Where the knowledge in this tree came from

This project is BSD-2-Clause. Every file says so, and LICENSE is the whole of it.

Almost nothing here could have been written without material that is licensed differently, and a reader is entitled to know exactly what was taken and what was not. That is what this page is for.

It is not legal advice. It is a statement of practice, written so that anyone reviewing the project can check the practice against the tree.

This page covers the ARMADA family only - the Marvell CN913x material, which reached us in

Sophos's GPL drop as the pcie_ep_armada release and, separately, as its umsd and netagent

components. The OCTEON TX family came out of a different drop, under a different licence, and is accounted for separately in octeontx/provenance.md. They are deliberately not merged: mixing two upstreams with different origins into one page is exactly how the boundary this project keeps would stop being checkable.

The rule

Facts are transcribed. Expression is not copied.

A register offset, a command number, a field width, a byte order, the value a device expects in a particular word — these are facts about a piece of hardware. Learning them from a published source and writing them down in a different form is how every clean driver for undocumented hardware has ever been written.

The code that acts on those facts is written here, from nothing.

contrib/npuep/npugiu.h is where the rule is visible. Marvell's giu_nic_hw.h opens by asking that it be copied verbatim into any host driver, which is the sensible request for a C header shared by two ends of a link. It was not copied. That file is a list of #defined offsets with no structure declarations at all, and the header explains at length why the offsets are the right form and the structures would have been wrong here anyway — a C struct laid over MMIO invites the compiler to merge, split, reorder or elide accesses that must happen exactly once.

The result is more tedious to write against. It is also the only form that could be published under this licence.

What was read

Source	Licence	What was taken
Marvell's giu_nic_hw.h, in Sophos's GPL drop	GPL	Offsets, command numbers, flag bits, descriptor layouts
Marvell's UMSD switch driver, same drop	GPL	Register names, TCAM entry semantics, the switch's initial state
Sophos's NetAgent sources, same drop	GPL	The network-agent message and attribute numbering, request and reply shapes
usfp_rh.ko, shipped on the appliance	proprietary binary	Structure layouts and enumerator names, read from the DWARF the vendor left in it
Sophos's kernel-side usfp sources, in the open-source disc of the same firmware release	GPL-2.0 or BSD-2-Clause, at the reader's choice, as each file states	The frame metadata and table-entry layouts, the command numbers, what the handlers for association, microflow and connection commands check and in what order, and the conditions and limit of the rule by which the fast path gives a TCP connection back
xgs-usb-spi-flash, Sophos's own host tool
proprietary binary	The MCP2210 command bytes and field offsets, and the per-board GPIO hold and
release masks, recovered by disassembling it
The appliance's own boot logs and captures	—	Confirmation that a reading was right

Every one of those is a description of an interface. None of it is in this tree as code.

Quotation

Four short comments are quoted verbatim, each attributed to its source where it appears:

• contrib/npuep/npuep.c and docs/npuep/facility-protocol.md — Marvell's "Workaround for MSI NMP zero MSIX ID entry issue", which is the only explanation anywhere for a doorbell count that makes no sense otherwise.
• contrib/npuep/npugiu.h — Marvell's request that the header be "copied AS-IS", quoted precisely because this project declined to do that.
• contrib/npuep/npunwa.c — UMSD's "On this stage, only Broadcast mode is enabled, for other packets TCAM HIT will be only on Drop entry, which drops the packet", which is the whole explanation for why unicast did not work for weeks.
• contrib/npuep/npugiu.c - Marvell's "Meanwhile, we support only strict prio", which is why the transmit queues are given weight zero rather than one.

These are single sentences, used to show the evidence for a claim rather than to reproduce a work. Paraphrasing them would be worse for the reader: the point of quoting is that you can check the claim against what the vendor actually wrote.

If any rights holder disagrees with that judgement, the remedy is one line each and the surrounding explanation stands without them. Open an issue.

What is deliberately not here

• No vendor source. Not a file, not a function, not a structure declaration.
• No firmware. The coprocessor runs the vendor's own, which stays on the appliance.
• No binaries. Nothing is redistributed. What the appliance ships, the appliance keeps.
• No credentials, no real addresses, no hardware identifiers. tools/check-private-data.py runs in CI and fails the build over any of them, including a MAC folded inside an IPv6 link-local. It has caught real leaks, including one from the author of this page.

Why BSD-2-Clause and not GPL

Three reasons, in order of weight.

It matches FreeBSD. This is a FreeBSD kernel module. The kernel it loads into is BSD-2-Clause, and a driver under the same terms is one that could be offered upstream rather than one that forever cannot.

It keeps the boundary legible. A GPL project built beside GPL sources invites exactly the confusion this page exists to prevent — did that come from there, or was it written here? Under a different licence the question has to be answered file by file, and it has been.

It does not restrict the reader. Anyone who wants to put this on an appliance and ship it can. That was the point of writing it.

Discussion

The discussion of this page is held in the project's GitHub Discussions. Showing it here loads giscus from giscus.app and GitHub, and writing in it needs a GitHub account - what that means.

Show the discussion