the primary problem with the unix architecture (other than random legacy cruft) is that there is no good way for programs to provide services to eachother. d-bus, wayland, x11, all rely on sockets and shared memory to communicate, sacrificing structure to allow these programs to exist outside the kernel.
kernel services have much easier, structured communications with userspace through filesystem-mapped devices. the /sys subsystem, even most of the ancient protocols of /dev, all provide elegant and discoverable interfaces to kernel services that userspace can barely mimic.
all of this encourages an ever-ballooning kernel size. the worst part imo isn't even that all this random shit like networking is running in kernel space, it's that the kernel and all the kernel modules are just such massive components now that they're difficult to work on without prior experience in the codebase.
now FUSE is definitely a step in the right direction, but it is unequivocally the pinnacle of legacy cruft. even the simplest readdir calls require messing with e.g. hardlink reference counting to maintain correctness in edge cases like e.g. using a rename() call to move a file between different mountpoints.
on the other hand, the solution to this has already been demonstrated: plan 9 from bell labs solves the crux of the issue (although still not entirely escaping the legacy cruft from back when it was just designed to export a unix filesystem). it still has the issue of drivers being merged into kernel space, but this seems to be out of inertia rather than out of necessity. and it doesn't seem to take much of a performance hit from serialising and deserialising 9p messages into a stream for communicating with the server program (since userspace file servers are just a managing program reading from and writing to a stream that is mounted into a kernel-managed 9p device file).
i have a kernel + os project (that i'm not yet ready to share, but soon) where i'm trying to fix this by shrinking the kernel to just a standard microkernel + 9p-like multiplexer. and i am not under the impression that this will ever be used for anything.
yousharknotpass
2087
gracieabrams
origamiaround