Skip to content

Contributing SPD dumps from your own memory #2

Open

Description

@danalec

The spdr project can only get better by being tested against real modules, and right now that runs into three walls: reading a DDR5 SPD out of a module needs an EFI tool or administrator rights, the file arrives in whatever format the tool happened to write, and it carries a serial number that identifies the person who owns the module.

spdr-collect is a small intake tool that clears those three. It is a companion to the spdr library, not a replacement for it: the decoding is still done by spdr.

What it does

It reads the dump you have, removes the one field that identifies your module, checks that the result is still a faithful image of the module, and prints the provenance record that a docs/validated-against.md entry is built from.

Three ways to use it, in order of how little they ask of you

In the browser, with nothing installed → https://danalec.github.io/spdr-collect/

Drag your dump onto the page. It is one self-contained HTML file: no server, no network request, and your file never leaves your computer. You get back the sanitised file to download and the provenance text to paste into a pull request. It works offline too, so you can save it and open it locally.

As a command → the releases page has prebuilt binaries for Windows, Linux and macOS, so Rust does not have to be installed:

spdr-collect check dump.bin        # what is in it
spdr-collect sanitize dump.bin -o fixtures/my-module.spd
spdr-collect submit dump.bin       # opens the pull request for you

submit needs the gh CLI, which already holds your GitHub credentials. Without it, it prints the steps to do by hand.

From a checkout, with cargo install --path ..

Getting the dump out of the module

This part is a hardware problem rather than a software one, and no tool can do it for you.

Linux. The SPD hub on a DDR5 module is bound to the spd5118 driver, which the kernel has carried since 6.11, and the EEPROM shows up as a file:

dd if=/sys/bus/i2c/drivers/spd5118/0-0050/eeprom of=dump.bin

The bus and device address differ from machine to machine; the address is usually 0x50 for the first slot, and ls /sys/bus/i2c/drivers/spd5118/ lists what is there. ee1004 is the older driver, for DDR4. On a kernel older than 6.11 there is no DDR5 SPD support at all, and the UEFI route below is the way.

Windows. An EFI tool such as RWEverything can write the SMBus to a file from a UEFI shell. Booting into the UEFI shell and running it from there is usually the easiest route.

Either way. If you would rather not produce a binary file, hexdump -C dump.bin > dump.txt produces a text dump that the tool reads just as well. A raw xxd or od dump works too.

What leaves your machine, and what does not

Only one field is removed: the module serial number, at bytes 517..=520. It is the only part of the SPD that identifies a particular device.

That removal is safe to vouch for, and this is the part worth checking: every identifying field lives in the manufacturing information block, at bytes 512..=554, and the base configuration CRC covers exactly bytes 0..=509. The two do not overlap. So after the serial number is replaced, the base CRC still recomputes to the value the module itself stores. The sanitised dump is a faithful image of the module, not a reconstruction, and it needs no recomputed checksum to be evidence.

Everything else is kept: the manufacturer, the manufacturing date, the part number (public product data) and the DRAM manufacturer. If you would rather the part number went too, --redact serial-and-part-number does that.

The report also carries the SHA-256 of the file you sent and of the sanitised file, so it is exactly checkable what arrived.

Sending it

The fixture has to land as a file in spdr/tests/fixtures/, so it needs a pull request. The browser page walks through it: fork the repository, upload the sanitised file into spdr/tests/fixtures/ on a branch, and open the pull request with the provenance text in the description. spdr-collect submit does those steps for you.

Whatever the module is, it is worth sending. The scope that is still deferred — SODIMM, RDIMM and LRDIMM module-specific decoding, the full JEDEC sub-grade tables, the tFAW ordering — is gated on real modules of those types, and a corpus of one UDIMM cannot get there.

Activity

  1. mertefesensoy commented on Nov 16, 2026

    @mertefesensoy
    Member

    Thanks for this, @danalec, and sorry it sat for three weeks without an answer.

    The claim you asked to have checked holds. I went through it against spdr's own source:

    • The base configuration CRC covers bytes 0..=509 (crc.rs).
    • The serial number is bytes 517..=520, inside the manufacturing block at 512..=554 (manufacturing.rs).
    • The XMP and EXPO sections start at 640 and 832, so their section CRCs do not reach it either.

    So replacing the serial number leaves every checksum spdr verifies unchanged, and a sanitised dump is still a faithful image of the module for everything the decoder reads. That is the property a fixture needs.

    One thing changed on our side that affects the tool. spdr has moved to a new organisation:

    https://github.com/Pactolus-Microelectronics/spdr

    The old address does not redirect any more; it now shows a fork kept for the record. spdr-collect's README, and anything in submit that names the upstream repository, should point at the new address, or fixture pull requests will land on the fork.

    What I would like to do next:

    1. Add a short "contributing a dump" section to the spdr docs that describes the fixture rules (where the file goes, what the provenance entry in docs/validated-against.md must contain) and points to spdr-collect as the easy way to prepare one.
    2. Take the first fixtures through it. Non-UDIMM modules are the most valuable, since SODIMM, RDIMM and LRDIMM decoding are all waiting on a real module of each type.

    Two things before I point people at it from the README:

    • I have not read spdr-collect's source yet. Because the tool's whole job is to remove an identifying field, I want to read the sanitising path myself first.
    • The repository does not show a licence. Could you add one? Apache-2.0 would match spdr.

    And thank you again for the quality pass in #1.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions