site_archiver@lists.apple.com Delivered-To: darwin-kernel@lists.apple.com Anton Kuzmin writes:

Hello,

I wonder if this is a known problem (or me missing something obvious) - outgoing TCP packets to be sent via the built-in ethernet device (GMace) appear to have only MBUF_CSUM_REQ_SUM16 hardware checksum request flag set as seen by an output KPI interface filter, which needs to modify the packet and therefore calls mbuf_outbound_finalize to do the delayed checksums. But from the xnu source it seems that this function has no knowledge of this flag and does nothing to calculate the checksum.

It sounds like a bug.

From what I understand, CSUM_TCP_SUM16 is just a special case of CSUM_TCP. See bsd/netinet/ip_output.c. You might try just "silently upgrading" to MBUF_CSUM_REQ_TCP when you call mbuf_outbound_finalize().

That brings up another annoying point, which is that mbuf_set_csum_requested() is marked KERNEL_PRIVATE, and not in the headers (though it is callable). I discovered this yesterday when implementing my own mbuf compression code to work around getPhysicalSegmentsWithCoalesce() being broken. Another bug in mbuf_outbound_finalize() (actually in in_delayed_cksum_offset()) is that it only works on big endian hosts. On little endian hosts, ip->ip_len must be byteswapped back into host byte order, otherwise in_delayed_cksum_offset() will call in_cksum_skip() with a bogus ip->ip_len. I learned this the hard way when writing my own sort of mbuf_outbound_finalize() for a driver on FreeBSD quite a while ago. Drew _______________________________________________ Do not post admin requests to the list. They will be ignored. Darwin-kernel mailing list (Darwin-kernel@lists.apple.com) Help/Unsubscribe/Update your Subscription: http://lists.apple.com/mailman/options/darwin-kernel/site_archiver%40lists.a... This email sent to site_archiver@lists.apple.com