> > > Oh, that existing logic is wrong too :( Another bug. > > > > I do not think there is a bug. > > > > > You can't skip pages in the pages[] array under the assumption they > > > are contiguous. ie the i+=step is wrong. > > > > If pages[i] is part of a compound page, the other parts of this page > > must be sequential in this array for this compound page > > That is true only if the PMD points to the page. If the PTE points to > a tail page then there is no requirement that other PTEs are > contiguous with the compount page. > > At this point we have no idea if the GUP logic got this compound page > as a head page in a PMD or as a tail page from a PTE, so we can't > assume a contiguous run of addresses. I see, I will fix this bug in an upstream as a separate patch in my series, and keep the fix when my fixes are applied. > > Look at split_huge_pmd() - it doesn't break up the compound page it > just converts the PMD to a PTE array and scatters the tail pages to > the PTE. Got it, unfortunately the fix will deoptimize the code by having to check every page if it is part of a previous compound page or not. > > I understand Matt is pushing on this idea more by having compound > pages in the page cache, but still mapping tail pages when required. > > > This is actually standard migration procedure, elsewhere in the kernel > > we migrate pages in exactly the same fashion: isolate and later > > migrate. The isolation works for LRU only pages. > > But do other places cause a userspace visible random failure when LRU > isolation fails? Makes sense, I will remove maximum retries for isolation, and retry indefinitely, the same as it is done during memory hot-remove. So, we will fail only when migration fails. > > I don't like it at all, what is the user supposed to do? > > Jason