Some sources call all exceptions “interrupts,” calling the interrupting-type of exception an “asynchronous interrupt” instead.

2.1.2 Faults

A fault is the result of an instruction failing to succeed in its execution. Examples include dividing by zero, dereferencing a null pointer, or attempting to execute a protected instruction while in user mode.

There are two basic families of responses to faults: either the kernel can fix the problem, then re-run the user code that generated the fault; or it cannot fix the problem and kills the process that cause the fault instead. In practice, fixable faults happen quite often, notably the page fault discussed later. Even faults the kernel can’t fix on its own are often handled by asking the user code if it knows how to fix them using a mechanism called signals, though many programs do not chose to handle the signals so the end result is often still termination of the fault-generating process.

2.1.3 Traps

A trap is an exception caused by a special instruction whose sole job is to generate exceptions. Traps are the main mechanism for intentionally switching from user mode to kernel mode and are the core of all system calls. System calls are used to interact with the file system, spawn and wait on threads, listen for key presses, and anything else that you cannot do with simple code alone.

2.2 Handling

When an exception occurs, the processor switches to kernel mode and jumps to a special function in kernel code called an “exception handler.” Because interrupts exist, this can happen without any particular instruction causing the jump, and thus might happen at any point during the code running. In order for the handler to later resume user code, the exception handling process must also save the processor state before executing the handler.

2.2.1 Save-Handle-Restore

The basic mechanism for any exception to be handled is

1. Save the current state of the processor (register contents, PC location, etc)
2. Enter kernel mode
3. Jump to code designed to react to the exception in question, called an exception handler
4. When the handler finishes, enter user mode and restore processor state (including register contents, PC location, etc)

These steps (except for the actual execution of the exception handler) are all done atomically by the hardware.

2.2.2 Which Handler?

In general, there can be as many different exception handlers as there are exception causes. To select which one to run, the hardware consults something called an exception table. The exception table is just an array of code addresses; the index into that array is determined by the kind of exception generated (e.g., divide-by-zero uses index 0, invalid instructions use index 6, etc.) The index is called an exception number or vector number and the array of code addresses is called the exception table.

switch(x) {
    case 0: f0();
    case 1: f1();
    case 2: f2();
    case 3: f3();
        break;
    case 4: f4();
    case 5: f5();
    case 6: f6();
}

    cmpl    $6, %eax
    ja      AfterSwitch
    jmpq    * Table(,%rax,8)
Case0:
    callq   f0
Case1:
    callq   f1
Case2:
    callq   f2
Case3:
    callq   f3
    jmp     AfterSwitch
Case4:
    callq   f4
Case5:
    callq   f5
Case6:
    callq   f6
AfterSwitch:

    .section    .rodata,"a",@progbits
Table:
    .quad   Case0
    .quad   Case1
    .quad   Case2
    .quad   Case3
    .quad   Case4
    .quad   Case5
    .quad   Case6

2.2.3 After the Handler

Handlers may either abort the user code or resume its operation. Aborting is primarily used when there is no obvious way to recover from the cause of the exception.

When resuming, the handler can either re-run the instruction that was running when it was generated or continue with the next instruction instead. A trap, for example, is similar to a callq in its behavior and thus resumes with the subsequent instruction. A fault handler, on the other hand, is supposed to remove the barrier to success that caused the fault and thus generally re-runs the faulting instruction.

2.2.4 Example: Linux system calls

In Linux, almost all communications from user- to kernel-code are handled by a trap with exception number 128. Instead of using exception number to select the particular action the user is requesting of the kernel, that action is put in a number in %rax; this decision allows Linux to have space for billions of different system calls instead of just the 200 or so that would fit in an exception number and are not reserved by the hardware.

To see a list of Linux system calls, see man 2 syscall. Most of these are wrapped in their own library function, listed in man 2 syscalls

socket:
    endbr64 
    mov    $0x29,%eax
    syscall 

2.3 Exception-Like Constructs

2.3.1 Signals

One view of exceptions is that they enable communication from user code to the kernel. Signals permit the opposite direction of communication.

Signals are roughly the kernel-to-user equivalent of an interrupt. At any time, while executing any line of code, a signal may appear. It will suspend the currently executing code and see if you’ve told the kernel about a signal handler, a special function you believe will handle that signal. After running the handler, the interrupted code will resume.

Linux defines many different signals, listed in man 7 signal. Each of them has a default action if no handler is registered, most commonly terminating the process.

#include <stdio.h>
#include <signal.h>

static void handler(int signum) {
    printf("Provide end-of-file to end program.\n");
}

int main() {
    struct sigaction sa;       // how we tell the OS what we want to handle
    sa.sa_handler = handler;   // the function pointer to invoke
    sigemptyset(&sa.sa_mask);  // we'll use sigaction, not signal, so set not needed
    sa.sa_flags = SA_RESTART;  // restart functions that can recognize they were interrupted

    if (sigaction(SIGINT, &sa, NULL) == -1) {
        fprintf(stderr, "unable to override ^C signal");
        return 1;
    }

    char buf[1024];
    while(scanf("%1023s", buf) && !feof(stdin)) {
        printf("token: %s\n", buf);
    }
}

#include <stdio.h>
#include <signal.h>

static void handler(int signum) {
    printf("Ignoring segfault.\n");
}

int main() {
    struct sigaction sa;       // how we tell the OS what we want to handle
    sa.sa_handler = handler;   // the function pointer to invoke
    sigemptyset(&sa.sa_mask);  // we'll use sigaction, not signal, so set not needed
    sa.sa_flags = SA_RESTART;  // restart functions that can recognize they were interrupted

    if (sigaction(SIGSEGV, &sa, NULL) == -1) {
        fprintf(stderr, "unable to override segfault signal");
        return 1;
    }

    // let's make a segfault -- enters infinite handler loop
    char * buf;
    buf[1234567] = 89;
}

There is a system call named “kill” that asks the kernel to send a signal to a different process. While this can be used for inter-process communication, better systems (like sockets and pipes) exist for that if it is to be a major part of an application’s design.

2.3.2 setjmp, longjmp, and software exceptions

Portions of the save- and restore-state functionality used by exception handlers is exposed by the C library functions setjmp and longjmp.

setjump, given a jmp_buf as an argument, fills that structure with information needed to resume execution at that location and then returns 0. Thereafter longjmp can be called with that same jmp_buf as an argument; longjmp never returns, instead “returning” from setjmp for a second time. longjmp also provides an alternative return value for setjmp.

#include <setjmp.h>
#include <stdio.h>

jmp_buf env;
int n = 0;

void b() {
    printf("b1: n = %d\n", n);
    n += 1;
    printf("b2: n = %d\n", n);
    longjmp(env, 1);
    printf("b3: n = %d\n", n);
    n += 1;
    printf("b4: n = %d\n", n);
}

void a() {
    printf("a1: n = %d\n", n);
    n += 1;
    printf("a2: n = %d\n", n);
    b();
    printf("a3: n = %d\n", n);
    n += 1;
    printf("a4: n = %d\n", n);
}

int main() {
    printf("main1: n = %d\n", n);
    n += 1;
    printf("main2: n = %d\n", n);
    int got = setjmp(env);
    if (got) {
        printf("main3: n = %d\n", n);
        n += 1;
        printf("main4: n = %d\n", n);
    } else {
        printf("main5: n = %d\n", n);
        n += 1;
        printf("main6: n = %d\n", n);
        a();
    }
    printf("end of main\n");
}

main1: n = 0
main2: n = 1
main5: n = 1
main6: n = 2
a1: n = 2
a2: n = 3
b1: n = 3
b2: n = 4
main3: n = 4
main4: n = 5
end of main

There was a time when setjmp/longjmp were seen as effective ways of achieving try/catch-like error recovery, which cannot be handled with simple goto because it may skip multiple function call/returns.

try {
    f();
} catch {
    g();
}

if (!setjmp(env)) {
    f();
} else {
    g();
}

More efficient approaches to this have since been developed. setjmp records much of the state of the program in advance, which gives it a significant cost in time; because try is generally assumed to be executed far more often than throw. Most of the information stored by setjmp into the jmp_buf can be found somewhere in the setjmp-invoking function’s stack frame, which many languages maintain with sufficient discipline that it can be “unwound” to restore a previous state upon a throw.

3 Virtual Memory

One of the most important functions provided by a collaboration between the hardware and the operating system is virtual memory. Virtual memory provides

1. each process with the illusion that it is the only process running.
2. each process with the illusion that all possible addresses exist, no matter how much physical memory is available.
3. memory protection such that user-mode code cannot access kernel memory.
4. an efficient mechanism for communicating with processes

3.1 Processes

A process is a software-only operating systems construct that roughly parallels the end-user idea of a running program. Each process has its own virtual address space. The hardware knows the difference between user and kernel mode, but not the difference between a web browser and a text editor.

The operating system maintains a variety of data structures inside kernel-only memory to help track different processes. Portions of those structures relating to one process at a time¹ are loaded into hardware-accessible registers and memory so that the hardware mechanisms will cause running code to interact with the appropriate process’s memory and other state.

Changing which process’s state is currently loaded is called a context switch. All major user operating systems (but not all embedded-system operating systems) use a hardware timer to create an exception every few dozen milliseconds to enabled automated context switching and facilitate the illusion that more processes are running at a time than there are processors in the computer.

3.2 Regions of Memory

To provide process separation, each memory access needs to be checked to ensure it is permitted. Specific sets of permissions may vary by system, but a common set includes

• Read-only vs. read-write memory (e.g., is mov %rax, address allowed?)
• Kernel-only vs. user-mode memory (e.g., is mov address, %rax allowed if not in kernel mode?)
• Executable vs. data-only memory (e.g., is jmp address allowed?)
• Valid vs. invalid addresses (e.g., no code can use address 0)

Because most instructions would need to check at least one of these constraints, there is reason to encode them very efficiently. Because each process may access a very large number of memory addresses, they also need to be efficient in space, in practice meaning that permissions are applied to large contiguous regions of memory.

3.2.1 Segments revisited

A common operating system internal representation of memory regions is a list if segements, pairs of starting- and stopping-addresses and a set of permissions for the intermediate addresses.

Segments are never directly visible to the hardware: instead, the operating system uses the segments to create initialize hardware-visible pages table entries and to react to hardware-generated page-related faults and potentially convert them into signals to convey to the user process.

struct segment_node {
    void *low_address;
    void *high_address;
    unsigned int permissions_flags;
    struct segment_node *next;
}

POSIX-compatible operating systems provide library functions mmap and munmap for requesting the operating system adjust its segments. Typical user applications use higher-level wrappers such as malloc instead.

3.2.2 Pages

Hardware access to memory protections is designed to be very simple so that it can be very rapid. Each address is split into two parts: a page number made up of the high-order bits and a page offset made up of the low-order bits.

Each time memory is accessed in any way—including in fetching the instruction itself, meaning at least once per instruction—the address is translated. Translation converts an address as it appears in code (called a virtual address) into an address that can be used to access the memory chip (called a physical address). This translation involves the following steps:

1. The virtual page number is used as a key or index into a data structure called a page table to retrieve a page table entry (see Page Tables). Conceptually this is what Java would call a Map<VirtualPageNumber, PageTableEntry>. It is stored in memory, using physical not virtual addresses and its location is stored in a register that is changed on each context switch so that the operating system maintain a different page table for each process.

2. The page table entry contains a variety of protection or flag bits. In addition to the various read/write/execute/kernel permission bits, it may also contain an “unmapped” flag indicating that this virtual page has no corresponding physical page.

   If the access cannot proceed for any reason, a fault is generated (a page fault if unmapped, otherwise a protection fault). The fault handler should either modify the page table so that the address will work on a second attempt or kill the program.

3. If the necessary permissions are present, the page table entry contains a physical page number. This physical page number, concatenated with the page offset from the virtual address, is the physical address.

Hardware uses various optimizations to speed up this process, notably using the Translation Lookaside Buffer.

3.2.3 Protection… but side channels?

Address translation is designed so that the hardware guarantees that

• Only kernel-mode code can change the page table.
• All memory accesses are guarded by the page table.

Combined with careful coding of the operating system, it is intended that these provide full separation of processes, so that no two programs can see the contents of the other’s memory unless they have both asked the operating system to enable that.

Since 2017, it has become known that these protections are not quite a complete as they were once thought. Although these protections do ensure that one process cannot directly access another’s memory, they does not guarantee the absence of side-effects of memory access that another process can detect. For example, many hardware accelerations are based on the principle that what happens once is likely to happen again, so timing how long a memory access takes can leak some information about what other memory accesses other processes may have made.

This is one example of how a side channel might exist, and why it is challenging to reason about their absence. Securing primary channels is easy to think about, even if not always easy to do, because the data in question is known. That the timing of individual memory accesses was information that needed protection was not even known for decades after it became a potential side channel.

3.3 Page Tables

A page table is conceptually just a map between virtual page numbers (which are fixed-size integers) and page table entries (which are each a physical page number and a set of Boolean flags).

3.3.1 Single-level page tables

In the early days of virtual memory, when the entire addressable space was not large, a single array could suffice for implementing an efficient page table. A special kernel-access-only register, called the page table base register or PTBR, stores the address of this array.

Single-level page tables are inefficient for 32-bit addresses and basically untenable for 64-bit addresses.

3.3.2 Multi-level page tables

Multi-level page tables are a particular instance of a linked data structure known as a high-arity (or wide node) fixed-depth tree, which is an efficient way for storing large, sparse arrays particularly when entries are clustered. Since this is not a data structure that is generally taught in data structure courses, this section provides a full overview of the data structure itself before discussing how they are used for virutal memory.

3.3.2.1 Fixed-depth high-arity trees

A tree is a linked data structure where each node contains an array of pointers to additional nodes. The best-known tree types use arity-2 (also known as binary or width-2) trees, where the arrays are all 2 elements long, but trees with higher-arity are common when hardware is involved.

Many trees have variable depth: the number of steps from the root to the node containing the desired data may be different for different data. That design allows the tree to scale in size, but also mean that conditional checks are needed at each node. Alternatively, a tree can be designed with a fixed depth, with all data stored the same number of steps from the root. Fixed-depth trees typically store all data in the leaves of the tree, with separate data design for the internal nodes and the leaf nodes.

#define WIDTH 16

struct leaf {
    PAYLOAD data[WIDTH];
};
struct height1 {
    struct leaf *kids[WIDTH];
};
struct height2 {
    struct height1 *kids[WIDTH];
};
struct height3 {
    struct height2 *kids[WIDTH];
};

/* ... */

struct height3 *root = /* ... */

Fixed-depth fixed-width trees can be accessed using bit masks to treat like an array. If the width of each node is a power of two, this is done by splitting up the bits of the index, using the high-order bits for the first index and lower-order bits as the tree is navigated.

PAYLOAD arr(struct height3 *root, unsigned short index) {
    struct height2 *child1 = root  [(index>>12)&(WIDTH-1)];
    struct height1 *child2 = child1[(index>> 8)&(WIDTH-1)];
    struct leaf    *child3 = child2[(index>> 4)&(WIDTH-1)];
    return child3[index&(WIDTH-1)];
}

Such a structure is significantly less efficient than a simple array: it requires more memory than a simple array because of the additional pointers; it requires more time because it requires multiple memory accesses per element access. However, it can save significant space if the majority of the indexes have no value, as entire subtrees can be omitted.

PAYLOAD *arr(struct height3 *root, unsigned short index) {
    if (!root) return NULL;
    struct height2 *child1 = root  [(index>>12)&(WIDTH-1)];
    if (!child1) return NULL;
    struct height1 *child2 = child1[(index>> 8)&(WIDTH-1)];
    if (!child2) return NULL;
    struct leaf    *child3 = child2[(index>> 4)&(WIDTH-1)];
    if (!child3) return NULL;
    return &child3[index&(WIDTH-1)];
}

3.3.2.2 Multi-level page table implementation

Most programs access only a few megabytes of memory, allocated in a few contiguous chunks, out of an address space of 2⁶⁴ = 18,446,744,073,709,551,616 possible addresses. This is, in many ways, and ideal scenario for using fixed-depth high-arity trees to store a very sparse array of page table entries.

x86-64-compatible processors handle 64-bit addresses as follows:

• They completely ignore the top 16 bits of each virtual address. 2⁴⁸ = 281,474,976,710,656 is seen as more than a large enough address space for the current generation of software applications.

• They use 64-bit page table entries, with larger physical page numbers than virtual page numbers. This does not mean that all physical page numbers are used: it is the OS’s job, when creating new page table entries, to ensure that the physical page number picked corresponds to an actual address in available RAM.

• They pick a page size such that every page table node exactly fits in one page.

  The most common such page size is 4KB pages with 2⁹ PTE per node (9+9+9+9+12 = 48). Other page sizes also work; for example 256K pages with 2¹⁵ PTE per node (15+15+18 = 48) fits perfectly, and at a slight loss of space efficiency other sizes can be used as well.

• Each address translation that the TLB does not optimize runs a process similar to the following:

  size_t vpn[4] = split_bits(va>>12, 9);
  size_t ppn = PTBR;
  for(int i=0; i<4; i+=1) {
      PTE pte = ram[(ppn<<12) | (vpn[i]<<3)];
      if (pte.unmapped) page fault;
      if (pte.flags & current_usage) protection fault;
      ppn = pte.ppn;
  }
  pa = (ppn<<12) | (va & ((1<<12)-1));

All of the above is performed by hardware, having been compiled down to actual physical transistors. It is not programmable.

The programmable parts are the fault handlers. A page fault handler (written in the OS software) might do something like

handle_page_fault(size_t va, int access_type) {
    int flags = permitted_actions(va, segment_list)
    if ((access_type & flags) != access_type))
        send_signal(SIGSEGV);
    ssize_t ppn = get_unused_physical_page();
    if (ppn < 0) ppn = swap_page();
    create_page_table_entry(va, ppn, flags);
}

The above pseudocode uses several function stubs:

permitted_actions

Check the OS list of segments to see what kinds of access (reading, writing, executing, etc) the segment that contains this address allows. If this address is not in a segment, returns that no actions are allowed.

get_unused_physical_page

The OS maintains a list of all of the physical pages in RAM and which process is using each. The idea here is to pull one page from the “unused” list onto the “used by current process” list.

swap_page

In the event that get_unused_physical_page fails to find an unused page, the OS pick a page currently being used by some process and “swaps it out”: that is,

1. Picks a physical page
2. Writes its contents onto a region of disk known as the “swap space”
3. Changes the PTE for that page such that
   1. the unmapped bit is set
   2. the rest of the bits tell where on disk the contents are located
4. Change that page to be owned by the current process
5. Return the newly-freed physical page number

See Page Swapping for more on swapping.

create_page_table_entry

Ensures that the page table entry is in the page table (possibly creating intermediate nodes in the tree) and sets the appropriate permissions and physical page number in the PTE. If the PTE already existed and noted where in the swap space the old contents of this page were located, loads the swapped-out contents back from disk into memory too.

An important thing to note about the above: software (the operating system) writes the page table in a format that the hardware understands and can read.

3.3.3 Translation Lookaside Buffer

Address translation needs to happen for every instruction, at least once to fetch the instruction and a second time if the instruction contains a memory access, so making it fast is vital to keeping the computer efficient.

Modern processors use a special cache called the translation lookaside buffer (or TBL) to keep the most-used address translations in high-speed memory.

Conceptually, the TLB is used as follows:

size_t vpn_full = va>>12;
if (tlb_has(vpn_full)) {
    pte = tlb_lookup(vpn_full);
    if (pte.flags & current_usage) protection fault;
    ppn = pte.ppn;
} else {
    size_t vpn[4] = split_bits(vpn_full, 9);
    size_t ppn = PTBR;
    for(int i=0; i<4; i+=1) {
        pte = ram[(ppn<<12) | (vpn[i]<<3)];
        if (pte.unmapped) page fault;
        if (pte.flags & current_usage) protection fault;
        ppn = pte.ppn;
    }
    tlb_store(vpn_full, pte);
}
pa = (ppn<<12) | (va & ((1<<12)-1));

For more details on how caches like the TLB work, see Caching.

3.4 Usage

Various uses of virtual memory are alluded to above to motivate the various features of virtual memory; this section focuses on those uses.

3.4.1 Page Swapping

Because addresses in code can be mapped to other addresses, with operating system intervention if needed, virtual memory lets code act beyond the constraints of available memory. This is traditionally handled by the operating system storing some pages on the disk instead of in memory. moving these pages between disk and memory is called “swapping”.

Because disks are vastly slower than memory, there is a desire to make this happen rarely, meaning that operating systems sometimes devote significant effort to selecting pages to swap out of memory that it believes are likely not to be swapped back in again anytime soon. Predicting which pages will be needed next is undecidable², so this is always at best a heuristic and sometimes incorrect. If it is consistently incorrect (or there is not enough physical memory available for accuracy to matter) and the process begins spending more time swapping pages in and out than it does doing actual work, it is said that the process is thrashing.

As memory has become less expensive and more plentiful, swap is arguably³ less important than it once was, but it is still a part of every major end-user OS.

3.4.2 Shared Memory

One of the goals of virtual memory is to allow each process to have its own address space, but because this is done by having a mapping between virtual and physical pages, it is quite possible to map the same physical page into multiple different processes’ virtual address space—potentially even to have the same physical page mapped to different virtual pages in each process.

Shared memory, as this is called, is used for several purposes, notably including the following.

Shared libraries.

If several programs are going to load in the same shared library code, it can be more efficient to have its code in just one physical page shared in each process’s virtual address space.

Not all libraries are linked in as shared memory. because shared memory must come in in full page chunks, shared memory library linkage can be inefficient if library code is small or not shared by several running processes. Additionally, sharing library code requires all processes that use it to use the exact same version of the code; as library code is frequently updated, this can itself be a challenge to engineer around.

Static library code is linked during compilation, while shared (also called “dynamic”) library code is linked by the loader or in some cases even part-way through execution.

Kernel memory.

Consider the page of memory that contains the instruction that changes the PTBR. For the next instruction to be fetched properly, that instruction’s virtual page needs to map to the same physical page in both the old and new page table.

Different operating systems handle this differently; in Linux, for example, half of the entire address space (all with addresses 0x800000000000 or above) is reserved for the OS, and is mapped to the same set of physical pages in every process’s page table.

Shared-memory communication.

Processes can communicate with one another by each reading what others have written into a shared region of memory. This can be a very efficient way of sharing large quantities of data but does not, by itself, provide mechanisms for efficiently awaiting a response.

3.4.3 Process Switching

Switching between processes happens in the operating system’s code. When the OS was invoked (in an Exception Handler) the state of the original process was stored somewhere; when changing processes

• the saved user state is moved into an OS data structure to track the not-currently-running processes.
• the PTBR is changed to the page table of the new process and the user page table entries in the translation lookaside buffer are marked as invalid.
• the saved state from the OS data structure is moved back into the saved user state location to be restored when the exception handler returns.

As this happens in operating system software, other computation is often attached to it. It is not uncommon for the operating system to perform many different bookkeeping and other operations each time a context switch is performed.

3.4.4 Direct Memory Access

Physical memory pages can be given to systems other than software processes. It is common for various input/output operations to be performed by allocating a page (or more) of memory to be accessed directly by the I/O systems. This direct memory access model allows I/O to happen at the slower pace of I/O systems concurrently with the processor doing other work, and then once the transfer into or out of memory is completed the processor to access it at the higher speed of memory access. This model is called direct memory access (DMA).

Technically, DMA does not require virtual memory—disks could be given access to physical memory even if code access memory with physical addresses—but it is generally implemented as an extension of the virtual memory system.