Anonymous VMA merging improvements WIP
In this post I would like to raise the awareness a bit about an effort to reduce the limitations of anonymous VMA merging, in the form of an ongoing master thesis by Jakub Matena, which I'm supervising. I suspect there might be userspace projects that would benefit and maybe their authors are aware of the current limitations and would welcome if they were relaxed, but they don't read the linux-mm mailing list – the last version of the RFC posted there is here
In a high-level summary, merging of anonymous VMAs in Linux generally happens as soon as they become adjacent in the address space and have compatible access protection bits (and also mempolicies etc.). However due to internal implementation details (involving VMA and page offsets) some operations such as mremap() that moves VMAs around the address space can cause anonymous VMAs not to merge even if everything else is compatible. This is then visible as extra entries in /proc/pid/maps that could be in theory be one larger entry, the associated larger memory and CPU overhead of VMA operations, or even hitting the limit of VMAs per process, set by the vm.max_map_count sysctl. A related issue is that mremap() syscall itself cannot currently process multiple VMAs, so a process that needs to further mremap() the non-merged areas would need to somehow learn the extra boundaries first and perform a sequence of multiple mremap()'s to achieve its goal.
Does any of the above sound familiar because you found that out already while working on a Linux application? Then we would love your feedback on the RFC linked above (or privately). The issue is that while in many scenarios the merging limitations can be lifted by the RFC, it doesn't come for free in both of some overhead of e.g. mremap(), and especially the extra complexity of an already complex code. Thus identifying workloads that would benefit a lot would be helpful. Thanks!