For a quick fix, append this line to /etc/system and reboot:
set scsi_options & ~0x80
This turns off Tagged Command Queuing, a SCSI feature that is improperly implemented in many older drives and can cause problems between Suns and some RAID implementations (for a special note on RAID, see the end of this question)
NOTE: this will seriously degrade performance on disks that do properly support tagged command queuing. Setting the SCSI options per broken target is therefor the preferred solution.
In Solaris 2.4 and later you can set those options per SCSI bus. See isp(7) and esp(7).
For some disks, all you need to do is decrease the maximum number of queued commands:
set sd:sd_max_throttle=10
In later Solaris releases you can specify scsi_options per (broken) target or per SCSI bus. See esp(7d), isp(7d), from which this example /kernel/drv/esp.conf file is derived:
name="esp" parent="/iommu@f,e0000000/sbus@f,e0001000/espdma@f,400000"
reg=0xf,0x800000,0x40
target1-scsi-options=0x58
scsi-options=0x178;
Certain hardware RAIDs support a number of different LUNs (logical disks) but share a common set of I/O buffers between them. This can cause SCSI QFULL conditions on devices without any commands queued. Since the algorithm is to retry the command when a previous command is completed, Solaris doesn't handle this situation very well.
The workaround is to decrease sd_max_throttle such that there's always at least 1 slot available for each LUN. E.g., if you have 3 LUNs and your RAID supports upto 64 outstanding commands, sd_max_throttle must be at most 31. (Any two LUNs can get 31 requests and you still have two slots over for number 3)
Decreasing sd_max_throttle was seen to improve performance due to better load balancing among LUNs on some hardware RAIDs.