Re: [RFC PATCH 1/2] mm, mincore2(): retrieve dax and tlb-size attributes of an address range

From: Dan Williams
Date: Mon Sep 12 2016 - 13:25:32 EST


On Sun, Sep 11, 2016 at 11:29 PM, Oliver O'Halloran <oohall@xxxxxxxxx> wrote:
> On Mon, Sep 12, 2016 at 3:31 AM, Dan Williams <dan.j.williams@xxxxxxxxx> wrote:
>> As evidenced by this bug report [1], userspace libraries are interested
>> in whether a mapping is DAX mapped, i.e. no intervening page cache.
>> Rather than using the ambiguous VM_MIXEDMAP flag in smaps, provide an
>> explicit "is dax" indication as a new flag in the page vector populated
>> by mincore.
>>
>> There are also cases, particularly for testing and validating a
>> configuration to know the hardware mapping geometry of the pages in a
>> given process address range. Consider filesystem-dax where a
>> configuration needs to take care to align partitions and block
>> allocations before huge page mappings might be used, or
>> anonymous-transparent-huge-pages where a process is opportunistically
>> assigned large pages. mincore2() allows these configurations to be
>> surveyed and validated.
>>
>> The implementation takes advantage of the unused bits in the per-page
>> byte returned for each PAGE_SIZE extent of a given address range. The
>> new format of each vector byte is:
>>
>> (TLB_SHIFT - PAGE_SHIFT) << 2 | vma_is_dax() << 1 | page_present
>
> What is userspace expected to do with the information in vec? Whether
> PMD or THP mappings can be used is going to depend more on the block
> allocations done by the filesystem rather than anything the an
> application can directly influence. Returning a vector for each page
> makes some sense in the mincore() case since the application can touch
> each page to fault them in, but I don't see what they can do here.

It's not a "can huge pages be used?" question it's interrogating the
mapping that got established after the fact. If an
application/environment expects huge mappings, but pte mappings are
getting established

> Why not just get rid of vec entirely and make mincore2() a yes/no
> check over the range for whatever is supplied in flags? That would
> work for NVML's use case and it should be easier to extend if needed.

I think having a way to ask the kernel if an address range satisfies a
certain set of input attributes is a useful interface. Perhaps a
"MINCORE_CHECK" flag can indicate that the input vector contains a
single character that it wants the kernel to validate during the page
table walk, and return zero or the offset of the first mismatch.