Texas Instruments TMS320TCI648x manual Segmentation, Direct I/O RX Operation

Models: TMS320TCI648x

1 256
Download 256 pages 4.34 Kb
Page 42
Image 42

www.ti.com

SRIO Functional Description

Segmentation:

The LSU handles two types of segmentation of outbound requests. The first type is when the Byte_Count of Read/Write requests exceeds 256 bytes (up to 4K bytes). The second type is when Read/Write request RapidIO address is non-64-bit aligned. In both cases, the outgoing request is broken up into multiple RapidIO request packets. For example, assume that the CPU wants to perform a 1K-byte store operation to an external RapidIO device. After setting up the LSU registers, the CPU performs one write to the LSUn_REG5 register. The peripheral hardware then segments the store operation into four RapidIO write packets of 256 bytes each, and calculates the 64-bit-aligned RapidIO address, WRSIZE, and WDPTR as required for each packet. This example requires four outbound handles to be assigned and four DMA transmit requests. The LSU registers cannot be released until all posted request packets are passed to the TX FIFOs. Alternatively, for non-posted operations, such as CPU loads, all packet responses must be received before the LSU registers are released.

2.3.3.3Direct I/O RX Operation

Response packets are always type 13 RapidIO packets. All response packets with transaction types not equal to 0001b are routed to the LSU block sequentially in order of reception. These packets may have a payload, depending on the type of corresponding request packet that was originally sent. Due to the nature of RapidIO switch fabric systems, response packets can arrive in any order. The data payload, if any, and header data is moved from the RX FIFO to the shared RX buffer. The targetTID field of the packet is examined to determine which core and corresponding set of registers are waiting for the response. Remember, there can be only one outstanding request per core. Any payload data is moved from the shared RX buffer into memory through normal DMA bus operations.

Registers for all non-posted operations should only be held for a finite amount of time to avoid blocking resources when a request or response packet is somehow lost in the switch fabric. This time correlates to the 24-bit Port Response Time-out Control CSR value discussed in sections 5.10.1 and 6.1.2.4 of the RapidIO Physical Layer 1x/4x LP-Serial Specification. If the time expires, control/command register resources should be released, and an error is logged in the error-management RapidIO registers. The RapidIO Interconnect Specification states that the maximum time interval (all 1s) is between 3 and 6 seconds. A logical layer timeout occurs if the response packet is not received before a countdown timer (initialized to this CSR value) reaches zero.

Each outstanding packet response timer requires a 4-bit register. The register is loaded with the current timecode when the transaction is sent. The timecode comes from a 4 bit counter associated with the 24 bit down counter that continually counts down and is re-loaded with the value of SP_RT_CTL (Address offset 1124h) when it reaches 0. Each time the timecode changes, a 4-bit compare is done to the register. If the register becomes equal to the timecode again, without a response being seen, then the transaction has timed out. Essentially, instead of the 24-bit value representing the period of the response timer, the period is now defined as P = (2^24 x 16)/F. This means the countdown timer frequency needs to be 44.7 – 89.5Mhz for a 6 – 3 second response timeout. Because the needed timer frequency is derived from the DMA bus clock (which is device dependent), the hardware supports a programmable configuration register field to properly scale the clock frequency. This configuration register field is described in the Peripheral Setting Control register (Address offset 0020h).

If a response packet indicates ERROR status, the Load/Store module notifies the CPU by generating an error interrupt for the pending non-posted transaction. If the response has completed successfully, and the Interrupt Req bit is set in the control register, the module generates a CPU servicing interrupt to notify the CPU that the response is available. The control/command registers can be released as soon as the response packet is received by the logical layer. The hardware is not responsible for attempting a retransmission of the non-posted transaction.

If a Doorbell response packet indicates Retry status, the Load/Store module notifies the CPU by generating an interrupt. The control/command registers can be released as soon as the response packet is received by the logical layer. The hardware is not responsible for attempting retransmission of the Doorbell transaction.

42

Serial RapidIO (SRIO)

SPRUE13A –September 2006

Submit Documentation Feedback

Page 42
Image 42
Texas Instruments TMS320TCI648x manual Segmentation, Direct I/O RX Operation