On Thu, 23 Sep 2021, Kirill A. Shutemov wrote: > On Thu, Sep 23, 2021 at 12:40:14PM +0100, Matthew Wilcox wrote: > > On Thu, Sep 23, 2021 at 01:15:16AM -0400, Kent Overstreet wrote: > > > On Thu, Sep 23, 2021 at 04:23:12AM +0100, Matthew Wilcox wrote: > > > > (compiling that list reminds me that we'll need to sort out mapcount > > > > on subpages when it comes time to do this. ask me if you don't know > > > > what i'm talking about here.) > > > > > > I am curious why we would ever need a mapcount for just part of a page, tell me > > > more. > > > > I would say Kirill is the expert here. My understanding: > > > > We have three different approaches to allocating 2MB pages today; > > anon THP, shmem THP and hugetlbfs. Hugetlbfs can only be mapped on a > > 2MB boundary, so it has no special handling of mapcount [1]. Anon THP > > always starts out as being mapped exclusively on a 2MB boundary, but > > then it can be split by, eg, munmap(). If it is, then the mapcount in > > the head page is distributed to the subpages. > > One more complication for anon THP is that it can be shared across fork() > and one process may split it while other have it mapped with PMD. > > > Shmem THP is the tricky one. You might have a 2MB page in the page cache, > > but then have processes which only ever map part of it. Or you might > > have some processes mapping it with a 2MB entry and others mapping part > > or all of it with 4kB entries. And then someone truncates the file to > > midway through this page; we split it, and now we need to figure out what > > the mapcount should be on each of the subpages. We handle this by using > > ->mapcount on each subpage to record how many non-2MB mappings there are > > of that specific page and using ->compound_mapcount to record how many 2MB > > mappings there are of the entire 2MB page. Then, when we split, we just > > need to distribute the compound_mapcount to each page to make it correct. > > We also have the PageDoubleMap flag to tell us whether anybody has this > > 2MB page mapped with 4kB entries, so we can skip all the summing of 4kB > > mapcounts if nobody has done that. > > Possible future complication comes from 1G THP effort. With 1G THP we > would have whole hierarchy of mapcounts: 1 PUD mapcount, 512 PMD > mapcounts and 262144 PTE mapcounts. (That's one of the reasons I don't > think 1G THP is viable.) > > Note that there are places where exact mapcount accounting is critical: > try_to_unmap() may finish prematurely if we underestimate mapcount and > overestimating mapcount may lead to superfluous CoW that breaks GUP. It is critical to know for sure when a page has been completely unmapped: but that does not need ptes of subpages to be accounted in the _mapcount field of subpages - they just need to be counted in the compound page's total_mapcount. I may be wrong, I never had time to prove it one way or the other: but I have a growing suspicion that the *only* reason for maintaining tail _mapcounts separately, is to maintain the NR_FILE_MAPPED count exactly (in the face of pmd mappings overlapping pte mappings). NR_FILE_MAPPED being used for /proc/meminfo's "Mapped:" and a couple of other such stats files, and for a reclaim heuristic in mm/vmscan.c. Allow ourselves more slack in NR_FILE_MAPPED accounting (either count each pte as if it mapped the whole THP, or don't count a THP's ptes at all - you opted for the latter in the "Mlocked:" accounting), and I suspect subpage _mapcount could be abandoned. But you have a different point in mind when you refer to superfluous CoW and GUP: I don't know the score there (and I think we are still in that halfway zone, since pte CoW was changed to depend on page_count, but THP CoW still depending on mapcount). Hugh
