rather than SEEKCOMPLETE before retrying the operation. If implied seeks
are being used, the state is set to DOIO (no change). This is why I
couldn't reproduce the disk_unbusy() panic on my SS2; it uses implied
seeks. Patch from John F. Woods <jfw@jfwhome.funhouse.com>
[prevents disk_unbusy panic when disk is loaded (if no
free IOPBs, xdstrategy() would queue the buffer for pickup
by xdcintr() but xdcintr() would never call disk_busy().
xdc_startbuf() is a better place since all bufs are routed
through here] problem detected by girish@dworkin.wustl.edu,
diagnosed and corrected by me.
- move disk_unbusy() call in xdc_remove_iorq() before the call to
XDC_FREE() [don't want to access a data structure that was just put
on a free list]
- Better disklabel handling. While a disklabel isn't used
in the driver, some versions of the OpenPROM insist on
one being present in order to boot from floppy. These
changes provide a default label (in a way similar to how
the SCSI disk driver provides a default) so that a user
can more easily place the label on the disk.
- Fix semi-bug in bootpath handling. It appears as if the
bootpath can appear in a couple of formats: "/fd@0,0", which
is what bootpath_fake() creates on v0 proms and may be
passed by some v2 proms, and "/fd0" which is what the
v2 prom on my SS2 passes. We now handle both formats.
- Use a mountroot hook to eject the floppy and wait for
the user to insert a filesystem floppy if we're the boot/root
device.
zero since xycmatch no longer sets it). [if you call mapiodev()
with a zero size it will reuse the KVA it returns, swiping the device
out from under you!]
- make xyc->iopbase point in the kernels DVMA space rather than at the
normal malloc'd KVA. this isn't compatable with sun4m [doesn't
have a kernel DVMA space] and will need to be changed later (XXX).
- move disk_attach() to before reading the disk label as per Jason.
otherwise we are reading into an unallocated buffer (oops!)
zero since xdcmatch no longer sets it). [if you call mapiodev()
with a zero size it will reuse the KVA it returns, swiping the device
out from under you!]
- make xdc->iopbase point in the kernels DVMA space rather than at the
normal malloc'd KVA. this isn't compatable with sun4m [doesn't
have a kernel DVMA space] and will need to be changed later (XXX).
- move disk_attach() to before reading the disk label as per Jason.
otherwise we are reading into an unallocated buffer (oops!)
device address (bus_tmp, bus_map, and mapiodev now do this for us).
also, we handle all our mappings, so don't have obio.c do any for us
(i.e. don't set ra->ra_len in xycmatch).
device address (bus_tmp, bus_map, and mapiodev now do this for us).
also, we handle all our mappings, so don't have obio.c do any for us
(i.e. don't set ra->ra_len in xdcmatch).
and thus probed the wrong address on the 4/300 (but it worked because
it was probing the esp0 registers!).
now that bus_tmp() adds the offset in for us [as of obio 1.19] i
discoved that a byte access to the dma registers is not allowed.
so, i've change probeget to use a word access.
page boundaries:
- change bus_tmp() to include the offset from the start of page in the
returned KVA [rather than forcing each driver to add it back in
individually]
- changed bus_map() to include the offset from the start of page in the
the returned value if a mapping is found in the PROM's KVA area
- clarified a few comments
- New metrics handling. Metrics are now kept in the new
`struct disk'. Busy time is now stored as a timeval, and
transfer count in bytes.
- Storage for disklabels is now dynamically allocated, so that
the size of the disk structure is not machine-dependent.
- Several new functions for attaching and detaching disks, and
handling metrics calculation.
Old-style instrumentation is still supported in drivers that did it before.
However, old-style instrumentation is being deprecated, and will go away
once the userland utilities are updated for the new framework.
For usage and architectural details, see the forthcoming disk(9) manual
page.
utilizing David Jones' new MI NCR 5380 code. Ported from the sun3 ncr_si.c
and "sw" DMA code written by me.
This driver contains user-configurable "options", which can be set via the
"flags" directive in the kernel configuration file. By default, only
DMA is enabled. DMA completion interrupts and reselection may be enabled
by setting the appropriate bits with "flags". See si.c for details.
Note that DMA completion interrupts and reselection don't yet work on the
4/100 controller. I don't know why, and it's unlikely that I'll have
the opportunity to find out any time soon. DMA does work, and results
in a considerable performance increase.
DMA, DMA completion interrupts, and reslection all work on my 4/260 (VME)
system with modern SCSI-II disks.
Remove dvma_malloc/dvma_free; drivers should allocate kernel memory and
use dvma_mapin/dvma_mapout to double map it in DVMA space.
Make the resource map `dvmamap' responsible for all DVMA allocation.
The VM map `physmap' only serves the role of placeholder in the VM system.
of FIFO overflows on high baud rates.
However, doing so on all 4 ports would cost a whopping 64KB (at 4096 entries
per FIFO) of kernel memory. So, the FIFOs are now allocated at attach time
allowing the size for the keyboard and mouse ports to be reduced (to 128)
which should be adequate for the 1200 baud they use.
- properly do MSG_IN handshaking, so we can actually receive multi-byte msgs.
- do synch negotiation (now that the above works).
- handle disconnects.
There are a few trial-and-error bits at points where the docs I have are
particularly ambiguous about the state of chip and/or SCSI bus.
Things to do:
- more cleanup
- deal with MSG_OUT phase better
- keep some "config reg 3" bits per target (ie. FASTCLK and FASTSCSI).
Since uname currently does not allow to discriminate different sparc
models, we use `sysctl.hw.cpumodel' to do this somewhat heuristicly:
if the returned strings starts with "SUN4-" a sun4 architecture is
assumed. This information is used to determine whether or not to strip
the a.out header off the 1st-stage bootblocks.
transfer ends in stead of doing it before starting a new transfer.
Use a different CSR bit to wait for "chip not busy" (may be necessary to
make this conditional on chip revision).
- make esp_poll() approximate the given timeout value.
- introduce esp_abort(), and use it for timed out commands; make targets and driver less confused.
- make {free,ready,nexus}-list management somewhat more coherent.
- make sure we only proceed down the state machine in espintr()
if there really is an interrupt pending.
(should be DT_BLOCK). otherwise the bootxx boot block will fail if
you are on a system with a v2 openprom [because it will not do the
v2_seek in obp_strategy()]. fixes the problem recently noted on
the port-sparc list by matthew green <mrg@eterna.com.au>. i tested
fix on a ss2 with a v2 prom.
- don't let pcyl go unassigned if we have a non-sunos disklabel.
if nothing else, guess pcyl = ncyl + acyl.
- don't use bogus disklabel data if disk label is invalid
(xdgetdisklabel() now returns if label was ok or not).
- move xddriver decl. before start of actual code (cleanup).
have been clearing just B_BUSY (s/B_BUSY/~B_BUSY/). this
fixes hard hang when reading disklabel when booting system.
minor clean up: added a few prototypes and only fool with bootpath
if booting the system (in xdattach).
- v2 openprom disk boot generates "/sbus/esp/sd@3,0" which we translated
into "/sbus@0,0/esp@0,0/sd@3,0"... this caused esp.c to reject
the bootpath because it was looking for sbus slot/offset 0,0.
changed the code to generate "/sbus0/esp0/sd@3,0" to fix this
panic.
- when generating fake boot paths, generate them with unit numbers
rather than slot/offsets (thus, /obio@0,0 becomes /obio0, as
does vmel and vmes).
- fixed up comments to reflect these changes
tested new code on: 4/280 (diskless), 4/360 (diskless), 4/360 ("xd" disk),
4/370 (scsi disk), ss1 w/v1 openprom (diskless),
ss1 w/v1 openprom (scsi disk), ss2 w/v2 openprom (diskless),
ss2 w/v2 openprom (scsi disk).
to match `sd' devices against the PROM's boot path.
Detect DMA "overshoots" when handling odd-sized transfer sizes (e.g., xfers
from/to raw partitions), by using the "Transfer Pad" command when the
transfer count reaches zero.
pk