[LSF/MM/PPF TOPIC] PKS for the page cache and beyond

[LSF/MM/PPF TOPIC] PKS for the page cache and beyond

[Ira Weiny]

Hello,

Protection Key Supervisor (PKS) presents a way to control access to a large
domain of memory quickly, without a page table walk or TLB flush, as well as
with finer granularity; allowing protection control on individual threads.

Multiple areas of memory have been identified as candidates to be protected
with PKS.  These include the initial use case persistent memory (PMEM), page
tables[1], kernel secret keys[2], and the page cache.[3]  Like PMEM the page
cache presents a significant surface area where stray writes, or other bugs,
could corrupt data permanently.

I would like to discuss the ramifications of being able to change memory
permissions in this new way.  While PKS has a lot to offer it does not come for
free.  One trade off is the loss of direct access via page_address() in
!HIGHMEM builds.

Already PMEM's faced challenges in the leverage of kmap/kunmap.  While the page
cache should be able to leverage this work, this is driving a redefinition of
what kmap means.  Especially since the HIGHMEM use case is increasingly
meaningless on modern machines.

Ira Weiny

[1] https://lore.kernel.org/lkml/20210830235927.6443-2-rick.p.edgecombe@intel.com/
[2] https://lore.kernel.org/lkml/20201009201410.3209180-3-ira.weiny@intel.com/
[3] https://lwn.net/Articles/883352/

[LSF/MM/BPF TOPIC] PKS for the page cache and beyond

[Ira Weiny]

Resend due to typo in tag. s/PPF/BPF/

However, I also realized one should fill out the google form now.  I'll do that
straight away.

Thanks,
Ira

On Fri, Feb 25, 2022 at 08:10:30AM -0800, Ira Weiny wrote:
> Hello,
> 
> Protection Key Supervisor (PKS) presents a way to control access to a large
> domain of memory quickly, without a page table walk or TLB flush, as well as
> with finer granularity; allowing protection control on individual threads.
> 
> Multiple areas of memory have been identified as candidates to be protected
> with PKS.  These include the initial use case persistent memory (PMEM), page
> tables[1], kernel secret keys[2], and the page cache.[3]  Like PMEM the page
> cache presents a significant surface area where stray writes, or other bugs,
> could corrupt data permanently.
> 
> I would like to discuss the ramifications of being able to change memory
> permissions in this new way.  While PKS has a lot to offer it does not come for
> free.  One trade off is the loss of direct access via page_address() in
> !HIGHMEM builds.
> 
> Already PMEM's faced challenges in the leverage of kmap/kunmap.  While the page
> cache should be able to leverage this work, this is driving a redefinition of
> what kmap means.  Especially since the HIGHMEM use case is increasingly
> meaningless on modern machines.
> 
> Ira Weiny
> 
> [1] https://lore.kernel.org/lkml/20210830235927.6443-2-rick.p.edgecombe@intel.com/
> [2] https://lore.kernel.org/lkml/20201009201410.3209180-3-ira.weiny@intel.com/
> [3] https://lwn.net/Articles/883352/