS Bootstrap -> service manager

Note that in this context the OS bootstrap is not building a distribution and packages from source code. The OS bootstrap has nothing to do with reproducible builds.

State at the beginning of the bootstrap

Also consider reading about Serverboot V2, which is a new bootstrap proposal.

After initializing itself, GNU Mach sets up tasks for the various bootstrap translators (which were loader by the GRUB bootloader). It notably makes variables replacement on their command lines and boot script function calls (see the details in gnumach/kern/boot_script.c). For instance, the GRUB bootloader can have the following typical configuration:

multiboot       /boot/gnumach-1.8-486-dbg.gz root=device:hd1 console=com0
module          /hurd/pci-arbiter.static pci-arbiter
                --host-priv-port='${host-port}' \
                --device-master-port='${device-port}' \
                --next-task='${acpi-task}' \
                '$(pci-task=task-create)' '$(task-resume)'
module          /hurd/acpi.static acpi \
                --next-task='${disk-task}' \
                '$(acpi-task=task-create)'
module          /hurd/rumpdisk.static rumpdisk \
                --next-task='${fs-task}' \
                '$(disk-task=task-create)'
module          /hurd/ext2fs.static ext2fs --readonly \
                --multiboot-command-line='${kernel-command-line}' \
                --exec-server-task='${exec-task}' -T typed '${root}' \
                '$(fs-task=task-create)'
module          /lib/ld-x86-64.so.1 exec /hurd/exec '$(exec-task=task-create)'

Note: use ld.so.1 instead of ld-x86-64.so.1 on 32b systems.

GNU Mach will first make the $(task-create) function calls, and thus create a series of tasks for the various modules, and assign to the pci-task, acpi-task, disk-task, and fs-task variables the task ports for each of them. None of these tasks is started yet.

It will then replace the variables (${foo}), i.e.

• ${kernel-command-line} with its own command line (root=device:hd1 console=com0),
• ${host-port} with a reference to the GNU Mach host port,
• ${device-port} with a reference to the GNU Mach device port,
• ${acpi-task} with a reference to the acpi task port, and similarly for all other tasks.
• ${root} with device:hd1

This typically results in:

task loaded: pci-arbiter --host-priv-port=1 --device-master-port=2 --next-task=3
task loaded: acpi --next-task=1
task loaded: rumpdisk --next-task=1
task loaded: ext2fs --readonly --multiboot-command-line=root="device:sd1 console=com0" --exec-server-task=1 -T typed device:sd1
task loaded: exec /hurd/exec

(You will have noticed that /hurd/exec is not run directly, but through ld.so.1: Mach only knows to run statically-linked ELF binaries, so we could either load /hurd/exec.static directly, or load the dynamic loader ld.so.1 and tell it to load /hurd/exec, which will be readable once ext2fs.static is started)

GNU Mach will eventually make the $(task-resume) function calls, and thus resume the pci-arbiter task only.

Usually the bootstrap ports of translators is used when starting them, see fshelp_start_translator_long: the parent translator starts the child and sets its bootstrap port. The parent then waits for the child to call fsys_startup on the bootstrap port, for the child to provide its control port, and for the parent to provide the FS node that the child is translator for.

But here when pci-arbiter initializes itself, it notices that its bootstrap port is nul (it is started by the kernel, not a filesystem) so it knows that it is alone and can only rely on the kernel. It initializes itself and parses the arguments, and since it is given a next-task, it uses task_set_special_port to pass a send right to its own control port to that next task (here acpi) as bootstrap port, and uses task_resume to start it.

Similarly, acpi initializes itself, gives a send right to rumpdisk and starts it.

rumpdisk does the same, so that eventually ext2fs starts, with all of pci-arbiter, acpi and rumpdisk ready to reply to device_open requests on the pci, acpi, and disks device names.

Now that ext2fs starts, a dance begin between the remaining bootstrap processes: ext2fs, exec, startup, proc, and auth. Indeed, there are a few dependencies between them: exec needs ext2fs working to be able to start startup, proc and auth, and ext2fs needs to register itself to startup, proc and auth so as to appear as a normal process, running under uid 0.

They will register to each other the following way:

• Between ext2fs and startup: startup calls fsys_init, to provide ext2fs with proc and auth ports.
• Between startup and proc: proc just calls startup_procinit to hand over a proc port and get auth and priv ports.
• Between startup and auth: auth calls startup_authinit to hand over an auth port and get a proc port, then calls startup_essential_task to notify startup that the boot can proceed.
• For the series of translators before ext2fs, each task calls fsys_startup to pass over the control port of ext2fs to the previous task (instead of its own control port, which is useless for it). This is typically done in the S_fsys_startup stub, simply forwarding it. It also calls fsys_init to pass over the proc and auth ports. Again, this is typically done in the S_fsys_init stub, simply forwarding them.

With that in mind, the dance between the bootstrap translators is happening as described in the next sections.

ext2fs initialization

ext2fs's main function starts by calling diskfs_init_main.

diskfs_init_main parses the ext2fs command line with argp_parse, to record the parameters set up by the kernel. It makes sure to have a working stdout by opening the Mach console.

Since the multiboot command line is available, diskfs_init_main sets the ext2fs bootstrap port to MACH_PORT_NULL: it is the bootstrap filesystem which will be in charge of dancing with the exec and startup translator.

diskfs_init_main then initializes the libdiskfs library and spawns a thread to manage libdiskfs RPCs. It also notices that the filesystem is given a kernel command line, i.e. this is the bootstrap filesystem.

ext2fs continues its initialization: creating a pager, opening the hypermetadata, opening the root inode to be set as root by libdiskfs.

ext2fs then calls diskfs_startup_diskfs to really run the startup, implemented by the libdiskfs library.

libdiskfs bootstrap

Since this is the bootstrap filesystem, diskfs_startup_diskfs calls diskfs_start_bootstrap.

diskfs_start_bootstrap starts by creating a open port on itself for the current and root directory, all other processes will inherit it.

diskfs_start_bootstrap does have a port on the exec task, so it can dance with it. It calls start_execserver that sets the bootstrap port of the exec task to a port of the diskfs_execboot_class, and resumes the exec task. diskfs_start_bootstrap then waits for execstarted.

exec bootstrap

exec's main function starts and calls task_get_bootstrap_port to get its bootstrap port and getproc to get a port on the proc translator (thus MACH_PORT_NULL at this point since the proc translator is not started yet).

exec initializes the trivfs library, and eventually calls trivfs_startup on its bootstrap port.

trivfs_startup creates a control port for the exec translator, and calls fsys_startup on the bootstrap port to notify ext2fs that it is ready, give it its exec control port, and get back a port on the underlying node for the exec translator (we want to make it show up on /servers/exec).

libdiskfs taking back control

diskfs_execboot_fsys_startup is thus called. It calls dir_lookup on /servers/exec to return the underlying node for the exec translator, and stores the exec control port in diskfs_exec_ctl. It can then signal execstarted.

diskfs_start_bootstrap thus takes back control, It calls fsys_getroot on the control port of exec, and uses dir_lookup and file_set_translator to attach it to /servers/exec.

diskfs_start_bootstrap then looks for which startup process to run. It may be specified on the multiboot command line, but otherwise it will default to /hurd/startup.

Now that exec is up and running, the startup process can be created with exec_exec. diskfs_start_bootstrap takes a lot of care in this: this is the first unix-looking process, it notably inherits the root directory and current working directory initialized above, it gets stdin/out/err on the mach console. It is passed as bootstrap port a port from the diskfs_control_class.

diskfs_start_bootstrap is complete, we are back to diskfs_startup_diskfs, which checks whether ext2fs was given a bootstrap port, i.e. whether the rumpdisk translator was started before ext2fs. If so, it calls diskfs_call_fsys_startup which creates a new control port and passes it do a call to fsys_startup on the bootstrap port, so rumpdisk gets access to the ext2fs filesystem. Rumpdisk however does not return any realnode port, since we are not exposing the ext2fs filesystem in rumpdisk, but rather the converse. TODO: Rumpdisk forwards this fsys_startup call to pci-arbiter, so the latter also gets access to the ext2fs filesystem.

startup

startup's main function starts and calls task_get_bootstrap_port to get its bootstrap port, i.e. the control port of ext2fs, and fsys_getpriv on it to get the host privileged port and device master port. It clears the bootstrap port so children do not inherit it. It sets itself up with output on the Mach console, and wires itself against swapping. It requests notification for ext2fs translator dying to detect it and print a warning in case that happens during boot. It creates a startup port which it will get RPCs on.

startup can then complete the unixish initialization, and run /hurd/proc and /hurd/auth, giving them as bootstrap port the startup port.

proc

proc's main function starts. It initializes itself, and calls task_get_bootstrap_port to get a port on startup. It can then call startup_procinit on it to pass it the proc port that will represent the startup task, and get ports on the auth server, the host privileged port, and device master port.

Eventually, proc calls startup_essential_task to tell startup that it is ready.

auth

auth's main function starts. It creates the initial root auth handle (all permissions allowed). It calls task_get_bootstrap_port to get a port on startup. It can then call startup_authinit on it to pass the initial root auth handle, and get a port on the proc server. It can then register itself to proc.

Eventually, auth calls startup_essential_task to tell startup that it is ready.

startup getting back control

startup notices initialization of auth and proc from S_startup_procinit and S_startup_authinit. Once both have been called, it calls launch_core_servers.

launch_core_servers starts by calling startup_procinit_reply to actually reply to the startup_procinit RPC with a port on auth.

launch_core_servers then tells proc that the bootstrap processes are important, and how they relate to each other.

launch_core_servers then calls install_as_translator to show up in the filesystem on /servers/startup.

launch_core_servers calls startup_authinit_reply to actually reply to the startup_authinit RPC with a port on proc.

launch_core_servers eventually calls fsys_init on its bootstrap port, to give ext2fs the proc and auth ports.

diskfs' diskfs_S_fsys_init thus gets called. It first replies to startup, so startup is not stuck in its fsys_init call and not able to reply to RPCs. From then on, startup will be watching for startup_essential_task calls from the various bootstrap processes.

libdiskfs taking back control

In diskfs' diskfs_S_fsys_init, diskfs now knows that proc and auth are ready, and can call exec_init on the exec port.

exec getting initialized

exec's S_exec_init gets called from the exec_init call from ext2fs. Exec can register itself with proc, and eventually call startup_essential_task to tell startup that it is ready.

back to libdiskfs initialization

It also calls fsys_init on its bootstrap port, i.e. rumpdisk.

rumpdisk getting initialized

rumpdisk's trivfs_S_fsys_init gets called from the fsys_init call from ext2fs. It calls fsys_init on its bootstrap port.

acpi getting initialized

acpi's trivfs_S_fsys_init gets called from the fsys_init call from rumpdisk. It calls fsys_init on its bootstrap port.

pci-arbiter getting initialized

pci-arbiter's trivfs_S_fsys_init gets called from the fsys_init call from rumpdisk.

It gets the root node of ext2fs, sets all common ports, and install itself in the ext2fs FS as translator for /servers/bus/pci.

It eventually calls startup_essential_task to tell startup that it is ready, and requests shutdown notifications.

back to acpi initialization

It gets the root node of ext2fs, sets all common ports, and install itself in the ext2fs FS as translator for /servers/acpi.

It eventually calls startup_essential_task to tell startup that it is ready, and requests shutdown notifications.

back to rumpdisk initialization

It gets the root node of ext2fs, sets all common ports, and install itself in the ext2fs FS as translator for /dev/disk.

It eventually calls startup_essential_task to tell startup that it is ready, and requests shutdown notifications.

back to libdiskfs initialization

It initializes the default proc and auth ports to be given to processes.

It calls startup_essential_task on the startup port to tell startup that it is ready.

Eventually, it calls _diskfs_init_completed to finish its initialization, and notably call startup_request_notification to get notified by startup when the system is shutting down.

startup monitoring bootstrap progress

As mentioned above, the different essential tasks (pci-arbiter, acpi, rumpdisk, ext2fs, proc, auth, exec) call startup_essential_task when they are ready. startup's S_startup_essential_task function thus gets called each time, and startup records each of them as essential, monitoring their death to crash the whole system.

Once all of proc, auth, exec have called startup_essential_task, startup replies to their respective RPCs, so they actually start working altogether. It also calls init_stdarrays which sets the initial values of the standard exec data, and frob_kernel_process to plug the kernel task into the picture. It eventually calls launch_something, which "launches something", which by default is /libexec/runsystem, but if that can not be found, launches a shell instead, so the user can fix it.