Skip to content

Latest commit

 

History

History
58 lines (34 loc) · 2.9 KB

README.md

File metadata and controls

58 lines (34 loc) · 2.9 KB
Raw

Brian OS

Lots of inspiration drawn from the OSTEP xv6 projects.

Things added

  • Threads, accessible to users via thread_create() and join() library functions.
  • getpstat system call and corresponding ps user program to get summary of current process statuses.
  • Unmapped first page, to produce pagefaults when null pointers are dereferenced
  • mprotect() and munprotect() system calls to toggle page write permissions
  • Automatic read-only permission for any page loaded into process memory that contains only program text.

Notes

Summary of the address space

It's essentially the same as the standard xv6 address space, but shifted up by one page. RO protections are added to all pages which contain only program text. If there are NTEXT of these, and NELF pages total in the executable, the layout looks like this:

0x0000 - 0x1000: unmapped

0x1000 - NTEXT * 0x1000: RO text

NTEXT * 0x1000 - NELF * 0x1000: RW program data (some .text, as well as the data sections)

NELF * 0x1000 - (NELF+1) * 0x1000: guard page, user-inaccessible

(NELF+1) * 0x1000 - (NELF+2) * 0x1000: stack

How threading is done.

The API is mostly borrowed from the OSTEP project description. Under the head, each thread is implemented as an xv6 process and lives in the process table - from the point of view of the scheduler, it's indistinguishable from a process. The stack thread lives in user memory, obtained via malloc(). The thread is initialized in the clone() system call, which allocates a kernel stack to the new thread and copies the page directory of the parent process. The stack is set up with a single void * argument, and the instruction pointer is set to the function pointer passed in by the user before the thread is handed off to the scheduler.

Threads, like xv6 processes, must terminate via exit(). Zombie threads are collected by join(), which provides the stack pointer of each dead thread to the user for cleanup.

Spin locks are added to the umalloc library, to prevent race conditions when multiple threads in the same address space need to allocated memory.

How gettimes() works

The system call gettimes() fills in a times struct for a user process with two integer values:

  • procticks - the number of times this process has been forced to yield from the RUNNING state. Initialized to zero in exec() - this is a rough measure of time elapsed in a single active thread.
  • allticks - the total number of timer interrupts since the process was exec()'d.

TODO

  • clear
  • better ps (address space summary?)
  • mmap
  • demand zeroing, copy-on-write for pages allocated to processes
  • a time utility
  • swap space and a page cache

known issues

  • interrupts are disabled by default when acquiring spinlocks, which means clock-tick values based on timer interrupt counts are only approximate.
  • tick variables will overflow uint bounds.