How the SM Sends a Message and Handles a Response
The Normal Method Can Be Used
To Send a Message
The SM may cause the SMI of a port on the local device to send a message in the following manner:
The SM builds a 256-byte request MAD in the device's local memory.
The SM calls the Create Address Handle verb and supplies it with the following input parameters:
- HCA Handle.
- PD that the Address Handle is to be a member of.
- DLID of the destination port.
- Maximum Static Rate (i.e., the IPD; see “Static Rate Control” on page 589 for more information).
- Source Path Bits to select which of the source port's assigned LID addresses will be inserted in the SMP's SLID field.
The SM passes a WR to the port's SMI SQ Logic by executing the Post Send Request verb. The WR includes:
- The Gather Buffer List specifying the 256-byte buffer in local memory that contains the 256-byte request MAD.
- The Address Handle specifying the destination port address.
Note that the Address Handle and the QP to which the WR is posted must belong to the same PD.
The SQ Logic executes the SQ WQE. In doing so, it:
- Uses the specified Gather Buffer List to read the MAD from local memory.
- Uses the Address Handle specified in the WR to set up the request SMP's DLID.
The SMP is accepted into the local port Link Layer VL15 transmit buffer and is then transmitted.
To Receive a Message
To handle an inbound SMP (either a response or a Trap SMP), the following steps are taken:
1. | The SM must post a WR to the SMI's RQ (using the Post Receive Request verb). The WR specifies the Scatter Buffer List defining where the incoming 256-byte MAD is to be written in local memory. |
2. | When the inbound SMP arrives at the port, it is accepted into the Link Layer's VL15 receive buffer. |
3. | It is then passed to the RQ Logic of the port's SMI. |
4. | The RQ Logic uses the top RQ
WQE to determine where to write the packet's data payload (i.e., the MAD) in local memory. |
5. | The RQ Logic then retires the top RQ
WQE and creates a CQE. |
6. | The creation of the RQ
CQE causes the HCA to inform the SM (typically via an interrupt) that a MAD is in memory and ready to be processed. |
A Device-Specific Method Can Be Used
The specification contains the following statement:
“Queue pairs 0 and 1 have unique semantics with respect to processing of messages specifying one of them as the destination queue pair. Implementations of QP 0 and 1 are not required to follow the semantics associated with other queue pairs with respect to requirements such as posting and consumption of WQEs, manipulation of an associated completion queue, and so on.”