On Thu, Apr 16, 2020 at 03:03:12PM -0700, Dave Hansen wrote: > On 4/16/20 2:32 PM, Kirill A. Shutemov wrote: > > With current code it's one way road: kernel tries to avoid splitting > > large pages, but it doesn't restore them back even if page attributes > > got compatible again. > > Looks pretty sane to me, and sounds like something we've needed for a > long time. > > I'd having doubts in my ability to find nasty corner cases in this code, > though. Could you rig up some tests to poke at this thing further? Maybe: > > Record what the direct map looks like (even from userspace). Then, > allocate some memory, including odd-sized and aligned ranges. Try to do > things >4M like the hugetlbfs code does. Make the allocation (or a > piece of it) not-present (or whatever), which usually fractures some > large pages. Then put it back the way it was. All the large pages > should come back. > > If it survives that for an hour or two, it should be pretty good to go. > Basically, fuzz it. We already have it in kernel: CONFIG_CPA_DEBUG. It messes up with the mapping every 30 seconds. It is pretty good for the change too. It produces a lot of 2M/1G pages to be restored. I run it over night in my setup and it survives. -- Kirill A. Shutemov
