Hi Nathan,

> That the word "probably" has to be used in the explanation of what's
> going on here makes me think that this issue isn't fully understood
> and doesn't make a good case for per-architecture "optimizing".

I didn't explore the issue more after the fix (defining ALLOW_MEMIO in
-current) solved the problem. The reported problem by Michael Wolfson
was

Adaptec AHA-2940 Ultra (SCSI mass storage) at ? dev 15 function 0 
(intrswiz 0, intrpin 0x1, i/o off, mem on, no quirks): Adaptec 2940
Ultra SCSI adapter: unable to map registers

The particular code during resolution
pointed to (in src/sys/dev/pci/ahc.c) (and I apologize for formatting
errors, Sylpheed is _really_ starting to irritate me)

ahc_pci_attach(parent, self, aux)

....

#ifdef AHC_ALLOW_MEMIO
        memh_valid = 0;
        memtype = pci_mapreg_type(pa->pa_pc, pa->pa_tag,
AHC_PCI_MEMADDR);        

      switch (memtype) {
         case PCI_MAPREG_TYPE_MEM | PCI_MAPREG_MEM_TYPE_32BIT:
         case PCI_MAPREG_TYPE_MEM | PCI_MAPREG_MEM_TYPE_64BIT:
                memh_valid = (pci_mapreg_map(pa, AHC_PCI_MEMADDR, 
                          memtype, 0, &memt, &memh, NULL, NULL) == 0);
               break;
        default:
                memh_valid = 0;
        }
#endif
        ioh_valid = (pci_mapreg_map(pa, AHC_PCI_IOADDR,
                                    PCI_MAPREG_TYPE_IO, 0, &iot,
                                    &ioh, NULL, NULL) == 0);
#if 0

  printf("%s: bus info: memt 0x%lx, memh 0x%lx, iot 0x%lx, ioh 0x%lx\n",
        ahc_name(ahc),(u_long)memt,(u_long)memh,(u_long)iot, 
(u_long)ioh);

#endif

        if (ioh_valid) {
                st = iot;
                sh = ioh;
#ifdef AHC_ALLOW_MEMIO
        } else if (memh_valid) {
                st = memt;
                sh = memh;
#endif
        } else {
                printf(": unable to map registers\n");
                return;
        }

It appeared that I/O was unavailable, and without AHC_ALLOW_MEMIO, the
mapping couldn't occur. With it back in -current, the card has worked
fine, or at least MW reported it was fine and hasn't said anything
since.

From a performance point of view, is writing to I/O space faster or
slower than writing to memory space? Because the above code appears to
default to using I/O even if memory space is available.

tim