i2c-imc-skylake/tools/README.md

Hardware measurements still owed to the submission

Three scripts, meant to be run by hand on an X299 (or Cascade Lake-X) system with the PCU function 8086:2085. None of them is run by CI and none is run by make test.

Output of all three belongs in the cover letter or in a reply on the list, as raw numbers. Do not paraphrase them.

Script	Answers	Writes to hardware
watch-engine.sh	Is firmware using the engine on this system?	no
dump-pcu-config.sh	Which PCU registers move on their own? Is there a TSOD polling control?	no
probe-cf8-write.sh	Is the CF8/CFC write really dropped, and does it raise an SMI?	yes, requires --i-understand

Not part of the submission: pre-poweroff-load.sh — fixed, mixed SMBus load to run before every poweroff of the A5 hang observation window (TODO.md E1), identical in the use_ecam=1 phase and the use_ecam=0 counter-proof. 3-minute blocks: software effect stream (~2760 writes/s) → rapid hw/sw switch burst → hardware effect → full SPD dump via ee1004 + serial integrity check. Fully automatic: saves run context + any driver warnings to a per-run file in tools/pre-poweroff-runs/, appends one summary line to pre-poweroff-load.log, then — only with --poweroff — powers the machine off (15 s grace, Ctrl-C aborts); without the flag it just reports and exits. Run: sudo ./tools/pre-poweroff-load.sh [minutes] [--poweroff] (default 30; re-execs under sudo by itself). The one exception: an SPD serial mismatch saves everything and does NOT power off — investigate first.

1. watch-engine.sh — foreign activity (read-only)

Samples the six per-channel registers and reports every change that this system's Linux side did not cause. Run it with the driver unloaded, so that anything observed is firmware.

sudo ./tools/watch-engine.sh --seconds 3600

What the cover letter needs: a run of at least one hour, ideally across a suspend/resume cycle, with the observed change count. A non-zero count on a system means the driver must not be loaded there, and is itself a result worth reporting.

2. dump-pcu-config.sh — register survey (read-only)

Dumps the function's 4 KB config space twice and diffs them, then prints the neighbourhood of the offsets that matter.

sudo ./tools/dump-pcu-config.sh --settle 10

This is the search for the open item F2: the Broadwell-E iMC driver stops the PCU's TSOD polling for the duration of a transfer by writing 0 to TSODCNTL at offset 0xe0 of its PCU function, waiting 10 ms for in-flight transactions to drain, and restoring the value afterwards. No equivalent register has been identified on 8086:2085. A field that changes on its own with a fixed cadence is the first candidate.

3. probe-cf8-write.sh — the dropped write (WRITES TO HARDWARE)

This is the only script that writes. It changes the register/offset field of a command register while the GO bit is clear, which does not start an SMBus transaction, then reads it back and reports whether the write stuck. It also reads MSR_SMI_COUNT (0x34) on every CPU before and after.

sudo ./tools/probe-cf8-write.sh --i-understand

Two results matter:

• the value does not read back → the claim "CF8/CFC writes to this function do not take effect" is confirmed on the tested board;
• the SMI delta is 0 → the behaviour is not System Management Mode, and the driver's comments and the cover letter must not say it is.

Requirements: msr-tools (rdmsr), pciutils (setpci, lspci), and the msr module (modprobe msr).

4. probe-cf8-transaction.sh — does CF8/CFC start a transfer? (WRITES TO HARDWARE)

Follow-up to §3. probe-cf8-write.sh shows whether the register accepts a value; this shows whether the engine acts on it, by setting the GO bit for a read of an unpopulated address and watching STATUS.

sudo ./tools/probe-cf8-transaction.sh --i-understand

It does start one. That result retired the plan to reach these registers through ECAM with a PCI-core change, and the driver uses the ordinary config accessors.

5. probe-imc-documented.sh — the interface Intel documents (WRITES TO HARDWARE)

Issues a read through SMB_STAT/SMBCMD/SMBCNTL on the iMC channel functions (8086:2040), the interface the datasheet describes, rather than the PCU registers this driver uses.

sudo ./tools/probe-imc-documented.sh --i-understand

On the development board it reports RDO=1, SBE=0, RDATA=0x00 identically for a populated 0x50, a populated 0x52 and an absent 0x57: it completes but cannot tell a populated slot from an empty one, so it is not reaching this bus.

6. probe-pntr-sel.sh — find the address-only bit (read-only on the bus)

Finds the command-word bit that suppresses the register byte, which is what SMBus Receive Byte and Send Byte need.

sudo ./tools/probe-pntr-sel.sh --i-understand --channel 1

Not a blind sweep. An EEPROM auto-increments its pointer after a read, so priming the pointer with an ordinary read and then repeating it with the candidate bit set has a predicted answer: the next byte, not the one named. The script checks that prediction at two different offsets and then walks eight consecutive reads, which an ordinary dump must reproduce exactly.

The answer on this engine is bit 18, matching the position of PNTR_SEL relative to WORD_ACCESS in the documented SMBCMD layout. Reads only, and never to the page-select addresses, so the DDR4 page latch does not move.

7. probe-send-byte.sh — the write half (WRITES TO HARDWARE)

Confirms the same bit produces a Send Byte, against the page-select addresses 0x36/0x37. Those hold a latch and store no data, so the only thing a write there can do is choose which half of the SPD is visible — and that choice is itself the readout, since byte 0x00 is the JEDEC 0x23 on page 0 and something else on page 1. Nothing is written to 0x50-0x57.

sudo ./tools/probe-send-byte.sh --i-understand --channel 0 --restore

--restore is the safe first run: it targets a channel already stuck on page 1 and selects page 0, so a working Send Byte repairs that channel and a broken one leaves it exactly as found. There is no outcome worse than the starting state. --flip does the full round trip on a healthy channel, after --restore has shown the encoding works.

A channel found on page 1 is not a fault of this driver: ee1004, and OpenRGB's DDR4DirectAccessor::set_page(), both select a page and leave it selected.

8. lore-thread-check.py — "la patch è stata pubblicata? qualcuno ha risposto?"

Niente hardware: interroga l'archivio NNTP di lore, nntp.lore.kernel.org:119, e dice (a) se un message-id è nell'archivio della lista, con gli header veri, e (b) quali articoli successivi appartengono allo stesso thread. Sola lettura, il server rifiuta il posting comunque.

./tools/lore-thread-check.py [email protected]
./tools/lore-thread-check.py <msgid> --group org.kernel.vger.linux-kernel

Esce 0 se il messaggio è archiviato, 1 se non c'è, 2 su errore.

🚨 Esiste perché dal browser o con curl la risposta è ingannevole: lore.kernel.org e patchwork.kernel.org stanno dietro il bot-check Anubis e restituiscono 403 o la pagina "Making sure you're not a bot!" al posto del messaggio — che si legge facilmente come "la patch non è mai stata pubblicata". Il mirror NNTP non ha il challenge. Per patchwork non esiste equivalente: lì serve un browser vero.

9. probe-ene-receive-byte.sh — la Receive Byte a 0x27 (read-only)

L'esperimento di TODO.md §P2, e le sue varianti diagnostiche. Nessuna scrittura, su nessun indirizzo: solo Receive Byte e Read Byte Data verso 0x27 e letture SPD dai nodi sysfs di ee1004. Le quattro seriali SPD sono confrontate prima e dopo a ogni run e una differenza è errore fatale.

sudo systemctl stop furyd            # lo script rifiuta di partire con furyd attivo
sudo ./tools/probe-ene-receive-byte.sh --rounds 12   # l'esperimento
sudo ./tools/probe-ene-receive-byte.sh --latch-test  # eco del command byte + cross-address
sudo ./tools/probe-ene-receive-byte.sh --map         # ri-deriva l'artefatto N>>2
sudo ./tools/probe-ene-receive-byte.sh --addr-sweep  # stessa read su altri indirizzi
sudo systemctl start furyd

Esito del 2026-08-10 (log in docs/evidence/2026-08-10-ene-*.log): la Receive Byte restituisce l'eco dell'ultimo command byte mandato a quel chip (12/12, due canali), il latch è del device e non dell'engine (da 0x31 torna 0xff), gli indirizzi vuoti NACKano. L'ENE non ha stato leggibile — ora misurato.

🚨 Due trappole, valgono per qualunque misura su questo engine: la prima transazione di un canale restituisce un valore stale (per questo lo script fa un read di warm-up scartato), e la lettura successiva a una riuscita NACKa spesso — un solo retry la pulisce sempre, come dice P1.