Re: [rtl] PCI Transfert

["C. W. Wright (1698)" <wright@lidar.wff.nasa.gov>]

> For our application I will use a data acquisition board with 2
> dma on board. Do you think is realistic to write through the
> pci bus 64 Mbytes/s in mbuff share memory buffer ????

Hi Fred,

  I'm using PCI on-card DMA from a user-space program and I've achieved
89mb/second average transfer rate transferring groups of 120 1531 byte
blocks. I transfer from PCI to a reserved 4mb block above linux.  If you
use above-linux RAM you can be sure of contiguous memory and not have to
do scatter-gather or service larger numbers of dma interrupts resulting
from chained-lists.
	Once the data are in above-linux RAM, I process it and move it to a large
120mb queue held in mbuff space.  I use rt-linux to intercept an interrupt
from my special hardware, and schedule the user-space program via the
rt-linux FIFOs. 
	Mbuff is great, but the memory it returns is managed by the system
and it is composed of non-contiguous regions.  The memory for dma must be
given in terms of the true physical address, and must be contiguous.

-Wayne




--<<< We are committed to exceeding requirements >>>----->-+
C.W. Wright,wright@lidar.wff.nasa.gov,      Ph:757-824-1698
http://lidar.wff.nasa.gov                  Fax:603-925-6886      
NASA, Goddard Space Flight Center
Building N-159 Room E117, Code 972,Wallops Flight Facility
Wallops Island, Va. 23337   Telecommute phone: 410-742-7333
-------+-<-------------------------------------------------