On Wed, Dec 2, 2020 at 8:08 PM Jason Gunthorpe <jgg@xxxxxxxx> wrote: > > On Wed, Dec 02, 2020 at 07:19:45PM -0500, Pavel Tatashin wrote: > > > It is a good moment to say, I really dislike how this was implemented > > > in the first place. > > > > > > Scanning the output of gup just to do the is_migrate_movable() test is > > > kind of nonsense and slow. It would be better/faster to handle this > > > directly while gup is scanning the page tables and adding pages to the > > > list. > > > > Hi Jason, > > > > I assume you mean to migrate pages as soon as they are followed and > > skip those that are faulted, as we already know that faulted pages are > > allocated from nomovable zone. > > > > The place would be: > > > > __get_user_pages() > > while(more pages) > > get_gate_page() > > follow_hugetlb_page() > > follow_page_mask() > > > > if (!page) > > faultin_page() > > > > if (page && !faulted && (gup_flags & FOLL_LONGTERM) ) > > check_and_migrate this page > > Either here or perhaps even lower down the call chain when the page is > captured, similar to how GUP fast would detect it. (how is that done > anyhow?) Ah, thank you for pointing this out. I think I need to address it here: https://soleen.com/source/xref/linux/mm/gup.c?r=96e1fac1#94 static __maybe_unused struct page *try_grab_compound_head() if (unlikely(flags & FOLL_LONGTERM) && is_migrate_cma_page(page)) return NULL; I need to change is_migrate_cma_page() to all migratable pages. Will study, and send an update with this fix. > > > I looked at that function, and I do not think the code will be cleaner > > there, as that function already has a complicated loop. > > That function is complicated for its own reasons.. But complicated is > not the point here.. > > > The only drawback with the current approach that I see is that > > check_and_migrate_movable_pages() has to check once the faulted > > pages. > > Yes > > > This is while not optimal is not horrible. > > It is. > > > The FOLL_LONGTERM should not happen too frequently, so having one > > extra nr_pages loop should not hurt the performance. > > FOLL_LONGTERM is typically used with very large regions, for instance > we are benchmarking around the 300G level. It takes 10s of seconds for > get_user_pages to operate. There are many inefficiencies in this > path. This extra work of re-scanning the list is part of the cost. OK, I did not realize that pinning was for such large regions, the path must be optimized. > > Further, having these special wrappers just for FOLL_LONGTERM has a > spill over complexity on the entire rest of the callchain up to here, > we now have endless wrappers and varieties of function calls that > generally are happening because the longterm path needs to end up in a > different place than other paths. > > IMHO this is due to the lack of integration with the primary loop > above > > > Also, I checked and most of the users of FOLL_LONGTERM pin only one > > page at a time. Which means the extra loop is only to check a single > > page. > > Er, I don't know what you checked but those are not the cases I > see. Two big users are vfio and rdma. Both are pinning huge ranges of > memory in very typical use cases. What I meant is the users of the interface do it incrementally not in large chunks. For example: vfio_pin_pages_remote vaddr_get_pfn ret = pin_user_pages_remote(mm, vaddr, 1, flags | FOLL_LONGTERM, page, NULL, NULL); 1 -> pin only one pages at a time RDMA indeed can do it in one chunk though. Regardless, the VFIO should probably be optimized to do it in a larger chunk, and the code path should be optimized for the reasons you gave above. > > > However, those changes can come after this series. The current series > > fixes a bug where hot-remove is not working with making minimal amount > > of changes, so it is easy to backport it to stable kernels. > > This is a good point, good enough that you should probably continue as > is I will continue looking into this code, and see if I can address your concerns in a follow-up fixes. Thanks, Pasha > > Jason
