MFC of "Large set of CAM improvements" breaks I/O to Adaptec 29160 SCSI controller

[Andy Farkas]

Hi, firstly:

RELENG_8 csup'd with date=2010.02.14.00.00 works perfectly for days.

RELENG_8 csup'd with date=2010.02.15.00.00 dead-locks the disk I/O
subsystem. Network still operational but anything needing disk hangs.
Power-cycle required.

kernel config is GENERIC with KDB, DDB and BREAK_TO_DEBUGGER options added.

hardware:
ahc0: <Adaptec 29160 Ultra160 SCSI adapter> port 0x4000-0x40ff mem
0xefa00000-0xefa00fff irq 16 at device 0.0 on pci10
ahc0: [ITHREAD]
aic7892: Ultra160 Wide Channel A, SCSI Id=7, 32/253 SCBs

da0: <SEAGATE ST3146707LW 0005> Fixed Direct Access SCSI-3 device
da1: <SEAGATE ST3146707LW 0005> Fixed Direct Access SCSI-3 device


The dead-lock can happen at any time, but I can provoke it by running
a bonnie++ disk test. It happens doing rm -rf /usr/obj/usr and it has
happened doing a make installworld. It can survive a make buildworld
(the system runs normally until it decides to dead-lock).

The box (HP ProLiant ML 110) has 2 scsi disks and 4 sata disks. The
2010.02.15 kernel will run perfectly for days on the SATA disks. *Only*
when the scsi disks are accessed will the system dead-lock. Note that
the SATA disks do not work either if the system has dead-locked.

I can provide more details and a vmcore.0 if anyone is interested.

-andyf