Arria V Avalon-ST Interface for PCIe Solutions: User Guide

ID 683660
Date 5/12/2017
Public
Document Table of Contents

9.3.1. Using Relaxed Ordering

Transactions from unrelated threads are unlikely to have data dependencies. Consequently, you may be able to use relaxed ordering to improve system performance. The drawback is that only some transactions can be optimized for performance. Complete the following steps to decide whether to enable relaxed ordering in your design:

  1. Create a system diagram showing all PCI Express and legacy devices.
  2. Analyze the relationships between the components in your design to identify the following hazards:
    1. Race conditions: A race condition exists if a read to a location can occur before a previous write to that location completes. The following figure shows a data producer and data consumer on opposite sides of a PCI-to-PCI bridge. The producer writes data to the memory through a PCI-to-PCI bridge. The consumer must read a flag to confirm the producer has written the new data into the memory before reading the data. However, because the PCI-to-PCI bridge includes a write buffer, the flag may indicate that it is safe to read data while the actual data remains in the PCI-to-PCI bridge posted write buffer.
      Figure 42. Design Including Legacy PCI Buses Requiring Strong Ordering
    2. A shared memory architecture where more than one thread accesses the same locations in memory.

    If either of these conditions exists, relaxed ordering leads to incorrect results.

  3. If your analysis determines that relaxed ordering does not lead to possible race conditions or read or write hazards, you can enable relaxed ordering by setting the RO bit in the TLP header.
  4. The following figure shows two PCIe Endpoints and Legacy Endpoint connected to a switch. The three PCIe Endpoints are not likely to have data dependencies. Consequently, it would be safe to set the relaxed ordering bit for devices connected to the switch. In this system, if relax ordering is not enabled, a memory read to the legacy Endpoint is blocked. The legacy Endpoint read is blocked because an earlier posted write cannot be completed as the write buffer is full. .
    Figure 43. PCI Express Design Using Relaxed Ordering
  5. If your analysis indicates that you can enable relaxed ordering, simulate your system with and without relaxed ordering enabled. Compare the results and performance.
  6. If relaxed ordering improves performance without introducing errors, you can enable it in your system.