Re: [PATCH] hv_balloon: avoid touching uninitialized struct page during tail onlining

From: David Hildenbrand
Date: Tue Jan 08 2019 - 04:16:47 EST


On 07.01.19 14:44, Vitaly Kuznetsov wrote:
> David Hildenbrand <david@xxxxxxxxxx> writes:
>
>> On 04.01.19 15:19, Vitaly Kuznetsov wrote:
>>> Hyper-V memory hotplug protocol has 2M granularity and in Linux x86 we use
>>> 128M. To deal with it we implement partial section onlining by registering
>>> custom page onlining callback (hv_online_page()). Later, when more memory
>>> arrives we try to online the 'tail' (see hv_bring_pgs_online()).
>>>
>>> It was found that in some cases this 'tail' onlining causes issues:
>>>
>>> BUG: Bad page state in process kworker/0:2 pfn:109e3a
>>> page:ffffe08344278e80 count:0 mapcount:1 mapping:0000000000000000 index:0x0
>>> flags: 0xfffff80000000()
>>> raw: 000fffff80000000 dead000000000100 dead000000000200 0000000000000000
>>> raw: 0000000000000000 0000000000000000 0000000000000000 0000000000000000
>>> page dumped because: nonzero mapcount
>>> ...
>>> Workqueue: events hot_add_req [hv_balloon]
>>> Call Trace:
>>> dump_stack+0x5c/0x80
>>> bad_page.cold.112+0x7f/0xb2
>>> free_pcppages_bulk+0x4b8/0x690
>>> free_unref_page+0x54/0x70
>>> hv_page_online_one+0x5c/0x80 [hv_balloon]
>>> hot_add_req.cold.24+0x182/0x835 [hv_balloon]
>>> ...
>>>
>>> Turns out that we now have deferred struct page initialization for memory
>>> hotplug so e.g. memory_block_action() in drivers/base/memory.c does
>>> pages_correctly_probed() check and in that check it avoids inspecting
>>> struct pages and checks sections instead. But in Hyper-V balloon driver we
>>> do PageReserved(pfn_to_page()) check and this is now wrong.
>>>
>>> Switch to checking online_section_nr() instead.
>>>
>>> Signed-off-by: Vitaly Kuznetsov <vkuznets@xxxxxxxxxx>
>>> ---
>>> drivers/hv/hv_balloon.c | 10 ++++++----
>>> 1 file changed, 6 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/drivers/hv/hv_balloon.c b/drivers/hv/hv_balloon.c
>>> index 5301fef16c31..7c6349a50ef1 100644
>>> --- a/drivers/hv/hv_balloon.c
>>> +++ b/drivers/hv/hv_balloon.c
>>> @@ -888,12 +888,14 @@ static unsigned long handle_pg_range(unsigned long pg_start,
>>> pfn_cnt -= pgs_ol;
>>> /*
>>> * Check if the corresponding memory block is already
>>> - * online by checking its last previously backed page.
>>> - * In case it is we need to bring rest (which was not
>>> - * backed previously) online too.
>>> + * online. It is possible to observe struct pages still
>>> + * being uninitialized here so check section instead.
>>> + * In case the section is online we need to bring the
>>> + * rest of pfns (which were not backed previously)
>>> + * online too.
>>> */
>>> if (start_pfn > has->start_pfn &&
>>> - !PageReserved(pfn_to_page(start_pfn - 1)))
>>> + online_section_nr(pfn_to_section_nr(start_pfn)))
>>> hv_bring_pgs_online(has, start_pfn, pgs_ol);
>>>
>>> }
>>>
>>
>> I wonder if you should use pfn_to_online_page() and check for PageOffline().
>>
>> (I guess online_section_nr() should also do the trick)
>
> I'm worried a bit about racing with mm code here as we're not doing
> mem_hotplug_begin()/done() so I'd slightly prefer keeping
> online_section_nr() (pfn_to_online_page() also uses it but then it gets
> to the particular struct page). Moreover, with pfn_to_online_page() we
> will be looking at some other pfn - because the start_pfn is definitelly
> offline (pre-patch we were looking at start_pfn-1). Just looking at the
> whole section seems cleaner.

Fine with me. I guess the section can never be offlined as it still
contains reserved pages if not fully "fake-onlined" by HV code already.

But we could still consider mem_hotplug_begin()/done() as we could have
a online section although online_pages() has not completed yet. So we
could actually touch some "semi onlined section".

Acked-by: David Hildenbrand <david@xxxxxxxxxx>

>
> P.S. I still think about bringing mem_hotplug_begin()/done() to
> hv_balloon but that's going to be a separate discussion, here I want to
> have a small fix backportable to stable.
>
> Thanks,
>


--

Thanks,

David / dhildenb