> "In PCI, a transaction that cannot be completed immediately > I think this *is* the only reason for failure, and it's compatible with what I observe. The failure usually happens only when moving big files. one day ago, on a second C3600 (yes, we have two workstations), we spent 10 hours at moving small files (<10Mbyte) on a "SY-PCX40009" "Silicon Image Sil3124" SIL24 card, with a sleep of 2 seconds between each copy, and we registered zero failures. Whereas, when copying a file bigger than 8Gbyte files we had a probability of 0.97 to end with a PCI error on the bus, and repeating the test, this usually happened within 10 minutes. results were the same if the card was plugged in PCI_SLOT2 (3.3V), or PCI_SLOT1 (5V): all the same behavior! what do I conclude? the voltage of signals is OK, the onboard voltage-level-shifter was doing a good job, so I think that it all a timing problem: the less an I/O task takes the PCI bus busy, the best it is. maybe, interrupt problem? some got lost? how to investigate the exact reason for the failure? did a PCI transaction fail when recycled? or did it fail due to interrupt lost? The MSI stuff is too complex and I haven't understood it, the wiki seems confusing to me, and we do not have it enabled in the kernel config. anyway, if it's a firmware problem, all related to the firmware running the SATA card, it may be fixed by a kernel driver forcing the chip into PCI64 mode. I am surprised that a PCI-X chip isn't able to switch into PCI64 mode. Cards are usually declared "backward compatible", and it would be just a matter of "how" to handle the finite state machine that drives the PCI, plus the onboard PLL and synchronization. > Here are some Adaptec 64-bit cards that I found. > https://storage.microsemi.com/en-us/support/raid/sata/aar-2410sa/ > https://storage.microsemi.com/en-us/support/raid/sata/aar-21610sa/ as written here (1), we have already tested Adaptec 2410SA, and I have also recently received an email by AlanCox confirming that the card is x86-only! Subsystem: Adaptec AAR-2410SA PCI SATA 4ch (Jaguar II) It fails on every non-x86 machine because it relies on PC-BIOS extension for the initialization of the onboard Intel i960 chip, and we do not have any Linux kernel driver able to do what the PC-BIOS does. I had opened a BUG report to a Linux mailing list, and Alan Cox replied to me privately.
