linux-kernel.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: David Hildenbrand <david@redhat.com>
To: Baoquan He <bhe@redhat.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>,
	mgorman@suse.de, linux-kernel@vger.kernel.org,
	linux-mm@kvack.org, akpm@linux-foundation.org, cai@lca.pw,
	mhocko@kernel.org, steve.wahl@hpe.com
Subject: Re: [PATCH] mm/compaction: Fix the incorrect hole in fast_isolate_freepages()
Date: Thu, 28 May 2020 11:08:38 +0200	[thread overview]
Message-ID: <78919796-0c26-35f0-55fa-305932b7f992@redhat.com> (raw)
In-Reply-To: <20200528090731.GI20045@MiWiFi-R3L-srv>

On 28.05.20 11:07, Baoquan He wrote:
> On 05/26/20 at 01:49pm, David Hildenbrand wrote:
>> On 26.05.20 13:32, Mike Rapoport wrote:
>>> Hello Baoquan,
>>>
>>> On Tue, May 26, 2020 at 04:45:43PM +0800, Baoquan He wrote:
>>>> On 05/22/20 at 05:20pm, Mike Rapoport wrote:
>>>>> Hello Baoquan,
>>>>>
>>>>> On Fri, May 22, 2020 at 03:25:24PM +0800, Baoquan He wrote:
>>>>>> On 05/22/20 at 03:01pm, Baoquan He wrote:
>>>>>>>
>>>>>>> So let's add these unavailable ranges into memblock and reserve them
>>>>>>> in init_unavailable_range() instead. With this change, they will be added
>>>>>>> into appropriate node and zone in memmap_init(), and initialized in
>>>>>>> reserve_bootmem_region() just like any other memblock reserved regions.
>>>>>>
>>>>>> Seems this is not right. They can't get nid in init_unavailable_range().
>>>>>> Adding e820 ranges may let them get nid. But the hole range won't be
>>>>>> added to memblock, and still has the issue.
>>>>>>
>>>>>> Nack this one for now, still considering.
>>>>>
>>>>> Why won't we add  the e820 reserved ranges to memblock.memory during
>>>>> early boot as I suggested?
>>>>>
>>>>> diff --git a/arch/x86/kernel/e820.c b/arch/x86/kernel/e820.c
>>>>> index c5399e80c59c..b0940c618ed9 100644
>>>>> --- a/arch/x86/kernel/e820.c
>>>>> +++ b/arch/x86/kernel/e820.c
>>>>> @@ -1301,8 +1301,11 @@ void __init e820__memblock_setup(void)
>>>>>  		if (end != (resource_size_t)end)
>>>>>  			continue;
>>>>>  
>>>>> -		if (entry->type == E820_TYPE_SOFT_RESERVED)
>>>>> +		if (entry->type == E820_TYPE_SOFT_RESERVED ||
>>>>> +		    entry->type == E820_TYPE_RESERVED) {
>>>>> +			memblock_add(entry->addr, entry->size);
>>>>>  			memblock_reserve(entry->addr, entry->size);
>>>>> +		}
>>>>>  
>>>>>  		if (entry->type != E820_TYPE_RAM && entry->type != E820_TYPE_RESERVED_KERN)
>>>>>  			continue;
>>>>>
>>>>> The setting of node later  in numa_init() will assign the proper node
>>>>> for these regions as it does for the usable memory.
>>>>
>>>> Yes, if it's only related to e820 reserved region, this truly works.
>>>>
>>>> However, it also has ACPI table regions. That's why I changed to call
>>>> the problematic area as firmware reserved ranges later.
>>>>
>>>> Bisides, you can see below line, there's another reserved region which only
>>>> occupies one page in one memory seciton. If adding to memblock.memory, we also
>>>> will build struct mem_section and the relevant struct pages for the whole
>>>> section. And then the holes around that page will be added and initialized in
>>>> init_unavailable_mem(). numa_init() will assign proper node for memblock.memory
>>>> and memblock.reserved, but won't assign proper node for the holes.
>>>>
>>>> ~~~
>>>> [    0.000000] BIOS-e820: [mem 0x00000000fed80000-0x00000000fed80fff] reserved
>>>> ~~~
>>>>
>>>> So I still think we should not add firmware reserved range into
>>>> memblock for fixing this issue.
>>>>
>>>> And, the fix in the original patch seems necessary. You can see in
>>>> compaction code, the migration source is chosen from LRU pages or
>>>> movable pages, the migration target has to be got from Buddy. However,
>>>> only the min_pfn in fast_isolate_freepages(), it's calculated by
>>>> distance between cc->free_pfn - cc->migrate_pfn, we can't guarantee it's
>>>> safe, then use it as the target to handle.
>>>
>>> I do not object to your original fix with careful check for pfn validity.
>>>
>>> But I still think that the memory reserved by the firmware is still
>>> memory and it should be added to memblock.memory. This way the memory
>>
>> If it's really memory that could be read/written, I think I agree.
> 
> I would say some of them may not be allowed to be read/written, if I
> understand it correctly. I roughly went through the x86 init code, there
> are some places where mem region is marked as E820_TYPE_RESERVED so that
> they are not touched after initialization. E.g:
> 
> 1) pfn 0
> In trim_bios_range(), we set the pfn 0 as E820_TYPE_RESERVED. You can
> see the code comment, this is a BIOS owned area, but not kernel RAM.
> 
> 2)GART reserved region
> In early_gart_iommu_check(), GART IOMMU firmware will reserve a region
> in an area, firmware designer won't map system RAM into that area.
> 
> And also intel_graphics_stolen(), arch_rmrr_sanity_check(), these
> regions are not system RAM backed area, reading from or writting into
> these area may cause error.
> 
> Futhermore, there's a KASLR bug found by HPE, its triggering and root
> cause are written into below commit log. You can see that accessing to
> firmware reserved region caused BIOS to halt system when cpu doing
> speculative.
> 
> commit 2aa85f246c181b1fa89f27e8e20c5636426be624
> Author: Steve Wahl <steve.wahl@hpe.com>
> Date:   Tue Sep 24 16:03:55 2019 -0500
> 
>     x86/boot/64: Make level2_kernel_pgt pages invalid outside kernel area
> 
>     Our hardware (UV aka Superdome Flex) has address ranges marked
>     reserved by the BIOS. Access to these ranges is caught as an error,
>     causing the BIOS to halt the system.
> 
>>
>>> map will be properly initialized from the very beginning and we won't
>>> need init_unavailable_mem() and alike workarounds and. Obviously, the patch
>>
>> I remember init_unavailable_mem() is necessary for holes within
>> sections, where we actually *don't* have memory, but we still have have
>> a valid memmap (full section) that we have to initialize.
>>
>> See the example from 4b094b7851bf ("mm/page_alloc.c: initialize memmap
>> of unavailable memory directly"). Our main memory ends within a section,
>> so we have to initialize the remaining parts because the whole section
>> will be marked valid/online.
> 
> Yes, memory hole need be handled in init_unavailable_mem(). Since we
> have created struct page for them, need initialize them. We can't
> discard init_unavailable_mem() for now.
> 
>>
>> Any way to improve this handling is appreciated. In that patch I also
>> spelled out that we might want to mark such holes via a new page type,
>> e.g., PageHole(). Such a page is a memory hole, but has a valid memmap.
>> Any content in the memmap (zone/node) should be ignored.
> 
> As I said at above, I am a little conservative to add all those regions of
> E820_TYPE_RESERVED into memblock.memory and memblock.reserved, because
> most of them are firmware reserved region, they may be not backed by normal
> RAM.
> 
> I was thinking to step back to use mm_zero_struct_page() inside
> init_unavailable_range() as below. But it doesn't differ much
> from __init_single_page(), except of the _refcount and mapcount.
> Zeroing struct page equals to putting them into node 0, zero 0.
> 
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 3973b5fdfe3f..4e4b72cf5283 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -6901,7 +6901,7 @@ static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
>                  * (in memblock.reserved but not in memblock.memory) will
>                  * get re-initialized via reserve_bootmem_region() later.
>                  */
> -               __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
> +               mm_zero_struct_page(pfn_to_page(pfn));
>                 __SetPageReserved(pfn_to_page(pfn));
>                 pgcnt++;
>         }
> 
> About adding these unavailable ranges into node/zone, in the old code,
> it just happened to add them into expected node/zone. You can see in
> early_pfn_in_nid(), if no nid found from memblock, the returned '-1'
> will make it true ironically. But that is not saying the bad thing
> always got good result. If the last zone of node 0 is DMA32 zone, the
> deferred init will skip the only chance to add some of unavailable
> rnages into expected node/zone. Means they were not always added into
> appropriate node/zone before, the change of iterating memblock.memory in
> memmap_init() dones't introduce regression.
> 
> static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
> {
>         int nid;
> 
>         nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
>         if (nid >= 0 && nid != node)
>                 return false;
>         return true;
> }
> 
> So if no anybody need access them after boot, not adding them into any
> node/zone sounds better. Otherwise, better add them in the appropriate
> node/zone.

Yes, the node/zone is just completely irrelevant for these pages I'd say.

As I said, maybe we can flag these memmaps somehow as "while this is an
initialized memmap, the node/zone is garbage and this memmap should just
be ignored completely in any kind of node/zone aware code".


-- 
Thanks,

David / dhildenb


  reply	other threads:[~2020-05-28  9:08 UTC|newest]

Thread overview: 21+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-05-21  1:44 Baoquan He
2020-05-21  9:26 ` Mike Rapoport
2020-05-21 15:52   ` Baoquan He
2020-05-21 17:18     ` Mike Rapoport
2020-05-22  7:01       ` Baoquan He
2020-05-22  7:25         ` Baoquan He
2020-05-22 14:20           ` Mike Rapoport
2020-05-26  8:45             ` Baoquan He
2020-05-26  8:55               ` David Hildenbrand
2020-05-26 11:32               ` Mike Rapoport
2020-05-26 11:49                 ` David Hildenbrand
2020-05-28  9:07                   ` Baoquan He
2020-05-28  9:08                     ` David Hildenbrand [this message]
2020-05-28 15:15                     ` Steve Wahl
2020-06-01 11:42                       ` Mike Rapoport
2020-06-01 13:31                         ` Baoquan He
2020-05-21  9:36 ` Mel Gorman
2020-05-21 13:38   ` Mike Rapoport
2020-05-21 15:41   ` Baoquan He
2020-05-28  8:59 [PATCH] mm/compaction: Fix the incorrect hole in fast_isolate_freepages() ^[ Baoquan He
2020-05-28  9:08 ` Baoquan He

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=78919796-0c26-35f0-55fa-305932b7f992@redhat.com \
    --to=david@redhat.com \
    --cc=akpm@linux-foundation.org \
    --cc=bhe@redhat.com \
    --cc=cai@lca.pw \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-mm@kvack.org \
    --cc=mgorman@suse.de \
    --cc=mhocko@kernel.org \
    --cc=rppt@linux.ibm.com \
    --cc=steve.wahl@hpe.com \
    --subject='Re: [PATCH] mm/compaction: Fix the incorrect hole in fast_isolate_freepages()' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).