update

2026-08-18 09:20:44 +02:00
commit b826cd721a
522 changed files with 93730 additions and 0 deletions
@@ -0,0 +1,41 @@
+# repo22 legacy full-index fix manual report
+
+- Date: 2026-08-08 UTC
+- Source baseline: `377d467a652beb2124d4adc0d1f8e735fca22344`
+- FreeBSD guest: 15.0-RELEASE-p8 amd64, QEMU TCG
+- Module SHA256: `f3b9fa70bfb8e98e2cef436583fd13cb91787514a5be76f9f108698d7586ca3c`
+- Scope: legacy full-index and partial-reference minimum regression only
+
+## Root cause and fix
+
+The proven legacy Layout 1 fixture maps its second extent as `HEAD1` with
+`clusterofs=28`, `pblk=2`, `m_la=639004`, `m_llen=409572`, and
+`m_plen=4096`. The LZ4 stream completes after consuming 2627 bytes; the
+remaining 1469 bytes of the pcluster are zero padding. The repo22 decoder
+incorrectly required all 4096 input bytes to be consumed and returned `EIO`.
+
+The decoder now accepts complete output only when the remaining pcluster bytes
+are all zero, while continuing to reject nonzero trailing data. The mapping
+recorder also no longer clears `compressedblks` and `partialref` on every
+lcluster load, matching the Linux state machine and preserving lookback state.
+
+## Results
+
+| Test | Mount | Full SHA256 | Boundary read | Result |
+|---|---:|---|---|---|
+| Legacy full index `/shape.dat` | PASS | `370eb0a8df86868c4842ca535ed64670f0277ea2ed47a703f089bbb13ee4ac52` matches | offset 638972, 128 bytes: `941b6e3cb9a7384428d93ca248eb9630782538e04caba30dbfbe31baacca8f60` matches | **PASS** |
+| Partial reference `/b.dat` | PASS | read returns `EIO` | offset 641020 read returns no data | **KERNEL-FAIL** |
+
+`./build.sh`, `kldload`, and exact-name `kldunload` passed. The isolated dmesg
+delta was empty, with no panic, trap, or integrity diagnostic.
+
+## Remaining partial-reference hypothesis
+
+The minimal reproducer is the existing
+`lz4-partial-ref.erofs` fixture: mount it and read `/b.dat`; the companion
+compact `/a.dat` remains the known-good source extent. The full-index mapping
+blocker is removed, but `zdata.c` still invokes `z_erofs_decompress()` with
+`partial=false` for every mapped extent even when `EROFS_MAP_PARTIAL_REF` is
+set. Linux propagates this state to partial LZ4 decoding. Per the convergence
+instruction, this batch records that hypothesis without further debugging or
+additional source changes.