USB Device Fundamentals

When plugged in, the host only knows a data line was pulled high (the speed-negotiation signal — full-speed pulls D+ via a 1.5kΩ resistor). Everything else — what the device is, what it can do — is learned through control transfers on endpoint 0, the one endpoint every device must implement. This is enumeration: reset the bus, assign address 0, request the device descriptor in small chunks (to learn EP0's max packet size), assign a real address, then read the full device/config/interface/endpoint descriptors.

Descriptors are small, fixed-format structs exposed read-only over EP0 — the host only ever requests one of these specific structures. They nest strictly:

Device Descriptor          (1 per device — VID, PID, class, EP0 max packet size)
  └─ Configuration Descriptor  (1+ — power draw, number of interfaces)
       └─ Interface Descriptor    (1+ — which class driver handles this: HID/CDC/MIDI…)
            └─ Endpoint Descriptor    (0+ — direction, type, max packet size, polling interval)

A device can expose multiple interfaces under one configuration — how a board can be a virtual COM port and a mass-storage drive at once (see the TinyUSB stack lesson's cdc_msc example): each interface gets its own host-side driver, matched by class code.

Every device descriptor carries a 16-bit Vendor ID and Product ID. The VID is licensed from the USB-IF; hobbyist projects commonly borrow an open-source-friendly VID (e.g. TinyUSB's default 0xCafe) rather than fabricate one — a colliding pair can confuse a host with a driver already bound to that combination. For a class the host already has a built-in driver for (HID, CDC-ACM, mass storage), VID/PID mostly just distinguishes your device in the OS list — the class code selects the driver.

Every non-zero endpoint is one of four types, each trading off latency, bandwidth, and error handling differently:

This is why a keyboard feels instant (interrupt endpoint, guaranteed a slot every polling interval) while a mass-storage transfer can stall and resume (bulk endpoint — no timing guarantee, but every packet is retried until it arrives correctly). Isochronous is the odd one out: it guarantees bandwidth but never retries, because a late audio sample is worse than a dropped one.