lkml.org 
[lkml]   [2019]   [Apr]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 1/2] mm/gup.c: fix the wrong comments
On Tue, Apr 09, 2019 at 09:08:33AM +0800, Huang Shijie wrote:
> On Mon, Apr 08, 2019 at 07:13:13AM -0700, Matthew Wilcox wrote:
> > On Mon, Apr 08, 2019 at 10:37:45AM +0800, Huang Shijie wrote:
> > > When CONFIG_HAVE_GENERIC_GUP is defined, the kernel will use its own
> > > get_user_pages_fast().
> > >
> > > In the following scenario, we will may meet the bug in the DMA case:
> > > .....................
> > > get_user_pages_fast(start,,, pages);
> > > ......
> > > sg_alloc_table_from_pages(, pages, ...);
> > > .....................
> > >
> > > The root cause is that sg_alloc_table_from_pages() requires the
> > > page order to keep the same as it used in the user space, but
> > > get_user_pages_fast() will mess it up.
> >
> > I don't understand how get_user_pages_fast() can return the pages in a
> > different order in the array from the order they appear in userspace.
> > Can you explain?
> Please see the code in gup.c:
>
> int get_user_pages_fast(unsigned long start, int nr_pages,
> unsigned int gup_flags, struct page **pages)
> {
> .......
> if (gup_fast_permitted(start, nr_pages)) {
> local_irq_disable();
> gup_pgd_range(addr, end, gup_flags, pages, &nr); // The @pages array maybe filled at the first time.
> local_irq_enable();
> ret = nr;
> }
> .......
> if (nr < nr_pages) {
> /* Try to get the remaining pages with get_user_pages */
> start += nr << PAGE_SHIFT;
> pages += nr; // The @pages is moved forward.
>
> if (gup_flags & FOLL_LONGTERM) {
> down_read(&current->mm->mmap_sem);
> ret = __gup_longterm_locked(current, current->mm, // The @pages maybe filled at the second time
>

Neither this nor the get_user_pages_unlocked is filling the pages a second
time. It is adding to the page array having moved start and the page array
forward.

Are you doing a FOLL_LONGTERM GUP? Or are you in the else clause below when
you get this bug?

Ira

> start, nr_pages - nr,
> pages, NULL, gup_flags);
> up_read(&current->mm->mmap_sem);
> } else {
> /*
> * retain FAULT_FOLL_ALLOW_RETRY optimization if
> * possible
> */
> ret = get_user_pages_unlocked(start, nr_pages - nr, // The @pages maybe filled at the second time.
> pages, gup_flags);
> }
> }
>
>
> .....................
>
> BTW, I do not know why we mess up the page order. It maybe used in some special case.
>
> Thanks
> Huang Shijie
>

\
 
 \ /
  Last update: 2019-04-09 22:23    [W:0.733 / U:0.176 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site