SDONE, not HDONE.
In the data phase dma handler, mask off just the enable bits instead of
clearing the whole register. Clearing the direction bit could be bad.
Also don't stop a DMA until MREQPEND goes false. Doing this may cause
an ABORT on the PCI bus although I have yet to see this happen.
Add definitions for MREQPEND and the BRDCTL register. The BRDCTL register
is used to handle high byte termination and automatic termination testing.
Only enable reselections once the channel and SCSIRATE have been cleared.
Add a pause block around the test busy code in the non-tagged case to simplify
error recovery in the corner case of aborting an SCB that just got started.
Simplify reselection processing by removing the call to initialize_scsiid.
Clear the scsiseq re/select control bits and setup for catching bogus
busfrees earlier in the re/select process.
Improve the automatic PIO code. It turns out that SPIORDY is not a reliable
hardware condition bit, so use REQINIT intstead. Don't rely on PHASEMIS
either since it can take too long to come true. Use a brute force comparison
instead.
Remove some unnecessary overhead in the command complete processing. It
should be nearly impossible to overflow the QOUTFIFO (worst case 9 command
have to complete with at least 6 of them requiring paging on an aic7850),
so don't take the additional PIO hit to guard against this condition. If we
don't see our interrupt in time, the system has bigger problems elsewhere.
If this ever does happen, the timeout handler will notice and retry the
command.
This will make a number of things easier in the future, as well as (finally!)
avoiding the Id-smashing problem which has plagued developers for so long.
Boy, I'm glad we're not using sup anymore. This update would have been
insane otherwise.
to miss reselections from some devices and since the reselection response
timeout is only 200ns, enabling reselections too late may be the cause of our
problem.
Immediate SCBs, since they always send messages that tell the target to
transition to bus free now rely on the busfree interrupt instead of the
IMMEDDONE sequencer interrupt that was generated before.
Rearrange some code in the message out loop to give ATN a little more time
to drop before we ACK the last byte.
Use SPIORDY instead of REQINIT when snooping for a tag message on a reconnect.
This is done for the same reasons we use SPIORDY in the inb functions.
When going into BITBUCKET mode, turn off HDMAEN in the DFCNTRL register so
that we can "not care" what the value of HCNT is. If HCNT is 0, BITBUCKET
mode won't transfer any data if HDMAEN is set. Seeing as we don't want the
transfer to even think about touching the host, this seems more sane anyway.
Thanks to "Dan Willis" <dan@plutotech.com> for pointing out that this was
a problem.
SPIORDY should go active on any REQ of the bus, so testing for REQINIT is
not necessary. It also seems that testing for SPIORDY is more robust then
REQINIT since SPIORDY comes active after REQINIT and PHASEMIS seems to take
some time to come true after REQ is asserted if the phase has changed. Of
course, none of this is documented.
This should give the code savings of my original changes, without breaking the
driver on fast peripherals.
initial selection when entering the status phase. This is the same assertion
we use for all the other data transfer phases.
Hopefully fix the hangs in the mesgin and mesgout phases that I introduced
last week during some code cleanup. I need to get some of these 12MB/s
drives so I can reproduce these hangs here...
Add a pause disable in the SCB paging case around our manipulation of the
QOUTQCNT variable. This is simply extra sanity.
Set LASTPHASE to P_BUSFREE once we see a busfree so that the kernel driver can
differentiate this from a data out phase.
1) get_free_or_disc_scb was not being passed its argument correctly
in one case
2) Add protection in the form of the QOUTQCNT variable to prevent
overflowing the QOUTFIFO.
This should make SCB Paging work. Really, I mean it now. 8-)
used mvi instead of mov. Luckily this code is most likely never executed
since it is only there for sanity should a target goes into the data phase
twice during a single selection or reselection.
SCB paging is now handled almost entirely by the sequencer and also uses
DMA. This should make SCB paging at least an order of magnitude more
efficient and vastly simplifies the implementation.
Add a few space optimizations so this code still fits on aic7770 chips.
Update comments.
mode when this occurs and allow the target to complete the transaction.
Force a retry on overruns since they are usually caused by termination or
cable problems.
seeing SPIORDY and checking for PHASEMIS. My last change turned out to
be less cosmetic then I thought.
Pointed out by: Satoshi Asami <asami@cs.berkeley.edu>,
Faried Nawaz <fn@pain.csrv.uidaho.edu
Cosmetic change to p_mesgout code so that it "looks" the same as what
is done in the inb* routines.
NetBSD/OpenBSD support Submitted by:Noriyuki Soda <soda@sra.co.jp>,
Pete Bentley <pete@demon.net>,
Charles M. Hannum <mycroft@mit.edu>,
Theo de Raadt <deraadt@theos.com>
phasemiss to sneak by without detection. This should fix the
Wide/Narrow boot problems that have been reported since this bug
caused the driver ignore a narrow target rejecting wide negotiation.
Change #ifdef linux to #ifdef __linux__
aic7xxx_reg.h:
Remove unneeded BOFF_60BCLOCKS
define CHIPRSTACK to be the same as CHIPRST
define RESET_SCSI and CHANNEL_B_PRIMARY bits
All of these aer used during the setup of adapters.
aic7770 >= Rev E, aic7850, aic7860, aic7870, and ai7880 based controllers.
Make findSCB safer for non-tagged commands when tagged commands are
active on the controller. The symptoms of this problem were
"Overlapped commands attempted" messages during error recovery
attempts.
Compact scratch ram usage. This leaves 8 bytes free for future use.
Clean up some comments.
aic7xxx_reg.h:
Update my copyright.
Fix support for the aic7850 by looking only at the relavent bits of the
QINCNT. The 7850 puts random garbage in the high bits and all my attempts
to determine the cause of this failed. This approach does seem to work
around the problem.
Don't trust SCSIPERR to tell us when there is a parity error. On
some revs of the 7870 and the 7880, this bit follows the parity of
the current byte. Instead of using a SEQINT to tell the kernel,
re-enable the standard parity error interrupt since it seems to pause
the sequencer right at the time of the error which is the effect we were
looking for anyway.
aic7xxx_reg.h:
Remove PARITY_ERROR seqeuncer interrupt type, its no longer used.
Define QCOUNTMASK as the SRAM location for the mask to use on the
QINCNT register. QCOUNTMASK is determined by the number of SCBs
supported by the device we're working on.
aic7xxx_asm.c
Properly check the return value of fopen, and define the arg list
in getopt correctly.
Submitted by: Pete Bentley <pete@demon.net>
since setting up the DMA is too costly. Restructure for efficiency.
Pause the sequencer when a parity error occurs so that the kernel driver
knows during which phase the error was encountered.
SPIORDY just before we ack on the bus so that there is no chance to
see SPIORDY for the same byte twice.
Make some small modifications so that the Linux aic7xxx driver can use
our sequencer and register definition files verbatum.
Add the same type of safeguards we use in the mesg_in phase to the mesg_out
phase.
aic7xxx_reg.h:
Add definitions for the DSCommand register for PCI adapters.
1) Use cpp to preprocess the sequencer code.
2) Convert all "magic numbers" to #defines shared by the sequencer and
kernel driver via the aic7xxx_reg.h file. (The assembler still needs
to be re-written in lex/yacc to allow ~|& type constructions).
3) Raise ATN on parity errors for "in" phases and send an initiator detected
error or message-in parity error message as appropriate.
4) Turn off the reselection hardware from the time or a (re)connection to
busfree. It seems that some fast targets were able to reconnect before
the sequencer was able to see busfree.
5) The message buffer is considered "in-use" when there is a positive length
count. The ACTIVE_MSG flag was unnecesary.
6) Properly set SCB_NEXT_WAITING to SCB_LIST_HEAD in scbs being added to
the waiting scb list. This is a change in how the list code works to
facilitate some planned work in the reset code.
7) The fields in the SCB have be re-arranged to be quad-word aligned.
8) The inb code has been rewritten to catch phasemisses and be more efficient.
9) Go back to "snooping the bus" to determine if the incomming identify
message will be followed by a simple queue message. Its much faster than
doing a search through the SCBs.
10) Implement better tag range checking for incomming tags.
11) Make sdtr_to_rate more accurate (use 25 instead of 24 in calculations -
must have been asleep that night).
12) Rearrange some routines to reduce code complexity and size.
13) Update comments and formatting.
14) Fixed bugs I've forgotten about??
Reviewed by: David Greenman <davidg@FreeBSD.org>
in each phase routine. Saves a few instructions.
Be more careful in how we deal with SXFRCTL0. Or in the control bits of
interest instead of using mvi. The kernel driver will set the ULTRAEN
bit of SXFRCTL0 if we are using Ultra (20MHz) mode and we don't want to
clobber it.
In sdtr_to_rate divide by two if we are in ultra mode to get the correct
setting since its a 20MHz instead of 10MHz scale.
to replace the very poor, original implementation of Scatter/Gather operations.
Use a bit (that was freed up with the rewrite above) in the SCB control byte
to designate commands that should allow disconnection. The kernel driver
makes this decision now instead of the sequencer since the sequencer can't
do the indexing very efficiently.
This commit drops the sequencer from 426 instructions to 390 most likely
freeing enough space to do a target mode implementation.
The first could occur because the original code would continue to reset
the SCSIID register while waiting for a selection. This could potentially
conflict with a reconnect since a successfull reconnect will also set the
SCSIID register. The fix is to use a separate wait loop after starting
a selection (as was done a few revisions ago).
The second probably never happens, but it was possible for a target to
reconnect while there was a pending SCB on the waiting list and not get
noticed. The fix was to remove a supurflous check of the scb waiting
list.
is needed for 3940 support.
Have tagged commands look to see if a target is "busy" with a non tagged
command before executing. This prevents overlapped tagged and non tagged
commands which can happen since request sense commands are not tagged.
1) If a target initiated a sync negotiation with us and happened to chose a
value above 15, the old code inadvertantly truncated it with an "& 0x0f".
If the periferal picked something really bad like 0x32, you'd end up with
an offset of 2 which would hang the drive since it didn't expect to ever
get something so low. We now do a MIN(maxoffset, given_offset).
2) In the case of Wide cards, we were turning on sync transfers after a
sucessfull wide negotiation. Now we leave the offset alone in the per
target scratch space (which implies asyncronous transfers since we initialize
it that way) until a syncronous negotation occurs.
3) We were advertizing a max offset of 15 instead of 8 for wide devices.
4) If the upper level SCSI code sent down a "SCSI_RESET", it would hang the
system because we would end up sending a null command to the sequencer. Now
we handle SCSI_RESET correctly by having the sequencer interrupt us when it
is about to fill the message buffer so that we can fill it in ourselves.
The sequencer will also "simulate" a command complete for these "message only"
SCBs so that the kernel driver can finish up properly. The cdplay utility
will send a "SCSI_REST" to the cdplayer if you use the reset command.
5) The code that handles SCSIINTs was broken in that if more than one type
of error was true at once, we'd do outbs without the card being paused.
The else clause after the busfree case was also an accident waiting to
happen. I've now turned this into an if, else if, else type of thing, since
in most cases when we handle one type of error, it should be okay to ignore
the rest (ie if we have a SELTO, who cares if there was a parity error on
the transaction?), but the section should really be rewritten after 2.0.5.
This fix was the least obtrusive way to patch the problem.
6) Only tag either SDTR or WDTR negotiation on an SCB. The real problem is
that I don't account for the case when an SCB that is tagged to do a particular
type of negotiation completes or SELTOs (selection timeout) without the
negotiation taking place, so the accounting of sdtrpending and wdtrpending
gets screwed up. In the wide case, if we tag it to do both wdtr and sdtr,
it only performs wdtr (since wdtr must occur first and we spread out the
negotiation over two commands) so we always have sdtrpending set for that
target and we never do a real SDTR. I fill properly fix the accounting
after 2.0.5 goes out the door, but this works (as confirmed by Dan) on
wide targets.
Other stuff that is also included:
1) Don't do a bzero when recycling SCBs. The only thing that must explicitly
be set to zero is the scb control byte which is done in ahc_get_scb. We also
need to set the SG_list_pointer and SG_list_count to 0 for commands that do
not transfer data.
2) Mask the interrupt type printout for the aic7870 case. The bit we were
using to determine interrupt type is only valid for the aic7770.
Submitted by: Justin Gibbs
the adapter's selections. Many fast periferals were getting upset when
the sequencer decided to rearbitrate after the device had already won
arbitration. This also forced the creation of a list threaded through
the SCBs (since we don't have enough space anywhere else) of commands that
are awaiting reselection. This list is run down before any new transactions
from the input queue are allowed. The list is appened to whenever we begin
a selection (simple case since the selecting device is always at the head)
and by the kernel driver whenever a request sense occurs. In the common
case, the list is only one element long, but when a reselection wins out
over a selection and that reselection generates a request sense, the
outstanding selection required for the retreval of the sense code grows
the list. On machines with many targets, this might cause the list to grow
large, so this solution, which will allow up to the maximum number of I/O
requests capible of the card elements in the list, was chosen. The list
manipulation is trivial and adds three sequencer instructions of overhead
to the selection phase.
This fixes the "target busy" errors from micropolis drives and the bursty
I/O problem when performing I/O between a Quantum Grand Prix and any other
device. I anticipate that this will correct many of the problems that
have been reported with this driver.
Reviewed by: Wcarchive and David Greenman
is identical to the older version, just the copyright has changed. Many
thanks go to Dean Gehnert of the Linux camp who went the extra mile to make
this happen.
Other changes:
Update assembler man page to include the -v and -D options
Merge in Dean's latest changes to the assembler
Have the sequencer do a MSG_REJECT when the negotiated syncronous rate
is lower than the adapter supports. This forces asyncronous mode which
is faster at these rates anyway.
This code will be moved shortly to the non-gpld portion of the tree.
old value.
Remove unnecessary check for active messages in setup SCB. This same test
would also jump to p_mesgin_done which would "ACK" an extra time possibly
confusing the target.
Tell the kernel driver whenever we send an ABORT_TAG message.
- Report valid residual byte counts. We actually pause the sequencer
when the residual is non-zero. I thought about using DMA to do this,
bus sequencer program space is tight.
- Fix embarassing off by one error in the computation of a 2's
compliment variable. This was most likely the cause of the
many problems reported with the tagged queuing code.
- Handle "MAX_SYNC" as a special case (ie we are the ones starting
the sync negotiation sequence). This was done so that the target
scratch area can be initialed to 0 offset (asyncronous transfers)
safely. The initialization to 0 (was 15) is necessary since in
some cases a Wide negotiation could run into problems if SCSIRATE
was set wrong and we went into data(in/out).
- Trim the DMA routines a little by using some procedures. Net
effect is more functionality with 3 less instructions after this
update.
- Toggle the WIDEODD bit of the DFCNTRL whenever this is not the
last SG block. It has no effect in the 8bit bus configuration,
but in the Wide configuration ensures that the overlap byte is
held in the SCSI block if the transfer is odd so it will end
up in the next SG (the correct behavior).
Print out the length of the compiled sequencer program.
aic7xxx.seq:
More optimizations. Replace generic bcopy routine with bcopy_3
and bcopy_4 (ie unroll the loops) since these are the only two cases used.
Initialize SIMODE1 and SXFRCTL1 from the kernel in ahc_init instead of
at each selection/reselection since this is expensive and only needs to
be done once. Condense function returns into previous instruction if possible.
Reorder some sections to kill superflous jumps. These optimizations kill
the ~150k/s penalty adding support for Twin/Wide cards was costing since
the last place in the commaon path of execution where we had to do ugly,
convoluted testing for the type of card in the sequencer has gone away.
Next stop tagged queuing and target mode.
