Optimizing the pre-dump algorithm

During pre-dump in CRIU, the memory pages of target process are stored in pipe-pages, until the content is not flushed to disk image or page-server. Primary issue is bloat of irreclaimable memory due to pipe-pages. Pipe buffers are pinned in memory making them non swappable. Pipes have maximum size restriction, so there could be many pipes which results in memory pressure during pre-dump process. Replacing pipe-pages with userspace supplied buffer from CRIU will alleviate this memory pressure. Since, the buffer pages can be swapped out.

Another issue is duration for which pre-dump algorithm freezes the target task. Use case like live migration expects smallest possible freezing glitch. So, objective is to reduce this freeze time during pre-dump.

To achieve both of these objectives, we use process_vm_readv system call with a set of VMA-list gathered by freezing target task for minimal amount of time. There are challenges in handling memory areas with this approach, like race conditions resulting in false reads. Development of graceful solution for these challenges is objective of this project.