Lab notes

Hardware measurements

On this page

These board experiments settled decisions in the firmware; the reference pages state the resulting rules and link here for the evidence. Most ran on an Orange Pi Zero with throwaway images that wrote register values into the counters table (MIB 0x100c) during start-up or from the main loop. Run lengths and exact revisions were not always recorded.

Interrupts and sleep

Experiment Result Decision
Timer 1 waits of 1,000, 20,000 and 5 µs Lasted 1,001, 20,001 and 7 µs Timer 1 bounds sleeps; it is no delay primitive below about 7 µs
Sleeps while the RX producer moved Of 3,707, 3,681 ended early with no controller source pending, 9 ended early with one pending, and 17 ran their full length while a frame arrived The MAC’s FIQ wakes the CPU for received frames, invisibly to the controller
Every interrupt source enabled Sources 9, 24 and 28 stay pending; acknowledging 28, or every source, stopped reception within seconds Leave 9, 24 and 28 disabled; drain the MAC event FIFO from the main loop
UART transmit interrupt (source 0) enabled Ended every wait at once Leave it disabled
Pending sources after a pass under traffic 13 (HIF) and 20 (AES), both cleared by their owners; 26 never IRQ 26 needs no handling while the TSF is not adjusted
Replaying the nine interrupt routing words No measurable change The firmware does not write them
Reading the interrupt controller and timer registers through a host driver’s debugfs ahb file, with the CPU halted All read 0, the free-running clock included Vendor settings cannot be read back this way
Fixed 1 ms sleep bound, associated and idle 1,000 wakes a second Sleep until the next deadline: about 20 wakes a second

Sleeping between frames

Downlink loss at 30 Mbit/s with two host-driver builds used during development, recorded only as the “minimal” driver and the “stack” driver; which revisions they were cannot be recovered, so these results cannot be reproduced exactly.

Policy Host driver Loss
Sleep between frames, 1 ms bound Minimal 0.5–7 %
Same, with a 200 µs bound, or with the write buffer drained before each wait Minimal 0.5–7 %
Wait awake on the same conditions Minimal 0–0.1 %
Sleep between frames Stack 0.03–0.04 %
Vendor firmware Stack 0.01–0.04 %
Stay awake while a flow is active (current) Stack 0.03–0.06 %

Waiting awake on the same wake conditions cut the loss to 0–0.1 %, which points at the wait for interrupt itself rather than a missed wake inferred. Why only the minimal driver is affected is an open question.

A-MPDU command stream

Each change was made to an otherwise working aggregate.

Change Result Decision
Delimiter 0x6600_0001 Normal operation The delimiter’s low bit is ignored
Delimiter 0x6700_0000 Only the second member was lost: 52 selective retries of it and 8 terminal failures in 5 s A delimiter controls the boundary before the member that follows it
First transfer word 0x6400_0000 Aggregation never became operational; 8 host buffers were not returned The first member cannot be invalidated alone

Image loading and timing

Experiment Result Decision
The runtime repacked as three records: a 0x7500-byte copy at 0, a 0x2218-byte zero fill, an entry record Healthy WSM idle state, no leaked buffers, normal scan traffic The loader rebuilds the runtime from records; the host-visible file need not match the memory image
Runtime started without the five PHY load groups Booted, then failed in ways that looked like MAC or calibration bugs The bootloader applies them
MAC +0x0230 and +0x0260 written with anything but 0x0400_0004 and 0x0400_0000 Association took 13–41 s instead of about 2 s Keep these values. They look like DTCM addresses but act like timing values inferred
Stack in high SRAM across the CONTROL_04 write Failed at that write, looking like a packet-memory fault Keep the stack in DTCM