During the investigation of a kernel panic, we encountered a condition that triggered a kernel panic due to a large skb with an unusual geometry. Inside of the STT codepath, an effort is made to linearize such packets to avoid trouble during both fragment reassembly and segmentation in the linux networking core.
As currently implemented, kernels with CONFIG_SLUB defined will skip this process because it does not expect an skb with a frag_list to be present. This patch removes the assumption, and allows these skb to be linearized as intended. We confirmed this corrects the panic we encountered. Reported-by: Johannes Erdfelt <[email protected]> Reported-at: https://mail.openvswitch.org/pipermail/ovs-discuss/2018-May/046800.html Requested-by: Pravin Shelar <[email protected]> Signed-off-by: Neal Shrader <[email protected]> --- datapath/linux/compat/stt.c | 7 ------- 1 file changed, 7 deletions(-) diff --git a/datapath/linux/compat/stt.c b/datapath/linux/compat/stt.c index ca9f03988..64fc550fe 100644 --- a/datapath/linux/compat/stt.c +++ b/datapath/linux/compat/stt.c @@ -559,12 +559,6 @@ static int __try_to_segment(struct sk_buff *skb, bool csum_partial, static int try_to_segment(struct sk_buff *skb) { -#ifdef SKIP_ZERO_COPY - /* coalesce_skb() since does not generate frag-list no need to - * linearize it here. - */ - return 0; -#else struct stthdr *stth = stt_hdr(skb); bool csum_partial = !!(stth->flags & STT_CSUM_PARTIAL); bool ipv4 = !!(stth->flags & STT_PROTO_IPV4); @@ -572,7 +566,6 @@ static int try_to_segment(struct sk_buff *skb) int l4_offset = stth->l4_offset; return __try_to_segment(skb, csum_partial, ipv4, tcp, l4_offset); -#endif } static int segment_skb(struct sk_buff **headp, bool csum_partial, -- 2.14.1 _______________________________________________ dev mailing list [email protected] https://mail.openvswitch.org/mailman/listinfo/ovs-dev
