On 09/28/11 00:57, Steve Reinhardt wrote: > On Tuesday, September 27, 2011, Gabe Black wrote: > >> Oh, something I forgot to mention. The idea of an address space and how >> their managed needs to be generalzed. I'm thinking Workload objects will >> have an AddressSpace object associated with them that will take care of >> translating from virtual to physical addresses, either by looking up the >> page tables in memory in FS mode, or by keeping around fake page tables >> in SE mode. >> > I'm confused... to me the whole notion of a simulator-visible "address > space" concept is an SE-mode artifact that's already handled pretty capably > by the PageTable object. Address spaces and page tables in FS mode are > managed by the OS running inside the simulator, and are basically invisible > to the simulator itself. (The concept leaks through a little in that we do > have to implement the TLB and (when appropriate) hardware page-table walker > appropriately, and there's the 'vtophys' hook for introspecting into user > programs in FS mode, but those are black boxes that don't require the > simulator to actually understand what an "address space" is.) > > If we ever build the fabled "hypervisor emulation" layer then I can see > combining these, using the fake PageTable to manage the > guest-physical-to-machine mappings while the guest OSes use the current FS > mechanisms to manage guest-virtual-to-guest-physical mappings, but that > still doesn't require the simulator to understand what a guest virtual > address space is. > > Also I think the PageTable should stay with the Process object and not be > part of Workload... if we ever implement fork() in SE mode then we could end > up having multiple Process objects (and PageTables) associated with a single > Workload.
Yeah, I tried to merge the SE and FS versions of ports with translation a long time ago. There was a thread about it on the list. That was yet another instance where the little gotchas turned it into a worthless disaster. A PageTable object is an example of an address space. A guest page table in memory is an address space. Because we have two, incompatible abstractions for essentially the same thing, that isn't justification for not having one abstraction to represent both. A Workload is an abstraction. I should have to care that when I access address 0xffff0000 that it's looked up in a fake PageTable hash map, or if it's looked up in some guest memory on some NUMA node somewhere. I just want to know what's on the stack. A process is a workload, a pagetable is an address space, a process has a pagetable, a workload has an address space. It's all basically the same thing, but we don't have two of everything. Gabe _______________________________________________ gem5-dev mailing list [email protected] http://m5sim.org/mailman/listinfo/gem5-dev
