"qsort()".
The kernel's "qsort()" routine can in worst case spend O(N*N) amount of
comparisons before the input array is sorted. It can also recurse a
significant amount of times using up the kernel's interrupt thread
stack.
The custom sorting routine takes advantage of that the sorting key is
only 64 bits. Based on set and cleared bits in the sorting key it
partitions the array until it is sorted. This process has a recursion
limit of 64 times, due to the number of set and cleared bits which can
occur. Compiled with -O2 the sorting routine was measured to use
64-bytes of stack. Multiplying this by 64 gives a maximum stack
consumption of 4096 bytes for AMD64. The same applies to the execution
time, that the array to be sorted will not be traversed more than 64
times.
When serving roughly 80Gb/s with 80K TCP connections, the old method
consisting of "qsort()" and "tcp_lro_mbuf_compare_header()" used 1.4%
CPU, while the new "tcp_lro_sort()" used 1.1% for LRO related sorting
as measured by Intel Vtune. The testing was done using a sysctl to
toggle between "qsort()" and "tcp_lro_sort()".
Differential Revision: https://reviews.freebsd.org/D6472
Sponsored by: Mellanox Technologies
Tested by: Netflix
Reviewed by: gallatin, rrs, sephe, transport
This is kinda critical to the performance when the CPU is slow and
network bandwidth is high, e.g. in the hypervisor.
Reviewed by: rrs, gallatin, Dexuan Cui <decui microsoft com>
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5765
So that callers could react accordingly.
Reviewed by: gallatin (no objection)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5695
ACK aggregation limit is append count based, while the TCP data segment
aggregation limit is length based. Unless the network driver sets these
two limits, it's an NO-OP.
Reviewed by: adrian, gallatin (previous version), hselasky (previous version)
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5185
- Add optimizing LRO wrapper which pre-sorts all incoming packets
according to the hash type and flowid. This prevents exhaustion of
the LRO entries due to too many connections at the same time.
Testing using a larger number of higher bandwidth TCP connections
showed that the incoming ACK packet aggregation rate increased from
~1.3:1 to almost 3:1. Another test showed that for a number of TCP
connections greater than 16 per hardware receive ring, where 8 TCP
connections was the LRO active entry limit, there was a significant
improvement in throughput due to being able to fully aggregate more
than 8 TCP stream. For very few very high bandwidth TCP streams, the
optimizing LRO wrapper will add CPU usage instead of reducing CPU
usage. This is expected. Network drivers which want to use the
optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead
of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of
"tcp_lro_flush()". Further the LRO control structure must be
initialized using "tcp_lro_init_args()" passing a non-zero number
into the "lro_mbufs" argument.
- Make LRO statistics 64-bit. Previously 32-bit integers were used for
statistics which can be prone to wrap-around. Fix this while at it
and update all SYSCTL's which expose LRO statistics.
- Ensure all data is freed when destroying a LRO control structures,
especially leftover LRO entries.
- Reduce number of memory allocations needed when setting up a LRO
control structure by precomputing the total amount of memory needed.
- Add own memory allocation counter for LRO.
- Bump the FreeBSD version to force recompilation of all KLDs due to
change of the LRO control structure size.
Sponsored by: Mellanox Technologies
Reviewed by: gallatin, sbruno, rrs, gnn, transport
Tested by: Netflix
Differential Revision: https://reviews.freebsd.org/D4914
Add a last-modified timestamp to each LRO entry and provide an interface
to flush all inactive entries. Drivers decide when to flush and what
the inactivity threshold should be.
Network drivers that process an rx queue to completion can enter a
livelock type situation when the rate at which packets are received
reaches equilibrium with the rate at which the rx thread is processing
them. When this happens the final LRO flush (normally when the rx
routine is done) does not occur. Pure ACKs and segments with total
payload < 64K can get stuck in an LRO entry. Symptoms are that TCP
tx-mostly connections' performance falls off a cliff during heavy,
unrelated rx on the interface.
Flushing only inactive LRO entries works better than any of these
alternates that I tried:
- don't LRO pure ACKs
- flush _all_ LRO entries periodically (every 'x' microseconds or every
'y' descriptors)
- stop rx processing in the driver periodically and schedule remaining
work for later.
Reviewed by: andre
Significantly update tcp_lro for mostly two things:
1) introduce basic support for IPv6 without extension headers.
2) try hard to also get the incremental checksum updates right,
especially also in the IPv4 case for the IP and TCP header.
Move variables around for better locality, factor things out into
functions, allow checksum updates to be compiled out, ...
Leave a few comments on further things to look at in the future,
though that is not the full list.
Update drivers with appropriate #includes as needed for IPv6 data
type in LRO.
Sponsored by: The FreeBSD Foundation
Sponsored by: iXsystems
Reviewed by: gnn (as part of the whole)
MFC After: 3 days
