Re: [Linux-kernel-mentees] [PATCH v3] media/v4l2-core: Fix kernel-infoleak in video_put_user()

[Dan Carpenter <dan.carpenter@oracle.com>]

On Mon, Jul 27, 2020 at 04:05:38PM +0200, Arnd Bergmann wrote:
> On Mon, Jul 27, 2020 at 3:16 PM Dan Carpenter <dan.carpenter@oracle.com> wrote:
> >
> > On Mon, Jul 27, 2020 at 09:25:16AM +0200, Arnd Bergmann wrote:
> > > On Mon, Jul 27, 2020 at 12:28 AM Peilin Ye <yepeilin.cs@gmail.com> wrote:
> > > >
> > > > video_put_user() is copying uninitialized stack memory to userspace due
> > > > to the compiler not initializing holes in the structures declared on the
> > > > stack. Fix it by initializing `ev32` and `vb32` using memset().
> > > >
> > > > Reported-and-tested-by: syzbot+79d751604cb6f29fbf59@syzkaller.appspotmail.com
> > > > Link: https://syzkaller.appspot.com/bug?extid=79d751604cb6f29fbf59
> > > > Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
> > > > Signed-off-by: Peilin Ye <yepeilin.cs@gmail.com>
> > >
> > > Thanks a lot for addressing this! I now see that I actually created a similar
> > > bugfix for it back in January, but for some reason that got stuck in my
> > > backlog and I never wrote a proper description for it or sent it out to the
> > > list, sorry about that. I would hope we could find a way to have either
> > > the compiler or sparse warn if we copy uninitialized data to user space,
> > > but we now don't even check for that within the kernel any more.
> >
> > Here are my latest warnings on linux-next from Friday.
> 
> Ah, I forgot you had that kind of list already, thanks for checking!
> 
> > block/scsi_ioctl.c:707 scsi_put_cdrom_generic_arg() warn: check that 'cgc32' doesn't leak information (struct has a hole after 'data_direction')
> 
> I see no padding in this one, should be fine AFAICT. Any idea why you
> get a warning
> for this instance?
> 

The warning message only prints the first struct hole or uninitialized
struct memeber.  ->data_direction in this case.

block/scsi_ioctl.c
   646  #ifdef CONFIG_COMPAT
   647  struct compat_cdrom_generic_command {
   648          unsigned char   cmd[CDROM_PACKET_SIZE];
   649          compat_caddr_t  buffer;
   650          compat_uint_t   buflen;
   651          compat_int_t    stat;
   652          compat_caddr_t  sense;
   653          unsigned char   data_direction;

There is going to be 3 bytes of padding between this char and the next
int.

   654          compat_int_t    quiet;
   655          compat_int_t    timeout;
   656          compat_caddr_t  reserved[1];
   657  };
   658  #endif

The bug seems real, but it depends on the compiler of course.

> > drivers/input/misc/uinput.c:743 uinput_ff_upload_to_user() warn: check that 'ff_up_compat' doesn't leak information (struct has a hole after 'replay')
> 
> This one hs padding in it and looks broken.

No.  This a bug in smatch.  It's memcpy() over the hole, but the check
isn't capturing that.  The code is slightly weird...  I could try
silence the warning but it would only silence this one false positive so
I haven't investigated it.

> 
> > drivers/input/misc/uinput.c:958 uinput_ioctl_handler() warn: check that 'ff_up' doesn't leak information (struct has a hole after 'replay')
> 
> hard to tell.
> 

Looks ok, I think.

> > drivers/firewire/core-cdev.c:463 ioctl_get_info() warn: check that 'bus_reset' doesn't leak information (struct has a hole after 'generation')
> 
> broken, trivial to fix
> 
> > drivers/scsi/megaraid/megaraid_mm.c:824 kioc_to_mimd() warn: check that 'cinfo.base' doesn't leak information
> 
> Seems fine due to __packed annotation.
> 

It's not related __packed.  Smatch is saying that cinfo.base isn't
necessarily initialized.

drivers/scsi/megaraid/megaraid_mm.c
   816  
   817                  case MEGAIOC_QADAPINFO:
   818  
   819                          hinfo = (mraid_hba_info_t *)(unsigned long)
   820                                          kioc->buf_vaddr;
   821  
   822                          hinfo_to_cinfo(hinfo, &cinfo);

hinfo_to_cinfo() is a no-op if hinfo is NULL.  Smatch can't tell if
"hinfo" is non-NULL.

   823  
   824                          if (copy_to_user(kmimd.data, &cinfo, sizeof(cinfo)))
   825                                  return (-EFAULT);
   826  
   827                          return 0;
   828  

> > drivers/gpio/gpiolib-cdev.c:473 lineevent_read() warn: check that 'ge' doesn't leak information (struct has a hole after 'id')
> 
> The driver seems to initialize the elements correctly before putting
> them into the kfifo, so there is no infoleak. However the structure layout
> of "struct gpioevent_data" is incompatible between x86-32 and x86-64
> calling conventions, so this is likely broken in x86 compat mode,
> unless user space can explicitly deal with the difference.

Smatch isn't parsing kfifo_out() correctly.  It turns out that
kfifo_out() is slightly complicated for Smatch.

> 
> > drivers/gpu/drm/i915/i915_query.c:136 query_engine_info() warn: check that 'query.num_engines' doesn't leak information
> 
> I don't think this leaks any state, as it just copies data to user space that
> it copied from there originally.

Yeah.  copy_query_item() isn't parsed correctly.  I've added this one
to my TODO list because it should parse this correctly.

regards,
dan carpenter