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 i.e. 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]