Jakub Sitnicki wrote:
On Thu, Aug 01, 2024 at 03:13 PM -04, Willem de Bruijn wrote:
quoted
On Thu, Aug 1, 2024 at 10:09 AM Jakub Sitnicki [off-list ref] wrote:
quoted
In commit 10154dbded6d ("udp: Allow GSO transmit from devices with no
checksum offload") we have intentionally allowed UDP GSO packets marked
CHECKSUM_NONE to pass to the GSO stack, so that they can be segmented and
checksummed by a software fallback when the egress device lacks these
features.
What was not taken into consideration is that a CHECKSUM_NONE skb can be
handed over to the GSO stack also when the egress device advertises the
tx-udp-segmentation / NETIF_F_GSO_UDP_L4 feature.
This can happen in two situations, which we detect in __ip_append_data()
and __ip6_append_data():
1) when there are IPv6 extension headers present, or
2) when the tunnel device does not advertise checksum offload.
Note that in the latter case we have a nonsensical device configuration.
Device support for UDP segmentation offload requires checksum offload in
hardware as well.
Syzbot has discovered the first case, producing a warning as below:
ip6tnl0: caps=(0x00000006401d7869, 0x00000006401d7869)
WARNING: CPU: 0 PID: 5112 at net/core/dev.c:3293 skb_warn_bad_offload+0x166/0x1a0 net/core/dev.c:3291
Modules linked in:
CPU: 0 PID: 5112 Comm: syz-executor391 Not tainted 6.10.0-rc7-syzkaller-01603-g80ab5445da62 #0
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 06/07/2024
RIP: 0010:skb_warn_bad_offload+0x166/0x1a0 net/core/dev.c:3291
[...]
Call Trace:
<TASK>
__skb_gso_segment+0x3be/0x4c0 net/core/gso.c:127
skb_gso_segment include/net/gso.h:83 [inline]
validate_xmit_skb+0x585/0x1120 net/core/dev.c:3661
__dev_queue_xmit+0x17a4/0x3e90 net/core/dev.c:4415
neigh_output include/net/neighbour.h:542 [inline]
ip6_finish_output2+0xffa/0x1680 net/ipv6/ip6_output.c:137
ip6_finish_output+0x41e/0x810 net/ipv6/ip6_output.c:222
ip6_send_skb+0x112/0x230 net/ipv6/ip6_output.c:1958
udp_v6_send_skb+0xbf5/0x1870 net/ipv6/udp.c:1292
udpv6_sendmsg+0x23b3/0x3270 net/ipv6/udp.c:1588
sock_sendmsg_nosec net/socket.c:730 [inline]
__sock_sendmsg+0xef/0x270 net/socket.c:745
____sys_sendmsg+0x525/0x7d0 net/socket.c:2585
___sys_sendmsg net/socket.c:2639 [inline]
__sys_sendmmsg+0x3b2/0x740 net/socket.c:2725
__do_sys_sendmmsg net/socket.c:2754 [inline]
__se_sys_sendmmsg net/socket.c:2751 [inline]
__x64_sys_sendmmsg+0xa0/0xb0 net/socket.c:2751
do_syscall_x64 arch/x86/entry/common.c:52 [inline]
do_syscall_64+0xf3/0x230 arch/x86/entry/common.c:83
entry_SYSCALL_64_after_hwframe+0x77/0x7f
[...]
</TASK>
We are hitting the bad offload warning because when an egress device is
capable of handling segmentation offload requested by
skb_shinfo(skb)->gso_type, the chain of gso_segment callbacks won't produce
any segment skbs and return NULL. See the skb_gso_ok() branch in
{__udp,tcp,sctp}_gso_segment helpers.
To fix it, skip bad offload detection when gso_segment has returned
nothing. We know that in such case the egress device supports the desired
GSO offload, which implies that it can fill in L4 checksums. Hence we don't
need to check the skb->ip_summed value, which reflects the egress device
checksum capabilities.
Fixes: 10154dbded6d ("udp: Allow GSO transmit from devices with no checksum offload")
Reported-by: <redacted>
Closes: https://lore.kernel.org/all/000000000000e1609a061d5330ce@google.com/ (local)
Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
It's a bit odd, in that the ip_summed == CHECKSUM_NONE ends up just
being ignored and devices are trusted to always be able to checksum
offload when they can segment offload -- even when the device does not
advertise checksum offload.
I think we should have a follow-on that makes advertising
NETIF_F_GSO_UDP_L4 dependent on having at least one of the
NETIF_F_*_CSUM bits set (handwaving over what happens when only
advertising NETIF_F_IP_CSUM or NETIF_F_IPV6_CSUM).
I agree. I've also gained some clarity as to how the fix should
look. Let's circle back to it, if we still think it's relevant once we
hash out the fix.
After spending some quality time debugging the addded regression test
[1], I've realized this fix is wrong.
You see, with commit 10154dbded6d ("udp: Allow GSO transmit from devices
with no checksum offload"), I've opened up the UDP_SEGMENT API to two
uses, which I think should not be allowed:
1. Hardware USO for IPv6 dgrams with extension headers
Previously that led to -EIO, because __ip6_append_data won't annotate
such packets as CHECKSUM_PARTIAL.
I'm guessing that we do this because some drivers that advertise csum
offload can't actually handle checksumming when extension headers are
present.
Extension headers are not part of IPv6 pseudo header, but who knows what
limitations NIC firmwares have.
Either way, changing it just like that sounds risky, so I think we need
to fall back to software USO with software checksum in this case.
Alternatively, we could catch it in the udp layer and error out with EIO
as ealier. But that shifts some burden onto the user space (detect and
segment before sendmsg()).
2. Hardware USO when hardware csum is unsupported / disabled
That sounds like a pathological device configuration case, but since it
is possible today with some drivers to disable csum offload for one IP
version and not the other, it seems safest to just handle that
gracefully.
I don't know why one might want to do that. Perhaps as a workaround for
some firmware bug while waiting for a fix?
I doubt that this is actually used. But today it can be configured.
Which is why I suggested making NETIF_F_GSO_UDP_L4 dependent on csum
offload (in netdev_fix_features). I doubt that that will break any
real user.
In this scenario I think we also need to fall back to software USO and
checksum.
Code-wise that could look like below. WDYT?
Since this only affects USO, can we fix this is in __udp_gso_segment.
Basically, not taking the NETIF_F_GSO_ROBUST path.
skb_segment is so complicated already. Whatever we can do to avoid
adding to that.
quoted hunk ↗ jump to hunk
[1] I feel silly for submitting a broken patch. I've been running a
stale test prog in a VM. It seems that virtme-ng with a read+write
overlay for rootfs played a trick on me. Changes to the host files are
not (always?) visible to the guest.
---8<---
diff --git a/net/core/skbuff.c b/net/core/skbuff.c
index 83f8cd8aa2d1..819173807c81 100644
--- a/net/core/skbuff.c
+++ b/net/core/skbuff.c
@@ -4806,7 +4806,13 @@ struct sk_buff *skb_segment(struct sk_buff *head_skb,
goto perform_csum_check;
if (!sg) {
- if (!csum) {
+ /* Fall back to software checksum for segments if:
+ * 1) device can't checksum this network protocol, OR
+ * 2) we consider the packet to be not checksummable in
+ * hardware, for example IPv6 extension headers are
+ * present.
+ */
+ if (!csum || head_skb->ip_summed != CHECKSUM_PARTIAL) {
if (!nskb->remcsum_offload)
nskb->ip_summed = CHECKSUM_NONE;
SKB_GSO_CB(nskb)->csum =@@ -4896,7 +4902,7 @@ struct sk_buff *skb_segment(struct sk_buff *head_skb,
nskb->truesize += nskb->data_len;
perform_csum_check:
- if (!csum) {
+ if (!csum || head_skb->ip_summed != CHECKSUM_PARTIAL) {
if (skb_has_shared_frag(nskb) &&
__skb_linearize(nskb))
goto err;diff --git a/net/ipv4/udp_offload.c b/net/ipv4/udp_offload.c
index bc8a9da750fe..0a34b418b83c 100644
--- a/net/ipv4/udp_offload.c
+++ b/net/ipv4/udp_offload.c
@@ -282,7 +282,15 @@ struct sk_buff *__udp_gso_segment(struct sk_buff *gso_skb,
skb_transport_header(gso_skb)))
return ERR_PTR(-EINVAL);
- if (skb_gso_ok(gso_skb, features | NETIF_F_GSO_ROBUST)) {
+ /* Pass-through to device for segmentation only if:
+ * 1) we consider the packet checksummable in hardware, that is no
+ * IPv6 extension headers present, AND
+ * 2) device supports checksum offload for this network protocol
+ * (NETIF_F_{IP,IPV6}_CSUM or NETIF_F_HW_CSUM), AND
+ * 3) device supports the requested GSO kind.
+ */
+ if (gso_skb->ip_summed == CHECKSUM_PARTIAL &&
+ skb_gso_ok(gso_skb, features | NETIF_F_GSO_ROBUST)) {
/* Packet is from an untrusted source, reset gso_segs. */
skb_shinfo(gso_skb)->gso_segs = DIV_ROUND_UP(gso_skb->len - sizeof(*uh),
mss);