thingfish

For starters, I needed to put the code of my kernel emulation (libposix-6.1.0.dll) at a specific memory address in order to get that snappy feeling when you start an OpenBSD program. Normally, 32bit DLLs in Windows are loaded at 0x10000000 (called the "ImageBase"). This causes the Windows kernel reposition the DLL because in most cases, other DLLs are already loaded there. In my case these are netapi32.dll, ws2_32.dll, iphlpapi.dll and shell32.dll among others. Making my DLLs load at an address just below the 0x10000000 boundary makes MinC programs load faster, in effect creating a space for my kernel, e.g. "kernel space".

Secondly, I needed some mechanism to reset the contents of the CPU's registers after calling into the various Windows DLLs which are not part of the Windows NT kernel. For instance, the WriteFile() function changes the %edx register, which is used by the GCC compiler to contain the high part of 64bit variables. Turns out I needed what is called a "context switch", saving the contents of all registers onto the stack, calling my kernel emulation and then restoring the registers from the stack. In Linux, this mechanism is triggered by a special interrupt (int 0x80). I decided to name the function that does this in my kernel emulation "__kernel", so in assembly I can write "call ___kernel" [0].

There is more, but I won't go into that right now. I plan to do a write-up on the topic. By the way, Microsoft also has a user space implementation of their kernel, called kernel32.dll, implementing Win32 APIs in user space.

They expose a dark secret behind the fork() call and I felt it too when implementing it myself. Almost gave me a heart failure. So here's my simplified take: what if the parent did a malloc() and put the resulting pointer on the stack (which is what git does, BTW). A simple copy of the stack to the child wouldn't be sufficient. The kernel then would have to follow the pointer, malloc() new memory and copy the data. It's not hard to see where this is going. What if there is malloc()ed memory in this copy? It's madness. I suspect this could bring a whole system to its knees.

This is a problem a kernel should not try to solve. Only user space knows about its application of memory. While reading the source code of the software I include in MinC, I always thought this is bad programming, perpetuated by Torvalds' #1 rule "don't mess with user-space", leading to things like copy-on-write.

This all leads me to believe I can get away with doing flat copies of stack and globals during fork(), implement spawn(), never implement copy-on-write and patch the userland code if needed. Am I right?

Thank you. I needed that. I started the project in 2016 and have been obsessed with it since then. The weird thing is that I never had to use workarounds or hacks that are mentioned by the Cygwin team [0]. For instance, the select() call does map cleanly on top of the Win32 API. I just needed to use WSAWaitForMultipleEvents() instead of WaitForMultipleEvents() (the other "easteregg"). Why the Cygwin people didn't figure this out baffles me. I guess their current code base doesn't allow the rewrite. My big breakthrough was when I realized that the "inconsistent interfaces" [1] in Win32 file handles can be implemented as virtual file systems. One for each handle type (char, disk, pipe, etc). That was my "throwing away 1000 lines of code" [2] moment.

As to the weird file names, I use the file names OpenBSD uses. My rule is to always use the file name of the header (.h) file where the system call is declared in OpenBSD. I also use their struct and constant names, prefixed with "WIN_".

The "fork is evil" thing is discussed a lot in the programmers community. I myself find it quite clever. Threads are highly volatile and are very hard to program without running into race conditions. The solution is to make a copy of everything the child will be using: duplicate file descriptors, the stack, globals (rss). The kernel does all this for you in one system call. I often wonder how the people who complain about the absence of real concurrency in their programming languages [3] actually would use this feature. In my opinion the best way to use concurrency is to string individual programs into a pipeline. This will never go "evil" on you.