Message-ID: <10a7304e-41c5-434b-b349-c457b7f6b484@arm.com>
Date: Fri, 30 May 2025 16:41:44 +0530
From: Dev Jain <dev.jain@....com>
To: Ryan Roberts <ryan.roberts@....com>, akpm@...ux-foundation.org,
 david@...hat.com, catalin.marinas@....com, will@...nel.org
Cc: lorenzo.stoakes@...cle.com, Liam.Howlett@...cle.com, vbabka@...e.cz,
 rppt@...nel.org, surenb@...gle.com, mhocko@...e.com, linux-mm@...ck.org,
 linux-kernel@...r.kernel.org, suzuki.poulose@....com, steven.price@....com,
 gshan@...hat.com, linux-arm-kernel@...ts.infradead.org
Subject: Re: [PATCH 0/3] Enable huge-vmalloc permission change


On 30/05/25 4:21 pm, Ryan Roberts wrote:
> On 30/05/2025 11:42, Dev Jain wrote:
>> On 30/05/25 4:07 pm, Ryan Roberts wrote:
>>> On 30/05/2025 11:10, Dev Jain wrote:
>>>> On 30/05/25 3:33 pm, Ryan Roberts wrote:
>>>>> On 30/05/2025 10:04, Dev Jain wrote:
>>>>>> This series paves the path to enable huge mappings in vmalloc space by
>>>>>> default on arm64. > For this we must ensure that we can handle any permission
>>>>>> games on vmalloc space.
>>>>> And the linear map :)
>>>>>
>>>>>> Currently, __change_memory_common() uses
>>>>>> apply_to_page_range() which does not support changing permissions for
>>>>>> leaf mappings.
>>>>> nit: A "leaf mapping" is the lowest level entry in the page tables for a given
>>>>> address - i.e. it maps an address to some actual memory rather than to another
>>>>> pgtable. It includes what the Arm ARM calls "page mappings" (PTE level) and
>>>>> "block mappings" (PMD/PUD/.. level). apply_to_page_range() does support page
>>>>> mappings, so saying it doesn't support leaf mappings is incorrect. It doesn't
>>>>> support block mappings.
>>>> Sorry, again got confused by nomenclature : )
>>>>
>>>>>> We attempt to move away from this by using walk_page_range_novma(),
>>>>>> similar to what riscv does right now; however, it is the responsibility
>>>>>> of the caller to ensure that we do not pass a range, or split the range
>>>>>> covering a partial leaf mapping.
>>>>>>
>>>>>> This series is tied with Yang Shi's attempt [1] at using huge mappings
>>>>>> in the linear mapping in case the system supports BBML2, in which case
>>>>>> we will be able to split the linear mapping if needed without break-before-
>>>>>> make.
>>>>>> Thus, Yang's series, IIUC, will be one such user of my series; suppose we
>>>>>> are changing permissions on a range of the linear map backed by PMD-hugepages,
>>>>>> then the sequence of operations should look like the following:
>>>>>>
>>>>>> split_range(start, (start + HPAGE_PMD_SIZE) & ~HPAGE_PMD_MASK);
>>>>>> split_range(end & ~HPAGE_PMD_MASK, end);
>>>>> I don't understand what the HPAGE_PMD_MASK twiddling is doing? That's not
>>>>> right.
>>>>> It's going to give you the offset within the 2M region. You just want:
>>>>>
>>>>> split_range(start)
>>>>> split_range(end)
>>>>>
>>>>> right?
>>>> Suppose start = 2M + 4K, end = 8M + 5K. Then my sequence will compute to
>>> 8M + 5K is not a valid split point. It has to be at least page aligned.
>> Sorry, so consider 8M + 4K.
>>
>>>> split_range(2M + 4K, 3M)
>>>> split_range(8M, 8M + 5K)
>>> We just want to split at start and end. What are the 3M and 8M params supposed
>>> to be? Anyway, this is off-topic for this series.
>> I think we are both saying the same thing; yes we will split only the start and
>> the end,
>> so if the address 2Mb + 4Kb is mapped as a PMD-hugepage, we need to split this
>> PMD into
>> a PTE table, which will map 2Mb till 4Mb as base pages now.
> Not quite; if you start with [2M, 4M) mapped by PMD, then want to ensure a
> boundary at 2M+4K, I think you would end up with the first 64K (16 pages) mapped
> by base page, then the then the next 1984K mapped by contpte blocks (31 contpte
> blocks).
>
> But yes, we agree :)

Yes, I wasn't considering the contpte case : )

>
>>>> __change_memory_common(2M + 4K, 8M + 5K)
>>>>
>>>> So now __change_memory_common() wouldn't have to deal with splitting the
>>>> starts and ends. Please correct me if I am wrong.
>>>>
>>>>>> __change_memory_common(start, end);
>>>>>>
>>>>>> However, this series can be used independently of Yang's; since currently
>>>>>> permission games are being played only on pte mappings (due to
>>>>>> apply_to_page_range
>>>>>> not supporting otherwise), this series provides the mechanism for enabling
>>>>>> huge mappings for various kernel mappings like linear map and vmalloc.
>>>>> In other words, you are saying that this series is a prerequisite for Yang's
>>>>> series (and both are prerequisites for huge vmalloc by default). Your series
>>>>> adds a new capability that Yang's series will rely on (the ability to change
>>>>> permissions on block mappings).
>>>> That's right.
>>>>
>>>>> Thanks,
>>>>> Ryan
>>>>>
>>>>>> [1] https://lore.kernel.org/all/20250304222018.615808-1-
>>>>>> yang@...amperecomputing.com/
>>>>>>
>>>>>> Dev Jain (3):
>>>>>>      mm: Allow pagewalk without locks
>>>>>>      arm64: pageattr: Use walk_page_range_novma() to change memory
>>>>>>        permissions
>>>>>>      mm/pagewalk: Add pre/post_pte_table callback for lazy MMU on arm64
>>>>>>
>>>>>>     arch/arm64/mm/pageattr.c | 81 +++++++++++++++++++++++++++++++++++++---
>>>>>>     include/linux/pagewalk.h |  4 ++
>>>>>>     mm/pagewalk.c            | 18 +++++++--
>>>>>>     3 files changed, 94 insertions(+), 9 deletions(-)
>>>>>>

Powered by blists - more mailing lists