On 25/03/2015 20:10, Hervé Poussineau wrote: > Le 25/03/2015 15:45, Paolo Bonzini a écrit : >> >> >> On 05/03/2015 23:13, Hervé Poussineau wrote: >>> Add a new memory region in system address space where DMA address space >>> definition (the 'translation table') belongs, so we can update on the >>> fly >>> the DMA address space. >>> >>> Signed-off-by: Hervé Poussineau <hpous...@reactos.org> >> >> Would it make sense to just use an IOMMU region for the DMA address >> space? > > I don't really know. Currently, the first user is the dp8393x network > card (a sysbus device), which is connected through the DMA address space. > It means that *every* read or write access needs to be translated, even > those accessing 1 or 2 bytes. dp8393x has no direct connection to the > system memory. > On some other machines, like the m68k Quadra 800, the dp8393x is only > connected to the system memory. > > A second user is the ESP SCSI card in DMA mode, using the DMA address > space with the help of the DMA controller. > > As DMA address space mapping doesn't change much (except initialization > phases), and that it is used for small (1 or 2 bytes) and big transfers > (network packets and disk accesses), I thought that an address space was > better.
Both are okay. The IOMMU makes address space changes faster; your scheme is basically a form of caching, it trades update performance for improved translation performance. The trick you're using with shadowing the RAM is nifty, but it leaks a memory region. I'll reply to the main patch. Paolo
