Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1272394

Re: [PATCH 02/37] mm, frame_vector: do not use get_user_pages_locked()

Show all headers | View raw

On Wed, Nov 18, 2015 at 01:29:38PM +0100, Jan Kara wrote:
> On Mon 16-11-15 19:35:14, Dave Hansen wrote:
> > 
> > From: Dave Hansen <dave.hansen@linux.intel.com>
> > 
> > get_user_pages_locked() appears to be for use when a caller needs
> > to know that its lock on mmap_sem was invalidated by the gup
> > call.
> > 
> > But, get_vaddr_frames() is not one of those users.  It
> > unconditionally locks the mmap_sem and unconditionally unlocks it
> > after the gup call.  It takes no special action and does not need
> > to know whether its lock was invalidated or not.
> > 
> > Replace get_user_pages_locked() with a vanilla get_user_pages()
> > and save a few lines of code.
> > 
> > Note that this was the *ONLY* use of get_user_pages_locked() in
> > the entire kernel tree.
> 
> I've used get_user_pages_locked() because of a comment before that function
> saying:
> 
>  * We can leverage the VM_FAULT_RETRY functionality in the page fault
>  * paths better by using either get_user_pages_locked() or
>  * get_user_pages_unlocked().
>  *
>  * get_user_pages_locked() is suitable to replace the form:
>  *
>  *      down_read(&mm->mmap_sem);
>  *      do_something()
>  *      get_user_pages(tsk, mm, ..., pages, NULL);
>  *      up_read(&mm->mmap_sem);
>  *
>  *  to:
>  *
>  *      int locked = 1;
>  *      down_read(&mm->mmap_sem);
>  *      do_something()
>  *      get_user_pages_locked(tsk, mm, ..., pages, &locked);
>  *      if (locked)
>  *          up_read(&mm->mmap_sem);
> 
> So I understood it as a way to reduce mmap_sem hold time by doing a try
> first. Did I understand that comment wrong?

That is correct. get_user_pages_locked should not be downgraded to
get_user_pages or it can actually break userfaultfd, as userfaultfd
needs to be allowed to drop the mmap_sem within handle_mm_fault, to
function correctly. Furthermore get_user_pages_locked allows for
higher SMP scalability as it can take advantage of the optimized
pagecache code that drops the mmap_sem before waiting for I/O (and
then it retries the page fault after the I/O is complete).

The only case where get_user_pages without FOLL_FAULT_ALLOW_RETRY
makes any sense is the get_dump_page case which is actually not using
get_user_pages in the first place, so ideally get_user_pages should be
obsoleted and dropped.

Thanks,
Andrea
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Back to linux.kernel | Previous | Next — Previous in thread | Find similar | Unroll thread

Thread

[PATCH 02/37] mm, frame_vector: do not use get_user_pages_locked() Dave Hansen <dave@sr71.net> - 2015-11-17 04:50 +0100
  Re: [PATCH 02/37] mm, frame_vector: do not use  get_user_pages_locked() Jan Kara <jack@suse.cz> - 2015-11-18 13:30 +0100
    Re: [PATCH 02/37] mm, frame_vector: do not use  get_user_pages_locked() Andrea Arcangeli <aarcange@redhat.com> - 2015-11-18 18:10 +0100

csiph-web