Re: PCI DMA into USER space

From: Jes Sorensen (jes@linuxcare.com)
Date: Wed Jun 14 2000 - 10:58:39 EST


>>>>> "Richard" == Richard B Johnson <root@chaos.analogic.com> writes:

Richard> On 14 Jun 2000, Jes Sorensen wrote:
>> Rubbish, PCI shared memory is *slow* you want DMA. Granted, PCI
>> shared

Richard> Hardly rubbish.

PCI shared memory is slow

>> memory is faster than the brain dead IO ports but it's far from as
>> fast as streaming DMA or in particular host memory access. In a
>> modern PC you easily get 300-400MB/sec of real bandwidth between
>> the CPU and

Richard> You can easily get over 1,000 MB/sec between RAM and the
Richard> CPU. This is relevent only during a copy operation. It has
Richard> nothing to do with DMA.

For pure reads, I am talking copy speed - granted my numbers are
pretty old and based on P2 boxes. Easily get over 1GB/sec on a PCs
however ... well it's not easy.

>> system memory whereas you'll get far less than 100MB/sec when
>> accessing the PCI memory (maybe half if you're lucky). As such it
>> is

Richard> A 33MHz, 32-bit Revision 2.2-compliant PCI bus gives you 132
Richard> MB/second. That's the spec. There is no way to slow it down
Richard> except by a disconnect.

Would you do us all a favor and go read the PCI specs then? You'll
*theoretically* get 132MB/sec if you do *streaming* DMA, there is no
way in hell you'll get that with indiidual CPU transactions. PCI was
designed for streaming access not for random access.

>> quite often worth it to DMA to host memory and then let the CPU
>> access data from there.

Richard> Not ever. The DMA controllers can't work at CPU speed because
Richard> they can't keep the bus forever like the CPU can. The fly-by
Richard> mode supported by some chips allows you to get one transfer
Richard> per clock. The clock is only 16 MHz (max), not the 33 MHz of
Richard> PCI (some PCI even do 66 MHz although they are doing it on
Richard> "invented" specifications).

You're soo funny today, even if we for a second run around and
naiively believe we'll get 132MB/sec out of the PCI bus with CPU
random access then that is going to eat all your CPU cycles. If you on
the other hand DMA data to system memory and let the memory controller
fetch it in asynchronously with the CPU, it'll eat maybe a
significantly lower percentage of the total bandwidth and that will be
left over for the CPU to play with.

Jes

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/



This archive was generated by hypermail 2b29 : Fri Jun 23 2000 - 21:00:11 EST