如何将 Linux 内核缓冲区映射到用户空间？

Question

Let's say the buffer is allocated using a page based scheme. One way to implement mmap would be to use remap_pfn_range but LDD3 says this does not work for conventional memory. It appears we can work around this by marking the page(s) reserved using SetPageReserved so that it gets locked in memory. But isn't all kernel memory already non-swappable ie already reserved? Why the need to set the reserved bit explicitly?

Does this have something to do with pages allocated from HIGH_MEM?

Answer 1

The simplest way to map a set of pages from the kernel in your mmap method is to use the fault handler to map the pages. Basically you end up with something like:

static int my_mmap(struct file *filp, struct vm_area_struct *vma)
{
    vma->vm_ops = &my_vm_ops;
    return 0;
}

static const struct file_operations my_fops = {
    .owner  = THIS_MODULE,
    .open   = nonseekable_open,
    .mmap   = my_mmap,
    .llseek = no_llseek,
};

(where the other file operations are whatever your module needs). Also in my_mmap you do whatever range checking etc. is needed to validate the mmap parameters.

Then the vm_ops look like:

static int my_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
{
    vmf->page = my_page_at_index(vmf->pgoff);
    get_page(vmf->page);

    return 0;
} 

static const struct vm_operations_struct my_vm_ops = {
    .fault      = my_fault
}

where you just need to figure out for a given vma / vmf passed to your fault function which page to map into userspace. This depends on exactly how your module works. For example, if you did

my_buf = vmalloc_user(MY_BUF_SIZE);

then the page you use would be something like

vmalloc_to_page(my_buf + (vmf->pgoff << PAGE_SHIFT));

But you could easily create an array and allocate a page for each entry, use kmalloc, whatever.

[just noticed that my_fault is a slightly amusing name for a function]

Answer 2

Minimal runnable example and userland test

Kernel module:

#include <linux/fs.h>
#include <linux/init.h>
#include <linux/kernel.h> /* min */
#include <linux/mm.h>
#include <linux/module.h>
#include <linux/proc_fs.h>
#include <linux/uaccess.h> /* copy_from_user, copy_to_user */
#include <linux/slab.h>

static const char *filename = "lkmc_mmap";

enum { BUFFER_SIZE = 4 };

struct mmap_info {
    char *data;
};

/* After unmap. */
static void vm_close(struct vm_area_struct *vma)
{
    pr_info("vm_close\n");
}

/* First page access. */
static vm_fault_t vm_fault(struct vm_fault *vmf)
{
    struct page *page;
    struct mmap_info *info;

    pr_info("vm_fault\n");
    info = (struct mmap_info *)vmf->vma->vm_private_data;
    if (info->data) {
        page = virt_to_page(info->data);
        get_page(page);
        vmf->page = page;
    }
    return 0;
}

/* After mmap. TODO vs mmap, when can this happen at a different time than mmap? */
static void vm_open(struct vm_area_struct *vma)
{
    pr_info("vm_open\n");
}

static struct vm_operations_struct vm_ops =
{
    .close = vm_close,
    .fault = vm_fault,
    .open = vm_open,
};

static int mmap(struct file *filp, struct vm_area_struct *vma)
{
    pr_info("mmap\n");
    vma->vm_ops = &vm_ops;
    vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
    vma->vm_private_data = filp->private_data;
    vm_open(vma);
    return 0;
}

static int open(struct inode *inode, struct file *filp)
{
    struct mmap_info *info;

    pr_info("open\n");
    info = kmalloc(sizeof(struct mmap_info), GFP_KERNEL);
    pr_info("virt_to_phys = 0x%llx\n", (unsigned long long)virt_to_phys((void *)info));
    info->data = (char *)get_zeroed_page(GFP_KERNEL);
    memcpy(info->data, "asdf", BUFFER_SIZE);
    filp->private_data = info;
    return 0;
}

static ssize_t read(struct file *filp, char __user *buf, size_t len, loff_t *off)
{
    struct mmap_info *info;
    ssize_t ret;

    pr_info("read\n");
    if ((size_t)BUFFER_SIZE <= *off) {
        ret = 0;
    } else {
        info = filp->private_data;
        ret = min(len, (size_t)BUFFER_SIZE - (size_t)*off);
        if (copy_to_user(buf, info->data + *off, ret)) {
            ret = -EFAULT;
        } else {
            *off += ret;
        }
    }
    return ret;
}

static ssize_t write(struct file *filp, const char __user *buf, size_t len, loff_t *off)
{
    struct mmap_info *info;

    pr_info("write\n");
    info = filp->private_data;
    if (copy_from_user(info->data, buf, min(len, (size_t)BUFFER_SIZE))) {
        return -EFAULT;
    } else {
        return len;
    }
}

static int release(struct inode *inode, struct file *filp)
{
    struct mmap_info *info;

    pr_info("release\n");
    info = filp->private_data;
    free_page((unsigned long)info->data);
    kfree(info);
    filp->private_data = NULL;
    return 0;
}

static const struct file_operations fops = {
    .mmap = mmap,
    .open = open,
    .release = release,
    .read = read,
    .write = write,
};

static int myinit(void)
{
    proc_create(filename, 0, NULL, &fops);
    return 0;
}

static void myexit(void)
{
    remove_proc_entry(filename, NULL);
}

module_init(myinit)
module_exit(myexit)
MODULE_LICENSE("GPL");

GitHub upstream .

Userland test:

#define _XOPEN_SOURCE 700
#include <assert.h>
#include <fcntl.h>
#include <stdio.h>
#include <stdlib.h>
#include <stdint.h> /* uintmax_t */
#include <string.h>
#include <sys/mman.h>
#include <unistd.h> /* sysconf */

/* Format documented at:
 * https://github.com/torvalds/linux/blob/v4.9/Documentation/vm/pagemap.txt
 */
typedef struct {
    uint64_t pfn : 54;
    unsigned int soft_dirty : 1;
    unsigned int file_page : 1;
    unsigned int swapped : 1;
    unsigned int present : 1;
} PagemapEntry;

/* Parse the pagemap entry for the given virtual address.
 *
 * @param[out] entry      the parsed entry
 * @param[in]  pagemap_fd file descriptor to an open /proc/pid/pagemap file
 * @param[in]  vaddr      virtual address to get entry for
 * @return                0 for success, 1 for failure
 */
int pagemap_get_entry(PagemapEntry *entry, int pagemap_fd, uintptr_t vaddr)
{
    size_t nread;
    ssize_t ret;
    uint64_t data;

    nread = 0;
    while (nread < sizeof(data)) {
        ret = pread(pagemap_fd, ((uint8_t*)&data) + nread, sizeof(data),
                (vaddr / sysconf(_SC_PAGE_SIZE)) * sizeof(data) + nread);
        nread += ret;
        if (ret <= 0) {
            return 1;
        }
    }
    entry->pfn = data & (((uint64_t)1 << 54) - 1);
    entry->soft_dirty = (data >> 54) & 1;
    entry->file_page = (data >> 61) & 1;
    entry->swapped = (data >> 62) & 1;
    entry->present = (data >> 63) & 1;
    return 0;
}

/* Convert the given virtual address to physical using /proc/PID/pagemap.
 *
 * @param[out] paddr physical address
 * @param[in]  pid   process to convert for
 * @param[in] vaddr  virtual address to get entry for
 * @return           0 for success, 1 for failure
 */
int virt_to_phys_user(uintptr_t *paddr, pid_t pid, uintptr_t vaddr)
{
    char pagemap_file[BUFSIZ];
    int pagemap_fd;

    snprintf(pagemap_file, sizeof(pagemap_file), "/proc/%ju/pagemap", (uintmax_t)pid);
    pagemap_fd = open(pagemap_file, O_RDONLY);
    if (pagemap_fd < 0) {
        return 1;
    }
    PagemapEntry entry;
    if (pagemap_get_entry(&entry, pagemap_fd, vaddr)) {
        return 1;
    }
    close(pagemap_fd);
    *paddr = (entry.pfn * sysconf(_SC_PAGE_SIZE)) + (vaddr % sysconf(_SC_PAGE_SIZE));
    return 0;
}

enum { BUFFER_SIZE = 4 };

int main(int argc, char **argv)
{
    int fd;
    long page_size;
    char *address1, *address2;
    char buf[BUFFER_SIZE];
    uintptr_t paddr;

    if (argc < 2) {
        printf("Usage: %s <mmap_file>\n", argv[0]);
        return EXIT_FAILURE;
    }
    page_size = sysconf(_SC_PAGE_SIZE);
    printf("open pathname = %s\n", argv[1]);
    fd = open(argv[1], O_RDWR | O_SYNC);
    if (fd < 0) {
        perror("open");
        assert(0);
    }
    printf("fd = %d\n", fd);

    /* mmap twice for double fun. */
    puts("mmap 1");
    address1 = mmap(NULL, page_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
    if (address1 == MAP_FAILED) {
        perror("mmap");
        assert(0);
    }
    puts("mmap 2");
    address2 = mmap(NULL, page_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
    if (address2 == MAP_FAILED) {
        perror("mmap");
        return EXIT_FAILURE;
    }
    assert(address1 != address2);

    /* Read and modify memory. */
    puts("access 1");
    assert(!strcmp(address1, "asdf"));
    /* vm_fault */
    puts("access 2");
    assert(!strcmp(address2, "asdf"));
    /* vm_fault */
    strcpy(address1, "qwer");
    /* Also modified. So both virtual addresses point to the same physical address. */
    assert(!strcmp(address2, "qwer"));

    /* Check that the physical addresses are the same.
     * They are, but TODO why virt_to_phys on kernel gives a different value? */
    assert(!virt_to_phys_user(&paddr, getpid(), (uintptr_t)address1));
    printf("paddr1 = 0x%jx\n", (uintmax_t)paddr);
    assert(!virt_to_phys_user(&paddr, getpid(), (uintptr_t)address2));
    printf("paddr2 = 0x%jx\n", (uintmax_t)paddr);

    /* Check that modifications made from userland are also visible from the kernel. */
    read(fd, buf, BUFFER_SIZE);
    assert(!memcmp(buf, "qwer", BUFFER_SIZE));

    /* Modify the data from the kernel, and check that the change is visible from userland. */
    write(fd, "zxcv", 4);
    assert(!strcmp(address1, "zxcv"));
    assert(!strcmp(address2, "zxcv"));

    /* Cleanup. */
    puts("munmap 1");
    if (munmap(address1, page_size)) {
        perror("munmap");
        assert(0);
    }
    puts("munmap 2");
    if (munmap(address2, page_size)) {
        perror("munmap");
        assert(0);
    }
    puts("close");
    close(fd);
    return EXIT_SUCCESS;
}

GitHub upstream .

Tested on kernel 5.4.3.

Answer 3

Though the pages are reserved via a kernel driver, it is meant to be accessed via user space. As a result, the PTE (page table entries) do not know if the pfn belongs to user space or kernel space (even though they are allocated via kernel driver).

This is why they are marked with SetPageReserved .