On Wed, Jan 29, 2020 at 02:26:06PM -0800, John Hubbard wrote: > On 1/29/20 3:25 AM, Kirill A. Shutemov wrote: > > On Tue, Jan 28, 2020 at 07:24:10PM -0800, John Hubbard wrote: > >> When debugging a problem that involves compound pages, it is extremely > >> helpful if dump_page() reports not only the page->_refcount, but also > >> the refcount of the head page of the compound page. That's because the > >> head page collects refcounts for the entire compound page. > >> > >> Therefore, enhance dump_page() so as to print out the refcount of the > >> head page of a compound page. > >> > >> This approach (printing information about a struct page that is not the > >> struct page that was passed into dump_page()) has a precedent: > >> compound_mapcount is already being printed. > > > > refcount on a tail must always be 0. I think we should only print it when > > it is non-zero, emphasizing this fact with a standalone message. > > > > Hi Kirill, > > Yes, good point, that sounds like just the right balance. And it avoids adding > a new item to print in the common case, which is very nice. Here's what I've > changed it to for the next version (I'll also rewrite the commit description > accordingly): > > > diff --git a/mm/debug.c b/mm/debug.c > index a90da5337c14..3a45e2b77de0 100644 > --- a/mm/debug.c > +++ b/mm/debug.c > @@ -75,12 +75,17 @@ void __dump_page(struct page *page, const char *reason) > */ > mapcount = PageSlab(page) ? 0 : page_mapcount(page); > > - if (PageCompound(page)) > - pr_warn("page:%px refcount:%d mapcount:%d mapping:%px " > - "index:%#lx compound_mapcount: %d\n", > - page, page_ref_count(page), mapcount, > + if (PageCompound(page)) { > + pr_warn("page:%px compound refcount:%d mapcount:%d mapping:%px " > + "index:%#lx compound_mapcount:%d\n", > + page, page_ref_count(compound_head(page)), mapcount, > page->mapping, page_to_pgoff(page), > compound_mapcount(page)); > + > + if (page != compound_head(page) && page_ref_count(page) != 0) > + pr_warn("page:%px PROBLEM: non-zero refcount (==%d) on " > + "this tail page\n", page, page_ref_count(page)); > + } > else > pr_warn("page:%px refcount:%d mapcount:%d mapping:%px index:%#lx\n", > page, page_ref_count(page), mapcount, I have a hunk in my current tree which looks like this: @@ -77,6 +77,11 @@ void __dump_page(struct page *page, const char *reason) pr_warn("page:%px refcount:%d mapcount:%d mapping:%px index:%#lx \n", page, page_ref_count(page), mapcount, page->mapping, page_to_pgoff(page)); + if (PageTail(page)) { + struct page *head = compound_head(page); + pr_warn("head:%px mapping:%px index:%#lx\n", + head, head->mapping, page_to_pgoff(head)); + } if (PageKsm(page)) pr_warn("ksm flags: %#lx(%pGp)\n", page->flags, &page->flags); else if (PageAnon(page)) I wonder if we can combine these two patches in some more useful way? I also think we probably want a sanity check that 'head' and 'page' are within a sane range of each other (ie head < page and head + MAX_ORDER_NR_PAGES > page) to protect against a struct page that contains complete garbage.