Lab: xv6 lazy page allocation

One of the many neat tricks an O/S can play with page table hardware is lazy allocation of heap memory. Xv6 applications ask the kernel for heap memory using the sbrk() system call. In the kernel we've given you, sbrk() allocates physical memory and maps it into the process's virtual address space. There are programs that allocate memory but never use it, for example to implement large sparse arrays. Sophisticated kernels delay allocation of each page of memory until the application tries to use that page -- as signaled by a page fault. You'll add this lazy allocation feature to xv6 in this lab.

Part One: Eliminate allocation from sbrk()

Your first task is to delete page allocation from the sbrk(n) system call implementation, which is the function sys_sbrk() in sysproc.c. The sbrk(n) system call grows the process's memory size by n bytes, and then returns the start of the newly allocated region (i.e., the old size). Your new sbrk(n) should just increment the process's size (myproc()->sz) by n and return the old size. It should not allocate memory -- so you should delete the call to growproc() (but you still need to increase the process's size!).

Try to guess what the result of this modification will be: what will break?

Make this modification, boot xv6, and type echo hi to the shell. You should see something like this:

init: starting sh
$ echo hi
usertrap(): unexpected scause 0x000000000000000f pid=3
            sepc=0x00000000000011dc stval=0x0000000000004008
va=0x0000000000004000 pte=0x0000000000000000
panic: unmappages: not mapped

The "usertrap(): ..." message is from the user trap handler in trap.c; it has caught an exception that it does not know how to handle. Make sure you understand why this page fault occurs. The "stval=0x0..04008" indicates that the virtual address that caused the page fault is 0x4008.

Part Two: Lazy allocation

Modify the code in trap.c to respond to a page fault from user space by mapping a newly-allocated page of physical memory at the faulting address, and then returning back to user space to let the process continue executing. You should add your code just before the printf call that produced the "usertrap(): ..." message.

Hint: look at the printf arguments to see how to find the virtual address that caused the page fault.

Hint: steal code from allocuvm() in vm.c, which is what sbrk() calls (via growproc()).

Hint: use PGROUNDDOWN(va) to round the faulting virtual address down to a page boundary.

Hint: usertrapret() in order to avoid the printf and the myproc()->killed = 1.

Hint: you'll need to call mappages().

Hint: you can check whether a fault is a page fault by r_scause() is 13 or 15 in trap().

Hint: modify unmappages() to not free pages that aren't mapped.

Hint: if the kernel crashes, look up sepc in kernel/kernel.asm

Hint: if you see the error "imcomplete type proc", include "proc.h" (and "spinlock.h").

Hint: the first test in sbrk() allocates something large, this should succeed now.

If all goes well, your lazy allocation code should result in echo hi working. You should get at least one page fault (and thus lazy allocation) in the shell, and perhaps two.

If you have the basics working, now turn your implementation into one that handles the corner cases too:

Handle negative sbrk() arguments. sbrktest() in usertests will tests this.
Handle fork correctly. sbrktst() will test this.
Make sure that kernel use of not-yet-allocated user addresses works; for example, if a program passes an sbrk()-allocated address to write(). sbrktest() will test this.
Handle out of memory correctly. sbrktst() will test this.
Handle faults on the invalid page below the stack. stacktest() in usertests will tests this.

Run all tests in usertests() to make sure your solution doesn't break other tests.

Submit: The code that you added to trap.c in a file named hwN.c where N is the homework number as listed on the schedule.