On 9/8/2014 3:55 AM, Steve Capper wrote:
> On Thu, Sep 04, 2014 at 10:41:52AM +0100, Catalin Marinas wrote:
>> Hi Zhichang,
>>
>> (cc'ing Steve Capper for the huge page stuff)
>>
>> On Fri, Aug 22, 2014 at 01:38:26PM +0100, zhichang.yuan wrote:
>>> I am working to implement the DEBUG_PAGEALLOC on ARMv8.
>>
>> I assume that's the arm64 kernel.
>>
>>> After i investigated the DEBUG_PAGEALLOC implementation on x86 arch,
>>> some questions are standing in the way to start coding.
>>>
>>> 1. How to handle the large page when DEBUG_PAGEALLOC is enabled In
>>> ARMv8, the kernel direct memory page table entries will set the block
>>> flag for better performance. When DEBUG_PAGEALLOC is configured, if
>>> the size of freed page is not multiply of page block size, there is no
>>> corresponding page table entry. In the old x86 kernel version, the
>>> large page to be freed will be split into normal page size and build
>>> the corresponding PTEs. And afterwards, someone done a patch to remove
>>> the splitting process. It will make the code simpler and easily
>>> stable.
>>
>> Initially, you could either map everything as pages or implement
>> splitting of huge pages (if for example the huge page is at the pmd
>> level, you allocate and populate a pte).
>>
>>> I prefer the current design in x86, what are your thoughts here?
>>
>> I haven't looked at it yet.
>>
>>> 2. Does ARMv8 support HIBERNATION?
>>
>> Not yet.
>>
>>> The HIBERNATION has some dependency on DEBUG_PAGEALLOC.
>>
>> Like in DEBUG_PAGEALLOC "depends on !HIBERNATION"?
>>
>>> 3. Is the hypothesis of DEBUG_PAGEALLOC always true?
>>
>> Which hypothesis?
>>
>>> From the x86 code, DEBUG_PAGEALLOC use the invalid page table entries
>>> to catch the accesses to free pages. This mechanism is based on the
>>> hypothesis that all the corresponding page table entries that are
>>> corresponding to the free pages are cleared correctly. Supposed this
>>> condition is always true, what we need to do is just to clear the
>>> kernel linear mapping page entries, since those page tables are
>>> fixable after initialization. DEBUG_PAGEALLOC on x86 seems to do like
>>> that.
>>
>> I guess that's the ARCH_SUPPORTS_DEBUG_PAGEALLOC rather than just the
>> simple DEBUG_PAGEALLOC which can be enabled on arm64 as well, you just
>> get page poisoning rather than invalid mappings.
>>
>> It could be done on arm64 as well but you need to sort out huge page
>> splitting or just map everything as pages when the option is enabled.
>>
>
> (cc'ing Laura Abbott for info...)
>
> Hi,
> There is support for splitting pmd's and pud's in the direct kernel
> mapping in the following series from Laura Abbott:
>
> "[PATCHv3 7/7] arm64: add better page protections to arm64"
> http://lists.infradead.org/pipermail/linux-arm-kernel/2014-August/280782.html
>
> Perhaps some of the splitting logic there could be used by the
> kernel_map_pages arm64 implementation for ARCH_SUPPORTS_DEBUG_PAGEALLOC?
>
> Cheers,
>
The page splitting was originally written for a out of tree implementation
of something similar to ARCH_SUPPORTS_DEBUG_PAGEALLOC for both arm and
arm64 so yes it could be used. The approach taken was
- map all memory with sections initially
- walk all memblock and remap as 4K pages
There is a performance hit involved but for some issues the benefits certainly
outweigh the costs (One person described it as 'the best feature since
CONFIG_SLUB_DEBUG').
If there is interest, I can clean up the patches and submit them as a proof
of concept. The approach probably won't cover 64K/THP but it might be a
starting point.
Thanks,
Laura
--
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
hosted by The Linux Foundation
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
