On 20.01.22 18:22, Linus Torvalds wrote: > On Thu, Jan 20, 2022 at 5:46 PM David Hildenbrand <david@xxxxxxxxxx> wrote: >> >> I'm, not concerned about fork(), I'm concerned about other false positives. > > Without a fork(), you won't have the THP marked for COW, so is it > really an issue? Oh sorry, I should have been clearer. I absolutely agree that there is no way around COW if there are multiple sharers of a page :) If you fork() and write to a shared page, you have to expect that THP get split and that you get a copy. No change on that front. At this point I'm trying to assess the impact of the change and how to eventually mitigate it if it turns out to be a real problem. (I've been spending a lot of time trying to understand all the complexity, and sometimes my brain is ... a bit overloaded by all of that. Well, at least I learn a lot ...) One scenario I have in mind how we can end up easily with R/O mapped exclusive THP is a simple fork() with an immediate exec(), unmap() or exit() of the child. Might not be the most efficient way of doing things, but that doesn't mean that existing user space doesn't rely on it not happening. Then, of course, there are other ways to read-protect THP temporarily (mprotect() and friends), where you wouldn't expect to lose your THP, but it's somewhat a secondary concern most probably . Obviously, we need another (temporary/speculative) reference on the THP or writeback to actually split it. I'm *hoping* that it will be so rare that we really don't care. I decided to always lock the THP in do_wp_page() to at least handle page migration and swapcache more reliably. To answer you question: people optimized the reuse case heavily previously (reuse_swap_page()), and I can see how it might happen. I'm hoping that it won't matter in practice, but I'd like to at least document the impact and have some magic solution in sleeve that I can just pull out when reports start coming in. > >> Here is what I currently have, I hope that makes sense: > > From a quick look, that patch looks fine to me, but there might be > something I'm missing... And who knows what odd usage patterns there > might be in this area. The whole odd Android thing with forking that > zygote process. Exactly my thoughts. I'm trying to be very careful. -- Thanks, David / dhildenb
