mirror of https://lore.kernel.org/lkml/
 help / color / mirror / Atom feed
* [PATCH] ieee802154: 6lowpan: validate fragment offset against datagram size
@ 2026-11-07 22:32 Ryan Prinz
  2026-11-08 13:05 ` [PATCH RESEND] " Ryan Prinz
                   ` (2 more replies)
  0 siblings, 3 replies; 8+ messages in thread
From: Ryan Prinz @ 2026-11-07 22:32 UTC (permalink / raw)
  To: linux-wpan, netdev
  Cc: Alexander Aring, Stefan Schmidt, Miquel Raynal, David S . Miller,
	Eric Dumazet, Jakub Kicinski, Paolo Abeni, Simon Horman,
	linux-kernel, Ryan Prinz


[-- Attachment #1.1: Type: text/plain, Size: 3342 bytes --]

A 6LoWPAN fragment announces the total size of the datagram it belongs
to (d_size) and its own offset within that datagram. The reassembly
code never checks that the offset actually fits inside d_size, so a
fragment claiming to start far past the end of the datagram is
accepted the same as a legitimate one, allowing an attacker to forge
data of another sender.

lowpan_frag_queue() considers a datagram complete by counting bytes
received (fq->q.meat) against the declared size (fq->q.len), not by
verifying the bytes form a contiguous range. A fragment with an
out-of-range offset still has its length counted toward that total, so
a legitimate fragment exchange plus an attacker fragment can sum to
d_size, marking the datagram complete while a gap in the data was
never filled, and appending the out-of-range fragment's bytes to the
end of the reassembled buffer as if they were part of the message.

IPv6's reassembly (net/ipv6/reassembly.c, ip6_frag_queue()) computes
end = of
fset + payload length for every fragment and checks end
against IPV6_MAXPLEN, which bounds offset by construction. 6LoWPAN's
d_offset and d_size are independent wire fields with no such coupling,
and no separate bounds check was added for it.

The same class of bug was assigned CVE-2019-8359 (CVSS 9.8) in
Contiki-NG's 6LoWPAN stack before a 2019 fix; RIOT-OS's implementation
(gnrc_sixlowpan_frag_rb_add) rejects a fragment once offset + size
exceeds the reassembly buffer's datagram_size. Linux's 6LoWPAN
reassembly is the only one of these implementations without an
equivalent check.

Fix by rejecting any fragment whose offset, or offset plus length,
falls outside [0, d_size).

Fixes: 7240cdec60b1 ("6lowpan: handling 6lowpan fragmentation via inet_frag api")
Signed-off-by: Ryan Prinz <rc4d5a@protonmail.com>
---
Reproduced two ways: with a local kernel module calling
lowpan_frag_rcv() directly, and over mac802154_hwsim with three
independent processes standing in for s
ender, receiver and attacker.
An attacker in radio range who observes one fragment of an
in-progress transmission reads d_tag, d_size and the addressing in the
clear, then sends a single out-of-range fragment of its own. The
receiver's reassembly completes with the attacker's bytes appended to
the legitimate data. Both reproducers were re-run against this fix to
confirm the out-of-range fragment is now rejected and the datagram no
longer completes with it included.

AI Disclosure: An AI assistant (Claude) was used to assist with the
analysis and testing. I have verified the vulnerability, the fix, and
the test results myself and take responsibility for this contribution,
per Documentation/process/coding-assistants.rst.

 net/ieee802154/6lowpan/reassembly.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/net/ieee802154/6lowpan/reassembly.c b/net/ieee802154/6lowpan/reassembly.c
index ddb6a5817d09..905ba5f86844 100644
--- a/net/ieee802154/6lowpan/reassembly.c
++
+ b/net/ieee802154/6lowpan/reassembly.c
@@ -103,6 +103,9 @@ static int lowpan_frag_queue(struct lowpan_frag_queue *fq,
 	offset = lowpan_802154_cb(skb)->d_offset << 3;
 	end = lowpan_802154_cb(skb)->d_size;

+	if (offset > end || skb->len > end - offset)
+		goto err;
+
 	/* Is this the final fragment? */
 	if (offset + skb->len == end) {
 		/* If we already have some bits beyond end
--
2.51.2

[-- Attachment #1.2: publickey - rc4d5a@protonmail.com - 0xE775BF83.asc --]
[-- Type: application/pgp-keys, Size: 897 bytes --]

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 322 bytes --]

^ permalink raw reply	[flat|nested] 8+ messages in thread

* [PATCH RESEND] ieee802154: 6lowpan: validate fragment offset against datagram size
  2026-11-07 22:32 [PATCH] ieee802154: 6lowpan: validate fragment offset against datagram size Ryan Prinz
@ 2026-11-08 13:05 ` Ryan Prinz
  2026-11-08 14:20 ` [PATCH net " Ryan Prinz
  2026-11-09  9:03 ` Ryan Prinz
  2 siblings, 0 replies; 8+ messages in thread
From: Ryan Prinz @ 2026-11-08 13:05 UTC (permalink / raw)
  To: linux-wpan, netdev
  Cc: Alexander Aring, Stefan Schmidt, Miquel Raynal, David S. Miller,
	Eric Dumazet, Jakub Kicinski, Paolo Abeni, Simon Horman,
	linux-kernel, Ryan Prinz


[-- Attachment #1.1: Type: text/plain, Size: 3407 bytes --]

A 6LoWPAN fragment announces the total size of the datagram it belongs
to (d_size) and its own offset within that datagram. The reassembly
code never checks that the offset actually fits inside d_size, so a
fragment claiming to start far past the end of the datagram is
accepted the same as a legitimate one, allowing an attacker to forge
data of another sender.

lowpan_frag_queue() considers a datagram complete by counting bytes
received (fq->q.meat) against the declared size (fq->q.len), not by
verifying the bytes form a contiguous range. A fragment with an
out-of-range offset still has its length counted toward that total, so
a legitimate fragment exchange plus an attacker fragment can sum to
d_size, marking the datagram complete while a gap in the data was
never filled, and appending the out-of-range fragment's bytes to the
end of the reassembled buffer as if they were part of the message.

IPv6's reassembly (net/ipv6/reassembly.c, ip6_frag_queue()) computes
end = of
fset + payload length for every fragment and checks end
against IPV6_MAXPLEN, which bounds offset by construction. 6LoWPAN's
d_offset and d_size are independent wire fields with no such coupling,
and no separate bounds check was added for it.

The same class of bug was assigned CVE-2019-8359 (CVSS 9.8) in
Contiki-NG's 6LoWPAN stack before a 2019 fix; RIOT-OS's implementation
(gnrc_sixlowpan_frag_rb_add) rejects a fragment once offset + size
exceeds the reassembly buffer's datagram_size. Linux's 6LoWPAN
reassembly is the only one of these implementations without an
equivalent check.

Fix by rejecting any fragment whose offset, or offset plus length,
falls outside [0, d_size).

Reproduced two ways: with a local kernel module calling
lowpan_frag_rcv() directly, and over mac802154_hwsim with three
independent processes standing in for sender, receiver and attacker.
An attacker in radio range who observes one fragment of an
in-progress transmission reads d_tag, d_size and
 the addressing in the
clear, then sends a single out-of-range fragment of its own. The
receiver's reassembly completes with the attacker's bytes appended to
the legitimate data. Both reproducers were re-run against this fix to
confirm the out-of-range fragment is now rejected and the datagram no
longer completes with it included.

AI Disclosure: An AI assistant (Claude) was used to assist with the
analysis and testing. I have verified the vulnerability, the fix, and
the test results myself and take responsibility for this contribution,
per Documentation/process/coding-assistants.rst.

Fixes: 7240cdec60b1 ("6lowpan: handling 6lowpan fragmentation via inet_frag api")
Signed-off-by: Ryan Prinz <rc4d5a@protonmail.com>
---
Resending to fix a diff formatting issue in my previous mail.

 net/ieee802154/6lowpan/reassembly.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/net/ieee802154/6lowpan/reassembly.c b/net/ieee802154/6lowpan/reassembly.c
index ddb6a5817d09..9
05ba5f86844 100644
--- a/net/ieee802154/6lowpan/reassembly.c
+++ b/net/ieee802154/6lowpan/reassembly.c
@@ -103,6 +103,9 @@ static int lowpan_frag_queue(struct lowpan_frag_queue *fq,
 	offset = lowpan_802154_cb(skb)->d_offset << 3;
 	end = lowpan_802154_cb(skb)->d_size;
 
+	if (offset > end || skb->len > end - offset)
+		goto err;
+
 	/* Is this the final fragment? */
 	if (offset + skb->len == end) {
 		/* If we already have some bits beyond end
-- 
2.51.2

[-- Attachment #1.2: publickey - rc4d5a@protonmail.com - 0xE775BF83.asc --]
[-- Type: application/pgp-keys, Size: 897 bytes --]

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 322 bytes --]

^ permalink raw reply	[flat|nested] 8+ messages in thread

* [PATCH net RESEND] ieee802154: 6lowpan: validate fragment offset against datagram size
  2026-11-07 22:32 [PATCH] ieee802154: 6lowpan: validate fragment offset against datagram size Ryan Prinz
  2026-11-08 13:05 ` [PATCH RESEND] " Ryan Prinz
@ 2026-11-08 14:20 ` Ryan Prinz
  2026-11-09  9:03 ` Ryan Prinz
  2 siblings, 0 replies; 8+ messages in thread
From: Ryan Prinz @ 2026-11-08 14:20 UTC (permalink / raw)
  To: linux-wpan, netdev
  Cc: Alexander Aring, Stefan Schmidt, Miquel Raynal, David S. Miller,
	Eric Dumazet, Jakub Kicinski, Paolo Abeni, Simon Horman,
	linux-kernel, Ryan Prinz


[-- Attachment #1.1: Type: text/plain, Size: 3399 bytes --]

A 6LoWPAN fragment announces the total size of the datagram it belongs
to (d_size) and its own offset within that datagram. The reassembly
code never checks that the offset actually fits inside d_size, so a
fragment claiming to start far past the end of the datagram is
accepted the same as a legitimate one, allowing an attacker to forge
data of another sender.

lowpan_frag_queue() considers a datagram complete by counting bytes
received (fq->q.meat) against the declared size (fq->q.len), not by
verifying the bytes form a contiguous range. A fragment with an
out-of-range offset still has its length counted toward that total, so
a legitimate fragment exchange plus an attacker fragment can sum to
d_size, marking the datagram complete while a gap in the data was
never filled, and appending the out-of-range fragment's bytes to the
end of the reassembled buffer as if they were part of the message.

IPv6's reassembly (net/ipv6/reassembly.c, ip6_frag_queue()) computes
end = of
fset + payload length for every fragment and checks end
against IPV6_MAXPLEN, which bounds offset by construction. 6LoWPAN's
d_offset and d_size are independent wire fields with no such coupling,
and no separate bounds check was added for it.

The same class of bug was assigned CVE-2019-8359 (CVSS 9.8) in
Contiki-NG's 6LoWPAN stack before a 2019 fix; RIOT-OS's implementation
(gnrc_sixlowpan_frag_rb_add) rejects a fragment once offset + size
exceeds the reassembly buffer's datagram_size. Linux's 6LoWPAN
reassembly is the only one of these implementations without an
equivalent check.

Fix by rejecting any fragment whose offset, or offset plus length,
falls outside [0, d_size).

Reproduced two ways: with a local kernel module calling
lowpan_frag_rcv() directly, and over mac802154_hwsim with three
independent processes standing in for sender, receiver and attacker.
An attacker in radio range who observes one fragment of an
in-progress transmission reads d_tag, d_size and
 the addressing in the
clear, then sends a single out-of-range fragment of its own. The
receiver's reassembly completes with the attacker's bytes appended to
the legitimate data. Both reproducers were re-run against this fix to
confirm the out-of-range fragment is now rejected and the datagram no
longer completes with it included.

AI Disclosure: An AI assistant (Claude) was used to assist with the
analysis and testing. I have verified the vulnerability, the fix, and
the test results myself and take responsibility for this contribution,
per Documentation/process/coding-assistants.rst.

Fixes: 7240cdec60b1 ("6lowpan: handling 6lowpan fragmentation via inet_frag api")
Signed-off-by: Ryan Prinz <rc4d5a@protonmail.com>
---
Resending with correct tree designator in the subject

 net/ieee802154/6lowpan/reassembly.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/net/ieee802154/6lowpan/reassembly.c b/net/ieee802154/6lowpan/reassembly.c
index ddb6a5817d09..905ba5f86
844 100644
--- a/net/ieee802154/6lowpan/reassembly.c
+++ b/net/ieee802154/6lowpan/reassembly.c
@@ -103,6 +103,9 @@ static int lowpan_frag_queue(struct lowpan_frag_queue *fq,
 	offset = lowpan_802154_cb(skb)->d_offset << 3;
 	end = lowpan_802154_cb(skb)->d_size;
 
+	if (offset > end || skb->len > end - offset)
+		goto err;
+
 	/* Is this the final fragment? */
 	if (offset + skb->len == end) {
 		/* If we already have some bits beyond end
-- 
2.51.2

[-- Attachment #1.2: publickey - rc4d5a@protonmail.com - 0xE775BF83.asc --]
[-- Type: application/pgp-keys, Size: 897 bytes --]

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 322 bytes --]

^ permalink raw reply	[flat|nested] 8+ messages in thread

* [PATCH net RESEND] ieee802154: 6lowpan: validate fragment offset against datagram size
  2026-11-07 22:32 [PATCH] ieee802154: 6lowpan: validate fragment offset against datagram size Ryan Prinz
  2026-11-08 13:05 ` [PATCH RESEND] " Ryan Prinz
  2026-11-08 14:20 ` [PATCH net " Ryan Prinz
@ 2026-11-09  9:03 ` Ryan Prinz
  2026-11-12  9:05   ` netdev-bot+sashiko
  2026-11-12 15:31   ` [PATCH net v2] " Ryan Prinz
  2 siblings, 2 replies; 8+ messages in thread
From: Ryan Prinz @ 2026-11-09  9:03 UTC (permalink / raw)
  To: linux-wpan, netdev
  Cc: Alexander Aring, Stefan Schmidt, Miquel Raynal, David S. Miller,
	Eric Dumazet, Jakub Kicinski, Paolo Abeni, Simon Horman,
	linux-kernel, Ryan Prinz

A 6LoWPAN fragment announces the total size of the datagram it belongs
to (d_size) and its own offset within that datagram. The reassembly
code never checks that the offset actually fits inside d_size, so a
fragment claiming to start far past the end of the datagram is
accepted the same as a legitimate one, allowing an attacker to forge
data of another sender.

lowpan_frag_queue() considers a datagram complete by counting bytes
received (fq->q.meat) against the declared size (fq->q.len), not by
verifying the bytes form a contiguous range. A fragment with an
out-of-range offset still has its length counted toward that total, so
a legitimate fragment exchange plus an attacker fragment can sum to
d_size, marking the datagram complete while a gap in the data was
never filled, and appending the out-of-range fragment's bytes to the
end of the reassembled buffer as if they were part of the message.

IPv6's reassembly (net/ipv6/reassembly.c, ip6_frag_queue()) computes
end = offset + payload length for every fragment and checks end
against IPV6_MAXPLEN, which bounds offset by construction. 6LoWPAN's
d_offset and d_size are independent wire fields with no such coupling,
and no separate bounds check was added for it.

The same class of bug was assigned CVE-2019-8359 (CVSS 9.8) in
Contiki-NG's 6LoWPAN stack before a 2019 fix; RIOT-OS's implementation
(gnrc_sixlowpan_frag_rb_add) rejects a fragment once offset + size
exceeds the reassembly buffer's datagram_size. Linux's 6LoWPAN
reassembly is the only one of these implementations without an
equivalent check.

Fix by rejecting any fragment whose offset, or offset plus length,
falls outside [0, d_size).

Reproduced two ways: with a local kernel module calling
lowpan_frag_rcv() directly, and over mac802154_hwsim with three
independent processes standing in for sender, receiver and attacker.
An attacker in radio range who observes one fragment of an
in-progress transmission reads d_tag, d_size and the addressing in the
clear, then sends a single out-of-range fragment of its own. The
receiver's reassembly completes with the attacker's bytes appended to
the legitimate data. Both reproducers were re-run against this fix to
confirm the out-of-range fragment is now rejected and the datagram no
longer completes with it included.

AI Disclosure: An AI assistant (Claude) was used to assist with the
analysis and testing. I have verified the vulnerability, the fix, and
the test results myself and take responsibility for this contribution,
per Documentation/process/coding-assistants.rst.

Fixes: 7240cdec60b1 ("6lowpan: handling 6lowpan fragmentation via inet_frag api")
Signed-off-by: Ryan Prinz <rc4d5a@protonmail.com>
---
Resending with correct tree designator in the subject

 net/ieee802154/6lowpan/reassembly.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/net/ieee802154/6lowpan/reassembly.c b/net/ieee802154/6lowpan/reassembly.c
index ddb6a5817d09..905ba5f86844 100644
--- a/net/ieee802154/6lowpan/reassembly.c
+++ b/net/ieee802154/6lowpan/reassembly.c
@@ -103,6 +103,9 @@ static int lowpan_frag_queue(struct lowpan_frag_queue *fq,
 	offset = lowpan_802154_cb(skb)->d_offset << 3;
 	end = lowpan_802154_cb(skb)->d_size;
 
+	if (offset > end || skb->len > end - offset)
+		goto err;
+
 	/* Is this the final fragment? */
 	if (offset + skb->len == end) {
 		/* If we already have some bits beyond end
-- 
2.51.2


^ permalink raw reply	[flat|nested] 8+ messages in thread

* Re: [PATCH net RESEND] ieee802154: 6lowpan: validate fragment offset against datagram size
  2026-11-09  9:03 ` Ryan Prinz
@ 2026-11-12  9:05   ` netdev-bot+sashiko
  2026-11-12 15:29     ` Ryan Prinz
  2026-11-12 15:31   ` [PATCH net v2] " Ryan Prinz
  1 sibling, 1 reply; 8+ messages in thread
From: netdev-bot+sashiko @ 2026-11-12  9:05 UTC (permalink / raw)
  To: rc4d5a
  Cc: linux-wpan, netdev, alex.aring, stefan, miquel.raynal, davem,
	edumazet, kuba, pabeni, horms, linux-kernel

Thank you for your contribution! Sashiko AI review found 1 potential
issue(s) to consider.

Critical: 0 · High: 0 · Medium: 1 · Low: 0

If you believe none of the issues are legitimate, please reply to each
issue of Medium severity or higher and explain why the AI is wrong.
If any of them are legitimate and you plan to respin, please let patchwork
know by including "pw-bot: cr" as a separate line at the end of your reply
(one such reply per series is enough).

- [Medium] The commit message overstates the security impact and
  misstates the check it adds.

--

ieee802154: 6lowpan: validate fragment offset against datagram size

This commit adds a bounds check to lowpan_frag_queue() that rejects a
6LoWPAN fragment when its offset, or its offset plus length, goes past
the declared datagram size (d_size). The commit message says this stops
an attacker from forging another sender's data. It also compares the
bug to CVE-2019-8359 in Contiki-NG.

> fragment claiming to start far past the end of the datagram is
> accepted the same as a legitimate one, allowing an attacker to forge
> data of another sender.

[ ... ]

> The same class of bug was assigned CVE-2019-8359 (CVSS 9.8) in
> Contiki-NG's 6LoWPAN stack before a 2019 fix; RIOT-OS's implementation

[ ... ]

> Fix by rejecting any fragment whose offset, or offset plus length,
> falls outside [0, d_size).

[ ... ]

> diff --git a/net/ieee802154/6lowpan/reassembly.c b/net/ieee802154/6lowpan/reassembly.c
> index ddb6a5817d098..905ba5f868442 100644
> --- a/net/ieee802154/6lowpan/reassembly.c
> +++ b/net/ieee802154/6lowpan/reassembly.c
> @@ -103,6 +103,9 @@ static int lowpan_frag_queue(struct lowpan_frag_queue *fq,
>  	offset = lowpan_802154_cb(skb)->d_offset << 3;
>  	end = lowpan_802154_cb(skb)->d_size;
>  
> +	if (offset > end || skb->len > end - offset)
> +		goto err;
> +

[Severity: Medium]
Does the commit message match what this check rejects and what it
protects against?

The check accepts offset + skb->len == end, which it has to for the
final fragment. It also accepts a zero-length FRAGN with offset ==
d_size. So the fragment end has a closed upper bound, not the
"[0, d_size)" range the message gives.

ip6_frag_queue() drops zero-length fragments:

	if (end == offset)
		goto discard_fq;

lowpan_frag_queue() still accepts them. They add nothing to fq->q.meat,
so they can't append any data.

About "allowing an attacker to forge data of another sender": the
message assumes the attacker can read d_tag, d_size and the addressing
in the clear. With that, can't the attacker still send an in-range
fragment at the next expected offset? For example, with d_size = 200:

  FRAG1 [0, 64)       legitimate, queued
  FRAGN [64, 136)     attacker, passes the new check
  FRAGN [64, 136)     legitimate, rejected as an overlap
  FRAGN [136, 200)    legitimate, completes the datagram

The attacker's fragment takes the append path in
inet_frag_queue_insert():

	else if (FRAG_CB(last)->ip_defrag_offset + last->len < end) {
		/* This is the common case: skb goes to the end. */
		/* Detect and discard overlaps. */
		if (offset < FRAG_CB(last)->ip_defrag_offset + last->len)
			return IPFRAG_OVERLAP;
		if (offset == FRAG_CB(last)->ip_defrag_offset + last->len)
			fragrun_append_to_last(q, skb);

The legitimate [64, 136) fragment then gets IPFRAG_OVERLAP and is freed
through goto err in lowpan_frag_queue(). The reassembled datagram ends
up holding the attacker's bytes.

Also, 802.15.4 has no source authentication without llsec, so an
attacker can just spoof a whole datagram. This bug doesn't really enable
forging another sender's data, and the patch doesn't prevent it.

About "The same class of bug was assigned CVE-2019-8359 (CVSS 9.8)": the
Contiki-NG bug was an out-of-bounds write into a fixed reassembly
buffer. Linux never copies payload to a computed offset.
inet_frag_reasm_finish() links the fragment skbs to the head:

				head->data_len += fp->len;
				head->len += fp->len;
				head->truesize += fp->truesize;

				*nextp = fp;

So before this patch, an out-of-range offset only changed the contents
of the reassembled payload. It never touched memory outside it. Could
the comparison and the 9.8 score come out of the message? They may
mislead stable and CVE triage.

The change itself looks like a reasonable integrity and robustness fix:
a reassembled datagram can no longer contain bytes past the declared
d_size. Could the message describe it that way, and give the fragment
end's upper bound as closed?

-- 
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20261109090226.799973-1-rc4d5a%40protonmail.com

^ permalink raw reply	[flat|nested] 8+ messages in thread

* Re: [PATCH net RESEND] ieee802154: 6lowpan: validate fragment offset against datagram size
  2026-11-12  9:05   ` netdev-bot+sashiko
@ 2026-11-12 15:29     ` Ryan Prinz
  0 siblings, 0 replies; 8+ messages in thread
From: Ryan Prinz @ 2026-11-12 15:29 UTC (permalink / raw)
  To: netdev-bot+sashiko
  Cc: Alexander Aring, David S . Miller, Eric Dumazet, Simon Horman,
	Jakub Kicinski, linux-kernel, linux-wpan, Miquel Raynal, netdev,
	Paolo Abeni, Ryan Prinz, Stefan Schmidt

Thanks for the review! These are fair points, I'll respin.

- The accepted offset range has a closed upper bound (offset + length
  == d_size is allowed, as needed for the final fragment, including a
  zero-length one at offset == d_size), not the open [0, d_size) I
  described.
- 802.15.4 has no source authentication without llsec, so this bug
  doesn't let an attacker impersonate another sender as the message
  implied. The real effect is an attacker's fragment getting spliced
  into the reassembled datagram in place of a legitimate sender's
  bytes for a given span, via the overlap-eviction path you described.
- The CVE-2019-8359/Contiki-NG comparison doesn't match. That was an
  out-of-bounds write into a fixed buffer. This is a reassembly
  data-integrity issue only, since inet_frag_reasm_finish() links skbs
  by pointer rather than copying into a computed offset. Dropping that
  comparison.

Sending v2 with the commit message corrected; the fix itself is
unchanged.

pw-bot: cr


^ permalink raw reply	[flat|nested] 8+ messages in thread

* [PATCH net v2] ieee802154: 6lowpan: validate fragment offset against datagram size
  2026-11-09  9:03 ` Ryan Prinz
  2026-11-12  9:05   ` netdev-bot+sashiko
@ 2026-11-12 15:31   ` Ryan Prinz
  2026-11-16 15:44     ` netdev-bot+sashiko
  1 sibling, 1 reply; 8+ messages in thread
From: Ryan Prinz @ 2026-11-12 15:31 UTC (permalink / raw)
  To: linux-wpan, netdev
  Cc: Alexander Aring, Stefan Schmidt, Miquel Raynal, David S . Miller,
	Eric Dumazet, Jakub Kicinski, Paolo Abeni, Simon Horman,
	linux-kernel, Ryan Prinz

A 6LoWPAN fragment announces the total size of the datagram it belongs
to (d_size) and its own offset within that datagram. The reassembly
code never checks that the offset actually fits inside d_size, so a
fragment claiming to start past the end of the datagram is accepted
the same as a legitimate one.

lowpan_frag_queue() considers a datagram complete by counting bytes
received (fq->q.meat) against the declared size (fq->q.len), not by
verifying the bytes form a contiguous range. A fragment with an
out-of-range offset still has its length counted toward that total.
If an attacker's out-of-range fragment lands in a gap before the
legitimate fragment for that span arrives, inet_frag_queue_insert()
accepts the attacker's bytes into that span; when the legitimate
fragment for the same span then arrives, it collides with the
attacker's skb and is dropped as an overlap (IPFRAG_OVERLAP). The
datagram still completes, but with the attacker's bytes in the
reassembled buffer instead of the legitimate sender's.

IPv6's reassembly (net/ipv6/reassembly.c, ip6_frag_queue()) computes
end = offset + payload length for every fragment and checks end
against IPV6_MAXPLEN, which bounds offset by construction. 6LoWPAN's
d_offset and d_size are independent wire fields with no such coupling,
and no separate bounds check was added for it. RIOT-OS's 6LoWPAN
implementation (gnrc_sixlowpan_frag_rb_add) already rejects a fragment
once offset + size exceeds the reassembly buffer's datagram_size;
Linux's 6LoWPAN reassembly had no equivalent check.

Fixed by rejecting any fragment whose offset exceeds d_size, or whose
offset plus length would exceed d_size. A final fragment ending
exactly at d_size, including a zero-length one at offset == d_size, is
still accepted, matching the completion check right below it.

Fixes: 7240cdec60b1 ("6lowpan: handling 6lowpan fragmentation via inet_frag api")
Signed-off-by: Ryan Prinz <rc4d5a@protonmail.com>
---
v2:
 - Reworded the impact description. 802.15.4 has no source
   authentication without llsec, so this bug doesn't let an attacker
   impersonate another sender; the real effect is that an attacker's
   fragment can get spliced into the reassembled datagram in place of
   a legitimate sender's bytes, with the legitimate fragment for that
   span evicted as an overlap.
 - Dropped the CVE-2019-8359/Contiki-NG comparison. That bug was an
   out-of-bounds write into a fixed buffer; this one only corrupts the
   reassembled payload's contents, since inet_frag_reasm_finish()
   links fragment skbs by pointer rather than copying into a computed
   offset. The two aren't the same severity class.
 - Corrected the described accepted range: the check allows
   offset + length == d_size, which it must for the final fragment
   (including a zero-length one at offset == d_size), so the accepted
   range has a closed upper bound, not the open "[0, d_size)" the v1
   message stated.

No change to the fix itself; the diff in reassembly.c is unchanged
from v1.

Reproduced two ways: with a local kernel module calling
lowpan_frag_rcv() directly, and over mac802154_hwsim with three
independent processes standing in for sender, receiver and attacker.
An attacker in radio range who observes one fragment of an
in-progress transmission reads d_tag, d_size and the addressing in the
clear, then sends a single out-of-range fragment of its own before the
legitimate fragment for that span arrives. The receiver's reassembly
completes with the attacker's bytes in place of the legitimate
sender's for that span, and the legitimate fragment is dropped as an
overlap. Both reproducers were re-run against this fix to confirm the
out-of-range fragment is now rejected outright.

AI Disclosure: An AI assistant (Claude) was used to assist with the
analysis and testing. I have verified the vulnerability, the fix, and
the test results myself and take responsibility for this contribution,
per Documentation/process/coding-assistants.rst.

 net/ieee802154/6lowpan/reassembly.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/net/ieee802154/6lowpan/reassembly.c b/net/ieee802154/6lowpan/reassembly.c
index ddb6a5817d09..905ba5f86844 100644
--- a/net/ieee802154/6lowpan/reassembly.c
+++ b/net/ieee802154/6lowpan/reassembly.c
@@ -103,6 +103,9 @@ static int lowpan_frag_queue(struct lowpan_frag_queue *fq,
 	offset = lowpan_802154_cb(skb)->d_offset << 3;
 	end = lowpan_802154_cb(skb)->d_size;

+	if (offset > end || skb->len > end - offset)
+		goto err;
+
 	/* Is this the final fragment? */
 	if (offset + skb->len == end) {
 		/* If we already have some bits beyond end
--
2.51.2


^ permalink raw reply	[flat|nested] 8+ messages in thread

* Re: [PATCH net v2] ieee802154: 6lowpan: validate fragment offset against datagram size
  2026-11-12 15:31   ` [PATCH net v2] " Ryan Prinz
@ 2026-11-16 15:44     ` netdev-bot+sashiko
  0 siblings, 0 replies; 8+ messages in thread
From: netdev-bot+sashiko @ 2026-11-16 15:44 UTC (permalink / raw)
  To: rc4d5a
  Cc: linux-wpan, netdev, alex.aring, stefan, miquel.raynal, davem,
	edumazet, kuba, pabeni, horms, linux-kernel

Thank you for your contribution! Sashiko AI review found 2 potential
issue(s) to consider.

Critical: 0 · High: 1 (1) · Medium: 1 · Low: 0

If you believe none of the issues are legitimate, please reply to each
issue of Medium severity or higher and explain why the AI is wrong.
If any of them are legitimate and you plan to respin, please let patchwork
know by including "pw-bot: cr" as a separate line at the end of your reply
(one such reply per series is enough).

- [High] ieee802154: 6lowpan: zero-length fragments accepted in
  lowpan_frag_queue()
- [Medium] The commit message gets both the impact and the mechanism
  wrong, and overstates what the patch protects against.

Pre-existing issues:
- [High] This problem wasn't introduced by this patch, but it is in the
  code path the patch's scenario depends on.

--

ieee802154: 6lowpan: validate fragment offset against datagram size

This patch adds a bounds check to lowpan_frag_queue(). The check rejects
any 6LoWPAN fragment whose offset is greater than d_size, or whose offset
plus length is greater than d_size. The commit message says this stops an
attacker's out-of-range fragment from being spliced into a reassembled
datagram in place of the legitimate sender's bytes.

> If an attacker's out-of-range fragment lands in a gap before the
> legitimate fragment for that span arrives, inet_frag_queue_insert()
> accepts the attacker's bytes into that span; when the legitimate
> fragment for the same span then arrives, it collides with the
> attacker's skb and is dropped as an overlap (IPFRAG_OVERLAP). The
> datagram still completes, but with the attacker's bytes in the
> reassembled buffer instead of the legitimate sender's.

[Severity: Medium]

Does this describe what happens to a fragment that starts past d_size?

lowpan_frag_queue() passes end = d_size to inet_frag_queue_insert(). If
the attacker's node sits at an offset greater than d_size, a later
legitimate fragment always takes the left branch in the binary search:

net/ipv4/inet_fragment.c:inet_frag_queue_insert() {
    ...
			if (end <= FRAG_CB(curr)->ip_defrag_offset)
				rbn = &parent->rb_left;
    ...
}

So nothing is dropped as an overlap. The out-of-range fragment only adds
to fq->q.meat. In most cases meat == len is then never reached, and the
queue waits for the timeout.

A DUP/OVERLAP collision can only happen for the other half of the check
(offset < d_size, offset + len > d_size). Even then, one such fragment
pushes meat past d_size. Doesn't "the datagram still completes" need
more conditions than the message gives?

Also, the new check still accepts an in-range forged fragment that gives
the same splice. Take d_size = 96, with a legitimate FRAG1 covering
[0, 48) already queued. An attacker sends a FRAGN with the same d_tag,
d_size and addresses, d_offset = 6 and 48 bytes of payload:

    offset = 48, skb->len = 48, end = 96
    48 > 96 is false, 48 > 96 - 48 is false  -> check passes
    48 + 48 == 96                            -> final, LAST_IN set

inet_frag_queue_insert() appends it on the common path. meat reaches 96,
and lowpan_frag_reasm() completes the datagram with the attacker's bytes.

802.15.4 has no source authentication without llsec, so no reassembly
bounds check can prevent that.

The check still works as hardening against malformed fragments. Could the
commit message be reworded so readers and stable backporters don't take
it as a fix for third-party fragment splicing?

> diff --git a/net/ieee802154/6lowpan/reassembly.c b/net/ieee802154/6lowpan/reassembly.c
> index ddb6a5817d098..905ba5f868442 100644
> --- a/net/ieee802154/6lowpan/reassembly.c
> +++ b/net/ieee802154/6lowpan/reassembly.c
> @@ -103,6 +103,9 @@ static int lowpan_frag_queue(struct lowpan_frag_queue *fq,
>  	offset = lowpan_802154_cb(skb)->d_offset << 3;
>  	end = lowpan_802154_cb(skb)->d_size;
>  
> +	if (offset > end || skb->len > end - offset)
> +		goto err;
> +

[Severity: High]
Should this also reject skb->len == 0?

The new check lets a zero-length fragment through at any offset up to
d_size. The commit message also says a zero-length final fragment at
offset == d_size is meant to be accepted. ip_frag_queue(),
ip6_frag_queue() and nf_ct_frag6_queue() all treat end == offset as
invalid, and drop the fragment before it reaches
inet_frag_queue_insert().

Empty fragments are easy to produce here:

  - lowpan_get_cb() only checks !skb->len on the FRAG1 branch. A FRAGN
    frame carrying nothing but its 5-byte header therefore reaches
    lowpan_frag_queue() with skb->len == 0.

  - A FRAG1 whose only payload is the LOWPAN_DISPATCH_IPV6 byte passes
    lowpan_get_cb() with skb->len == 1. lowpan_rx_h_ipv6() then pulls
    that byte, so this fragment also arrives here empty.

Take an empty FRAGN with d_offset = 0 that arrives before FRAG1. It
becomes the first node in fq->q.rb_fragments, with ip_defrag_offset 0
and len 0. When the real FRAG1 arrives at offset 0, the common path in
inet_frag_queue_insert() sees

    offset == FRAG_CB(last)->ip_defrag_offset + last->len   (0 == 0)

It calls fragrun_append_to_last(), so FRAG1 is chained behind the
empty skb.

Once meat reaches len, inet_frag_reasm_prepare() uses the first
rb_fragments node as the head, which is the empty FRAGN.
inet_frag_reasm_finish() then does

    skb_push(head, head->data - skb_network_header(head));

That pushes the pulled FRAGN header back in front of the packet. The
decompressed IPv6 header from FRAG1 sits behind it in frag_list. The
dispatch-only FRAG1 case does the same with the 0x41 byte. In both
cases the reassembled packet passed up the stack no longer starts with
its IPv6 header.

Empty fragments at offset == d_size cause a different problem. The
first one sets INET_FRAG_LAST_IN before the real final fragment has
arrived. Repeats go down the binary search in inet_frag_queue_insert(),
where "end <= ip_defrag_offset" holds (d_size <= d_size) before the
IPFRAG_DUP test is reached. Each copy is linked into rb_fragments and
charged through add_frag_mem_limit(). fq->q.meat never moves, so the
memory stays held until the reassembly timer fires.

At an intermediate offset, an empty fragment can do one of two things.
It can be chained onto the tail run without advancing frag_run_len.
Or it can start a zero-length run, so that a later legitimate fragment
below it fails with IPFRAG_OVERLAP.

Would it make sense to add !skb->len to this check, and drop the
zero-length wording from the commit message? Putting it here rather
than in lowpan_get_cb() also covers the FRAG1 case, because that skb
only becomes empty after the rx handlers run.

This matters even more if the insert end is changed to
offset + skb->len, because inet_frag_queue_insert() assumes
offset < end.

[Severity: High]

This is a pre-existing issue and was not introduced by this patch, but
the scenario in the commit message depends on it. Further down,
lowpan_frag_queue() does:

	err = inet_frag_queue_insert(&fq->q, skb, offset, end);

Here end is d_size, the size of the whole datagram, rather than the
fragment's own end (offset + skb->len). ip_frag_queue() and
ip6_frag_queue() both pass the fragment's own end:

net/ipv6/reassembly.c:ip6_frag_queue() {
    ...
	end = offset + (ntohs(ipv6_hdr(skb)->payload_len) -
			((u8 *)(fhdr + 1) - (u8 *)(ipv6_hdr(skb) + 1)));
    ...
}

Doesn't this make inet_frag_queue_insert() treat every 6LoWPAN fragment
as covering [offset, d_size)?

On the common path, the "last_end < end" test turns into
"last_end < d_size". Any fragment with offset < last_end then returns
IPFRAG_OVERLAP.

On the binary search path, "end <= curr offset" is never true for an
in-range node. The walk can never go left, so it ends in IPFRAG_DUP or
IPFRAG_OVERLAP.

Trace 1, d_size = 96, fragments arrive out of order:

    FRAGN d_offset=6, len 48  -> final, run [48, 96)
    FRAG1 offset 0, len 48    -> last_end 96 is not < 96, binary search
                                 96 <= 48 false, 0 >= 96 false,
                                 0 >= 48 false -> IPFRAG_OVERLAP

FRAG1 is freed through goto err and INET_FRAG_FIRST_IN is never set, so
the datagram can never complete.

Trace 2, four 24-byte fragments at offsets 0/24/48/72, arriving in the
order 0, 48, 24:

    fragment 24 -> last_end 72 < 96, offset 24 < 72 -> IPFRAG_OVERLAP

So it looks like any datagram whose fragments arrive out of ascending
order, or with a gap, can never be reassembled. The partial queue keeps
its frag memory until the reassembly timer expires. This seems to date
back to 254c5dbe15d4 ("6lowpan: use rbtree for IP frag queue").

The "dropped as an overlap" behaviour in the commit message looks like a
side effect of this wrong interval, not real overlap detection.

Would it make sense to also pass offset + skb->len as the insert end,
and use d_size only for q.len and the completion check?

-- 
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20261112153034.1015181-1-rc4d5a%40protonmail.com

^ permalink raw reply	[flat|nested] 8+ messages in thread

end of thread, other threads:[~2026-11-16 15:44 UTC | newest]

Thread overview: 8+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-11-07 22:32 [PATCH] ieee802154: 6lowpan: validate fragment offset against datagram size Ryan Prinz
2026-11-08 13:05 ` [PATCH RESEND] " Ryan Prinz
2026-11-08 14:20 ` [PATCH net " Ryan Prinz
2026-11-09  9:03 ` Ryan Prinz
2026-11-12  9:05   ` netdev-bot+sashiko
2026-11-12 15:29     ` Ryan Prinz
2026-11-12 15:31   ` [PATCH net v2] " Ryan Prinz
2026-11-16 15:44     ` netdev-bot+sashiko

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox

all inboxes | Powered by JetHome®