Repository navigation
Don't split IP-reassembled packets in homa_softirq - #155
Stuartwastaken wants to merge 1 commit into
Conversation
homa_softirq assumed that any frag_list was a batch built by Homa's GRO and processed each entry as a separate Homa packet. IP reassembly also leaves fragments on frag_list, and those are pieces of one packet: their bytes are counted in the head's len, their payload was parsed as Homa headers, and their dev is NULL, so dev_net crashed in homa_dispatch_pkts. Any remote host can send fragmented Homa packets, and in the stripped build, which has no GRO, every frag_list comes from reassembly. Homa's GRO doesn't add batched packets to the head's len, so split the list only when skb->len == skb_pagelen(skb).
|
I don't this fix will work: I think there are other problems that will prevent Homa from correctly handling IP-fragmented packets. For example, Homa's GRO assumes it "owns" the frag list to use for collecting batches, so if the head packet is actually IP-fragmented I suspect that GRO will append additional full packets onto the end of the fragment list. I don't think it's worth Homa's trying to handle fragmented packets (I don't believe that TCP does this either). I believe there are ways to tell the IP stack "do not ever fragment this packet", but Homa doesn't currently attempt to do this. I've added this to notes.txt as a future task; it probably makes sense to wait until after the UDP conversion is complete so the fix can be done in that context. |
homa_softirq treats any frag_list as a batch from Homa's GRO, but IP reassembly also builds one when a Homa packet arrives in fragments. The fragments then get dispatched as separate Homa packets with skb->dev == NULL, so dev_net() faults in homa_dispatch_pkts. Any remote host can trigger this, and in the stripped build (no GRO) a frag_list can only come from reassembly.
Reassembly counts the fragments in skb->len and Homa's GRO doesn't, so this only splits the list when skb->len == skb_pagelen(skb). The new unit test homa_softirq__frag_list_from_ip_reassembly crashes in dev_net under ASan without the fix; all 887 unit tests pass with it.