ultra5/cmd646 hang

[Doug White]

In my continuing quest to get -current working on the ultra5 here, I've
been able to get it to break into the debugger (yay DEBUGGER_ON_POWERFAIL).

Initial details:

FreeBSD 5.1-CURRENT #2: Fri Nov 14 10:51:48 PST 2003
    dwhite at dwsparc.looksmart.com:/usr/src/sys/sparc64/compile/SPARC
Preloaded elf kernel "/boot/kernel/kernel" at 0xc0330000.
Timecounter "tick" frequency 270000000 Hz quality 0
real memory  = 134217728 (128 MB)
avail memory = 104366080 (99 MB)
cpu0: Sun Microsystems UltraSparc-IIi Processor (270.00 MHz CPU)
atapci0: <CMD 646 WDMA2 controller> port 0xc00020-0xc0002f,0xc00018-0xc0001b,0xc00010-0xc00017,0xc00008-0xc0000b,0xc00000-0xc00007 at device 3.0 on pci1

Note this system is running source as of 11/1 to eliminate the interrupt
code changes as any source of the problem.

The problem:

Timecounters tick every 10.000 msec
ad0: WARNING - SETFEATURES recovered from missing interrupt
ad0: WARNING - SETFEATURES recovered from missing interrupt
ad0: WARNING - SET_MULTI recovered from missing interrupt
ad0: WARNING - SETFEATURES recovered from missing interrupt
GEOM: create disk ad0 dp=0xfffff8001075eec0
ad0: 38182MB <MAXTOR 4K040H2> [77578/16/63] at ata2-master WDMA2
ad0: WARNING - READ_DMA recovered from missing interrupt
acd0: WARNING - MODE_SENSE_BIG recovered from missing interrupt
ad0: WARNING - READ_DMA recovered from missing interrupt
ata3: resetting devices ..
### we are hanging here ###

Analysis:

Compiling with WITNESS & INVARIANTS causes none of these messages, the
system locks solid just after the "timecounter" message.  I'm thinking
there is a timing issue a la the log message for ata-lowlevel.c rev 1.21.

Setting the loader tunable hw.ata.ata_dma=0 has no effect on the problem
(although the disk comes up at PIO4 as expected).

Using the power switch and checking "show intr", it appears the ATA chip
is getting stuck with the interrupt asserted:

db> show intr
fast pil13      5
ithrd pil2      3016775
pcib0 vec2021   2
sab0 vec2027    5
atapci0 vec2016 3016775
tick pil14      5067

Next steps:

ata-chipset.c has a special interrupt override for 648 and newer chipsets,
into the SiI bits. I'll try changing the interrupt handler registration to
use SIIINTR.

Hints appreciated :-)

-- 
Doug White                    |  FreeBSD: The Power to Serve
dwhite at gumbysoft.com          |  www.FreeBSD.org