userfaultfd - Man Page

The userfaultfd mechanism is designed to allow a thread in a multithreaded program to perform user-space paging for the other threads in the process. When a page fault occurs for one of the regions registered to the userfaultfd object, the faulting thread is put to sleep and an event is generated that can be read via the userfaultfd file descriptor. The fault-handling thread reads events from this file descriptor and services them using the operations described in ioctl_userfaultfd(2). When servicing the page fault events, the fault-handling thread can trigger a wake-up for the sleeping thread.

It is possible for the faulting threads and the fault-handling threads to run in the context of different processes. In this case, these threads may belong to different programs, and the program that executes the faulting threads will not necessarily cooperate with the program that handles the page faults. In such non-cooperative mode, the process that monitors userfaultfd and handles page faults needs to be aware of the changes in the virtual memory layout of the faulting process to avoid memory corruption.

Since Linux 4.11, userfaultfd can also notify the fault-handling threads about changes in the virtual memory layout of the faulting process. In addition, if the faulting process invokes fork(2), the userfaultfd objects associated with the parent may be duplicated into the child process and the userfaultfd monitor will be notified (via the UFFD_EVENT_FORK described below) about the file descriptor associated with the userfault objects created for the child process, which allows the userfaultfd monitor to perform user-space paging for the child process. Unlike page faults which have to be synchronous and require an explicit or implicit wakeup, all other events are delivered asynchronously and the non-cooperative process resumes execution as soon as the userfaultfd manager executes read(2). The userfaultfd manager should carefully synchronize calls to UFFDIO_COPY with the processing of events.

The current asynchronous model of the event delivery is optimal for single threaded non-cooperative userfaultfd manager implementations.

Since Linux 5.7, userfaultfd is able to do synchronous page dirty tracking using the new write-protect register mode. One should check against the feature bit UFFD_FEATURE_PAGEFAULT_FLAG_WP before using this feature. Similar to the original userfaultfd missing mode, the write-protect mode will generate a userfaultfd notification when the protected page is written. The user needs to resolve the page fault by unprotecting the faulted page and kicking the faulted thread to continue. For more information, please refer to the "Userfaultfd write-protect mode" section.

After the userfaultfd object is created with userfaultfd(), the application must enable it using the UFFDIO_API ioctl(2) operation. This operation allows a two-step handshake between the kernel and user space to determine what API version and features the kernel supports, and then to enable those features user space wants. This operation must be performed before any of the other ioctl(2) operations described below (or those operations fail with the EINVAL error).

After a successful UFFDIO_API operation, the application then registers memory address ranges using the UFFDIO_REGISTER ioctl(2) operation. After successful completion of a UFFDIO_REGISTER operation, a page fault occurring in the requested memory range, and satisfying the mode defined at the registration time, will be forwarded by the kernel to the user-space application. The application can then use various (e.g., UFFDIO_COPY, UFFDIO_ZEROPAGE, or UFFDIO_CONTINUE) ioctl(2) operations to resolve the page fault.

Since Linux 4.14, if the application sets the UFFD_FEATURE_SIGBUS feature bit using the UFFDIO_API ioctl(2), no page-fault notification will be forwarded to user space. Instead a SIGBUS signal is delivered to the faulting process. With this feature, userfaultfd can be used for robustness purposes to simply catch any access to areas within the registered address range that do not have pages allocated, without having to listen to userfaultfd events. No userfaultfd monitor will be required for dealing with such memory accesses. For example, this feature can be useful for applications that want to prevent the kernel from automatically allocating pages and filling holes in sparse files when the hole is accessed through a memory mapping.

The UFFD_FEATURE_SIGBUS feature is implicitly inherited through fork(2) if used in combination with UFFD_FEATURE_FORK.

Details of the various ioctl(2) operations can be found in ioctl_userfaultfd(2).

Since Linux 4.11, events other than page-fault may enabled during UFFDIO_API operation.

Up to Linux 4.11, userfaultfd can be used only with anonymous private memory mappings. Since Linux 4.11, userfaultfd can be also used with hugetlbfs and shared memory mappings.

Since Linux 5.7, userfaultfd supports write-protect mode for anonymous memory. The user needs to first check availability of this feature using UFFDIO_API ioctl against the feature bit UFFD_FEATURE_PAGEFAULT_FLAG_WP before using this feature.

Since Linux 5.19, the write-protection mode was also supported on shmem and hugetlbfs memory types. It can be detected with the feature bit UFFD_FEATURE_WP_HUGETLBFS_SHMEM.

To register with userfaultfd write-protect mode, the user needs to initiate the UFFDIO_REGISTER ioctl with mode UFFDIO_REGISTER_MODE_WP set. Note that it is legal to monitor the same memory range with multiple modes. For example, the user can do UFFDIO_REGISTER with the mode set to UFFDIO_REGISTER_MODE_MISSING | UFFDIO_REGISTER_MODE_WP. When there is only UFFDIO_REGISTER_MODE_WP registered, user-space will not receive any notification when a missing page is written. Instead, user-space will receive a write-protect page-fault notification only when an existing but write-protected page got written.

After the UFFDIO_REGISTER ioctl completed with UFFDIO_REGISTER_MODE_WP mode set, the user can write-protect any existing memory within the range using the ioctl UFFDIO_WRITEPROTECT where uffdio_writeprotect.mode should be set to UFFDIO_WRITEPROTECT_MODE_WP.

When a write-protect event happens, user-space will receive a page-fault notification whose uffd_msg.pagefault.flags will be with UFFD_PAGEFAULT_FLAG_WP flag set. Note: since only writes can trigger this kind of fault, write-protect notifications will always have the UFFD_PAGEFAULT_FLAG_WRITE bit set along with the UFFD_PAGEFAULT_FLAG_WP bit.

To resolve a write-protection page fault, the user should initiate another UFFDIO_WRITEPROTECT ioctl, whose uffd_msg.pagefault.flags should have the flag UFFDIO_WRITEPROTECT_MODE_WP cleared upon the faulted page or range.

Since Linux 5.13, userfaultfd supports minor fault mode. In this mode, fault messages are produced not for major faults (where the page was missing), but rather for minor faults, where a page exists in the page cache, but the page table entries are not yet present. The user needs to first check availability of this feature using the UFFDIO_API ioctl with the appropriate feature bits set before using this feature: UFFD_FEATURE_MINOR_HUGETLBFS since Linux 5.13, or UFFD_FEATURE_MINOR_SHMEM since Linux 5.14.

To register with userfaultfd minor fault mode, the user needs to initiate the UFFDIO_REGISTER ioctl with mode UFFD_REGISTER_MODE_MINOR set.

When a minor fault occurs, user-space will receive a page-fault notification whose uffd_msg.pagefault.flags will have the UFFD_PAGEFAULT_FLAG_MINOR flag set.

To resolve a minor page fault, the handler should decide whether or not the existing page contents need to be modified first. If so, this should be done in-place via a second, non-userfaultfd-registered mapping to the same backing page (e.g., by mapping the shmem or hugetlbfs file twice). Once the page is considered "up to date", the fault can be resolved by initiating an UFFDIO_CONTINUE ioctl, which installs the page table entries and (by default) wakes up the faulting thread(s).

Minor fault mode supports only hugetlbfs-backed (since Linux 5.13) and shmem-backed (since Linux 5.14) memory.

Each read(2) from the userfaultfd file descriptor returns one or more uffd_msg structures, each of which describes a page-fault event or an event required for the non-cooperative userfaultfd usage:

struct uffd_msg {
    __u8  event;            /* Type of event */
    ...
    union {
        struct {
            __u64 flags;    /* Flags describing fault */
            __u64 address;  /* Faulting address */
            union {
                __u32 ptid; /* Thread ID of the fault */
            } feat;
        } pagefault;

        struct {            /* Since Linux 4.11 */
            __u32 ufd;      /* Userfault file descriptor
                               of the child process */
        } fork;

        struct {            /* Since Linux 4.11 */
            __u64 from;     /* Old address of remapped area */
            __u64 to;       /* New address of remapped area */
            __u64 len;      /* Original mapping length */
        } remap;

        struct {            /* Since Linux 4.11 */
            __u64 start;    /* Start address of removed area */
            __u64 end;      /* End address of removed area */
        } remove;
        ...
    } arg;

    /* Padding fields omitted */
} __packed;

If multiple events are available and the supplied buffer is large enough, read(2) returns as many events as will fit in the supplied buffer. If the buffer supplied to read(2) is smaller than the size of the uffd_msg structure, the read(2) fails with the error EINVAL.

The fields set in the uffd_msg structure are as follows:

event

The type of event. Depending of the event type, different fields of the arg union represent details required for the event processing. The non-page-fault events are generated only when appropriate feature is enabled during API handshake with UFFDIO_API ioctl(2).

The following values can appear in the event field:

UFFD_EVENT_PAGEFAULT (since Linux 4.3)

A page-fault event. The page-fault details are available in the pagefault field.

UFFD_EVENT_FORK (since Linux 4.11)

Generated when the faulting process invokes fork(2) (or clone(2) without the CLONE_VM flag). The event details are available in the fork field.

UFFD_EVENT_REMAP (since Linux 4.11)

Generated when the faulting process invokes mremap(2). The event details are available in the remap field.

UFFD_EVENT_REMOVE (since Linux 4.11)

Generated when the faulting process invokes madvise(2) with MADV_DONTNEED or MADV_REMOVE advice. The event details are available in the remove field.

UFFD_EVENT_UNMAP (since Linux 4.11)

Generated when the faulting process unmaps a memory range, either explicitly using munmap(2) or implicitly during mmap(2) or mremap(2). The event details are available in the remove field.

pagefault.address

The address that triggered the page fault.

pagefault.flags

A bit mask of flags that describe the event. For UFFD_EVENT_PAGEFAULT, the following flag may appear:

UFFD_PAGEFAULT_FLAG_WP

If this flag is set, then the fault was a write-protect fault.

UFFD_PAGEFAULT_FLAG_MINOR

If this flag is set, then the fault was a minor fault.

UFFD_PAGEFAULT_FLAG_WRITE

If this flag is set, then the fault was a write fault.

If neither UFFD_PAGEFAULT_FLAG_WP nor UFFD_PAGEFAULT_FLAG_MINOR are set, then the fault was a missing fault.

pagefault.feat.pid

The thread ID that triggered the page fault.

fork.ufd

The file descriptor associated with the userfault object created for the child created by fork(2).

remap.from

The original address of the memory range that was remapped using mremap(2).

remap.to

The new address of the memory range that was remapped using mremap(2).

remap.len

The original length of the memory range that was remapped using mremap(2).

remove.start

The start address of the memory range that was freed using madvise(2) or unmapped

remove.end

The end address of the memory range that was freed using madvise(2) or unmapped

A read(2) on a userfaultfd file descriptor can fail with the following errors:

EINVAL

The userfaultfd object has not yet been enabled using the UFFDIO_API ioctl(2) operation

If the O_NONBLOCK flag is enabled in the associated open file description, the userfaultfd file descriptor can be monitored with poll(2), select(2), and epoll(7). When events are available, the file descriptor indicates as readable. If the O_NONBLOCK flag is not enabled, then poll(2) (always) indicates the file as having a POLLERR condition, and select(2) indicates the file descriptor as both readable and writable.

/* userfaultfd_demo.c

   Licensed under the GNU General Public License version 2 or later.
*/
#define _GNU_SOURCE
#include <err.h>
#include <errno.h>
#include <fcntl.h>
#include <inttypes.h>
#include <linux/userfaultfd.h>
#include <poll.h>
#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <sys/syscall.h>
#include <unistd.h>

static int page_size;

static void *
fault_handler_thread(void *arg)
{
    int                 nready;
    long                uffd;   /* userfaultfd file descriptor */
    ssize_t             nread;
    struct pollfd       pollfd;
    struct uffdio_copy  uffdio_copy;

    static int      fault_cnt = 0; /* Number of faults so far handled */
    static char     *page = NULL;
    static struct uffd_msg  msg;  /* Data read from userfaultfd */

    uffd = (long) arg;

    /* Create a page that will be copied into the faulting region. */

    if (page == NULL) {
        page = mmap(NULL, page_size, PROT_READ | PROT_WRITE,
                    MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
        if (page == MAP_FAILED)
            err(EXIT_FAILURE, "mmap");
    }

    /* Loop, handling incoming events on the userfaultfd
       file descriptor. */

    for (;;) {

        /* See what poll() tells us about the userfaultfd. */

        pollfd.fd = uffd;
        pollfd.events = POLLIN;
        nready = poll(&pollfd, 1, -1);
        if (nready == -1)
            err(EXIT_FAILURE, "poll");

        printf("\nfault_handler_thread():\n");
        printf("    poll() returns: nready = %d; "
               "POLLIN = %d; POLLERR = %d\n", nready,
               (pollfd.revents & POLLIN) != 0,
               (pollfd.revents & POLLERR) != 0);

        /* Read an event from the userfaultfd. */

        nread = read(uffd, &msg, sizeof(msg));
        if (nread == 0) {
            printf("EOF on userfaultfd!\n");
            exit(EXIT_FAILURE);
        }

        if (nread == -1)
            err(EXIT_FAILURE, "read");

        /* We expect only one kind of event; verify that assumption. */

        if (msg.event != UFFD_EVENT_PAGEFAULT) {
            fprintf(stderr, "Unexpected event on userfaultfd\n");
            exit(EXIT_FAILURE);
        }

        /* Display info about the page-fault event. */

        printf("    UFFD_EVENT_PAGEFAULT event: ");
        printf("flags = %"PRIx64"; ", msg.arg.pagefault.flags);
        printf("address = %"PRIx64"\n", msg.arg.pagefault.address);

        /* Copy the page pointed to by 'page' into the faulting
           region. Vary the contents that are copied in, so that it
           is more obvious that each fault is handled separately. */

        memset(page, 'A' + fault_cnt % 20, page_size);
        fault_cnt++;

        uffdio_copy.src = (unsigned long) page;

        /* We need to handle page faults in units of pages(!).
           So, round faulting address down to page boundary. */

        uffdio_copy.dst = (unsigned long) msg.arg.pagefault.address &
                                           ~(page_size - 1);
        uffdio_copy.len = page_size;
        uffdio_copy.mode = 0;
        uffdio_copy.copy = 0;
        if (ioctl(uffd, UFFDIO_COPY, &uffdio_copy) == -1)
            err(EXIT_FAILURE, "ioctl-UFFDIO_COPY");

        printf("        (uffdio_copy.copy returned %"PRId64")\n",
               uffdio_copy.copy);
    }
}

int
main(int argc, char *argv[])
{
    int        s;
    char       c;
    char       *addr;   /* Start of region handled by userfaultfd */
    long       uffd;    /* userfaultfd file descriptor */
    size_t     len, l;  /* Length of region handled by userfaultfd */
    pthread_t  thr;     /* ID of thread that handles page faults */
    struct uffdio_api       uffdio_api;
    struct uffdio_register  uffdio_register;

    if (argc != 2) {
        fprintf(stderr, "Usage: %s num-pages\n", argv[0]);
        exit(EXIT_FAILURE);
    }

    page_size = sysconf(_SC_PAGE_SIZE);
    len = strtoull(argv[1], NULL, 0) * page_size;

    /* Create and enable userfaultfd object. */

    uffd = syscall(SYS_userfaultfd, O_CLOEXEC | O_NONBLOCK);
    if (uffd == -1)
        err(EXIT_FAILURE, "userfaultfd");

    /* NOTE: Two-step feature handshake is not needed here, since this
       example doesn't require any specific features.

       Programs that *do* should call UFFDIO_API twice: once with
       `features = 0` to detect features supported by this kernel, and
       again with the subset of features the program actually wants to
       enable. */
    uffdio_api.api = UFFD_API;
    uffdio_api.features = 0;
    if (ioctl(uffd, UFFDIO_API, &uffdio_api) == -1)
        err(EXIT_FAILURE, "ioctl-UFFDIO_API");

    /* Create a private anonymous mapping. The memory will be
       demand-zero paged--that is, not yet allocated. When we
       actually touch the memory, it will be allocated via
       the userfaultfd. */

    addr = mmap(NULL, len, PROT_READ | PROT_WRITE,
                MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
    if (addr == MAP_FAILED)
        err(EXIT_FAILURE, "mmap");

    printf("Address returned by mmap() = %p\n", addr);

    /* Register the memory range of the mapping we just created for
       handling by the userfaultfd object. In mode, we request to track
       missing pages (i.e., pages that have not yet been faulted in). */

    uffdio_register.range.start = (unsigned long) addr;
    uffdio_register.range.len = len;
    uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
    if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register) == -1)
        err(EXIT_FAILURE, "ioctl-UFFDIO_REGISTER");

    /* Create a thread that will process the userfaultfd events. */

    s = pthread_create(&thr, NULL, fault_handler_thread, (void *) uffd);
    if (s != 0) {
        errc(EXIT_FAILURE, s, "pthread_create");
    }

    /* Main thread now touches memory in the mapping, touching
       locations 1024 bytes apart. This will trigger userfaultfd
       events for all pages in the region. */

    l = 0xf;    /* Ensure that faulting address is not on a page
                   boundary, in order to test that we correctly
                   handle that case in fault_handling_thread(). */
    while (l < len) {
        c = addr[l];
        printf("Read address %p in %s(): ", addr + l, __func__);
        printf("%c\n", c);
        l += 1024;
        usleep(100000);         /* Slow things down a little */
    }

    exit(EXIT_SUCCESS);
}