Lab 4: preemptive multitasking

  • Part A due: Tue, 08 Nov 2016 23:00:00 -0800
  • Part B due: Tue, 15 Nov 2016 23:00:00 -0800
  • Part C due: Tue, 22 Nov 2016 23:00:00 -0800

Introduction

In this lab you will implement preemptive multitasking among multiple simultaneously active user-mode environments.

In part A you will add multiprocessor support to JOS, implement round-robin scheduling, and add basic environment management system calls that create and destroy environments.

In part B, you will implement a Unix-like fork(), which allows a user-mode environment to create copies of itself.

Finally, in part C you will add support for inter-process communication (IPC), allowing different user-mode environments to communicate and synchronize with each other explicitly. You will also add support for hardware clock interrupts and preemption.

Getting started

Use Git to commit your Lab 3 source, fetch the latest version of the course repository, and then create a local branch called lab4 based on our branch, origin/lab4:

$ cd lab
$ git pull
Already up-to-date.
$ git checkout -b lab4 origin/lab4
Branch lab4 set up to track remote branch refs/remotes/origin/lab4.
Switched to a new branch "lab4"
$ git merge lab3
Merge made by recursive.
...
$ 

Lab 4 contains a number of new source files, some of which you should browse before you start:

  • kern/cpu.h: kernel-private definitions for multiprocessor support
  • kern/mpconfig.c, kern/acpi.c: code to read the multiprocessor configuration from ACPI tables
  • kern/lapic.c: kernel code driving the local APIC unit in each processor
  • kern/ioapic.c: kernel code driving the IO APIC unit
  • kern/mpentry.S: assembly-language entry code for non-boot CPUs
  • kern/spinlock.h: kernel-private definitions for spin locks, including the big kernel lock
  • kern/spinlock.c: kernel code implementing spin locks
  • kern/sched.c: code skeleton of the scheduler that you are about to implement

Lab requirements

This lab is divided into three parts, A, B, and C. We have allocated one week in the schedule for each part.

As before, you will need to do all of the regular exercises described in the lab. Write up your answers to the questions a file called answers-lab4.txt and place it in the top level of your lab directory before handing in your work.

Part A: Cooperative multitasking

In the first part of this lab, you will first extend JOS to run on a multiprocessor system, and then implement some new JOS kernel system calls to allow user-level environments to create additional new environments. You will also implement cooperative round-robin scheduling, allowing the kernel to switch from one environment to another when the current environment voluntarily relinquishes the CPU (or exits). Later in part C you will implement preemptive scheduling, which allows the kernel to re-take control of the CPU from an environment after a certain time has passed even if the environment does not cooperate.

Multiprocessor support

We are going to make JOS support “symmetric multiprocessing” (SMP), a multiprocessor model in which all CPUs have equivalent access to system resources such as memory and I/O buses. While all CPUs are functionally identical in SMP, during the boot process they can be classified into two types: the bootstrap processor (BSP) is responsible for initializing the system and for booting the operating system; and the application processors (APs) are activated by the BSP only after the operating system is up and running. Which processor is the BSP is determined by the hardware and the BIOS. Up to this point, all your existing JOS code has been running on the BSP.

In an SMP system, each CPU has an accompanying local APIC (LAPIC) unit. The LAPIC units are responsible for delivering interrupts throughout the system. The LAPIC also provides its connected CPU with a unique identifier. In this lab, we make use of the following basic functionality of the LAPIC unit (in kern/lapic.c):

  • Reading the LAPIC identifier (APIC ID) to tell which CPU our code is currently running on (see cpunum()).
  • Sending the STARTUP interprocessor interrupt (IPI) from the BSP to the APs to bring up other CPUs (see lapic_startap()).
  • In part C, we program LAPIC’s built-in timer to trigger clock interrupts to support preemptive multitasking (see apic_init()).

A processor accesses its LAPIC using memory-mapped I/O (MMIO). In MMIO, a portion of physical memory is hardwired to the registers of some I/O devices, so the same load/store instructions typically used to access memory can be used to access device registers. You’ve already seen one IO hole at physical address 0xA0000 (we use this to write to the VGA display buffer). The LAPIC lives in a hole starting at physical address 0xFE000000 (32MB short of 4GB), so it’s too high for us to access using our usual direct map at KERNBASE. The JOS virtual memory map leaves a 4MB gap at MMIOBASE so we have a place to map devices like this. Since later labs introduce more MMIO regions, you’ll write a simple function to allocate space from this region and map device memory to it.

Exercise

Implement mmio_map_region in kern/pmap.c. To see how this is used, look at the beginning of lapic_init in kern/lapic.c, or ioapic_init in kern/ioapic.c. You’ll have to do the next exercise, too, before the tests for mmio_map_region will run.

Application processor bootstrap

Before booting up APs, the BSP should first collect information about the multiprocessor system, such as the total number of CPUs, their APIC IDs and the MMIO address of the LAPIC unit. The mp_init() function in kern/mpconfig.c retrieves this information by reading the Advanced Configuration and Power Interface (ACPI) tables provided by the BIOS.

The boot_aps() function (in kern/init.c) drives the AP bootstrap process. APs start in real mode, much like how the bootloader started in boot/boot.S, so boot_aps() copies the AP entry code (kern/mpentry.S) to a memory location that is addressable in the real mode. Unlike with the bootloader, we have some control over where the AP will start executing code; we copy the entry code to 0x7000 (MPENTRY_PADDR), but any unused, page-aligned physical address below 640KB would work.

After that, boot_aps() activates APs one after another, by sending STARTUP IPIs to the LAPIC unit of the corresponding AP, along with an initial CS:IP address at which the AP should start running its entry code (MPENTRY_PADDR in our case). The entry code in kern/mpentry.S is quite similar to that of boot/boot.S. After some brief setup, it puts the AP into protected mode with paging enabled, and then calls the C setup routine mp_main() (also in kern/init.c). boot_aps() waits for the AP to signal a CPU_STARTED flag in cpu_status field of its struct CpuInfo before going on to wake up the next one.

Exercise

Read boot_aps() and mp_main() in kern/init.c, and the assembly code in kern/mpentry.S. Make sure you understand the control flow transfer during the bootstrap of APs. Then modify your implementation of page_init() in kern/pmap.c to avoid adding the page at MPENTRY_PADDR to the free list, so that we can safely copy and run AP bootstrap code at that physical address. Your code should pass the updated check_page_free_list() test (but might fail the updated check_kern_pgdir() test, which we will fix soon).

Question

Compare kern/mpentry.S side by side with boot/boot.S. Bearing in mind that kern/mpentry.S is compiled and linked to run above KERNBASE just like everything else in the kernel, what is the purpose of macro MPBOOTPHYS? Why is it necessary in kern/mpentry.S but not in boot/boot.S? In other words, what could go wrong if it were omitted in kern/mpentry.S?

Hint: recall the differences between the link address and the load address that we have discussed in Lab 1.

Per-CPU state and initialization

When writing a multiprocessor OS, it is important to distinguish between per-CPU state that is private to each processor, and global state that the whole system shares. kern/cpu.h defines most of the per-CPU state, including struct CpuInfo, which stores per-CPU variables. cpunum() always returns the ID of the CPU that calls it, which can be used as an index into arrays like cpus (note that cpu_apicid cannot be directly used as an index, as it is not guaranteed to be continuous). Alternatively, the macro thiscpu is shorthand for the current CPU’s struct CpuInfo.

Here is the per-CPU state you should be aware of:

  • Per-CPU kernel stack.

    Because multiple CPUs can trap into the kernel simultaneously, we need a separate kernel stack for each processor to prevent them from interfering with each other’s execution. The array percpu_kstacks[NCPU][KSTKSIZE] reserves space for NCPU’s worth of kernel stacks.

    In Lab 2, you mapped the physical memory that bootstack refers to as the BSP’s kernel stack just below KSTACKTOP. Similarly, in this lab, you will map each CPU’s kernel stack into this region with guard pages acting as a buffer between them. CPU 0’s stack will still grow down from KSTACKTOP; CPU 1’s stack will start KSTKGAP bytes below the bottom of CPU 0’s stack, and so on. inc/memlayout.h shows the mapping layout.

  • Per-CPU TSS and TSS descriptor.

    A per-CPU task state segment (TSS) is also needed in order to specify where each CPU’s kernel stack lives. The TSS for CPU i is stored in cpus[i].cpu_ts, and the corresponding TSS descriptor is defined in the GDT entry gdt[(GD_TSS0 >> 3) + i]. The global ts variable defined in kern/trap.c will no longer be useful.

  • Per-CPU current environment pointer.

    Since each CPU can run different user process simultaneously, we redefined the symbol curenv to refer to cpus[cpunum()].cpu_env (or thiscpu->cpu_env), which points to the environment currently executing on the current CPU (the CPU on which the code is running).

  • Per-CPU system registers.

    All registers, including system registers, are private to a CPU. Therefore, instructions that initialize these registers, such as lcr3(), ltr(), lgdt(), lidt(), etc., must be executed once on each CPU. Functions env_init_percpu() and trap_init_percpu() are defined for this purpose.

Exercise

Modify mem_init_mp() (in kern/pmap.c) to map per-CPU stacks starting at KSTACKTOP, as shown in inc/memlayout.h. The size of each stack is KSTKSIZE bytes plus KSTKGAP bytes of unmapped guard pages. Your code should pass the new check in check_kern_pgdir().

Exercise

The code in trap_init_percpu() (in kern/trap.c) initializes the TSS and TSS descriptor for the BSP. It worked in Lab 3, but is incorrect when running on other CPUs. Change the code so that it can work on all CPUs. (Note: your new code should not use the global ts variable any more.)

When you finish the above exercises, run JOS in QEMU with 4 CPUs using make qemu CPUS=4 (or make qemu-nox CPUS=4), you should see output like this:

...
E820: physical memory 128MB
check_page_alloc() succeeded!
check_page() succeeded!
check_kern_pgdir() succeeded!
check_page_installed_pgdir() succeeded!
ACPI: RSDP 0x000f6430 000014 v00 BOCHS 
ACPI: RSDT 0x07fe2085 000038 v01 BOCHS  BXPCRSDT 01 BXPC 01
ACPI: FACP 0x07fe1b78 000074 v01 BOCHS  BXPCFACP 01 BXPC 01
ACPI: SSDT 0x07fe1bec 000395 v01 BOCHS  BXPCSSDT 01 BXPC 01
ACPI: APIC 0x07fe1f81 000090 v01 BOCHS  BXPCAPIC 01 BXPC 01
ACPI: HPET 0x07fe2011 000038 v01 BOCHS  BXPCHPET 01 BXPC 01
ACPI: MCFG 0x07fe2049 00003c v01 BOCHS  BXPCMCFG 01 BXPC 01
SMP: 4 CPU(s)
SMP: LAPIC 0xfee00000 v14
SMP: IOAPIC 0xfec00000 v11 [global_irq 00-23]
SMP: BSP #0 [apicid 00]
  AP #1 [apicid 01] starting
  AP #2 [apicid 02] starting
  AP #3 [apicid 03] starting

Challenge: fast cpunum()

The current implementation of cpunum() is slow: it first reads the APIC ID of the current CPU, and then scans the cpus array to return the index of a matching CPU. How to make this faster?

For example, you may use a register (e.g., %gs) to store CPU-specific data, such as a CPU’s index. Another approach is to make use of a model specific register (MSR), IA32_TSC_AUX, which is available on recent x86 processors. Implement a faster cpunum() using one of the approaches.

Note that you’ll need the RDTSCP instruction to read the IA32_TSC_AUX register, which is not supported by default in QEMU. You’ll need to specify a newer CPU model (e.g., QEMUEXTRA="-cpu Broadwell") or explicitly enable it (e.g., QEMUEXTRA="-cpu qemu32,+rdtscp). Make sure you see the CPU feature rdtscp in JOS’s output before proceeding. See Section 17.14.2, IA32_TSC_AUX Register and RDTSCP Support of the Intel 64 and IA-32 Architectures Software Developer’s Manual for detail.

Locking

Our current code spins after initializing the AP in mp_main(). Before letting the AP get any further, we need to first address race conditions when multiple CPUs run kernel code simultaneously. The simplest way to achieve this is to use a big kernel lock. The big kernel lock is a single global lock that is held whenever an environment enters kernel mode, and is released when the environment returns to user mode. In this model, environments in user mode can run concurrently on any available CPUs, but no more than one environment can run in kernel mode; any other environments that try to enter kernel mode are forced to wait.

kern/spinlock.h declares the big kernel lock, namely kernel_lock. It also provides lock_kernel() and unlock_kernel(), shortcuts to acquire and release the lock. You should apply the big kernel lock at four locations:

  • In i386_init(), acquire the lock before the BSP wakes up the other CPUs.
  • In mp_main(), acquire the lock after initializing the AP, and then call sched_yield() to start running environments on this AP.
  • In trap(), acquire the lock when trapped from user mode. To determine whether a trap happened in user mode or in kernel mode, check the low bits of the tf_cs.
  • In env_run(), release the lock right before switching to user mode. Do not do that too early or too late, otherwise you will experience races or deadlocks.

Exercise

Apply the big kernel lock as described above, by calling lock_kernel() and unlock_kernel() at the proper locations.

How to test if your locking is correct? You can’t at this moment! But you will be able to after you implement the scheduler in the next exercise.

Question

It seems that using the big kernel lock guarantees that only one CPU can run the kernel code at a time. Why do we still need separate kernel stacks for each CPU? Describe a scenario in which using a shared kernel stack will go wrong, even with the protection of the big kernel lock.

Challenge: fine-grained locking

The big kernel lock is simple and easy to use. Nevertheless, it eliminates all concurrency in kernel mode. Most modern operating systems use different locks to protect different parts of their shared state, an approach called fine-grained locking. Fine-grained locking can increase performance significantly, but is more difficult to implement and error-prone. If you are brave enough, drop the big kernel lock and embrace concurrency in JOS!

It is up to you to decide the locking granularity (the amount of data that a lock protects). As a hint, you may consider using spin locks to ensure exclusive access to these shared components in the JOS kernel:

  • The page allocator.
  • The console driver.
  • The scheduler.
  • The inter-process communication (IPC) state that you will implement in the part C.

Round-robin scheduling

Your next task in this lab is to change the JOS kernel so that it can alternate between multiple environments in “round-robin” fashion. Round-robin scheduling in JOS works as follows:

  • The function sched_yield() in the new kern/sched.c is responsible for selecting a new environment to run. It searches sequentially through the envs[] array in circular fashion, starting just after the previously running environment (or at the beginning of the array if there was no previously running environment), picks the first environment it finds with a status of ENV_RUNNABLE (see inc/env.h), and calls env_run() to jump into that environment.

  • sched_yield() must never run the same environment on two CPUs at the same time. It can tell that an environment is currently running on some CPU (possibly the current CPU) because that environment’s status will be ENV_RUNNING.

  • We have implemented a new system call for you, sys_yield(), which user environments can call to invoke the kernel’s sched_yield() function and thereby voluntarily give up the CPU to a different environment.

Exercise

Implement round-robin scheduling in sched_yield() as described above. Don’t forget to modify syscall() to dispatch sys_yield().

Make sure to invoke sched_yield() in mp_main.

Modify kern/init.c to create three (or more!) environments that all run the program user/yield.c.

Run make qemu. You should see the environments switch back and forth between each other five times before terminating, like below.

Test also with several CPUS: make qemu CPUS=2.

...
Hello, I am environment 00001000.
Hello, I am environment 00001001.
Hello, I am environment 00001002.
Back in environment 00001000, iteration 0.
Back in environment 00001001, iteration 0.
Back in environment 00001002, iteration 0.
Back in environment 00001000, iteration 1.
Back in environment 00001001, iteration 1.
Back in environment 00001002, iteration 1.
...

After the yield programs exit, there will be no runnable environment in the system, the scheduler should invoke the JOS kernel monitor. If any of this does not happen, then fix your code before proceeding.

If you use CPUS=1 at this point, all environments should successfully run. Setting CPUS larger than 1 at this time may result in a general protection or kernel page fault once there are no more runnable environments due to unhandled timer interrupts (which we will fix below!).

Question

In your implementation of env_run() you should have called lcr3(). Before and after the call to lcr3(), your code makes references (at least it should) to the variable e, the argument to env_run. Upon loading the %cr3 register, the addressing context used by the MMU is instantly changed. But a virtual address (namely e) has meaning relative to a given address context–the address context specifies the physical address to which the virtual address maps. Why can the pointer e be dereferenced both before and after the addressing switch?

Question

Whenever the kernel switches from one environment to another, it must ensure the old environment’s registers are saved so they can be restored properly later. Why? Where does this happen?

Challenge: scheduling

Add a less trivial scheduling policy to the kernel, such as a fixed-priority scheduler that allows each environment to be assigned a priority and ensures that higher-priority environments are always chosen in preference to lower-priority environments. If you’re feeling really adventurous, try implementing a Unix-style adjustable-priority scheduler or even a lottery or stride scheduler. (Look up “lottery scheduling” and “stride scheduling” in Google.)

Write a test program or two that verifies that your scheduling algorithm is working correctly (i.e., the right environments get run in the right order). It may be easier to write these test programs once you have implemented fork() and IPC in parts B and C of this lab.

Challenge: floating point

The JOS kernel currently does not allow applications to use the x86 processor’s x87 floating-point unit (FPU), MMX instructions, or Streaming SIMD Extensions (SSE). Extend the Env structure to provide a save area for the processor’s floating point state, and extend the context switching code to save and restore this state properly when switching from one environment to another. The FXSAVE and FXRSTOR instructions may be useful, but note that these are not in the old i386 user’s manual because they were introduced in more recent processors. Write a user-level test program that does something cool with floating point.

System calls for environment creation

Although your kernel is now capable of running and switching between multiple user-level environments, it is still limited to running environments that the kernel initially set up. You will now implement the necessary JOS system calls to allow user environments to create and start other new user environments.

Unix provides the fork() system call as its process creation primitive. Unix fork() copies the entire address space of calling process (the parent) to create a new process (the child). The only differences between the two observable from user space are their process IDs and parent process IDs (as returned by getpid and getppid). In the parent, fork() returns the child’s process ID, while in the child, fork() returns 0. By default, each process gets its own private address space, and neither process’s modifications to memory are visible to the other.

You will provide a different, more primitive set of JOS system calls for creating new user-mode environments. With these system calls you will be able to implement a Unix-like fork() entirely in user space, in addition to other styles of environment creation. The new system calls you will write for JOS are as follows:

  • sys_exofork: This system call creates a new environment with an almost blank slate: nothing is mapped in the user portion of its address space, and it is not runnable. The new environment will have the same register state as the parent environment at the time of the sys_exofork call. In the parent, sys_exofork will return the envid_t of the newly created environment (or a negative error code if the environment allocation failed). In the child, however, it will return 0. (Since the child starts out marked as not runnable, sys_exofork will not actually return in the child until the parent has explicitly allowed this by marking the child runnable using….)
  • sys_env_set_status: Sets the status of a specified environment to ENV_RUNNABLE or ENV_NOT_RUNNABLE. This system call is typically used to mark a new environment ready to run, once its address space and register state has been fully initialized.

For all of the system calls above that accept environment IDs, the JOS kernel supports the convention that a value of 0 means “the current environment.” This convention is implemented by envid2env() in kern/env.c.

We have provided a very primitive implementation of a Unix-like fork() in the test program user/dumbfork.c. This test program uses the above system calls to create and run a child environment with a copy of its own address space. The two environments then switch back and forth using sys_yield as in the previous exercise. The parent exits after 10 iterations, whereas the child exits after 20.

Exercise

Implement the system calls described above in kern/syscall.c. You will need to use various functions in kern/pmap.c and kern/env.c, particularly envid2env(). For now, whenever you call envid2env(), pass 1 in the checkperm parameter. Be sure you check for any invalid system call arguments, returning -E_INVAL in that case. Test your JOS kernel with user/dumbfork and make sure it works before proceeding.

Challenge: checkpointing

Add the additional system calls necessary to read all of the vital state of an existing environment as well as set it up. Then implement a user mode program that forks off a child environment, runs it for a while (e.g., a few iterations of sys_yield()), then takes a complete snapshot or checkpoint of the child environment, runs the child for a while longer, and finally restores the child environment to the state it was in at the checkpoint and continues it from there. Thus, you are effectively “replaying” the execution of the child environment from an intermediate state. Make the child environment perform some interaction with the user using sys_cgetc() or readline() so that the user can view and mutate its internal state, and verify that with your checkpoint/restart you can give the child environment a case of selective amnesia, making it “forget” everything that happened beyond a certain point.

This completes Part A of the lab; check it using make grade and hand in the tarball from make handin as usual. If you are trying to figure out why a particular test case is failing, run ./grade-lab4 -v, which will show you the output of the kernel builds and QEMU runs for each test, until a test fails. When a test fails, the script will stop, and then you can inspect jos.out to see what the kernel actually printed.

Part B: Copy-on-write fork

As mentioned earlier, Unix provides the fork() system call as its primary process creation primitive. The fork() system call copies the address space of the calling process (the parent) to create a new process (the child).

The approach that dumbfork() takes is essentially copying all data from the parent’s pages into new pages allocated for the child. The copying of the parent’s address space into the child is the most expensive part of the fork() operation.

However, a call to fork() is frequently followed almost immediately by a call to exec() in the child process, which replaces the child’s memory with a new program. This is what the the shell typically does, for example. In this case, the time spent copying the parent’s address space is largely wasted, because the child process will use very little of its memory before calling exec().

For this reason, later versions of Unix took advantage of virtual memory hardware to allow the parent and child to share the memory mapped into their respective address spaces until one of the processes actually modifies it. This technique is known as copy-on-write. To do this, on fork() the kernel would copy the address space mappings from the parent to the child instead of the contents of the mapped pages, and at the same time mark the now-shared pages read-only. When one of the two processes tries to write to one of these shared pages, the process takes a page fault. At this point, the Unix kernel realizes that the page was really a “virtual” or “copy-on-write” copy, and so it makes a new, private, writable copy of the page for the faulting process. In this way, the contents of individual pages aren’t actually copied until they are actually written to. This optimization makes a fork() followed by an exec() in the child much cheaper: the child will probably only need to copy one page (the current page of its stack) before it calls exec().

In the next piece of this lab, you will implement a “proper” Unix-like fork() with copy-on-write, as a user space library routine. Implementing fork() and copy-on-write support in user space has the benefit that the kernel remains much simpler and thus more likely to be correct. It also lets individual user-mode programs define their own semantics for fork(). A program that wants a slightly different implementation (for example, the expensive always-copy version like dumbfork(), or one in which the parent and child actually share memory afterward) can easily provide its own.

User-level page fault handling

A user-level copy-on-write fork() needs to know about page faults on write-protected pages, so that’s what you’ll implement first. Copy-on-write is only one of many possible uses for user-level page fault handling.

It’s common to set up an address space so that page faults indicate when some action needs to take place. For example, most Unix kernels initially map only a single page in a new process’s stack region, and allocate and map additional stack pages later “on demand” as the process’s stack consumption increases and causes page faults on stack addresses that are not yet mapped. A typical Unix kernel must keep track of what action to take when a page fault occurs in each region of a process’s space. For example, a fault in the stack region will typically allocate and map new page of physical memory. A fault in the program’s BSS region will typically allocate a new page, fill it with zeroes, and map it. In systems with demand-paged executables, a fault in the text region will read the corresponding page of the binary off of disk and then map it.

This is a lot of information for the kernel to keep track of. Instead of taking the traditional Unix approach, you will decide what to do about each page fault in user space, where bugs are less damaging. This design has the added benefit of allowing programs great flexibility in defining their memory regions; you’ll use user-level page fault handling later for mapping and accessing files on a disk-based file system.

Setting the page fault handler

In order to handle its own page faults, a user environment will need to register a page fault handler entrypoint with the JOS kernel. The user environment registers its page fault entrypoint via the new sys_env_set_pgfault_upcall system call. We have added a new member to the Env structure, env_pgfault_upcall, to record this information.

Exercise

Implement the sys_env_set_pgfault_upcall system call. Be sure to enable permission checking when looking up the environment ID of the target environment, since this is a “dangerous” system call.

Normal & exception stacks in user environments

During normal execution, a user environment in JOS will run on the normal user stack: its ESP register starts out pointing at USTACKTOP, and the stack data it pushes resides on the page between USTACKTOP-PGSIZE and USTACKTOP-1 inclusive. When a page fault occurs in user mode, however, the kernel will restart the user environment running a designated user-level page fault handler on a different stack, namely the user exception stack. In essence, we will make the JOS kernel implement automatic “stack switching” on behalf of the user environment, in much the same way that the x86 processor already implements stack switching on behalf of JOS when transferring from user mode to kernel mode!

The JOS user exception stack is also one page in size, and its top is defined to be at virtual address UXSTACKTOP, so the valid bytes of the user exception stack are from UXSTACKTOP-PGSIZE through UXSTACKTOP-1 inclusive. While running on this exception stack, the user-level page fault handler can use JOS’s regular system calls to map new pages or adjust mappings so as to fix whatever problem originally caused the page fault. Then the user-level page fault handler returns, via an assembly language stub, to the faulting code on the original stack.

Each user environment that wants to support user-level page fault handling will need to allocate memory for its own exception stack, using the sys_page_alloc() system call.

Invoking the user page fault handler

You will now need to change the page fault handling code in kern/trap.c to handle page faults from user mode as follows. We will call the state of the user environment at the time of the fault the trap-time state.

If there is no page fault handler registered, the JOS kernel destroys the user environment with a message as before. Otherwise, the kernel sets up a trap frame on the exception stack that looks like a struct UTrapframe from inc/trap.h:

                    <-- UXSTACKTOP
trap-time esp
trap-time eflags
trap-time eip
trap-time eax       start of struct PushRegs
trap-time ecx
trap-time edx
trap-time ebx
trap-time esp
trap-time ebp
trap-time esi
trap-time edi       end of struct PushRegs
tf_err (error code)
fault_va            <-- %esp when handler is run

The kernel then arranges for the user environment to resume execution with the page fault handler running on the exception stack with this stack frame; you must figure out how to make this happen. The fault_va is the virtual address that caused the page fault.

If the user environment is already running on the user exception stack when an exception occurs, then the page fault handler itself has faulted. In this case, you should start the new stack frame just under the current tf->tf_esp rather than at UXSTACKTOP. You should first push an empty 32-bit word, then a struct UTrapframe.

To test whether tf->tf_esp is already on the user exception stack, check whether it is in the range between UXSTACKTOP-PGSIZE and UXSTACKTOP-1, inclusive.

Exercise

Implement the code in page_fault_handler in kern/trap.c required to dispatch page faults to the user-mode handler. Be sure to take appropriate precautions when writing into the exception stack. (What happens if the user environment runs out of space on the exception stack?)

User-mode page fault entrypoint

Next, you need to implement the assembly routine that will take care of calling the C page fault handler and resume execution at the original faulting instruction. This assembly routine is the handler that will be registered with the kernel using sys_env_set_pgfault_upcall().

Exercise

Implement the _pgfault_upcall routine in lib/pfentry.S. The interesting part is returning to the original point in the user code that caused the page fault. You’ll return directly there, without going back through the kernel. The hard part is simultaneously switching stacks and re-loading the EIP.

Finally, you need to implement the C user library side of the user-level page fault handling mechanism.

Exercise

Finish set_pgfault_handler() in lib/pgfault.c.

Testing

Run user/faultread (make run-faultread). You should see:

...
[00000000] new env 00001000
[00001000] user fault va 00000000 ip 0080003a
TRAP frame ...
[00001000] free env 00001000

Run user/faultdie. You should see:

...
[00000000] new env 00001000
i faulted at va deadbeef, err 6
[00001000] exiting gracefully
[00001000] free env 00001000

Run user/faultalloc. You should see:

...
[00000000] new env 00001000
fault deadbeef
this string was faulted in at deadbeef
fault cafebffe
fault cafec000
this string was faulted in at cafebffe
[00001000] exiting gracefully
[00001000] free env 00001000

If you see only the first “this string” line, it means you are not handling recursive page faults properly.

Run user/faultallocbad. You should see:

...
[00000000] new env 00001000
[00001000] user_mem_check assertion failure for va deadbeef
[00001000] free env 00001000

Make sure you understand why user/faultalloc and user/faultallocbad behave differently.

Challenge: user-mode exception handler

Extend your kernel so that not only page faults, but all types of processor exceptions that code running in user space can generate, can be redirected to a user-mode exception handler. Write user-mode test programs to test user-mode handling of various exceptions such as divide-by-zero, general protection fault, and illegal opcode.

Implementing copy-on-write fork

You now have the kernel facilities to implement copy-on-write fork() entirely in user space.

We have provided a skeleton for your fork() in lib/fork.c. Like dumbfork(), fork() should create a new environment, then scan through the parent environment’s entire address space and set up corresponding page mappings in the child. The key difference is that, while dumbfork() copied pages, fork() will initially only copy page mappings. fork() will copy each page only when one of the environments tries to write it.

The basic control flow for fork() is as follows:

  1. The parent installs pgfault() as the C-level page fault handler, using the set_pgfault_handler() function you implemented above.

  2. The parent calls sys_exofork() to create a child environment.

  3. For each writable or copy-on-write page in its address space below UTOP, the parent calls duppage, which should map the page copy-on-write into the address space of the child and then remap the page copy-on-write in its own address space. duppage sets both PTEs so that the page is not writeable, and to contain PTE_COW in the “avail” field to distinguish copy-on-write pages from genuine read-only pages.

    The exception stack is not remapped this way, however. Instead you need to allocate a fresh page in the child for the exception stack. Since the page fault handler will be doing the actual copying and the page fault handler runs on the exception stack, the exception stack cannot be made copy-on-write: who would copy it?

    fork() also needs to handle pages that are present, but not writable or copy-on-write.

  4. The parent sets the user page fault entrypoint for the child to look like its own.

  5. The child is now ready to run, so the parent marks it runnable.

Each time one of the environments writes a copy-on-write page that it hasn’t yet written, it will take a page fault. Here’s the control flow for the user page fault handler:

  1. The kernel propagates the page fault to _pgfault_upcall, which calls fork()’s pgfault() handler.

  2. pgfault() checks that the fault is a write (check for FEC_WR in the error code) and that the PTE for the page is marked PTE_COW. If not, panic.

  3. pgfault() allocates a new page mapped at a temporary location and copies the contents of the faulting page into it. Then the fault handler maps the new page at the appropriate address with read/write permissions, in place of the old read-only mapping.

The user-level lib/fork.c code must consult the environment’s page tables for several of the operations above (e.g., that the PTE for a page is marked PTE_COW). The kernel maps the environment’s page tables at UVPT exactly for this purpose. It uses a clever mapping trick to make it to make it easy to lookup PTEs for user code. lib/entry.S sets up uvpt and uvpd so that you can easily lookup page-table information in lib/fork.c.

Exercise

Implement fork, duppage and pgfault in lib/fork.c.

Test your code with the forktree program. It should produce the following messages, with interspersed ‘new env’, ‘free env’, and ‘exiting gracefully’ messages. The messages may not appear in this order, and the environment IDs may be different.

1000: I am ''
1001: I am '0'
2000: I am '00'
2001: I am '000'
1002: I am '1'
3000: I am '11'
3001: I am '10'
4000: I am '100'
1003: I am '01'
5000: I am '010'
4001: I am '011'
2002: I am '110'
1004: I am '001'
1005: I am '111'
1006: I am '101'

Challenge: shared-memory forking

Implement a shared-memory fork() called sfork(). This version should have the parent and child share all their memory pages (so writes in one environment appear in the other) except for pages in the stack area, which should be treated in the usual copy-on-write manner. Modify user/forktree.c to use sfork() instead of regular fork(). Also, once you have finished implementing IPC in part C, use your sfork() to run user/pingpongs. You will have to find a new way to provide the functionality of the global thisenv pointer.

Challenge: system call batching

Your implementation of fork makes a huge number of system calls. On x86, switching into the kernel using interrupts has non-trivial cost. Augment the system call interface so that it is possible to send a batch of system calls at once. Then change fork to use this interface.

How much faster is your new fork?

You can answer this (roughly) by using analytical arguments to estimate how much of an improvement batching system calls will make to the performance of your fork: How expensive is an int 0x30 instruction? How many times do you execute int 0x30 in your fork? Is accessing the TSS stack switch also expensive? And so on…

Alternatively, you can boot your kernel on real hardware and really benchmark your code. See the RDTSC (read time-stamp counter) instruction, defined in the Intel manual, which counts the number of clock cycles that have elapsed since the last processor reset. QEMU doesn’t emulate this instruction faithfully (it can either count the number of virtual instructions executed or use the host TSC, neither of which reflects the number of cycles a real CPU would require).

This ends part B. As usual, you can grade your submission with make grade and hand in the tarball from make handin.

Part C: Preemptive multitasking & IPC

In the final part of lab 4 you will modify the kernel to preempt uncooperative environments and to allow environments to pass messages to each other explicitly.

Clock interrupts and preemption

Run the user/spin test program. This test program forks off a child environment, which simply spins forever in a tight loop once it receives control of the CPU. Neither the parent environment nor the kernel ever regains the CPU. This is obviously not an ideal situation in terms of protecting the system from bugs or malicious code in user-mode environments, because any user-mode environment can bring the whole system to a halt simply by getting into an infinite loop and never giving back the CPU. In order to allow the kernel to preempt a running environment, forcefully retaking control of the CPU from it, we must extend the JOS kernel to support external hardware interrupts from the clock hardware.

Interrupt discipline

External interrupts (i.e., device interrupts) are referred to as IRQs. There are 16 standard IRQs, numbered 0 through 15 (newer systems typically have 24 or more IRQs; we only deal with 16 IRQs in this lab). The mapping from IRQ number to IDT entry is not fixed. The code in picirq.c and ioapic.c maps IRQs 0–15 to IDT entries IRQ_OFFSET through IRQ_OFFSET+15.

In inc/trap.h, IRQ_OFFSET is defined to be decimal 32. Thus the IDT entries 32–47 correspond to the IRQs 0–15. For example, the clock interrupt is IRQ 0. Thus, IDT[IRQ_OFFSET+0] (i.e., IDT[32]) contains the address of the clock’s interrupt handler routine in the kernel. This IRQ_OFFSET is chosen so that the device interrupts do not overlap with the processor exceptions, which could obviously cause confusion. (In fact, in the early days of PCs running MS-DOS, the IRQ_OFFSET effectively was zero, which indeed caused massive confusion between handling hardware interrupts and handling processor exceptions!)

In JOS, we make a key simplification: external device interrupts are always disabled when in the kernel (and enabled when in user space). External interrupts are controlled by the FL_IF flag bit of the %eflags register (see inc/mmu.h). When this bit is set, external interrupts are enabled. While the bit can be modified in several ways, because of our simplification, we will handle it solely through the process of saving and restoring %eflags register as we enter and leave user mode.

You will have to ensure that the FL_IF flag is set in user environments when they run so that when an interrupt arrives, it gets passed through to the processor and handled by your interrupt code. Otherwise, interrupts are masked, or ignored until interrupts are re-enabled. We masked interrupts with the very first instruction of the bootloader, and so far we have never gotten around to re-enabling them.

Exercise

Modify kern/trapentry.S and kern/trap.c to initialize the appropriate entries in the IDT and provide handlers for IRQs 0 through 15. Then modify the code in env_alloc() in kern/env.c to ensure that user environments are always run with interrupts enabled.

The processor never pushes an error code or checks the Descriptor Privilege Level (DPL) of the IDT entry when invoking a hardware interrupt handler. You might want to re-read Section 9.2, “Enabling and Disabling Interrupts” of the 80386 Reference Manual (or Section 6.8 of the Intel 64 and IA-32 Architectures Software Developer’s Manuals) at this time.

After doing this exercise, if you run your kernel with any test program that runs for a non-trivial length of time (e.g., user/spin), you should see the kernel print trap frames for hardware interrupts. While interrupts are now enabled in the processor, JOS isn’t yet handling them, so you should see it misattribute each interrupt to the currently running user environment and destroy it. Eventually it should run out of environments to destroy and drop into the monitor.

Handling clock interrupts

In the user/spin program, after the child environment was first run, it just spun in a loop, and the kernel never got control back. We need to program the hardware to generate clock interrupts periodically, which will force control back to the kernel where we can switch control to a different user environment.

The calls to lapic_init, pic_init, and ioapic_init (from i386_init in init.c), which we have written for you, set up the clock and the interrupt controller to generate interrupts. You now need to write the code to handle these interrupts.

Exercise

Modify the kernel’s trap_dispatch() function so that it calls sched_yield() to find and run a different environment whenever a clock interrupt takes place.

You should now be able to get the user/spin test to work: the parent environment should fork off the child, sys_yield() to it a couple times but in each case regain control of the CPU after one time slice, and finally kill the child environment and terminate gracefully.

This is a great time to do some regression testing. Make sure that you haven’t broken any earlier part of that lab that used to work (e.g., forktree) by enabling interrupts. Also, try running with multiple CPUs using make CPUS=2 target. You should also be able to pass stresssched now. Run make grade to see for sure.

Time

Your kernel does not have a notion of time, so let’s add it. Since a clock interrupt is generated by the hardware every 10ms, on every clock interrupt we can increment a variable to indicate that time has advanced by 10ms. This is implemented in kern/sysinfo.c, but is not yet fully integrated into your kernel.

Exercise

Add a call to time_tick for every clock interrupt in kern/trap.c. Implement sys_sysinfo and add it to syscall in kern/syscall.c so that user space has access to the time. Use make run-testtime to test your time code. You should see the environment count down from 5 in 1 second intervals.

Don’t worry about other fields in struct sysinfo for now. You’ll fill in them in later labs.

Challenge: HPET

As you may have noticed, the time precision of your JOS is 10ms. A more precise timing technique for implementing time is through the High Precision Event Timer (HPET) chip. Read the HPET specification in the readings page. You should be able to find the base address of the HPET registers through the ACPI “HPET” table. Use MMIO to program the HPET chip.

Challenge: fast sys_sysinfo

The system call sys_sysinfo does not change the kernel state; it merely reads the values of a few kernel variables and fills in struct sysinfo. There is really no need to pay the cost of switching into the kernel.

Here’s one way to improve the performance. First, map a page to a user read-only address (look at inc/memlayout.h and find an unused page). Second, update values to that page in the kernel. Then, change sys_sysinfo() in lib/syscall.c to read from that address, instead of making a system call. Implement this optimization or design your own. You may also be interested in Linux’s vsyscalls and vDSO.

Inter-process communication (IPC)

(Technically in JOS this is “inter-environment communication” or “IEC,” but everyone else calls it IPC, so we’ll use the standard term.)

We’ve been focusing on the isolation aspects of the operating system, the ways it provides the illusion that each program has a machine all to itself. Another important service of an operating system is to allow programs to communicate with each other when they want to. It can be quite powerful to let programs interact with other programs. The Unix pipe model is the canonical example.

There are many models for interprocess communication. Even today there are still debates about which models are best. We won’t get into that debate. Instead, we’ll implement a simple IPC mechanism and then try it out.

IPC in JOS

You will implement a few additional JOS kernel system calls that collectively provide a simple interprocess communication mechanism. You will implement two system calls, sys_ipc_recv and sys_ipc_try_send. Then you will implement two library wrappers ipc_recv and ipc_send.

The “messages” that user environments can send to each other using JOS’s IPC mechanism consist of two components: a single 32-bit value, and optionally a single page mapping. Allowing environments to pass page mappings in messages provides an efficient way to transfer more data than will fit into a single 32-bit integer, and also allows environments to set up shared memory arrangements easily.

Sending and receiving messages

To receive a message, an environment calls sys_ipc_recv. This system call de-schedules the current environment and does not run it again until a message has been received. When an environment is waiting to receive a message, any other environment can send it a message - not just a particular environment, and not just environments that have a parent/child arrangement with the receiving environment. In other words, the permission checking that you implemented in Part A will not apply to IPC, because the IPC system calls are carefully designed so as to be “safe”: an environment cannot cause another environment to malfunction simply by sending it messages (unless the target environment is also buggy).

To try to send a value, an environment calls sys_ipc_try_send with both the receiver’s environment id and the value to be sent. If the named environment is actually receiving (it has called sys_ipc_recv and not gotten a value yet), then the send delivers the message and returns 0. Otherwise the send returns -E_IPC_NOT_RECV to indicate that the target environment is not currently expecting to receive a value.

A library function ipc_recv in user space will take care of calling sys_ipc_recv and then looking up the information about the received values in the current environment’s struct Env.

Similarly, a library function ipc_send will take care of repeatedly calling sys_ipc_try_send until the send succeeds.

Transferring pages

When an environment calls sys_ipc_recv with a valid dstva parameter (below UTOP), the environment is stating that it is willing to receive a page mapping. If the sender sends a page, then that page should be mapped at dstva in the receiver’s address space. If the receiver already had a page mapped at dstva, then that previous page is unmapped.

When an environment calls sys_ipc_try_send with a valid srcva (below UTOP), it means the sender wants to send the page currently mapped at srcva to the receiver, with permissions perm. After a successful IPC, the sender keeps its original mapping for the page at srcva in its address space, but the receiver also obtains a mapping for this same physical page at the dstva originally specified by the receiver, in the receiver’s address space. As a result this page becomes shared between the sender and receiver.

If either the sender or the receiver does not indicate that a page should be transferred, then no page is transferred. After any IPC the kernel sets the new field env_ipc_perm in the receiver’s Env structure to the permissions of the page received, or zero if no page was received.

Implementing IPC

Exercise

Implement sys_ipc_recv and sys_ipc_try_send in kern/syscall.c. Read the comments on both before implementing them, since they have to work together. When you call envid2env in these routines, you should set the checkperm flag to 0, meaning that any environment is allowed to send IPC messages to any other environment, and the kernel does no special permission checking other than verifying that the target envid is valid.

Then implement the ipc_recv and ipc_send functions in lib/ipc.c.

Use the user/pingpong and user/primes functions to test your IPC mechanism. user/primes will generate for each prime number a new environment until JOS runs out of environments. You might find it interesting to read user/primes.c to see all the forking and IPC going on behind the scenes.

Challenge: fast IPC

Why does ipc_send have to loop? Change the system call interface so it doesn’t have to. Make sure you can handle multiple environments trying to send to one environment at the same time.

Furthermore, make JOS’s IPC mechanism more efficient by applying some of the techniques from Liedtke’s paper, Improving IPC by Kernel Design, or any other tricks you may think of. Feel free to modify the kernel’s system call API for this purpose, as long as your code is backwards compatible with what our grading scripts expect.

Challenge: CSP

The prime sieve is only one neat use of message passing between a large number of concurrent programs. Read C. A. R. Hoare’s Communicating Sequential Processes and implement the matrix multiplication example. Or, implement Doug McIlroy’s power series calculator described in Squinting at Power Series and compute the power series for sin(x+x3).

This ends part C. Make sure you pass all of the make grade tests. In answers-lab4.txt, write up your answers to the questions, how much time you spend on this lab, and the names of your team members if you work in pairs.

Before handing in, use git status and git diff to examine your changes. When you’re ready, commit your changes, type make handin, and upload the tarball through the course dropbox.