Re: [PATCH 2/2] arm64/mm: fix incorrect file_map_count for invalid pmd/pud

From: Anshuman Khandual <anshuman.khandual@arm.com>
To: Mark Rutland <mark.rutland@arm.com>,
	David Hildenbrand <david@redhat.com>
Cc: Liu Shixin <liushixin2@huawei.com>,
	Catalin Marinas <catalin.marinas@arm.com>,
	Will Deacon <will@kernel.org>,
	Denys Vlasenko <dvlasenk@redhat.com>,
	Kefeng Wang <wangkefeng.wang@huawei.com>,
	David Hildenbrand <dhildenb@redhat.com>,
	Rafael Aquini <raquini@redhat.com>,
	Pasha Tatashin <pasha.tatashin@soleen.com>,
	linux-arm-kernel@lists.infradead.org,
	linux-kernel@vger.kernel.org
Subject: Re: [PATCH 2/2] arm64/mm: fix incorrect file_map_count for invalid pmd/pud
Date: Thu, 17 Nov 2022 09:54:06 +0530	[thread overview]
Message-ID: <53c0c955-7afe-905b-468a-cd7ac81238c5@arm.com> (raw)
In-Reply-To: <Y3UF87Z/a7UVThVF@FVFF77S0Q05N.cambridge.arm.com>

On 11/16/22 21:16, Mark Rutland wrote:
> On Wed, Nov 16, 2022 at 10:08:27AM +0100, David Hildenbrand wrote:
>> On 16.11.22 09:38, Liu Shixin wrote:
>>> The page table check trigger BUG_ON() unexpectedly when split hugepage:
>>>
>>>   ------------[ cut here ]------------
>>>   kernel BUG at mm/page_table_check.c:119!
>>>   Internal error: Oops - BUG: 00000000f2000800 [#1] SMP
>>>   Dumping ftrace buffer:
>>>      (ftrace buffer empty)
>>>   Modules linked in:
>>>   CPU: 7 PID: 210 Comm: transhuge-stres Not tainted 6.1.0-rc3+ #748
>>>   Hardware name: linux,dummy-virt (DT)
>>>   pstate: 20000005 (nzCv daif -PAN -UAO -TCO -DIT -SSBS BTYPE=--)
>>>   pc : page_table_check_set.isra.0+0x398/0x468
>>>   lr : page_table_check_set.isra.0+0x1c0/0x468
>>> [...]
>>>   Call trace:
>>>    page_table_check_set.isra.0+0x398/0x468
>>>    __page_table_check_pte_set+0x160/0x1c0
>>>    __split_huge_pmd_locked+0x900/0x1648
>>>    __split_huge_pmd+0x28c/0x3b8
>>>    unmap_page_range+0x428/0x858
>>>    unmap_single_vma+0xf4/0x1c8
>>>    zap_page_range+0x2b0/0x410
>>>    madvise_vma_behavior+0xc44/0xe78
>>>    do_madvise+0x280/0x698
>>>    __arm64_sys_madvise+0x90/0xe8
>>>    invoke_syscall.constprop.0+0xdc/0x1d8
>>>    do_el0_svc+0xf4/0x3f8
>>>    el0_svc+0x58/0x120
>>>    el0t_64_sync_handler+0xb8/0xc0
>>>    el0t_64_sync+0x19c/0x1a0
>>> [...]
>>>
>>> On arm64, pmd_present() will return true even if the pmd is invalid.
>>
>> I assume that's because of the pmd_present_invalid() check.
>>
>> ... I wonder why that behavior was chosen. Sounds error-prone to me.
> 
> That seems to be down to commit:
> 
>   b65399f6111b03df ("arm64/mm: Change THP helpers to comply with generic MM semantics")
> 
> ... apparently because Andrea Arcangelli said this was necessary in:
> 
>   https://lore.kernel.org/lkml/20181017020930.GN30832@redhat.com/
> 
> ... but that does see to contradict what's said in:
> 
>   Documentation/mm/arch_pgtable_helpers.rst
> 
> ... which just says:
> 
>   pmd_present  Tests a valid mapped PMD

It should be as follows instead, will update. Not sure about PUD level though,
where anon THP is not supported (AFAIK).

+---------------------------+--------------------------------------------------+
| pmd_present               | Tests if pmd_page() points to valid memory page  |
+---------------------------+--------------------------------------------------+

> 
> ... and it's not clear to me why this *only* applies to the PMD level.
> 
> Anshuman?

Because THP is supported at PMD level. As Andrea had explained earlier, pmd_present()
should return positive if pmd_page() on the entry points to valid memory irrespective
of whether the entry is valid/mapped or not. That is the semantics expected in generic
THP during PMD split, collapse, migration etc and other memory code walking past such
PMD entries. That was my understanding.