On 3/24/21 10:04 AM, Brijesh Singh wrote: > The spliting of the physmap is a temporary solution until we work to > improve the kernel page fault handler to split the pages on demand. > One of the disadvtange of splitting is that eventually, we will end up > breaking down the entire physmap unless we combine the split pages back to > a large page. I am open to the suggestation on various approaches we could > take to address this problem.
Other than suggesting that the hardware be fixed to do the fracturing itself? :) I suspect that this code is trying to be *too* generic. I would expect that very little of guest memory is actually shared with the host. It's also not going to be random guest pages. The guest and the host have to agree on these things, and I *think* the host is free to move the physical backing around once it's shared. So, let's say that there a guest->host paravirt interface where the guest says in advance, "I want to share this page." The host can split at *that* point and *only* split that one page's mapping. Any page faults would occur only if the host screws up, and would result in an oops. That also gives a point where the host can say, "nope, that hugetlbfs, I can't split it".
