Re: uninitialized pmem struct pages

[David Hildenbrand]

On 05.01.21 10:05, Michal Hocko wrote:
> On Tue 05-01-21 00:57:43, Dan Williams wrote:
>> On Tue, Jan 5, 2021 at 12:42 AM Michal Hocko <mhocko@xxxxxxxx> wrote:
>>>
>>> On Tue 05-01-21 00:27:34, Dan Williams wrote:
>>>> On Tue, Jan 5, 2021 at 12:17 AM Michal Hocko <mhocko@xxxxxxxx> wrote:
>>>>>
>>>>> On Tue 05-01-21 09:01:00, Michal Hocko wrote:
>>>>>> On Mon 04-01-21 16:44:52, David Hildenbrand wrote:
>>>>>>> On 04.01.21 16:43, David Hildenbrand wrote:
>>>>>>>> On 04.01.21 16:33, Michal Hocko wrote:
>>>>>>>>> On Mon 04-01-21 16:15:23, David Hildenbrand wrote:
>>>>>>>>>> On 04.01.21 16:10, Michal Hocko wrote:
>>>>>>>>> [...]
>>>>>>>>>> Do the physical addresses you see fall into the same section as boot
>>>>>>>>>> memory? Or what's around these addresses?
>>>>>>>>>
>>>>>>>>> Yes I am getting a garbage for the first struct page belonging to the
>>>>>>>>> pmem section [1]
>>>>>>>>> [    0.020161] ACPI: SRAT: Node 0 PXM 0 [mem 0x100000000-0x603fffffff]
>>>>>>>>> [    0.020163] ACPI: SRAT: Node 4 PXM 4 [mem 0x6060000000-0x11d5fffffff] non-volatile
>>>>>>>>>
>>>>>>>>> The pfn without the initialized struct page is 0x6060000. This is a
>>>>>>>>> first pfn in a section.
>>>>>>>>
>>>>>>>> Okay, so we're not dealing with the "early section" mess I described,
>>>>>>>> different story.
>>>>>>>>
>>>>>>>> Due to [1], is_mem_section_removable() called
>>>>>>>> pfn_to_page(PHYS_PFN(0x6060000)). page_zone(page) made it crash, as not
>>>>>>>> initialized.
>>>>>>>>
>>>>>>>> Let's assume this is indeed a reserved pfn in the altmap. What's the
>>>>>>>> actual address of the memmap?
>>>>>>>>
>>>>>>>> I do wonder what hosts pfn_to_page(PHYS_PFN(0x6060000)) - is it actually
>>>>>>>> part of the actual altmap (i.e. > 0x6060000) or maybe even self-hosted?
>>>>>>>>
>>>>>>>> If it's not self-hosted, initializing the relevant memmaps should work
>>>>>>>> just fine I guess. Otherwise things get more complicated.
>>>>>>>
>>>>>>> Oh, I forgot: pfn_to_online_page() should at least in your example make
>>>>>>> sure other pfn walkers are safe. It was just an issue of
>>>>>>> is_mem_section_removable().
>>>>>>
>>>>>> Hmm, I suspect you are right. I haven't put this together, thanks! The memory
>>>>>> section is indeed marked offline so pfn_to_online_page would indeed bail
>>>>>> out:
>>>>>> crash> p (0x6060000>>15)
>>>>>> $3 = 3084
>>>>>> crash> p mem_section[3084/128][3084 & 127]
>>>>>> $4 = {
>>>>>>   section_mem_map = 18446736128020054019,
>>>>>>   usage = 0xffff902dcf956680,
>>>>>>   page_ext = 0x0,
>>>>>>   pad = 0
>>>>>> }
>>>>>> crash> p 18446736128020054019 & (1UL<<2)
>>>>>> $5 = 0
>>>>>>
>>>>>> That makes it considerably less of a problem than I thought!
>>>>>
>>>>> Forgot to add that those who are running kernels without 53cdc1cb29e8
>>>>> ("drivers/base/memory.c: indicate all memory blocks as removable") for
>>>>> some reason can fix the crash by the following simple patch.
>>>>>
>>>>> Index: linux-5.3-users_mhocko_SLE15-SP2_for-next/drivers/base/memory.c
>>>>> ===================================================================
>>>>> --- linux-5.3-users_mhocko_SLE15-SP2_for-next.orig/drivers/base/memory.c
>>>>> +++ linux-5.3-users_mhocko_SLE15-SP2_for-next/drivers/base/memory.c
>>>>> @@ -152,9 +152,14 @@ static ssize_t removable_show(struct dev
>>>>>                 goto out;
>>>>>
>>>>>         for (i = 0; i < sections_per_block; i++) {
>>>>> -               if (!present_section_nr(mem->start_section_nr + i))
>>>>> +               unsigned long nr = mem->start_section_nr + i;
>>>>> +               if (!present_section_nr(nr))
>>>>>                         continue;
>>>>> -               pfn = section_nr_to_pfn(mem->start_section_nr + i);
>>>>> +               if (!online_section_nr()) {
>>>>
>>>> I assume that's onlince_section_nr(nr) in the version that compiles?
>>>
>>> Yup.
>>>
>>>> This makes sense because the memory block size is larger than the
>>>> section size. I suspect you have 1GB memory block size on this system,
>>>> but since the System RAM and PMEM collide at a 512MB alignment in a
>>>> memory block you end up walking the back end of the last 512MB of the
>>>> System RAM memory block and run into the offline PMEM section.
>>>
>>> Sections are 128MB and memory blocks are 2GB on this system.
>>>
>>>> So, I don't think it's pfn_to_online_page that necessarily needs to
>>>> know how to disambiguate each page, it's things that walk sections and
>>>> memory blocks and expects them to be consistent over the span.
>>>
>>> Well, memory hotplug code is hard wired to sparse memory model so in
>>> this particular case asking about the section is ok. But pfn walkers
>>> shouldn't really care and only rely on pfn_to_online_page. But that will
>>> do the right thing here. So we are good as long as the section is marked
>>> properly. But this would become a problem as soon as the uninitialized
>>> pages where sharing the same memory section as David pointed out.
>>> pfn_to_online_page would then return something containing garbage. So we
>>> should still think of a way to either initialize all those pages or make
>>> sure pfn_to_online_page recognizes them. The former is preferred IMHO.
>>
>> The former would not have saved the crash in this case because
>> pfn_to_online_page() is not used in v5.3:removable_show() that I can
>> see, nor some of the other paths that might walk pfns and the wrong
>> thing with ZONE_DEVICE.
> 
> If the page was initialized properly, and by that I mean also have it
> reserved, then the old code would have properly reported is as not
> removable.
> 
>> However, I do think pfn_to_online_page() should be reliable, and I
>> prefer to just brute force add a section flag to indicate whether the
>> section might be ZONE_DEVICE polluted and fallback to the
>> get_dev_pagemap() slow-path in that case.
> 
> Do we have some spare room to hold that flag in a section?
> 
>> ...but it would still require hunting to find the places where
>> pfn_to_online_page() is missing for assumptions like this crash which
>> assumed memblock-online + section-present == section-online.
> 
> Yes, but most users should be using pfn_to_online_page already.
> 

Quite honestly, let's not hack around this issue and just fix it
properly - make pfn_to_online_page() only ever return an initialized,
online (buddy) page, just as documented.

What speaks against not adding early sections in case they have a
relevant hole at the end, besides wasting some MB? Relevant setups most
probably just don't care. Print a warning.

-- 
Thanks,

David / dhildenb