On Sun, Jun 15, 2014 at 08:01:27PM -0700, Hugh Dickins wrote: > On Fri, 6 Jun 2014, Sasha Levin wrote: > > On 06/06/2014 02:49 PM, Kirill A. Shutemov wrote: > > > On Fri, Jun 06, 2014 at 11:26:14AM -0700, Linus Torvalds wrote: > > >> > On Fri, Jun 6, 2014 at 10:43 AM, Dave Jones <davej@xxxxxxxxxx> wrote: > > >>> > > > > >>> > > RIP: 0010:[<ffffffff8b3287b5>] [<ffffffff8b3287b5>] copy_page_rep+0x5/0x10 > > >> > > > >> > Ok, it's the first iteration of "rep movsq" (%rcx is still 0x200) for > > >> > copying a page, and the pages are > > >> > > > >> > RSI: ffff880052766000 > > >> > RDI: ffff880014efe000 > > >> > > > >> > which both look like reasonable kernel addresses. So I'm assuming it's > > >> > DEBUG_PAGEALLOC that makes this trigger, and since the error code is > > >> > 0, and the CR2 value matches RSI, it's the source page that seems to > > >> > have been freed. > > >> > > > >> > And I see absolutely _zero_ reason for wht your 64k mmap_min_addr > > >> > should make any difference what-so-ever. That's just odd. > > >> > > > >> > Anyway, can you try to figure out _which_ copy_user_highpage() it is > > >> > (by looking at what is around the call-site at > > >> > "handle_mm_fault+0x1e0". The fact that we have a stale > > >> > do_huge_pmd_wp_page() on the stack makes me suspect that we have hit > > >> > that VM_FAULT_FALLBACK case and this is related to splitting. Adding a > > >> > few more people explicitly to the cc in case anybody sees anything > > >> > (original email on lkml and linux-mm for context, guys). > > > Looks like a known false positive from DEBUG_PAGEALLOC: > > > > > > https://lkml.org/lkml/2013/3/29/103 > > > > > > We huge copy page in do_huge_pmd_wp_page() without ptl taken and the page > > > can be splitted and freed under us. Once page is copied we take ptl again > > > and recheck that PMD is not changed. If changed, we don't use new page. > > > Not a bug, never triggered with DEBUG_PAGEALLOC disabled. > > > > > > It would be nice to have a way to mark this kind of speculative access. > > > > FWIW, this issue makes fuzzing with DEBUG_PAGEALLOC nearly impossible since > > this thing is so common we never get to do anything "fun" before this issue > > triggers. > > > > A fix would be more than welcome. > > Please give this a try: I think it's right, but I could easily be wrong. > > > [PATCH] thp: fix DEBUG_PAGEALLOC oops in copy_page_rep > > Trinity has for over a year been reporting a CONFIG_DEBUG_PAGEALLOC > oops in copy_page_rep() called from copy_user_huge_page() called from > do_huge_pmd_wp_page(). > > I believe this is a DEBUG_PAGEALLOC false positive, due to the source > page being split, and a tail page freed, while copy is in progress; and > not a problem without DEBUG_PAGEALLOC, since the pmd_same() check will > prevent a miscopy from being made visible. > > Fix by adding get_user_huge_page() and put_user_huge_page(): reducing > to the usual get_page() and put_page() on head page in the usual config; > but get and put references to all of the tail pages when DEBUG_PAGEALLOC. > > Signed-off-by: Hugh Dickins <hughd@xxxxxxxxxx> Ugly, but should do the job: Acked-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx> BTW, we will not need this with new THP refcounting I'm playing with: reference on the page will be enough to protect against splitting. -- Kirill A. Shutemov -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>