Intel® L- and H-tile Avalon® Memory-mapped+ IP for PCI Express* User Guide

ID 683527
Date 10/19/2021
Public

A newer version of this document is available. Customers should click here to go to the newest version.

Document Table of Contents

7.1.1.1.2. Deadlock Risk and Avoidance

Under certain circumstances, it is possible for the DMA engine in the example design hardware to get into a deadlock. This section describes the conditions that may lead to a deadlock, and how to avoid them.

When you program the DMA Controller to use the Read Data Mover to fetch too many descriptors for the Read Data Mover descriptor queue, the following loop of back-pressure that leads to a deadlock can occur.

Once the Read Data Mover has transferred enough descriptors through the DMA Controller to its own descriptor queue to fill up the queue, it deasserts its ready output. The DMA Controller in turn asserts its waitrequest output, thus preventing the Read Data Mover from writing any remaining descriptor to its own queue. After this situation occurs, the Read Data Mover continues to issue MRd read requests, but because the completions can no longer be written to the DMA Controller, the tags associated with these MRd TLPs are not released. The Read Data Mover eventually runs out of tags and stops, having gotten into a deadlock situation.

To avoid this deadlock situation, you can limit the number of descriptors that are fetched at a time. Doing so ensures that the Read Data Mover's descriptor queue never fills up when it is trying to write to its own descriptor queue.