virtual-sx1262

A virtual Semtech SX1262, with a C ABI. This is the chip real MeshCore firmware talks to when it runs under MeshBench, and the only thing between an unmodified radio driver and a simulated sky.

This is not a model of MeshCore's radio driver. The driver is real and runs unmodified: CustomSX1262 over RadioLib, clocking bytes at a chip select. What is ours is the part that answers those bytes. A simulator that answered them approximately would be deciding listen-before-talk for the firmware, which is answering a different question from the one anybody asked.

Where the real ends and the simulation begins

Worth being exact about, because the interesting bugs live on the boundary.

Layer	What runs
MeshCore application and mesh logic	real, unmodified
MeshCore radio driver, CustomSX1262 and RadioLibWrapper	real, unmodified
RadioLib	real, vendored by the host, unmodified
The SX1262 chip	this repository
Arduino, board, filesystem, RTC, sensors, RNG	the host's, per host
The air: path loss, noise, collisions, capture effect	the simulator's

The line matters in both directions. This part never decides whether a packet gets through: it is told "a carrier is present" and handed frames. And it never guesses at airtime: vsx_est_airtime_ms is Semtech's own formula, because MeshCore's CSMA is built on the figure the firmware computes and a second formula beside it would drift.

Why this is its own repository

It used to be a file inside MeshBench/meshcore-native, beside the host firmware build, with a small server process wrapping it so emulators could reach it down a socket. That put a shared model inside a repository named after one of its consumers, and left every other consumer talking to it through a pipe.

It has four, and none is more entitled to it than the others:

Host	How it links	Its licence
MeshBench/meshcore-native, the native firmware build	the C++ class, in process	MIT
MeshBench/qemu, the sx1262 device	this ABI, extern "C"	GPLv2
MeshBench/meshbench, the simulator	this ABI, via cgo	GPL-3.0-or-later
Renode's SX1262 peripheral	this ABI, via DllImport	MIT

One model, four hosts, no server. That table is also the whole of the licence argument below.

The ABI

include/virtual_sx1262.h is the entire surface, and two shapes in it were learned from the socket it replaces.

Whole transactions, not single bytes. vsx_spi_transaction takes one chip-select framed exchange. A call per byte across a P/Invoke boundary costs more than the socket it was meant to remove, and the chip has to know where a command ends regardless: the SX1262 wire protocol carries no length, so chip select is the only frame there is.

The chip raises DIO1. Nobody asks it. The socket was request-response, so the model had no way to call back and its hosts sampled the line on a one millisecond timer instead. That is a millisecond of latency on every received packet, for a pin a real part asserts the instant it has something to say. vsx_set_dio1_callback pushes the edge. vsx_dio1_asserted is still there for a host that would rather ask.

The C++ class is not part of the ABI. Hosts hold an opaque vsx_chip *, so the model can be rearranged without breaking a DllImport written by hand.

What the tests are for

test/test_model.cpp is not coverage for its own sake. Every case is a bug that previously took a four minute emulated boot to see, because the model had no bench of its own:

• A carrier that has ended is not still present. The model raised PreambleDetected, HeaderValid and SyncWordValid at RxDone, which asserts that a signal is arriving at the instant it stopped. MeshCore's CustomSX1262::isReceiving reads HEADER_VALID as "the channel is busy", so a repeater holding a packet to forward never saw a clear channel and relayed nothing at all. Detection flags are raised as the signal arrives and cleared when it goes.
• A packet handed to a deaf chip does not turn up later. A frame delivered while the chip was not listening waited in the inbox until it started listening, however long that took, so a packet could arrive seconds after the air that carried it went quiet.
• A brief re-arm gap still receives. The bound on that wait is a short grace rather than an immediate drop, because the driver's own startReceive is standby, configure, SetRx: the chip is briefly not listening inside a perfectly healthy receiver, and a frame landing in that gap is receivable.
• The chip only lies when asked to. Latching a detection flag past the signal is what real SX1262s sometimes do, and what MeshCore 1.17's recovery exists to survive, so it is a deliberate variant and never the default.

test/test_c_abi.c is compiled by the C compiler on purpose. Two hosts are not C++, and a header that only builds under a C++ compiler passes every other test here and then fails in the one place it matters.

Building

./build.sh            everything, then run both test binaries
./build.sh static     libvirtualsx1262.a, for C and C++ hosts
./build.sh shared     libvirtualsx1262.{so,dylib,dll}, for hosts loading at runtime
./build.sh test       build and run the tests
./build.sh sanitize   the tests under AddressSanitizer and UBSan

STRICT=1 turns warnings into errors, which is what CI uses. No dependencies beyond a C++17 compiler, which is not an accident: see the licence.

What it does not model

Real errata, and a real preamble detector's behaviour under interference. It is our best understanding of an SX1262, which is a different thing from an SX1262, and every result that rests on it is a best case.

Licence

MIT, and deliberately rather than by inheritance.

Two of the four hosts above are copyleft, and they are not the same copyleft: QEMU is GPLv2 as upstream ships it, MeshBench is GPL-3.0-or-later. Those two cannot both link one copyleft library, because GPLv2-only and GPL-3.0 are mutually incompatible. A permissive licence is the only thing that lets the same chip answer a native node, a QEMU device and a Renode peripheral, and one chip answering all of them is the entire reason those backends are comparable.

Nothing here constrains the choice from below: the model links the C++ standard library and nothing else, so no third-party terms are inherited. It was already MIT in meshcore-native, so this is where it stays rather than where it moved. See NOTICE.md.