Re: [PATCH] staging: comedi: drivers: Fix memory leak in gsc_hpdi_auto_attach

[Navid Emamdoost]

Ian, thanks for your feedback.

On Mon, Dec 16, 2019 at 4:36 AM Ian Abbott <abbotti@xxxxxxxxx> wrote:
>
> On 15/12/2019 01:33, Navid Emamdoost wrote:
> > In the implementation of gsc_hpdi_auto_attach(), the allocated dma
> > description is leaks in case of alignment error, or failure of
> > gsc_hpdi_setup_dma_descriptors() or comedi_alloc_subdevices(). Release
> > devpriv->dma_desc via dma_free_coherent().
> >
> > Signed-off-by: Navid Emamdoost <navid.emamdoost@xxxxxxxxx>
>
> Actually, there is no memory leak (although there is another problem
> that I'll mention below).  If the "auto_attach" handler
> gsc_hpdi_auto_attach() returns an error, then the "detach" handler
> gsc_hpdi_detach() will be called automatically to clean up.  (This is
> true for all comedi drivers).  gsc_hpdi_detach() calls
> gsc_hpdi_free_dma() to free the DMA buffers and DMA descriptors.
>
I was aware that comedi_alloc_devpriv() is a resource managed
allocation, but was not sure how subsequent dma_desc allocation will
be handled when device detach.

> > ---
> >   drivers/staging/comedi/drivers/gsc_hpdi.c | 16 +++++++++++++---
> >   1 file changed, 13 insertions(+), 3 deletions(-)
> >
> > diff --git a/drivers/staging/comedi/drivers/gsc_hpdi.c b/drivers/staging/comedi/drivers/gsc_hpdi.c
> > index 4bdf44d82879..c0c7047a6d1b 100644
> > --- a/drivers/staging/comedi/drivers/gsc_hpdi.c
> > +++ b/drivers/staging/comedi/drivers/gsc_hpdi.c
> > @@ -633,16 +633,17 @@ static int gsc_hpdi_auto_attach(struct comedi_device *dev,
> >       if (devpriv->dma_desc_phys_addr & 0xf) {
> >               dev_warn(dev->class_dev,
> >                        " dma descriptors not quad-word aligned (bug)\n");
> > -             return -EIO;
> > +             retval = -EIO;
> > +             goto release_dma_desc;
> >       }
> >
> >       retval = gsc_hpdi_setup_dma_descriptors(dev, 0x1000);
> >       if (retval < 0)
> > -             return retval;
> > +             goto release_dma_desc;
> >
> >       retval = comedi_alloc_subdevices(dev, 1);
> >       if (retval)
> > -             return retval;
> > +             goto release_dma_desc;
> >
> >       /* Digital I/O subdevice */
> >       s = &dev->subdevices[0];
> > @@ -660,6 +661,15 @@ static int gsc_hpdi_auto_attach(struct comedi_device *dev,
> >       s->cancel       = gsc_hpdi_cancel;
> >
> >       return gsc_hpdi_init(dev);
> > +
> > +release_dma_desc:
> > +     if (devpriv->dma_desc)
> > +             dma_free_coherent(&pcidev->dev,
> > +                               sizeof(struct plx_dma_desc) *
> > +                             NUM_DMA_DESCRIPTORS,
> > +                             devpriv->dma_desc,
> > +                             devpriv->dma_desc_phys_addr);
> > +     return retval;
> >   }
> >
> >   static void gsc_hpdi_detach(struct comedi_device *dev)
> >
>
> This patch could actually result in devpriv->dma_desc being freed twice
> - once in the 'release_dma_desc:' code and again when gsc_hpdi_detach()
> is called externally as part of the clean-up.
>
> The real bug in the original code is that it does not check whether any
> of the calls to dma_alloc_coherent() returned NULL.  If any of the calls
> to dma_alloc_coherent() returns NULL, gsc_hpdi_auto_attach() needs to
> return an error (-ENOMEM).  The subsequent call to gsc_hpdi_detach()
> will then free whatever DMA coherent buffers where allocated.
>
Yes, this potential null deref is another type of bug, which I will
send a patch for separately.

> --
> -=( Ian Abbott <abbotti@xxxxxxxxx> || Web: www.mev.co.uk )=-
> -=( MEV Ltd. is a company registered in England & Wales. )=-
> -=( Registered number: 02862268.  Registered address:    )=-
> -=( 15 West Park Road, Bramhall, STOCKPORT, SK7 3JZ, UK. )=-



-- 
Navid.