get_user_pages() on an mmap()ed file allowed? What to do if 0 < get_user_pages(..., nr_pages, ...) < nr_pages?

[Leon Woestenberg <leon.woestenberg@gmail.com>]

Hello,


I have a PCI device driver performing DMA to a scattered user-space buffer.
Given a malloc()ed buffer, get_user_pages(..., buffer, nr_pages, ...)
always returns to requested number of pages and everything works as
expected.
So far so good.

Since that I changed userspace to mmap() a file, instead of
malloc()ing a buffer.
The mmap() in userspace works.

However, in the driver get_user_pages() starts to return less pages
than I requested, in an undeterministic fashion (most of the times I
get the expected number,
sometimes I get only a part of the requested pages).

Reading the get_user_pages() implementation dazzles me too much,
still.  I wonder if I am violating the kernel API?

- is it allowed to have a PCI device DMA-read from memory pages, that
belong to a file mmap()'d by userspace?
- what are valid reasons for get_user_pages() to fail?
- what should a driver do when get_user_pages() returns less pages
than requested?


A snippet of the code:


my user space does:

    int fd = open(filename, O_RDONLY);
    assert(fd >= 0);

    /* map the file in memory */
    char *buffer = mmap(0, buffersize, PROT_READ, MAP_SHARED, fd, 0);
    assert(buffer != MAP_FAILED);

    /* advice sequential access */
    int rc = madvise(buffer, buffersize, MADV_SEQUENTIAL);
    assert(rc == 0);

my driver does:

    const unsigned long first = (boe & PAGE_MASK) >> PAGE_SHIFT;
    const unsigned long last  = ((boe + count - 1) & PAGE_MASK) >> PAGE_SHIFT;
    const int nr_pages = last - first + 1;
    ...
    down_read(&current->mm->mmap_sem);
    rc = get_user_pages(current, current->mm, start & PAGE_MASK,
nr_pages, 0 /* do not write*/, 1 /* do force */, pages, NULL);
    up_read(&current->mm->mmap_sem);

    BUG_ON(rc < nr_pages);

Thanks,
-- 
Leon