This is a completely rewritten scsipi_xfer execution engine, and the
associated changes to HBA drivers. Overview of changes & features:
- All xfers are queued in the mid-layer, rather than doing so in an
ad-hoc fashion in individual adapter drivers.
- Adapter/channel resource management in the mid-layer, avoids even trying
to start running an xfer if the adapter/channel doesn't have the resources.
- Better communication between the mid-layer and the adapters.
- Asynchronous event notification mechanism from adapter to mid-layer and
peripherals.
- Better peripheral queue management: freeze/thaw, sorted requeueing during
recovery, etc.
- Clean separation of peripherals, adapters, and adapter channels (no more
scsipi_link).
- Kernel thread for each scsipi_channel makes error recovery much easier
(no more dealing with interrupt context when recovering from an error).
- Mid-layer support for tagged queueing: commands can have the tag type
set explicitly, tag IDs are allocated in the mid-layer (thus eliminating
the need to use buggy tag ID allocation schemes in many adapter drivers).
- support for QUEUE FULL and CHECK CONDITION status in mid-layer; the command
will be requeued, or a REQUEST SENSE will be sent as appropriate.
Just before the merge syssrc has been tagged with thorpej_scsipi_beforemerge
and the buffer is not 8-byte aligned, it leaves a partial transfer in the
SDRx registers. This could cause data corruption in dump(8) because the
read-ahead buffers are allocated on a non-aligned address. When the read
DMA transfer is interrupted and the current DMA position is not aligned,
call the dma_end() routine to flush the SDRx register contents to the
buffer.
timeout()/untimeout() API:
- Clients supply callout handle storage, thus eliminating problems of
resource allocation.
- Insertion and removal of callouts is constant time, important as
this facility is used quite a lot in the kernel.
The old timeout()/untimeout() API has been removed from the kernel.
- Nuke tc_badaddr() check from asc_tc because the existency check is already
done for real TC devices.
- Nuke tc_badaddr() check from asc_ioasic because it is a builtin device
found in all of IOASIC DECstations.
starts running can occur on both DMA in and DMA out. I missed the little comment
in the Mach driver for the DMA out case. Also, only ignore the interrupt if the
TC is non-zero (to match the Mach driver).
is written. This significantly reduces, if not eliminates, the problem reported in
PR#5510. It stops the problem on my 5000/200, and two other people have verified that
it fixes the problem on their 5000/200s. I'm waiting for confirmation from the PR
originator before closing the PR.
struct scsipi_adapter; they were not used.
Add a scsipi_ioctl entry point to struct scsipi_adapter. This will be
used to issue ioctl commands to the host adapters.
Inspired by PR #6090, from Matt Jacob.
problem was. A collision between a select and reselect would leave TC
non-zero from the command-out DMA count, which could later be considered
a fatal condition, causing a reboot. The message for that error was
only displayed with DEBUG. Fixed by clearing TC on a reselect. The
non-zero TC detection won't occur in this case, so unconditionally
display the message if it occurs.
Workaround for another problem that resulted from an "Illegal Command"
status from the 53c94 which would get ignored and result in a timeout
(which also reboots the system). Added the missing check for the
illegal command status, and add the workaround of resending the "accept
message" command to the 53c94. Correct fix will be to determine why the
message wasn't sent in the first place. Abort if the resending the
command doesn't work.
Correctly detect a spurious interrupt and ignore it. This was taken
from a newer Mach driver, but did not get the check converted for the
design difference between the current NetBSD driver and the Mach driver.
does a "restore data pointers" when reselected after disconnecting in
the middle of a DMA transfer). The driver needs a different way to know
which script to continue the DMA transfer. The message-in for the "restore
data pointers" loses the original "resume" script, and the driver would
attempt to continue the DMA transfer at the beginning of the current DMA
chunk, rather than at the point the disconnect occured. The result was a
spurious console message, and a trashed filesystem.
(currently only CD-ROM drives on i386). The sys/dev/scsipi system provides 2
busses to which devices can attach (scsibus and atapibus). This needed to
change some include files and structure names in the low level scsi drivers.
re-enabled with ASC_IOASIC_BOUNCE.
All DMA buffered processing is now done in the bus-specific DMA routines
in asc_ioasic.c and asc_tc.c.
Disable several informational messages dealing with non-empty FIFO conditions,
but allow enabling with ASC_DIAGNOSTIC for troubleshooting.
after synch negotation during device probes, the 53C94 stops with the target
reqesting more command data. The fifo appears to still contain a couple of
bytes, so transfer them. If the target still wants command data after that,
do a transfer pad. The target seems to function normally after this. A
much better solution thant simply rebooting!
Also fix a problem when compiling with DIAGNOSTIC - a mismatch between using
DIAGNOSTIC and MACH_DIAGNOSTIC caused the compile to fail.
Try to start another target and a device disconnects.
* Don't print DMA_OUT message in the `normal' case of a 16-byte residual.
* ioasic-connected 53c94 devices are always clocked at 25MHz even on
machines with 12.5MHz TurboChannel.
* When we are transferring in DATA (in asc_dma_in) and the target
is an async device, there is sometimes an extra byte in the FIFO.
If so, we need to drain that byte out of the fifo, but if and only
if the target is async. See also the comments in asc_dma_in()
in the related Mach mk84 asc driver (scsi_53C94_hdw.c), which
has an identical fix but applied in more restrictive conditions
than we need, with async *disk* targets, as well as async tapes.
* Add a watchdog and timeout active SCSI requests, to eliminate any
potential for deadlock due to applying the fix above on newer
silicon versions of the 53c94 which may not have the above problem.
Should use the MI scsi per-target timeout instead, when available.
cfattach code for TC SCSI option cards and ioasic 53c94 baseboard SCSI.
ascvar.h: shared softc declarations
asc_ioasic.c: ioasic front-end code.
asc_tc.c: Turbochannel option (and 5000/200 basebard) front-end code.
* ioasic_attach meeds more work to eliminate pmax_type dependency
and to verify the clocks speed passed to 53c94.
* Add prototypes for asc script entry points; should compile with
-Wstrict-prototypes -Wmissing-prototypes.
* Use tcvar.h interface. The usage of tc_syncbus() and tc_mb() may
not be quite stylistically on an Alhpa, but it apparently makes no
difference on the eerly-generation Alpha CPUs in TC Alphas.
multi-channel driver), or to SCSI_CHANNEL_ONLY_ONE if a
single-channel driver.
(2) use scsiprint() rather than a locally-defined autoconfig print
function, and kill any locally-defined print function.
a char *, because that's what was really intended, and because
if the print function modifies the string, various things could become
unhappy (so the string should _not_ be modified).
a boot string for firmware that can do this, such as the SPARC and
the sun3 models. It is currently silently ignored on all other
hardware now, however. The MD function "boot()" has been changed to
also take a char *.
Enable reselections as soon as possible after a disconnect - prevents
losing a reselecting device.
Check for and ignore a spurious interrupt during a DMA input (from the
Mach driver).
Turbochannel as well as the IOASIC. It should now work on the 5000/200.
Removed the "aborts" which could leave the disk trashed when the abort
rebooted the system. Fix the data corruption problem by clearing the
FIFO before starting a data transfer.
* Change MachEmptyWriteBuffer() to wbflush(). Should use TC mi names tc_mb(),
tc_wmb, tc_syncbus() but I'm not sure which each wbflush() should be.
* Add prototyped forward decl for asc_Dumplog() and add an explicit void
return type.
* Remove unused variables.
Rename the ioctl asic register and slot macros from ASIC_<xxx> to
IOASIC_<xxx>, to be compatible with the machine-indpendent names in
sys/dev/tc/ioasicvar.h. The pmax code still uses
sys/arch/pmax/pmax/asic.h, as some of the registers and offsets
defined there are not yet defined in sys/dev/tc/ioasicvar.h.
Rename the ioctl asic base-address pointer from `asic_base' to `ioasic_base'.
Use the device address in the attach_args structure, instead of
using the deprecated BUS_CVTADDR macro.
Change the Mach derived asc driver to use "SCSI_PHASE_xxx" instead of
"ASC_PHASE_xxx", as the latest version of the Mach driver does.
"struct pmax_device" to avoid conflict with <sys/device.h>.
Change the signature of interrupt-handlers to take a void *
(a pointer to the softc) and return an int (indicating spurious
interrupts or other conditions.)
Concomitant changes to code that prints driver/unit name: use dv_xname
and dv_unit, instead of doing pointer arithmetic on elements of the static
softc array.
Remove support for old config. The old-config "driver" structure
is still present, because the pmax non-MI SCSI driver needs it.
Merge some off Per Fogelstrom's changes for the Pica driver,
which uses the machine-independent SCSI code. This is #ifdef'ed
out until the DMA is fixed to work on Decstations, too.
bouyer