On Tue 20-10-20 17:19:42, Guilherme Piccoli wrote: > When I first wrote that, the design was a bit different, the flag was > called __GFP_HTLB_PAGE or something like that. The design was to > signal/mark the composing pages of hugetlb as exactly this: they are > pages composing a huge page of hugetlb "type". Then, I skipped the > "init_on_alloc" thing for such pages. As pointed out in the other email. This is not about hugetlb although this might be visible more than other because they just add a tiny bit to an overall overhead. Each page cache read, CoW and many many other !__GFP_ZERO users are in the same position when they double initialize. A dedicated __GFP_HTLB_PAGE is really focusing on a wrong side of the problem. We do have __GFP_ZERO for a good reason and that is to optimize the initialization. init_on_alloc goes effectively against this approach with a "potentially broken code" philosophy in mind and that is good as a hardening measure indeed. But that comes with an increased overhead and/or shifted layer when the overhead happens. Sure there is some room to optimize the code here and there but the primary idea of the hardening is to make the initialization dead trivial and clear that nothing can sneak out. -- Michal Hocko SUSE Labs
