linux-kernel - Re: [PATCH] mm: extend prefault helpers to fault in more than PAGE

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:	Fri, 24 Feb 2012 14:34:31 +0100
From:	Daniel Vetter <daniel@...ll.ch>
To:	Andrew Morton <akpm@...ux-foundation.org>
Cc:	Daniel Vetter <daniel.vetter@...ll.ch>,
	Intel Graphics Development <intel-gfx@...ts.freedesktop.org>,
	DRI Development <dri-devel@...ts.freedesktop.org>,
	LKML <linux-kernel@...r.kernel.org>,
	Linux MM <linux-mm@...ck.org>
Subject: Re: [PATCH] mm: extend prefault helpers to fault in more than
 PAGE_SIZE

On Thu, Feb 23, 2012 at 02:36:58PM -0800, Andrew Morton wrote:
> On Thu, 16 Feb 2012 13:01:36 +0100
> Daniel Vetter <daniel.vetter@...ll.ch> wrote:
> 
> > drm/i915 wants to read/write more than one page in its fastpath
> > and hence needs to prefault more than PAGE_SIZE bytes.
> > 
> > I've checked the callsites and they all already clamp size when
> > calling fault_in_pages_* to the same as for the subsequent
> > __copy_to|from_user and hence don't rely on the implicit clamping
> > to PAGE_SIZE.
> > 
> > Also kill a copy&pasted spurious space in both functions while at it.
> >
> > ...
> >
> > --- a/include/linux/pagemap.h
> > +++ b/include/linux/pagemap.h
> > @@ -408,6 +408,7 @@ extern void add_page_wait_queue(struct page *page, wait_queue_t *waiter);
> >  static inline int fault_in_pages_writeable(char __user *uaddr, int size)
> >  {
> >  	int ret;
> > +	char __user *end = uaddr + size - 1;
> >  
> >  	if (unlikely(size == 0))
> >  		return 0;
> > @@ -416,17 +417,20 @@ static inline int fault_in_pages_writeable(char __user *uaddr, int size)
> >  	 * Writing zeroes into userspace here is OK, because we know that if
> >  	 * the zero gets there, we'll be overwriting it.
> >  	 */
> > -	ret = __put_user(0, uaddr);
> > +	while (uaddr <= end) {
> > +		ret = __put_user(0, uaddr);
> > +		if (ret != 0)
> > +			return ret;
> > +		uaddr += PAGE_SIZE;
> > +	}
> 
> The callsites in filemap.c are pretty hot paths, which is why this
> thing remains explicitly inlined.  I think it would be worth adding a
> bit of code here to avoid adding a pointless test-n-branch and larger
> cache footprint to read() and write().
> 
> A way of doing that is to add another argument to these functions, say
> "bool multipage".  Change the code to do
> 
> 	if (multipage) {
> 		while (uaddr <= end) {
> 			...
> 		}
> 	}
> 
> and change the callsites to pass in constant "true" or "false".  Then
> compile it up and manually check that the compiler completely removed
> the offending code from the filemap.c callsites.
> 
> Wanna have a think about that?  If it all looks OK then please be sure
> to add code comments explaining why we did this.

I wasn't really happy with the added branch either, but failed to come up
with a trick to avoid it. Imho adding new _multipage variants of these
functions instead of adding a constant argument is simpler because the
functions don't really share much thanks to the block below. I'll see what
it looks like (and obviously add a comment explaining what's going on).

> >  	if (ret == 0) {
> > -		char __user *end = uaddr + size - 1;
> > -
> >  		/*
> >  		 * If the page was already mapped, this will get a cache miss
> >  		 * for sure, so try to avoid doing it.
> >  		 */
> > -		if (((unsigned long)uaddr & PAGE_MASK) !=
> > +		if (((unsigned long)uaddr & PAGE_MASK) ==
> >  				((unsigned long)end & PAGE_MASK))
> 
> Maybe I'm having a dim day, but I don't immediately see why != got
> turned into ==.

Because of the loop uaddr will now point one page beyond the last
prefaulted page. To check whether end spilled into a new page we therefore
need to check whether uaddr and end are in the same pfn. Before uaddr
wasn't changed and hence the checking for a different pfn worked
correctly.

> Once we have this settled I'd suggest that the patch be carried in
> whatever-git-tree-needs-it.

Thanks for the comments.

Yours, Daniel
-- 
Daniel Vetter
Mail: daniel@...ll.ch
Mobile: +41 (0)79 365 57 48
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/