Repository navigation
os/v3/puppy: ship 1bit-selftest in the appliance, and let it read the log this image writes - #1166
Conversation
…t is still waiting (engine#1159)
The cycle-26 board boot printed 'Waiting for partition [1BIT-DATA]............... failed' on a
healthy stick: Puppy's wait_for_blkid() polls blkid 15 times, one second apart, while the kernel
did not publish sda1/sda2 until 45.7 s - the stick's SCSI probe sat behind two 'usb 3-2: reset
high-speed USB device' cycles from the JetKVM's emulated media on the same xHCI controller. Every
gate was green because QEMU publishes the partition at 3.7 s.
patch-init.sh now raises the wait to 120 s and prints '*** 1bit: still waiting for [<token>] - a
slow medium is not a failure' at 15 s, and --verify refuses an initrd whose wait is under 90 s or
that waits silently - build.sh and mkimage.sh both run --verify on the packed initrd, so no image
can be cut with the 15 s wait again. usbgate.sh is the new row: it asserts that budget from the
shipped image's own initrd and then boots the image with its medium on USB mass storage on xhci.
Measured: the cycle-26 initrd fails the verify row ('the partition wait is the stock 15 s'),
cycle 27's passes; the cycle-27 image is 75 PASS / 0 FAIL / 3 SKIP on vm-check.sh (Secure Boot,
UEFI) and passes usbgate.sh. closes #1159
… log this image writes The install sheet and os/README.md both name `1bit-selftest` as the one command that proves the engine answers from THIS boot, and os/mini/mkimage.sh has always installed it - but the Puppy image never did, so pi-release's install gate has been reporting 'NOT SHIPPED' since the appliance took over at c15. container-build.sh now installs it beside 1bit-release-check, and the build REFUSES an image without it (the same presence list that already catches the other tools). The file always read /tmp/lemond.log for the engine's own NPU lines; the appliance's /etc/init.d/10lemonade runs lemond with /var/log/lemond-onebit.log, so those rows would have failed on the appliance for a path reason. It now reads whichever of the two exists and, when neither does, names both. Its mode row gets release-check.sh's measured fallback (/data/lemonade present, no /etc/1bit marker = LIVE): the appliance has no os/mini init, so nothing writes /etc/1bit/live there and the tool said 'mode: UNKNOWN' on its own image. Both images ship the same file, like release-check.sh: the paths are layout-aware rather than forked. Verified with a known-answer probe (~/wt/os3i/selftest-probe.sh: the real script in a mount namespace, /data + /etc/1bit + /var/log as fixtures, a stub wget for Lemonade). Appliance arm: mode LIVE, RESULT: PASS, RESULT-NPU PASS. os/mini arm (/tmp/lemond.log): still PASS. No-log arm: FAIL naming both candidates. The same probe on the pre-fix script (origin/main) gives mode=UNKNOWN and RESULT-NPU=FAIL on the appliance arm - the fixture bites in both directions. Not in cycle 27: that cut is initrd-only and is already handed off; this rides the next full build.
|
pi-prs: #1166 ready for review at |
bong-water-water-bong
left a comment
bong-water-water-bong
left a comment
There was a problem hiding this comment.
VERDICT: READY
Both defects are real on the shipped artifacts, and I checked them on the artifacts rather than on the description.
"NOT SHIPPED" is true of c26's filesystem, not just of the build script. unsquashfs -l on ~/wt/puppy-out-c26/puppy_upup64r_26.04.sfs (491,606,016 B):
squashfs-root/usr/local/bin/1bit-install
squashfs-root/usr/local/bin/1bit-release-check
<- no 1bit-selftest
and git show origin/main:os/v3/puppy/container-build.sh | grep -c 1bit-selftest -> 0, so the install -D -m 755 /repo/os/mini/selftest.sh "$skeleton/usr/local/bin/1bit-selftest" line this PR adds at container-build.sh:272 is the missing half.
The log path is the one the appliance writes. selftest.sh now tries both (for L in /tmp/lemond.log /var/log/lemond-onebit.log, :34-38) and the appliance's own init is LOG=/var/log/lemond-onebit.log (embed/etc/init.d/10lemonade:33), so RESULT-NPU can pass on a booted appliance instead of failing on a path.
All touched scripts parse: bash -n on container-build.sh, selftest.sh, patch-init.sh, usbgate.sh, and sh -n on selftest.sh, all clean.
One ordering fact that needs to be in the merge decision, not just in this comment: this PR already contains #1160. gh pr diff 1166 and gh pr diff 1160 produce byte-identical 79-line diffs for os/v3/puppy/patch-init.sh, and both add the same new os/v3/puppy/usbgate.sh (+0,0 +1,114), while both branches declare base=main. So they are stacked in content and presented as siblings to main:
- merge #1160 first, then #1166 is a much smaller PR (its
patch-init.sh/usbgate.shhunks disappear) and should be re-verdicted at the rebased head; - or merge #1166 and close #1160, since main would then already carry it.
Merging both as they stand would either conflict in those two files or land the same hunks twice, and git will pick whichever it can - the point of saying it here is that nobody should have to discover it in the merge.
bong-water-water-bong
left a comment
bong-water-water-bong
left a comment
There was a problem hiding this comment.
VERDICT: READY at f99f8775 — the head move does not touch this PR: its two files (os/mini/selftest.sh, os/v3/puppy/container-build.sh) are byte-identical to the head I reviewed (4d5f893e), so everything in my earlier verdict carries — 1bit-selftest shipped in the appliance, and it reads the log this image writes. Merge after #1160 (it contains that commit).
PR Reviewer Guide 🔍Here are some key observations to aid the review process:
|
…ree rows that could skip themselves (#1218) c29 ships /usr/local/bin/1bit-selftest (engine #1166) - the one-command proof the install sheet asks the operator to run - and no row ever executed it. The file was proved present and byte-identical to os/mini/selftest.sh, and proved by nobody that it runs in the appliance: the same class the U4 release-check row was added for in cycle 21. The new row does not demand 'RESULT: PASS'. The image boots with an empty /data/models, where the tool's own documented answer is 'RESULT: FAIL - no model to test', so demanding PASS would red a healthy image - and a row that reds a healthy image gets deleted. It requires a VERDICT: PASS with a model, SKIP naming the no-model reason without one, FAIL when the binary is absent or dies without a RESULT line. Its first version printed selftest-broken-missing and came back SKIP, and that is how the second half of this commit was found: Puppy's shell echoes the typed command into the row's own output, vm_check_verdict.skipping() matches a *-skip-* token anywhere in the body, and the command carried the marker literally - so the row was skipped by the echo of itself. Two rows already in the tree carry a literal marker and are exposed to the same thing. U1 shipped shells (#1186, merged tonight) was measured self-skipping, i.e. the row that exists to catch an unparsable script had never run; U2 console heals (#1093) is latent. All three now assemble the marker from m/sk, so the token exists only in the output. No row's semantics change. Proved: four arms under bash and sh on fixtures (absent / no-model / pass / dies silently); the module's own verdict mapping on the echoed command before and after (SKIP with token m-skip-no-model -> no token, so the output decides); and end to end on two shipped artifacts, c28 0c371992 (no such file -> FAIL selftest-broken-missing, and U1 now RUNS: 'shells-parse: 45 scripts') and c29 8266aff9 (SKIP selftest-skip-no-model on its blank models dir, 77 PASS / 0 FAIL / 4 SKIP). Co-authored-by: pi-os3 <os3@1bit.invalid>
Why
The install sheet (
~/wt/release-rc/vminstall-test.sh) andos/README.mdboth name1bit-selftestas the one command that proves the engine answers from the boot you are on.os/mini/mkimage.shhas always installed it — the Puppy appliance never did, so pi-release's install gate reports1bit-selftest: NOT SHIPPED in this image(its own note: "stale step, not a gate failure", which is exactly the kind of step that stays stale for ten cycles).Two defects, both measured on the shipped artifact rather than assumed:
container-build.shinstalls1bit-release-check,1bit-install,1bit-board, … and neverselftest.sh. Confirmed in the cycle-26 SFS:usr/local/bin/1bit-release-checkis there,1bit-selftestis not.selftest.shgreps/tmp/lemond.logfor the engine's own NPU lines; the appliance's/etc/init.d/10lemonaderuns lemond withLOG=/var/log/lemond-onebit.log. So even once installed,RESULT-NPUwould have failed on the appliance for a path reason. And its mode row has no marker to read:/etc/1bit/liveis written byos/mini/init.sh, which the Puppy image does not run, so it printedmode: UNKNOWNon its own image —release-check.shhas carried the/data/lemonadefallback since c12 and passes on the board with it, so that rule is reused, not invented.What changed
os/v3/puppy/container-build.sh: installselftest.shas/usr/local/bin/1bit-selftestbeside1bit-release-check, and add it to the presence list that makes the build REFUSE an image without it.os/mini/selftest.sh: engine log is whichever of/tmp/lemond.log(os/mini's init) and/var/log/lemond-onebit.log(the appliance's unit) exists, and when neither does the FAIL names both instead of guessing. Mode row gains release-check.sh's/data/lemonadefallback.One file, both images: the paths are layout-aware rather than forked, exactly as
release-check.shalready is.Verified — known-answer probe, both directions
~/wt/os3i/selftest-probe.shruns the real script in a mount namespace with/data,/etc/1bitand/var/logreplaced by fixtures and a stubwgetstanding in for Lemonade; three arms with their answers known before it runs.origin/main's script/etc/1bitmarker,/data/lemonade, log at/var/log/lemond-onebit.logRESULT: PASS, RESULT-NPU PASS/tmp/lemond.logprobe rc=0on this branch,rc=1onorigin/main's script — the fixture bites in both directions.sh -n,bash -n,shellcheck -S warningandtools/copyright.py --checkall clean.Not in cycle 27
Cycle 27 (
~/briefs/OS3-C27-HANDOFF.md, imagee3026513…) is an initrd-only cut and is already handed to pi-release; this rides the next full build. Until then the install gate keeps printingNOT SHIPPED, which is honest.Nothing merged, tagged or published.