[alsa-devel] [PATCH] sound/oss/midi_synth: prevent underflow, use of uninitialized value, and signedness issue

[Dan Rosenberg]

On Wed, 2011-03-23 at 07:23 -0400, Dan Rosenberg wrote:
> On Wed, 2011-03-23 at 08:39 +0100, Takashi Iwai wrote:
> > At Tue, 22 Mar 2011 14:31:08 -0400,
> > Dan Rosenberg wrote:
> > > 
> > > The offset passed to midi_synth_load_patch() can be essentially
> > > arbitrary.  If it's greater than the header length, this will result in
> > > a copy_from_user(dst, src, negative_val).  While this will just return
> > > -EFAULT on x86, on other architectures this may cause memory corruption.
> > > Additionally, the length field of the sysex_info structure may not be
> > > initialized prior to its use.  Finally, a signed comparison may result
> > > in an unintentionally large loop.
> > > 
> > > This patch fixes all these issues, as well as some cleanup to prevent
> > > checkpatch.pl from complaining.
> > > 
> > > Signed-off-by: Dan Rosenberg <drosenberg at vsecurity.com>
> > > Cc: stable at kernel.org
> > 
> > Well, the whole load_patch mechanism doesn't look working right with
> > ofs != 0.  The only caller of this callback is sound/oss/sequencer.c,
> > and it assumes that the whole chunk is passed once without splitting.
> > Thus the offset calculation in this code is obviously wrong, and
> > passing offset itself doesn't make any sense.
> > 
> > A similar problem (uninitialized struct fields) is found in another
> > load_patch callback in sound/oss/opl3.c.
> > 
> > That is, the best fix would be to rip off the offset argument from
> > this callback.
> > 
> 
> Thanks, I'll resend a new patch series that uses this approach as well
> as addresses a few more security issues.
> 
> On a related note, is there any interest in removing OSS (actual, not
> emulation) entirely?  It's been marked as deprecated since before the
> git epoch (2005), and it's seen little development other than bug fixes
> and security issues.
> 

Feel free to ignore this uneducated suggestion, as I clearly hadn't done
my homework in looking up history of the sound tree.  No need to bring
up the ALSA vs. OSS arguments again.

-Dan

> Regards,
> Dan
> 
> > 
> > thanks,
> > 
> > Takashi
> > 
> > > ---
> > >  sound/oss/midi_synth.c |   27 ++++++++++++++-------------
> > >  1 files changed, 14 insertions(+), 13 deletions(-)
> > > 
> > > diff --git a/sound/oss/midi_synth.c b/sound/oss/midi_synth.c
> > > index 3c09374..3500f80 100644
> > > --- a/sound/oss/midi_synth.c
> > > +++ b/sound/oss/midi_synth.c
> > > @@ -491,16 +491,18 @@ midi_synth_load_patch(int dev, int format, const char __user *addr,
> > >  	if (!prefix_cmd(orig_dev, 0xf0))
> > >  		return 0;
> > >  
> > > +	/* Invalid patch format */
> > >  	if (format != SYSEX_PATCH)
> > > -	{
> > > -/*		  printk("MIDI Error: Invalid patch format (key) 0x%x\n", format);*/
> > >  		  return -EINVAL;
> > > -	}
> > > +
> > > +	/* Patch header too short */
> > >  	if (count < hdr_size)
> > > -	{
> > > -/*		printk("MIDI Error: Patch header too short\n");*/
> > >  		return -EINVAL;
> > > -	}
> > > +
> > > +	/* Offset too high */
> > > +	if (offs > offsetof(struct sysex_info, len) || offs < 0)
> > > +		return -EINVAL;
> > > +
> > >  	count -= hdr_size;
> > >  
> > >  	/*
> > > @@ -510,14 +512,13 @@ midi_synth_load_patch(int dev, int format, const char __user *addr,
> > >  
> > >  	if(copy_from_user(&((char *) &sysex)[offs], &(addr)[offs], hdr_size - offs))
> > >  		return -EFAULT;
> > > - 
> > > - 	if (count < sysex.len)
> > > -	{
> > > -/*		printk(KERN_WARNING "MIDI Warning: Sysex record too short (%d<%d)\n", count, (int) sysex.len);*/
> > > +
> > > +	/* Sysex record too short */
> > > +	if ((unsigned)count < (unsigned)sysex.len)
> > >  		sysex.len = count;
> > > -	}
> > > -  	left = sysex.len;
> > > -  	src_offs = 0;
> > > +
> > > +	left = sysex.len;
> > > +	src_offs = 0;
> > >  
> > >  	for (i = 0; i < left && !signal_pending(current); i++)
> > >  	{
> > > 
> > > 
>