On Fri, May 24, 2019 at 05:00:32PM +0300, Kirill Tkhai wrote: > On 24.05.2019 14:52, Kirill A. Shutemov wrote: > > On Fri, May 24, 2019 at 01:45:50PM +0300, Kirill Tkhai wrote: > >> On 22.05.2019 18:22, Kirill A. Shutemov wrote: > >>> On Mon, May 20, 2019 at 05:00:01PM +0300, Kirill Tkhai wrote: > >>>> This patchset adds a new syscall, which makes possible > >>>> to clone a VMA from a process to current process. > >>>> The syscall supplements the functionality provided > >>>> by process_vm_writev() and process_vm_readv() syscalls, > >>>> and it may be useful in many situation. > >>> > >>> Kirill, could you explain how the change affects rmap and how it is safe. > >>> > >>> My concern is that the patchset allows to map the same page multiple times > >>> within one process or even map page allocated by child to the parrent. > >>> > >>> It was not allowed before. > >>> > >>> In the best case it makes reasoning about rmap substantially more difficult. > >>> > >>> But I'm worry it will introduce hard-to-debug bugs, like described in > >>> https://lwn.net/Articles/383162/. > >> > >> Andy suggested to unmap PTEs from source page table, and this make the single > >> page never be mapped in the same process twice. This is OK for my use case, > >> and here we will just do a small step "allow to inherit VMA by a child process", > >> which we didn't have before this. If someone still needs to continue the work > >> to allow the same page be mapped twice in a single process in the future, this > >> person will have a supported basis we do in this small step. I believe, someone > >> like debugger may want to have this to make a fast snapshot of a process private > >> memory (when the task is stopped for a small time to get its memory). But for > >> me remapping is enough at the moment. > >> > >> What do you think about this? > > > > I don't think that unmapping alone will do. Consider the following > > scenario: > > > > 1. Task A creates and populates the mapping. > > 2. Task A forks. We have now Task B mapping the same pages, but > > write-protected. > > 3. Task B calls process_vm_mmap() and passes the mapping to the parent. > > > > After this Task A will have the same anon pages mapped twice. > > Ah, sure. > > > One possible way out would be to force CoW on all pages in the mapping, > > before passing the mapping to the new process. > > This will pop all swapped pages up, which is the thing the patchset aims > to prevent. > > Hm, what about allow remapping only VMA, which anon_vma::rb_root contain > only chain and which vma->anon_vma_chain contains single entry? This is > a vma, which were faulted, but its mm never were duplicated (or which > forks already died). The requirement for the VMA to be faulted (have any pages mapped) looks excessive to me, but the general idea may work. One issue I see is that userspace may not have full control to create such VMA. vma_merge() can merge the VMA to the next one without any consent from userspace and you'll get anon_vma inherited from the VMA you've justed merged with. I don't have any valid idea on how to get around this. -- Kirill A. Shutemov
