Robert Allan JamesandClaude Sonnet 5
71bcb72a59
sk_repl_idle()'s messaging pump (repl.c) walked the entire live-VM registry every idle beat (~1Hz) and dispatched a full VM-EXEC "MSG-TICK" -- a dictionary lookup plus a 32-slot arena scan -- into every live VM, every tick, unconditionally, forever. Fine at Tripod's original 3-VM scale; a full 9-identity turn-attractor campaign exposed it as a genuine wall on riscv64 specifically (its TCG makes each dispatch cost more): the same campaign that completed in 194s/ 339s on amd64/aarch64 never finished on riscv64 at 9 VMs across three attempts, while 8 VMs there was fine in 159s. Ruled out capacity explanations before touching anything: bumping riscv64's QEMU RAM 1024->4096 changed nothing (reverted), and a live STADIUM-RES@/MSG-STATUS probe with all 9 VMs attached showed no depletion. Host memory pressure was also ruled out directly (one background task did get OOM-killed once during the investigation, but the identical stall reproduced again with 9.2GB free). The real mistake was three premature kills under 4 minutes with no way to tell "slow" from "stuck" from outside the guest -- fixed by having run_doe_batch.sh sample the qemu process's own /proc/<pid>/stat utime every 60s; with that signal, riscv64 at 9 VMs was unambiguously alive (climbing utime, no hang), just disproportionately slow going from 8 VMs (159s) to 9 (600s+ and climbing). This was never really a riscv64-only bug: an O(N) per-second walk over the full VM population doesn't scale to the hundreds of VMs this fleet is headed toward, on any architecture -- riscv64 just made it visible first, at N=9, because its per-dispatch cost is highest. Fixed by round-robin batching: the pump now dispatches to at most SK_MSG_PUMP_BATCH (4) live VMs per idle beat via a persistent cursor that resumes where the previous beat left off, instead of all of them every time. Bounds both the scan and dispatch cost to O(K) regardless of total VM count; any single VM's queue now drains roughly every ceil(N/K) beats instead of every beat, still bounded and still matching the pump's own existing best-effort contract. No new C primitives, no messaging/Stadium changes. Verified: clean build on all 3 architectures, then the full 3x9x3 campaign re-run on all 3 (not just riscv64) per the standing rule that a defect repair requires a clean re-run everywhere before anything counts as closed: amd64 198s 0 faults 99/99 tokens x8 656/656 K-conserved aarch64 339s 0 faults 99/99 tokens x8 659/659 K-conserved riscv64 178s 0 faults 99/99 tokens x8 659/659 K-conserved riscv64 went from "never completes" to faster than aarch64, same campaign, same seed, same identity set. No regression on amd64/ aarch64. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EXieurDfDSsDFdnSyusuWo
disk/
QEMU disk images used for Artemis (block-storage VM) persistence testing.
Mounted via Makefile.starkernel's ARTDISK variable (default
disk/artemis.img) as a virtio-blk-pci device on all three
architectures' kernel QEMU boots — not scripts/rundisk.sh, which targets
a separate, currently-unused disks/ (plural) directory for the hosted
VM's --disk-img= flag instead.
artemis.img— standard Artemis persistence test disk. Reformatted 2026-08-02: this image had been stuck in a corrupted state (validLithosAnankemagic header, but data not matching whatART-READ-TESTexpects) since before this repo's own git history begins (git logshows it already broken at the initial commit, carried over from the pre-split monorepo). Chasing down the resulting persistentFAIL: persist-readtraced to the data, not the code — the write→reboot→read round trip works correctly on a fresh image (seeartemis-debug-roundtrip.imgbelow). Reformatted by blanking the file and letting a normal boot format+write-test it; verifiedPASS: persist-readon amd64, aarch64, and riscv64 against the same image afterward (cross-arch resume, matching the arch-neutral on-disk format.claude/ARTEMIS.mdspecifies).artemis-debug-roundtrip.img— round-trip regression fixture created during that investigation. Known-good: format → self-test → write-test → reboot → resume →PASS: persist-read, confirmed 3 times in a row. Keep this in a passing state; if a future change breaks it, that's a real regression, not a stale-fixture artifact likeartemis.imgwas.artemis-persist-test.img— persistence round-trip test image (pre-existing; history/state not re-verified during the above investigation).artemis-poison.img— a separate test image (exact scenario not documented elsewhere in the repo as of this writing; name suggests an adversarial/corruption test, not confirmed).artemis-unrecognized-test.img— exercisesART-HALT-UNRECOG(.claude/ARTEMIS.mdacceptance criterion #6). Regenerated 2026-08-02: the previous copy of this file had itself been silently reformatted by a since-fixed bug in the generic block subsystem (src/block_subsystem.c) — it carried a valid low-level'STFR'/v2 header despite being meant to represent foreign disk content, direct forensic evidence of the bug described in.claude/ARTEMIS.md's Build Status item 6. Regenerated as 30MB of a repeatingPOISON-UNRECOGNIZED-DISK-TEST-FIXTURE--NOT-BLANK-NOT-STFR-NOT-ARTEMIS--ASCII pattern — deliberately neither blank, nor the block subsystem's own'STFR'magic, nor Artemis's"ARTEMIS\0"marker. Verified on amd64 and riscv64 post-fix: boot correctly halts (ARTEMIS HALT: unrecognized disk content) and the file's sha256 is now byte-for-byte identical before and after boot. Keep this fixture in this poisoned state — if a future change makes its sha256 change across a boot, that is exactly the regression this fixture exists to catch.
Note on incidental header churn: artemis.img picks up a few changed
header bytes on every ordinary boot even though no user data changes —
blk_subsys_attach_device() always records a fresh mounted_time on a
successfully recognized disk, which gets flushed at shutdown. This is
expected bookkeeping, not a bug; revert it before committing rather than
carrying timestamp noise in git history.
-
usb-thumbdrive-test.img— 64MB raw image backing a QEMUusb-storagedevice attached to the xHCI controller's bus (xhci0.0) for Milestone 2e/ 2h hotplug testing, added 2026-08-22. Blank (all zero) —blkio_usb.c+blk_subsys_attach_device()wiring (Milestone 2h, done 2026-08-25) attach it asBLK_FMT_PROVISIONALevery time, which is the intended, exercised state; not yetBLK_FMT_FORMATTEDviaBLK-CONFIRM-FORMAT. -
usb-thumbdrive-test2.img— 64MB raw image, added 2026-08-25 for Milestone 2h hot-detach/re-attach verification. Filled with a repeatingHOTDETACH-REATTACH-FIXTURE-2026-08-25--ASCII pattern, deliberately distinguishable fromusb-thumbdrive-test.img's all-zero content — the point is proving a block read after detachingusb-thumbdrive-test.imgand re-attaching this one actually returns this pattern rather than silently replaying the old device's cached (all-zero) content, which is exactly the class of bug a same-LBN-range device swap can cause if the VM block window cache (vm->blk_vm_cbuf[]) isn't re-validated on a hit. -
artemis-reloc-test.img— 64MB raw image, added 2026-08-25 for single-block relocation (RELOCATE-BLOCK) persistence verification. Blank at creation; a genuinely fresh volume was required (notartemis.img) because relocation-table capacity (reloc_start/reloc_devblocksinblk_volume_meta_t) is only reserved byblk_compute_fresh_geometry()on a fresh format —artemis.imgpredates the feature and correctly reads backreloc_devblocks=0from what was previously unused header padding. Attach viamake -f Makefile.starkernel ARTDISK=disk/artemis-reloc-test.img ...(the Makefile'sARTDISKvar is?=-overridable). Not yetBLK-CONFIRM-FORMAT-committed in the repo copy — commit that step live if reusing this fixture for further reloc-table testing. -
artemis-metafence-fresh.img— 30MB raw image, blank at creation, added 2026-08-26 for the top-of-device system-metadata fence (meta_fence_blocksinblk_volume_meta_t, FABRIC-2.md Phase 8 §C). Same reasoning asartemis-reloc-test.imgabove: the fence is only initialized toBLK_META_FENCE_INIT(128) byblk_compute_fresh_geometry()on a fresh format, so a genuinely blank volume was needed to exercise that path. Verified round-trip: fresh format writes 128 (confirmed via direct byte read at header offset 184, independent of kernel self-report), a second boot without reformatting reads it back unchanged. Keep in its formatted (meta_fence_blocks=128) state — a future change that resets or corrupts this on reload is a real regression. -
artemis-metafence-test.img— 30MB raw image, a direct copy of the pre-existing (pre-fence)artemis.img, added 2026-08-26 alongside the fixture above to verify the other direction: an old volume that predatesmeta_fence_blockscorrectly reads it back as 0 (graceful default via zeroed former padding, not corruption) rather than crashing or misreading adjacent fields. -
zuse.img— 64MB raw image simulating the physical Zuse superuser thumbdrive for QEMU testing (FABRIC-2.md, Phase 8:zuse.img"bleach" mechanism, added 2026-08-26). 64MB is only this fixture's size, not a constraint on real home-blocks thumbdrives (Captain Bob, 2026-08-26) — the design is not bound to any particular drive size;homeblocks_sig_t's ownmetadata_devblocksfield records whatever size is actually observed on a real drive, nothing in the format hardcodes one. 64MB here is purely an arbitrary QEMU-test convenience (matchingusb-thumbdrive-test.img's existing precedent for BOT driver testing scale); usescripts/bleach_zuse_img.sh --size-mb Nfor a different test size. Blank (all zero) at creation — reads back asHOMEBLOCKS_SIG_BLANKviahomeblocks_sig_check(), confirmed live (xhci: USB drive not recognized (blank or foreign media)), simulating a genuine first boot for exercising the still-to-be-built one-time mint-Zuse flow. "Bleach" it back to this pristine/unminted state withscripts/bleach_zuse_img.shbefore each first-boot test run, rather than hand-regenerating the file — same all-zero content either way, the script just makes the reset a single documented, repeatable command. Deliberately a flat/raw image, not GPT-partitioned, matchinghomeblocks_sig_check()'s current call site (repl.c,sig_start_fblock=0) — both will move to a real GPT-partition-relative offset together once a GPT parser exists, not attempted ahead of that.
Convention, standing as of 2026-08-22: every virtual disk/thumb-drive image
used for testing — Artemis persistence disks above, and USB Mass Storage
backing images alike — lives in this directory and is a tracked, committed
file, never scratchpad. This was already artemis.img's convention;
usb-thumbdrive-test.img and any future USB test images follow the same
rule. Confirmed no .gitignore in this repo excludes disk/*.img.
These are regenerable QEMU raw disk images, not source — see
.claude/ARTEMIS.md for the storage model they exercise.