On 5/23/23 14:25, Sean Christopherson wrote: >> There are consequences for converting pages between shared and private. >> Doing it on a vmalloc() mapping is guaranteed to fracture the underlying >> EPT/SEPT mappings. >> >> How does this work with load_unaligned_zeropad()? Couldn't it be >> running around poking at one of these vmalloc()'d pages via the direct >> map during a shared->private conversion before the page has been accepted? > Would it be feasible and sensible to add a GFP_SHARED or whatever, to communicate > to the core allocators that the page is destined to be converted to a shared page? > I assume that would provide a common place (or two) for initiating conversions, > and would hopefully allow for future optimizations, e.g. to keep shared allocation > in the same pool or whatever. Sharing memory without any intelligence as to what > memory is converted is going to make both the guest and host sad. I don't think we want a GFP flag. This is still way too specialized to warrant one of those. It sounds like a similar problem to what folks want for modules or BPF. There are a bunch of allocations that are related and can have some of their setup/teardown costs amortized if they can be clumped together. For BPF, the costs are from doing RW=>RO in the kernel direct map, and fracturing it in the process. Here, the costs are from the private->shared conversions and fracturing both the direct map and the EPT/SEPT. I just don't know if there's anything that we can reuse from the BPF effort.
