SchOS/docs/architecture/ldls.md

# LDLS — running somebody else's drivers

SchOS cannot write a driver for every device that exists. No project can. The
only route to the hardware support an established system has is to run Linux's
drivers, and LDLS is what does that.

It works, and on more than one driver. `8139too` and `e1000` — unmodified GPL
kernel modules — load into a SchOS process, drive their hardware through
`/dev/pci`, and carry packets:

```
# ldlsd /lib/modules/8139too.ko /lib/modules/btf &
  kpi: the driver read a MAC off the hardware: 52:54:00:ab:cd:ef
  ndo_open returned 0
  carrying frames. the interface is ldls0.

# ping 10.9.0.2 4
64 bytes from 10.9.0.2: icmp_seq=1 time=30 ms
...
4 packets transmitted, 4 received, 0% packet loss
```

`10.9.0.0/24` is reachable only through `ldls0`. The address on it came from
DHCP, over the same driver.

The e1000 run is the more interesting of the two, because it is a **handover**:
the kernel was driving that NIC, and gave it up.

```
net: eth1 is gone; interface 1 removed
e1000: released to userspace; the interface is down
  rehomed 23 %gs per-CPU reference(s)
  calling probe, with 82 of 155 symbols implemented
  kpi: the driver read a MAC off the hardware: 52:54:00:12:34:57
net: ldls0 attached as interface 2, asking for an address
net: ldls0 10.9.0.15/255.255.255.0 via 10.9.0.2
```

Everything the second driver needed that the first did not is below, and the
pattern in it is worth naming up front: **none of it was a missing feature.**
Every one was a call that succeeded, returned zero, and did nothing.

---

## The rules it is built on

Five, and everything else is an implementation of them.

### 1. Nothing GPL ever enters an image we build

The driver is **the user's**: obtained by them, loaded on their machine. SchOS
ships the layer and never the thing it loads. `.gitignore` refuses `*.ko`,
`make ldls-stage` is the one command that puts a module into a local image and
is deliberately not part of `all`, and `make ldls-unstage` takes it out again.

[../licensing.md](../licensing.md) §6 is the whole argument. The short form is
that the combination happens in a user process on the user's machine, so no
combined work is distributed by anyone. Since SchOS moved to GPL-2.0-or-later
that is no longer the only way it *could* be done — it is the way it is done,
because a module we never ship is a module we never have to answer for.

### 2. It runs in a process, never in the kernel

A module linked into our kernel is the strongest derivative-work argument that
exists. A module combined with our shim inside a user process is a different
thing at arm's length — the position FreeBSD has shipped from for a decade.

And practically: this is foreign code. A driver we did not write, cannot fully
audit and did not choose the quality of should not be able to take the machine
down.

### 3. Facts, not expression

A symbol's signature and a struct's field offsets are facts a compiled driver
depends on. Comments, macro bodies and implementations are expression and are
never carried over. Every file in `ldls/kpi/` records where its shape came
from.

Mostly the question does not arise, because of rule 3½:

### 3½. The layouts are not in our source at all

A kernel built with `CONFIG_DEBUG_INFO_BTF` publishes
`/sys/kernel/btf/vmlinux`: every struct layout and every function prototype, as
generated type data describing the kernel that is installed. `ldls/btf.c` reads
it at load time, on the user's machine, from the user's own kernel.

```
$ ldls-types /sys/kernel/btf/vmlinux --need
  pci_driver   probe      offset    16  size 8
  pci_dev      resource   offset  1000  size 1088
  resource     start      offset     0  size 8
  all present
```

Nothing there is ours to have transcribed, and it is right for the kernel
actually present rather than one pinned by guessing.

### 4. One kernel version at a time

Linux states in writing that it has no stable in-kernel ABI. BTF mostly
dissolves this, because it describes the kernel that is there — but a kernel
without BTF needs hand-written layouts for a named version, and that is when
this rule bites.

### 5. The unresolved-symbol list is the specification

Do not guess the API surface. Load a real module and let the loader say what is
missing. Every symbol LDLS does not implement still *resolves*, to a stub that
reports itself and returns zero, so the driver carries on and the next thing it
wants becomes visible. The work queue keeps writing itself.

---

## The four pieces

### The loader — `ldls/ko.c`

A `.ko` is an ELF64 relocatable object. Index the sections, place the
allocatable ones, resolve the undefined symbols, apply the relocations. Written
from the ELF-64 gABI and the x86-64 psABI, both published specifications, so
there is no licence question in it at all.

Measured over a stock 6.18.46 tree: **2,440 modules, every one lays out and
relocates to completion**, five relocation types, nothing unhandled. That sweep
is also what retired the fixed section table — `nvidia-modeset` has 7,069
sections against a median of 50 — so the caller supplies it and there is no
limit to run into.

It is **freestanding**: no libc, no allocation, no I/O. The same source builds
as a host tool, where a format failure costs a second, and as the SchOS-side
host. That is the arrangement `mkschfs` and the kernel already share.

#### Three things it has to fix in the object

**The `%gs` stack guard.** Modules are built `-mstack-protector-guard-reg=gs`,
so the canary is fetched through the GS segment — an ABI between the kernel and
its own per-CPU area that a host which is not that kernel cannot provide.
`ko_defuse_gs()` rewrites those references to `xor %reg, %reg`. There are two
forms: the prologue *loads* the guard with `mov`, the epilogue *compares* it
with `sub` before a `jne`. `xor` is right for both — it zeroes the canary and
sets ZF, so the branch falls through as a passing check does. It checks every
site before patching any and refuses the whole operation on anything it does
not recognise.

**Every other per-CPU variable.** The guard turned out to be one instance of a
general problem. `current` is another — `mov %gs:const_current_task(%rip), %r8`
— and e1000 has 23 such references across four variables. Defusing does not
work here: the guard's *value* is arbitrary, and `current`'s is not, because
the driver dereferences what it reads.

What does work is that **a per-CPU variable in a single-threaded host is just a
variable**. `ko_fixup_percpu()` takes the `%gs` off, leaving the same opcode
reading the same width in the same direction, from storage the host supplies
per name. Eight bytes become seven and a `nop`. The displacement is measured
from the end of the *shortened* instruction, and the displacement is not always
last — `cmpb $0x0, %gs:softnet_data+0x75(%rip)` has an immediate after it — so
there is just enough of an x86 decoder in there to know how long the
instruction is. Anything it does not recognise is refused with the symbol
named, which is rule 5 again.

`ldlsd` hands out zeroed blobs, except for `current`, which points at the rest
of its own: a driver that walks into the task it thinks it has reads zeroes.

**The compiler's thunks.** `__x86_indirect_thunk_rN` and `__x86_return_thunk`
are retpolines, not API, and they are pure control flow — `jmp *%r8` and `ret`.
A stub cannot stand in for one, because a stub's whole behaviour is to log a
name and return, and what a thunk is *for* is the jump. `ldls/kpi/thunk.c` is
fifteen lines of assembly and is the single highest-value file in the layer.

That one cost a day. e1000 refills its receive ring through a function pointer
and the compiler made it a *tail* call:

```
mov  0x2a0(%rbx),%r8          ; adapter->alloc_rx_buf
...                           ; compute E1000_DESC_UNUSED(ring) == 255
jmp  __x86_indirect_thunk_r8
```

so the driver asked for 255 receive buffers, the shim answered "yes", and the
symptom was a NIC that transmitted perfectly and received nothing with `RDT`
sitting at zero.

**Where it is loaded.** `-mcmodel=kernel` means `R_X86_64_32S` relocations, so
the object has to live where a signed 32-bit absolute address reaches. A
process's mmap area starts at `0x40000000`, which is inside that window; on the
host harness `MAP_32BIT` reaches it from the other end.

### The KPI — `ldls/kpi/`

The Linux symbols, implemented. Register access to a mapped BAR; the PCI calls
to `/dev/pci`; allocation to `mmap` and `PCIIOC_DMA`; `register_netdev` to
`/dev/netdev`. Organised under Linux's own header names so a reader who knows
where something lives in Linux knows where it is here.

`kpi_io_ok()` bounds every `ioread`/`iowrite` against the mapping `pci_iomap`
returned. A stray access would otherwise be a page fault with no context, and
with no IOMMU a hosted driver's wild writes have to be caught early or not at
all.

Three pieces of it are not register access and are worth naming, because each
one is something a driver does that a poll loop does not obviously provide:

- **The interrupt handler runs.** Not to receive packets — LDLS was already
  calling the NAPI poll by hand — but because a driver's handler is where it
  does everything *else*. e1000 queues its watchdog from there, and the
  watchdog is what sets `TCTL.EN`. A driver whose handler never ran was a
  driver whose transmitter was never switched on. Calling it from the loop is
  safe: a handler's first act is to read its cause register, and with nothing
  pending it returns `IRQ_NONE`.
- **Deferred work runs.** `queue_delayed_work_on` keeps the item and
  `kpi_work_run()` calls what is due from the same loop. `jiffies` had to start
  moving for that to mean anything; it comes from `uptime()`, at 100 Hz.
- **`netif_carrier_on`/`off` write the bit.** `netif_carrier_ok()` is an inline
  that tests `__LINK_STATE_NOCARRIER` in `net_device.state`, so a no-op means
  the driver reads "the carrier is already up" and skips its whole link-up
  path. The bit number comes from the kernel's BTF (`btf_enum_value()`), not
  from us — rule 3½ applied to a constant rather than an offset.

And **DMA in small pieces**. `dma_alloc_coherent` maps onto `PCIIOC_DMA`
exactly; `dma_map_single` does not. 8139too never needed it, because it copies
each frame into a ring it allocated coherently up front; e1000 hands the device
individual buffers. Since a process cannot know the physical address of an
ordinary page, those buffers come from the arena too, a chunk at a time behind
a bump pointer.

The address arithmetic for that is not a call at all. `dma_map_single` is a
static inline, so what reaches the shim is already expanded: the module
computes `pa = va - page_offset_base`, then `page = vmemmap_base + (pa >> 12) *
64`. Both bases are undefined *data* symbols — which means LDLS supplies them,
and setting both to zero makes the whole thing reversible with no lookup table.

### The back end — `/dev/pci`

Already built, for `vgpud`, before any of this existed.
[hybrid.md](../plan/hybrid.md) insisted the interface be about *PCI devices*
rather than about a driver model, on the argument that an interface whose first
consumers are native drivers is a general mechanism rather than a hole shaped
like something foreign. A Linux driver's `pci_iomap`, `dma_alloc_coherent` and
`request_irq` land on it unaltered, which is that argument being collected.

### Giving a device up — `release`

A driver in the kernel now has a `release` hook, and `/dev/pci` calls it before
handing its device to a process. `docs/plan/hybrid.md` had said the honest
answer to "may I have the one you are using" is *no* until a driver could be
told to let go; e1000 is the first that can.

Letting go is three things, and the third was missed at first:

1. quiesce the hardware — mask interrupts, clear `RCTL` and `TCTL`;
2. **mask the interrupt line.** The kernel's handler is gone and the device is
   about to be brought up by a driver in a process, so nothing will acknowledge
   the cause. The console said what that looks like: `irq 11: 449905
   interrupts/second -- a device on this line is not acknowledging it`. It
   comes back when the process asks for it with `PCIIOC_IRQ`.
3. **take the interface out of the stack.** `net_detach()` is the missing half
   of `net_attach_late()`, and it went missing because nothing had ever taken a
   device back. An interface left behind keeps its address and the route to its
   subnet, so the stack goes on choosing it, and every frame reaches a
   `transmit` that now correctly refuses. Packets stop and nothing reports an
   error — which is indistinguishable from a driver in a process that does not
   work, and was mistaken for exactly that.

### Ports — the TSS I/O permission bitmap

A Linux driver does not always ask. `ioread32` is a call the KPI can implement;
`outl` is an *instruction* compiled into the module, and e1000's reset path
picks it deliberately — "these controllers can't ack the 64-bit write when
issuing the reset, so use IO-mapping as a workaround". On QEMU's 82540EM there
is no route to a working NIC that avoids it.

So a process holding a device gets that device's I/O BARs, **and no other
ports**. x86 offers two ways to grant this and they are far apart: `IOPL=3` is
one flag and hands over all 65536 ports, including the interrupt controller,
the timer and the disk; the TSS I/O permission bitmap is per port. A foreign
driver already has DMA with no IOMMU, and the answer to "this is not a security
boundary" is not to move it further — so it is the bitmap, at 8 KB per CPU of
resident memory. See [../../include/ioperm.h](../../include/ioperm.h).

Writing it found a latent hole. The TSS's I/O map base is at offset 102 and the
old code wrote offset 64; the field it meant to set stayed 0, and a base of 0
with a limit of `0x67` does not mean "no bitmap" — it means the bitmap starts at
the top of the TSS. The CPU had been checking ports 0–831 against the TSS's own
bytes, which are `RSP0`: a kernel stack pointer whose zero bits are permission.

### The class bridge — `/dev/netdev`

The one piece the kernel had to grow. `netdev_ops` is four entries —
`transmit`, `receive`, `mac`, `link_up` — so a process claims the device,
supplies a MAC, and `read()`/`write()` carry the frames. `/dev/pci` with the
arrows reversed: that hands a device *to* a process, this lets a process *be*
one. See [devices.md](devices.md).

lwIP never learns that the thing behind the vtable is not a driver.

---

## What it costs

**It polls.** A Linux driver expects its interrupt handler to be called from an
interrupt. `/dev/pci` does deliver them — by blocking `read()`, which needs a
thread `ldlsd` does not have — so the handler, the NAPI poll and the driver's
deferred work are all called from one loop instead. The same work, in a context
LDLS can provide, at the cost of latency and a spinning process.

**The stack protector is gone**, per `ko_defuse_gs()` above, and so are the
retpolines, per `thunk.c`. Foreign code, no IOMMU, three hardening measures
fewer.

**Per-CPU means per-process.** Correct with one thread, and it is the same
assumption as the no-op locks below, arrived at from the other side.

**A lock is a no-op.** In one process with no preemption inside the driver
there is nobody to race with. That is correct right up until a driver has an
interrupt thread, and it is the deepest unfixed assumption in the layer.

---

## It can be killed

The whole H2 machinery applies to a foreign driver unchanged, which was the
point of having built it against our own:

| | |
|---|---|
| `ldlsd` running | `ping` works; it holds the device and two DMA blocks |
| killed | interface **down**, device released, bus mastering cleared, **DMA reclaimed** |
| `ping` | fails honestly — `sendto: -5` |
| a second `ldlsd` | `ping` works again, and the arena hands back the same addresses |

`pciuser_release_all()` and `netuser_release_all()` sit beside each other in
`exit()` for the same reason: a process holding a device is one that may crash.

---

## What it does not do

- **Two drivers, one kernel version, one machine.** `8139too` and `e1000` on
  6.18.46 against QEMU. The second driver was worth more than the first: it
  found the per-CPU problem, the retpoline problem, the port-I/O problem, the
  interrupt storm and a seven-year-old hole in the TSS. Nothing here has been
  tried on a second kernel or on real silicon, and "LDLS runs Linux drivers" is
  still a claim about a category from a sample of two.
- **No interrupts**, per the polling note above.
- **`jiffies` is not `HZ`.** It comes from a 100 Hz `uptime()`, which is a
  plausible `HZ` and not necessarily the one the module was built for. A driver
  measuring a timeout against it gets a longer timeout, which is the safe
  direction and is still wrong.
- **One work item at a time, on the main loop.** A driver whose deferred work
  blocks stops everything, including the transmit path it is probably waiting
  for.
- **It is not a security boundary, and here that is worse than usual.** There
  is no IOMMU, so a hosted driver can point its device at any physical address.
  For SchOS's own drivers that isolates bugs and not intent; for *foreign*
  drivers the distinction between bug and intent is one we are in no position
  to make. This is the strongest argument for an IOMMU the project has
  produced.
- **It relies on SchOS having no NX.** Every mapping is executable because
  `EFER.NXE` is never set. If that is ever fixed — and it should be — `ldlsd`
  needs a way to ask for executable memory, and `mmap` does not offer one.
- **Nothing is ever freed.** A hosted driver's allocations last as long as the
  process, which is a leak in the strict sense and the right trade for a
  program that will be killed to restart it.
- **BTF has to come from somewhere.** SchOS has no `/sys/kernel/btf/vmlinux`,
  so the user brings it with the module, from the kernel the module was built
  for. A kernel built without `CONFIG_DEBUG_INFO_BTF` has neither.
- **No config-space writes.** `/dev/pci` has none, deliberately — a write is
  how a BAR gets moved out from under the kernel — so a driver that needs one
  is refused loudly rather than quietly succeeded.
- **No module-to-module dependencies, no firmware loading, no sysfs, no
  unload.** `request_firmware` fails, `module_exit` is never called, and the
  way to stop a driver is to kill the process.
- **Some symbols are still logging stubs.** Deliberate, per rule 5 — and the
  single most expensive thing in this layer. **A stub that returns zero is not
  safe when zero is a meaningful answer**, and the list of ways that bit is
  now long enough to be a rule rather than an anecdote:

  | stub | what zero meant | what it looked like |
  |---|---|---|
  | `memcpy` | the copy did not happen | frames arrived as the right *length* of zeroes |
  | `_find_next_bit` | "bit 0 is set" | `for_each_cpu` never terminated; `ndo_open` hung |
  | `netif_carrier_off` | the carrier is already up | driver skipped its link-up path; `TCTL.EN` never set |
  | `__x86_indirect_thunk_r8` | the jump was discarded | 255 receive buffers requested, none allocated, `RDT` at 0 |
  | `dma_map_page_attrs` | bus address 0 | the NIC pointed at the bottom of physical memory |

  Not one of these reported an error, and only the last is a fault waiting to
  happen. A search function, a predicate, a jump and an address are all things
  for which zero is a specific answer rather than an absent one.