Re: [PATCH v3 01/12] mm: dump_page(): better diagnostics for compound pages

[John Hubbard]

On 2/3/20 5:16 AM, Kirill A. Shutemov wrote:
> On Fri, Jan 31, 2020 at 07:40:18PM -0800, John Hubbard wrote:
>> A compound page collects the refcount in the head page, while leaving
>> the refcount of each tail page at zero. Therefore, when debugging a
>> problem that involves compound pages, it's best to have diagnostics that
>> reflect that situation. However, dump_page() is oblivious to these
>> points.
>>
>> Change dump_page() as follows:
>>
>> 1) For tail pages, print relevant head page information: refcount, in
>>    particular. But only do this if the page is not corrupted so badly
>>    that the pointer to the head page is all wrong.
>>
>> 2) Do a separate check to catch any (rare) cases of the tail page's
>>    refcount being non-zero, and issue a separate, clear pr_warn() if
>>    that ever happens.
>>
>> Suggested-by: Matthew Wilcox <willy@xxxxxxxxxxxxx>
>> Suggested-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
>> Signed-off-by: John Hubbard <jhubbard@xxxxxxxxxx>
> 
> Acked-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>

Thanks for looking through all of these!

> 
> Few nit-picks below.
> 
>> ---
>>  mm/debug.c | 34 ++++++++++++++++++++++++++++------
>>  1 file changed, 28 insertions(+), 6 deletions(-)
>>
>> diff --git a/mm/debug.c b/mm/debug.c
>> index ecccd9f17801..beb1c59d784b 100644
>> --- a/mm/debug.c
>> +++ b/mm/debug.c
>> @@ -42,6 +42,32 @@ const struct trace_print_flags vmaflag_names[] = {
>>  	{0, NULL}
>>  };
>>  
>> +static void __dump_tail_page(struct page *page, int mapcount)
>> +{
>> +	struct page *head = compound_head(page);
>> +
>> +	if ((page < head) || (page >= head + MAX_ORDER_NR_PAGES)) {
> 
> I'm not sure if we want to use compound_nr() here instead of
> MAX_ORDER_NR_PAGES. Do you have any reasonaing about it?


Yes: compound_nr(page) reads from the struct page, whereas MAX_ORDER_NR_PAGES
is an independent, immutable limit. When checking a struct page for corruption,
it's ideal to avoid relying on data within the struct page, as compound_nr()
would have to do.


> 
>> +		/*
>> +		 * Page is hopelessly corrupted, so limit any reporting to
>> +		 * information about the page itself. Do not attempt to look at
>> +		 * the head page.
>> +		 */
>> +		pr_warn("page:%px refcount:%d mapcount:%d mapping:%px "
>> +			"index:%#lx (corrupted tail page case)\n",
>> +			page, page_ref_count(page), mapcount, page->mapping,
>> +			page_to_pgoff(page));
>> +	} else {
>> +		pr_warn("page:%px compound refcount:%d mapcount:%d mapping:%px "
>> +			"index:%#lx compound_mapcount:%d\n",
>> +			page, page_ref_count(head), mapcount, head->mapping,
>> +			page_to_pgoff(head), compound_mapcount(page));
>> +	}
>> +
>> +	if (page_ref_count(page) != 0)
>> +		pr_warn("page:%px PROBLEM: non-zero refcount (==%d) on this "
>> +			"tail page\n", page, page_ref_count(page));
> 
> Wrap into {}, please.


Fixed, thanks.


thanks,
-- 
John Hubbard
NVIDIA