• No timeouts occur during DMA arbitration.
9.4.2.2 DMA Arbitration (400/500/600/700/800/900 Models)
Priority is determined by means of a binary tree of priority selectors. Based on
its own state, each selector (p#) determines which of its two inputs has priority.
Whenever ACK is generated, each selector in the priority path sets its next state
to give priority to the opposite input request.
Request 7 is given highest priority (IOCTL ASIC): although it has a low
bandwidth, it requires low latency because of LANCE chip constraints. Thus
if all devices request DMA transactions at once, devices 0-5 would each get 12.5%
and device 6 (SCSI) would get 25% of the available arbitration slots, after device
7 (IOCTL) has been serviced.
Figure 10 illustrates the arbitration scheme on 400/500/600/700/800/900 models.
In the illustration, rq0-5 stand for requests from options 0-5 and rq6 stands for
requests from the SCSI.
Figure 10 400/500/600/700/800/900 Models: DMA Arbitration Scheme
RQ0
RQ1
RQ2
RQ3
RQ4
RQ5
RQ6
P3
P4
P5
P1
P2
P0
MR−0103−93RAGS
9.4.3 I/O Timeout
The timeout period for I/O transactions on the system varies according to the
model:
• On the 300 models, the timeout period is 175 80-ns TURBOchannel cycles,
or 14 s. If a device takes longer than 14 s to return data on an I/;O read
operation or to signal its readiness to perform an I/O write operation, the
system logs a tcTimeout error and signals a hard error interrupt to the CPU
on IRQ4.
• On the 400/500/600/700/800/900 models, the timeout period is 256
TURBOchannel cycles, or 10.24 s. If a device takes longer than 10.24
s to return data on an I/O read operation or to signal its readiness to
perform an I/O write operation, the system logs a Timeout error and signals
an Uncorrectible Error interrupt to the CPU.
9–6 I/O Programming
Commentaires sur ces manuels