On Fri, Nov 25, 2016 at 09:42:07AM +0100, Vlastimil Babka wrote:
> On 11/24/2016 03:23 PM, Dmitry Vyukov wrote:
> > On Thu, Nov 24, 2016 at 2:49 PM, Vlastimil Babka <vbabka@xxxxxxx> wrote:
> >> On 11/18/2016 11:19 AM, Dmitry Vyukov wrote:
> >>>
> >>> Hello,
> >>>
> >>> I've got the following BUG while running syzkaller on
> >>> a25f0944ba9b1d8a6813fd6f1a86f1bd59ac25a6 (4.9-rc5). Unfortunately it's
> >>> not reproducible.
> >>>
> >>> kernel BUG at ./include/linux/mm.h:1743!
> >>> invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC KASAN
> >>
> >>
> >> Shouldn't there be also dump_page() output? Since you've hit this:
> >> VM_BUG_ON_PAGE(page->pmd_huge_pte, page);
> >
> > Here it is:
> >
> > [ 250.326131] page:ffffea0000e196c0 count:1 mapcount:0 mapping:
> > (null) index:0x0
> > [ 250.343393] flags: 0x1fffc0000000000()
> > [ 250.345328] page dumped because: VM_BUG_ON_PAGE(page->pmd_huge_pte)
> > [ 250.346780] ------------[ cut here ]------------
> > [ 250.347742] kernel BUG at ./include/linux/mm.h:1743!
>
> Yeah, as expected, not very useful for this particular BUG_ON :/
>
> >> Anyway the output wouldn't contain the value of pmd_huge_pte or stuff that's
> >> in union with it. I'd suggest adding a local patch that prints this in the
> >> error case, in case the fuzzer hits it again.
> >>
> >> Heck, it might even make sense to print raw contents of struct page in
> >> dump_page() as a catch-all solution? Should I send a patch?
> >
> > Yes, please send.
> > We are moving towards continuous build without local patches.
>
> Something like this?
> -------8<-------
> From 2ac2c9b83d7c4c8be076c24246865a2ed01f9032 Mon Sep 17 00:00:00 2001
> From: Vlastimil Babka <vbabka@xxxxxxx>
> Date: Fri, 25 Nov 2016 09:08:05 +0100
> Subject: [PATCH] mm, debug: print raw struct page data in __dump_page()
>
> The __dump_page() function is used when a page metadata inconsistency is
> detected, either by standard runtime checks, or extra checks in CONFIG_DEBUG_VM
> builds. It prints some of the relevant metadata, but not the whole struct page,
> which is based on unions and interpretation is dependent on the context.
>
> This means that sometimes e.g. a VM_BUG_ON_PAGE() checks certain field, which
> is however not printed by __dump_page() and the resulting bug report may then
> lack clues that could help in determining the root cause. This patch solves
> the problem by simply printing the whole struct page word by word, so no part
> is missing, but the interpretation of the data is left to developers. This is
> similar to e.g. x86_64 raw stack dumps.
>
> Example output:
>
> page:ffffea00000475c0 count:1 mapcount:0 mapping: (null) index:0x0
> flags: 0x100000000000400(reserved)
> raw struct page data:
> 0100000000000400 0000000000000000 0000000000000000 00000001ffffffff
> ffffea00000475e0 ffffea00000475e0 0000000000000000 0000000000000000
> page dumped because: VM_BUG_ON_PAGE(1)
>
> Signed-off-by: Vlastimil Babka <vbabka@xxxxxxx>
> ---
> mm/debug.c | 17 +++++++++++++++++
> 1 file changed, 17 insertions(+)
>
> diff --git a/mm/debug.c b/mm/debug.c
> index 9feb699c5d25..9f67ad74d036 100644
> --- a/mm/debug.c
> +++ b/mm/debug.c
> @@ -48,6 +48,8 @@ void __dump_page(struct page *page, const char *reason)
> * encode own info.
> */
> int mapcount = PageSlab(page) ? 0 : page_mapcount(page);
> + int i;
> + const int words_per_line = (sizeof(unsigned long) == 8) ? 4 : 8;
>
> pr_emerg("page:%p count:%d mapcount:%d mapping:%p index:%#lx",
> page, page_ref_count(page), mapcount,
> @@ -59,6 +61,21 @@ void __dump_page(struct page *page, const char *reason)
>
> pr_emerg("flags: %#lx(%pGp)\n", page->flags, &page->flags);
>
> + pr_alert("raw struct page data:");
Do we really need this line? I would like to keep dump_page() output as
compact as possible.
> + for (i = 0; i < sizeof(struct page) / sizeof(unsigned long); i++) {
> + unsigned long *word_ptr;
> +
> + word_ptr = ((unsigned long *) page) + i;
> +
> + if ((i % words_per_line) == 0) {
> + pr_cont("\n");
> + pr_alert(" %016lx", *word_ptr);
> + } else {
> + pr_cont(" %016lx", *word_ptr);
16 is a waste on 32-bit system. And it will produce too long lines.
Maybe 'unsigned long long' a time?
> + }
> + }
> + pr_cont("\n");
> +
> if (reason)
> pr_alert("page dumped because: %s\n", reason);
>
> --
> 2.10.2
>
>
> --
> 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>
--
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>