TMS320TCI6487_1 Datasheet, PDF (67/90 Pages) Texas Instruments – Digital Signal Processor Silicon Revisions 1.3, 1.2, 1.1, 1.0

www.ti.com
Silicon Revision 1.0 Known Design Exceptions to Functional Specifications
Advisory 1.0.1
SRIO: Packet Forwarding Cannot Be Used With NREAD Response Packets Greater
Than 16 Bytes
Revision(s) Affected:
Details:
Workaround: 1:
Workaround: 2:
1.0; Fixed in revision 1.1
Packet forwarding uses programmable look-up tables to direct incoming packets to an
outbound port when the packets do not belong to the local device. Packet forwarding is
carried out at the logical layer of the serial RapidIO (SRIO) without the interaction of the
CPU. The SRIO logical layer copies incoming packets from an inbound buffer to an
outbound buffer. When used for packet forwarding, it forwards all types of packets,
including response, maintenance, DOORBELL, and message packets.
The current SRIO design fails to correctly copy response packets with a payload greater
than 16 bytes from the inbound to the outbound buffer. The first 16 bytes are correctly
copied, but the remainder of the payload is discarded. This issue affects only NREAD
response packets since their payloads can be up to 256 bytes. Packet types with small
responses, such as NWRITE_R, maintenance, message, and DOORBELL packets are
not affected by this issue.
Use a data push model, where each device in the daisy chain only submits write
requests. Using this approach will avoid the issue and provide the lowest latency
solution.
Two options exist if NREAD response packets cannot be avoided; for example, when
reading core dump information from an unresponsive processor which is unable to
initiate traffic by itself. The first option is to use software to segment read requests into
16-byte NREADs. Note that this option will work functionally, but may take too much
time.
The second option is illustrated in Figure 9.
Network/
System
Host
Port 0
Port 1 Port 0
Port 1 Port 0
DSP 1
DSP 2
DSP 3
Figure 9. Daisy-Chain Example
In this example, assume that DSP3 is down and the system host wants to do a large
NREAD of DSP3 to examine the core dump. The issue discussed above prohibits the
NREAD from completing correctly because, as the response packets from DSP3 are
sent back, they are corrupted by DSP2 and DSP1 packet forwarding. Instead, the
system host needs to request that the adjacent DSP (DSP2) generates the NREAD
request to DSP3. The NREAD responses are sent to DSP2 and temporarily stored in
memory. Then, DSP2 can generate NWRITE/NWRITE_R/SWRITE packets to the
system host with the needed payload. These packets are correctly forwarded by DSP1
to the system host since they are request packets and not responses.
SPRZ248D – September 2007 – Revised August 2008
Submit Documentation Feedback
TMS320TCI6487/8 DSP
67
Silicon Revisions 1.3, 1.2, 1.1, 1.0