tiny-wb/docs/usb-protocol.md

13 KiB
Raw Blame History

USB console and scanraw protocol

This document is the implementation reference for talking to tinySA firmware. It combines the official USB-interface description with the behavior visible in the upstream C source and in src/protocol.rs, src/transport.rs, and src/acquisition.rs.

Transport

Connecting a powered tinySA by a data-capable USB cable exposes a virtual serial port (USB CDC): commonly /dev/ttyACM* on Linux and COM* on Windows. Drivers are built into most Linux systems and are normally installed automatically on Windows.

The on-device shell is a command stream, not a packet protocol:

  • commands are ASCII and are normally terminated by carriage return (\r);
  • terminal input is commonly echoed, so a reply can begin with the command;
  • ordinary output is text;
  • the ready prompt is ch>;
  • some commands (capture, display bulk data, and scanraw) place binary data in the same byte stream;
  • firmware performs limited parameter validation, so a prompt does not prove that the intended operation occurred.

The baud selector in tiny-wb controls how the host opens the serial device. USB CDC and the optional internal TTL USART are different transports; usart_cfg configures/reports the internal serial interface in firmware. Do not infer the physical USB transfer speed from that value.

Shell grammar

Common scalar syntax is:

  • frequency: integer/floating value, optionally suffixed k, M, or G;
  • level: floating-point dB/dBm value;
  • time: seconds, optionally suffixed m (milliseconds) or u (microseconds).

For software, plain integer Hz is the least ambiguous frequency representation. Commands and availability vary by build; issue help, then record version and info before relying on optional commands.

Commands relevant to tiny-wb

Command Meaning and notes
version Firmware version/build text.
info Model/platform information when enabled in the build.
help Commands compiled into this firmware. Prefer this over a hard-coded global list.
mode low input Select Basic low input; on Ultra, firmware also accepts mode input and has no Basic high path.
mode high input Select Basic high input. Not an Ultra input mode.
pause / resume Stop/restart normal display sweeping.
repeat 1..1000 Measurements per frequency. More repeats reduce random noise and increase time.
zero N Set the additive raw-sample encoding offset used by scanraw; not an RF zero/calibration operation.
sweep start F, sweep stop F Change display sweep boundaries. sweep alone reports settings.
scan START STOP [POINTS] [OUTMASK] Perform a scan with at most 290 points and optional textual output. Outmask bits are 1 frequencies, 2 measured data, and 4 stored data.
scanraw START STOP [POINTS] [OPTIONS] Perform an arbitrary-point binary scan. Option bit 0 requests unbuffered writes; bit 1 repeats continuously.
rbw auto or rbw VALUE Set resolution bandwidth in kHz in current firmware syntax. Valid ranges differ by model.
attenuate auto or attenuate 0..31 Select automatic/manual front-end attenuation where supported.
spur off|on[|auto] Control spur removal; auto is Ultra-specific.
lna off|on Ultra optional LNA; reduces overload headroom and disables its step attenuator path.
ultra off|on|auto|start|harm ... Ultra-family build only; control extended-frequency behavior.
leveloffset, vbat, threads, usart_cfg, marker Diagnostic/status queries used by tiny-wb; exact output is firmware-dependent.

Avoid sending mutating calibration, output, correction, or configuration-save commands until the model and active port are known.

scanraw request

The request is:

scanraw START_HZ STOP_HZ POINTS [OPTIONS]\r

Constraints visible in upstream firmware:

  • at least start and stop are required;
  • start must not exceed stop;
  • points defaults to the current display sweep-point count;
  • option values are bit flags: 1 unbuffered, 2 continuous, 3 both;
  • waterfall is disabled before acquisition in current firmware;
  • a continuous request must be explicitly aborted and needs a streaming parser.

tiny-wb omits OPTIONS, so ordinary acquisition requests one buffered sweep and --live repeats that complete request over one configured connection. The firmware buffers in small blocks internally; “buffered” does not mean the entire sweep is delivered in a single USB transfer. Native option-2 continuous framing remains unsupported until explicit abort and resynchronization are verified.

Binary response framing

Ignoring command echo, line endings, and the final prompt, the firmware emits:

0x7b                                      # '{'
0x78 LOW_BYTE HIGH_BYTE                  # 'x', sample 0, little-endian
0x78 LOW_BYTE HIGH_BYTE                  # 'x', sample 1
...
0x7d                                      # '}'

Thus the framed body length is 1 + 3 * points + 1 bytes. Each level is a 16-bit raw value. Although the official USB page describes the bytes in one place as “MSB LSB,” the firmware source writes val & 0xff first and val >> 8 second. The repository correctly unpacks the value as little-endian.

The x byte is a per-record alignment marker, not padding. Basic and Ultra use the same braces, marker, two-byte value and exact body length; there is no model-specific padding in the documented or reviewed scanraw format. Their model-specific difference is the numeric level bias described below.

Payload bytes are arbitrary. A payload byte can equal {, }, x, c, h, or >. Framing logic must only interpret { before the records, x at every three-byte record boundary, and } after exactly the requested record count (or after an explicitly aborted short frame). Searching binary payload for text tokens is unsafe.

Level conversion and zero

The official USB interface documents these native conversions:

Basic dBm = raw_u16 / 32 - 128
Ultra dBm = raw_u16 / 32 - 174

This is a fixed-point value with 1/32 dB numeric steps. The detector itself has coarser effective resolution, so the encoding precision is not the same as measurement accuracy.

The reviewed firmware defaults set config.ext_zero_level to 128 for Basic and 174 for Ultra. The scanraw implementation adds that offset to the measured value, and the zero handler lets zero N replace it. tiny-wb detects the family and selects its native offset automatically, then requires matching zero readback before scanning. For Basic this gives:

dBm = raw_u16 / 32 - 128

Ultra uses 174. Manual normalization is still available with scan.auto_zero_offset = false; readback must agree with scan.zero_offset. A mismatch would shift levels by 46 dB between native defaults, so acquisition now rejects missing or mismatched readback. See support and migration.

Worked record

Suppose the three bytes are:

78 00 08

0x0800 is 2048. With zero 128:

2048 / 32 - 128 = -64 dBm

Frequency assignment

The official USB command page specifies the scanraw request and record count, but does not define a point-to-frequency formula. The upstream cmd_scanraw implementation, reviewed on 2026-10-05, is the byte-level authority used here. Its C expression has three relevant conversions:

integer_step_hz = integer_divide(stop_hz - start_hz, points)
firmware_step_hz = float32(integer_step_hz)
frequency_hz[i] = start_hz + truncate_to_freq_t(firmware_step_hz * float32(i))
                    for i = 0 .. points - 1

The denominator is points, not points - 1. Integer division discards any remainder before the assignment to the 32-bit float; multiplication can then round at large values before conversion back to freq_t. For example, a 0–10 Hz, three-point request measures nominally at 0, 3, and 6 Hz. The requested stop is an upper boundary and the final point is normally roughly one step below it. tiny-wb's firmware axis reproduces these conversions, including the single-precision rounding.

The default is now scan.frequency_axis = "firmware". The compatibility-only inclusive setting retains labels produced by older tiny-wb releases:

host_step = (stop - start) / (points - 1)
frequency[i] = start + host_step * i

inclusive labels both requested endpoints and disagrees with the reviewed firmware. The discrepancy grows linearly through a sweep to almost one firmware step at the last point. Existing saved configuration remains compatible because an explicit inclusive value is still accepted. Record the connected device's version when validating a narrow carrier: the synthetic tests establish source parity, not RF frequency accuracy or behavior of every firmware revision.

Prompt and synchronization behavior

A reliable request/response state machine should:

  1. retain a persistent receive buffer across OS reads;
  2. write the complete command plus \r;
  3. parse optional echo and line endings;
  4. for text commands, recognize ch> even when split across reads and retain any bytes after it;
  5. for scanraw, find the frame start in shell text, then consume exactly the expected number of aligned records and the closing brace before looking for the prompt;
  6. use an overall deadline plus idle deadlines rather than treating one serial timeout as a successfully completed short reply;
  7. make cancellation explicit with the firmware's abort behavior;
  8. return protocol errors that include command, stage, byte counts, and a bounded hex/text diagnostic.

Sleeping a fixed duration is not a protocol guarantee. Sweep time changes with span, RBW, repeat, mode, signal-processing options, and firmware.

Current tiny-wb transaction

The current code sends the following logical sequence (frequency placeholders are decimal Hz):

version
info
help
leveloffset
vbat
usart_cfg
threads
marker
repeat REPEAT
zero MODEL_OFFSET
zero
pause
mode LOW_OR_HIGH input
sweep start START
sweep stop STOP
sweep
scan START STOP 290 101
scanraw START STOP POINTS
resume

Every item except scanraw goes through the text helper and waits for ch>. The transport retains unread bytes and recognizes prompts split across serial reads. It consumes the complete length-delimited raw frame before returning to text prompt parsing; arbitrary sample bytes cannot prematurely end the frame. The pure decoder requires the exact record count, alignment and closing brace.

After attempting pause, acquisition attempts resume even if frame decoding or a later command fails. A failed recovery is included with the original error. There is no cancellation protocol or full state restoration, and process exit can prevent recovery. The offset requires readback; other settings are checked for recognizable command errors. Optional diagnostic queries require help advertisement. ZS406 Ultra acquisition additionally requires explicit opt-in, an advertised ultra command and exact requested sweep boundary readback. See model-specific planning.

The scan outmask value 101 remains outside the documented three-bit mask and its returned data is discarded. The legacy preparation is enabled by default for compatibility, controlled by scan.legacy_pre_scan. Establish its hardware purpose before changing the default or copying it into another client.

Manual diagnostic session

Use a serial terminal that can send carriage return, with every RF source disconnected or safely attenuated. A minimally invasive identification session is:

version
info
help
sweep
vbat

Save the exact transcript. Do not test scanraw in a terminal that assumes all input is UTF-8; it returns binary and may corrupt the display or its own log. Use a byte-oriented program and a bounded point count instead.

Compatibility rules for implementers

  • Identify capabilities; do not infer Ultra merely from a version prefix if info gives an explicit model.
  • Treat the command list as a firmware-build property.
  • Keep Basic low/high path selection separate from Ultra normal/Ultra/harmonic range selection.
  • Make raw offset, valid frequency range, RBW limits, and legal mode syntax model capabilities.
  • Do not use UTF-8 strings for the mixed shell/binary receive buffer.
  • Never use an unescaped textual prompt as a delimiter inside a binary frame.
  • Restore or clearly report every instrument setting that the client changes.
  • Include model, firmware version, configured offset, start/stop, point count, repeat, RBW, attenuation, LNA, and Ultra state in future output metadata.

See Project architecture for remaining implementation limits and tiny-wb operations for what users can safely expect today.

Configuration command rejection

Shell::command_checked rejects text lines beginning with error:, usage:, or unknown command (case-insensitive). Acquisition uses it for settings and pause/resume; diagnostic queries retain raw replies. This catches known failure forms, not every firmware-specific error or acknowledgement. Binary scan errors still fail framing/timeout checks. Model-aware replies remain roadmap work.