On 04/11/20 2:16 AM, Jakub Kicinski wrote:
On Fri, 30 Oct 2020 23:32:20 +0530 Rohit Maheshwari wrote:
Creating SKB per tls record and freeing the original one causes
panic. There will be race if connection reset is requested. By
freeing original skb, refcnt will be decremented and that means,
there is no pending record to send, and so tls_dev_del will be
requested in control path while SKB of related connection is in
queue.
Better approach is to use same SKB to send one record (partial
data) at a time. We still have to create a new SKB when partial
last part of a record is requested.
This fix introduces new API cxgb4_write_partial_sgl() to send
partial part of skb. Present cxgb4_write_sgl can only provide
feasibility to start from an offset which limits to header only
and it can write sgls for the whole skb len. But this new API
will help in both. It can start from any offset and can end
writing in middle of the skb.
You never replied to my question on v2.
If the problem is that the socket gets freed, why don't you make the
new skb take a reference on the socket?
650 LoC is really a rather large fix.
This whole skb alloc and copy record frags was written under the
assumption that there will be zero data copy (no linear skb was
expected) but that isn't correct. Continuing with the same change
requires more checks and will be more complicated. That's why I
made this change. I think using same SKB to send out multiple
records is better than allocating new SKB every time.
