[PATCH v2 net] net: skbuff: do not leave stale header offsets after pskb_carve()

[Eric Dumazet <edumazet@google.com>]

pskb_carve_inside_header() and pskb_carve_inside_nonlinear() remove
the first bytes of a packet and reallocate skb->head.

All the headers that were present before the operation are gone,
but both functions call skb_headers_offset_update(skb, 0), which
is a no-op : skb->mac_header, skb->network_header,
skb->transport_header and skb->csum_start keep their old values and
now describe bytes which are no longer there.

Both helpers size the new head from the old skb_end_offset(), so the
stale offsets still land inside the new allocation. They point past
skb_tail_pointer() though, to bytes that were never initialized.

pskb_carve_inside_nonlinear() is the worst case, because it leaves a
zombie skb with an empty linear part (skb->data ==
skb_tail_pointer(skb), skb_headlen(skb) == 0), while
skb_mac_header_was_set() is still true and skb->mac_header is way
ahead of skb->data.

The only user of pskb_extract() is rds_tcp_data_recv(), and the
carved skb is queued on tinc->ti_skb_list. When the RDS incoming
message is released, rds_tcp_inc_free() calls skb_queue_purge(),
which frees the skbs with SKB_DROP_REASON_QUEUE_PURGE. This is
visible from drop_monitor, which then tries to pull back to the
(bogus) mac header :

skbuff: __skb_pull(len=234)
skb len=6968 data_len=6968 headroom=0 headlen=0 tailroom=0
end-tail=384 mac=(234,14) mac_len=14 net=(248,40) trans=288
shinfo(txflags=0 nr_frags=1 gso(size=1428 type=16 segs=5))
csum(0x100120 start=288 offset=16 ip_summed=3 complete_sw=0 valid=1 level=0)
hash(0x7b446c6c sw=0 l4=1) proto=0x86dd pkttype=0 iif=60
------------[ cut here ]------------
kernel BUG at ./include/linux/skbuff.h:2847!

Add skb_carve_reset_headers() to mark the mac and transport headers
as not set, reset the network header, clear skb->mac_len, and drop
a now meaningless CHECKSUM_PARTIAL (csum_start no longer describes
anything).

Invalidate the inner offsets as well. Unlike mac_header and
transport_header they have no "unset" sentinel, so a leftover
non-zero value still looks like a real header. Zero
skb->inner_mac_header, skb->inner_network_header,
skb->inner_transport_header, skb->inner_protocol and
skb->encapsulation, so that all the header state is invalidated in
one place.

v2: fixed an inaccurate changelog. The stale offsets stay inside the
    new skb->head, which is never smaller than the old one, they
    simply point past skb_tail_pointer() to bytes that are gone.
    Thanks to Xuanqiang Luo for insisting on this.
    Also invalidate the inner header state, as suggested by the
    netdev AI review :
    https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260911114922.621937-1-edumazet%40google.com

Fixes: 6fa01ccd8830 ("skbuff: Add pskb_extract() helper function")
Reported-by: syzbot+586af68eb819833c2d91@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/netdev/6aa3e9d3.f2639fcc.29487d.0028.GAE@google.com/
Cc: Xuanqiang Luo <xuanqiang.luo@linux.dev>
Cc: Allison Henderson <achender@kernel.org>
Cc: rds-devel@oss.oracle.com
Signed-off-by: Eric Dumazet <edumazet@google.com>
---
 net/core/skbuff.c | 32 ++++++++++++++++++++++++++++++--
 1 file changed, 30 insertions(+), 2 deletions(-)

diff --git a/net/core/skbuff.c b/net/core/skbuff.c
index cc3b4b70288b4e3f17984cd4f4e4b7a330353e15..609f2c7f4a47ad7a81426af149727fa47be12bb5 100644
--- a/net/core/skbuff.c
+++ b/net/core/skbuff.c
@@ -6832,6 +6832,34 @@ struct sk_buff *alloc_skb_with_frags(unsigned long header_len,
 }
 EXPORT_SYMBOL(alloc_skb_with_frags);
 
+/* pskb_carve_inside_header() and pskb_carve_inside_nonlinear()
+ * remove the first bytes of a packet and reallocate skb->head.
+ *
+ * Whatever headers were present before the operation are gone,
+ * we must not leave stale offsets, otherwise users of this skb
+ * (skb_dump(), drop_monitor, taps, ...) would read or pull garbage.
+ */
+static void skb_carve_reset_headers(struct sk_buff *skb)
+{
+	skb_unset_mac_header(skb);
+	skb_unset_transport_header(skb);
+	skb_reset_network_header(skb);
+	skb->mac_len = 0;
+
+	/* Inner offsets have no "unset" marker, zero them so that
+	 * skb_inner_network_header_was_set() becomes false and no
+	 * consumer mistakes them for a real (and long gone) header.
+	 */
+	skb->inner_mac_header = 0;
+	skb->inner_network_header = 0;
+	skb->inner_transport_header = 0;
+	skb->inner_protocol = 0;
+	skb->encapsulation = 0;
+
+	if (skb->ip_summed == CHECKSUM_PARTIAL)
+		skb->ip_summed = CHECKSUM_NONE;
+}
+
 /* carve out the first off bytes from skb when off < headlen */
 static int pskb_carve_inside_header(struct sk_buff *skb, const u32 off,
 				    const int headlen, gfp_t gfp_mask)
@@ -6887,7 +6915,7 @@ static int pskb_carve_inside_header(struct sk_buff *skb, const u32 off,
 	skb->head_frag = 0;
 	skb_set_end_offset(skb, size);
 	skb_set_tail_pointer(skb, skb_headlen(skb));
-	skb_headers_offset_update(skb, 0);
+	skb_carve_reset_headers(skb);
 	skb->cloned = 0;
 	skb->hdr_len = 0;
 	skb->nohdr = 0;
@@ -7027,7 +7055,7 @@ static int pskb_carve_inside_nonlinear(struct sk_buff *skb, const u32 off,
 	skb->data = data;
 	skb_set_end_offset(skb, size);
 	skb_reset_tail_pointer(skb);
-	skb_headers_offset_update(skb, 0);
+	skb_carve_reset_headers(skb);
 	skb->cloned   = 0;
 	skb->hdr_len  = 0;
 	skb->nohdr    = 0;
-- 
2.55.0.1032.g73a4cd73de-goog