62.6.3.9 Priority Queueing in the DMA

The DMA by default uses a single transmit and receive queue. This means the list of transmit/receive buffer descriptors point to data buffers associated with a single transmit/receive data stream. The GMAC can select up to 6 priority queues. Each queue has an independent list of buffer descriptors pointing to separate data streams.

The table below gives the memory size associated with each queue:

In the transmit direction, higher priority queues are always serviced before lower priority queues, with Q0 as lowest priority and Q5 as highest priority. This strict priority scheme requires the user to ensure that high priority traffic is constrained so that lower priority traffic will have required bandwidth. The GMAC DMA will determine the next queue to service by initiating a sequence of buffer descriptor reads interrogating the ownership bits of each. The buffer descriptor corresponding to the highest priority queue is read first. As an example, if the ownership bit of this descriptor is set, then the DMA will progress to reading the 2nd highest priority queue’s descriptor. If that ownership bit read of this lower priority queue is set, then the DMA will read the 3rd highest priority queue’s descriptor. If all the descriptors return an ownership bit set, then a resource error has occurred, an interrupt is generated and transmission is automatically halted. Transmission can only be restarted by setting the START bit in the Network Control register. The GMAC DMA will need to identify the highest available queue to transmit from when the START bit in the Network Control register is written to and the TX is in a halted state, or when the last word of any packet has been fetched from system memory.

The GMAC transmit DMA maximizes the effectiveness of priority queuing by ensuring that high priority traffic be transmitted as early as possible after being fetched from the system bus. High priority traffic fetched from the system bus is pushed to the MAC layer, depending on traffic shaping being enabled and the associated credit value for that queue, before any lower priority traffic that may pre-exist in the transmit SRAM-based packet buffer. This is achieved by separating the transmit GMAC local memory packet buffer into regions, one region per queue. The size of each region determines the amount of memory space allocated per queue.

For each queue, there is an associated Transmit Buffer Queue Base Address register. For the lowest priority queue (or the only queue when only one queue is selected), the Transmit Buffer Queue Base Address is located at address 0x1C. For all other queues, the Transmit Buffer Queue Base Address registers are located at sequential addresses starting at address 0x440.

In the receive direction each packet is written to system memory data buffers in the order that it is received. For each queue, there is an independent set of receive buffers for each queue. There is therefore a separate Receive Buffer Queue Base Address register for each queue. For the lowest priority queue (or the only queue when only one queue is selected), the Receive Buffer Queue Base Address is located at address 0x18. For all other queues, the Receive Buffer Queue Base Address registers are located at sequential addresses starting at address 0x480. Every received packet will pass through a programmable screening algorithm which will allocate a particular queue to that frame. The user interface to the screeners is through two types of programmable registers:

• Screening Type 1 registers—The module features 4 Screening Type 1 registers (GMAC_ST1RPQ). Screening Type 1 registers hold values to match against specific IP and UDP fields of the received frames. The fields matched against are DS (Differentiated Services field of IPv4 frames), TC (Traffic class field of IPv6 frames) and/or the UDP destination port.
• Screening Type 2 registers—The module features 8 Screening Type 2 registers (GMAC_ST2RPQ). Screening Type 2 registers operate independently of Screening Type 1 registers and offer additional match capabilities. Screening Type 2 allows a screen to be configured that is the combination of all or any of the following comparisons:

1. An enable bit VLAN priority, VLANE. A VLAN priority match will be performed if the VLAN priority enable is set. The extracted priority field in the VLAN header is compared against VLANP in the GMAC_ST2RPQ register itself.
2. An enable bit EtherType, ETHE. The EtherType field I2ETH in GMAC_ST2RPQ maps to one of 4 EtherType match registers, GMAC_ST2ER. The extracted EtherType is compared against GMAC_ST2ER designated by this EtherType field.
3. An enable bit Compare A, COMPAE. This bit is associated with a Screening Type 2 Compare Word 0/1 register x, GMAC_ST2CW0R/1R.
4. An enable bit Compare B, COMPBE. This bit is associated with a Screening Type 2 Compare Word 0/1 register x, GMAC_ST2CW0R/1R.
5. An enable bit Compare C, COMPCE. This bit is associated with a Screening Type 2 Compare Word 0/1 register x, GMAC_ST2CW0R/1R.

Each screener type has an enable bit, a match pattern and a queue number. If a received frame matches on an enabled Screening register, then the frame will be tagged with the queue value in the associated Screening register, and forwarded onto the DMA and subsequently into the external memory associated with that queue. If two screeners are matched, then the one which resides at the lowest register address will take priority so care must be taken on the selection of the screener location.

When the priority queuing feature is enabled, the number of interrupt outputs from the GMAC core is increased to match the number of supported queues. The number of Interrupt Status registers is increased by the same number. Only DMA related events are reported using the individual interrupt outputs, as the GMAC can relate these events to specific queues. All other events generated within the GMAC are reported in the interrupt associated with the lowest priority queue. For the lowest priority queue (or the only queue when only 1 queue is selected), the Interrupt Status register is located at address 0x24. For all other queues, the Interrupt Status register is located at sequential addresses starting at address 0x400.

The additional screening done by the functions Compare A, B, and C each have an enable bit and compare register field. COMPA, COMPB and COMPC in GMAC_ST2RPQ are pointers to a configured offset (OFFSVAL), value (COMPVAL), and mask (MASKVAL). If enabled, the compare is true if the data at the offset into the frame, ANDed with MASKVAL, is equal to the value of COMPVAL ANDed with MASKVAL. A 16-bit word comparison is done. The byte at the offset number of bytes from the index start is compared to bits 7:0 of the configured COMPVAL and MASKVAL. The byte at the offset number of bytes + 1 from the index start is compared to bits 15:8 of the configured COMPVAL and MASKVAL.

The offset value in bytes, OFFSVAL, ranges from 0 to 127 bytes from either the start of the frame, the byte after the EtherType field, the byte after the IP header (IPv4 or IPv6) or the byte after the TCP/UDP header. Note the logic to decode the IP header or the TCP/UDP header is reused from the TCP/UDP/IP checksum offload logic and therefore has the same restrictions on use (the main limitation is that IP fragmentation is not supported). Refer to the Checksum Offload for IP, TCP and UDP section of this documentation for further details.

Compare A, B, and C use a common set of 24 GMAC_ST2CW0R/1R registers, thus all COMPA, COMPB and COMPC fields in the registers GMAC_ST2RPQ point to a single pool of 24 GMAC_ST2CW0R/1R registers.

Note that Compare A, B and C together allow matching against an arbitrary 48 bits of data and so can be used to match against a MAC address.

All enabled comparisons are ANDed together to form the overall type 2 screening match.