a hashed port list. In the new scheme, in_pcblookup() goes away and is
replaced by a new routine, in_pcblookup_local() for doing the local port
check. Note that this implementation is space inefficient in that the PCB
struct is now too large to fit into 128 bytes. I might deal with this in the
future by using the new zone allocator, but I wanted these changes to be
extensively tested in their current form first.
Also:
1) Fixed off-by-one errors in the port lookup loops in in_pcbbind().
2) Got rid of some unneeded rehashing. Adding a new routine, in_pcbinshash()
to do the initialial hash insertion.
3) Renamed in_pcblookuphash() to in_pcblookup_hash() for easier readability.
4) Added a new routine, in_pcbremlists() to remove the PCB from the various
hash lists.
5) Added/deleted comments where appropriate.
6) Removed unnecessary splnet() locking. In general, the PCB functions should
be called at splnet()...there are unfortunately a few exceptions, however.
7) Reorganized a few structs for better cache line behavior.
8) Killed my TCP_ACK_HACK kludge. It may come back in a different form in
the future, however.
These changes have been tested on wcarchive for more than a month. In tests
done here, connection establishment overhead is reduced by more than 50
times, thus getting rid of one of the major networking scalability problems.
Still to do: make tcp_fastimo/tcp_slowtimo scale well for systems with a
large number of connections. tcp_fastimo is easy; tcp_slowtimo is difficult.
WARNING: Anything that knows about inpcb and tcpcb structs will have to be
recompiled; at the very least, this includes netstat(1).
fix PR#3618 weren't sufficient since malloc() can block - allowing the
net interrupts in and leading to the same problem mentioned in the
PR (a panic). The order of operations has been changed so that this
is no longer a problem.
Needs to be brought into the 2.2.x branch.
PR: 3618
socket addresses in mbufs. (Socket buffers are the one exception.) A number
of kernel APIs needed to get fixed in order to make this happen. Also,
fix three protocol families which kept PCBs in mbufs to not malloc them
instead. Delete some old compatibility cruft while we're at it, and add
some new routines in the in_cksum family.
in_setsockaddr and in_setpeeraddr.
Handle the case where the socket was disconnected before the network
interrupts were disabled.
Reviewed by: Garrett Wollman <wollman@khavrinen.lcs.mit.edu>
This commit includes the following changes:
1) Old-style (pr_usrreq()) protocols are no longer supported, the compatibility
glue for them is deleted, and the kernel will panic on boot if any are compiled
in.
2) Certain protocol entry points are modified to take a process structure,
so they they can easily tell whether or not it is possible to sleep, and
also to access credentials.
3) SS_PRIV is no more, and with it goes the SO_PRIVSTATE setsockopt()
call. Protocols should use the process pointer they are now passed.
4) The PF_LOCAL and PF_ROUTE families have been updated to use the new
style, as has the `raw' skeleton family.
5) PF_LOCAL sockets now obey the process's umask when creating a socket
in the filesystem.
As a result, LINT is now broken. I'm hoping that some enterprising hacker
with a bit more time will either make the broken bits work (should be
easy for netipx) or dike them out.
cache lines. Removed the struct ip proto since only a couple of chars
were actually being used in it. Changed the order of compares in the
PCB hash lookup to take advantage of partial cache line fills (on PPro).
Discussed-with: wollman
the quality of the hash distribution. This does not fix a problem dealing
with poor distribution when using lots of IP aliases and listening
on the same port on every one of them...some other day perhaps; fixing
that requires significant code changes.
The use of xor was inspired by David S. Miller <davem@jenolan.rutgers.edu>
pr_usrreqs. Collapse duplicates with udp_usrreq.c and
tcp_usrreq.c (calling the generic routines in uipc_socket2.c and
in_pcb.c). Calling sockaddr()_ or peeraddr() on a detached
socket now traps, rather than harmlessly returning an error; this
should never happen. Allow the raw IP buffer sizes to be
controlled via sysctl.
This will make a number of things easier in the future, as well as (finally!)
avoiding the Id-smashing problem which has plagued developers for so long.
Boy, I'm glad we're not using sup anymore. This update would have been
insane otherwise.
to TAILQs. Fix places which referenced these for no good reason
that I can see (the references remain, but were fixed to compile
again; they are still questionable).
"high" and "secure"), we can't use a single variable to track the most
recently used port in all three ranges.. :-] This caused the next
transient port to be allocated from the start of the range more often than
it should.
callers of it to take advantage of this. This reduces new connection
request overhead in the face of a large number of PCBs in the system.
Thanks to David Filo <filo@yahoo.com> for suggesting this and providing
a sample implementation (which wasn't used, but showed that it could be
done).
Reviewed by: wollman
IPPORT_RESERVED that is used for selection when bind() is told to allocate
a reserved port.
Also, implement simple sanity checking for all the addresses set, to make
it a little harder for a user/sysadmin to shoot themselves in the feet.
systems (my last change did not mix well with some firewall
configurations). As much as I dislike firewalls, this is one thing I
I was not prepared to break by default.. :-)
Allow the user to nominate one of three ranges of port numbers as
candidates for selecting a local address to replace a zero port number.
The ranges are selected via a setsockopt(s, IPPROTO_IP, IP_PORTRANGE, &arg)
call. The three ranges are: default, high (to bypass firewalls) and
low (to get a port below 1024).
The default and high port ranges are sysctl settable under sysctl
net.inet.ip.portrange.*
This code also fixes a potential deadlock if the system accidently ran out
of local port addresses. It'd drop into an infinite while loop.
The secure port selection (for root) should reduce overheads and increase
reliability of rlogin/rlogind/rsh/rshd if they are modified to take
advantage of it.
Partly suggested by: pst
Reviewed by: wollman
to 20000 through 30000. These numbers are used for local IP port numbers
when an explicit address is not specified.
The values are sysctl modifiable under: net.inet.ip.port_{first|last}_auto
These numbers do not overlap with any known server addresses, without going
above 32768 which are "negative" on some other implementations.
20000 through 30000 is 2.5 times larger than the old range, but some have
suggested even that may not be enough... (gasp!) Setting a low address
of 10000 should be plenty.. :-)
the lookup fails. Updated callers to deal with this. Call in_pcblookuphash
instead of in_pcblookup() in in_pcbconnect; this improves performance of
UDP output by about 17% in the standard case.
