> I don't understand. The current driver will get exactly one > interrupt per frame (or maybe less if frames are coming in fast > and they are queued), regardless of the frame size.
Sorry, I didn't figure it out clearly. As you know the I bit in BDs is just to let the cpm interrupt the cpu when a buffer not a frame is received. (yes, in init_fcc_param(), set the rx BDs to bdp->cbd_sc = BD_ENET_RX_EMPTY | BD_ENET_RX_INTR; or bdp->cbd_sc = BD_ENET_RX_EMPTY; has the same result in the current driver. That's to say we will receive an interrupt when a "frame" is received. The cpm will issue RXF interrupt wheneven a frame is received if it's not masked whether you set the I bit or not) If I want to use the I bit control the interrupt, I can only mask the RXF interrupt, and unmask the RXB interrupt. When a buffer is full filled with the I bit set in the BD, a RXB interrupt will occur. But there is a problem, if the buffer is lager enough to hold an frame (that's the case of the current driver: ep->fen_genfcc.fcc_mrblr = PKT_MAXBLR_SIZE; /* 1536 */) no RXB interrupt is issued since the RXF will be isssed though it is masked by me. My solution is set the mrblr to a small value(64), and the interrupts will be issued whenever an buffer is used up by the cpm. The trick is not to set the I bit in all BDs, but set it one by one or one by more. With this way the interrupt number will be controled by the I bits in the BDs. Yes, it still has problems, if the last buffer of a frame has the I bit set in the BD, no RXB interrupt will be issued for the above reason. The worst case is the I bits are all set only in the last buffers of receiving frames, no interrupts are issued. I hope I have explained all now, but maybe my poor Englist won't let me do so:) > > The purpose for using first/last is so you can split a frame across multiple > BDs, making the use of memory a little more efficient. Using this method, > you can't dma directly into sk buffers, you will have to check for the > buffer sequences and copy the data into the sk buffers. This method also I do can use dma into small buffers, and then copy to a sk buffer.Since the current driver copys data to sk buffer too, it won't cost much more time, though it does incur more bus accesses to the BDs(both the cpu & cpm). Shenrong ** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/
