On Fri, Aug 11, 2023 at 12:11:55PM -0400, Zi Yan wrote: > On 11 Aug 2023, at 12:08, David Hildenbrand wrote: > > > On 11.08.23 17:58, Peter Xu wrote: > >> On Fri, Aug 11, 2023 at 05:32:37PM +0200, David Hildenbrand wrote: > >>> On 11.08.23 17:18, Peter Xu wrote: > >>>> On Fri, Aug 11, 2023 at 12:27:13AM +0200, David Hildenbrand wrote: > >>>>> On 10.08.23 23:48, Matthew Wilcox wrote: > >>>>>> On Thu, Aug 10, 2023 at 04:57:11PM -0400, Peter Xu wrote: > >>>>>>> AFAICS if that patch was all correct (while I'm not yet sure..), you can > >>>>>>> actually fit your new total mapcount field into page 1 so even avoid the > >>>>>>> extra cacheline access. You can have a look: the trick is refcount for > >>>>>>> tail page 1 is still seems to be free on 32 bits (if that was your worry > >>>>>>> before). Then it'll be very nice if to keep Hugh's counter all in tail 1. > >>>>>> > >>>>>> No, refcount must be 0 on all tail pages. We rely on this in many places > >>>>>> in the MM. > >>>>> > >>>>> Very right. > >>>> > >>>> Obviously I could have missed this in the past.. can I ask for an example > >>>> explaining why refcount will be referenced before knowing it's a head? > >>> > >>> I think the issue is, when coming from a PFN walker (or GUP-fast), you might > >>> see "oh, this is a folio, let's lookup the head page". And you do that. > >>> > >>> Then, you try taking a reference on that head page. (see try_get_folio()). > >>> > >>> But as you didn't hold a reference on the folio yet, it can happily get > >>> freed + repurposed in the meantime, so maybe it's not a head page anymore. > >>> > >>> So if the field would get reused for something else, grabbing a reference > >>> would corrupt whatever is now stored in there. > >> > >> Not an issue before large folios, am I right? Because having a head page > >> reused as tail cannot happen iiuc with current thps if only pmd-sized, > >> because the head page is guaranteed to be pmd aligned physically. > > > > There are other users of compound pages, no? THP and hugetlb are just two examples I think. For example, I can spot __GFP_COMP in slab code. > > > > Must such compound pages would not be applicable to GUP, though, but to PFN walkers could end up trying to grab them. > > > For FS supporting large folios, their page cache pages can be any order <= PMD_ORDER. > See page_cache_ra_order() in mm/readahead.c Ah yes.. > > >> > >> I don't really know, where a hugetlb 2M head can be reused by a 1G huge > >> later right during the window of fast-gup walking. But obviously that's not > >> common either if that could ever happen. > >> > >> Maybe Matthew was referring to something else (per "in many places")? > > > > There are some other cases where PFN walkers want to identify tail pages to skip over them. See the comment in has_unmovable_pages(). Indeed. Thanks! -- Peter Xu
