On 12/13/2011 03:10 PM, Christoffer Dall wrote: > >> the question is just which mappings are the > >> most efficient to reclaim. > > > > Do you have accessed bits in those PTEs? > > > > nope. We can protect the underlying target pages though, but... Yeah, we have the same issue with one of the vendors. Fortunately only 90% of the market is affected. > > It's not really critical to have efficient reclaim here, since it > > happens so rarely. It just needs to do something. > > > > when would you trigger it - when it reaches a certain limit, or? And > then what, free the lot and re-allocate what's needed? The kernel triggers it based on internal pressure. It tells you how much pressure to apply, so you just translate it to a number of pages to free. > >> The other problem, the actual guest memory consuming too much memory, > >> I assumed this limit would be set by the user when creating his/her > >> VM, or can we do something smarter? (again, forgive my ignorance). > >> What is the alternative to pinning actual guest pages > > > > mmu notifiers - pages aren't pinned; instead, Linux calls back into kvm > > when modifying a host pte, and kvm responds by dropping or modifying its > > translation (second stage pte in your case). > > > > ah ok, so this works across VM boundary. Based on hyper-calls I presume? No, it's completely internal to the host. See for example kvm_mmu_notifier_invalidate_page() (in common code). It's called when Linux-as-host wants to change a pte (say to swap a page). kvm responds by translating the host virtual address into a guest physical address (via the memory slots table), then zapping the relevant pte and flushing and TLBs which may have cached the pte. > > mmu notifiers are also useful for other optimizations, like ksm, > > ballooning, and transparent huge pages. > > > > I know those features have to be supported eventually. The question is > if all this must be in place before a merge upstream? It doesn't have to be there for the merge but I recommend giving it high priority. At least read and understand the code so the addition will follow naturally. -- error compiling committee.c: too many arguments to function
