On 22.08.19 00:43, Richard Henderson wrote: > On 8/21/19 3:31 PM, Richard Henderson wrote: >>> Yes, that's what I mean, TARGET_PAGE_SIZE, but eventually crossing a >>> page boundary. The longer I stare at the MVCL code, the more broken it >>> is. There are more nice things buried in the PoP. MVCL does not detect >>> access exceptions beyond the next 2k. So we have to limit it there >>> differently. >> That language is indeed odd. >> >> The only reading of that paragraph that makes sense to me is that the >> hardware >> *must* interrupt MVCL after every 2k bytes processed. The idea that the user >> can magically write to a read-only page simply by providing length = 2MB and >> page that is initially writable is dumb. I cannot imagine that is a correct >> reading. >> >> Getting clarification from an IBM engineer on that would be good; otherwise I >> would just ignore that and proceed as if all access checks are performed. >> > > FWIW, splitting the operation at every aligned 2k boundary is exactly what the > Hercules emulator does: > > len3 = NOCROSS2KL(addr1,len1) ? len1 : (int)(0x800 - (addr1 & 0x7FF)); > len4 = NOCROSS2KL(addr2,len2) ? len2 : (int)(0x800 - (addr2 & 0x7FF)); > len = len3 < len4 ? len3 : len4; > /* Use concpy to ensure Concurrent block update consistency */ > concpy (regs, dest, source, len); > > After this it writes back the lengths and addresses to the > register file, and then if necessary loops back to the address > translation step.
That's almost exactly how I planned to fix MVCL :) -- Thanks, David / dhildenb
