limited in the amount of data they can handle at once.
Drivers can set ifp->if_hw_tsomax before calling ether_ifattach() to
change the limit.
The lowest allowable size is IP_MAXPACKET / 8 (8192 bytes) as anything
less wouldn't be very useful anymore. The upper limit is still at
IP_MAXPACKET (65536 bytes). Raising it requires further auditing of
the IPv4/v6 code path's as the length field in the IP header would
overflow leading to confusion in firewalls and others packet handler on
the real size of the packet.
The placement into "struct ifnet" is a bit hackish but the best place
that was found. When the stack/driver boundary is updated it should
be handled in a better way.
Submitted by: cperciva (earlier version)
Reviewed by: cperciva
Tested by: cperciva
MFC after: 1 week (using spare struct members to preserve ABI)
reside on their own cache line to prevent false sharing with other
nearby structures, especially for those in the .bss segment.
NB: Those mutexes and rwlocks with variables next to them that get
changed on every invocation do not benefit from their own cache line.
Actually it may be net negative because two cache misses would be
incurred in those cases.
While here, also make the code that enforces power-of-two more
forgiving, instead of just resetting to 512, graciously round-down
to the next lower power of two.
in network byte order. Any host byte order processing is
done in local variables and host byte order values are
never[1] written to a packet.
After this change a packet processed by the stack isn't
modified at all[2] except for TTL.
After this change a network stack hacker doesn't need to
scratch his head trying to figure out what is the byte order
at the given place in the stack.
[1] One exception still remains. The raw sockets convert host
byte order before pass a packet to an application. Probably
this would remain for ages for compatibility.
[2] The ip_input() still subtructs header len from ip->ip_len,
but this is planned to be fixed soon.
Reviewed by: luigi, Maxim Dounin <mdounin mdounin.ru>
Tested by: ray, Olivier Cochard-Labbe <olivier cochard.me>
- Stateful TCP offload drivers for Terminator 3 and 4 (T3 and T4) ASICs.
These are available as t3_tom and t4_tom modules that augment cxgb(4)
and cxgbe(4) respectively. The cxgb/cxgbe drivers continue to work as
usual with or without these extra features.
- iWARP driver for Terminator 3 ASIC (kernel verbs). T4 iWARP in the
works and will follow soon.
Build-tested with make universe.
30s overview
============
What interfaces support TCP offload? Look for TOE4 and/or TOE6 in the
capabilities of an interface:
# ifconfig -m | grep TOE
Enable/disable TCP offload on an interface (just like any other ifnet
capability):
# ifconfig cxgbe0 toe
# ifconfig cxgbe0 -toe
Which connections are offloaded? Look for toe4 and/or toe6 in the
output of netstat and sockstat:
# netstat -np tcp | grep toe
# sockstat -46c | grep toe
Reviewed by: bz, gnn
Sponsored by: Chelsio communications.
MFC after: ~3 months (after 9.1, and after ensuring MFC is feasible)
headers for TSO but also for generic checksum offloading. Ideally we
would only have one common function shared amongst all drivers, and
perhaps when updating them for IPv6 we should introduce that.
Eventually we should provide the meta information along with mbufs to
avoid (re-)parsing entirely.
To not break IPv6 (checksums and offload) and to be able to MFC the
changes without risking to hurt 3rd party drivers, duplicate the v4
framework, as other OSes have done as well.
Introduce interface capability flags for TX/RX checksum offload with
IPv6, to allow independent toggling (where possible). Add CSUM_*_IPV6
flags for UDP/TCP over IPv6, and reserve further for SCTP, and IPv6
fragmentation. Define CSUM_DELAY_DATA_IPV6 as we do for legacy IP and
add an alias for CSUM_DATA_VALID_IPV6.
This pretty much brings IPv6 handling in line with IPv4.
TSO is still handled in a different way and not via if_hwassist.
Update ifconfig to allow (un)setting of the new capability flags.
Update loopback to announce the new capabilities and if_hwassist flags.
Individual driver updates will have to follow, as will SCTP.
Reported by: gallatin, dim, ..
Reviewed by: gallatin (glanced at?)
MFC after: 3 days
X-MFC with: r235961,235959,235958
Add code to handle pre-checked TCP checksums as indicated by mbuf
flags to save the entire computation for validation if not needed.
In the IPv6 TCP output path only compute the pseudo-header checksum,
set the checksum offset in the mbuf field along the appropriate flag
as done in IPv4.
In tcp_respond() just initialize the IPv6 payload length to 0 as
ip6_output() will properly set it.
Sponsored by: The FreeBSD Foundation
Sponsored by: iXsystems
Reviewed by: gnn (as part of the whole)
MFC After: 3 days
proposed MTU value from it and update the TCP host cache. Then
tcp_mss_update() is called on the corresponding tcpcb. It finds the
just allocated entry in the TCP host cache and updates MSS on the
tcpcb. And then we do a fast retransmit of what we have in the tcp
send buffer.
This sequence gets broken if the TCP host cache is exausted. In this
case allocation fails, and later called tcp_mss_update() finds nothing
in cache. The fast retransmit is done with not reduced MSS and is
immidiately replied by remote host with new ICMP datagrams and the
cycle repeats. This ping-pong can go up to wirespeed.
To fix this:
- tcp_mss_update() gets new parameter - mtuoffer, that is like
offer, but needs to have min_protoh subtracted.
- tcp_mtudisc() as notification method renamed to tcp_mtudisc_notify().
- tcp_mtudisc() now accepts not a useless error argument, but proposed
MTU value, that is passed to tcp_mss_update() as mtuoffer.
Reported by: az
Reported by: Andrey Zonov <andrey zonov.org>
Reviewed by: andre (previous version of patch)
the original IPv4 implementation from r178888:
- Use RT_DEFAULT_FIB in the IPv4 implementation where noticed.
- Use rt*fib() KPI with explicit RT_DEFAULT_FIB where applicable in
the NFS code.
- Use the new in6_rt* KPI in TCP, gif(4), and the IPv6 network stack
where applicable.
- Split in6_rtqtimo() and in6_mtutimo() as done in IPv4 and equally
prevent multiple initializations of callouts in in6_inithead().
- Use wrapper functions where needed to preserve the current KPI to
ease MFCs. Use BURN_BRIDGES to indicate expected future cleanup.
- Fix (related) comments (both technical or style).
- Convert to rtinit() where applicable and only use custom loops where
currently not possible otherwise.
- Multicast group, most neighbor discovery address actions and faith(4)
are locked to the default FIB. Individual IPv6 addresses will only
appear in the default FIB, however redirect information and prefixes
of connected subnets are automatically propagated to all FIBs by
default (mimicking IPv4 behavior as closely as possible).
Sponsored by: Cisco Systems, Inc.
This makes pf find the wrong state and cause errors reported with state mismatches.
Clear the cached state link on the pf(4) tag to avoid the state mismatches.
Approved by: bz
struct inpcbgroup. pcbgroups, or "connection groups", supplement the
existing inpcbinfo connection hash table, which when pcbgroups are
enabled, might now be thought of more usefully as a per-protocol
4-tuple reservation table.
Connections are assigned to connection groups base on a hash of their
4-tuple; wildcard sockets require special handling, and are members
of all connection groups. During a connection lookup, a
per-connection group lock is employed rather than the global pcbinfo
lock. By aligning connection groups with input path processing,
connection groups take on an effective CPU affinity, especially when
aligned with RSS work placement (see a forthcoming commit for
details). This eliminates cache line migration associated with
global, protocol-layer data structures in steady state TCP and UDP
processing (with the exception of protocol-layer statistics; further
commit to follow).
Elements of this approach were inspired by Willman, Rixner, and Cox's
2006 USENIX paper, "An Evaluation of Network Stack Parallelization
Strategies in Modern Operating Systems". However, there are also
significant differences: we maintain the inpcb lock, rather than using
the connection group lock for per-connection state.
Likewise, the focus of this implementation is alignment with NIC
packet distribution strategies such as RSS, rather than pure software
strategies. Despite that focus, software distribution is supported
through the parallel netisr implementation, and works well in
configurations where the number of hardware threads is greater than
the number of NIC input queues, such as in the RMI XLR threaded MIPS
architecture.
Another important difference is the continued maintenance of existing
hash tables as "reservation tables" -- these are useful both to
distinguish the resource allocation aspect of protocol name management
and the more common-case lookup aspect. In configurations where
connection tables are aligned with hardware hashes, it is desirable to
use the traditional lookup tables for loopback or encapsulated traffic
rather than take the expense of hardware hashes that are hard to
implement efficiently in software (such as RSS Toeplitz).
Connection group support is enabled by compiling "options PCBGROUP"
into your kernel configuration; for the time being, this is an
experimental feature, and hence is not enabled by default.
Subject to the limited MFCability of change dependencies in inpcb,
and its change to the inpcbinfo init function signature, this change
in principle could be merged to FreeBSD 8.x.
Reviewed by: bz
Sponsored by: Juniper Networks, Inc.
- The existing ipi_lock continues to protect the global inpcb list and
inpcb counter. This lock is now relegated to a small number of
allocation and free operations, and occasional operations that walk
all connections (including, awkwardly, certain UDP multicast receive
operations -- something to revisit).
- A new ipi_hash_lock protects the two inpcbinfo hash tables for
looking up connections and bound sockets, manipulated using new
INP_HASH_*() macros. This lock, combined with inpcb locks, protects
the 4-tuple address space.
Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb
connection locks, so may be acquired while manipulating a connection on
which a lock is already held, avoiding the need to acquire the inpcbinfo
lock preemptively when a binding change might later be required. As a
result, however, lookup operations necessarily go through a reference
acquire while holding the lookup lock, later acquiring an inpcb lock --
if required.
A new function in_pcblookup() looks up connections, and accepts flags
indicating how to return the inpcb. Due to lock order changes, callers
no longer need acquire locks before performing a lookup: the lookup
routine will acquire the ipi_hash_lock as needed. In the future, it will
also be able to use alternative lookup and locking strategies
transparently to callers, such as pcbgroup lookup. New lookup flags are,
supplementing the existing INPLOOKUP_WILDCARD flag:
INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb
INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb
Callers must pass exactly one of these flags (for the time being).
Some notes:
- All protocols are updated to work within the new regime; especially,
TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely
eliminated, and global hash lock hold times are dramatically reduced
compared to previous locking.
- The TCP syncache still relies on the pcbinfo lock, something that we
may want to revisit.
- Support for reverting to the FreeBSD 7.x locking strategy in TCP input
is no longer available -- hash lookup locks are now held only very
briefly during inpcb lookup, rather than for potentially extended
periods. However, the pcbinfo ipi_lock will still be acquired if a
connection state might change such that a connection is added or
removed.
- Raw IP sockets continue to use the pcbinfo ipi_lock for protection,
due to maintaining their own hash tables.
- The interface in6_pcblookup_hash_locked() is maintained, which allows
callers to acquire hash locks and perform one or more lookups atomically
with 4-tuple allocation: this is required only for TCPv6, as there is no
in6_pcbconnect_setup(), which there should be.
- UDPv6 locking remains significantly more conservative than UDPv4
locking, which relates to source address selection. This needs
attention, as it likely significantly reduces parallelism in this code
for multithreaded socket use (such as in BIND).
- In the UDPv4 and UDPv6 multicast cases, we need to revisit locking
somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which
is no longer sufficient. A second check once the inpcb lock is held
should do the trick, keeping the general case from requiring the inpcb
lock for every inpcb visited.
- This work reminds us that we need to revisit locking of the v4/v6 flags,
which may be accessed lock-free both before and after this change.
- Right now, a single lock name is used for the pcbhash lock -- this is
undesirable, and probably another argument is required to take care of
this (or a char array name field in the pcbinfo?).
This is not an MFC candidate for 8.x due to its impact on lookup and
locking semantics. It's possible some of these issues could be worked
around with compatibility wrappers, if necessary.
Reviewed by: bz
Sponsored by: Juniper Networks, Inc.
keep constant ISN growth rate, do the same directly inside tcp_new_isn(),
taking into account how much time (ticks) passed since the last call.
On my test systems this decreases idle interrupt rate from 140Hz to 70Hz.
Add some comments at #endifs given more nestedness. To make the compiler
happy, some default initializations were added in accordance with the style
on the files.
Reviewed by: gnn
Sponsored by: The FreeBSD Foundation
Sponsored by: iXsystems
MFC after: 4 days
As long as this is a costy function, even when compiled in (along with
the option TCP_SIGNATURE), it can be disabled via the
net.inet.tcp.signature_verify_input sysctl.
Sponsored by: Sandvine Incorporated
Reviewed by: emaste, bz
MFC after: 2 weeks
access inbound/outbound events and associated data for established TCP
connections. The hooks only run if at least one hook function is registered
for the hook point, ensuring the impact on the stack is effectively nil when
no TCP Khelp modules are loaded. struct tcp_hhook_data is passed as contextual
data to any registered Khelp module hook functions.
- Add an OSD (Object Specific Data) pointer to struct tcpcb to allow Khelp
modules to associate per-connection data with the TCP control block.
- Bump __FreeBSD_version and add a note to UPDATING regarding to ABI changes
introduced by this commit and r216753.
In collaboration with: David Hayes <dahayes at swin edu au> and
Grenville Armitage <garmitage at swin edu au>
Sponsored by: FreeBSD Foundation
Reviewed by: bz, others along the way
MFC after: 3 months
DPCPU_DEFINE and VNET_DEFINE macros, as these cause problems for various
people working on the affected files. A better long-term solution is
still being considered. This reversal may give some modules empty
set_pcpu or set_vnet sections, but these are harmless.
Changes reverted:
------------------------------------------------------------------------
r215318 | dim | 2010-11-14 21:40:55 +0100 (Sun, 14 Nov 2010) | 4 lines
Instead of unconditionally emitting .globl's for the __start_set_xxx and
__stop_set_xxx symbols, only emit them when the set_vnet or set_pcpu
sections are actually defined.
------------------------------------------------------------------------
r215317 | dim | 2010-11-14 21:38:11 +0100 (Sun, 14 Nov 2010) | 3 lines
Apply the STATIC_VNET_DEFINE and STATIC_DPCPU_DEFINE macros throughout
the tree.
------------------------------------------------------------------------
r215316 | dim | 2010-11-14 21:23:02 +0100 (Sun, 14 Nov 2010) | 2 lines
Add macros to define static instances of VNET_DEFINE and DPCPU_DEFINE.
runs on boot and each time a vnet jail is created. Running cc_init() multiple
times results in a panic when attempting to initialise the cc_list lock again,
and so r215166 effectively broke the use of vnet jails.
Switch to using a SYSINIT to run cc_init() on boot. CC algorithm modules loaded
on boot register in the same SI_SUB_PROTO_IFATTACHDOMAIN category as is used in
this patch, so cc_init() is run at SI_ORDER_FIRST to ensure the framework is
initialised before module registration is attempted.
Sponsored by: FreeBSD Foundation
Reported and tested by: Mikolaj Golub <to.my.trociny at gmail com>
MFC after: 11 weeks
X-MFC with: r215166
Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details
about the project are available at: http://caia.swin.edu.au/freebsd/5cc/
- Add a KPI and supporting infrastructure to allow modular congestion control
algorithms to be used in the net stack. Algorithms can maintain per-connection
state if required, and connections maintain their own algorithm pointer, which
allows different connections to concurrently use different algorithms. The
TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to
programmatically query or change the congestion control algorithm respectively
from within an application at runtime.
- Integrate the framework with the TCP stack in as least intrusive a manner as
possible. Care was also taken to develop the framework in a way that should
allow integration with other congestion aware transport protocols (e.g. SCTP)
in the future. The hope is that we will one day be able to share a single set
of congestion control algorithm modules between all congestion aware transport
protocols.
- Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack
and use it to decouple the meaning of recovery from a congestion event and
recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based
congestion control protocols don't generally need to recover from packet loss
and need a different way to note a congestion recovery episode within the
stack.
- Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code
and ensures the stack always uses the appropriate mechanisms for recovering
from packet loss during a congestion recovery episode.
- Extract the NewReno congestion control algorithm from the TCP stack and
massage it into module form. NewReno is always built into the kernel and will
remain the default algorithm for the forseeable future. Implementations of
additional different algorithms will become available in the near future.
- Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code
that relies on the size of "struct tcpcb" is required.
Many thanks go to the Cisco University Research Program Fund at Community
Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work
at the Centre for Advanced Internet Architectures, Swinburne University of
Technology is greatly appreciated.
In collaboration with: David Hayes <dahayes at swin edu au> and
Grenville Armitage <garmitage at swin edu au>
Sponsored by: Cisco URP, FreeBSD Foundation
Reviewed by: rpaulo
Tested by: David Hayes (and many others over the years)
MFC after: 3 months
not be used outside of the reassembly queue implementation. Provide a new
function to flush all segments from a reassembly queue and call it from the
appropriate places instead of manipulating the queue directly.
Sponsored by: FreeBSD Foundation
Reviewed by: andre, gnn, rpaulo
MFC after: 2 weeks
to give way for the pluggable congestion control framework. It is
the task of the congestion control algorithm to set the congestion
window and amount of inflight data without external interference.
In 'struct tcpcb' the variables previously used by the inflight
limiter are renamed to spares to keep the ABI intact and to have
some more space for future extensions.
In 'struct tcp_info' the variable 'tcpi_snd_bwnd' is not removed to
preserve the ABI. It is always set to 0.
In siftr.c in 'struct pkt_node' the variable 'snd_bwnd' is not removed
to preserve the ABI. It is always set to 0.
These unused variable in the various structures may be reused in the
future or garbage collected before the next release or at some other
point when an ABI change happens anyway for other reasons.
No MFC is planned. The inflight bandwidth limiter stays disabled by
default in the other branches but remains available.
sysctl's and remove any side effects.
Both sysctl's share the same backend infrastructure and due to the
way it was implemented enabling net.inet.tcp.log_in_vain would also
cause log_debug output to be generated. This was surprising and
eventually annoying to the user.
The log output backend is kept the same but a little shim is inserted
to properly separate log_in_vain and log_debug and to remove any side
effects.
PR: kern/137317
MFC after: 1 week
number of syncache entries into account for the surplus we add to account
for a possible increase of records in the re-entry window.
Discussed with: jhb, silby
MFC after: 1 week
memory size estimate to userland for pcb list sysctls. The previous
behavior of a "slop" of n/8 does not work well for small values of n
(e.g. no slop at all if you have less than 8 open UDP connections).
Reviewed by: bz
MFC after: 1 week
path MTU discovery and the tcp_minmss limiter for very small MTU's.
When the MTU suggested by the gateway via ICMP, or if there isn't
any the next smaller step from ip_next_mtu(), is lower than the
floor enforced by net.inet.tcp.minmss (default 216) the value is
ignored and the default MSS (512) is used instead. However the
DF flag in the IP header is still set in tcp_output() preventing
fragmentation by the gateway.
Fix this by using tcp_minmss as the MSS and clear the DF flag if
the suggested MTU is too low. This turns off path MTU dissovery
for the remainder of the session and allows fragmentation to be
done by the gateway.
Only MTU's smaller than 256 are affected. The smallest official
MTU specified is for AX.25 packet radio at 256 octets.
PR: kern/146628
Tested by: Matthew Luckie <mjl-at-luckie org nz>
MFC after: 1 week
It was experimental and interferes with the normal congestion control
algorithms by instating a separate, possibly lower, ceiling for the
amount of data that is in flight to the remote host. With high speed
internet connections the inflight limit frequently has been estimated
too low due to the noisy nature of the RTT measurements.
This code gives way for the upcoming pluggable congestion control
framework. It is the task of the congestion control algorithm to
set the congestion window and amount of inflight data without external
interference.
Reviewed by: lstewart
MFC after: 1 week
Removal after: 1 month
"Whitspace" churn after the VIMAGE/VNET whirls.
Remove the need for some "init" functions within the network
stack, like pim6_init(), icmp_init() or significantly shorten
others like ip6_init() and nd6_init(), using static initialization
again where possible and formerly missed.
Move (most) variables back to the place they used to be before the
container structs and VIMAGE_GLOABLS (before r185088) and try to
reduce the diff to stable/7 and earlier as good as possible,
to help out-of-tree consumers to update from 6.x or 7.x to 8 or 9.
This also removes some header file pollution for putatively
static global variables.
Revert VIMAGE specific changes in ipfilter::ip_auth.c, that are
no longer needed.
Reviewed by: jhb
Discussed with: rwatson
Sponsored by: The FreeBSD Foundation
Sponsored by: CK Software GmbH
MFC after: 6 days
their calling contexts in {IP divert, raw IP sockets, TCP, UDP} and
create new helper functions: in_pcbinfo_init() and in_pcbinfo_destroy()
to do this work in a central spot. As inpcbinfo becomes more complex
due to ongoing work to add connection groups, this will reduce code
duplication.
MFC after: 1 month
Reviewed by: bz
Sponsored by: Juniper Networks
to not leak them, otherwise making UMA/vmstat unhappy with every stoped vnet.
We will still leak pages (especially for zones marked NOFREE).
Reshuffle cleanup order in tcp_destroy() to get rid of what we can
easily free first.
Sponsored by: ISPsystem
Reviewed by: rwatson
MFC after: 5 days
feature when you have a seemingly stuck socket and want to figure
out why it has not been closed yet.
No plans to MFC this, as it changes the netstat sysctl ABI.
Reviewed by: andre, rwatson, Eric Van Gyzen
