Lab 2: Memory Management

Handed out Wednesday, September 14, 2016
Due Thursday, September 29, 2016

Introduction

In this lab, you will write the memory management code for your operating system. Memory management has two components.

The first component is a physical memory allocator for the kernel, so that the kernel can allocate memory and later free it. Your allocator will operate in units of 4096 bytes, called pages. Your task will be to maintain data structures that record which physical pages are free and which are allocated, and how many processes are sharing each allocated page. You will also write the routines to allocate and free pages of memory.

The second component of memory management is virtual memory, which maps the virtual addresses used by kernel and user software to addresses in physical memory. The x86 hardware's memory management unit (MMU) performs the mapping when instructions use memory, consulting a set of page tables. You will modify JOS to set up the MMU's page tables according to a specification we provide.

Getting started

In this and future labs you will progressively build up your kernel. We will also provide you with some additional source. To fetch that source, use Git to commit changes you've made since handing in lab 1 (if any), fetch the latest version of the course repository, and then create a local branch called lab2 based on our lab2 branch, origin/lab2:

athena% cd ~/6.828/lab
athena% add git
athena% git pull
Already up-to-date.
athena% git checkout -b lab2 origin/lab2
Branch lab2 set up to track remote branch refs/remotes/origin/lab2.
Switched to a new branch "lab2"
athena% 

The git checkout -b command shown above actually does two things: it first creates a local branch lab2 that is based on the origin/lab2 branch provided by the course staff, and second, it changes the contents of your lab directory to reflect the files stored on the lab2 branch. Git allows switching between existing branches using git checkout branch-name, though you should commit any outstanding changes on one branch before switching to a different one.

You will now need to merge the changes you made in your lab1 branch into the lab2 branch, as follows:

athena% git merge lab1
Merge made by recursive.
 kern/kdebug.c  |   11 +++++++++--
 kern/monitor.c |   19 +++++++++++++++++++
 lib/printfmt.c |    7 +++----
 3 files changed, 31 insertions(+), 6 deletions(-)
athena% 

In some cases, Git may not be able to figure out how to merge your changes with the new lab assignment (e.g. if you modified some of the code that is changed in the second lab assignment). In that case, the git merge command will tell you which files are conflicted, and you should first resolve the conflict (by editing the relevant files) and then commit the resulting files with git commit -a.

Lab 2 contains the following new source files, which you should browse through:

memlayout.h describes the layout of the virtual address space that you must implement by modifying pmap.c. memlayout.h and pmap.h define the PageInfo structure that you'll use to keep track of which pages of physical memory are free. kclock.c and kclock.h manipulate the PC's battery-backed clock and CMOS RAM hardware, in which the BIOS records the amount of physical memory the PC contains, among other things. The code in pmap.c needs to read this device hardware in order to figure out how much physical memory there is, but that part of the code is done for you: you do not need to know the details of how the CMOS hardware works.

Pay particular attention to memlayout.h and pmap.h, since this lab requires you to use and understand many of the definitions they contain. You may want to review inc/mmu.h, too, as it also contains a number of definitions that will be useful for this lab.

Before beginning the lab, don't forget to add -f 6.828 to get the 6.828 version of QEMU.

Lab Requirements

In this lab and subsequent labs, do all of the regular exercises described in the lab and at least one challenge problem. (Some challenge problems are more challenging than others, of course!) Additionally, write up brief answers to the questions posed in the lab and a short (e.g., one or two paragraph) description of what you did to solve your chosen challenge problem. If you implement more than one challenge problem, you only need to describe one of them in the write-up, though of course you are welcome to do more. Place the write-up in a file called answers-lab2.txt in the top level of your lab directory before handing in your work.

Hand-In Procedure

When you are ready to hand in your lab code and write-up, add your answers-lab2.txt to the Git repository, commit your changes, and then run make handin.

athena% git add answers-lab2.txt
athena% git commit -am "my answer to lab2"
[lab2 a823de9] my answer to lab2
 4 files changed, 87 insertions(+), 10 deletions(-)
athena% make handin

As before, we will be grading your solutions with a grading program. You can run make grade in the lab directory to test your kernel with the grading program. You may change any of the kernel source and header files you need to in order to complete the lab, but needless to say you must not change or otherwise subvert the grading code.

Part 1: Physical Page Management

The operating system must keep track of which parts of physical RAM are free and which are currently in use. JOS manages the PC's physical memory with page granularity so that it can use the MMU to map and protect each piece of allocated memory.

You'll now write the physical page allocator. It keeps track of which pages are free with a linked list of struct PageInfo objects (which, unlike xv6, are not embedded in the free pages themselves), each corresponding to a physical page. You need to write the physical page allocator before you can write the rest of the virtual memory implementation, because your page table management code will need to allocate physical memory in which to store page tables.

Exercise 1. In the file kern/pmap.c, you must implement code for the following functions (probably in the order given).

boot_alloc()
mem_init() (only up to the call to check_page_free_list(1))
page_init()
page_alloc()
page_free()

check_page_free_list() and check_page_alloc() test your physical page allocator. You should boot JOS and see whether check_page_alloc() reports success. Fix your code so that it passes. You may find it helpful to add your own assert()s to verify that your assumptions are correct.

This lab, and all the 6.828 labs, will require you to do a bit of detective work to figure out exactly what you need to do. This assignment does not describe all the details of the code you'll have to add to JOS. Look for comments in the parts of the JOS source that you have to modify; those comments often contain specifications and hints. You will also need to look at related parts of JOS, at the Intel manuals, and perhaps at your 6.004 or 6.033 notes.

Part 2: Virtual Memory

Before doing anything else, familiarize yourself with the x86's protected-mode memory management architecture: namely segmentation and page translation.

Exercise 2. Look at chapters 5 and 6 of the Intel 80386 Reference Manual, if you haven't done so already. Read the sections about page translation and page-based protection closely (5.2 and 6.4). We recommend that you also skim the sections about segmentation; while JOS uses paging for virtual memory and protection, segment translation and segment-based protection cannot be disabled on the x86, so you will need a basic understanding of it.

Virtual, Linear, and Physical Addresses

In x86 terminology, a virtual address consists of a segment selector and an offset within the segment. A linear address is what you get after segment translation but before page translation. A physical address is what you finally get after both segment and page translation and what ultimately goes out on the hardware bus to your RAM.

           Selector  +--------------+         +-----------+
          ---------->|              |         |           |
                     | Segmentation |         |  Paging   |
Software             |              |-------->|           |---------->  RAM
            Offset   |  Mechanism   |         | Mechanism |
          ---------->|              |         |           |
                     +--------------+         +-----------+
            Virtual                   Linear                Physical

A C pointer is the "offset" component of the virtual address. In boot/boot.S, we installed a Global Descriptor Table (GDT) that effectively disabled segment translation by setting all segment base addresses to 0 and limits to 0xffffffff. Hence the "selector" has no effect and the linear address always equals the offset of the virtual address. In lab 3, we'll have to interact a little more with segmentation to set up privilege levels, but as for memory translation, we can ignore segmentation throughout the JOS labs and focus solely on page translation.

Recall that in part 3 of lab 1, we installed a simple page table so that the kernel could run at its link address of 0xf0100000, even though it is actually loaded in physical memory just above the ROM BIOS at 0x00100000. This page table mapped only 4MB of memory. In the virtual memory layout you are going to set up for JOS in this lab, we'll expand this to map the first 256MB of physical memory starting at virtual address 0xf0000000 and to map a number of other regions of virtual memory.

Exercise 3. While GDB can only access QEMU's memory by virtual address, it's often useful to be able to inspect physical memory while setting up virtual memory. Review the QEMU monitor commands from the lab tools guide, especially the xp command, which lets you inspect physical memory. To access the QEMU monitor, press Ctrl-a c in the terminal (the same binding returns to the serial console).

Use the xp command in the QEMU monitor and the x command in GDB to inspect memory at corresponding physical and virtual addresses and make sure you see the same data.

Our patched version of QEMU provides an info pg command that may also prove useful: it shows a compact but detailed representation of the current page tables, including all mapped memory ranges, permissions, and flags. Stock QEMU also provides an info mem command that shows an overview of which ranges of virtual memory are mapped and with what permissions.

From code executing on the CPU, once we're in protected mode (which we entered first thing in boot/boot.S), there's no way to directly use a linear or physical address. All memory references are interpreted as virtual addresses and translated by the MMU, which means all pointers in C are virtual addresses.

The JOS kernel often needs to manipulate addresses as opaque values or as integers, without dereferencing them, for example in the physical memory allocator. Sometimes these are virtual addresses, and sometimes they are physical addresses. To help document the code, the JOS source distinguishes the two cases: the type uintptr_t represents opaque virtual addresses, and physaddr_t represents physical addresses. Both these types are really just synonyms for 32-bit integers (uint32_t), so the compiler won't stop you from assigning one type to another! Since they are integer types (not pointers), the compiler will complain if you try to dereference them.

The JOS kernel can dereference a uintptr_t by first casting it to a pointer type. In contrast, the kernel can't sensibly dereference a physical address, since the MMU translates all memory references. If you cast a physaddr_t to a pointer and dereference it, you may be able to load and store to the resulting address (the hardware will interpret it as a virtual address), but you probably won't get the memory location you intended.

To summarize:

C typeAddress type
T*  Virtual
uintptr_t  Virtual
physaddr_t  Physical

Question

  1. Assuming that the following JOS kernel code is correct, what type should variable x have, uintptr_t or physaddr_t?
    	mystery_t x;
    	char* value = return_a_pointer();
    	*value = 10;
    	x = (mystery_t) value;

The JOS kernel sometimes needs to read or modify memory for which it knows only the physical address. For example, adding a mapping to a page table may require allocating physical memory to store a page directory and then initializing that memory. However, the kernel, like any other software, cannot bypass virtual memory translation and thus cannot directly load and store to physical addresses. One reason JOS remaps all of physical memory starting from physical address 0 at virtual address 0xf0000000 is to help the kernel read and write memory for which it knows just the physical address. In order to translate a physical address into a virtual address that the kernel can actually read and write, the kernel must add 0xf0000000 to the physical address to find its corresponding virtual address in the remapped region. You should use KADDR(pa) to do that addition.

The JOS kernel also sometimes needs to be able to find a physical address given the virtual address of the memory in which a kernel data structure is stored. Kernel global variables and memory allocated by boot_alloc() are in the region where the kernel was loaded, starting at 0xf0000000, the very region where we mapped all of physical memory. Thus, to turn a virtual address in this region into a physical address, the kernel can simply subtract 0xf0000000. You should use PADDR(va) to do that subtraction.

Reference counting

In future labs you will often have the same physical page mapped at multiple virtual addresses simultaneously (or in the address spaces of multiple environments). You will keep a count of the number of references to each physical page in the pp_ref field of the struct PageInfo corresponding to the physical page. When this count goes to zero for a physical page, that page can be freed because it is no longer used. In general, this count should be equal to the number of times the physical page appears below UTOP in all page tables (the mappings above UTOP are mostly set up at boot time by the kernel and should never be freed, so there's no need to reference count them). We'll also use it to keep track of the number of pointers we keep to the page directory pages and, in turn, of the number of references the page directories have to page table pages.

Be careful when using page_alloc. The page it returns will always have a reference count of 0, so pp_ref should be incremented as soon as you've done something with the returned page (like inserting it into a page table). Sometimes this is handled by other functions (for example, page_insert) and sometimes the function calling page_alloc must do it directly.

Page Table Management

Now you'll write a set of routines to manage page tables: to insert and remove linear-to-physical mappings, and to create page table pages when needed.

Exercise 4. In the file kern/pmap.c, you must implement code for the following functions.

        pgdir_walk()
        boot_map_region()
        page_lookup()
        page_remove()
        page_insert()
	

check_page(), called from mem_init(), tests your page table management routines. You should make sure it reports success before proceeding.

Part 3: Kernel Address Space

JOS divides the processor's 32-bit linear address space into two parts. User environments (processes), which we will begin loading and running in lab 3, will have control over the layout and contents of the lower part, while the kernel always maintains complete control over the upper part. The dividing line is defined somewhat arbitrarily by the symbol ULIM in inc/memlayout.h, reserving approximately 256MB of virtual address space for the kernel. This explains why we needed to give the kernel such a high link address in lab 1: otherwise there would not be enough room in the kernel's virtual address space to map in a user environment below it at the same time.

You'll find it helpful to refer to the JOS memory layout diagram in inc/memlayout.h both for this part and for later labs.

Permissions and Fault Isolation

Since kernel and user memory are both present in each environment's address space, we will have to use permission bits in our x86 page tables to allow user code access only to the user part of the address space. Otherwise bugs in user code might overwrite kernel data, causing a crash or more subtle malfunction; user code might also be able to steal other environments' private data. Note that the writable permission bit (PTE_W) affects both user and kernel code!

The user environment will have no permission to any of the memory above ULIM, while the kernel will be able to read and write this memory. For the address range [UTOP,ULIM), both the kernel and the user environment have the same permission: they can read but not write this address range. This range of address is used to expose certain kernel data structures read-only to the user environment. Lastly, the address space below UTOP is for the user environment to use; the user environment will set permissions for accessing this memory.

Initializing the Kernel Address Space

Now you'll set up the address space above UTOP: the kernel part of the address space. inc/memlayout.h shows the layout you should use. You'll use the functions you just wrote to set up the appropriate linear to physical mappings.

Exercise 5. Fill in the missing code in mem_init() after the call to check_page().

Your code should now pass the check_kern_pgdir() and check_page_installed_pgdir() checks.

Question

  1. What entries (rows) in the page directory have been filled in at this point? What addresses do they map and where do they point? In other words, fill out this table as much as possible:
    Entry Base Virtual Address Points to (logically):
    1023?Page table for top 4MB of phys memory
    1022??
    .??
    .??
    .??
    20x00800000?
    10x00400000?
    00x00000000[see next question]
  2. We have placed the kernel and user environment in the same address space. Why will user programs not be able to read or write the kernel's memory? What specific mechanisms protect the kernel memory?
  3. What is the maximum amount of physical memory that this operating system can support? Why?
  4. How much space overhead is there for managing memory, if we actually had the maximum amount of physical memory? How is this overhead broken down?
  5. Revisit the page table setup in kern/entry.S and kern/entrypgdir.c. Immediately after we turn on paging, EIP is still a low number (a little over 1MB). At what point do we transition to running at an EIP above KERNBASE? What makes it possible for us to continue executing at a low EIP between when we enable paging and when we begin running at an EIP above KERNBASE? Why is this transition necessary?

Challenge! We consumed many physical pages to hold the page tables for the KERNBASE mapping. Do a more space-efficient job using the PTE_PS ("Page Size") bit in the page directory entries. This bit was not supported in the original 80386, but is supported on more recent x86 processors. You will therefore have to refer to Volume 3 of the current Intel manuals. Make sure you design the kernel to use this optimization only on processors that support it!

Challenge! Extend the JOS kernel monitor with commands to:

Address Space Layout Alternatives

The address space layout we use in JOS is not the only one possible. An operating system might map the kernel at low linear addresses while leaving the upper part of the linear address space for user processes. x86 kernels generally do not take this approach, however, because one of the x86's backward-compatibility modes, known as virtual 8086 mode, is "hard-wired" in the processor to use the bottom part of the linear address space, and thus cannot be used at all if the kernel is mapped there.

It is even possible, though much more difficult, to design the kernel so as not to have to reserve any fixed portion of the processor's linear or virtual address space for itself, but instead effectively to allow user-level processes unrestricted use of the entire 4GB of virtual address space - while still fully protecting the kernel from these processes and protecting different processes from each other!

Challenge! Write up an outline of how a kernel could be designed to allow user environments unrestricted use of the full 4GB virtual and linear address space. Hint: the technique is sometimes known as "follow the bouncing kernel." In your design, be sure to address exactly what has to happen when the processor transitions between kernel and user modes, and how the kernel would accomplish such transitions. Also describe how the kernel would access physical memory and I/O devices in this scheme, and how the kernel would access a user environment's virtual address space during system calls and the like. Finally, think about and describe the advantages and disadvantages of such a scheme in terms of flexibility, performance, kernel complexity, and other factors you can think of.

Challenge! Since our JOS kernel's memory management system only allocates and frees memory on page granularity, we do not have anything comparable to a general-purpose malloc/free facility that we can use within the kernel. This could be a problem if we want to support certain types of I/O devices that require physically contiguous buffers larger than 4KB in size, or if we want user-level environments, and not just the kernel, to be able to allocate and map 4MB superpages for maximum processor efficiency. (See the earlier challenge problem about PTE_PS.)

Generalize the kernel's memory allocation system to support pages of a variety of power-of-two allocation unit sizes from 4KB up to some reasonable maximum of your choice. Be sure you have some way to divide larger allocation units into smaller ones on demand, and to coalesce multiple small allocation units back into larger units when possible. Think about the issues that might arise in such a system.

This completes the lab. Make sure you pass all of the make grade tests and don't forget to write up your answers to the questions and a description of your challenge exercise solution in answers-lab2.txt. Commit your changes (including adding answers-lab2.txt) and type make handin in the lab directory to hand in your lab.