XNU startup sequence

General Principles

XNU Startup sequence is driven by the <kern/startup.h> module.

The startup sequence is made of individual subsystems (the STARTUP_SUB_* values of the startup_subsystem_id_t type) that get initialized in sequence.

A subsystem can use ranks to order the various initializers that make up its initialization sequence. Usage of ranks is custom to each subsystem and must be documented in this file.

The subsystem module will basically run hooks in that order:

for (subsystem 0 -> N) {
  for (rank 0 -> N) {
    // run in no particular order for a given rank in the given subsystem
    init(subsystem, rank);
  }
}

Extending the startup sequence

When extending the startup sequence:

add a new value to the startup_subsystem_id_t enum in the right order
document what services this phase provides, and how it uses ranks in this file.

When hooking with a given subsystem, consult this documentation to use the proper rank for your callback.

If a new rank needs to be used, update this documentation in the proper section.

`STARTUP_SUB_TUNABLES`

Description

Initializes various globals that alter the behavior of the kernel, lookup tables, ... Available hooks are:

TUNABLES: parses a boot arg into a global that will become read-only at lockdown time,
TUNABLE_WRITEABLE: same as TUNABLE but the global will not be locked down.

Rank usage

Rank 1: TUNABLE, TUNABLE_WRITEABLE
Middle: globals that require complex initialization (e.g. SFI classes).

`STARTUP_SUB_LOCKS_EARLY`

Description

Initializes early locks that do not require any memory allocations to be initialized. Available hooks are:

LCK_GRP_DECLARE*: automatically initialized lock groups,
LCK_GRP_ATTR_DECLARE: automatically initialized lock group attributes,
LCK_ATTR_DECLARE: automatically initialized lock attributes,
LCK_SPIN_DECLARE*: automatically initialized spinlocks,
LCK_RW_DECLARE: automatically initialized reader/writer lock,
LCK_MTX_EARLY_DECLARE*: automatically initialized mutexes, with statically allocated buffers for statistics/tracing,
SIMPLE_LOCK_DECLARE*: automatically initialized simple locks.

Rank usage

Rank 1: Initializes the module (lck_mod_init),
Rank 2: LCK_GRP_ATTR_DECLARE, LCK_ATTR_DECLARE,
Rank 3: LCK_GRP_DECLARE*
Rank 4: LCK_SPIN_DECLARE*, LCK_MTX_EARLY_DECLARE*, LCK_RW_DECLARE, SIMPLE_LOCK_DECLARE*.

`STARTUP_SUB_KPRINTF`

Description

Initializes the kprintf subsystem.

Rank usage

Rank 1: calls the module initializer (PE_init_kprintf).

`STARTUP_SUB_PMAP_STEAL`

Description

Allows for subsystems to steal early memory.

Rank usage

N/A.

`STARTUP_SUB_VM_KERNEL`

Description

Denotes that the early kernel VM is initialized.

Rank usage

N/A.

`STARTUP_SUB_KMEM`

Description

Denotes that kernel_memory_allocate is now usable.

Rank usage

N/A.

`STARTUP_SUB_KMEM_ALLOC`

Description

Denotes that kmem_alloc is now usable.

Rank usage

N/A.

`STARTUP_SUB_ZALLOC`

Description

Initializes the zone allocator.

ZONE_DECLARE, ZONE_INIT: automatically initialized permanent zones.
ZONE_VIEW_DEFINE, KALLOC_HEAP_DEFINE: zone and kalloc heap views.

Rank usage

Rank 1: zone_init: setup the zone subsystem, this allows for the already created VM/pmap zones to become dynamic.
Rank 2: vm_page_module_init: create the "vm pages" zone. The vm_page_zone must be created prior to kalloc_init; that routine can trigger zalloc()s (for e.g. mutex statistic structure initialization).

The vm_page_zone must exist to satisfy fictitious page allocations (which are used for guard pages by the guard mode zone allocator).
Rank 3: Initialize kalloc.
Rank 4: Enable zone caching (uses kalloc)
Middle: for any initialization that only requires kalloc/zalloc runs ZONE_DECLARE and ZONE_INIT.
Last: zone and kalloc heaps (ZONE_VIEW_DEFINE, KALLOC_HEAP_DEFINE).

`STARTUP_SUB_PERCPU`

Description

Initializes the percpu subsystem.

Rank usage

Rank 1: allocates the percpu memory, percpu_foreach_base and percpu_foreach become usable.

Rank 2: sets up static percpu counters.

`STARTUP_SUB_LOCKS`

Description

Initializes kernel locks that might require allocations (due to statistics and tracing features). Available hooks are:

LCK_MTX_DECLARE: automatically initialized mutex,

Rank usage

Rank 1: LCK_MTX_DECLARE.

`STARTUP_SUB_CODESIGNING`

Description

Initializes the codesigning subsystem.

Rank usage

Rank 1: calls the module initializer (cs_init).

`STARTUP_SUB_OSLOG`

Description

Initializes the os_log facilities.

Rank usage

Rank 1: Calls the module initializer (oslog_init).

`STARTUP_SUB_MACH_IPC`

Description

Initializes the Mach IPC subsystem.

Rank usage

Rank 1: Initializes IPC submodule globals (ipc tables, voucher hashes, ...)
Rank last: Final IPC initialization.

`STARTUP_SUB_SYSCTL`

Description

Initializes the sysctl kernel subsystem

Rank usage

Rank 1: automatic SYSCTL_NODE registration.
Rank 2: automatic SYSCTL_OID registration.
Middle: other manual early registrations.
Last: registrations of dummy nodes in the constant nodes to allow extension.

`STARTUP_SUB_EARLY_BOOT`

Description

Denotes that subsystems that expect to operate with interrupts or preemption enabled may begin enforcement.

Rank usage

N/A.

`STARTUP_SUB_LOCKDOWN`

Description

Denotes that the kernel is locking down, this phase should never be hooked. When the kernel locks down:

data marked __startup_data and code marked __startup_func is unmapped,
data marked __security_const_late or SECURITY_READ_ONLY_LATE becomes read-only.

Rank usage

N/A.