This is an experimental version of XPMEM based on a version provided by Cray and uploaded to https://code.google.com/p/xpmem. This version supports any Linux kernel 3.12 and newer (tested up to 5.8.x). Keep in mind there may be bugs and this version may cause kernel panics, code crashes, eat your cat, etc.
XPMEM is a Linux kernel module that enables a process to map the memory of another process into its virtual address space. Source code can be obtained by cloning the Git repository, original Mercurial repository or by downloading a tarball from the link above.
The XPMEM API has three main functions:
xpmem_make()
xpmem_get()
xpmem_attach()
A process calls xpmem_make() to export a region of its virtual address space. Other processes can then attach to the region by calling xpmem_get() and xpmem_attach(). After a memory region is attached, it is accessed via direct loads and stores. This enables upper-level protocols such as MPI and SHMEM to perform single-copy address-space to address-space transfers, completely at user-level.
Note, there is a limitation to the usage of an attached region. Any system call that will call get_user_pages() on the region from the non-owning process with get EFAULT. This include pthread mutexes and condition variable, and SYS V semaphores. We intend to address this limitation in a future release.
XPMEM regions are free to have "holes" in them, meaning virtual memory regions that are not allocated. This makes XPMEM somewhat more flexible than mmap(). A process could, for example, export a region via XPMEM starting at address 0 and extending 4 GB. Accesses to allocated (valid) virtual addresses in this region proceed normally, and pages are mapped between address spaces on demand. A segfault will occur if the source process or any other process mapping the region tries to access an unallocated (invalid) virtual address in the region.