Re: [PATCH] vduse: Fix a possible warning in vduse_create_dev()

[Harshit Mogalapalli]

Hi Micheal,

On 27/11/22 10:04 pm, Michael S. Tsirkin wrote:
> On Sun, Nov 27, 2022 at 08:16:24AM +0530, Harshit Mogalapalli wrote:
> > Hi Micheal,
> >
> > On 27/11/22 4:52 am, Michael S. Tsirkin wrote:
> > > On Fri, Nov 25, 2022 at 07:59:58PM -0800, Harshit Mogalapalli wrote:
> > > > As 'dev->vq_num' is user-controlled data, if user tries to allocate
> > > > memory larger than(>=) MAX_ORDER, then kcalloc() will fail, it
> > > > creates a stack trace and messes up dmesg with a warning.
> > > >
> > > > Call trace:
> > > > -> vduse_ioctl
> > > > --> vduse_create_dev
> > > > 'config->vq_num' is user data as it comes from ioctl, which is
> > > > assigned to 'dev->vq_num'.
> > > >
> > > > Add __GFP_NOWARN in order to avoid too large allocation warning.
> > > > This is detected by static analysis using smatch.
> > > >
> > > > Fixes: c8a6153b6c59 ("vduse: Introduce VDUSE - vDPA Device in Userspace")
> > > > Signed-off-by: Harshit Mogalapalli <harshit.m.mogalapalli@xxxxxxxxxx>
> > > > ---
> > > >    drivers/vdpa/vdpa_user/vduse_dev.c | 3 ++-
> > > >    1 file changed, 2 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > index 35dceee3ed56..5e9546b16165 100644
> > > > --- a/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > +++ b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > @@ -1512,7 +1512,8 @@ static int vduse_create_dev(struct vduse_dev_config *config,
> > > >    	dev->config_size = config->config_size;
> > > >    	dev->vq_align = config->vq_align;
> > > >    	dev->vq_num = config->vq_num;
> > > > -	dev->vqs = kcalloc(dev->vq_num, sizeof(*dev->vqs), GFP_KERNEL);
> > > > +	dev->vqs = kcalloc(dev->vq_num, sizeof(*dev->vqs),
> > > > +			   GFP_KERNEL | __GFP_NOWARN);
> > > >    	if (!dev->vqs)
> > > >    		goto err_vqs;
> >
> > Thanks for checking the patch.
> >
> > > This is insufficient - the real source of the problem is that
> > > vq_num is not validated.
> > > The thing to do is to validate config and limit vq_num to 0xffff;
> >
> > 1557 static long vduse_ioctl(struct file *file, unsigned int cmd,
> > 1558                         unsigned long arg)
> > 1559 {
> > 1560         int ret;
> > 1561         void __user *argp = (void __user *)arg;
> > 1564         mutex_lock(&vduse_lock);
> > 1565         switch (cmd) {
> >               ....
> > 1584         case VDUSE_CREATE_DEV: {
> > 1585                 struct vduse_dev_config config;
> > 1587                 void *buf;
> > 1588
> > 1589                 ret = -EFAULT;
> > 1590                 if (copy_from_user(&config, argp, size))
> > 1591                         break;
> > 1592
> > 1593                 ret = -EINVAL;
> > 1594                 if (vduse_validate_config(&config) == false)
> > 1595                         break;
> > 1596
> > 1597                 buf = vmemdup_user(argp + size, config.config_size);
> > 1598                 if (IS_ERR(buf)) {
> > 1599                         ret = PTR_ERR(buf);
> > 1600                         break;
> > 1601                 }
> > 1602                 config.name[VDUSE_NAME_MAX - 1] = '\0';
> > 1603                 ret = vduse_create_dev(&config, buf,
> > control->api_version);
> > 1604                 if (ret)
> > 1605                         kvfree(buf);
> > 1606                 break;
> > 1607         }
> >
> > we have vduse_validate_config() being called in vduse_ioctl() which is the
> > caller of vduse_create_dev(), so validate_config() is not necessary in
> > vduse_create_dev() ?
> >
> > Thanks,
> > Harshit
>
> OK but I don't see vduse_validate_config checking vq_num.

right, I have added a limit of 0xffff to vq_num as you suggested in V2.

The reason for keeping the limit as 0xffff is the max number of virt 
queues is the size of vring buffer?

Thanks,
Harshit

> > > > --
> > > > 2.38.1