the lazy-transmit-interrupt logic, fixing a few minor logic problems.
Now unable to reproduce the lockup problem described in PR #6767. Changing
PR's state to "feedback".
Soren S. Jorvang <soren@t.dk>, with a lower match level.
Display the other accented characters which are not present in the IBM
character set by their non-accented equivalents, with an even lower
match level.
* Macroize hardware and software descriptor access (much easier to read).
* Simplify and optimize the transmit loop a bit, and use IFF_OACTIVE as
appropriate.
* Fix a potential race condition in the transmit loop. This change has
made the driver significantly more stable (almost completely eliminated
the "device timeout" errors that have plagued this driver).
* Implement transmit interrupt pacing.
* Add missing bus_dmamap_sync() calls (on transmit and receive descriptors
and fraglists). (Draining the write buffer when accessing these structures
may have also contributed to the increased stability of this driver on
the Alpha.)
like the SMC83C100 EPIC/100 driver:
* Rather than using pointers to the head and tail of the transmit and
receive rings, use wrapping indexes into arrays. This is a little more
obvious when reading the code.
* More cleanly separate the hardware descriptor from the software descriptor.
* bus_dma it everywhere.
* Implement interrupt pacing and avoid a potential race in the transmit
loop.
Now this looks more or less like the Rhine driver I was working on when
this driver was committed :-) Update copyright notice to reflect that.
Use fast memory timing NVRAM parameter. Clean up and fix establishment
of default target parameters. Don't use NVRAM if are flagged as not to
do so (I had a busted NVRAM setup which I couldn't edit that enabled SYNC
mode but disabled disconnect/reconnect and wide!!). Fix delays after
resets. BUS resets not done in isp_init anymore- relegated to OS
specific outer layers. Fix a buglet where you can get in a loop for
a NULL xs in the completion list in isp_intr. Add in some defines that
can disable fast posting. Add in code for Loop Up/Loop Down events that
call into the outer layers as to what to do.
define. Add a new config flag param (ISP_CFG_NONVRAM) whose intent it is
to cause NVRAM to be ignored. Add ISPASYNC_LOOP_DOWN and ISPASYNC_LOOP_UP
isp_async enums.
(we get LOOP DOWN events, and we'll hang on that at this time).
Add other isp_async cases- ISPASYNC_LOOP_DOWN and ISPASYNC_LOOP_UP. DOWN
will cause internal queuing until UP, whereupon a timeout will fire up
any pending xfers. It doesn't really keep commands from getting destroyed
by loop down events, but at least minimizes the damage. This was much
easier to implement with CAM.
command (unlock for sd and cd) if no other partitions are open, return
EBUSY otherwise. DIOCEJECT will have the old semantic if its argument is not
0. The old ioctl has been renamed to ODIOCEJECT for binary compatibility.
the disk is ready to transfer data, and in ATA-5 the DSC has been obsoleted.
So only wait for DRQ to transfer data. This can be made conditional on the
ATA version if it's proven to break with some drives (worked with all the
drives I have access to).
While I'm there correct a few typos.
contiguous chunk of outq to emulation routine. Fix based on patch
from M. Drochner, modified to call output routine once with each
contiguous chunk when the ring buffer wraps around.
improvement in the "cat /usr/share/misc/termcap" benchmark on 486/75
ttwrite queues data for us in OBUFSIZ chunks. If we only consume
OBUFSIZ-1 bytes per call to wsdisplaystart(), we're *guaranteed* to
fall behind when given maximum-sized reads (which will defer
processing the leftovers until ttrstrt gets called, after a timeout)
Note that OBUFSIZ*4 may be overkill; OBUFSIZ*3 may be sufficient to
get maximal performance; OBUFSIZ*2 resulted in only about a 2.5x
performance improvement, but it's 2am and I've done enough tweaking
for the night.
on the 8 port card Simon Gerraty has. In general, cards which have
this lots of ports also have a separate interrupt status register, but
this change is just to talk to the various ports independently. It works,
but it's not optimal. (XXX still need a good name for the card in the
comments, and to update the manual page.)
has the same 4-byte alignment requirement that the transmit side does. This
causes the packet payload to be misaligned. So, on systems which require
strict alignment, we must copy the incoming frame to a new packet buffer,
suitably aligned.
