Ways to reclaim unused page-table pages

david.hildenbrand — Wed, 11 May 2022 11:00:13 +0000

Hi,

considering cgroups, page tables are charged that way already, towards the total memory consumption of a cgroup. In cgroup v1 there is an additional "kmem" accounting to track+limit such kernel allocations separately, which seems to no longer exist in v2. So in v2 you can only limit the total memory consumption of a cgroup.

In general, the issue is that page tables are unmovable and unswappable. In environments where neither matters (e.g., no swap, no memory hotunplug, no THP, no memory compaction), there is no real difference to movable and swappable user pages.

However, usually we care about the difference, for example, when using swap for memory overcommit or having memory in the system managed by ZONE_MOVABLE for memory hotunplug or memory defragmentation. As page tables cannot go to swap or cannot be placed on ZONE_MOVABLE, malicious processes can trigger OOM or negatively affect the performance of other workloads in setups, simply by primarily consuming primarily unmovable and unswappable memory via page tables.

Ways to reclaim unused page-table pages

nickodell — Mon, 09 May 2022 14:47:06 +0000

There must be something I'm missing here with regard to malicious processes.

>These are all legitimate use cases, and they can be problematic enough; it is also easy to write a malicious program that fills memory with empty page tables and brings the system to a halt, which is rather less legitimate.

Are the page tables used by a process not charged to it in memory accounting? If so, it seems like a malicious process which makes lots of page tables can be dealt with in the same way as a malicious process which allocates lots of memory. If not, why not?

LWN: Comments on "Ways to reclaim unused page-table pages"

Ways to reclaim unused page-table pages

Ways to reclaim unused page-table pages