process_enumerate snapshots the task table (the device_enumerate shape, so
ps is a user program); system_spawn returns the child id, records the caller
as supervisor, and takes an exit endpoint; process_kill is allowed only for
the supervisor. Every death — exit, fault, or kill — posts a child-exit badge
to that endpoint (the IRQ-as-IPC pattern as SIGCHLD). A target caught off-CPU
is reaped in place; a running one is condemned and finished at its next
system call or tick, guarded so teardown never lands mid-kernel-operation.
Tested by process-list, process-kill, and supervision (a ring-3 supervisor
exercising the whole surface); design notes in docs/process-management.md.
Processes now start with C-compatible arguments: the kernel builds the
System V AMD64 entry block (argc, argv, empty envp, auxiliary vector)
at the top of the stack, argv[0] is the path or initial-ramdisk name
the process was spawned as, and system_spawn carries an optional
NUL-separated blob that becomes argv[1..]. The runtime parses the block
(runtime.argumentCount/argument) and its spawn wrappers pass arguments
through. The name is also recorded on the task, so a fault report says
which binary died, not just its id.
The user stack grows from one page to eight (32 KiB,
parameters.user_stack_pages), with the page below left unmapped as a
guard so an overflow faults into a clean process kill rather than
corrupting the image. Task.name_buffer is zero-initialised, not
undefined: an undefined default is materialised as a 0xAA fill that
moved the static task pool out of .bss and made the whole kernel ~7x
slower under QEMU TCG (caught by the affinity test).
Proven end to end by the new args test: args-echo respawns itself with
arguments via the syscall blob, burns more stack than one page could
hold, and echoes its argv intact. Full suite: 44/44.
A CPU exception raised in ring 3 by a scheduled process now kills that
process - IRQ bindings, IPC handles, and address space reclaimed, a
client it owed a reply to failed with the new -EPEER instead of hung -
and the core reschedules (docs/resilience.md step 2). Kernel-mode
faults, NMI, double fault, and machine check stay terminal, as does the
borrowed-thread isolation probe. Proven by the new fault-recovery QEMU
test: init keeps heartbeating after a process page-faults to death.
Correct the abi.zig header (and the matching build.zig / README lines): the syscall ABI
is the private contract between the kernel and the runtime library, not something every
user program speaks. danos applications call the `runtime` (the stable, danos-native
ABI); the runtime is the only thing that issues system calls, and POSIX layers over the
runtime — the same split as libSystem on macOS or win32 over the NT syscalls. The call
numbers here are an implementation detail the runtime hides and may renumber, not a
public interface.
Ring 3 still has no direct in/out (no TSS I/O bitmap, IOPL never raised — a #GP), but a
driver no longer needs it: io_read(device_id, resource_index, offset, width) and
io_write(..., value) grant port access the same way mmio_map grants memory. The claim
plus the device's discovered io_port resource are the capability — resolveIoPort checks
the device is claimed by the caller, the resource is io_port, and [offset, offset+width)
stays inside it, then issues the in/out via architecture.pioRead/pioWrite. So a PS/2 or
16550 driver is now writable; the low-rate legacy hardware that needs port I/O is fine
with a syscall per access. io_port resources were recorded by discovery and ignored —
now they're used. Runtime: device.ioRead/ioWrite.
New `ioport` test claims QEMU's PS/2 controller (io_port 0x64, discovered via ACPI) and
checks the gate admits an in-range access, refuses over-wide / out-of-range / unclaimed,
and that the kernel actually reads the status port (0x1c). Suite 41/41 plus host tests.
Docs refreshed for the whole M13-M16 + port-I/O reality: drivers.md "Limits" no longer
lists port I/O, DMA memory, or barriers as missing (they exist) and its "what's next"
reflects that; the FSH doc's character-device and block-device sections are corrected
("cannot host a block driver at all" is no longer true — writable now, not yet memory-
safe pending IOMMU enforcement); device-interrupts.md unblocks the keyboard and narrows
the interrupt gap to MSI-X.
Detect the IOMMU: discovery now parses the ACPI DMAR table, finds the first VT-d
DMA-remapping unit (DRHD), maps its register block, and records its version and
capabilities (iommu_present/base/version/capabilities in the platform info). On QEMU's
emulated intel-iommu this reads back a real unit (base 0xfed90000, version 1.0).
This is detection only, and deliberately so. A full VT-d bring-up — per-device
translation domains that confine a driver's DMA to the buffers it dma_alloc'd — is the
real device-side safety guarantee, but it cannot be verified without a DMA-capable
device driver (none exist yet) and QEMU's intel-iommu to fault against. Writing that
enforcement now would be a large body of unverifiable page-table code; it belongs with
the first DMA driver, which is both the natural order and the only way to test it. Until
then the caveat stands in full: device_claim on a DMA-capable device is still equivalent
to granting ring 0. The docs say so plainly.
New `iommu` test (harness boots it with -device intel-iommu via a new per-case qemu_extra
hook) confirms the DMAR is parsed and the unit's registers read. Suite 40/40 plus host
tests.
Two parts, both blockers for real PCI drivers.
ECAM config space per function: enumeratePci now gives every pci_device its own 4 KiB
configuration window as resource 0. A claimed PCI driver mmio_maps that to reach its
command register, BARs, and — the point — its capability list (MSI/MSI-X, PCIe extended
caps), with no new syscall. Verified in the discovery test (QEMU q35's functions each
carry it).
MSI: msi_bind(device_id, endpoint) -> address (rax), data (rdx) allocates a per-device
edge-triggered vector, binds it to the endpoint, and returns the (address, data) the
driver programs into its own MSI capability. Unlike irq_bind there's no GSI, no I/O APIC
entry, no sharing, and no ack cycle — dispatch recognises an MSI vector (vector_gsi ==
none, msi_bound set), EOIs, and notifies. Owner-keyed release drops the binding on exit.
Legacy INTx (_PRT parsing + shared lines) is deliberately skipped; MSI is the real
answer.
QEMU's HPET has no MSI, so the new `msi` test proves the vector-routing path with a
self-IPI (new apic.selfIpi) standing in for the device's MSI write: bind a vector,
fire it, the bound endpoint is notified. The msi_bind syscall wraps irq.msiBind with
the claim check and lands its first real use with the first PCI driver. Suite 39/39
plus host tests.
An HCD programs a bus-master engine: it needs a descriptor ring that is physically
contiguous, at a physical address it knows, uncacheable, and pinned. mmap gives none
of those. Add dma_alloc(len, flags) -> vaddr (rax), paddr (rdx) and dma_free(vaddr,
len): grant contiguous, zeroed, pinned, strong-uncacheable memory in a per-process DMA
arena (PML4[228]) and hand back both addresses.
Pieces: pmm.allocContiguous(count, max_phys) finds a run of contiguous free frames
below a cap (dma_below_4g for 32-bit engines); mapUserDmaInto maps them uncacheable
(PCD|PWT) but WITHOUT device_grant, so unlike an MMIO grant these frames are real RAM
and freeSubtree returns them on teardown — a driver that dies leaks nothing. dma_free
is bounded to the DMA arena so it can never unmap the caller's stack/heap/MMIO.
dma_write_combining is accepted but falls back to coherent (WC needs PAT programming).
Runtime: runtime.dma.alloc/free (a two-return-value stub, like replyWait). New `dma`
kernel test drives the mechanism directly — contiguity, the below-4G cap, coherent
mapping, and reclaim-on-teardown (no leak). The thin syscall wrappers follow the tested
mmap/mmio_map shape and land their first real use with the first DMA driver. Suite
38/38 plus host tests.
ipc_call and ipc_reply_wait grow a `send_cap` argument (r9) and a `received_cap`
return (r8): an endpoint travels alongside a message, installed into the receiver's
handle table. The transfer is a share, not a move — the endpoint's refcount is bumped
and the sender keeps its handle. If the receiver's table is full the call fails
-ENOSPC and the message is NOT delivered (a half-delivered capability is worse than a
failed send); a bad handle fails -EBADF. Both directions carry a cap: a client's call
hands one to the server (seen in the server's replyWait), and the server's reply hands
one back (seen in the client's call return).
This is the "open" primitive the driver model was blocked on: a bus driver mints a
per-device endpoint and hands it to a class driver, giving it a private channel to one
device without the 8-slot global name registry.
Kernel: shareCapability in ipc-synchronous.zig at both copy points; new
setSystemCallResult3 (r8, saved/restored by the syscall stub); Task gains
ipc_send_cap / ipc_received_cap. Runtime: callCap + Reply, replyWait gains send_cap
and Received.cap; plain call/replyWait delegate with no_cap. New abi.no_cap.
New ipc-cap test (two kernel tasks exercise both directions, each verifying the
endpoint it received is the same object shared, refcount bumped to 2). No class driver
consumes callCap yet — it lands with the first one. Suite 37/37 plus host tests.
drivers.md gains a "How a driver gets started" section — the doc explained what a
running driver does but never who starts it. It lays out the three-level supervision
hierarchy (kernel spawns init; init spawns the services; the device-manager discovers,
matches, and spawns the drivers) and names the two user-space policies that "configure"
drivers today: init's service list and the device-manager's match table.
driver-model.md's "what exists today" adds system_spawn and the supervision model, and
the drivers.md restart bullet is refreshed: a spawning supervisor now exists, a
restarting one still doesn't.
The reorg moved user binaries under /system (init -> /system/services/init,
drivers -> /system/drivers/<name>, vfs-test -> /system/services/vfs/vfs-test), but
many comments and log strings still named the old /sbin/ home. Retarget them all:
kernel/loader/test comments and the two boot log lines, plus vision.md and the
driver-model.md proposed tree (also dropped the stale `d` suffixes and rt->runtime
there). The initial-ramdisk spawn log no longer fakes a /sbin/ prefix, since those
binaries live in different homes (services vs drivers).
Left the FSH design doc's /sbin and /lib rows alone — whether /sbin stays a
directory at all is a design call for its owner, not a stale-comment fix.
The `system` module (formerly `danos`) had become a grab-bag: it held the
loader<->kernel handoff *and* the kernel<->user ABI *and* the device wire types, in
one module three different audiences imported. Usage proved the seam — the
bootloader never touched the syscall/device ABI, and user space never touched the
boot handoff — so split it by audience, one module per contract:
system/boot-handoff.zig loader <-> kernel: BootInformation, Framebuffer,
MemoryMap, the VM layout + physicalToVirtual, kernel_abi
system/abi.zig kernel <-> user, core: SystemCall, mmap prot flags,
page_size, notify_badge_bit, ServiceId
system/devices/device-abi.zig kernel <-> user, devices: DeviceDescriptor,
DeviceClass, ResourceDescriptor, ResourceKind, ...
device-abi is the devices sub-project's public interface, exposed as its own module
the way vfs exposes vfs-protocol — importable by user space, unlike the
kernel-internal device model it also feeds. That collapses a real duplication:
DeviceClass and ResourceKind were defined twice (device-model.zig and the contract,
kept "in sync by hand"); device-model now re-exports them from device-abi, so the
enum a driver matches on and the one the kernel classifies with are one type.
Each import now declares which contract it speaks: the bootloader imports only
boot-handoff; a driver only abi + device-abi (via the runtime); the kernel all
three. This also retires the `system` / `runtime.system` name overlap. page_size
lands in abi (it's part of the mmap contract user space aligns to); the bootloader
keeps its own local 4 KiB constant so it depends on nothing but the handoff.
All 21 importers rewired, docs updated to keep /system mapping to source. Build,
host tests, and the QEMU suite (36/36) all green.
The device manager is the ring-3 process that turns the device tree into a running
system — the udev-analog. It is mechanism-vs-policy done right: the kernel
enumerates the hardware and enforces the claim capability; this decides which
driver serves which device, using no special privilege (the same device_enumerate
any process could call).
This first increment does the discovery + matching half: system/services/
device-manager enumerates /system/devices, matches each device to a driver by
class (a small static policy table), and logs the decision — finding the HPET
(a timer) and deciding `hpet` serves it. It does not spawn yet: spawning needs a
`system_spawn` system call (the kernel spawns every initial-ramdisk binary in a
loop today), which is the next increment. New `device-manager` test; suite 36/36
plus host tests.
The shared kernel<->user ABI contract (BootInformation, the SystemCall numbers,
DeviceDescriptor, page_size, ...) is now the `system` module at
system/system.zig, following the convention that a directory's root file takes
the directory's name.
One overlap to note: the runtime's syscall wrappers are already `runtime.system`,
so the single file that uses both the contract and those wrappers
(library/runtime/heap.zig) aliases the wrappers locally as `system_calls`. The
two are distinct (top-level `system` vs `runtime.system`); everywhere else the
contract is just `system`.
Also: the QEMU run's serial capture now lands in the FHS log location,
zig-out/var/log/system/serial0-<timestamp>.log — a stand-in for the kernel's own
logging system, which will eventually write there itself.
Suite 35/35 plus host tests green.
`zig build` now installs into a FHS-shaped zig-out that *is* the danos filesystem
and the boot volume — no more zig-out/bin or a separate esp/:
zig-out/EFI/BOOT/BOOTX64.efi (firmware entry; UEFI fixes this path)
zig-out/boot/initial-ramdisk.img
zig-out/system/kernel (the kernel binary)
zig-out/system/services/init vfs
zig-out/system/drivers/hpet bus
Binaries land at their addressed, leaf-collapsed paths per the sub-project
resolution rule (system/services/init/init.zig -> system/services/init); vfs, hpet,
and bus are installed to their FHS homes too, so the image is complete even though
at boot they arrive inside the initial-ramdisk.
The bootloader (boot/efi.zig) now loads each artifact from its FHS path
(system\kernel, system\services\init, boot\initial-ramdisk.img); run-x86-64 mounts
zig-out directly; the QEMU test harness assembles its ESP from the FHS zig-out.
Also renames system/kernel/main.zig -> kernel.zig so the kernel follows the
name/name.zig convention (kernel/ = ring-0 code, services/ = ring-3 OS services).
Documents the resolution rule in the repository-layout section (README + coding
standard). Suite 35/35 plus host tests green.
Follow-up to the monorepo re-org. Suite 35/35 plus host tests green.
POSIX compatibility is now its own library, library/posix/ (unistd, stdio),
layered strictly over the runtime — it calls the runtime's IPC/heap, never
system calls directly. The runtime is now POSIX-free (the danos-native
application ABI). The VFS wire protocol is danos-native throughout
(Stat -> FileStatus, .stat -> .status, O_CREAT -> create); the POSIX layer
maps the POSIX spellings at the boundary. The coding standard's ABI-name
exception is scoped to one place: a file is allowed POSIX spellings only if it
lives under library/posix/ — everywhere else, danos naming with no exception.
Naming fixes, all mechanical:
- initrd -> initial-ramdisk: the source file, the module, the tool
(make-initial-ramdisk.py), the artifact (initial-ramdisk.img, including the
bootloader's load path), and the identifiers.
- system/kernel/device-service.zig -> devices-broker.zig: it is ring-0 kernel
code (the trusted device table + claim capability), not a ring-3 service. The
future user-space device *manager* (policy) will live in system/services/.
- Dropped the daemon `d` suffix: hpetd -> hpet, busd -> bus. A driver lives in
system/drivers/, so the folder already says what it is; encoding the role in
the name too is redundant. The coding standard drops that exception.
- system/devices/aml/interp.zig -> interpreter.zig (the type was already
Interpreter).
The source layout now mirrors the runtime filesystem hierarchy
(docs/danos-file-system-hierarchy-FSH.md): what lives under system/ in the
source is what a running danos represents under /system. Each service and
driver is a sub-project directory that is its own Zig module — cross-project
references go by module name, never by a path into another project's files.
Moves (all git mv, history preserved):
- src/ -> system/ (danos internals; the self-representation)
root.zig -> danos.zig (the kernel<->user contract module)
kernel/arch/ -> kernel/architecture/ (arch -> architecture)
device/ -> devices/ (what /system/devices reflects)
boot/ -> /boot (the loaders, top level)
- sbin/ -> split by role:
init, vfs -> system/services/<name>/<name>.zig
hpetd, busd -> system/drivers/<name>/<name>.zig
vfs-test -> system/services/vfs/vfs-test.zig (inside the vfs project)
- lib/ -> library/runtime/ (room for other libraries beside runtime)
The VFS wire protocol becomes its own module, system/services/vfs/protocol.zig
("vfs-protocol"): the vfs sub-project exposes its interface, and the runtime's
file layer imports it by name. First instance of the "protocol module" pattern
(docs/driver-model.md); usb/block will expose theirs the same way.
Also: fix a naming-standard violation in the protocol — Op -> Operation (and
req -> request, _pad -> _padding). Docs updated: /system/services added to the
FHS doc, a repository-layout section added to the docs index, and stale source
paths swept across comments and docs.
Runtime boot paths are unchanged (the bootloader still loads /sbin/init);
aligning the runtime filesystem to the FHS is a separate follow-up. Suite 35/35
plus host tests green.
Two driver-model milestones plus a tree-wide naming pass. Suite 35/35
(QEMU) + host tests green.
M11 — IRQ-as-IPC. A ring-3 driver now sleeps until its device interrupts
it. New src/kernel/irq.zig: per-GSI endpoint bindings, comptime per-vector
trampolines, dispatch = mask GSI -> LAPIC EOI -> notifyLocked, all under one
lock region. irq_bind/irq_ack syscalls, gated by the device claim like
mmio_map. interruptDispatch no longer EOIs — each handler owns its EOI,
because a level line must be masked before it is acknowledged (irq_ack is
the unmask). Bindings are keyed on the owning task and released on exit
(a shared endpoint's siblings survive). hpetd rewritten interrupt-driven.
Tests: hpet (rewritten, reads back the I/O APIC routing) and irqfree.
M12 — bus drivers. DeviceDesc gains a parent, making the device table a
tree. dev_register (device_register) lets a process publish children below
a device it claimed; the kernel enforces resource containment (a child's
resources must nest in its parent's), so a descriptor can't fabricate a
window over kernel RAM. Descriptor copied in via copyFromUser (physmap
walk — an unmapped user pointer fails the call instead of faulting the
kernel). Per-parent child cap bounds table exhaustion. sbin/busd.zig is a
worked bus driver. Test: bus.
Naming — per docs/coding-standards.md: non-acronym abbreviations spelled
out (message, descriptor, device_service, scheduler, runtime, physical,
interpreter, ...); acronyms kept (IPC, MMIO, DMA, HCD, ...); files are
kebab-case (ipc-synchronous.zig, device-service.zig, vfs-protocol.zig, ...).
Exceptions: POSIX/C ABI names and Zig idioms (init/len/ptr) kept. Module
collisions resolved by specific naming (config -> parameters, device.zig
alias -> device_model). AML op/Op disambiguated: op = opcode, Op =
operation; per-opcode parse handlers renamed opX -> parseX.
New driver docs: drivers.md, driver-model.md (bus/class/HCD shapes + the
proposed M13–M16 ABI), coding-standards.md.
Start the user-space driver track (VFS + IPC + heap). This lays the
process/syscall foundation the runtime heap will grow on.
- Rename usermode.zig -> process.zig; drop the retired hello/ping blob and
its `user` test (subsumed by the real /sbin/init exerciser). Keep the
isolation-proof pf blob and the user-pf test.
- Add danos.Syscall as the single source of truth for syscall numbers, shared
by the kernel dispatcher and (later) the user runtime lib. Dispatch on the
enum. New calls: 1=yield, 4=mmap, 5=munmap. Widen debug_write's bounds
check to the whole user low half so heap buffers are writable.
- mmap grants zeroed RW+NX pages from a per-process bump arena
(Task.heap_next, PML4[224] above image+stack); munmap frees the frames.
Add paging.translateIn / unmapInto (+ arch.translate / unmapUserPageInto)
as the primitives munmap and future cross-AS copies need.
- New `usermem` test: grant three pages into a fresh AS, translate them,
release via the munmap path, tear down, and assert no frames leak.
Suite 28/28 (user -> usermem).
init is now a real scheduled ring-3 process: it prints a heartbeat and
sleeps, forever. The kernel spawns it at boot via spawnProcess (its own
address space, preemption on) and the boot context drops to idle — the
system's steady state is "kernel idle, init alive in ring 3", beating
~1 Hz. Adds the sleep(ms) syscall. The init/process tests are reshaped
around the heartbeat (repeated syscalls, still-alive, two concurrent
processes on distinct address spaces). Pins init to LLD and folds the
.large code model's .ltext/.lrodata/.ldata into the linker script so its
code segment is R+X. Removes the now-dead borrowed-thread ELF loader
(runInitElf); spawnProcess is the one path. Docs updated. Suite 28/28.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Per-process address spaces (AddressSpace = a PML4 with an empty user
half and the shared kernel half copied in; create/destroy in paging.zig)
with CR3 switched on context switch and TSS.rsp0/kernel_rsp published per
switch. The GS base now points at an arch per-CPU block and every ring
transition observes the swapgs discipline, so a ring-3 `mov %ax,%gs` can
no longer poison per-CPU access. syscall/sysret is the primary user entry
(int 0x80 kept as a test path); one handler, installed once at boot,
serves both and dispatches on whether the caller is a scheduled process
or a borrowed test thread. spawnProcess loads an ELF into a fresh address
space and schedules it; exit frees the address space after switching to
the kernel tables. New `process` test: init runs twice as a real process
(create/exit/recreate) on its own page tables, coexisting with a kernel
task under preemption. Suite 28/28.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
paging.init now maps only the physmap, the framebuffer/LAPIC windows,
and the kernel's own segments; the entire low canonical half is left to
user space. Every higher-half PML4 entry is pre-created so a per-process
address space can share the kernel half by copying PML4[256..512), with
an assert against late top-half entries and a 4 GiB guard on
pre-switch table frames. The AP trampoline's low identity page is now
created transiently by arm() and unmapped by disarm(); startAp asserts
the page-table root is 32-bit addressable. Docs (paging.md) updated.
Suite 27/27; 4-core normal boot reaches /sbin/init.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Ring 3 works: user GDT descriptors (sysret-ready layout), TSS.rsp0,
U/S-bit user mappings (W^X preserved), an int 0x80 syscall gate with a
mutable trap frame, and a setjmp-style enter/exit path. /sbin/init is a
real freestanding Zig binary built from sbin/, shipped on the ESP,
loaded by the bootloader (BootInfo.init_base/len), validated and mapped
by an in-kernel user-ELF loader, and run at CPL 3 — syscalls: exit,
ping, write. Tests: user, user-pf (U/S isolation proof, error code
0x5), init. Suite 27/27.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Hoist max_cpus into danos (root.zig), raise it to 128, and stop reserving a static [max_cpus][16 KiB] IST array. Only the BSP's IST stack is static (needed before the allocator); each AP's is heap-allocated when it comes online. Shrinks the kernel from ~1.1 MB to ~139 KiB in ReleaseSmall.
fault-ap-df pins a #DF to an AP (its own IST must catch it); affinity checks a pinned task never migrates. onException now names the core, so an AP fault is attributed and shown contained. Both teeth-checked.
spawnOn(entry, priority, cpu) routes to a per-core pinned queue, merged with the global queue at O(1) selection. Falls back to unpinned for an offline/invalid core.
Arm the low frame (copy blob, make executable) only while a core climbs, then zero it and restore RW+NX; the frame stays reserved so cores can be re-woken. startSecondary is one re-runnable attempt; boot retries a non-responding core 3x. Validated the retry path by forcing a first-attempt failure.
Daniel Samson
Daniel Samson