Re: [RFC v4][PATCH part-2 00/13] ASI - Part II (Decorated Page-Table)

[Alexandre Chartre <alexandre.chartre@oracle.com>]

On 5/14/20 11:29 AM, Mike Rapoport wrote:
> Hello Alexandre,
> 
> On Mon, May 04, 2020 at 04:57:57PM +0200, Alexandre Chartre wrote:
>> This is part II of ASI RFC v4. Please refer to the cover letter of
>> part I for an overview the ASI RFC.
>>
>>    https://lore.kernel.org/lkml/20200504144939.11318-1-alexandre.chartre@oracle.com/
>>
>> This part introduces decorated page-table which encapsulate native page
>> table (e.g. a PGD) in order to provide convenient page-table management
>> functions, such as tracking address range mapped in a page-table or
>> safely handling references to another page-table.
>>
>> Decorated page-table can then be used to easily create and manage page
>> tables to be used with ASI. It will be used by the ASI test driver (see
>> part III) and later by KVM ASI.
>>
>> Decorated page-table is independent of ASI, and can potentially be used
>> anywhere a page-table is needed.

Hi Mike,

> This is very impressive work!
> 
> I wonder why did you decide to make dpt x86-specific? Unless I've missed
> simething, the dpt implementation does not rely on anything architecture
> specific and can go straight to linux/mm.

Correct, this is not x86 specific. I put it in arch/x86 because that's currently
the only place were I use it, but it can be moved to linux/mm.

> Another thing that comes to mind is that we already have a very
> decorated page table, which is mm_struct.

mm_struct doesn't define a generic page-table encapsulation. mm_struct references
a page table (i.e. PGD) and adds all kind of attributes needed for mm management but
not necessarily related to page-table.

> I admit that my attempt to
> split out the core page table bits from the mm_struct [1] didn't  went
> far, but I still think we need a first class abstraction for the page
> table that will be used by both user memory management and the
> management of the reduced kernel address spaces.

Agree. I remember your attempt to extract the page-table from mm_struct; this is
not a simple work! For ASI, I didn't need mm, so it was simpler to built a simple
decorated page-table without attempting to use with mm (at least for now).

Thanks,

alex.

PS: if you want to play with dpt, there's a bug in dpt_destroy(), patch 08 adds a
a double free of dpt->backend_pages pages.

> 
> [1] https://git.kernel.org/pub/scm/linux/kernel/git/rppt/linux.git/log/?h=pg_table/v0.0
> 
>> Thanks,
>>
>> alex.
>>
>> -----
>>
>> Alexandre Chartre (13):
>>    mm/x86: Introduce decorated page-table (dpt)
>>    mm/dpt: Track buffers allocated for a decorated page-table
>>    mm/dpt: Add decorated page-table entry offset functions
>>    mm/dpt: Add decorated page-table entry allocation functions
>>    mm/dpt: Add decorated page-table entry set functions
>>    mm/dpt: Functions to populate a decorated page-table from a VA range
>>    mm/dpt: Helper functions to map module into a decorated page-table
>>    mm/dpt: Keep track of VA ranges mapped in a decorated page-table
>>    mm/dpt: Functions to clear decorated page-table entries for a VA range
>>    mm/dpt: Function to copy page-table entries for percpu buffer
>>    mm/dpt: Add decorated page-table remap function
>>    mm/dpt: Handle decorated page-table mapped range leaks and overlaps
>>    mm/asi: Function to init decorated page-table with ASI core mappings
>>
>>   arch/x86/include/asm/asi.h |    2 +
>>   arch/x86/include/asm/dpt.h |   89 +++
>>   arch/x86/mm/Makefile       |    2 +-
>>   arch/x86/mm/asi.c          |   57 ++
>>   arch/x86/mm/dpt.c          | 1051 ++++++++++++++++++++++++++++++++++++
>>   5 files changed, 1200 insertions(+), 1 deletion(-)
>>   create mode 100644 arch/x86/include/asm/dpt.h
>>   create mode 100644 arch/x86/mm/dpt.c
>>
>> -- 
>> 2.18.2
>>
>