Assignment 3. Virtual Memory

Initial design spec due on March 29, Monday, 11:59 PM.

Design review meetings on March 30.

Complete code and documentation due on April 8, Thursday, 11:59 PM.

Introduction

The third phase of Nachos is to investigate the use of caching. In this assignment we use caching for two purposes. First, we use a software-managed translation lookaside buffer (TLB) as a cache for page tables to provide the illusion of fast access to virtual page translation over a large address space. Second, we use memory as a cache for disk, to provide the abstraction of an (almost) unlimited virtual memory size, with performance close to that provided by physical memory. We provide no new code for this assignment (the only change is that you need to compile with the VM and USE_TLB flags); your job is to write the code to manage the TLB and to implement virtual memory.

Page tables were used in Assignment 2 to simplify memory allocation and to isolate failures from one address space from affecting other programs. For this assignment, the hardware knows nothing about page tables. Instead it only deals with a software-loaded cache of page table entries, called the TLB. On almost all modern processor architectures, a TLB is used to speed address translation. Given a memory address (an instruction to fetch, or data to load or store), the processor first looks in the TLB to determine if the mapping of virtual page to physical page is already known. If so (a TLB ``hit''), the translation can be done quickly. But if the mapping is not in the TLB (a TLB ``miss''), page tables and/or segment tables are used to determine the correct translation. On several architectures, including Nachos, the DEC MIPS, and the HP Snakes, a ``TLB miss'' simply causes a trap to the OS kernel, which does the translation, loads the mapping into the the TLB and re-starts the program. This allows the OS kernel to choose whatever combination of page table, segment table, inverted page table, etc., it needs to do the translation. On systems without software-managed TLB's, the hardware does the same thing as the software, but in this case, the hardware must specify the exact format for page and segment tables. Thus, software managed TLB's are more flexible, at a cost of being somewhat slower for handling TLB misses. If TLB misses are very infrequent, the performance impact of software managed TLB's can be minimal.

The illusion of unlimited memory is provided by the operating system by using main memory as a cache for the disk. For this assignment, page translation allows us the flexibility to get pages from disk as they are needed. Each entry in the TLB has a valid bit: if the valid bit is set, the virtual page is in memory. If the valid bit is clear or if the virtual page is not found in the TLB, a software page table is needed to tell whether the the page is in memory (with the TLB to be loaded with the translation), or the page must be brought in from disk. In addition, the hardware sets the use bit in the TLB entry whenever a page is referenced and the dirty bit whenever the page is modified.

When a program references a page that is not in the TLB, the hardware generates a TLB exception, trapping to the kernel. The operating system kernel then checks its own page table. If the page is not in memory, it reads the page in from disk, sets the page table entry to point to the new page, and then resumes the execution of the user program. Of course, the kernel must first find space in memory for the incoming page, potentially writing some other page back to disk, if it has been modified.

As with any caching system, performance depends on the policy used to decide which things are kept in memory and which are only stored on disk. On a page fault, the kernel must decide which page to replace; ideally, it will throw out a page that will not be referenced for a long time, keeping pages in memory those that are soon to be referenced. Another consideration is that if the replaced page has been modified, the page must be first saved to disk before the needed page can be brought in; many virtual memory systems (such as UNIX) avoid this extra overhead by writing modified pages to disk in advance, so that any subsequent page faults can be completed more quickly.

Problem 1

Implement software-management of the TLB. For this, you will need to implement some kind of software page translation, for handling TLB misses. Note that with the compile time flag -DUSE_TLB, the hardware no longer deals with page tables; thus, you need to do something about making sure the TLB state is set up properly on a context switch. Most systems simply invalidate all the TLB entries on a context switch; the entries get re-loaded as the pages are referenced. For Problem 2, your page translation scheme should keep track of the dirty and use flags for each page set by hardware in the TLB entry.

Problem 2

Implement virtual memory. For this, you will need routines to move a page from disk to memory and from memory to disk. We recommend that you use the Nachos file system as backing store -- this way, when we implement the file system in Assignment 4, we'll be able to use the virtual memory system as a test case. In order to find unreferenced pages to throw out on page faults, you will need to keep track of all of the pages in the system which are currently in use. A simple way to do this is to keep a ``core map'', which is basically a reverse page table -- instead of translating virtual page numbers to physical pages, a core map translates physical page numbers to the virtual pages that are stored there.

Problem 3 (Extra Credit)

Implement memory-mapped files. The traditional way to access the file system is via Read and Write system calls, but that requires an extra level of copying between the kernel and the user level (since Read and Write transfer data through a temporary kernel buffer). A different interface is simply to "map" the file into the virtual address space; the program can then use load and store instructions directly on the file data. (An alternative way of viewing the file system is as "durable memory"; files just store data structures. If you access data structures in memory using load and store instructions, why not access data structures in files the same way?)

In order to provide memory-mapped files to user programs, you need to implement an "mmap" system call, that takes an open file descriptor, and a location in the address space to put the file:

char* mmap(OpenFileId id, char* requestedAddress, int* length);

Here, id is the OpenFileId of the file to mmap; requestedAddress points to the virtual address at which the program would like to map the file. If this is NULL (i.e. 0), the kernel may either (a) choose the virtual address on its own (through complicated dynamic virtual address space allocation that most of you will probably not want to do), or (b) fail the call. If the program picks the virtual address, it must be aligned on a page boundary, and all the pages that would be needed by the mmap must not overlap any other allocation (e.g. the program's code, data, and stack, or another mmap'ed region). The kernel does not need to mmap more pages as the file grows.

length points to a 32-bit integer that the kernel should fill with the length of the region mapped (beware endianness!).

mmap returns the virtual address of the beginning of the mapped region. If requestedAddress was non-NULL, mmap must only try to use this address. Otherwise, mmap is free to allocate any virtual address region that does not overlap with any other allocation.

From part 2, you already have the mechanism to bring missing pages of the mapped file into memory -- in fact, you can use the same mechanism to bring in code and data pages from the executable on program startup.

Note that the consequence of supporting mmap is that address spaces will be sparsely populated with (potentially lots of) segments, one per open file. For parts 1 and 2, you should design your software translation mechanisms with this in mind.

Implement a utility program (such as UNIX "cp") that exploits the new interface.

Modifications to the file are committed to disk when the file is closed. Note that you may have consistency problems if you call Read or Write on a mmap'ed file. We do not require that you deal with this, but you should know how you could deal with it.

Problem 4 (Extra Credit)

Implement an sbrk system call that allows you to increase the size of a mmap-ed file:

   int sbrk(OpenFileId id, int increase);

Here, id must have previously been passed to mmap, and is the OpenFileId of the file to grow. increase must be a non-negative integer specifying the amount by which the file should be grown. sbrk should return the new size (in bytes) of the region.

Implement a memory heap through mmap and sbrk. To create the heap, call mmap with id set to zero; the kernel should use the length pointer to tell the program the initial heap size. To grow the heap, call sbrk with id set to zero. The C++ "new" implementation uses "sbrk" to increase the amount of available virtual memory, whenever it runs out of free space.