Regarding the problem of writes to mmaped files being problematic to scan whilst keeping them functional -- would it be a reasonable solution to let each file-mapped VMA to get COW pages, and defer scanning and committing the results to the filesystem until the mapping is released?
This would let most uses of writeable mapped files work unimpeded.
But it does raise the question, are there any applications that need to use a disk-file-backed shared memory area (as opposed to non-disk-backed objects as provided by shmfs or tmpfs) for interprocess communication?
Posted Feb 12, 2009 13:07 UTC (Thu) by nix (subscriber, #2304)
[Link]
glibc nscd, for one. (That's how it can be so damn fast: nscd maintains the mmap()ed area, and every user of nscd just needs to yank stuff straight out of the cache, with no IPC as such at all.)