LWN: Comments on "Memory management notifiers"

Memory management notifiers

roelofs — 2008-01-30T00:49:19+00:00

But what the article doesn't say is why a guest kernel would be interested.

Seems like primarily a performance issue to me. If the guest kernel doesn't know when its "RAM" is really swap, it's not going to be able to manage its memory as effectively as it might like. For example, it might be able to predict memory-usage patterns where the host kernel can't. Wasn't there a recent article(s) about a patch to do speculative read-in of swapped-out memory, specifically for the use-case where some automated overnight process pushes out OpenOffice/Firefox/etc., causing the user significant delays upon his/her return in the morning? (Perhaps even one of Con Kolivas' patches?)

Greg

Memory management notifiers

giraffedata — 2008-01-25T20:11:39+00:00

This would only benefit guest kernels that have been modified to take avantage of it, right?

I think that's obvious in, "The interested code then registers its notifier with:"

But what the article doesn't say is why a guest kernel would be interested. It says that because the guest kernel can't know when the host has invalidated a page, the host must never invalidate a page (i.e. keep the memory pinned). I guess I don't know how KVM works, but I've worked with virtual machines that don't have this issue.

That swapped out page should still be virtually resident. The guest's page table says so, and, consistent with that, when the guest does a load from its virtual address, the instruction completes without the guest seeing any page fault (because the host takes a page fault, reads the data in, and updates the real page table).

Memory management notifiers

bronson — 2008-01-24T18:39:18+00:00

This would only benefit guest kernels that have been modified to take avantage of it, right?