On Tue, 18 May 2004 11:52:45 +0200, Matthias Müller <(E-Mail Removed)> wrote:
>Hello,
>I have developed a fully PCIX compliant card which performs a DMA into
>the systems memory (64bit/133Mhz). The PCIX controller on my card is a
>Xilinx PCIX core, running on a Virtex2P FPGA. I use the PCI command
>x"07" to write burst of 4Kb. The OS is Linux Redhat 2.4.26. The DMA
>works properly on a PC with Intel E7505 chipset. But actually I want to
>use a HP Prolinat DL380 system with serverworks chipset.
>If I want to do the DMA there the system hangs up. ONLY if the burst
>size is reduced to 1Kb, than it works with several systems warnings like
>"dazed and confused...probably you have a problem with your RAM chips"
>and the first 64bit word is always wrong.
>Because the DMA runs properly on the PC with Intel chipset I think there
>is a problem with the serverworks chipset and the 4Kb-DMA. Does anyone
>know the problem?
>Thank you, Matthias
You're probably going to need an NDA one way or the other, but the
sightings/errata reports are what you need to see. I can say I have not heard
of similar misbehavior on our GC-LE or -HE platforms.
But...we only certify a limited breadth of PCI agents, and users are not
allowed to stick random PCI cards into our systems, all of which cuts down the
range of exposure to nasty bugs like yours. It's a blessing :-)
And we do not use the CSB5 the way most GC designs use it: most use the thin
IMB link to tie it to a CMIC, and then have the CSB5 cough out a 32b/33mhz PCI
bus to embedded devices and occasionally to a slot or two. But all of our
slots are behind a CIOB, and our CSB5 simply connects to a PCI bus as a peer
and not a bridge.
If the slot you're using on the HP is behind a CSB5 you'd be using
functionality we never use, covered by errata/sightings reports I'd likely not
be able to remember...
/daytripper (old dudes only have so much functional memory to spare ;-)
|
Similar Threads
