libusrp/doc/traffic-cop-dma

129 lines
4.6 KiB
Plaintext
Raw Normal View History

Matt,
Here's my idea on the interface to the traffic cop. Basically I'm
thinking about treating it as 4 separate DMA channels, one for each of
the four possible "flows". In the interest of simplicity, I think we can
assign buffers 0,1 to channel 0; 2,3 to channel 1, etc...
port assignments
-----------------
0 SERDES
1 DSP pipeline
2 Gigabit ethernet MAC
3 RAM
registers
---------
MBZ == Must Be Zero
3 2 1
1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0
TC_DMA_SRC_{0,3} [WR] ("traffic cop DMA source, channel N")
Specifies where the writing port adapter writes info the buffer, and
the maximum number of lines to write.
5 9 9 9
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| MBZ | start | end (max) | step |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
start: starting line number for transfer (32-bit lines)
end: maximum number of lines to write
step: normally 1.
TC_DMA_DST_{0,3} [WR]
Specifies the range of lines that the reading port adapter accesses.
The number of lines to be transferred is controlled by the source.
5 9 9 9
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| MBZ | start | MBZ | step |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
start: starting line number tranfer (32-bit lines)
step: normally 1.
TC_DMA_CTRL_{0,3} [WR]
4 4 1 19 4
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| src | dst |A| MBZ | cmd |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
src: source port number. E.g., 2 = ethernet MAC (the buffer writer)
dst: destination port number. E.g., 1 = DSP pipeline (the buffer reader)
A: Set if "Host Approval" is required before beginning xfer to dst.
Used to allow processor to inspect packet for s/w dispatch. If set,
traffic cop sets PENDING_APPROVAL bit (and causes interrupt?) after the
2nd line has been written into the buffer. Reader is held off
until hosts APPROVES or DROPS the buffer.
cmd: command
TC_DMA_CMD_RESET 0 // abort active tranfers now; reset to idle state
TC_DMA_CMD_START 1 // begin transfers according to src, dst, A
TC_DMA_CMD_STOP 2 // stop transfers at completion of current buffer
TC_DMA_CMD_APPROVE_0 3 // host approves xfer on even buffer, continue
TC_DMA_CMD_APPROVE_1 4 // host approves xfer on odd buffer, continue
TC_DMA_CMD_DROP_0 5 // host naks xfer on even buffer, drop buffer and continue
TC_DMA_CMD_DROP_1 6 // host naks xfer on even buffer, drop buffer and continue
TC_DMA_STATUS_{0,3} [RD]
10 10
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| nwritten even | nwritten odd | state? | flags |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
nwritten even: number of lines written into even buffer
nwritten odd: number of lines written into odd buffer
flag bits:
bmTCDS_PENDING_APPROVAL_0 (1 << 0) // pending host approval on even buffer
bmTCDS_PENDING_APPROVAL_1 (1 << 1) // pending host approval on odd buffer
bmTCDS_WRITE_DONE_0 (1 << 2) // the even buffer write is complete
bmTCDS_WRITE_DONE_1 (1 << 3) // the odd buffer write is complete
I think the combination of the "host approval" and WRITE_DONE bits
will allow us to handle the cases where the host looks and doesn't
care, and the case where the host looks, cares, and needs to wait
until it sees the whole packet.
WRITE_DONE_* should be cleared when the corresponding buffer is
selected to be written into (e.g., when swapping buffers, and at init)
prior to writing the first line. WRITE_DONE_* is set when the
requested number of lines have been written into the buffer.
I also want a "global status register" that pulls the 4 flag bits
from each of the 4 status registers into a single word. This should
allow me to read a single word to figure out what to do.
TC_DMA_STATUS_GLOBAL [RD]
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| MBZ |flags3 |flags2 |flags1 |flags0 |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
I'm sure I've missed something...
What do you think?
Eric