This slide confused me for a while so here's my understanding. The problem here is DMA, Direct Memory Access. Communication between applications (running on the processor) and an external device (e.g. a network card) often communicate by reading from and writing to specific addresses in memory.
Case 1: This is the case when the processor wants to send data to the network card (say we want to send a packet over the network). The problem is that the network card will access the memory directly. But if we have a write-back cache, our data may not have been written to memory yet. The key point is that the network card does not access memory with the same pattern as the processor (read() -> cache -> main memory). Instead, the network card reads from a direct address.
Case 2: In this case, data is flowing in the other direction. The network card received a message and wants to pass it to the processor. It writes to the memory address and tells the processor that the data is ready. But if the processor has the addresses corresponding to the message buffer still in cache, it will read that data instead.
This comment was marked helpful 1 times.
fangyihua
Processor will not go to the main memory if the address is in the cache. However, other devices may update main memory directly, so the cache will have outdated value. In this way, the memory and cache are not synchronized.
This slide confused me for a while so here's my understanding. The problem here is DMA, Direct Memory Access. Communication between applications (running on the processor) and an external device (e.g. a network card) often communicate by reading from and writing to specific addresses in memory.
Case 1: This is the case when the processor wants to send data to the network card (say we want to send a packet over the network). The problem is that the network card will access the memory directly. But if we have a write-back cache, our data may not have been written to memory yet. The key point is that the network card does not access memory with the same pattern as the processor (read() -> cache -> main memory). Instead, the network card reads from a direct address.
Case 2: In this case, data is flowing in the other direction. The network card received a message and wants to pass it to the processor. It writes to the memory address and tells the processor that the data is ready. But if the processor has the addresses corresponding to the message buffer still in cache, it will read that data instead.
This comment was marked helpful 1 times.
Processor will not go to the main memory if the address is in the cache. However, other devices may update main memory directly, so the cache will have outdated value. In this way, the memory and cache are not synchronized.
This comment was marked helpful 0 times.