Re: [PATCH v3] staging: vt6655: check for memory allocation failures

From: Ji-Hun Kim
Date: Wed Apr 04 2018 - 03:24:23 EST


On Tue, Apr 03, 2018 at 01:40:52PM +0300, Dan Carpenter wrote:
> > desc->rd_info = kzalloc(sizeof(*desc->rd_info), GFP_KERNEL);
> > -
> > + if (!desc->rd_info) {
> > + ret = -ENOMEM;
> > + goto error;
> > + }
> > if (!device_alloc_rx_buf(priv, desc))
> > dev_err(&priv->pcid->dev, "can not alloc rx bufs\n");
> >
>
> We need to handle the case where device_alloc_rx_buf() fails as well...

Hi Dan, thanks for your comments! I will write a patch v5 including this fix.

> Some years back, I wrote a post about error handling that might be
> helpful:
> https://plus.google.com/106378716002406849458/posts/dnanfhQ4mHQ

This post is very helpful to me, Thank you.

> You are using "one err" and "do nothing" style error handling which are
> described in the post.

Sorry, I didn't fully understand "one err" and "do nothing" style error
handling of this code. The reason using goto instead of returns directly
was that for freeing previously allocated "rd_info"s in the for loop.
Please see below comment.


> > @@ -550,20 +554,29 @@ static void device_init_rd0_ring(struct vnt_private *priv)
> > if (i > 0)
> > priv->aRD0Ring[i-1].next_desc = cpu_to_le32(priv->rd0_pool_dma);
> > priv->pCurrRD[0] = &priv->aRD0Ring[0];
> > +
> > + return 0;
> > +error:
> > + device_free_rd0_ring(priv);
> > + return ret;
> > }
>
> Of course, Jia-Ju Bai is correct to say that this is a layering
> violation. Each function should only clean up after its self.
>
> Also, this is a very typical "one err" style bug which I explain about
> in my g+ post. The rule that applies here is that you should only free
> things which have been allocated. Since we only partially allocated the
> rd0 ring, device_free_rd0_ring() will crash when we do:
>
> dma_unmap_single(&priv->pcid->dev, rd_info->skb_dma,
> priv->rx_buf_sz, DMA_FROM_DEVICE);
>
> "rd_info" is NULL so rd_info->skb_dma is a NULL dereference.

For dealing with this problem, I added below code on patch v3. I think it
would not make Null dereferencing issues, because it will not enter
dma_unmap_single(), if "rd_info" is Null.

+ if (rd_info) {
+ dma_unmap_single(&priv->pcid->dev, rd_info->skb_dma,
+ priv->rx_buf_sz, DMA_FROM_DEVICE);
+ dev_kfree_skb(rd_info->skb);
+ kfree(desc->rd_info);
+ }

I would appreciate for your opinions what would be better way for freeing
allocated "rd_info"s in the loop previously.

Best regards,
Ji-Hun