вт, 13 нояб. 2018 г. в 20:59, Pavel Tatashin <pasha.tatashin@xxxxxxxxxx>: > > On 18-11-13 15:23:50, Oleksandr Natalenko wrote: > > Hi. > > > > > Yep. However, so far, it requires an application to explicitly opt in > > > to this behavior, so it's not all that bad. Your patch would remove > > > the requirement for application opt-in, which, in my opinion, makes > > > this way worse and reduces the number of applications for which this > > > is acceptable. > > > > The default is to maintain the old behaviour, so unless the explicit > > decision is made by the administrator, no extra risk is imposed. > > The new interface would be more tolerable if it honored MADV_UNMERGEABLE: > > KSM default on: merge everything except when MADV_UNMERGEABLE is > excplicitly set. > > KSM default off: merge only when MADV_MERGEABLE is set. > > The proposed change won't honor MADV_UNMERGEABLE, meaning that > application programmers won't have a way to prevent sensitive data to be > every merged. So, I think, we should keep allow an explicit opt-out > option for applications. > We just did not have VM/Madvise flag for that currently. Same as THP. Because all logic written with assumption, what we have exactly 2 states. Allow/Disallow (More like not allow). And if we try to add, that must be something like: MADV_FORBID_* to disallow something completely. And same for THP (because currently apps just refuse to start if THP enabled, because of no way to forbid thp). Thanks. > > > > > As far as I know, basically nobody is using KSM at this point. There > > > are blog posts from several cloud providers about these security risks > > > that explicitly state that they're not using memory deduplication. > > > > I tend to disagree here. Based on both what my company does and what UKSM > > users do, memory dedup is a desired option (note "option" word here, not the > > default choice). > > Lightweight containers is a use case for KSM: when many VMs share the > same small kernel. KSM is used in production by large cloud vendors. > > Thank you, > Pasha >
