linking NIC Receive Side Scaling (RSS) to the network stack's
connection-group implementation. This prototype (and derived patches)
are in use at Juniper and several other FreeBSD-using companies, so
despite some reservations about its maturity, merge the patch to the
base tree so that it can be iteratively refined in collaboration rather
than maintained as a set of gradually diverging patch sets.
(1) Merge a software implementation of the Toeplitz hash specified in
RSS implemented by David Malone. This is used to allow suitable
pcbgroup placement of connections before the first packet is
received from the NIC. Software hashing is generally avoided,
however, due to high cost of the hash on general-purpose CPUs.
(2) In in_rss.c, maintain authoritative versions of RSS state intended
to be pushed to each NIC, including keying material, hash
algorithm/ configuration, and buckets. Provide software-facing
interfaces to hash 2- and 4-tuples for IPv4 and IPv6 using both
the RSS standardised Toeplitz and a 'naive' variation with a hash
efficient in software but with poor distribution properties.
Implement rss_m2cpuid()to be used by netisr and other load
balancing code to look up the CPU on which an mbuf should be
processed.
(3) In the Ethernet link layer, allow netisr distribution using RSS as
a source of policy as an alternative to source ordering; continue
to default to direct dispatch (i.e., don't try and requeue packets
for processing on the 'right' CPU if they arrive in a directly
dispatchable context).
(4) Allow RSS to control tuning of connection groups in order to align
groups with RSS buckets. If a packet arrives on a protocol using
connection groups, and contains a suitable hardware-generated
hash, use that hash value to select the connection group for pcb
lookup for both IPv4 and IPv6. If no hardware-generated Toeplitz
hash is available, we fall back on regular PCB lookup risking
contention rather than pay the cost of Toeplitz in software --
this is a less scalable but, at my last measurement, faster
approach. As core counts go up, we may want to revise this
strategy despite CPU overhead.
Where device drivers suitably configure NICs, and connection groups /
RSS are enabled, this should avoid both lock and line contention during
connection lookup for TCP. This commit does not modify any device
drivers to tune device RSS configuration to the global RSS
configuration; patches are in circulation to do this for at least
Chelsio T3 and Intel 1G/10G drivers. Currently, the KPI for device
drivers is not particularly robust, nor aware of more advanced features
such as runtime reconfiguration/rebalancing. This will hopefully prove
a useful starting point for refinement.
No MFC is scheduled as we will first want to nail down a more mature
and maintainable KPI/KBI for device drivers.
Sponsored by: Juniper Networks (original work)
Sponsored by: EMC/Isilon (patch update and merge)
the protocol specific mbuf flags are shared between them.
- Move all M_FOO definitions into a single place: netinet/in6.h, to
avoid future clashes.
- Resolve clash between M_DECRYPTED and M_SKIP_FIREWALL which resulted
in a failure of operation of IPSEC and packet filters.
Thanks to Nicolas and Georgios for all the hard work on bisecting,
testing and finally finding the root of the problem.
PR: kern/186755
PR: kern/185876
In collaboration with: Georgios Amanakis <gamanakis gmail.com>
In collaboration with: Nicolas DEFFAYET <nicolas-ml deffayet.com>
Sponsored by: Nginx, Inc.
- Use counter(9) for rt_pksent (former rt_rmx.rmx_pksent). This
removes another cache trashing ++ from packet forwarding path.
- Create zini/fini methods for the rtentry UMA zone. Via initialize
mutex and counter in them.
- Fix reporting of rmx_pksent to routing socket.
- Fix netstat(1) to report "Use" both in kvm(3) and sysctl(3) mode.
The change is mostly targeted for stable/10 merge. For head,
rt_pksent is expected to just disappear.
Discussed with: melifaro
Sponsored by: Netflix
Sponsored by: Nginx, Inc.
were primarily used to size the sysctl name list macros that were removed
in r254295. A few other constants either did not have an associated
sysctl node, or the associated node used OID_AUTO instead.
PR: ports/184525 (exp-run)
tested and is unfinished. However, I've tested my version,
it works okay. As before it is unfinished: timeout aren't
driven by TCP session state. To enable the HASH_ALL mode,
one needs in kernel config:
options FLOWTABLE_HASH_ALL
o Reduce the alignment on flentry to 64 bytes. Without
the FLOWTABLE_HASH_ALL option, twice less memory would
be consumed by flows.
o API to ip_output()/ip6_output() got even more thin: 1 liner.
o Remove unused unions. Simply use fle->f_key[].
o Merge all IPv4 code into flowtable_lookup_ipv4(), and do same
flowtable_lookup_ipv6(). Stop copying data to on stack
sockaddr structures, simply use key[] on stack.
o Move code from flowtable_lookup_common() that actually works
on insertion into flowtable_insert().
Sponsored by: Netflix
Sponsored by: Nginx, Inc.
Currently we have 3 usage patterns:
1) nd6_output (most traffic flow, no lle supplied, lle RLOCK sufficient)
2) corner cases for output (no lle, STALE lle, so on). lle WLOCK needed.
3) nd* iunternal machinery (WLOCK'ed lle provided, perform packet queing).
We separate case 1 and implement it inside its only customer - nd6_output.
This leads to some code duplication (especialy SEND stuff, which should be
hooked to output in a different way), but simplifies locking and control
flow logic fir nd6_output_lle.
Reviewed by: ae
MFC after: 3 weeks
Sponsored by: Yandex LLC
* Check ND6_IFF_IFDISABLED before acquiring any locks
* Assume m is always non-NULL
* remove 'bad' case not used anymore
* Simply if_output conditional
MFC after: 2 weeks
Sponsored by: Yandex LLC
- ip_output() and ip_output6() simply call flowtable_lookup(),
passing mbuf and address family. That's the only code under
#ifdef FLOWTABLE in the protocols code now.
o Revamp statistics gathering and export.
- Remove hand made pcpu stats, and utilize counter(9).
- Snapshot of statistics is available via 'netstat -rs'.
- All sysctls are moved into net.flowtable namespace, since
spreading them over net.inet isn't correct.
o Properly separate at compile time INET and INET6 parts.
o General cleanup.
- Remove chain of multiple flowtables. We simply have one for
IPv4 and one for IPv6.
- Flowtables are allocated in flowtable.c, symbols are static.
- With proper argument to SYSINIT() we no longer need flowtable_ready.
- Hash salt doesn't need to be per-VNET.
- Removed rudimentary debugging, which use quite useless in dtrace era.
The runtime behavior of flowtable shouldn't be changed by this commit.
Sponsored by: Netflix
Sponsored by: Nginx, Inc.
* Set ia address/mask values BEFORE attaching to address lists.
Inet6 address assignment is not atomic, so the simplest way to
do this atomically is to fill in ia before attach.
* Validate irfa->ia_addr field before use (we permit ANY sockaddr in old code).
* Do some renamings:
in6_ifinit -> in6_notify_ifa (interaction with other subsystems is here)
in6_setup_ifa -> in6_broadcast_ifa (LLE/Multicast/DaD code)
in6_ifaddloop -> nd6_add_ifa_lle
in6_ifremloop -> nd6_rem_ifa_lle
* Split working with LLE and route announce code for last two.
Add temporary in6_newaddrmsg() function to mimic current rtsock behaviour.
* Call device SIOCSIFADDR handler IFF we're adding first address.
In IPv4 we have to call it on every address change since ARP record
is installed by arp_ifinit() which is called by given handler.
IPv6 stack, on the opposite is responsible to call nd6_add_ifa_lle() so
there is no reason to call SIOCSIFADDR often.
included.. netstat uses -DKERNEL=1 to get these parts and breaks the
build w/o it...
melifaro@ says that ae@ is probably asleep, and the PR doesn't have
this part of the patch... Probably a local change got in by accident..
PR: 185148
Pointy hat to: ae@
Thus we can use IF_AFDATA_RLOCK() instead of IF_AFDATA_LOCK() when doing
lla_lookup() without LLE_CREATE flag.
Reviewed by: glebius, adrian
MFC after: 1 week
Sponsored by: Yandex LLC
to lla_lookup().
This drastically reduces the very high lock contention when doing parallel
TCP throughput tests (> 1024 sockets) with IPv6.
Tested:
* parallel IPv6 TCP bulk data exchange, 8192 sockets
MFC after: 1 week
Sponsored by: Netflix, Inc.
option, unbreak the lock tracing release semantic by embedding
calls to LOCKSTAT_PROFILE_RELEASE_LOCK() direclty in the inlined
version of the releasing functions for mutex, rwlock and sxlock.
Failing to do so skips the lockstat_probe_func invokation for
unlocking.
- As part of the LOCKSTAT support is inlined in mutex operation, for
kernel compiled without lock debugging options, potentially every
consumer must be compiled including opt_kdtrace.h.
Fix this by moving KDTRACE_HOOKS into opt_global.h and remove the
dependency by opt_kdtrace.h for all files, as now only KDTRACE_FRAMES
is linked there and it is only used as a compile-time stub [0].
[0] immediately shows some new bug as DTRACE-derived support for debug
in sfxge is broken and it was never really tested. As it was not
including correctly opt_kdtrace.h before it was never enabled so it
was kept broken for a while. Fix this by using a protection stub,
leaving sfxge driver authors the responsibility for fixing it
appropriately [1].
Sponsored by: EMC / Isilon storage division
Discussed with: rstone
[0] Reported by: rstone
[1] Discussed with: philip
unlocked route. Use in6_rtalloc() instead of in6_rtalloc1. This helps
simplify the code and remove several now unused variables.
PR: 156283
MFC after: 2 weeks
to this event, adding if_var.h to files that do need it. Also, include
all includes that now are included due to implicit pollution via if_var.h
Sponsored by: Netflix
Sponsored by: Nginx, Inc.
four counters to struct ifaddr. This kills '+=' on a variables shared
between processors for every packet.
- Nuke struct if_data from struct ifaddr.
- In ip_input() do not put a reference on ifaddr, instead update statistics
right now in place and do IN_IFADDR_RUNLOCK(). These removes atomic(9)
for every packet. [1]
- To properly support NET_RT_IFLISTL sysctl used by getifaddrs(3), in
rtsock.c fill if_data fields using counter_u64_fetch().
- Accidentially fix bug in COMPAT_32 version of NET_RT_IFLISTL, which
took if_data not from the ifaddr, but from ifaddr's ifnet. [2]
Submitted by: melifaro [1], pluknet[2]
Sponsored by: Netflix
Sponsored by: Nginx, Inc.
