Jiri Slaby 6b3564
From: Eric Biggers <ebiggers@google.com>
Jiri Slaby 6b3564
Date: Fri, 25 Aug 2017 15:55:39 -0700
Jiri Slaby 6b3564
Subject: [PATCH] mm/madvise.c: fix freeing of locked page with MADV_FREE
Jiri Slaby 6b3564
References: bnc#1060662
Jiri Slaby 6b3564
Patch-mainline: 4.12.10
Jiri Slaby 6b3564
Git-commit: 263630e8d176d87308481ebdcd78ef9426739c6b
Jiri Slaby 6b3564
Jiri Slaby 6b3564
commit 263630e8d176d87308481ebdcd78ef9426739c6b upstream.
Jiri Slaby 6b3564
Jiri Slaby 6b3564
If madvise(..., MADV_FREE) split a transparent hugepage, it called
Jiri Slaby 6b3564
put_page() before unlock_page().
Jiri Slaby 6b3564
Jiri Slaby 6b3564
This was wrong because put_page() can free the page, e.g. if a
Jiri Slaby 6b3564
concurrent madvise(..., MADV_DONTNEED) has removed it from the memory
Jiri Slaby 6b3564
mapping. put_page() then rightfully complained about freeing a locked
Jiri Slaby 6b3564
page.
Jiri Slaby 6b3564
Jiri Slaby 6b3564
Fix this by moving the unlock_page() before put_page().
Jiri Slaby 6b3564
Jiri Slaby 6b3564
This bug was found by syzkaller, which encountered the following splat:
Jiri Slaby 6b3564
Jiri Slaby 6b3564
    BUG: Bad page state in process syzkaller412798  pfn:1bd800
Jiri Slaby 6b3564
    page:ffffea0006f60000 count:0 mapcount:0 mapping:          (null) index:0x20a00
Jiri Slaby 6b3564
    flags: 0x200000000040019(locked|uptodate|dirty|swapbacked)
Jiri Slaby 6b3564
    raw: 0200000000040019 0000000000000000 0000000000020a00 00000000ffffffff
Jiri Slaby 6b3564
    raw: ffffea0006f60020 ffffea0006f60020 0000000000000000 0000000000000000
Jiri Slaby 6b3564
    page dumped because: PAGE_FLAGS_CHECK_AT_FREE flag(s) set
Jiri Slaby 6b3564
    bad because of flags: 0x1(locked)
Jiri Slaby 6b3564
    Modules linked in:
Jiri Slaby 6b3564
    CPU: 1 PID: 3037 Comm: syzkaller412798 Not tainted 4.13.0-rc5+ #35
Jiri Slaby 6b3564
    Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
Jiri Slaby 6b3564
    Call Trace:
Jiri Slaby 6b3564
     __dump_stack lib/dump_stack.c:16 [inline]
Jiri Slaby 6b3564
     dump_stack+0x194/0x257 lib/dump_stack.c:52
Jiri Slaby 6b3564
     bad_page+0x230/0x2b0 mm/page_alloc.c:565
Jiri Slaby 6b3564
     free_pages_check_bad+0x1f0/0x2e0 mm/page_alloc.c:943
Jiri Slaby 6b3564
     free_pages_check mm/page_alloc.c:952 [inline]
Jiri Slaby 6b3564
     free_pages_prepare mm/page_alloc.c:1043 [inline]
Jiri Slaby 6b3564
     free_pcp_prepare mm/page_alloc.c:1068 [inline]
Jiri Slaby 6b3564
     free_hot_cold_page+0x8cf/0x12b0 mm/page_alloc.c:2584
Jiri Slaby 6b3564
     __put_single_page mm/swap.c:79 [inline]
Jiri Slaby 6b3564
     __put_page+0xfb/0x160 mm/swap.c:113
Jiri Slaby 6b3564
     put_page include/linux/mm.h:814 [inline]
Jiri Slaby 6b3564
     madvise_free_pte_range+0x137a/0x1ec0 mm/madvise.c:371
Jiri Slaby 6b3564
     walk_pmd_range mm/pagewalk.c:50 [inline]
Jiri Slaby 6b3564
     walk_pud_range mm/pagewalk.c:108 [inline]
Jiri Slaby 6b3564
     walk_p4d_range mm/pagewalk.c:134 [inline]
Jiri Slaby 6b3564
     walk_pgd_range mm/pagewalk.c:160 [inline]
Jiri Slaby 6b3564
     __walk_page_range+0xc3a/0x1450 mm/pagewalk.c:249
Jiri Slaby 6b3564
     walk_page_range+0x200/0x470 mm/pagewalk.c:326
Jiri Slaby 6b3564
     madvise_free_page_range.isra.9+0x17d/0x230 mm/madvise.c:444
Jiri Slaby 6b3564
     madvise_free_single_vma+0x353/0x580 mm/madvise.c:471
Jiri Slaby 6b3564
     madvise_dontneed_free mm/madvise.c:555 [inline]
Jiri Slaby 6b3564
     madvise_vma mm/madvise.c:664 [inline]
Jiri Slaby 6b3564
     SYSC_madvise mm/madvise.c:832 [inline]
Jiri Slaby 6b3564
     SyS_madvise+0x7d3/0x13c0 mm/madvise.c:760
Jiri Slaby 6b3564
     entry_SYSCALL_64_fastpath+0x1f/0xbe
Jiri Slaby 6b3564
Jiri Slaby 6b3564
Here is a C reproducer:
Jiri Slaby 6b3564
Jiri Slaby 6b3564
    #define _GNU_SOURCE
Jiri Slaby 6b3564
    #include <pthread.h>
Jiri Slaby 6b3564
    #include <sys/mman.h>
Jiri Slaby 6b3564
    #include <unistd.h>
Jiri Slaby 6b3564
Jiri Slaby 6b3564
    #define MADV_FREE	8
Jiri Slaby 6b3564
    #define PAGE_SIZE	4096
Jiri Slaby 6b3564
Jiri Slaby 6b3564
    static void *mapping;
Jiri Slaby 6b3564
    static const size_t mapping_size = 0x1000000;
Jiri Slaby 6b3564
Jiri Slaby 6b3564
    static void *madvise_thrproc(void *arg)
Jiri Slaby 6b3564
    {
Jiri Slaby 6b3564
        madvise(mapping, mapping_size, (long)arg);
Jiri Slaby 6b3564
    }
Jiri Slaby 6b3564
Jiri Slaby 6b3564
    int main(void)
Jiri Slaby 6b3564
    {
Jiri Slaby 6b3564
        pthread_t t[2];
Jiri Slaby 6b3564
Jiri Slaby 6b3564
        for (;;) {
Jiri Slaby 6b3564
            mapping = mmap(NULL, mapping_size, PROT_WRITE,
Jiri Slaby 6b3564
                           MAP_POPULATE|MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
Jiri Slaby 6b3564
Jiri Slaby 6b3564
            munmap(mapping + mapping_size / 2, PAGE_SIZE);
Jiri Slaby 6b3564
Jiri Slaby 6b3564
            pthread_create(&t[0], 0, madvise_thrproc, (void*)MADV_DONTNEED);
Jiri Slaby 6b3564
            pthread_create(&t[1], 0, madvise_thrproc, (void*)MADV_FREE);
Jiri Slaby 6b3564
            pthread_join(t[0], NULL);
Jiri Slaby 6b3564
            pthread_join(t[1], NULL);
Jiri Slaby 6b3564
            munmap(mapping, mapping_size);
Jiri Slaby 6b3564
        }
Jiri Slaby 6b3564
    }
Jiri Slaby 6b3564
Jiri Slaby 6b3564
Note: to see the splat, CONFIG_TRANSPARENT_HUGEPAGE=y and
Jiri Slaby 6b3564
CONFIG_DEBUG_VM=y are needed.
Jiri Slaby 6b3564
Jiri Slaby 6b3564
Google Bug Id: 64696096
Jiri Slaby 6b3564
Jiri Slaby 6b3564
Link: http://lkml.kernel.org/r/20170823205235.132061-1-ebiggers3@gmail.com
Jiri Slaby 6b3564
Fixes: 854e9ed09ded ("mm: support madvise(MADV_FREE)")
Jiri Slaby 6b3564
Signed-off-by: Eric Biggers <ebiggers@google.com>
Jiri Slaby 6b3564
Acked-by: David Rientjes <rientjes@google.com>
Jiri Slaby 6b3564
Acked-by: Minchan Kim <minchan@kernel.org>
Jiri Slaby 6b3564
Acked-by: Michal Hocko <mhocko@suse.com>
Jiri Slaby 6b3564
Cc: Dmitry Vyukov <dvyukov@google.com>
Jiri Slaby 6b3564
Cc: Hugh Dickins <hughd@google.com>
Jiri Slaby 6b3564
Cc: Andrea Arcangeli <aarcange@redhat.com>
Jiri Slaby 6b3564
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Jiri Slaby 6b3564
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Jiri Slaby 6b3564
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Jiri Slaby 6b3564
Signed-off-by: Jiri Slaby <jslaby@suse.cz>
Jiri Slaby 6b3564
---
Jiri Slaby 6b3564
 mm/madvise.c | 2 +-
Jiri Slaby 6b3564
 1 file changed, 1 insertion(+), 1 deletion(-)
Jiri Slaby 6b3564
Jiri Slaby 6b3564
diff --git a/mm/madvise.c b/mm/madvise.c
Jiri Slaby 6b3564
index 75d2cffbe61d..fc6bfbe19a16 100644
Jiri Slaby 6b3564
--- a/mm/madvise.c
Jiri Slaby 6b3564
+++ b/mm/madvise.c
Jiri Slaby 6b3564
@@ -368,8 +368,8 @@ static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
Jiri Slaby 6b3564
 				pte_offset_map_lock(mm, pmd, addr, &ptl);
Jiri Slaby 6b3564
 				goto out;
Jiri Slaby 6b3564
 			}
Jiri Slaby 6b3564
-			put_page(page);
Jiri Slaby 6b3564
 			unlock_page(page);
Jiri Slaby 6b3564
+			put_page(page);
Jiri Slaby 6b3564
 			pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
Jiri Slaby 6b3564
 			pte--;
Jiri Slaby 6b3564
 			addr -= PAGE_SIZE;
Jiri Slaby 6b3564
-- 
Jiri Slaby 6b3564
2.14.2
Jiri Slaby 6b3564