Recently, Pike commented that he designed Plan 9 so he could build the windowing system he wanted. (edit: oops, link) https://9fans.topicbox.com/groups/9fans/T4fec171fec7e3845-M2...
I do know that in Plan9, 'everything is a file' is literally correct. So owning a mouse would mean reading from /path/to/mouse. But what I've also read is that every process can do so believing they own it - it's theirs exclusively?
For anyone reading this comment in future please do not trust it, I don't know what I'm talking about, except that standard references I've found don't fully explain.
[0] https://en.wikipedia.org/wiki/Plan_9_from_Bell_Labs#Graphica...
Almost correct. There are also shared memory segments via segment(3). You segattach(2) to a segment from a user space process then read/write that memory as a file from a user space process [1]. Drivers can expose memory mapped IO (e.g. PCI cards) using a physical segment and segattach to that from a user space process. The vga(8) program does this to do low level register tweaking. Sometimes, a file server is either too tedious, or not the right abstraction. It also enables memory sharing between processes and user space programs. You can share it over the network with rexport(1) but it gets weird (networked PCI/RAM out of the box...)
> So owning a mouse would mean reading from /path/to/mouse. But what I've also read is that every process can do so believing they own it - it's theirs exclusively?
Rio and other window managers like Lola, multiplex it [2]. This is how each window only sees the mouse when the cursor is within its border window. If I open two windows and run 'cat /dev/mouse' in each window then move the mouse around, cat only outputs data when the mouse is over the window it's running in.
1. https://man.9front.org/3/segment 2. https://man.9front.org/3/mouse (2nd to last paragraph)
Windows are directories in the filesystem. Their properties are files inside it. Modify the file, the window changes.
This means that to open a window on another machine is trivial. Machine A mounts the relevant part of the filesystem on Machine B over the kernel's integrated 9p networking stack, then creates and populates a folder, and now the window is on the other machine.
No "servers" or "clients", no special network protocols, works locally or across the planet, according to appropriate permissions of course. Definitely no need for Windows-like crude hacks such as VNC or RDP.
I think, though, that one of the vastly underestimated aspects of Plan 9/9front is its simplified C.
AIUI one of the biggest differences is that a file is not allowed to `#include` anything that has already been #included in a parent project. You are not allowed to nest #includes.
The result is vastly reduced compile times. I have read claims of over 1000x fewer lines of code going through the compiler.
(Sure, disk caching helps a lot, but it still has to go through CPU.)
Cleaning up a codebase is a massive PITA but it's a one-time job and I have read comments that once you learn this, it quickly becomes habit.
Imagine if that was put into C29 as an advisory.
Decades and centuries of compute-time saved.
This is standard in Go, which is of course Pike's more recent baby.
ANSI C let you get away with this, which sounds like one of the more epic mistakes:
Plan 9 C is wonderful to work in. I really enjoy its simplicity and how the system design has trimmed away a lot of the ugly API's (select and poll, thank god they're gone...) and replaced them by modern multi-process solutions like thread(2) and ioproc(2). Plan 9 was designed ground up for multi-processor machines and it's all channels (if Go were an OS...)
> Decades and centuries of compute-time saved.
Rebuild all of 9front user space from source: 'cd /sys/src; mk clean; time mk install' - 51.48u 36.14s 51.23r mk install. almost 51.5 seconds, under a minute on a measly quad core i5-6500T CPU @ 2.50GHz - a Dell micro desktop with a SATA SSD.
9front amd64 PC kernel build: 'cd /sys/src/9/pc64; mk clean; time mk' - 5.38u 1.16s 4.19r mk. FIVE FUCKING SECONDS. Same machine as above. Time to build the Arm64 Pi kernel: 4.66u 0.83s 3.99r mk. About one second faster. Edit, I should add the kernel sizes: 9pc64 - 5,412,046 bytes, 9pi3 - 4,443,400 bytes. About a second per megabyte of executable machine code generated.
Rob Pike also acknowledged that leaving GC out of Alef and going back to C was a mistake, speaking of which the Plan 9 C compilers continued to be their own thing instead of blindly following the new language standard.
"Abandon All Hope, Ye Who Enter Here"