Skip to contentSkip to conversation

ISDN: fix double free bug in isdn_net

PATCH In review Jesper Juhl 3 participants 10 messages Aug 12, 2006 – Aug 16, 2006

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ISDN: fix double free bug in isdn_net

There's double-free bug in drivers/isdn/i4l/isdn_net.c::isdn_net_writebuf_skb().

If isdn_writebuf_skb_stub() doesn't handle the entire skb, then it will have freed the skb that was passed to it and when the code then jumps to the error label it'll result in a double free of the skb.

The easy way to fix this is to insert an assignment of skb = NULL in the 'if' following the call to isdn_writebuf_skb_stub() so that when the code at the error label calls dev_kfree_skb(skb); the skb will be NULL and nothing will happen since dev_kfree_skb() just does a return if passed a NULL.

---

 drivers/isdn/i4l/isdn_net.c |    1 +
 1 files changed, 1 insertion(+)
drivers/isdn/i4l/isdn_net.c +1 -0
@@ -1023,6 +1023,7 @@ void isdn_net_writebuf_skb(isdn_net_loca
 	if (ret != len) {
 		/* we should never get here */
 		printk(KERN_WARNING "%s: HL driver queue full\n", lp->name);
+		skb = NULL;
 		goto error;
 	}
 	
#1 lore raw

From: Jesper Juhl <[email protected]>
Date: Sat, 12 Aug 2006 22:48:22 +0200

Quoted 7 lines

There's double-free bug in drivers/isdn/i4l/isdn_net.c::isdn_net_writebuf_skb().

If isdn_writebuf_skb_stub() doesn't handle the entire skb, then it will have freed the skb that was passed to it and when the code then jumps to the error label it'll result in a double free of the skb.

Quoted 7 lines

The easy way to fix this is to insert an assignment of skb = NULL in the 'if' following the call to isdn_writebuf_skb_stub() so that when the code at the error label calls dev_kfree_skb(skb); the skb will be NULL and nothing will happen since dev_kfree_skb() just does a return if passed a NULL.

I can't ascertain that this is exactly true.

For example, in isdn_writebuf_skb_stub() we have this:

	if (ret > 0) {
 ...
		if (dev->v110[idx]) {
 ...
			if (ret == skb->len)
				dev_kfree_skb(skb);

So if ret > 0 and we're doing V.110, we only free the skb if isdn_v110_encode put a value at *((int *)nskb->data) equal to skb->len.

So in the V.110 case it's not purely the ->writebuf_skb() return value that gets returned, rather we return that integer that isdn_v110_encode() put there.

Therefore I don't think this SKB leak is %100 correct. What did I miss?

BTW, this ISDN code is rotting a bit and could use some cleanups if not a rewrite. :-/

#2 lore raw
Quoted 20 lines · David Miller

From: Jesper Juhl <[email protected]>
Date: Sat, 12 Aug 2006 22:48:22 +0200

Quoted 7 lines

There's double-free bug in drivers/isdn/i4l/isdn_net.c::isdn_net_writebuf_skb().

If isdn_writebuf_skb_stub() doesn't handle the entire skb, then it will have freed the skb that was passed to it and when the code then jumps to the error label it'll result in a double free of the skb.

Quoted 7 lines

The easy way to fix this is to insert an assignment of skb = NULL in the 'if' following the call to isdn_writebuf_skb_stub() so that when the code at the error label calls dev_kfree_skb(skb); the skb will be NULL and nothing will happen since dev_kfree_skb() just does a return if passed a NULL.

I can't ascertain that this is exactly true.

Quoted 9 lines

For example, in isdn_writebuf_skb_stub() we have this:

        if (ret > 0) {
 ...
                if (dev->v110[idx]) {
 ...
                        if (ret == skb->len)
                                dev_kfree_skb(skb);

Hmm, perhaps I made a mistake and missed a path. Maybe it would be better to fix if by making isdn_writebuf_skb_stub() always set the skb to NULL when it does free it. That would add a few more assignments but should ensure the right result always.
What do you say?

Quoted 8 lines

So if ret > 0 and we're doing V.110, we only free the skb if isdn_v110_encode put a value at *((int *)nskb->data) equal to skb->len.

So in the V.110 case it's not purely the ->writebuf_skb() return value that gets returned, rather we return that integer that isdn_v110_encode() put there.

Quoted 6 lines

Therefore I don't think this SKB leak is %100 correct. What did I miss?

BTW, this ISDN code is rotting a bit and could use some cleanups if not a rewrite. :-/

A rewrite would be beyond me, but I try to clean up bits and pieces...

Signature · 3 lines
Jesper Juhl <[email protected]>
Don't top-post  http://www.catb.org/~esr/jargon/html/T/top-post.html
Plain text mails only, please      http://www.expita.com/nomime.html
#3 lore raw

From: "Jesper Juhl" <[email protected]>
Date: Tue, 15 Aug 2006 11:08:35 +0200

Quoted 5 lines

Hmm, perhaps I made a mistake and missed a path. Maybe it would be better to fix if by making isdn_writebuf_skb_stub() always set the skb to NULL when it does free it. That would add a few more assignments but should ensure the right result always.
What do you say?

Do we know if the ->writebuf_skb() method ever frees the skb? If it never does, then yes your suggestion would be one way to handle this.

#4 lore raw
Quoted 12 lines · David Miller

From: "Jesper Juhl" <[email protected]>
Date: Tue, 15 Aug 2006 11:08:35 +0200

Quoted 5 lines

Hmm, perhaps I made a mistake and missed a path. Maybe it would be better to fix if by making isdn_writebuf_skb_stub() always set the skb to NULL when it does free it. That would add a few more assignments but should ensure the right result always.
What do you say?

Do we know if the ->writebuf_skb() method ever frees the skb? If it never does, then yes your suggestion would be one way to handle this.

I'll look at it tonight and if writebuf_skb() method never frees the skb I'll send a patch.

Signature · 3 lines
Jesper Juhl <[email protected]>
Don't top-post  http://www.catb.org/~esr/jargon/html/T/top-post.html
Plain text mails only, please      http://www.expita.com/nomime.html
#5 lore raw
In reply to David Miller
Quoted 11 lines · David Miller

From: "Jesper Juhl" <[email protected]>
Date: Tue, 15 Aug 2006 11:08:35 +0200

Quoted 5 lines

Hmm, perhaps I made a mistake and missed a path. Maybe it would be better to fix if by making isdn_writebuf_skb_stub() always set the skb to NULL when it does free it. That would add a few more assignments but should ensure the right result always.
What do you say?

Do we know if the ->writebuf_skb() method ever frees the skb? If it never does, then yes your suggestion would be one way to handle this.

It does if it consumes the skb (then it returns skb->len). But the skb have not to be freed imediately in this case, it maybe queued or used until all bytes are written to the physical device.

If it returns any other value the skb is not freed.

This logic came from using skb for transparent data too. Here it was possible, that the hw driver only take some bytes from the skb (so it returns < skb->len), then the isdn layer should requeue the skb so no transparent data get lost.

But this mechanism was never used in drivers, only 3 states:

The driver accept the packet then it is responsible for the skb and return skb->len or the driver do not accept it (e.g. buffer full, conntection is going down), then it return 0 and does not free the skb.

If some internal error in the HW driver occur, it should return a negative value and it also do not free the skb.

Signature · 3 lines
Karsten Keil
SuSE Labs
ISDN development
#6 lore raw
Quoted 27 lines · Karsten Keil
Quoted 11 lines · David Miller

From: "Jesper Juhl" <[email protected]>
Date: Tue, 15 Aug 2006 11:08:35 +0200

Quoted 5 lines

Hmm, perhaps I made a mistake and missed a path. Maybe it would be better to fix if by making isdn_writebuf_skb_stub() always set the skb to NULL when it does free it. That would add a few more assignments but should ensure the right result always.
What do you say?

Do we know if the ->writebuf_skb() method ever frees the skb? If it never does, then yes your suggestion would be one way to handle this.

It does if it consumes the skb (then it returns skb->len). But the skb have not to be freed imediately in this case, it maybe queued or used until all bytes are written to the physical device.

If it returns any other value the skb is not freed.

This logic came from using skb for transparent data too. Here it was possible, that the hw driver only take some bytes from the skb (so it returns < skb->len), then the isdn layer should requeue the skb so no transparent data get lost.

But this mechanism was never used in drivers, only 3 states:

Quoted 8 lines

The driver accept the packet then it is responsible for the skb and return skb->len or the driver do not accept it (e.g. buffer full, conntection is going down), then it return 0 and does not free the skb.

If some internal error in the HW driver occur, it should return a negative value and it also do not free the skb.

Ok, if I understand you correctly, then there's no actual problem here. right?

Signature · 3 lines
Jesper Juhl <[email protected]>
Don't top-post  http://www.catb.org/~esr/jargon/html/T/top-post.html
Plain text mails only, please      http://www.expita.com/nomime.html
#7 lore raw
Quoted 33 lines · Jesper Juhl
Quoted 32 lines · Karsten Keil
Quoted 11 lines · David Miller

From: "Jesper Juhl" <[email protected]>
Date: Tue, 15 Aug 2006 11:08:35 +0200

Quoted 5 lines

Hmm, perhaps I made a mistake and missed a path. Maybe it would be better to fix if by making isdn_writebuf_skb_stub() always set the skb to NULL when it does free it. That would add a few more assignments but should ensure the right result always.
What do you say?

Do we know if the ->writebuf_skb() method ever frees the skb? If it never does, then yes your suggestion would be one way to handle this.

It does if it consumes the skb (then it returns skb->len). But the skb have not to be freed imediately in this case, it maybe queued or used until all bytes are written to the physical device.

If it returns any other value the skb is not freed.

This logic came from using skb for transparent data too. Here it was possible, that the hw driver only take some bytes from the skb (so it returns < skb->len), then the isdn layer should requeue the skb so no transparent data get lost.

But this mechanism was never used in drivers, only 3 states:

The driver accept the packet then it is responsible for the skb and return skb->len or the driver do not accept it (e.g. buffer full, conntection is going down), then it return 0 and does not free the skb.

Quoted 7 lines
Quoted 3 lines

If some internal error in the HW driver occur, it should return a negative value and it also do not free the skb.

Ok, if I understand you correctly, then there's no actual problem here. right?

I not aware of any problems in this code (besides it should be cleaned up and rewritten).

Was here any trigger for your patch ?

Signature · 3 lines
Karsten Keil
SuSE Labs
ISDN development
#8 lore raw

From: Karsten Keil <[email protected]>
Date: Wed, 16 Aug 2006 22:39:35 +0200

Quoted 1 line

Was here any trigger for your patch ?

I think Jesper didn't have a test case, but rather spotted this visually.

#9 lore raw
In reply to Karsten Keil
Quoted 38 lines · Karsten Keil
Quoted 37 lines · Jesper Juhl
Quoted 35 lines · Karsten Keil
Quoted 11 lines · David Miller

From: "Jesper Juhl" <[email protected]>
Date: Tue, 15 Aug 2006 11:08:35 +0200

Quoted 5 lines

Hmm, perhaps I made a mistake and missed a path. Maybe it would be better to fix if by making isdn_writebuf_skb_stub() always set the skb to NULL when it does free it. That would add a few more assignments but should ensure the right result always.
What do you say?

Do we know if the ->writebuf_skb() method ever frees the skb? If it never does, then yes your suggestion would be one way to handle this.

It does if it consumes the skb (then it returns skb->len). But the skb have not to be freed imediately in this case, it maybe queued or used until all bytes are written to the physical device.

If it returns any other value the skb is not freed.

This logic came from using skb for transparent data too. Here it was possible, that the hw driver only take some bytes from the skb (so it returns < skb->len), then the isdn layer should requeue the skb so no transparent data get lost.

But this mechanism was never used in drivers, only 3 states:

The driver accept the packet then it is responsible for the skb and return skb->len or the driver do not accept it (e.g. buffer full, conntection is going down), then it return 0 and does not free the skb.

If some internal error in the HW driver occur, it should return a negative value and it also do not free the skb.

Quoted 9 lines
Quoted 3 lines

Ok, if I understand you correctly, then there's no actual problem here. right?

I not aware of any problems in this code (besides it should be cleaned up and rewritten).

Was here any trigger for your patch ?

Well, the trigger for the initial patch was Coverity bug #1060 . I looked at it and thought it looked valid, so I tried to cook up a patch for was I believed looked like an actual problem.

Here's what Coverity had to say :

CID: 1060
Checker: USE_AFTER_FREE (help)
File: base/src/linux-2.6/drivers/isdn/i4l/isdn_net.c

...
1006 void isdn_net_writebuf_skb(isdn_net_local *lp, struct sk_buff *skb)
1007 {
1008 int ret;
1009 int len = skb->len; /* save len */
1010
1011 /* before obtaining the lock the caller should have checked that
1012 the lp isn't busy */
1013 if (isdn_net_lp_busy(lp)) {
1014 printk("isdn BUG at %s:%d!\n", __FILE__, __LINE__);
1015 goto error;
1016 }
1017
1018 if (!(lp->flags & ISDN_NET_CONNECTED)) {
1019 printk("isdn BUG at %s:%d!\n", __FILE__, __LINE__);
1020 goto error;
1021 }

Event freed_arg: Pointer "skb" freed by function
"isdn_writebuf_skb_stub" [model]
Also see events: [double_free]

1022 ret = isdn_writebuf_skb_stub(lp->isdn_device, lp->isdn_channel, 1, skb);

At conditional (1): "ret != len" taking true path

1023 if (ret != len) {
1024 /* we should never get here */
1025 printk(KERN_WARNING "%s: HL driver queue full\n", lp->name);
1026 goto error;
1027 }
1028
1029 lp->transcount += len;
1030 isdn_net_inc_frame_cnt(lp);
1031 return;
1032
1033 error:

Event double_free: Double free of pointer "skb" in call to "kfree_skb" [model] Also see events: [freed_arg]

1034 dev_kfree_skb(skb);
1035 lp->stats.tx_errors++;
1036
1037 }
...

Signature · 3 lines
Jesper Juhl <[email protected]>
Don't top-post  http://www.catb.org/~esr/jargon/html/T/top-post.html
Plain text mails only, please      http://www.expita.com/nomime.html