QP Context Defines QP's Operational Characteristics
Refer to Figure 3-1 on this page. Before a QP can be used to send and/or receive messages with a QP of the same type in another CA, software creates the QP and supplies it with the basic operational characteristics that it will use to send and receive messages with its companion QP in the remote CA. In this example, the QP Context of each of the two companion QPs have been programmed (when the QP was setup by software) with the information described in the subsections that follow. The QP Context items described in the following subsections assume that the QP type is Reliable Connected (RC).
Figure 3-1. QP Context
Local Port Number
The local CA port number through which the QP will send and receive message packets is programmed into the QP's context during QP setup.
QP Type
The type of QP must be specified when the QP is created. There are five types:
SQ Start PSN (Packet Sequence Number)
When the SQ Logic starts transmitting request packets, this is the PSN that will be inserted in the first packet generated. This start number is stored in the QP Context. The SQ's current PSN (cPSN; see items L and P) is the PSN that will be inserted in the current request packet to be transmitted. Initially this is set equal to the start PSN and is then updated as each request packet is issued by the SQ Logic.
RQ Logic's Expected PSN (ePSN)
See items K and Q. When using the RC protocol, upon receiving a request packet from the SQ Logic in the remote CA, the RQ Logic in the target QP is required to validate that the packet's PSN is the next expected PSN.
If not, then one or more request packets may have been lost somewhere along the flight path (e.g., a switch or a router may have dropped the packet due to a CRC error) and a PSN Sequence Error Nak is returned.
If the packet's PSN is correct (i.e., it equals the RQ Logic's ePSN), then the RQ Logic returns an Ack (Acknowledge) back to the sender's SQ Logic.
It should be obvious that the RQ Logic in each QP must be initialized with the start PSN that will be used by the SQ Logic in the other QP.
Maximum Data Payload Size
Refer to Figure 3-2 on this page. As a packet traverses the link(s) between the source and destination CA ports, it is sent from a transmit buffer in one port to the receive buffer in the port on the other end of that link. Although the maximum allowable size of a packet's data payload field is 4KB, one or more of the transmit and/or receive buffers along the path between the source and destination ports may not be able to handle packets with 4KB data payloads. The buffers implemented in a port may be limited to a maximum packet data payload size of:
256 bytes.
512 bytes.
1024 bytes.
2048 bytes.
4096 bytes.
Figure 3-2. Example Path Maximum Transfer Unit Size
During configuration, the configuration software (i.e., the SM) discovers whether this limitation exists. When a pair of QPs are subsequently created in the two CAs, the QP Context of each is initialized with the maximum permissible data payload size that can be used in packets. This is referred to as the Path Maximum Transfer Unit, or PMTU. Additional information can be found in “Packet Length Field (PktLen)” on page 614.
Destination Local ID (DLID) Address
This is the destination Local ID address of the port on the other CA behind which the remote QP resides. In the illustration, the QP Context of the QP in HCA “X” (item M) is initialized with the LID address assigned to the port on HCA “Y” behind which its companion QP resides. Likewise, the QP Context of the QP in HCA “Y” (item N) is initialized with the LID address assigned to the port on HCA “X” behind which its companion QP resides. The DLID address stored in a QP's Context is inserted in each request packet generated by the QP's SQ Logic.
Desired Local Quality of Service
In order for a request packet to get from the QP SQ that generates it to the target QP's RQ Logic, it may have to traverse a series of links interconnected by switches. Some applications may require that messages be transferred through the network as quickly as possible in order to achieve performance adequate to the task at hand. In other words, some applications require a high QoS (Quality of Service).
On the other hand, another application may not require that the network expedite transmission of its messages through the fabric to their destination. Said another way, the application doesn't require a high QoS.
When a QP is initially set up by software, the programmer indicates the desired QoS by specifying a desired Service Level (SL). This 4-bit value is placed in each packet generated by the QP. As will be seen later, the packet's SL value determines how quickly the packet will begin transmission from the source port. Likewise, if the packet arrives at a switch, the switch looks at the packet's SL value to determine how quickly the packet must be forwarded out through one of the switch's egress ports.
Packet Injection Delay
Packet injection delay is also referred to as the Inter-Packet Delay, or IPD, and as the Maximum Static Rate. Refer to Figure 3-3 on page 44. Some links implement only one transmit/receive signal pair, while others may implement four or twelve signal pairs. It should be obvious that a wider link can receive data significantly faster than a thinner link. The path between the source and destination CA ports may encompass a number of links and multiple switches. The links are not necessarily the same width. To prevent faster links from overrunning slower links with traffic, a QP is supplied with an IPD that defines the interval that must be observed between sending packets to the destination QP.
Figure 3-3. Packet Injection Delay Example
Ack Receipt Timeout (Local Ack Timeout)
The Local Ack Timeout value assigned to a QP when it is set up defines the amount of time the QP's SQ Logic will wait for an Ack packet before it retries the transmission of the corresponding request packet.
Ack Timeout and Missing Packet(s) Retry Counter
The official name of this item is the Retry Count and it defines the number of times the QP's SQ Logic will retry the transmission of a request packet due to any of the following:
a timeout while awaiting an Ack packet,
the receipt of an RDMA Read response packet with a PSN higher than the expected PSN. This indicates that one or more RDMA read response packets were lost in the fabric.
the receipt of a response packet with a PSN higher than that of an expected Atomic response packet. This indicates that an Atomic response packet was lost in the fabric.
the receipt of a Nak packet from the remote QP's RQ indicating that it detected a missing packet (i.e., a request packet was received with a PSN > the ePSN).
Receiver Not Ready (RNR) Retry Count
This value is supplied by the remote QP's RQ Logic when the two QPs are first set up. It defines how many times this QP's SQ Logic should retry the transmission of a request packet that receives an RNR Nak from the remote QP's RQ Logic. The remote QP's RQ Logic will respond to a request packet with an RNR Nak if it is temporarily unable to handle that request. A classic example would be if a request packet is received and there are currently no WQEs posted to the receiver's RQ to handle the inbound request.
Source Port's LID Address
At a minimum, the configuration software (i.e., the SM) will assign a single Local ID (LID) address to a port. This is referred to as its base LID address. The configuration software may optionally assign a range of LID addresses to a port by assigning its base LID address and also telling it how many LIDs starting at the base LID are assigned to it.
When a QP's SQ Logic sends a request packet to its assigned port for transmission, it must indicate to the port which of the port's assigned LID addresses to insert in the packet's Source LID (SLID) field. This item is supplied in the form of an offset from the source port's base LID address. It is programmed into the QP Context of a RC, UC, or UD QP when the QP is set up and is referred to as the Source Path Bits.
Global Source/Destination Addresses
Introduction
If the destination CA is not in the same subnet as the source CA, then the packets generated by the QP will have to traverse one or more routers to get to the destination CA. In this case, the packet must contain a Global Router Header (GRH) that contains the 128-bit Source Global ID (SGID) of the source CA port and the 128-bit Destination Global ID (DGID) of the destination CA port.
DGID Address Identification
When a QP is set up, software provides its QP Context with the DGID assigned to the destination CA port behind which the remote QP resides. This DGID address is inserted in the DGID field in each request packet generated by the QP's SQ Logic.
SGID Address Identification
Background
Each port has at least one 64-bit GUID address assigned to it by the device manufacturer (see “Port's Default GUID Is Hardwired” on page 156). This GUID address resides in entry 0 of the port's GUIDInfo attribute. The GUIDInfo attribute is a table with one or more entries. The SM can use entries 1 through n of the table to assign additional GUIDS to the port.
SGID Selection
The upper 64 bits of the SGID is provided from the attribute that contains the port's 64-bit subnet ID (the GIDPrefix attribute). When a RC, UC, or UD QP is set up, software indicates which of the GUIDs assigned to the local port must be inserted into the SLID field of each packet generated by the QP. This information is supplied to the QP Context in the form of an index into the port's GUIDInfo attribute.
Additional Global Address Information
In the case of a global destination, software also supplies the QP Context with the following items (which are inserted into their respective fields in each request packet's GRH):
Traffic Class (TClass). This value indicates the QoS desired as the request packets cross multiple subnets to get to the destination CA port.
Flow Label. If non-zero, this instructs all routers along the path to ensure that all packets with the same Flow Label value are delivered in the correct order to the destination CA port.
Hop Limit. Each router along the path decrements the Hop Limit and, if it is exhausted, the packet is discarded (because it has lost its way and is wandering throughout the global fabric).