When compiled with gcc 4.6.1, 32-bit Sortix would triple fault during
early boot: When the TLB is being flushed, somehow a garbage value had
sneaked into Sortix::Memory::currentdir, and a non-page aligned (and
garbage) page directory is loaded. (Triple fault, here we come!)
However, adding a volatile addr_t foo after the currentdir variable
actually caused the system to boot correctly - the garbage was written
into that variable instead. To debug the problem, I set the foo value
to 0: as long as !foo (hence the name nofoo) everything was alright.
After closer examination I found that the initrd open code wrote to a
pointer supplied by kernel.cpp. The element pointed to was on the
stack. Worse, its address was the same as currentdir (now foo).
Indeed, the stack had gone into the kernel's data segment!
Turns out that this gcc configuration stores variables in the data
segment in the reverse order they are defined in, whereas previous
compilers did the opposite. The hack used to set up the stack during
early boot relied on this (now obviously incorrect) fact.
In effect, the stack was initialized to the end of the stack, not
the start of it: completely ignoring all the nice stack space
allocated in kernel.cpp.
I did not see that one coming.
This lets the kernel use any memory not directly used by it or the
init ramdisk. Although, now we test whether the kernel fits into
the identitymapped area. It can't really grow down there, unless it
wants to collide with user-space. Instead, modules and the like
(when they are invented), should be put in the upper memory. Or in
their own user-space process, yay, microkernel!
This fixes issues where it did not fit into the first few MiB,
or that GRUB loaded it someplace weird.
The kernel heap is now also protected against growing into the
ramdisk and the kernel stack.
This commit got completely out of control.
Added the fork(), getpid(), getppid(), sleep(), usleep() system calls, and
aliases in the Maxsi:: namespace.
Fixed a bug where zero-byte allocation would fail.
Worked on the DescriptorTable class which now works and can fork.
Got rid of some massive print-registers statements and replaced them with
the portable InterruptRegisters::LogRegisters() function.
Removed the SysExecuteOld function and replaced it with Process::Execute().
Rewrote the boot sequence in kernel.cpp such that it now loads the system
idle process 'idle' as PID 0, and the initization process 'init' as PID 1.
Rewrote the SIGINT hack.
Processes now maintain a family-tree structure and keep track of their
threads. PIDs are now allocated using a simple hack. Virtual memory
per-process can now be allocated using a simple hack. Processes can now be
forked. Fixed the Process::Execute function such that it now resets the
stack pointer to where the stack actually is - not just a magic value.
Removed the old and ugly Process::_endcodesection hack.
Rewrote the scheduler into a much cleaner and faster version. Debug code is
now moved to designated functions. The noop kernel-thread has been replaced
by a simple user-space infinite-loop program 'idle'.
The Thread class has been seperated from the Scheduler except in Scheduler-
related code. Thread::{Save,Load}Registers has been improved and has been
moved to $(CPU)/thread.cpp. Threads can now be forked. A new CreateThread
function creates threads properly and portably.
Added a MicrosecondsSinceBoot() function.
Fixed a crucial bug in MemoryManagement::Fork().
Added an 'idle' user-space program that is a noop infinite loop, which is
used by the scheduler when there is nothing to do.
Rewrote the 'init' program such that it now forks off a shell, instead of
becoming the shell.
Added the $$ (current PID) and $PPID (parent PPID) variables to the shell.
1) The PML2 was not initialized to zeroes, thus leaving some bits behind that
caused the fork code to go crazy, forking the unforkable, and mapping addresses
that never, ever, should have been mapped, leaving behind a trail of page faults
and general protection faults on some computers, while other computers worked
because the uninitalized memory just wasn't uninitialized enough. Yep, this was
a schrödinbug!
2) Fixed a time bomb. The kernel heap was accidentally put such that whenever a
few megabytes were allocated, it would begin overwriting the physical page stack
causing unthinkable events to unfold and would probably be even more obscure to
debug than 1).
Oh, and some string errors fixed and removed RunApplication from kernel.cpp,
funny thing that even linked in the first place. Guess, the optimizer actually
did work for once. :)
