When under pressure for mbufs or we have too many fragments in the IP
reassembly queue, drop half of all fragments. This multiplicative-drop
strategy ensures we return to a healthy state, even under borderline
denial-of-service from extremely lossy NFS-over-UDP peers.
The multiplicative-drop phase currently drops 50% of fragments, but
has pre-placed support for implementing drop-fractions other than 50%
The threshhold for the `drop-half' phase is the new variable,
ip_maxfrags which is calculated as nmbclusters/4.
ip_input.c now keeps ip_nmbclusters, a cached copy of nmbclusters.
Before using limits derived from nmbclusters, we check if nmbclusters
and ip_nmclusters are equal. If not, we recompute Ip parameters
derived from nmbclusters. Based on a suggestion by Jason Thorpe.
ip_maxfrags is currently auto-recalcuated.
The counters ip_nfrags and ip_nfragpacketsr are now declared static
and uninitialized (bss), to discourage tampering with them.
The idea is that we only clear M_CANFASTFWD if an SPD exists
for the packet. Otherwise, it's safe to add a fast-forward
cache entry for the route.
To make this work properly, we invalidate the entire ipflow
cache if a fast-ipsec key is added or changed.
to check if interface exists, as (1) if_index has different meaning
(2) ifindex2ifnet could become NULL when interface gets destroyed,
since when we have introduced dynamically-created interfaces. from kame
- seed2 is necessary, but use it as "seed2 + x" not "seed2 ^ x".
- skipping number is not needed, so disable it for 16bit generator (makes
the repetition period to 30000)
Gone are the old kern_sysctl(), cpu_sysctl(), hw_sysctl(),
vfs_sysctl(), etc, routines, along with sysctl_int() et al. Now all
nodes are registered with the tree, and nodes can be added (or
removed) easily, and I/O to and from the tree is handled generically.
Since the nodes are registered with the tree, the mapping from name to
number (and back again) can now be discovered, instead of having to be
hard coded. Adding new nodes to the tree is likewise much simpler --
the new infrastructure handles almost all the work for simple types,
and just about anything else can be done with a small helper function.
All existing nodes are where they were before (numerically speaking),
so all existing consumers of sysctl information should notice no
difference.
PS - I'm sorry, but there's a distinct lack of documentation at the
moment. I'm working on sysctl(3/8/9) right now, and I promise to
watch out for buses.
XXX: The decision whether or not to fast forward should be made
XXX: dynamically. Using the current approach seriously reduces
XXX: routing performance on gateways with IPsec enabled.
* Include "opt_inet.h" everywhere IP-ids are generated with ip_newid(),
so the RANDOM_IP_ID option is visible. Also in ip_id(), to ensure
the prototype for ip_randomid() is made visible.
* Add new sysctl to enable randomized IP-ids, provided the kernel was
configured with RANDOM_IP_ID. (The sysctl defaults to zero, and is
a read-only zero if RANDOM_IP_ID is not configured).
Note that the implementation of randomized IP ids is still defective,
and should not be enabled at all (even if configured) without
very careful deliberation. Caveat emptor.
Revert the (default) ip_id algorithm to the pre-randomid algorithm,
due to demonstrated low-period repeated IDs from the randomized IP_id
code. Consensus is that the low-period repetition (much less than
2^15) is not suitable for general-purpose use.
Allocators of new IPv4 IDs should now call the function ip_newid().
Randomized IP_ids is now a config-time option, "options RANDOM_IP_ID".
ip_newid() can use ip_random-id()_IP_ID if and only if configured
with RANDOM_IP_ID. A sysctl knob should be provided.
This API may be reworked in the near future to support linear ip_id
counters per (src,dst) IP-address pair.
due to demonstrated low-period repeated IDs from the randomized IP_id
code. Consensus is that the low-period repetition (much less than
2^15) is not suitable for general-purpose use.
Allocators of new IPv4 IDs should now call the function ip_newid().
Randomized IP_ids is now a config-time option, "options RANDOM_IP_ID".
ip_newid() can use ip_random-id()_IP_ID if and only if configured
with RANDOM_IP_ID. A sysctl knob should be provided.
This API may be reworked in the near future to support linear ip_id
counters per (src,dst) IP-address pair.
mbuf chains which are recycled (e.g., ICMP reflection, loopback
interface). A consensus was reached that such recycled packets should
behave (more-or-less) the same way if a new chain had been allocated
and the contents copied to that chain.
Some packet tags may in future be marked as "persistent" (e.g., for
mandatory access controls) and should persist across such deletion.
NetBSD as yet hos no persistent tags, so m_tag_delete_nonpersistent()
just deletes all tags. This should not be relied upon.
in_ifaddrhead. Recent changes in struct names caused a namespace
collision in fast-ipsec, which are most cleanly fixed by using
"in_ifaddrhead" as the listhead name.
sysctl. Add a protocol-independent sysctl handler to show the per-protocol
"struct ifq' statistics. Add IP(v4) specific call to the handler.
Other protocols can show their per-protocol input statistics by
allocating a sysclt node and calling sysctl_ifq() with their own struct ifq *.
As posted to tech-kern plus improvements/cleanup suggested by Andrew Brown.
initialized. Update the txp(4) to compensate.
- Statically initialize the TCP timer callout handles in the tcpcb
template. We still use callout_setfunc(), but that call is now much
less expensive. Add a comment that the compiler is likely to unroll
the loop (so don't sweat that it's there).
During testing the prediction counters show a hit-rate on about 85% for
packets sent on a local LAN, and better than 99% for intercontinental
high-speed bulk traffic (!).
close sockets on address changes, which was deemed to be a bad idea and was
summarily removed, so there is no point in wasting effort on maintaining it
any more.
individually, create a tcpcb template pre-initialized (and pre-zero'd)
with the static and mostly-static tcpcb parameters. The template is
now copied into the new tcpcb, which zeros and initializes most of the
tcpcb in one pass. The template is kept up-to-date as TCP sysctl
variables are changed.
Combined with the previous sb_max change, TCP socket creation is now
25% faster.
throughput significantly in a wide variety of test cases, including
local gigabit ethernet with both jumbo and standard frames,
transcontinental (U.S.) connections with e2e bandwidths ranging from
10Mbit/sec to 155Mbit/sec, and on a variety of test connections
between the NetBSD Project public servers and machines in Australia.
The impact of this change is less dramatic for high-delay connections
when Path MTU is in use but still measurable.
For optimal performance on local gigabit networks, a higher socket
buffer size (at least 64K) will still yield a substantial improvement
in performance, but 32K gets us most of the way there in my test
cases, with only a cost of _doubling_ memory use per socket rather
than _quadrupling_ it.
N.B. Windows NT, at least since Win2k SP2, uses a default socket buffer
size (or their analogue thereof) of 64K, which is a useful data
point.
the purpose of this code appears to be on crack -- it's talking about
end-to-end authentication, but the purpose of an AH tunnel is NOT end-to-end
authentication; it's authentication of the tunnel endpoints.
NB: This does not fix the fact that IPsec leaks "packet tags."
argument. So check so is non-NULL before doing the pointer-chasing
dance to find the PCB. (Unless and until we rework fast-ipsec and
KAME, to pass a struct in_pcbhdr * instead of the struct socket *).
argument to ip6_output() with a new explicit struct in6pcb* argument.
(The underlying socket can be obtained via in6pcb->inp6_socket.)
In preparation for fast-ipsec. Reviewed by itojun.
wiz