This commit is contained in:
2026-08-18 09:20:44 +02:00
commit b826cd721a
522 changed files with 93730 additions and 0 deletions
+344
View File
@@ -0,0 +1,344 @@
# Pre13 Final Report
## Conclusion
Pre13 is **scoped PASS** for the planned and admitted Pre13 work:
- DUT baseline snapshot: `a587d489b829ef8e7fac94c87130b61d47aa3e44`.
- Final DUT commit: `b804d7b77d1c76b8e844dc4a758c6ab3381ca006`.
- At final report time, `HEAD`, local `main`, fetched `xdm/main`, remote
`refs/heads/main`, and the only DUT worktree all resolve to
`b804d7b77d1c76b8e844dc4a758c6ab3381ca006`.
- DUT working tree is clean.
This is **not** a full-filesystem certification. No full 161-case feature
suite was run, no positive V02 explicit-extent fixture was produced, and Pre13
does not claim that unrun cases such as TC004, TC011, TC143, TC144, TC145,
TC150, or other non-targeted feature cases passed.
## Commit Inventory
The audited DUT range is
`a587d489b829ef8e7fac94c87130b61d47aa3e44..b804d7b77d1c76b8e844dc4a758c6ab3381ca006`.
### Planning and Stage0
- `76a4c82`, `fd300cb`, `2f487f0`: Pre13 planning documents and low-risk batch
accounting.
- `8ed80c3ddbedc64b8b80daf1be16b28c4f24f10d`: recorded Stage0 gate decisions
in `repo-pre-13/docs/pre13-stage0-decisions.md` and the initial execution
status.
### Low-Risk Source Alignment
- `2d0041ba3a195c184a52bdccae3bd00af520d4b7`: L-A1 helper and guard alignment
across `compress.h`, `data.c`, `erofs_fs.h`, and `inode.c`, with
`pre13-low-a1.md`.
- `190a0b908d25e62b5c78ba82377f53d359fc5c4f`: L-A2 local helper ordering in
`decompressor.c` and `xattr.c`, with `pre13-low-a2.md`.
- `f236836a55ad7ec7d536df483460ea394e38b6c3`: L-B semantic `erofs_nid_t` and
`erofs_off_t` alignment across data, directory, inode, namei, super, xattr,
zdata, and zmap paths, with `pre13-low-b.md`.
- `217b7a3032ec917c5ea63652b0f3cad734990d48`: L-C1 on-disk `__leXX` field
alignment in `erofs_fs.h` and initial `pre13_ondisk_layout_probe.c`, with
`pre13-low-c1.md`.
- `e72ecd7cc024047f0ce7c1855e6e44ae7f5350e1`: L-C2 flexible on-disk arrays
and layout probe updates, with `pre13-low-c2.md`.
- `fe338f701059de8dfd6336dea7f9288bc0f12e9f`: L-D LZ4 offset endian helper in
`decompressor_lz4.c`, with `pre13-low-d.md`.
- `9028b92343ec7f5bbc1b07477b8d4a669c436955`: L-E localized private
constants, removed `erofs_defs.h`, and updated consumers/docs.
### Behavior Changes and Fixups
- `ee9dc4715436eeee21ae2d84eab2677b27d87ddd`: H04 rejects non-directory root
inodes during mount, with `pre13-h04-root-validation.md`.
- `f51b0fe1ceb27cac641667aa5688d371a60d2d99`: H03a introduced the bounded
qstr comparator.
- `3313d305a2193ef6ffa26c6a189e47187b9c0d70`: H03b passed qstr through block
and dirent lookup search.
- `83a5bef73b329c967bd8390a0966d30005295758`: H03c adapted VOP lookup to the
bounded qstr path.
- `dc47e8f99c947d2929d3528a6c139a85eb4251d5`: fixed const-correct build inputs
and the userspace layout probe `bool` include issue.
- `b804d7b77d1c76b8e844dc4a758c6ab3381ca006`: retired stale decompression test
harnesses that no longer matched the current module surface; this did not hide
a DUT test failure.
## Stage0 Gate Results
Stage0 was a mixed gate result, not an overall PASS:
- Gate A, ROOT-NONDIR: **READY**. The baseline accepted a non-directory root;
Pre13 froze the rejection as `EINTEGRITY` and allowed H04.
- Gate B, ZERO-RUN: **STOP/no-op**. The analysis covered 1,360 legal states and
found no reachable zero-progress state for H07.
- Gate C, map tuple oracle: **STOP**. The oracle covered only 26 plain,
inline, and hole tuples; it did not cover chunk, multidevice, bounds, or
overflow behavior, so H01a/H01b/H01c were not implemented.
- Gate D, qstr comparator: **PASS/READY**. The Python comparator corpus passed
11,562 cases and the independent C harness passed 1,048,576 cases.
- Gate E, device option probe: **blocked/non-gating**. It produced no runtime
behavior claim and did not block Pre13.
## Intentional Stops and Rejects
- H01 map objectization: **STOP** because the tuple oracle was incomplete for
chunk, multidevice, bounds, and overflow coverage.
- H07 zero-run: **STOP/no-op** because 1,360 legal states produced no reachable
zero-progress path.
- H05 header-only xattr: **rejected**; Linux treats the format as undefined and
BSD already returns `EOPNOTSUPP`.
- H06 duplicate `device.N`: **rejected/no-op** for source changes because
`vfs_sanitizeopts()` makes the EROFS duplicate-option check unreachable; any
probe is non-gating.
- H02 inode split: **rejected** because it would diverge from Linux's single
`erofs_read_inode()` switch and create low-value churn.
- H08 pager errno propagation: **rejected** because the target
`vnode_create_vobject()` path observably returns 0.
- V02 explicit extent positive coverage: **PARTIAL**. No positive mapped
fixture was available, and TC157 negative coverage is not a substitute.
## Build, Layout, and Probe Evidence
Primary evidence:
`/work/tests-dev/temp/pre13-targeted-final-runnerfix-retry-20260814T0138Z`.
After `dc47e8f` and `b804d7b`, the targeted retry produced:
- `WITH_ZSTDIO=0` build: PASS, module SHA256
`0b1244cf9b7fe70649bdcc1c6ef12e6151c77b10c2755b366f960b101ccb7220`.
- `WITH_ZSTDIO=1` build: PASS, module SHA256
`15db1fbf9450259d030e88602e1c5ba2e452a731c641b359bd5a2d7823da2735`.
- Loaded module in that targeted run: SHA256
`15db1fbf9450259d030e88602e1c5ba2e452a731c641b359bd5a2d7823da2735`.
- Userspace layout probe: build and run return code 0.
- Four userspace probe builds: `g3_vfs_probe`, `mount_errno_probe`,
`nfs_fh_tool`, and `readdir_probe` all returned 0.
The first targeted runner attempt still ended host-side FAIL because TC141
included an extra empty `umount ''` command. That was classified as a test
runner false negative, not a DUT failure: the 25 real TC141 commands all had
return code 0, and only the 26th synthetic empty unmount failed.
## H04 Root Validation
Evidence path:
`/work/tests-dev/temp/pre13-targeted-final-runnerfix-retry-20260814T0138Z/guest-evidence/cases/H04`.
H04 custom fixture result: **PASS**.
- Valid root fixture SHA256:
`d0baba501ae12fcdb8869a98908486473d2bcfb6b332a47cb6a6c1f856a3ab34`.
- Non-directory root fixture SHA256:
`60bf0d0d93baabdfa23b71f48283d317f6b7218852f3d19aee77575c41500db7`.
- Root proof payload SHA256:
`0ebc44a6c5a02ba458d1a115bc338fa4acb4d973a37804c9b926eafc0c53575f`.
- Valid root mounted, root mode was `040755`, payload matched, and unmount plus
`md` detach succeeded.
- Non-directory root returned exact `mount_errno=97` (`EINTEGRITY`) with message
`erofs: root inode nid=36 is not a directory`; the failed mount left the
provider detachable.
- The valid fixture remounted successfully after the non-directory failure.
- Per-case dmesg delta and suspect logs were empty.
TC011 and TC150 were not run and are not claimed as passing. They are not
Pre13 blockers because the dedicated H04 fixture directly exercised the actual
root-type change, including the valid control, exact errno, detach, and remount
properties.
## H03 Targeted Feature Evidence
H03 targeted cases exercised the qstr lookup changes and passed for:
TC042, TC043, TC044, TC045, TC053, TC054, TC123, TC141, and TC148.
Evidence:
- TC042/043/044/045/053/054/123 and TC141 raw evidence:
`/work/tests-dev/temp/pre13-targeted-final-runnerfix-retry-20260814T0138Z`.
- TC148 valid copied evidence and final smoke summary:
`/work/tests-dev/temp/pre13-missing-and-smoke-20260814T021330Z/final-summary.md`.
TC141 details:
- The runner logged 26 commands, but the last was the invalid empty
`umount ''` cleanup bug.
- The 25 real commands passed: valid-base and valid-padding cold lookup/read,
repeated missing lookups returning `ENOENT`, `entries=322` wide-directory
checks, and three corrupt lookup fixtures returning repeated
`EINTEGRITY` (`errno=97`).
- The only failure was the empty unmount runner command. Its stderr was
`statfs: No such file or directory` / `unknown file system`, with no DUT
dmesg delta.
TC148 details:
- Outcome: PASS.
- Fixture SHA256:
`9a94e9af2cab264b9c11975a20d78e615d6fe1e6f86267173cb5ac86aecb2b17`.
- Mount command return code: 0.
- Expected and actual regular-file count: 320.
- Directory oracle: 322 entries, with two directories and 320 regular files.
- `d_off` restart count and `seekdir` restart count: 322 each.
- Target content matched.
- Per-case dmesg delta and suspect log were empty.
## Final Smoke Evidence
Evidence path:
`/work/tests-dev/temp/pre13-missing-and-smoke-20260814T021330Z/final-summary.md`.
Final smoke outcome: **PASS**.
- DUT archive SHA256:
`d0c052bfef024d0acb40e107800cc8e8227fab6b691e8825af5cf01f19672fe4`.
- Assets SHA256:
`a9741ea468493bdc72d399615892d7f6269226467d09fa6613fa01352f011908`.
- Probe bundle SHA256:
`7b10d23c8a780fb35810fe85626305a8c49f196d270f98e5224045bd79f51066`.
- Clean tests-dev worktree commit used for the smoke:
`f86799654a37cdb81025b9e703635f7c7812fecb`.
- `WITH_ZSTDIO=0` build: PASS, module SHA256
`0b1244cf9b7fe70649bdcc1c6ef12e6151c77b10c2755b366f960b101ccb7220`.
- `WITH_ZSTDIO=1` build: PASS, module SHA256
`15db1fbf9450259d030e88602e1c5ba2e452a731c641b359bd5a2d7823da2735`.
- Loaded module for final smoke: `WITH_ZSTDIO=0`, SHA256
`0b1244cf9b7fe70649bdcc1c6ef12e6151c77b10c2755b366f960b101ccb7220`.
- `kldload`, `kldstat -n erofs.ko`, `kldunload`, and final KLD cleanup: PASS.
- Plain fixture SHA256:
`3785ca07e7bd16f6c611191596ae0314253c0ae9b7217a4b22de25799b0f08ac`;
payload SHA256 matched
`056f8f7585667dc695e2edf936deaf84cfee0f88671ba6cd5be22ce890763433`.
- LZ4 fixture SHA256:
`967cc1b625546f9f9f881472e71cc3d849c65feb4e98e67fbfdc9b90a4be6dda`;
payload SHA256 matched
`3ff012b76087c4da65ce0b69813a76f47ea367e95782edf8b8cd6cd3ec4d1880`.
- LZMA fixture SHA256:
`32107a084b27362a093768b88746c37c2e99a74b9f1301a9d4046988479defe9`;
payload SHA256 matched
`ddda39737f0f6093e828a032ec161511fefbb1fa361bc6cbffdbc91e48e4c461`.
- Every attach, mount, hash, unmount, and detach command returned 0.
- Per-case and global dmesg suspect logs were empty.
- Global mount, md, KLD, QEMU, ControlMaster, overlay, and ports
`10056`, `10057`, `10058` cleanup passed.
- Guard QEMU PID `26318` and port `9222` were unchanged and reachable.
- The main tests-dev worktree was dirty before and after with unchanged diff
SHA256 `92d3808ab33ff9b22fa2ee3f8a854c1fc36effecb4dd6c0917ae7e80c57fc23a`;
the clean tests worktree, not the dirty main worktree, was used as evidence.
The final summary explicitly states that no other TC, full feature suite,
TC004, TC143, TC144, or TC145 was run. Pre13 preserves that limitation.
## V01 Race Evidence
Raw evidence:
`/work/tests-dev/temp/pre13-v01-race-20260814T031032Z/retry-fresh-20260814T032500Z`.
tests-dev evidence commits were verified on `xdm/main`:
- `2269908cc3f721d57ae8138faef664a6bc4c7cc9`: added the V01 concurrent unmount
race runner.
- `374c9d8473e6638fa0c8e865fcfcfee141b13f0d`: fixed runner SSH command
execution.
- `22c0bea73b9b81cc44385d1ba7103fa820c56070`: corrected forced-unmount deadfs
errno classification and added the offline analyzer/unit tests. Fetched
`xdm/main` and remote `refs/heads/main` both resolve to this commit.
Original raw run:
- Raw host/guest outcome: FAIL, `guest race failed rc=41`.
- DUT commit: `b804d7b77d1c76b8e844dc4a758c6ab3381ca006`.
- Module SHA256:
`15db1fbf9450259d030e88602e1c5ba2e452a731c641b359bd5a2d7823da2735`.
- Fixture root-valid SHA256:
`d0baba501ae12fcdb8869a98908486473d2bcfb6b332a47cb6a6c1f856a3ab34`.
Normal group:
- Duration 90 seconds, 8 workers, 154 loops.
- `umount` returned accepted `EBUSY` 154 times; no successful unmount/remount.
- Worker read/stat/open/fstat success counts were each 79,164.
- Worker content mismatch, short reads, rejected errors, and dmesg suspect bytes
were all 0.
Forced group:
- Duration 90 seconds, 8 workers, 125 loops.
- Successful forced unmount/remount cycles: 125/125.
- `stat ENOENT`: 102,165; `open ENOENT`: 16.
- Raw rejected errors: 1,772, consisting of 882 `fstat EBADF` and 890
`read ENXIO`.
- The final mounted SHA matched the expected payload SHA in both normal and
forced groups.
Final analyzer result from
`scripts/analyze-pre13-v01-race-evidence.py` at
`22c0bea73b9b81cc44385d1ba7103fa820c56070`:
- `normal_verdict=PASS`.
- `forced_verdict=PASS`.
- `forced_expected_deadfs_errors=1772`.
- `forced_expected_deadfs_fstat_ebadf=882`.
- `forced_expected_deadfs_read_enxio=890`.
- `dut_verdict=PASS`.
- `test_verdict=false-negative-corrected`.
- `overall_verdict=DUT PASS / TEST false-negative-corrected`.
The `EBADF` and `ENXIO` results occur only after successful forced unmounts,
when existing descriptors are taken over by `deadfs`; that is standard FreeBSD
semantics and not a DUT instability.
The analyzer's four offline unit tests all passed:
- forced deadfs errors are corrected;
- normal deadfs errno is rejected;
- forced unrelated errno is rejected;
- count mismatch is rejected.
V01 cleanup and integrity:
- QEMU stopped, V01 port released, overlay removed.
- Base image mode, size, and mtime remained `444`, `16515530752`, and
`1786418212`.
- Guard VM was unchanged and reachable.
- Guest mount/md/KLD owned-resource cleanup passed.
- Total dmesg delta was 24 lines / 1,266 bytes, all known non-suspect
`mangled entry` lines; suspect bytes were 0. This is not a zero-total-delta
run.
- The dirty main tests-dev worktree was unchanged; the independent clean
worktree was clean.
## Static Review and Evidence Boundaries
Final static review of the Pre13 source delta found no production-source
blocker, test backdoor, out-of-bounds issue, ABI/layout issue,
lock/lifecycle issue, or errno issue requiring another DUT fix. The one
dynamic item that was pending at that review point was documentation/status
backfill; that was a reporting-sequence issue, not an unperformed test gate.
The source/test diff boundary was also reviewed:
- Production changes are limited to the planned low-risk alignment work, H04,
H03, and build-input fixups.
- The only new tracked test asset in the DUT repo is the userspace layout probe.
- The stale decompression harness removals in `b804d7b` were documented and do
not hide a failing Pre13 DUT test.
## Unrun or Non-Claimed Coverage
Pre13 deliberately does **not** claim:
- full feature-suite PASS;
- TC004 PASS;
- TC011 or TC150 PASS;
- TC143, TC144, or TC145 PASS;
- TC146/V02 positive explicit-extent PASS;
- H01 map objectization behavior;
- H07 zero-progress behavior;
- H05/H06/H02/H08 source behavior changes.
The custom H04 fixture directly covers the actual H04 source change, so the
absence of TC011/TC150 is not a blocker for the scoped Pre13 result. Likewise,
the final smoke covers Plain/LZ4/LZMA smoke only; Deflate, ZSTD, partial, and
other codec feature cases remain outside the Pre13 claim unless specifically
listed above.