[pdfinn]

InferNode on the Raspberry Pi 5 (BCM2712 and RP1), native, no host OS. os/bcm2712/README.md is the guide and os/bcm2712/INVENTORY.md says what is proved, feature by feature.

On the bench board (Pi 5 8GB, rev 1.0): four cores, the JIT, all 8GB, multi-core Dis, SD, Ethernet at 10/100/1000, USB, Wi-Fi, Bluetooth (pairing, L2CAP, LE scan), GPIO with edges, I2C (#J), SPI (#Q), PWM (#W), the real-time clock, the fan, the power button, NVMe at gen 2, boot from the drive, the release image, Lucifer on either HDMI port.

	Pi 5	Pi 4
Ethernet in / out, Mbit/s	946 / 930	935 / 863
Wi-Fi in / out, Mbit/s	52 / 51	47 / not measured
SD read, MB/s	19	19
NVMe read / write, MB/s	370 / 297	none
jitbench, s	6.0	10.4

Bugs this found in shared code

• arm64: core 0 idled in wfi, and ready() wakes cores with sev, which does not end that. A process that had to run on core 0 waited for a clock tick each time. The IP hardening in fix(security): complete report hardening and 9front audit #1032 removed a lock that had been hiding it: gigabit input fell to 300 Mbit/s with nothing lost. All cores now wfe: 946, steady, with that hardening in place.
• runproc spun taking and freeing the run queue lock at splhi on a proc still marked with a core, starving ready(). It now lets go, with interrupts on.
• lock(): a proc preempted between winning a lock and raising its priority, pinned to a core with a spinner at PriLock, deadlocked (lock loop ... held by pc 0x0, within 6 to 70 minutes of the all-devices soak). lock() and canlock() now keep interrupts off across those instructions.
• Audio: the jack's fuzz was a fractional clock divider and dropped rounding error; fixed, with a spectrum test.
• kfs marks itself clean when idle, so a reboot or an idle power cut needs no check.
• devletters.py --check failed on master for #K, #x, #X, #h, #z.

Gated to this board: NVME0, RAMDISCOVER (also BCM2711), SDCARD_HISPEED, WORDMEM, POOLBYNEED. Not gated: the two fixes above, poolsize taking a ulong, the heapreserve test, and osinit's button watcher and drive fallback.

Tests: QEMU bcm2837 + bcm2711 330/0/0 with the lock fix; audiolevel, wordmem, kernelpins and devletters pass. On the Pi 5, 24 hours with every device busy on a diagnostic build of this code: 0 panics, 0 failed compares, 0 media errors, kernel memory 14-17 MB. A 24-hour run on the release kernel is in progress. No QEMU models a Pi 5, so CI only builds it. The 3B+ booted and carried traffic on these fixes before the lock change, from a USB disk, and has not soaked them; the lock change has not run on a real Pi 3 or Pi 4.

Known limits

• D0 stepping pin tables have run only as a dry run; RAM above 4GB on a Pi 4 has not run; 1 to 4GB Pi 5s are simulated with memlimit=.
• Pools keep what they grow to, so after one VM's runaway the others have little.
• Bluetooth throughput is unmeasured; I2C and SPI have no real device on them yet.
• The bench Pi 5 logs under-voltage at boot.
• main and image ceilings come from a few measured pages, not a survey; pools keep what they grow to.
• Charon's heap is badly fragmented by something that grows in small steps; not in this PR.

🤖 Generated with Claude Code

https://claude.ai/code/session_0195u6z8JjsAy9yrE5TqKxzx

[pdfinn]

Branch rebuilt as master plus the Pi 5 work, and marked ready for review. CI is green: 36 checks pass, 1 skipped.

What changed on feat/pi5-port (a0787c3)

• Current master is merged in. Conflicts were resolved for master, and the merge commit says how:
  • os/port/taslock.c: master's os/port is a copy of emu/port from 2026-08-24: reconcile the 26 shared files, then stop having two copies #640 fix (PriLock before the test-and-set) covers the race this branch had closed by turning interrupts off.
  • libkern memmove and memset: master's eight-byte copies replace this branch's Pi 5-only WORDMEM ones.
• The 22 Pi 5 commits from the integration line are added:
  • RP1 Ethernet transmitter stall recovery;
  • NVMe recovery in three steps and a bridge re-setup;
  • Wi-Fi re-join without dropping the address;
  • kfs check and fsbench;
  • the gefs port, with its crash-safety fixes and tests;
  • #S/sdctl flush;
  • root on gefs (rootpath gefs part), off by default;
  • from the soak: osinit sets up again a USB device that dropped off the bus and came back, and keeps asking for the Wi-Fi address after a failed lease.
• The Pi 5 half of Bare metal: a key is made only from a hardware generator #1051 (../bcm/rng200.c) is added; the cherry-pick had missed it.
• kernelpins_test now accepts either fix for the lock window. Removing the PriLock raise still fails it.
• wordmem_test is removed: libkern_copy_test covers master's routines.
• brick(1): styxbrick client — TLS pinning, sdg transport, brick d2 safety outputs #1015 (styxbrick), which the integration line also carried, is not included.

Checked against the soaked branch: every file that differs differs through master's own changes or styxbrick's.

Evidence

• Local: the dis manifest builds (1,184 modules); device letters, mkfile dependencies, the bcm2712, bcm2711, bcm2837 and virt kernels, and the gefs, kfs, kernrandom, libkern, kernelpins and audiolevel tests all pass.
• soak33 (12 h, Pi 5, on the integration line plus Bare metal: one ChaCha20 generator, keyed from the hardware, behind every reader #1059 and dossrv: a 1.1 MB sector cache, not a floppy's 90 KB #1062, so the same Pi 5 code without master's newer commits):
  • no reboots or crashes, memory flat;
  • Ethernet ~850/550 Mbit/s, Wi-Fi ~43/75, USB ~190/125;
  • every device loop failure-free (SD, NVMe, kfs, SPI, I2C, GPIO, Bluetooth, drawing, both VMs, memory, RNG, RSA);
  • 2 TLS failures early in the run, with output lost. They have not recurred: 457 runs since, 193 of them at ten times the rate under full load. The soak harness now keeps each failure's output.
• soak34 (12 h, Pi 5, on this branch's build) started 00:36 and is clean at the start. Its result will be posted here.

Merge is yours. soak34 is the soak of exactly this build.

🤖 Generated with Claude Code

https://claude.ai/code/session_0195u6z8JjsAy9yrE5TqKxzx