The last claimant. The kernel seeds the acpi-tables node, so it sits in the
same boot snapshot the manager already scans to find the PCI host bridge —
there was never a bootstrap problem, only a lookup nobody had written. The
manager claims it and names it in the spawn; the service stops claiming.
Every driver in the system now receives its hardware rather than taking it.
Two failures on the way, both mine. addDriver puts the device id in argv[1],
and the acpi service read argv[1] as a self-verify device-count floor — so
handed device 7 it decided it was in test mode, printed "acpi-parse: ok",
and never reported a device. The test argument is now floor:N, which a bare
id cannot be mistaken for.
And acpi-parse spawns the service directly rather than through the manager,
so nothing handed it the node. That test now claims and transfers it exactly
as the manager does, which is the right shape: the test plays the manager's
role instead of the service reaching for hardware.
device_claim now has two callers left: the manager, which is the acquirer
and should have it, and the display service's GOP path. That is recorded as
question 10 — the framebuffer is not a device, so the answer is likely that
it leaves the device table rather than being exempted from its rules.
Suite 118/118.
The 8042 is a single controller described by two ACPI nodes — PNP0303
carries the 0x60/0x64 ports, PNP0F13 is the mouse — so it cannot be split
across two processes without them fighting over the same registers. That is
why ps2-bus is a singleton, and why it used to find and claim both nodes
itself.
The manager now gives it every matching node instead. The keyboard node
rides the spawn, because it holds the ports and is needed immediately; the
mouse node is transferred to the already-running instance. Late arrival is
safe here and the ordering is natural rather than lucky: the mouse is not
touched until after the controller handshakes and identify. Measured, the
handover lands at 0.336 and the driver reaches the mouse at 0.456.
Because the count is however many matched, a machine with no PS/2 ports or
only one works without a special case — which matters, since the bus is
mostly emulated now and machines vary.
ps2-bus claims nothing. irq_bind on the mouse node is the proof it holds it:
that call is ownership-gated, so a failure means the handover did not land
rather than a hardware fault, and the log says so.
acpi-ps2 asserts both delegations with the spawned device backreferenced, so
the node that rides the spawn must be the one the driver was spawned for.
Disabling the second delegation fails it.
Two things worth recording. The first discrimination patch was not valid Zig,
so nothing ran and a stale binary reported a pass — checked the build before
believing it. And with the second delegation disabled, acpi-ps2 fails while
input still passes: the mouse works without its IRQ binding, so exactly one
case covers that path.
Suite 118/118.
The second invented ceiling. It was written to stop a driver looping
device_register and exhausting a shared table — but there is no shared table
to exhaust any more, and each registrar already has its own allowance, so a
runaway costs only itself.
It never bounded a determined caller in the first place: 16 children per
parent, and nothing stopped it claiming more parents. What it reliably did
was refuse a real PCI bus with more than 16 functions, which is how an AMD
Ryzen booted with a working display, no USB and no storage.
The constant, its check, and the now-unused childCount all go. TooManyChildren
survives with one meaning instead of two: the caller is at its per-registrar
allowance.
This was unblocked from the moment D9 landed. The plan said so — "once the
quota exists the per-parent cap is redundant whether or not D6 has landed" —
in the same edit that left the step tagged "blocked on D6". Three iterations
were then spent re-reading that tag instead of the sentence beside it.
The containment test now asserts 64 children under one parent, four times
the old ceiling; restoring the cap fails it.
Suite 118/118.
reassign was added at D4 to fix a regression and has been proven only
indirectly since — three IOMMU+USB cases going green. That covered the
visible symptom (a driver's DMA rings unbound) and neither of the latent
ones: the confinement still naming the giver, so the giver's death would
tear down a domain a live driver was using, and the receiver's death would
leave one behind. Those are now asserted.
confinementOwner exposes the record's owner so the suite can see it. The
sequence is the delegation in miniature: unconfined, confine as this task,
reassign to another, confirm the new holder owns it and the old one does
not, then kill the new holder and confirm the domain goes with it.
Two attempts at this test could not have failed. The first found no PCI
function to confine — pciAddressOf needs a pci_device entry and this case
runs no pci-bus — so every assertion skipped silently while the case stayed
green. It now synthesizes a function the way pci-bus does, a 4 KiB config
window inside the bridge's ECAM, and asserts that precondition explicitly so
a skip is a failure.
Verified to discriminate: making reassign a no-op flips three assertions,
including the domain surviving its holder's death.
Suite 118/118.
Third driver converted. It no longer claims the id from argv[1] — the
manager holds the device and names it in the call that creates the process,
so it is held before the driver's first instruction.
display-reattach is the case that matters here: it kills the driver and
watches the compositor re-attach to the fresh scanout. It passes, so the
restart path survives the fused grant — the manager re-takes the device when
the driver dies and hands it to the replacement.
ps2-bus and discovery are NOT converted, and the reason is recorded as open
question 9 rather than worked around. Both need a device nobody assigned
them. ps2-bus ignores its argv[1] entirely: it finds the controller by
walking the table for PNP0303, then claims a second device, the PNP0F13
mouse node, which it also finds itself — so it holds two devices and was
assigned at most one, while system_spawn carries one. discovery claims the
acpi-tables node it locates itself, because it is what produces the device
tree and there is nothing to assign at that point.
One thing worth checking before designing an answer: devices.csv maps both
PS/2 hardware ids to ps2-bus, so the manager may already be spawning two
instances where the driver expects one. If so the fix is smaller than it
looks.
D6 stays blocked — closing device_claim with these two still depending on it
would stop the machine booting.
Suite 118/118.
Delegation moves out of onHello and into the spawn itself. The manager holds
the hardware and names it in the call that creates the driver; the kernel
checks the device is the caller's to give, then hands it over as part of
making the child.
The reason is the window. A transfer after spawning always leaves an
interval in which the child is running and does not yet hold its device. It
would close on QEMU every time and open occasionally on a machine with
different core counts and timing — the exact failure shape this track exists
to delete, and not one worth introducing while removing the others. Fused
into the spawn there is no interval: the child does not exist until it holds
the device.
Ownership is checked BEFORE the child is created, so a refusal leaves
nothing running rather than a driver without the hardware it was spawned
for. The IOMMU confinement moves with the device, as it does on the transfer
path. systemCall6 is added for the sixth argument; r9 was free, and abi
gains a no_device sentinel matching the protocol's.
No driver had to change to receive a device, which is what makes this
better than requiring every driver to hello: ps2-bus keeps its legacy
status, and discovery — which has no assignment at all, since it is what
produces the device tree — is unaffected.
The attacker fixture now tries the spawn as a back door: name someone else's
device, and both the spawn and any child must be refused. Verifying that
assertion exposed a bug in the fixture itself. The kernel case's pass marker
was "device-authority: ok", which matches the FIRST per-assertion line, so
its wait loop exited before any failure was printed — the case would have
passed with failures in it, and had been able to since D2. The verdict lines
now carry a distinct VERDICT prefix, and with the ownership check removed
the case genuinely fails. A green test that cannot go red is worse than no
test.
Suite 118/118.
Questions 6 and 7 both dissolved on inspection, so what was left was only
where the grant is delivered. Three candidates: transfer after spawn, every
driver hellos, or fuse the device into system_spawn.
Take the third. The manager cannot transfer before the child exists, so a
separate transfer always leaves a window in which the child is running and
does not yet hold its device. That window would close on QEMU every time and
open occasionally on a machine with different timing — the exact failure
shape this track exists to delete, and not worth introducing while removing
the others. Fusing it into the spawn removes the window by construction: the
child does not exist until it holds the device. It adds no knowledge to the
kernel, only atomicity — the same rule, you may give away what you hold,
fused with the call that creates the recipient. system_spawn uses five of
six argument registers, and no_device is already the sentinel.
Making every driver hello is a good idea on its own merits — uniform
liveness, the deadline applied to all rather than some, and the
speaks_protocol two-class split leaving the manager, since a wedged ps2-bus
is invisible to its supervisor today. Kept as its own step so grant delivery
does not force it.
Order is now D0 -> D5 -> D6 -> D8, which deletes maximum_children_per_parent.
Only D7 remains blocked, on question 8, and it is needed for neither ceiling.
Question 7 read the comment "Best-effort: standalone bring-up has no
manager" as a boot path that delegation would delete. It is not one.
virtio-gpu's main requires argv[1] and exits without it, and the only source
of that argument is the device manager: ACPI, then pci-bus reports the
function, then devices.csv matches 1AF4:1050, then the manager spawns the
driver with the id. Without a manager the driver prints and returns, and
never reaches the hello at all.
The comment is about resilience, not boot: the hello is best-effort so a
driver whose manager has died keeps serving, which device-manager.md states
outright.
So the question dissolves. The residual is one narrow window — the manager
spawns a driver and dies before transferring — where today the driver could
still claim because claiming is free-for-all, and after D6 it would exit for
the restarted manager to respawn. That is the better behaviour: a driver
holding hardware nobody assigned it is exactly what D6 exists to stop.
Taken with the transfer-at-spawn delivery point, D5, D6 and D8 are
unblocked. D7 remains blocked on question 8, and is not needed for either
ceiling.
Question 6 lumped the acpi service in with ps2-bus as "a claimant that does
not hello". That was wrong about what it is. The manager starts it as
addDriver("discovery", no_device, false): the discovery service, one per
firmware, with no device assignment at all, which finds and claims the
acpi-tables node itself because it is the thing that produces the device
tree. The manager cannot hand it a device — at that point there is nothing
to match against.
So its question is not whether it should hello. It is whether the bootstrap
is exempt from D6, or whether the manager claims the acpi-tables node and
passes it on.
Recorded alongside it: a delivery point that needs no hello is already
implied by the code, since spawnSupervised returns the child's pid and the
manager could transfer immediately after spawn. If that is acceptable,
question 6 largely dissolves — ps2-bus would not need to leave "legacy"
either. The cost is ordering: the child may reach for the device before the
transfer lands, where hello guarantees it cannot because the child is the
one asking. That trade is the real decision, and it is now written down as
such.
maximum_devices = 64 is gone. It was a guess about someone else's computer,
and because it was shared, one driver's enumeration starved every other —
which is how an AMD Ryzen booted with a working display, no USB and no
storage. The table now grows from the kernel heap. It was always built after
heap.init; nothing ever prevented this except it having been written static
first.
What replaces it is an allowance charged to the registrar, so a driver
looping device_register exhausts its own and every other driver carries on.
It is declared as what it is — a runaway detector, NOT a security boundary.
A quota generous enough never to bite a real machine is still generous
enough to be unpleasant, and it is not trying to be the defence; delegation
is. What this catches is a legitimate driver in a loop, early, attributably,
and without collateral. Reaching 4096 is a bug report, not a tuning request.
The initial block is 8, deliberately small. Sizing it for a typical machine
would mean the growth path never ran on the hardware we test on and only
woke up on someone else's larger machine — the exact failure shape this
track exists to stop. At 8 it grows several times every boot; disabling
growth now fails the suite with the HPET not fitting, which is the Ryzen
failure in miniature.
The comptime coupling assert added earlier fired, and was right to. confined
(one slot per device id) and domains (the IOMMU's own translation pool) were
sized by the same constant only because device ids happened to stop at 64
too. Two unrelated quantities: confined now grows with the device table,
while maximum_domains stays as the hardware's number — both VT-d and AMD-Vi
report how many domains they support, and reading it is phase 4. The assert
existed for exactly this and did its job.
Suite 118/118.
D7 would remove zero-resource devices from the kernel. It cannot proceed
because nothing else mints their ids. A USB interface registers with
resource_count = 0, and the id device_register hands back is load-bearing in
three places: the child_added packet's target, the class driver's argv[1],
and the device_token of the usb-transfer WIRE protocol — so the id space is
visible on the wire, not merely internal. usb-xhci-bus records the fourth
constraint itself: the kernel's idempotency is what makes the same port and
interface map back to the same id across a bus restart, which is what stops
a respawned bus spawning duplicate class drivers.
Moving that out means answering who mints the id, how it survives a bus
restart, how it survives a manager restart, and whether device_token changes
meaning. A design step, not a mechanical one.
D8 is reordered to run after D9, correcting the original sequencing. D8's
justification was that the authorisation the per-parent cap stood in for now
exists — but D6 is blocked, so it does not, and deleting the shared cap now
would reopen the exhaustion hole it was written for. D9's per-holder quota
closes that hole independently of authorisation, and closes it better: a
rogue exhausts its own allowance instead of the table everyone shares. Once
the quota exists the per-parent cap is redundant either way.
D9 also no longer depends on D7. Its rationale was that zero-resource
children are the case that sidesteps containment — true, but a per-holder
quota bounds them as well as anything else, because it counts entries per
holder rather than per parent.
pci-bus joins usb-xhci-bus in receiving its device from the manager rather
than claiming the id it found in argv[1]. Its hello moves ahead of the ECAM
mapping, since that is where the bridge now arrives, and its hello was
already mandatory so nothing about its failure behaviour changes.
isDelegated compared whole strings, which silently missed this driver: the
manager records the boot-snapshot match as the bare "pci-bus" and a
devices.csv match as the full "/system/drivers/pci-bus". pci-bus was then
neither claiming nor delegated and died on "ECAM mmio_map failed". It now
matches on the last path component. Reintroducing the whole-string compare
breaks usb-xhci-bus instead of pci-bus — the two spellings swap which driver
loses — so usb-hid is the case that catches it, not pci-scan.
pci-scan asserts the delegation on the initial bring-up AND after the
restart drill, with the device id backreferenced so both must name the same
device. That is what proves the manager re-takes a device when its driver
dies and hands it to the replacement, which is the property the whole
supervision design rests on.
The remaining three claimants are NOT converted, and the plan records why
rather than working around it. ps2-bus and the acpi service never hello at
all, which device-manager.md states deliberately ("legacy drivers ... not
yet required to hello"), so delegating to them means either promoting them
out of legacy or giving the grant a delivery point that is not hello.
virtio-gpu hellos best-effort by design — "standalone bring-up has no
manager" — and delegation would make it mandatory. Both are decisions, not
mechanical steps.
Consequence: D6 is blocked, because device_claim cannot be closed off while
three claimants still depend on it. D7-D9 are unaffected — they concern what
the kernel stores and how its table is sized.
Suite 118/118.
The first driver to stop claiming its own hardware. The device manager holds
the controller and transfers it in the hello reply, so its matching becomes
authoritative instead of advisory — until now the driver claimed the id it
found in argv[1], and any process could have claimed the same integer first.
The manager claims before it spawns, so there is no window in which anything
else could take the device, and transfers in onHello using invocation.sender
— the kernel-stamped task id, which cannot be forged by the caller. hello is
synchronous, so the transfer has completed before the reply lands: no gap
between being told yes and holding the thing.
usb-xhci-bus's hello moves from after controller bring-up to before anything
that needs the device, which is the bring-up reorder the design predicted.
It is the first member of an explicit delegated set, so every unconverted
driver keeps claiming exactly as before and the suite stays green; the set
and device_claim both go at D6. D3 and D4 could not be separated and the
plan records why: the moment the manager claims, any driver still calling
device_claim is refused, and D3 applied to nothing changes no behaviour and
cannot be tested.
This step introduced a regression and the incremental conversion is what
caught it. confineDevice runs inside systemDeviceClaim, so a device arriving
by transfer was never confined for its new owner. Three IOMMU+USB cases
failed on the driver's DMA rings going unbound, and two worse consequences
were latent: a manager death would have torn down a domain a live driver was
using, and a driver death would have leaked one. iommu.reassign now moves
the confinement with the device, keeping the domain and its attachment
intact so it never translates through nothing. Converting all five drivers
at once would have produced the same three failures with five suspects.
A log line of mine claimed "holding controller device N" before anything
verified it — it printed even in the failure case, where the driver held
nothing. Reworded to state only what is known there: where the registers
are.
usb-hid asserts the delegation with the device id backreferenced, so the id
delegated and the id the driver ends up with must match. Emptying the
delegated set fails it with "hello acknowledged" then "mmio_map failed".
usb-hub failed once in a full run and has passed six times since (four
isolated, two full) — recorded in the plan as a suspected instance of the
known intermittent AP fault, not dismissed, since this step did shift boot
timing.
Suite 118/118.
The audit's sharpest finding was structural, not a bug: a fully green suite
had hidden six real defects because it contains no attacker. Every device
case asserts that a driver handed its own hardware can drive it. None asked
what a process handed NOTHING can do.
device-authority-test is that process. It is spawned with no device and
asserts what it therefore cannot do: it cannot give away a device another
task holds, nor a free one, because the kernel's rule is that you may give
away what you hold and the device's state is irrelevant to a process holding
nothing. Asserted across every device the machine actually has, so it cannot
pass by accident of which one happened to be free at boot — six on QEMU,
none of them its.
A positive control runs first. device_enumerate works from this process, so
the refusals below it are decisions rather than a syscall path that is
simply broken here; without it, "everything failed" would read identically
to "the assertions are meaningless". A nonexistent device is refused as
NoSuchDevice rather than NotHeld, because a refusal that cannot name its own
rule is what cost a debugging session on the Ryzen.
What it deliberately does not assert, and says so in its header:
device_claim is still first-come-first-served at this point in the run. That
is the hole D6 closes, and the claim half of the invariant joins this
fixture then. Asserting it now would be writing a test that documents the
bug.
Verified to discriminate: removing the holder check flips "every transfer by
a non-holder is refused" while the positive control keeps passing.
Suite 117 -> 118.
The mechanism behind delegation, which device-manager.md named as the step
after hello: the device manager claims what discovery seeded and hands each
device to the driver it matched, so assignment stops being
first-come-first-served.
It is a MOVE, not a copy. A claim is exclusive (driver-model.md, invariant
1), so the giver stops holding the device the instant the receiver starts.
That is why this is a new syscall rather than the M13 capability path, where
a passed handle is shared refcounted — exclusivity cannot be expressed that
way.
The kernel's whole rule is that you may give away what you hold. It has no
notion of which task is the device manager and deliberately gains none: a
binary name inside the kernel is not something that cannot safely live in
user space. A recipient that does not exist is refused, because a device
moved to nobody would be unreachable for the rest of the boot — nothing
un-holds a device but task death.
Three errnos, each naming its own rule: ENODEV no such device, EPERM you do
not hold it, ESRCH no such recipient.
Nothing uses it yet. The five claimants move across one at a time in D4-D5,
so the suite stays green throughout and a regression names the driver that
caused it.
Ten assertions, verified to discriminate: removing the ownership check flips
four of them, including the giveaway that an illegal transfer then blocks
the legitimate claim behind it.
Suite 116 -> 117.
device-authority.md is rewritten as an implementation design rather than a
rival to device-manager.md. The what was already settled there in 2026-07:
structure in the manager, authority in the kernel, and delegation as the
step after hello. This is the how, plus the two decisions that paragraph
leaves open.
Decision 1: the manager claims, it is not granted. device-manager.md says
"claims (or is granted)"; claiming wins because the manager runs before any
driver exists and takes the seeded devices unopposed, leaving nothing unheld
to race for. One new call, device_transfer(device_id, task_id), checks only
that the caller holds the device — no names in the kernel, no attestation.
The alternative put a binary path inside the kernel, and the kernel should
hold only what cannot safely live in user space. The residual is stated
rather than hidden: authority rests on the manager claiming first, which
init.csv makes an operator-visible ordering rather than an attacker-
controlled one, and the enforced version arrives with the spawn capability
drivers.md already names as missing.
Decision 2: the kernel stops holding inventory. It reads three things out of
a descriptor — physical ranges, interrupt numbers, one PCI BDF — and stores
the rest only so device_enumerate can hand it back. Devices with no
resources leave the kernel entirely: a USB device conveys no mapping
authority, so there is nothing to enforce. That is also the case which
sidesteps containment, and therefore the reason a shared cap existed.
Decision 3: no shared ceiling. The table becomes dynamic — it is built after
heap.init, so nothing ever prevented it — and the two invented numbers go. A
per-holder quota replaces them, because dynamic storage with no bound moves
the ceiling to the kernel heap, which is shared and fatal rather than
partial. A bound charged to whoever caused it is isolation.
Two earlier drafts of this document are gone: one gave init the root grants,
the other proposed extracting a firmware-framebuffer driver. Both were wrong
and both are recorded as wrong in the run plan's settled list — the
framebuffer is not a device, it is where pixels go until a real display
driver announces itself.
Run 2 is nine steps, ordered so the suite stays green throughout: build and
prove the transfer mechanism, move the five claimants across one at a time,
then the flag day, then the inventory, then the ceilings.
The design still had the display service receiving a device grant, which
keeps the wrong layering and just moves who hands it over. Drivers talk
hardware. Services talk to no hardware at all — they receive device events
and send data and commands over a protocol, which is the OS abstraction
between them.
So the answer to "which services need device grants" is none, and the design
collapses: every holder of a device is a driver, and every driver is spawned
by the device manager, which is what grants it. No exceptions to
accommodate.
display already shows both halves. Its VirtioGpu backend speaks
scanout-protocol over an IPC handle and touches no device — the driver holds
the hardware, the service speaks to it, and it works today. Its Gop backend
calls device.enumerate, device.claim and device.mmioMap: the service
reaching into hardware itself, because the firmware framebuffer has no
driver to talk to. That is the only such case in the tree; every other
claimant is a device-manager child spawned with its device id.
A firmware-framebuffer driver is therefore a prerequisite of this phase
rather than a consequence. It is small — hold the display node, map the
framebuffer, serve the same scanout-protocol virtio-gpu already serves — and
the compositor needs no new path, since not caring which backend is behind
it is what it was designed for.
Two earlier drafts are recorded as wrong in the document: display asking the
device manager for a grant, and before that init minting grants because init
starts display. Both accommodated an exception instead of removing it.
The design had init minting device grants and handing them on, because init
is what starts display and display claims the framebuffer. That was the
wrong shape. init has nothing to do with devices: it is the first process
and it starts the rest of the system. The device manager owns hardware.
So the kernel mints the root grants to the device manager, and display asks
the manager for its framebuffer like any other driver. The exception that
drove the earlier draft disappears instead of being accommodated — one
authority for hardware, not two.
The kernel recognises the manager by the chain attestation the security
track already settled on: the binary path it stamped itself, AND a
supervisor of PID 1. Path alone is forgeable, because system_spawn is
deliberately ungated and any process may spawn any bundled binary — but a
rogue copy's supervisor is the rogue, and PID 1 is the kernel's own first
process.
Costs restated for the corrected model. display gains a boot-order
dependency on the manager, mitigated by its existing ability to attach a
backend late. The kernel gains one piece of knowledge about one binary,
which is the minimum: authority has to enter somewhere, and the
alternatives are a file parser in the kernel or first-to-ask, which is the
hole again.
The phase 2 design called devices.csv "policy". It is not. The CSV is
configuration — declarative data an operator edits. The policy is the
component that decides using it: the device manager matching a device to a
driver, init reading protocol.csv and granting a binding. The kernel holds
mechanism, the check that a grant exists before a mapping is made.
Corrected where the document conflated them, and the three are now tabulated
so the rest of the design can lean on the distinction.
It also sharpens the open question. It was "manifest or not"; it is really
whether the grant list should be configuration from the start, or whether
init should hand the device manager the root grants and let it match by the
devices.csv it already reads. A device-grant file adds a second place an
operator must keep correct, and earns itself only when something needs to
differ from "the manager gets the hardware" — holding a device back for a
test or a bare-metal driver, say.
device_claim checks that a device exists and is free. That is all. Any
process may claim any unclaimed device, and a claim is what gates mmio_map
and irq_bind — a licence to map physical memory and take interrupts. The
device manager's matching is real but advisory: it spawns a driver with the
device id in argv[1] and nothing binds that decision to the kernel's grant.
maximum_children_per_parent is the visible cost. It exists because a driver
that claimed one device could loop device_register under it, and it is a
poor defence — an attacker burns 16 slots, claims another device, burns 16
more — while reliably refusing a legitimate PCI bus with more than 16
functions. Closing the hole is what retires the constant.
The principles decide the split: matching is policy and stays in the device
manager; enforcing that a driver holds only what it was given is security
and stays in the kernel, and is the whole of what the kernel needs.
The mechanism is decided by an awkward fact. Five of six claimants are
device-manager children spawned with their device id. display is not — init
spawns it, and it finds its framebuffer by enumerating for a display-class
node and claiming whatever it finds. So "record the device named at spawn"
closes the hole for five and breaks the sixth, and the sixth is not an
oddity to special-case: it shows authority must be delegable rather than
welded to the moment of spawn.
So: a device grant is a capability, minted by the kernel to init for the
devices firmware discovery found, delegated by init to the device manager
and to display, and passed by the manager to each driver it spawns. The
cap-passing path already exists and already carries shared memory and DMA
regions. init is already the grantor for /protocol, and protocol.csv already
records init granting the device manager its binding.
Costs named rather than buried: five drivers must hello before claiming
(pci-bus claims first today), and init grows a device role on top of the
protocol registry. A device-grant manifest mirroring protocol.csv would be
the natural symmetry and is deliberately not proposed yet.
Both device-setup paths issued a successful Enable Slot and then returned
null if allocateDevice failed, without disabling it. A slot the driver
forgets is one the controller never reissues, so each attempt lost one
permanently for the boot. The hub path did it with no log line at all.
Both now release the slot through a shared disableSlot, extracted from
tearDownDevice, and the hub path warns like the root-port path does.
tearDownDevice also now frees the interface list. That allocation arrived
with the previous commit, so an unplug would have leaked it — found while
reading the teardown path for this fix rather than by a test.
No regression test, and it is recorded as open question 5 rather than
implied. After the slot count became the controller's own figure, reaching
this path needs more devices than the controller has slots: QEMU offers four
against sixty-four. What was verified is that the new path RUNS correctly —
pinning tracking to 2 with four devices attached produced "port 6 setup: no
free device slot", the first two devices enumerated normally, and no Disable
Slot error or timeout appeared, which is how disableSlot reports failure.
Suite 116/116.
max_interfaces was 4. A composite device — a headset, a webcam with audio, a
dock, a multifunction printer — routinely has more, and the fifth did not
merely go missing. parseConfiguration's cap branch had no `else`, so when the
count was reached `current` kept pointing at interface 3 and the fifth
interface's endpoint descriptors were appended to interface 3's array. A
class driver bound to interface 3 could then be handed an endpoint belonging
to something else entirely, and subscribe or bulk-transfer on it. The
alternate-setting arm one line above cleared `current` correctly, which is
what the cap branch should have done.
Interfaces are now counted from the block in a first pass and allocated to
exactly that number, so the ceiling is bNumInterfaces' u8 — the USB
specification's. The missing `else` is added too, though after this the bug
is unreachable by construction: interface_count cannot reach interfaces.len
mid-parse when the list was sized from the same walk.
max_configured_endpoints was max_interfaces * max_endpoints_per_interface =
16, a derived guess that moved whenever either input moved. It is now 31,
which is the xHCI specification's own limit: a Device Context holds a slot
context plus at most 31 endpoint contexts, because the Context Entries field
addressing them is 5 bits.
max_endpoints_per_interface stays at 4 with its reason recorded — the
usb-transfer wire protocol reports exactly max_reported_endpoints (4) per
interface, so widening it alone would change nothing a class driver sees.
Lifting it is a protocol change.
No direct test, and that is written down as open question 5 rather than
glossed. The parser is pure and wants a host unit test, but
usb-xhci-library.zig imports memory, mmio and time so it cannot be a
standalone test root, and QEMU offers nothing that reaches the path — the
largest device available is usb-audio,multi=on at 2 interfaces and 211
bytes. The alternate-setting path that shares the same `current = null`
logic is exercised by that device.
Suite 116/116.
The driver read the first 512 bytes of a configuration block into a fixed
buffer and parsed those. The block's length is the device's own choice
(wTotalLength, a u16), so anything larger was silently cut: interfaces past
the cut did not exist as far as the host was concerned, while the
SET_CONFIGURATION that follows still configured the device for all of them.
A headset is 500-900 bytes, a UVC webcam 1-3 KB, a multifunction printer
600+.
Now allocated at the declared length, so the ceiling is the field's u16 —
the specification's number rather than one of ours. A block shorter than its
own 9-byte header is refused rather than trusted.
The bring-up line reports the declared length and the bytes actually read,
so a truncation can never again be invisible, and usb-large-descriptor
asserts they match with a backreference.
That case has an honest limit, recorded in its comment: QEMU cannot produce
a block over 512 bytes. The boot keyboard, mouse and stick are 34-44, and
the largest device available is usb-audio in multi-channel mode at 211 —
which is exactly why the suite never caught this, and why it cannot now
reproduce the original trigger. What it does catch is the class: any clamp
below the attached device's block fails it, verified by pinning the buffer
to 128 and watching "config block 211 bytes, read 128" turn the case red.
Suite 115 -> 116.
max_devices was 8, with the comment "QEMU presents a handful; a fuller
machine would grow this" — a number chosen against the test rig, waiting for
a real machine, which is the pattern docs/bounds-track-plan.md exists to
stop.
The driver already knew the true figure. It reads HCSPARAMS1.MaxSlots at
bring-up and writes it straight into op_config, so every slot the controller
offers has always been *enabled*; only the array tracking them was 8. QEMU's
xHCI reports 64, so seven eighths of the controller was live and invisible,
and the ninth device — a keyboard, mouse, webcam, headset, hub and two
sticks reach that without trying — disappeared on a hub-attached path that
logs nothing at all.
The array becomes a slice allocated from max_slots at bring-up. A controller
claiming zero slots cannot address anything, so that is now a dead
controller rather than an empty allocation failing mysteriously later. The
Device Context Base Address Array is a page, 511 usable entries, so it
already covered the 255-slot maximum.
The bring-up line reports both numbers, and usb-hid asserts they are equal
with a backreference rather than a magic number, so the test cannot drift
from the hardware. Pinning tracking back to 8 fails it: "64 slots,
tracking 8".
Suite 115/115.
The convention that tunables live in system/parameters.zig with their
reasoning attached predates this and got 2% compliance — 5 of 235. A
convention with no teeth is how a bare `const maximum_devices = 64` reached
an AMD desktop and cost it USB and storage. This is the same rule with a
gate behind it.
tools/check-bounds.py finds every bound-shaped declaration — a `maximum_*`
const with a literal value, or a type with a literal array length — and
requires the five-field block above it: what it counts, who decides its
size, what it protects, what happens at the limit, and how anyone finds out.
The at-limit vocabulary is closed: refuse, degrade, truncate, grow. There is
deliberately no way to spell "silent", no way to spell "drop", and nothing
meaning "allow", so the behaviours that did the damage cannot be written
down. Truncation is legal only carrying a marker the reader can see, which
is why klog_maximum_message qualifies and a USB descriptor cut at 512 bytes
does not.
An array length that names a declared bound is not itself a bound; only
literal lengths are flagged, which pushes ceilings toward having names.
The 273 that predate the rule are allowlisted so this lands without a
tree-wide sweep in front of it, and that list may only shrink: declaring a
bound means deleting its line, and the check fails on a stale entry too.
Nothing may be added.
Wired into `zig build test` and available alone as `zig build bounds`. Not
in the default build — it reads the whole tree, and a red bounds check
should not stop you booting a kernel.
Five are now declared rather than allowlisted. Writing them out is its own
argument: maximum_devices reads "protects: nothing — this is a sizing guess
about someone else's computer", and maximum_tasks now carries the fact that
it has been raised twice, each time by something that outgrew it.
Verified the gate refuses an undeclared bound, a declared one using
forbidden vocabulary, and an allowlist entry that has since been declared.
Suite 115/115.
The first step of the unattended run was "a dead task's registrations die
with its claims". Implementing it would have broken the restart path it was
meant to protect.
The broker keeps entries on purpose: they describe hardware, which did not go
away when a driver died. And device ids must stay stable across a bus
restart, because device-manager dedupes re-reports by device_id so a
restarted bus does not spawn a second driver instance — stability that comes
from the idempotency scan returning the existing id. Removing entries on
death would hand a restarted bus fresh ids and duplicate every driver.
Everything else a task holds is already reclaimed on every path out: IRQ
bindings, IOMMU domains, DMA regions, then its claims.
The leak the audit found is real but has two other sources: a device that
genuinely goes away has no retirement path, and a bus that enumerates
differently on restart strands its old entries. Both belong to the device
manager's inventory, and both need the id-stability question settled first —
tombstone-and-reuse aliases ids another process still holds, generation
tagging changes the id encoding, which is ABI. Recorded as an open question
rather than guessed at.
Six steps an agent can execute: reclamation, the bounds build check, and
the four user-space USB/xHCI bounds where the hardware already reports the
number we guessed.
Deliberately excluded: the authorisation gate and moving the device
inventory to the manager. Both decide whether the OS is secure and both are
a direction rather than a specification — what a device capability is, which
syscalls change, what replaces device_claim for its seven callers. They want
a design session, not an agent.
Three open questions are written down rather than guessed: device_enumerate
most likely narrows to the firmware-discovered roots rather than retiring
(the manager cannot ask itself for the PCI host bridge); a manager restart
has no re-enumerate handshake, so it comes back blind while its buses live;
and "add adversarial tests" is not executable until the attacks are named.
An AMD Ryzen booted to a working compositor with no USB and no storage,
and the log said only "register refused". A tree-wide audit of every
compile-time ceiling followed: 235 of them, 139 on quantities the machine
or a file decides rather than us, 5 documented anywhere, 171 silent when
reached. docs/fixed-bounds-audit.md has the inventory.
Errno attribution. The errno space was split between the kernel and the
envelope, free to drift; it is now one list in system/abi.zig, restated on
both sides, with a comptime check in library/device/driver where the two
halves are visible. device_register's six refusals and device_claim's three
are distinct codes, so a bus driver can say which rule stopped it, and
BadParent splits into NoSuchParent and NotYourParent. pci-bus reconciles
found against registered instead of counting refused functions as found.
Idempotency ordering. The child cap was checked before the identity match,
so a restarted bus was refused its own devices — the supervision restart the
system leans on ratcheted toward a degraded machine. A re-registration
consumes no slot and is now admitted first.
IOMMU fail-closed. confineDevice returned success for a device id past the
confinement table, leaving the device outside every domain while the caller
believed it confined — unreachable only while ids stop at 64, which both the
inventory move and a hardware-reported domain count would change. It refuses
now, and the coupling to the broker's device cap is a comptime assert rather
than a sentence in a comment.
PCI apertures. The bridge's MMIO apertures are derived from the holes in the
firmware memory map, and the derivation copied sub-4 GiB entries into a
fixed [64] array and skipped the rest. A skipped region is not merely lost:
the gap finder concludes it is free, so a real machine's 60-200 entry map
yields an aperture over live RAM, and containment then admits a child BAR
covering kernel memory. Rewritten to walk the map in place, with the hole
finder extracted as a pure function and driven by a synthetic 100-entry map
in a new test case. Both new tests were verified to fail on the old code.
parameters.zig gains the rationale it was missing and loses a stale sentence
pointing at the wrong file; vdso.md documents the errno space, including
EPEER, which had no written meaning anywhere.
docs/os-development/bounds.md is how a ceiling is declared from here.
docs/bounds-track-plan.md is the plan to remove the ones we invented.
Suite 114 -> 115.
A Ryzen 3 3200G triple-faulted on its first timer tick after reaching init.
Three defects in a chain, each hiding the one beneath it.
STAR's SYSRET base was 0x10, so SS came back as base+8 = 0x18 with RPL 0
while CS carried RPL 3. Intel ORs RPL 3 into SS on SYSRET; AMD only does so
for CS. Ring 3 ran fine — RPL is not checked on data access — and died the
moment an interrupt tried to IRETQ back, where SS.RPL must equal CS.RPL.
The base now carries the RPL (0x13), as Linux does.
Two fixes below it, both of which made the first one unreadable:
scheduler() read IA32_GS_BASE and dereferenced it without testing for zero,
so every fault reporter faulted in turn — a panic inside a panic, and the
machine reset before printing anything. Cast after the null test, plus a
re-entrancy guard in the panic handler.
NT is now masked in SFMASK alongside the rest, and isr.s exports
isr_return_iretq at the faulting instruction so a frame dump can say which
IRETQ died and print the CS/SS it was about to load. That dump is what
identified the RPL mismatch.
SMAP makes the rule the copy layer has followed since it was written into a
rule the hardware keeps. A ring-0 read or write of a user page now faults,
so any code that reaches for a user pointer directly fails the first time it
runs rather than the first time someone attacks it — and the suite becomes
the enforcement test, because every case exercises the kernel with the bit
on. Nothing had to be fixed to turn it on, which is the retrospective proof
that the nine stragglers converted earlier were all of them.
The interrupt entry needed one instruction first. Hardware does not clear
the alignment-check flag on its way into a handler, and ring 3 sets that
flag freely, so a process could have taken an interrupt with SMAP suspended
for the duration. The system call path was already covered — its flag mask
clears it — but the interrupt path needed a `clac`, which cannot simply be
assembled in: it is an invalid instruction on a processor without SMAP, and
danos boots on those too. So the entry ships as a three-byte NOP and is
patched at boot, through the physmap, because the kernel maps its own text
read-only.
The ordering that makes that safe is enforced rather than described: the
patch sets a flag, and no core will set the SMAP bit until it is true. A
translation that fails, or bytes that read back wrong through the address
they will actually be fetched from, leave the machine unhardened and saying
so — which is the same posture the IOMMU takes, and better than enforcing
over an entry path that cannot comply. The patch runs before interrupts are
enabled and before any second core exists; a comment says so, because the
three bytes pass through an encoding that must never be executed and a
future change that moves this later has to deal with that first.
Suite 114/114, with a case that reads a user page from ring 0 and requires
the fault, and the multi-core case asserting every core that ran work had
the bit — the same shape SMEP got, for the same reason: CR4 is per-core, and
a hardening is only as wide as its narrowest core.
SYSRETQ with a non-canonical RIP raises a general protection fault in ring
0 — on the kernel stack, an instruction after the swapgs that installed the
user's GS base. It is one of the better-known escalation primitives, and
ring 3 reaches it without any kernel bug at all: the processor saves the
address of the instruction after SYSCALL, so a program whose SYSCALL is the
last two bytes of the last canonical page returns to the first
non-canonical address. The new test does exactly that.
The exit path now sign-extends the return address from bit 47 and compares;
if the value changed, it returns through IRETQ instead, which commits the
privilege change before fetching the new address, so the fault arrives from
ring 3 and the process dies like any other. Four register-only operations
and a branch that a correct program can never take — it could not have
executed at a non-canonical address in the first place. Bit 47 is the right
pivot because danos builds four-level page tables and nothing sets the
five-level bit; a future port must move the pivot, and the comment says so.
SFMASK grows one bit while we are here. SYSCALL, unlike an interrupt gate,
does not clear the nested-task flag, so the kernel had been running every
system call with whatever ring 3 last chose — harmless while the only exit
was SYSRETQ, and a question worth not having now that one exit is IRETQ.
The kernel is never nested; ring 3 still gets its own flag back.
Suite 113/113. The new case asserts the refusal counter rather than the
dying process: the emulator we test on kills it either way, so only the
counter distinguishes a guard that ran from one that did not.
SMEP turns the classic escalation — divert kernel control flow into a page
the attacker wrote — from a silent takeover into an immediate fault with
the offending address in the log. The bit is per-core state, so it is set
where the syscall MSRs already are: in the per-CPU bring-up both the boot
processor and every application processor run on their way in. A core that
climbed the trampoline without it would be a hole no boot log would show,
which is why the SMP case now reads CR4 on each core it lands on and
requires every one of them to be hardened, not just the one that printed
the banner.
Enabling it that early is only safe because nothing ring 0 executes is
mapped for ring 3, and that had to be established rather than assumed:
kernel text carries only its ELF flags, the physmap is no-execute, the
trampoline page is mapped supervisor and the core running it has not
enabled the bit yet, and the boot processor turns it on while still on the
loader's tables — which map nothing user-accessible at all. The one
indirect call in the kernel takes a kernel address.
The CPUID probing that was scattered across the timer code becomes a small
shared helper, since the feature question is now asked from two places and
each wanted the same maximum-leaf guard. Absence is tolerated and reported,
like the IOMMU: danos still boots on a machine without the feature, and
says which one it is.
The test harness starts asking QEMU for a CPU that has the bit at all —
its default model has neither SMEP nor SMAP, so the code would otherwise
have been unreachable in every run. No case behaved differently under the
richer model.
Suite 112/112, with a new case that maps an executable user page, calls
into it from the kernel, and requires the fault the CPU is supposed to
raise.
Three services had each written the same thing and got it three different
ways: input polled the process list to notice a dead subscriber, and only
when someone else subscribed; the power service never noticed at all; the
device manager noticed drivers but not subscribers. The harness owns the
table now, driven by the events a protocol declares — it registers on the
reserved verb, frames each event once, posts to everyone interested without
waiting on any of them, and reclaims a slot when the kernel says its owner
died. Interest masks moved to the envelope, so a subscriber that wants only
mice asks the same way everywhere.
Two consequences the plan had not foreseen. The device manager now hears a
supervised child's death twice, once as its supervisor and once as a
subscriber, so restart backoff counted every crash twice and gave up after
half as many; it retires the id before counting. And the kernel's published
exit table had eight slots for what is now six subscriptions in a plain
boot, so it holds sixteen.
The other half is a hole the design named early and left standing: a
backend handed out a small integer and then honoured it from anyone. A
process that guessed a file's node id read another client's file; a display
layer had no owner at all, so any client could reconfigure or destroy any
layer; a USB device token was never checked against the client that opened
it. Each is now bound to the task that opened it, and a wrong owner gets
exactly what an unknown id gets — the refusal must not become the oracle
the identical answers elsewhere were designed to remove. Closing a file
changed with it: it used to succeed unconditionally, which would have told
a caller which ids existed.
Suite 111/111, with a new case in which one process holds a file and a
layer, hands both ids to a second process, and finds them untouched after
that process has tried everything with them.
These were the awkward ones. Each began with an operation packed into a
single byte — two of them with a version wedged in beside it — so there was
no wrapping them: the layouts had to be rebuilt. The device manager's own
enumerate and subscribe become the reserved verbs that mean the same thing
everywhere, its replies lose three status structs the envelope already
carries, and a device id becomes the packet's target. Power drops the
version it repeated on every request, because describe is the handshake,
and stops claiming a 64-byte ceiling it never needed for calls. USB moves a
control transfer's data to the packet tail in both directions, which makes
the status length the transferred length and retires a field that had been
saying the same thing twice.
The danger in this one was not the protocols but their readers. Init
recognised a power button by two bytes at the head of a message, the ACPI
service dispatched on the first byte, the xHCI driver read its operation
with a raw integer load, and the HID drivers reinterpreted a report
wholesale — none of which would have failed to compile once the layouts
moved. They would simply have stopped: no shutdown on the power button, no
reports from the keyboard. Every one of them now reads through the
generated types, and the shutdown gate that answers only a subscriber is
the same code it was.
Two sizes were decided by measuring rather than assuming. The child-added
message is both a request and the event broadcast to subscribers, and
alignment rounds it to 48 bytes, which puts its packet exactly on the
64-byte push floor — a test pins that, because a field added carelessly
would now overflow it. The interrupt report gives up eight bytes of inline
room to make space for the header; the two drivers that produce reports
send eight and four.
Suite 110/110.
The folded header stops being a rule in a document and becomes the layout
on the wire. Verbs number from sixteen, leaving describe, enumerate,
subscribe and unsubscribe reserved and answered the same way by every
provider — none of them writes a line to do it. What each protocol used to
carry in a field of its own now travels in the header: a vfs node and a
display layer are the packet's target, and a reply opens with a status the
envelope stamps rather than one each protocol spelled for itself.
Display gains the most. One forty-byte request had served eleven verbs, so
attach_scanout smuggled stride through x, refresh through y and format
through colour, and every coordinate crossed as a bitcast. Per-operation
structs end all three: the fields have their own names and their own signs,
and the tile payload grows to 224 bytes because the prefix shrank. Scanout
loses a message maximum of 64 it had no business declaring — it answers
calls, and the floor for a call is 256 — and virtio-gpu stops hard-coding
that number at its harness.
Two changes are semantic rather than notational. A directory now ends at an
entry with no name, because the fixed part of a reply always travels and a
zero-length reply no longer exists to mean anything. And input joins the
service harness, the last loop in the tree that answered no ping and heard
no terminate; its subscriber table, its pruning and its fan-out are the
same code, and a shutdown now asks it to stop instead of killing it.
A new conformance case reads the registry's own listing and asks every
protocol it finds for its name, its version and its verb count, then offers
a verb nobody defines and requires -ENOSYS — the envelope's promise,
checked against providers rather than against itself. What it cannot reach
in that boot it names on the serial line instead of passing quietly.
Suite 110/110.
The registry consults the open rows it has been parsing since P2, so
reaching a contract now takes a grant as well as a binding. A caller
without one is answered exactly as it would be for a name nobody ever
bound: same status, same empty reply, same absent capability, byte for
byte, and no log line on either path — klog_read is ungated, so a line on
one and not the other would be the oracle the design set out to remove.
Refusal and absence being one answer is what lets a supervisor later
narrow, fake or park a child's namespace without the child learning what
it was denied.
The manifest gains a third permission for a shape the plan did not
foresee: attestation is one hop, but the driver tree is three deep — the
PS/2 keyboard and mouse are spawned by ps2-bus, which the device manager
spawned — so no row could name them and PS/2 input would simply stop.
A supervise grant lets a delegate vouch for what its children *reach*,
never for what they claim; the bind path is untouched, and the laundering
deputy is still refused.
The review found the receive side of a rule this track had already
written down. Every process holds a sendable handle to the registrar —
resolve installs one for anyone who asks — and ipc_reply_wait never asked
who owned the endpoint, so a stranger could dequeue there: take the
provider endpoints riding bind requests, and answer other clients' opens
in the registrar's name. Receiving is the owner's privilege, like binding
a signal or a timer; sending remains anyone's.
Suite 109/109.
The plan is the only progress view from a plain main checkout, and the work
lands on a feature branch that reaches main a group at a time — so between
merges the file said less than was true. It now opens with where the work
is, which branch carries it, and what main itself holds, refreshed on main
after every phase rather than at merges alone.
A protocol is reached by name now, not by a compile-time integer. Init is
PID 1 and already knows which binary it started, so init serves /protocol
as a vfs backend: bind claims a contract with the provider's endpoint
attached, open answers with that endpoint as the reply's capability, and
readdir lists what is bound with the task and binary behind it. The kernel
reserves the prefix — nothing may mount over it, under it, or unmount it —
and ServiceId, ipc_register and ipc_lookup are gone, their syscall numbers
left vacant.
A bind is authorized by who the caller *is*: the kernel-stamped binary
together with the supervising task's identity, matched against
/system/configuration/protocol.csv. Identity, not spelling — spawn is
ungated, so an attacker can run any bundled binary, and a name-only rule
would have let it launder grants through an init of its own making. A name
a live process holds is refused to everyone else; a dead one's is released.
Three review rounds against a hostile ring-3 process found what 108 green
tests could not, because the suite contains no attacker. Publishing init's
supervision endpoint as the registry put PID 1's mailbox in every process's
hands, where two forged bytes reached the shutdown path: privileged traffic
is now believed only from the task that holds the contract it speaks for.
A capability arriving on a request outlived every path that ignored it,
one handle per call until the table was full — in init, and in the harness
ten services share — so the arriving capability is owned by the turn and
released unless a handler says otherwise. And the kernel let anyone holding
an endpoint handle aim signals, timers, exit notices and interrupts at it:
binding now requires having created it.
Suite 108/108. The new protocol-registry case asserts eleven properties,
each one an attack that must fail.
A protocol is now defined through envelope.Define rather than beside it.
Every packet begins with the same 16-byte prefix: a Header of {operation,
target} for a request or event, a Status for a reply. The header is folded
into each protocol's own layout, never stacked on top of it — Define walks
a spec's types and refuses one that carries a Header or Status field, so
the rule cannot be missed by reading. Operations number from 16, leaving
describe, enumerate, subscribe and unsubscribe reserved and answerable the
same way everywhere; describe the envelope answers itself, and an unknown
verb is -ENOSYS without a provider writing a line.
The size floor becomes something protocols can compile against: 256 bytes
for a call, 64 for a push, matching the kernel's own limits, checked at the
Define call with the prefix counted in. A protocol that outgrows either
fails the build with the reason and the remedy — shrink, or move the bulk
to shared memory, because packets never fragment. Events carry the header
too and number in their own space, so appending an operation can never
renumber a shipped event.
Beside it, the client half of the model: a Channel is what opening a
/protocol name yields, wrapping the endpoint capability so a program holds
a conversation rather than a handle. Nothing speaks through either yet —
the vfs protocol gains NodeKind.protocol and the convention that an open
reply may carry a capability, and that is the whole runtime change.
Suite 107/107, unchanged: P1 converts nothing.
Daniel Samson