FEAST Fast Ethernet Controller for PCMCIA and Generic 16-Bit Applications
MEMORY PARTITIONING
Unlike other controllers, the LAN91C110 does not require a fixed memory partitioning between transmit and receive
resources. The MMU allocates and de-allocates memory upon different events. An additional mechanism allows the
CPU to prevent the receive process from starving the transmit memory allocation.
Memory is always requested by the side that needs to write into it, that is: the CPU for transmit or the MAC for
receive. The CPU can control the number of bytes it requests for transmit but it cannot determine the number of
bytes the receive process is going to demand. Furthermore, the receive process requests will be dependent on
network traffic, in particular on the arrival of broadcast and multicast packets that might not be for the node, and that
are not subject to upper layer software flow control.
In order to prevent unwanted traffic from using too much memory, the CPU can program a "memory reserved for
transmit" parameter. If the free memory falls below the "memory reserved for transmit" value, MMU requests from
the MAC block will fail and the packets will overrun and be ignored. Whenever enough memory is released, packets
can be received again. If the reserved value is too large, the node might lose data which is an abnormal condition. If
the value is kept at zero, memory allocation is handled on first-come first-served basis for the entire memory capacity.
Note that with the memory management built into the LAN91C110, the CPU can dynamically program this parameter.
For instance, when the driver does not need to enqueue transmissions, it can allow more memory to be allocated for
receive (by reducing the value of the reserved memory). Whenever the driver needs to burst transmissions it can
reduce the receive memory allocation. The driver program the parameter as a function of the following variables:
1. Free memory (read only register)
2. Memory size (read only register)
The reserved memory value can be changed on the fly. If the MEMORY RESERVED FOR TX value is increased
above the FREE MEMORY, receive packets in progress are still received, but no new packets are accepted until the
FREE MEMORY increases above the MEMORY RESERVED value.
INTERRUPT GENERATION
The interrupt strategy for the transmit and receive processes is such that it does not represent the bottleneck in the
transmit and receive queue management between the software driver and the controller. For that purpose there is no
register reading necessary before the next element in the queue (namely transmit or receive packet) can be handled
by the controller. The transmit and receive results are placed in memory.
The receive interrupt will be generated when the receive queue (FIFO of packets) is not empty and receive interrupts
are enabled. This allows the interrupt service routine to process many receive packets without exiting, or one at a
time if the ISR just returns after processing and removing one.
There are two types of transmit interrupt strategies:
1. One interrupt per packet.
2. One interrupt per sequence of packets.
The strategy is determined by how the transmit interrupt bits and the AUTO RELEASE bit are used.
TX INT bit - Set whenever the TX completion FIFO is not empty.
TX EMPTY INT bit - Set whenever the TX FIFO is empty.
AUTO RELEASE - When set, successful transmit packets are not written into completion FIFO, and their memory is
released automatically.
1. One interrupt per packet: enable TX INT, set AUTO RELEASE=0. The software driver can find the completion
result in memory and process the interrupt one packet at a time. Depending on the completion code the driver will
take different actions. Note that the transmit process is working in parallel and other transmissions might be taking
place. The LAN91C110 is virtually queuing the packet numbers and their status words.
SMSC DS – LAN91C110 REV. B
Page 44
Rev. 09/05/02
SMSC [ SMSC CORPORATION ]
