CMSC 412 Project #4

Demand Paging and Virtual Memory

Intermediate submission due Friday April 8th, at 11:59pm
Final submission due Friday, April 15, at 11:59pm (Taxes are due April 18 this year, due to Emancipation Day)

Introduction

The purpose of this project is to add paging to your GeekOS kernel. This will require many small, but difficult, changes. More than any previous project, it will be important to implement one thing, test it, and then move to the next one. A successful implementation of earlier projects is not required for this project. You can implement Project 4 on top of any earlier version of the kernel (or directly on the provided distribution).

This project will involve an intermediate submission and a final project submission. The intermediate submission should be your implementation of the kernel memory mapping. It is worth 10 points (out of 100 for the project). To submit the intermediate submission just follow the same submission instructions as the final submission of project 4. For details, check the Intermediate submission section in the grading criteria.

Background

In this project you will add paging to the GeekOS kernel, combined with the segmentation that is already present. You may wish to refer to earlier discussion on segmentation from Project 1. For detailed information on paging, segmentation, and how they are combined on the Intel architecture, please refer to sections 8.4-8.7 in the text (pps. 288-309), and the Intel documentation specified above.

In a system that combines paging and segmentation, there are three kinds of addresses: logical addresses, linear addresses, and physical addresses. Logical addresses are those issued by a process. Logical addresses are mapped, via segmentation, to linear addresses by merely adding the base address of the relevant process segment to the logical address. A linear address is mapped to a physical address, which will be used by the processor to actually read from main memory, by using page tables set up and maintained by the operating system.  Paging is similar to segmentation because it allows each process to run in its own memory space. However, paging allows a much finer granularity of control by specifiying per-page mappings rather than a constant value offset. This will allow us to implement two useful features: demand paging, which allows a user program's stack to grow dynamically, and paging to disk, which allows memory to be "paged out" to disk, enabling the system to use more physical memory than it actually has.

In GeekOS, your paging system will use a page directory and page tables which form what as known as a two-level page table scheme (see Section 8.5.1, and notably Fig. 8.14, in the text).  Each memory address in the x86 system is 32 bits long.  To translate this linear address to a physical one, the processor will use the current page directory (this is analagous to finding an LDT for a process):

  1. It will use the most significant 10 bits of the linear address to index into the page directory to obtain a page directory entry. 10 bits allows you to write numbers between 0 and 1023, so therefore the page directory must have 1024 entries.
  2. The page directory entry is then used to find the appropriate page table. The next 10 bits are used to index into the page table to obtain a page table entry. As above, the page table must have 1024 entries.
  3. The page table entry is used to find the base address of the physical memory page. The last 12 bits of the linear memory address are used to index into the physical page. 12 bits allows you to write numbers between 0 and 4095, so therefore a page must be 4096 bytes in size.

Each page directory and page table are themselves pages in the system. Each page directory entry and page table entry is 4 bytes, and as explained above, each table contains 1024 entries. Therefore, each page directory and page table is 4096 bytes, which allows you to use pages of memory to represent the directories and tables. Each page directory contains a pointer to a page table which in turn contains a pointer to the physical memory.

Part I

The first step is to modify your project to use page tables and segmentation rather than just segments to provide memory protection.

Kernel Memory Mapping

The kernel currently starts running without paging enabled, and the segmentation hardware maps logical addresses directly to physical addresses.  Adding paging introduces a level in between: a logical address is mapped by segmentation to a linear address, which paging then maps to a physical address.  As a first step, we will introduce a page table that maps all linear addresses to themselves; that is each linear address X is mapped to physical address X. In Part II, this will become our page table for kernel-only threads.

To set up page tables, you will need to allocate a page directory (via Alloc_Page) and then allocate page tables for the entire region that will be mapped into this memory context. You will do this in the Init_VM in paging.c. You will need to fill out the appropriate fields in the page tables and page directories. The definition of paging tables and directories are found in paging.h (structs pte_t and pde_t). Finally, as mentioned earlier, GeekOS does not use paging by default, so to enable it, you will need to call the routine Enable_Paging which is already defined for you in lowlevel.asm. It takes the base address of your page directory as a parameter.

The final step of this function is to add a handler for page faults. Currently, a page fault can occur only when a user program attempts to access an invalid address. Therefore, we have provided a default handler, Page_Fault_Handler in paging.c, to terminate a user program that does this. You should register it for the page fault interrupt, interrupt 14, by calling Install_Interrupt_Handler. You should then add a call the Init_VM function from your main.c (after Init_Interrupts).

You should be able to complete this part and test it by temporarily giving user mode access to these pages by setting the flags field in the page table entries to include VM_USER.  This allows user programs complete access to the pages referenced by the table. Refer to item 1 in the grading criteria to understand how to test what you should have running at this point.  You can then submit this as your intermediate submission.

Part II

The next step is to make user processes have their own linear address space, as well as implementing demand paging and paging to disk.

User Memory Mapping

We will use the page directory and page tables you set up in Part I for all kernel-only threads, and will now add a page directory for each user process.  As such, you should change the flags fields in this directory to not include VM_USER.

The page directory for user processes will contain entries mapping user logical memory to linear memory, but will also contain entries to address the kernel memory.  This is not so user processes can access kernel memory directly (we will set the access flags of the memory to prevent them from doing so), but rather so that when an interrupt occurs, the page table does not need to be changed for the kernel to access its own memory; it will simply use the page table of the user process that was running when the interrupt occurred.

The memory layout for a user process is shown below, where the addresses on the left are linear addresses (which are mapped to physical addresses by the paging system).  User processes still refer to logical addresses, which are mapped by the segmentation system to the linear addresses shown here.  These addresses will be important in your paging implementation, and thus have constants defined in paging.h to represent them.

0x0000 0000 Kernel Memory

Start of kernel memory

(map all physical memory here)



<gap>


0x8000 0000 User Memory User address space begins here (unmapped page)
0x8000 1000 User Memory Text segment usually loaded here (segment->startAddress)


<gap>


0xFFFF E000 User Memory Initial stack at top of this page
0xFFFF F000 User Memory Args in this page
0xFFFF FFFF   Memory space ends here

The next step is to modify your user processes to use pages in the user range of memory. To help you do this, there is a new file called uservm.c that will replace your userseg.c from previous projects. You should start by taking the implementations from userseg.c and copying them to uservm.c and then modifying them for paging. There is a line in the Makefile.common in the build directory which specifies whether it should use userseg.c or uservm.c. You can switch between them by modifying USER_IMP_C variable.

Setting up paging for user processes occurs in Load_User_Program, and takes two steps.  First, you need to allocate a page directory for the process. You should copy all of the entries from the kernel page directory for the physical pages at the bottom of the address space.

Next, allocate page table entries for the user process's text, data, and stack regions.  Each of these regions will consist of some number of pages allocated by the routine Alloc_Pageable_Page.  This routine differs from Alloc_Page in that the allocated page it returns will have a special flag PAGE_PAGEABLE set in the flags field of its entry in the struct Page data structure (see mem.h).  This marks the page as eligible for being stolen and paged out to disk by the kernel when a page of memory is needed elsewhere, but no free pages are available. All pages (but not page directories and page tables) for a user space process should be allocated using this routine.

Contrast this with current implementation: for segmentation, one big chunk of memory was allocated for the entire user process. Paging allows per-page mappings for user memory so that each page of the user process is now allocated and mapped individually, and so it need not be contiguous.

Although there is a much better way, if you're stuck, do this as follows: calculate the size of the text and data segment of the process as is done now, without including the size of the stack (do round to PAGE_SIZE, though), and Malloc it.  Load the program into this memory, as now.  Copy the image, page by page, into the newly allocated pages.  Allocate two pages of memory at the end of the virtual address range (i.e., the last two entries in the last page table, as shown in the figure above). One is for the arguments, the other one is for stack.  Make sure the flags bits in both the page directory and page table entries allow user mode access (contain the VM_USER flag). Unmap the first page to support null pointer checking.

Finally, you will need to tweak some aspects of the current segmentation implementation so that it works with paging. The base linear address for the user mode process (that is, the base address for the code and data segments set in Create_User_Context) should be 0x80000000 (USER_VM_START in paging.h), and the limit should be (USER_VM_END - USER_VM_START) / PAGE_SIZE , rounded up. This will allow the user process to think that its logical address 0 is the linear address at 2GB and will simplify your kernel compared to traditional paged systems. You will also need to add code to switch the PDBR (cr3) register as part of a context switch. For this, in Switch_To_Address_Space you should add a call to Set_PDBR (provided for you in lowlevel.asm), after you load the LDT. You will use the pageDir field in the User_Context structure that will store the address of the process's page directory.

At this point, you should test your memory mapping by running QEMU. If you are able to load and run shell, you have completed this correctly.

Demand Paging

A nice benefit of paging is that it is straightforward to add to the memory allocation of a process by simply allocating a new (physical) page and mapping it into a free region in the process's address space.  For example, this can be done to allow the process's stack to grow beyond its initial allocation.  To implement stack growth, you must modify the default page fault handler from Part I.  The fault handler reads register cr2 to determine the faulting address. It also prints the errorCode defined in InterruptState and the fault defined in the struct faultcode_t in paging.h.

You will modify this page fault handler to determine an appropriate action to take based on the address.  In the case that the address is within one page of the current stack limit, you should allocate a new page, map it into the appropriate address, and then return normally from the handler.  The program will now be able to use the memory that you just allocated for it.  In order to test your new page fault handler, run the provided program rec.c.

Paging to Disk

While paging is useful for efficiently managing memory, the processes that can be active are still limited by the amount of physical memory.  To remedy this, you will implement paging to disk as part of GeekOS, in which memory pages can be temporarily stored on disk so that the freed physical memory can be used by another process.  To do this, the OS will need to pick a page to evict from memory and write it to the paging file, stored on disk. You should implement a version of pseudo-LRU algorithm (see section 9.4.5 in text, pp. 336). Use the accessed bit in the page tables to keep track of how frequently pages are accessed. To do this, maintain a global (static) clock value that points to a page in memory (a struct Page in the array of pages). When a page needs to be paged out, increment that clock, and clear the accessed bit or claim the page if the accessed bit was already clear.

The paging file consists of consecutive disk blocks of SECTOR_SIZE bytes.  Calling the routine Get_Paging_Device in vfs.h will return a Paging_Device structure containing the first disk block number of the paging file and the number of disk blocks in the paging file.  Each page will consume eight consecutive disk blocks (PAGE_SIZE/SECTOR_SIZE). To read and write the paging file, use the functions Block_Read and Block_Write provided. These functions write SECTOR_SIZE bytes at a time. How you manage your paging file is up to you. You may want to write a Init_Pagefile function in paging.c and call it from main.c

The code to page out a page is implemented for you in Alloc_Pageable_Page in mem.c, and works as follows:

Eventually, the page that was put on disk will be needed by some process again. At this point you will have to read it back off disk into memory (possibly while paging out another page to fit it into memory). Since the page that is paged out has its present bit set to 0 in the page table, an access to it will cause a page fault. Your page fault handler should then realize that this page is actually stored on disk and bring it back from disk (the kernelInfo field in the page table entry). When you bring a page in off disk, you may free the disk space used by the page. This will simplify your paging system, but will require that when a page is removed from memory it must always be written to the backing store. You will rely on the information stored when a page is paged out (such as pageBaseAddr) to find it on disk and page it back in.

The following table summarizes the actions of your page fault handler.

Cause Indication Action
Stack growing to new page Fault is within one page of the current stack limit Allocate a new page and continue.
Fault for paged out page Bits in page table indicate page is on disk Read page from paging device (sector indicated in PTE) and continue.
Fault for invalid address None of the other conditions apply Terminate user process

Copying Data Between Kernel and User Memory

Because the GeekOS kernel is preemptible and user memory pages can be stolen at any time, some subtle issues arise when copying data between the kernel and user memory spaces. Specifically, the kernel must never read or write data on a user memory page if that page has the PAGE_PAGEABLE bit set at any time that a thread switch could occur. The reason is simple; if a thread switch did occur, another process could run and steal the page. When control returns to the original thread, it would be reading or writing the wrong data, causing serious memory corruption.

There are two general approaches to dealing with this problem. One is that interrupts (and thus preemption) should be disabled while touching user memory. This approach is not a complete solution, because it is not legal to do I/O (i.e., Block_Read and Block_Write) while interrupts are disabled. The second approach is to use page locking. Before touching a user memory page, the kernel will atomically clear the PAGE_PAGEABLE flag for the page; this is referred to as locking the page. Once a page is locked, the kernel can then freely modify the page, safe in the knowledge that the page will not be stolen by another process. When it is done reading or writing the page, it can unlock the page by clearing the PAGE_PAGEABLE flag. Note that page flags should only be modified while interrupts are disabled.

Process Termination

As part of process termination, you will need to free the memory associated with a process. This includes freeing the pages used by the process, freeing the page tables and page directories. In addition, you will need to release the backing store space used by any pages of the terminating process. You should modify your Destroy_User_Context function in uservm.c to do this.

Extra Credit

As mentioned in recitation (and in recitation slides), you should use Alloc_Page() for allocating page tables and page directories and Alloc_Pageable_Page() for everything else (program memory, arguments, stack). This is intended to make your task easier, since you know the page tables are always in memory. For extra credit (10 points), make page tables pageable too, i.e., use Alloc_Pageable_Page() for page tables. (NOTE: we have no tests for this (yet), so if you do complete the extra credit task, please send mail to the TA's and Professor so your score can be adjusted.)