CVS log for floppy-darwin/MkLinux_floppy/portdef.h

Some minor bug fixes.  Formatting now works for 720k MFM floppies as
well as the 1440k that I got working last night.  In addition, the
format code now properly checks to make sure the end of the track
didn't overwrite the first sector's address mark.  This should
dramatically reduce the misformats.

Note that not finding sector 0/1 can also be a sign of a bad disk.
It is believed that if a disk can be formatted, this code should
be able to do the job, but there are no guarantees.

Two major formatting issues still remain:

* 800k(GCR) support
* Setting DMAByteBoundryAlignment back to 8

The latter probably won't hurt anything, but I didn't want to make
that change until I could test it.  I may test tomorrow before
releasing an updated package.

Note that the updated package (0.95) will not include formatting
tools, since this support is so experimental, and requires a great
deal of work on the command line.  Currently, the format tools are
a total hack.  Before I make formatting support available in binary
form, the tools will have to, at the very least, not suck.


dgatwood@mklinux.org

This update documents a number of timing details that are essential
to the correct operation of the SWIM3 controller.  Many if not most
of these details are relevant to other SWIM and IWM controllers on
Macintosh hardware.  Developers, take note.  :-)

Fixed a bug where the interrupt handler routines could read the status
of the wrong SWIM chip.  I'm not sure if this could ever actually occur,
though, since the chip should only be sending an interrupt when asked
to do something, and the two driver instances are wrapped by a mutex
to keep them from ever doing an operation at the same time.  This is
one of those "theoretical" bugs.

Fixed a bug where if you write to a block in a given track without
reading from the track first, then read from a different block in
that track, you could get back the data from a different track,
since writing doesn't necessarily cause the track to be read.  This
is a deferred read on write bug introduced late in the life of the
Copland driver, or possibly after it moved to Mac OS X Server, since
the comments in the code still talked about reading the track
before writing until recently.  :-)

Fixed a race condition in signal handler as triggered by a hardware
race condition between the DMA controller's interrupt and the floppy
controller's interrupt.  The DMA interrupt is now ignored.

Found actual timing information and determined that the drive reports
itself ready after seeking and after power-up before it is actually
ready, much as we already suspected.  Corrected the amount of delay
in the workaround.

Fixed a timing bug where the driver was out of spec, since the drive
requires a period of rest after writing before a head switch or seek
is permitted.

Fixed a bug where the wrong variable was checked for return status
of a function when determining the reason for failure of a read or write.

Forced the driver to report 720k and 800k as (potentially) valid formats
for disks formatted single-sided, since there's no easy way to tell
if the back side can be correctly formatted without trying it.  (Okay,
there is.  You could switch heads and wait for an index mark.  I'll
fix this for real eventually, time permitting.)

Improved detection of unformatted/bad disks by test-reading block 0
immediately on insertion as part of the media test.  There might be a
better way, but my first attempt caused the I/O Kit to wedge.  If
that turns out to no longer be an issue, I'll take out this hack
later.

Fixed bug in dbdma support code (in two places) where data was being sent
in weird sizes for large DMA transactions.  This was caused by some bad
math.

Fixed an issue in the old (unused) dbdma support code that was
inherited by MkLinux along with the floppy driver (but only used for
a few weeks until I could write a better one).  Anyway, I wanted to
make it a little better just in case someone was brave enough to try
to use it....  :-)  The issue was that it could set dbdma interrupts
too often (in theory) for really long requests.  Not that it matters,
since we ignore DMA interrupts anyway, but why waste the PCI
bandwidth....  Also fixed the DMA block size bug in the old code.

Fixed numerous missing cast warnings and incorrect long vs. int errors in
printf and IOLog statements.

Fixed a number of missing casts in other places to get rid of some warnings.

Fixed bug where polling for a disk could be very slow (potentially
causing mount delays) in rare cases.

Fixed an issue where excessive use of delay() for multi-millisecond delays
was causing some mouse jerkiness.  Replaced all delays over a millisecond
with FloppyTimedSleep.

Cleaned up PrintRegs() to print certain registers on the SWIM III.
Turned it off by default to avoid slowing things down unnecessarily.

Finished format support.

Fixed a bug in format support where you couldn't change from GCR to MFM
encoding on a disk and vice-versa.

Fixed a bug in format support where you couldn't reformat a single-sided
disk as double-sided.  If I ever add support for index mark detection
on the back side to determine double-sidedness, we will want to add
the ability to format double-sided floppies as single-sided.  We
do not want to do that currently, though, since a disk must be
formatted as double-sided in order to show up as double-sided currently
(meaning that the disk would end up being formatted double-sided,
but the driver would only know about one side, which would be bad. :-)

Fixed a huge delay on disk insertion caused by doing a RecalDrive from
track 40.  This only occurred for the second and subsequent disks
inserted.  Fixed by believing the SWIM 3's current track register
and seeking from the current head position to track zero, and -then-
doing a RecalDrive.  :-)

Added a delay after recalibrate drive to match the one used after
seeking, with a little extra just to be safe (since a Recal only
occurs after an error condition or when formatting, both of which
are more delicate states than a normal read or write).

Made the gap length be an unsigned int instead of an unsigned char.
Another driver's code suggests that it isn't unreasonable to allow
for a longer postamble after you've written a track, although from
what I can see, they were just being cautious/overly pessimistic.

Changed cachedTrack to a signed char, since it can go from -1
through 80.

Changed the GCR track format to match that used in Mac OS X
Server 1.0.  This still needs to be tested and verified.

Removed all the retry code when waiting for an interrupt, since
if an interrupt is missed, it won't ever happen anyway.  Added
diagnostic printf calls to indicate when a wait for an interrupt
times out.

Removed the hack where the GCR track setup code was basically
return'ed out.

Removed unnecessary waits.  Fixed waits in mode set code.  Turns
out the pulse between phase changes must be at -least- 1 msec, not
exactly 1 msec.  And it should be less than 2 msec.  When I had
it "exactly" 1 msec.  according to the SWIM3 chip's internal
timer, the chip started consistently failing to notice the state
changes.  Oops.

Fixed a theoretical timing bug in format switch code.  SWIM needs
a delay when switching between GCR and MFM.  I'm not sure if this is
still true for SWIM3.

Fixed a timing bug where (in theory) the drive power could be
turned back on immediately after it is turned off.  The spec
requires a rest period.

Fixed a timing bug in RecalDrive which could potentially cause
stepping beyond track 0 (thrashing) on faster machines.  The
track 0 sensor lies to us for a few milliseconds after track
zero is reached.  The amount of time differs according to the
drive in question.  The smaller timing values used in newer
drives are used, but the older drives' timing information is
also recorded in the comments.

Fixed several sub-millisecond timing bugs that are not worth noting
in detail.

Updated copyright dates at top of files, added MkLinux copyright
and license notices to a bunch of files that were missing it.


dgatwood@mklinux.org

Fix for the random stack corruption panic.  The panic occurred when
doAsyncReadWrite requests were handled synchronously.  Changing back
to spawning a thread made the panic go away.  Dunno.  :-)


dgatwood@mklinux.org

Fixed broken build style changes.  I didn't notice this before because
pbxbuild appears to ignore them....


dgatwood@mklinux.org

Oops.  Turning debugging back off.


dgatwood@mklinux.org

Some timing fixes for GCR disks (that may not actually be necessary --
my test failed because the SWIM3 chip was somehow wedged, and I had
to power cycle it.  In any case, we should probably be doing this
anyway as best I can gather.


dgatwood@mklinux.org

Major hack.  I've replaced the printf statements en masse in one particular
file with a delay(100) followed by the printfs.  The driver works much
better now.  Apparently, some printf statement happens to be in a spot
where a delay is needed, but is missing.  I'm not sure where.  In any
case, it works now, and that's what matters, right?  ;-)


dgatwood@mklinux.org

Turned debugging back on.

Apparently, either I wrapped a bit too much code with a "#if (MACH_DEBUG)"
directive (unlikely, since I don't think I added any this round) or there
is a depressingly timing-sensitive bug in this code.  Either way, the
driver works when all the debug printfs are turned on and fails when they
aren't.  Miserably.  I'll try to debug this more soon.

Until then, enjoy the driver, and sorry for all the noise in your syslog.
In the words of God (and indeed, the final words of God to man), "We
apologize for the Inconvenience."  (Obscure HGG reference.)


dgatwood@mklinux.org

re-disabled most of the debugging output.  Apparently missed a little bit.
Will try back later.


dgatwood@mklinux.org

Turns out the write support issues were a hardware problem.  While
really noisy, this code reads and writes.  Cleanup pending.


dgatwood@mklinux.org

Addded a small delay that looks like it really isn't necessary while
trying to track down the bogus write protect issues.   Also, fixed
the write protect issue.  It was caused by differing opinions of the
correct size for type Boolean between the C and C++ portions of the
code.  And they say size doesn't matter....


dgatwood@mklinux.org

Cleanup of debug messages.


dgatwood@mklinux.org

Hideously messy stuff.  This was "working" before the latest changes --
it couldn't quite eject, though, because the event was occurring before
the wait.  Moved to semaphores to get preposting behaviour.  Praying.


dgatwood@mklinux.org

Working on trying to keep it from panicking when the driver unloads, but
still not there.  For now, you still have to reboot each test, which is
a pain in the....  Anyway, added code to try to help figure out why the
I/O Kit never creates an IOMedia nub.  Also added code to allow callback
into the c++ parts of the code for media change notification, should
that become necessary.  Also, a subtle one-line fix for detection of
disk ejection.  Also, changed the device to be non-removable (this is
-not- the same as ejectable, AFAICT, but refers to whether the drive
itself can be removed) since we don't support removal of the floppy
drive on PowerBooks in the current rev.

The current state is that the floppy core code detects the media, determines
correctly whether it is GCR or MFM encoded, and basically plays with the
drive a bit.  It does not currently show up in /dev, so it is not possible
to test eject or other functionality.  Read/write is still not implemented.

Best guess: another three or four hours of debugging, and it should be
working.  :-)


dgatwood@mklinux.org

Fixed a NULL pointer dereference in probe code.  Fixed a second NULL pointer dereference caused by temporarily removed code that I'd never reenabled.  Reenabled everything that had been temporarily enabled except for the pieces that have not yet been done (correct sleep/delay behaviour for long sleeps, and add support for read, write, and format).  Changed a bunch of NOTYETDAG ifdefs to IOKITREPLACED or PDMSUPPORT, as appropriate.  For basic support, the only requirements left are the NOTYETDAG items.  Also fixed the arguments so that it can spawn a media scan thread.

dgatwood@mklinux.org

Major work on this little puppy.  It took until just now to get it
buildable again.  The fd.c file still needs to be made buildable.

Remaining work:

1.  Make fd.c buildable
2.  Write interface routines to look like block_io() and friends
3.  Write read/write routines
4.  Write interrupt handler pass-thrus to match read/write routines
5.  Write some trivial glue code for dbdma support
6.  Anything else I haven't thought of....


dgatwood@mklinux.org

Initial revision

My pet project.  Working on porting MkLinux's floppy driver to Mac OS X
to support the beige G3 machines, the Wallstreet PowerBook, and any older
machines as appropriate.  This is the first import.  It builds, but does
not actually do anything yet.


dgatwood@mklinux.org