Received-SPF: pass (google.com: domain of linux-kernel+bounces-154819-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.80.249 as permitted sender) client-ip=147.75.80.249;
Date: Tue, 23 Apr 2024 11:55:38 +0200
From: Nam Cao <namcao@linutronix.de>
To: Thomas Zimmermann <tzimmermann@suse.de>
Cc: Jaya Kumar <jayalk@intworks.biz>, Daniel Vetter <daniel@ffwll.ch>,
	Helge Deller <deller@gmx.de>,
	Javier Martinez Canillas <javierm@redhat.com>,
	linux-fbdev@vger.kernel.org, dri-devel@lists.freedesktop.org,
	linux-kernel@vger.kernel.org, tiwai@suse.de, bigeasy@linutronix.de,
	patrik.r.jakobsson@gmail.com,
	Vegard Nossum <vegard.nossum@oracle.com>,
	George Kennedy <george.kennedy@oracle.com>,
	Darren Kenny <darren.kenny@oracle.com>, chuansheng.liu@intel.com,
	Harshit Mogalapalli <harshit.m.mogalapalli@oracle.com>,
	stable@vger.kernel.org
Subject: Re: [PATCH] fbdev: fix incorrect address computation in deferred IO
Message-ID: <20240423095538.m79ML6a0@linutronix.de>
References: <20240419190032.40490-1-namcao@linutronix.de>
 <666d986e-5227-4b6d-829c-95ff16115488@suse.de>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <666d986e-5227-4b6d-829c-95ff16115488@suse.de>

On Tue, Apr 23, 2024 at 10:37:41AM +0200, Thomas Zimmermann wrote:
> Hi,
> 
> thanks for following through with the bug and sending the patch
> 
> Am 19.04.24 um 21:00 schrieb Nam Cao:
> > With deferred IO enabled, a page fault happens when data is written to the
> > framebuffer device. Then driver determines which page is being updated by
> > calculating the offset of the written virtual address within the virtual
> > memory area, and uses this offset to get the updated page within the
> > internal buffer. This page is later copied to hardware (thus the name
> > "deferred IO").
> > 
> > This calculation is only correct if the virtual memory area is mapped to
> > the beginning of the internal buffer. Otherwise this is wrong. For example,
> > if users do:
> >      mmap(ptr, 4096, PROT_WRITE, MAP_FIXED | MAP_SHARED, fd, 0xff000);
> > 
> > Then the virtual memory area will mapped at offset 0xff000 within the
> > internal buffer. This offset 0xff000 is not accounted for, and wrong page
> > is updated. This will lead to wrong pixels being updated on the device.
> > 
> > However, it gets worse: if users do 2 mmap to the same virtual address, for
> > example:
> > 
> >      int fd = open("/dev/fb0", O_RDWR, 0);
> >      char *ptr = (char *) 0x20000000ul;
> >      mmap(ptr, 4096, PROT_WRITE, MAP_FIXED | MAP_SHARED, fd, 0xff000);
> >      *ptr = 0; // write #1
> >      mmap(ptr, 4096, PROT_WRITE, MAP_FIXED | MAP_SHARED, fd, 0);
> >      *ptr = 0; // write #2
> > 
> > In this case, both write #1 and write #2 apply to the same virtual address
> > (0x20000000ul), and the driver mistakenly thinks the same page is being
> > written to. When the second write happens, the driver thinks this is the
> > same page as the last time, and reuse the page from write #1. The driver
> > then lock_page() an incorrect page, and returns VM_FAULT_LOCKED with the
> > correct page unlocked. It is unclear what will happen with memory
> > management subsystem after that, but likely something terrible.
> 
> Please tone down the drama. :)

Sorry, that wasn't intentional. Writing is hard :(

Let me just cut this out, this info is not really needed to justify the
changes.

> 
> > 
> > Fix this by taking the mapping offset into account.
> > 
> > Reported-and-tested-by: Harshit Mogalapalli <harshit.m.mogalapalli@oracle.com>
> > Closes: https://lore.kernel.org/linux-fbdev/271372d6-e665-4e7f-b088-dee5f4ab341a@oracle.com
> > Fixes: 56c134f7f1b5 ("fbdev: Track deferred-I/O pages in pageref struct")
> > Cc: stable@vger.kernel.org
> > Signed-off-by: Nam Cao <namcao@linutronix.de>
> > ---
> >   drivers/video/fbdev/core/fb_defio.c | 3 ++-
> >   1 file changed, 2 insertions(+), 1 deletion(-)
> > 
> > diff --git a/drivers/video/fbdev/core/fb_defio.c b/drivers/video/fbdev/core/fb_defio.c
> > index dae96c9f61cf..d5d6cd9e8b29 100644
> > --- a/drivers/video/fbdev/core/fb_defio.c
> > +++ b/drivers/video/fbdev/core/fb_defio.c
> > @@ -196,7 +196,8 @@ static vm_fault_t fb_deferred_io_track_page(struct fb_info *info, unsigned long
> >    */
> >   static vm_fault_t fb_deferred_io_page_mkwrite(struct fb_info *info, struct vm_fault *vmf)
> >   {
> > -	unsigned long offset = vmf->address - vmf->vma->vm_start;
> > +	unsigned long offset = vmf->address - vmf->vma->vm_start
> > +			+ (vmf->vma->vm_pgoff << PAGE_SHIFT);
> 
> The page-fault handler at [1] use vm_fault.pgoff to retrieve the page
> structure. Can we do the same here and avoid that computation?

Yes, thanks for the suggestion.

It will change things a bit: offset will not be the exact value anymore,
but will be rounded down to multiple of PAGE_SIZE. But that doesn't matter,
because it will only be used to calculate the page offset later on.

We can clean this up and rename this "offset" to "pg_offset". But that's
for another day.

Best regards,
Nam