Skip to content

sx1262: a frame the receiver locked onto and lost reaches the chip - #24

Closed
A13xB0 wants to merge 1 commit into
meshbench-mainfrom
failed-frames-reach-the-chip
Closed

A13xB0 wants to merge 1 commit into
meshbench-mainfrom
failed-frames-reach-the-chip

Conversation

@A13xB0

@A13xB0 A13xB0 commented Oct 16, 2026

Copy link
Copy Markdown
Collaborator

The QEMU half of MeshBench/meshbench#757.

The engine now sends kind 0x0A, [failure][frame...], for a frame this node locked onto and lost. The device hands it to vsx_deliver_frame_failed, so the chip raises HeaderErr alone for a header it could not read, or RxDone beside CrcErr for a payload it could not.

vsx_deliver_frame_failed is bound optionally: it arrived in virtual-sx1262 1.3.2, and refusing to load an older chip model over an entry point only this kind needs would break boards that work today. With an older model the kind is skipped, which is what happened to it before.

Checked: hw/ssi/sx1262.c compiles with this tree's own flags from compile_commands.json (-fsyntax-only, warnings as errors). Not booted on a board yet.

The engine sends kind 0x0A, [failure][frame...], for a frame this node's
receiver locked onto and could not trust, and the device hands it to
vsx_deliver_frame_failed. The symbol is bound only if the chip model has it,
so a model older than 1.3.2 loads as before and the kind is skipped.
@A13xB0

A13xB0 commented Oct 16, 2026

Copy link
Copy Markdown
Collaborator Author

Closing for now. Today's native bridge hangs up on a message kind it does not know, so MeshBench/meshbench#757 now sends 0x0A only to a node that says it takes it, in byte 47 of its radio stats record. This device's record is 43 bytes, so it cannot say so yet. It comes back with the record grown to carry the IRQ suppressed count and the capability byte, tested on a board.

@A13xB0 A13xB0 closed this Oct 16, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants