devsoak.quirks

# devsoak.quirks -- known-driver-behaviour database (implementation brief S16)
#
# Plain text, hand-parseable on Kickstart 1.3. One "quirk" block per entry:
#
#   quirk   <id>
#   match   key=value key<value key>=value ...   (repeatable; ALL conditions
#                                                  across ALL match lines in
#                                                  the entry must hold - AND)
#   because <free text, may wrap onto further indented lines>
#   action  <verb> <args...>                     (repeatable)
#   status  confirmed|reported|suspected
#
# '#' comments and blank lines are ignored. Match keys understood by
# devsoak's parser: name=STR (case-insensitive exact match on the driver's
# ln_Name), version</<=/>/>=/= N and revision</<=/>/>=/= N (the driver's
# lib_Version/lib_Revision), id~STR (case-insensitive substring of
# lib_IdString), kick</<=/>/>=/= N (the running Kickstart's exec.library
# lib_Version). Any other key (e.g. a SCSI INQUIRY string, an AttnFlags
# bit, a specific expansion board) is not evaluable by this parser: an
# entry using one is flagged unmatchable at load time and printed as a
# warning, and can then only ever be applied via -k.
#
# Actions: skip CMD... (trailing '*' wildcard prefix, e.g. NSCMD_*),
# expect CMD ERROR, warn TEST, skiptest TEST, align N, nochip,
# maxinflight N, maxxfer BYTES (decimal or 0x hex), norandomcmd, tier N.
#
# Status policy (S16.2): only "confirmed" entries are applied automatically
# against a reference driver. "reported" (seen in the wild, not reproduced
# by us) and "suspected" (a guess, or a first crash devsoak itself hit)
# entries are inert until named with -k on the command line. Nothing in
# this file may be promoted to "confirmed" on the strength of a document
# alone -- only after it has actually been reproduced with devsoak against
# real hardware or a real emulator build. Every entry below except the
# last is exactly the S16.5 seed table, left at the status that table
# gives it; the implementer of this file did not promote any of them.

quirk   cbm-scsi-pre40-no-nsd
match   name=scsi.device version<40
because Commodore scsi.device before V40 predates NSD. NSCMD_DEVICEQUERY
        correctly returns IOERR_NOCMD, but the TD64 dialect (24-27) may be
        dispatched through a jump table that was never range-checked
        against these command numbers.
action  skip TD_READ64 TD_WRITE64 TD_SEEK64 TD_FORMAT64
action  skip NSCMD_*
action  expect NSCMD_DEVICEQUERY IOERR_NOCMD
status  suspected

quirk   cbm-scsi-1.3-no-getgeometry
match   name=scsi.device version<36
because TD_GETGEOMETRY is a V36 command; on the 6.x-ROM scsi.device found
        on an A590/A2091 it is unimplemented. devsoak should expect
        IOERR_NOCMD and fall back to the RDB partition size for its
        bounds tests rather than treating the failure as a bug.
action  expect TD_GETGEOMETRY IOERR_NOCMD
status  reported

quirk   cbm-ide-cf-geometry
match   name=scsi.device version>=40
because On some IDE units with a CompactFlash card fitted, TD_GETGEOMETRY
        has been observed to fail even on a V40+ scsi.device (noted in the
        devtest README). This match can only key on the driver name/
        version, not on "IDE unit" or "CF medium fitted" -- those would
        need an id~ or inquiry~ match this parser does not support, so
        this entry is broader than the real condition. Downgrade a
        geometry-stable failure to a warning rather than skip it outright.
action  warn geometry-stable
status  reported

quirk   a3000-sdmac-word-align
match   name=scsi.device
because On an A3000 (SDMAC + Ramsey), non-LONGWORD-aligned DMA buffers
        are unsafe. devtest's README reported odd-address trouble (last
        two bytes unwritten); devsoak reproduced something stronger
        under Copperline's SDMAC emulation, scsi.device 47.4: a read
        into a word-aligned (+2) buffer arrives shifted left by two
        bytes wholesale. align 4 (not 2) makes every soak clean. There
        is no A3000-vs-other-machine match key in this parser, so this
        entry matches every scsi.device everywhere; it is deliberately
        left at a status that never auto-applies. Force it with
        -k a3000-sdmac-word-align when the run is on an A3000.
action  align 4
action  warn payload
status  reported

quirk   a2091-dma-align
match   name=scsi.device
because The A2091/A590 DMA controller needs word-aligned buffers. As with
        the A3000 entry above, there is no way to key specifically on
        "this scsi.device instance is driving an A2091" from ln_Name/
        lib_Version/lib_IdString alone, so this is broad-matching and
        force-only: use -k a2091-dma-align when the run is known to be on
        an A2091/A590.
action  align 2
status  reported

quirk   a2091-37-unaligned-length-hang
match   name=scsi.device version=37 id~SCSI/XT
because scsi.device 37.64 from the A2091 PDMA-07 boot ROM hard-wedges
        the machine (black screen, no recovery) on a CMD_READ whose
        io_Length is not a sector multiple. Reproduced with devsoak
        under Copperline emulation 2026-09-02: the run stops exactly at
        the unaligned-length matrix test every time, and skipping that
        one test yields a clean full soak (RESULT PASS, -Z included --
        random command numbers and unadvertised TD64 dialects are all
        rejected cleanly, so the classic V37 jump-table fear did not
        materialise on this ROM). Also: TD_GETGEOMETRY and
        NSCMD_DEVICEQUERY both fail with -1 (predate the driver), no
        ETD dialect.
action  skiptest unaligned-length
status  reported

quirk   removable-stop-ejects
match   kick>=0
because CMD_STOP / TD_EJECT may physically eject media or spin a drive
        down and not restart it, on any driver, whenever the medium is
        removable (DGF_REMOVABLE) -- a property of the medium, not of the
        driver identity, so it has no expressible match key here.
        `kick>=0` is always true; it exists only so this entry has a
        match block at all; the file's default policy (status is not
        "confirmed") means it is otherwise inert. Force it with
        -k removable-stop-ejects whenever running against a removable
        drive without -R, to keep CMD_STOP/TD_EJECT out of the mix.
action  skip CMD_STOP
action  skip TD_EJECT
status  reported

quirk   old-jumptable-random-cmd
match   version<37
because A driver built before V37 may index its command dispatch table
        directly off the command number without a range check, so an
        unknown command number in the random-cmd test (S8 test 21) can
        walk off the end of the table instead of falling through to
        IOERR_NOCMD. Applies to any driver below that version, regardless
        of name.
action  norandomcmd
status  suspected

quirk   uaehf-no-stop-start
match   name=uaehf.device
because The emulated uaehf.device (WinUAE/Amiberry hardfile device) may
        not implement CMD_STOP/CMD_START gating semantics the way real
        trackdisk-alike hardware does. Downgrade the stop-start test to a
        warning against this driver rather than failing the run over
        emulator-only behaviour.
action  warn stop-start
status  suspected

# --- entries added from this project's own testing -------------------------

quirk   lide-format-etd-nocmd
match   name=lide.device version<=40
because lide.device 40.12 advertises ETD_FORMAT in NSCMD_DEVICEQUERY but
        its BeginIO dispatch omits it (IOERR_NOCMD), and TD_FORMAT reads
        stale io_Actual as an offset high word (IOERR_BADADDRESS when the
        caller leaves garbage there). Reported upstream; reproduced with
        devsoak under Copperline emulation 2026-09-01.
action  expect ETD_FORMAT IOERR_NOCMD
status  reported

quirk   floppy-chip-buffers
match   name=trackdisk.device
because Floppy trackdisk.device DMA goes through Paula, which can only
        address chip RAM. devsoak's default MEMF_PUBLIC buffers may land
        in fast RAM, where a read completes with no data transferred
        (verify sees the untouched buffer). On Kickstart 2.0+ the
        driver's TD_GETGEOMETRY dg_BufMemType already advertises
        MEMF_CHIP and devsoak honours it automatically; Kickstart 1.3
        trackdisk predates TD_GETGEOMETRY, so force it there with
        -k floppy-chip-buffers. Applies to the internal drive and any
        external floppy on trackdisk.device.
action  chipbuffers
status  confirmed