When epoch(9) was introduced to network stack, it was basically
dropped in place of existing locking, which was mutexes and
rwlocks. For the sake of performance mutex covered areas were
as small as possible, so became epoch covered areas.
However, epoch doesn't introduce any contention, it just delays
memory reclaim. So, there is no point to minimise epoch covered
areas in sense of performance. Meanwhile entering/exiting epoch
also has non-zero CPU usage, so doing this less often is a win.
Not the least is also code maintainability. In the new paradigm
we can assume that at any stage of processing a packet, we are
inside network epoch. This makes coding both input and output
path way easier.
On output path we already enter epoch quite early - in the
ip_output(), in the ip6_output().
This patch does the same for the input path. All ISR processing,
network related callouts, other ways of packet injection to the
network stack shall be performed in net_epoch. Any leaf function
that walks network configuration now asserts epoch.
Tricky part is configuration code paths - ioctls, sysctls. They
also call into leaf functions, so some need to be changed.
This patch would introduce more epoch recursions (see EPOCH_TRACE)
than we had before. They will be cleaned up separately, as several
of them aren't trivial. Note, that unlike a lock recursion the
epoch recursion is safe and just wastes a bit of resources.
Reviewed by: gallatin, hselasky, cy, adrian, kristof
Differential Revision: https://reviews.freebsd.org/D19111
- Perform ifp mismatch checks (to determine if a send tag is allocated
for a different ifp than the one the packet is being output on), in
ip_output() and ip6_output(). This avoids sending packets with send
tags to ifnet drivers that don't support send tags.
Since we are now checking for ifp mismatches before invoking
if_output, we can now try to allocate a new tag before invoking
if_output sending the original packet on the new tag if allocation
succeeds.
To avoid code duplication for the fragment and unfragmented cases,
add ip_output_send() and ip6_output_send() as wrappers around
if_output and nd6_output_ifp, respectively. All of the logic for
setting send tags and dealing with send tag-related errors is done
in these wrapper functions.
For pseudo interfaces that wrap other network interfaces (vlan and
lagg), wrapper send tags are now allocated so that ip*_output see
the wrapper ifp as the ifp in the send tag. The if_transmit
routines rewrite the send tags after performing an ifp mismatch
check. If an ifp mismatch is detected, the transmit routines fail
with EAGAIN.
- To provide clearer life cycle management of send tags, especially
in the presence of vlan and lagg wrapper tags, add a reference count
to send tags managed via m_snd_tag_ref() and m_snd_tag_rele().
Provide a helper function (m_snd_tag_init()) for use by drivers
supporting send tags. m_snd_tag_init() takes care of the if_ref
on the ifp meaning that code alloating send tags via if_snd_tag_alloc
no longer has to manage that manually. Similarly, m_snd_tag_rele
drops the refcount on the ifp after invoking if_snd_tag_free when
the last reference to a send tag is dropped.
This also closes use after free races if there are pending packets in
driver tx rings after the socket is closed (e.g. from tcpdrop).
In order for m_free to work reliably, add a new CSUM_SND_TAG flag in
csum_flags to indicate 'snd_tag' is set (rather than 'rcvif').
Drivers now also check this flag instead of checking snd_tag against
NULL. This avoids false positive matches when a forwarded packet
has a non-NULL rcvif that was treated as a send tag.
- cxgbe was relying on snd_tag_free being called when the inp was
detached so that it could kick the firmware to flush any pending
work on the flow. This is because the driver doesn't require ACK
messages from the firmware for every request, but instead does a
kind of manual interrupt coalescing by only setting a flag to
request a completion on a subset of requests. If all of the
in-flight requests don't have the flag when the tag is detached from
the inp, the flow might never return the credits. The current
snd_tag_free command issues a flush command to force the credits to
return. However, the credit return is what also frees the mbufs,
and since those mbufs now hold references on the tag, this meant
that snd_tag_free would never be called.
To fix, explicitly drop the mbuf's reference on the snd tag when the
mbuf is queued in the firmware work queue. This means that once the
inp's reference on the tag goes away and all in-flight mbufs have
been queued to the firmware, tag's refcount will drop to zero and
snd_tag_free will kick in and send the flush request. Note that we
need to avoid doing this in the middle of ethofld_tx(), so the
driver grabs a temporary reference on the tag around that loop to
defer the free to the end of the function in case it sends the last
mbuf to the queue after the inp has dropped its reference on the
tag.
- mlx5 preallocates send tags and was using the ifp pointer even when
the send tag wasn't in use. Explicitly use the ifp from other data
structures instead.
- Sprinkle some assertions in various places to assert that received
packets don't have a send tag, and that other places that overwrite
rcvif (e.g. 802.11 transmit) don't clobber a send tag pointer.
Reviewed by: gallatin, hselasky, rgrimes, ae
Sponsored by: Netflix
Differential Revision: https://reviews.freebsd.org/D20117
Mainly focus on files that use BSD 2-Clause license, however the tool I
was using misidentified many licenses so this was mostly a manual - error
prone - task.
The Software Package Data Exchange (SPDX) group provides a specification
to make it easier for automated tools to detect and summarize well known
opensource licenses. We are gradually adopting the specification, noting
that the tags are considered only advisory and do not, in any way,
superceed or replace the license texts.
No functional change intended.
Implement get_pcpu() for amd64/sparc64/mips/powerpc, and use it to
replace pcpu_find(curcpu) in MI code.
Reviewed by: andreast, kan, lidl
Tested by: lidl(mips, sparc64), andreast(powerpc)
Differential Revision: https://reviews.freebsd.org/D9587
The switch to get_pcpu() in MI code seems to cause hangs on MIPS.
Back out until we can get a better idea of what's happening there.
Reported by: kan, lidl
Use netisr_get_cpuid() in netisr_select_cpuid() to limit cpuid value
returned by protocol to be sure that it is not greather than nws_count.
PR: 211836
Reviewed by: adrian
MFC after: 3 days
Add accessor functions to toggle the state per VNET.
The base system (vnet0) will always enable itself with the normal
registration. We will share the registered protocol handlers in all
VNETs minimising duplication and management.
Upon disabling netisr processing for a VNET drain the netisr queue from
packets for that VNET.
Update netisr consumers to (de)register on a per-VNET start/teardown using
VNET_SYS(UN)INIT functionality.
The change should be transparent for non-VIMAGE kernels.
Reviewed by: gnn (, hiren)
Obtained from: projects/vnet
MFC after: 2 weeks
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D6691
Currently, Application Processors (non-boot CPUs) are started by
MD code at SI_SUB_CPU, but they are kept waiting in a "pen" until
SI_SUB_SMP at which point they are released to run kernel threads.
SI_SUB_SMP is one of the last SYSINIT levels, so APs don't enter
the scheduler and start running threads until fairly late in the
boot.
This change moves SI_SUB_SMP up to just before software interrupt
threads are created allowing the APs to start executing kernel
threads much sooner (before any devices are probed). This allows
several initialization routines that need to perform initialization
on all CPUs to now perform that initialization in one step rather
than having to defer the AP initialization to a second SYSINIT run
at SI_SUB_SMP. It also permits all CPUs to be available for
handling interrupts before any devices are probed.
This last feature fixes a problem on with interrupt vector exhaustion.
Specifically, in the old model all device interrupts were routed
onto the boot CPU during boot. Later after the APs were released at
SI_SUB_SMP, interrupts were redistributed across all CPUs.
However, several drivers for multiqueue hardware allocate N interrupts
per CPU in the system. In a system with many CPUs, just a few drivers
doing this could exhaust the available pool of interrupt vectors on
the boot CPU as each driver was allocating N * mp_ncpu vectors on the
boot CPU. Now, drivers will allocate interrupts on their desired CPUs
during boot meaning that only N interrupts are allocated from the boot
CPU instead of N * mp_ncpu.
Some other bits of code can also be simplified as smp_started is
now true much earlier and will now always be true for these bits of
code. This removes the need to treat the single-CPU boot environment
as a special case.
As a transition aid, the new behavior is available under a new kernel
option (EARLY_AP_STARTUP). This will allow the option to be turned off
if need be during initial testing. I plan to enable this on x86 by
default in a followup commit in the next few days and to have all
platforms moved over before 11.0. Once the transition is complete,
the option will be removed along with the !EARLY_AP_STARTUP code.
These changes have only been tested on x86. Other platform maintainers
are encouraged to port their architectures over as well. The main
things to check for are any uses of smp_started in MD code that can be
simplified and SI_SUB_SMP SYSINITs in MD code that can be removed in
the EARLY_AP_STARTUP case (e.g. the interrupt shuffling).
PR: kern/199321
Reviewed by: markj, gnn, kib
Sponsored by: Netflix
to specify exact number of cpus in loader.conf which get annoying when you have
mix of machines which don't have equal number of total cpus. I propose "-1" as
that value. When loader.conf has net.isr.maxthreads = -1, netisr will use all
available cpus.
In collaboration with: davide
Reviewed by: gnn
Differential Revision: https://reviews.freebsd.org/D2318
MFC after: 2 weeks
Sponsored by: Limelight Networks
device polling rather than having it only be controlled by the compile
time option.
Summary: Rubicon Communications (Netgate)
Reviewers: #network, hiren
Reviewed By: #network, hiren
Subscribers: hiren
Differential Revision: https://reviews.freebsd.org/D2258
from the FreeBSD network code. The flag is still kept around in the
"sys/mbuf.h" header file, but does no longer have any users. Instead
the "m_pkthdr.rsstype" field in the mbuf structure is now used to
decide the meaning of the "m_pkthdr.flowid" field. To modify the
"m_pkthdr.rsstype" field please use the existing "M_HASHTYPE_XXX"
macros as defined in the "sys/mbuf.h" header file.
This patch introduces new behaviour in the transmit direction.
Previously network drivers checked if "M_FLOWID" was set in "m_flags"
before using the "m_pkthdr.flowid" field. This check has now now been
replaced by checking if "M_HASHTYPE_GET(m)" is different from
"M_HASHTYPE_NONE". In the future more hashtypes will be added, for
example hashtypes for hardware dedicated flows.
"M_HASHTYPE_OPAQUE" indicates that the "m_pkthdr.flowid" value is
valid and has no particular type. This change removes the need for an
"if" statement in TCP transmit code checking for the presence of a
valid flowid value. The "if" statement mentioned above is now a direct
variable assignment which is then later checked by the respective
network drivers like before.
Additional notes:
- The SCTP code changes will be committed as a separate patch.
- Removal of the "M_FLOWID" flag will also be done separately.
- The FreeBSD version has been bumped.
MFC after: 1 month
Sponsored by: Mellanox Technologies
These changes prevent sysctl(8) from returning proper output,
such as:
1) no output from sysctl(8)
2) erroneously returning ENOMEM with tools like truss(1)
or uname(1)
truss: can not get etype: Cannot allocate memory
there is an environment variable which shall initialize the SYSCTL
during early boot. This works for all SYSCTL types both statically and
dynamically created ones, except for the SYSCTL NODE type and SYSCTLs
which belong to VNETs. A new flag, CTLFLAG_NOFETCH, has been added to
be used in the case a tunable sysctl has a custom initialisation
function allowing the sysctl to still be marked as a tunable. The
kernel SYSCTL API is mostly the same, with a few exceptions for some
special operations like iterating childrens of a static/extern SYSCTL
node. This operation should probably be made into a factored out
common macro, hence some device drivers use this. The reason for
changing the SYSCTL API was the need for a SYSCTL parent OID pointer
and not only the SYSCTL parent OID list pointer in order to quickly
generate the sysctl path. The motivation behind this patch is to avoid
parameter loading cludges inside the OFED driver subsystem. Instead of
adding special code to the OFED driver subsystem to post-load tunables
into dynamically created sysctls, we generalize this in the kernel.
Other changes:
- Corrected a possibly incorrect sysctl name from "hw.cbb.intr_mask"
to "hw.pcic.intr_mask".
- Removed redundant TUNABLE statements throughout the kernel.
- Some minor code rewrites in connection to removing not needed
TUNABLE statements.
- Added a missing SYSCTL_DECL().
- Wrapped two very long lines.
- Avoid malloc()/free() inside sysctl string handling, in case it is
called to initialize a sysctl from a tunable, hence malloc()/free() is
not ready when sysctls from the sysctl dataset are registered.
- Bumped FreeBSD version to indicate SYSCTL API change.
MFC after: 2 weeks
Sponsored by: Mellanox Technologies
These were used to control/export dispatch policy but they're not anymore.
This commit cannot be MFC'ed to 9 because old netstat(9) binary relies
on such sysctl to work. On the other hand, there's no real reason to
keep'em around in 10.
The SYSCTL_NODE macro defines a list that stores all child-elements of
that node. If there's no SYSCTL_DECL macro anywhere else, there's no
reason why it shouldn't be static.
be brought up in the order they are enumerated in the device tree (in
particular, that thread 0 on each core be brought up first). The SLIST
through which we loop to start the CPUs has all of its entries added with
SLIST_INSERT_HEAD(), which means it is in reverse order of enumeration
and so AP startup would always fail in such situations (causing a machine
check or RTAS failure). Fix this by changing the SLIST into an STAILQ,
and inserting new CPUs at the end.
Reviewed by: jhb
be represented:
- A single policy namespace is defined, consisting of four possible
policies: "default" to use the global default, "deferred" to force
deferred dispatch, "direct" to employ direct dispatch where possible, and
"hybrid" which makes a dynamic decision based on CPU affinity, ordering,
etc. Routines are implemented to convert between strings and an integer
namespace.
- A new global variable, netisr_dispatch_policy, subsumes existing global
variables for direct dispatch, forced direct dispatch, etc, and is used
for explicit policy interpretation and composition. Old variables remain
so that they can be exported by legacy sysctls for use by old netstat(1)
binaries. A new sysctl and tunable, netisr.dispatch.policy, accepts the
above strings for specifying a global policy default.
- The protocol registration structure, netisr_handler, grows an nh_dispatch
field, which accepts a per-policy policy override. The default value is
'0', which corresponds to "default", meaning that protocols will accept
the global default policy unless otherwise specified.
- Policies are now interpreted and composed explicitly at various points in
packet dispatch; protocol policies override global policies.
- Protocols grow the ability to express a non-opinion about affinity even
when implenting m2cpuid by returning NETISR_CPUID_NONE. In that case, the
framework falls back on source ordering, rather than simply using the
current CPU.
These changes are in support of allowing link layer re-dispatch based on
RSS or similar hashes provided by NICs, especially in the case where the
number of hardware receive queues matches hardware core count, rather than
hardware thread count, requiring further software redistributeon. (i.e.,
on RMI XLR).
MFC after: 3 weeks
Reviewed by: bz
Sponsored by: Juniper Networks, Inc.
Make VNET_ASSERT() available with either VNET_DEBUG or INVARIANTS.
Change the syntax to match KASSERT() to allow more flexible panic
messages rather than having a printf with hardcoded arguments
before panic.
Adjust the few assertions we have to the new format (and enhance
the output).
Sponsored by: The FreeBSD Foundation
Sponsored by: CK Software GmbH
Reviewed by: jhb
MFC after: 2 weeks
- Rename the netisr protocol registration array, 'np' to 'netisr_proto',
in order to reduce the chances of symbol name collisions. It remains
statically defined, but it will be looked up by netstat(1).
- Move certain internal structure definitions from netisr.c to
netisr_internal.h so that netstat(1) can find them. They remain
private, and should not be used for any other purpose (for example,
they should not be used by kernel modules, which must instead use the
public interfaces in netisr.h).
- Store a kernel-compiled version of NETISR_MAXPROT in the global variable
netisr_maxprot, and export via a sysctl, so that it is available for use
by netstat(1). This is especially important for crashdump
interpretation, where the size of the workstream structure is determined
by the maximum number of protocols compiled into the kernel.
MFC after: 1 week
Sponsored by: Juniper Networks
counters have not gone about MAXCPU or NETISR_MAXPROT. These problems
caused panics on UP kernels with INVARIANTS when using sysctl -a, but
would also have caused problems for 32-core boxes or if the netisr
protocol vector was fully populated.
Reported by: nwhitehorn, Neel Natu <neelnatu@gmail.com>
MFC after: 4 days
vnet.h, we now use jails (rather than vimages) as the abstraction
for virtualization management, and what remained was specific to
virtual network stacks. Minor cleanups are done in the process,
and comments updated to reflect these changes.
Reviewed by: bz
Approved by: re (vimage blanket)
MAXCPU to mp_maxid, and handling and reporting of requests to use more
threads than we have CPUs to run them on.
Reviewed by: bz
Approved by: re (kib)
MFC after: 6 weeks
arrays to [MAXCPU], offering moderate memory savings. In some places,
this requires using CPU_ABSENT() to handle less common platforms with
sparse CPU IDs. In several places, assert that the selected CPUID for
work placement or statistics is not CPU_ABSENT() to be on the safe side.
Discussed with: bz, jeff
queue was drained. It will never fire for a directly dispatched packet.
You will most likely never want to use this for any ordinary netisr usage
and you will never blame netisr in case you try to use it and it does
not work as expected.
Reviewed by: rwatson
required for options DEVICE_POLLING.
De-fragment the NETISR_ constant space and lower NETISR_MAXPROT from
32 to 16 -- when sizing queue arrays using this compile-time constant,
significant amounts of memory are saved.
Warn on the console when tunable values for netisr are automatically
adjusted during boot due to exceeding limits, invalid values, or as a
result of DEVICE_POLLING.
threads:
- Support up to one netisr thread per CPU, each processings its own
workstream, or set of per-protocol queues. Threads may be bound
to specific CPUs, or allowed to migrate, based on a global policy.
In the future it would be desirable to support topology-centric
policies, such as "one netisr per package".
- Allow each protocol to advertise an ordering policy, which can
currently be one of:
NETISR_POLICY_SOURCE: packets must maintain ordering with respect to
an implicit or explicit source (such as an interface or socket).
NETISR_POLICY_FLOW: make use of mbuf flow identifiers to place work,
as well as allowing protocols to provide a flow generation function
for mbufs without flow identifers (m2flow). Falls back on
NETISR_POLICY_SOURCE if now flow ID is available.
NETISR_POLICY_CPU: allow protocols to inspect and assign a CPU for
each packet handled by netisr (m2cpuid).
- Provide utility functions for querying the number of workstreams
being used, as well as a mapping function from workstream to CPU ID,
which protocols may use in work placement decisions.
- Add explicit interfaces to get and set per-protocol queue limits, and
get and clear drop counters, which query data or apply changes across
all workstreams.
- Add a more extensible netisr registration interface, in which
protocols declare 'struct netisr_handler' structures for each
registered NETISR_ type. These include name, handler function,
optional mbuf to flow ID function, optional mbuf to CPU ID function,
queue limit, and ordering policy. Padding is present to allow these
to be expanded in the future. If no queue limit is declared, then
a default is used.
- Queue limits are now per-workstream, and raised from the previous
IFQ_MAXLEN default of 50 to 256.
- All protocols are updated to use the new registration interface, and
with the exception of netnatm, default queue limits. Most protocols
register as NETISR_POLICY_SOURCE, except IPv4 and IPv6, which use
NETISR_POLICY_FLOW, and will therefore take advantage of driver-
generated flow IDs if present.
- Formalize a non-packet based interface between interface polling and
the netisr, rather than having polling pretend to be two protocols.
Provide two explicit hooks in the netisr worker for start and end
events for runs: netisr_poll() and netisr_pollmore(), as well as a
function, netisr_sched_poll(), to allow the polling code to schedule
netisr execution. DEVICE_POLLING still embeds single-netisr
assumptions in its implementation, so for now if it is compiled into
the kernel, a single and un-bound netisr thread is enforced
regardless of tunable configuration.
In the default configuration, the new netisr implementation maintains
the same basic assumptions as the previous implementation: a single,
un-bound worker thread processes all deferred work, and direct dispatch
is enabled by default wherever possible.
Performance measurement shows a marginal performance improvement over
the old implementation due to the use of batched dequeue.
An rmlock is used to synchronize use and registration/unregistration
using the framework; currently, synchronized use is disabled
(replicating current netisr policy) due to a measurable 3%-6% hit in
ping-pong micro-benchmarking. It will be enabled once further rmlock
optimization has taken place. However, in practice, netisrs are
rarely registered or unregistered at runtime.
A new man page for netisr will follow, but since one doesn't currently
exist, it hasn't been updated.
This change is not appropriate for MFC, although the polling shutdown
handler should be merged to 7-STABLE.
Bump __FreeBSD_version.
Reviewed by: bz
direct dispatch policy for specific protocols (NETISR_USB). We leave
the additional 'flags' argument to netisr_register() for the time being,
even though it is no longer required.
previously always pointing to the default vnet context, to a
dynamically changing thread-local one. The currvnet context
should be set on entry to networking code via CURVNET_SET() macros,
and reverted to previous state via CURVNET_RESTORE(). Recursions
on curvnet are permitted, though strongly discuouraged.
This change should have no functional impact on nooptions VIMAGE
kernel builds, where CURVNET_* macros expand to whitespace.
The curthread->td_vnet (aka curvnet) variable's purpose is to be an
indicator of the vnet context in which the current network-related
operation takes place, in case we cannot deduce the current vnet
context from any other source, such as by looking at mbuf's
m->m_pkthdr.rcvif->if_vnet, sockets's so->so_vnet etc. Moreover, so
far curvnet has turned out to be an invaluable consistency checking
aid: it helps to catch cases when sockets, ifnets or any other
vnet-aware structures may have leaked from one vnet to another.
The exact placement of the CURVNET_SET() / CURVNET_RESTORE() macros
was a result of an empirical iterative process, whith an aim to
reduce recursions on CURVNET_SET() to a minimum, while still reducing
the scope of CURVNET_SET() to networking only operations - the
alternative would be calling CURVNET_SET() on each system call entry.
In general, curvnet has to be set in three typicall cases: when
processing socket-related requests from userspace or from within the
kernel; when processing inbound traffic flowing from device drivers
to upper layers of the networking stack, and when executing
timer-driven networking functions.
This change also introduces a DDB subcommand to show the list of all
vnet instances.
Approved by: julian (mentor)
dispatched without Giant, and add NETISR_FORCEQUEUE, which allows specific
netisr handlers to always be dispatched via a queue (deferred). Mark the
usb and if_ppp netisr handlers as NETISR_FORCEQUEUE, and explicitly
acquire Giant in those handlers.
Previously, any netisr handler not marked NETISR_MPSAFE would necessarily
run deferred and with Giant acquired. This change removes Giant
scaffolding from the netisr infrastructure, but NETISR_FORCEQUEUE allows
non-MPSAFE handlers to continue to force deferred dispatch so as to avoid
lock order reversals between their acqusition of Giant and any calling
context.
It is likely we will be able to remove NETISR_FORCEQUEUE once
IFF_NEEDSGIANT is removed, as non-MPSAFE usb and if_ppp drivers will no
longer be supported.
Reviewed by: bz
MFC after: 1 month
X-MFC note: We can't remove NETISR_MPSAFE from stable/7 for KPI reasons,
but the rest can go back.
after each SYSINIT() macro invocation. This makes a number of
lightweight C parsers much happier with the FreeBSD kernel
source, including cflow's prcc and lxr.
MFC after: 1 month
Discussed with: imp, rink
previously conditionally acquired Giant based on debug.mpsafenet. As that
has now been removed, they are no longer required. Removing them
significantly simplifies error-handling in the socket layer, eliminated
quite a bit of unwinding of locking in error cases.
While here clean up the now unneeded opt_net.h, which previously was used
for the NET_WITH_GIANT kernel option. Clean up some related gotos for
consistency.
Reviewed by: bz, csjp
Tested by: kris
Approved by: re (kensmith)
framework for non-MPSAFE network protocols:
- Remove debug_mpsafenet variable, sysctl, and tunable.
- Remove NET_NEEDS_GIANT() and associate SYSINITSs used by it to force
debug.mpsafenet=0 if non-MPSAFE protocols are compiled into the kernel.
- Remove logic to automatically flag interrupt handlers as non-MPSAFE if
debug.mpsafenet is set for an INTR_TYPE_NET handler.
- Remove logic to automatically flag netisr handlers as non-MPSAFE if
debug.mpsafenet is set.
- Remove references in a few subsystems, including NFS and Cronyx drivers,
which keyed off debug_mpsafenet to determine various aspects of their own
locking behavior.
- Convert NET_LOCK_GIANT(), NET_UNLOCK_GIANT(), and NET_ASSERT_GIANT into
no-op's, as their entire behavior was determined by the value in
debug_mpsafenet.
- Alias NET_CALLOUT_MPSAFE to CALLOUT_MPSAFE.
Many remaining references to NET_.*_GIANT() and NET_CALLOUT_MPSAFE are still
present in subsystems, and will be removed in followup commits.
Reviewed by: bz, jhb
Approved by: re (kensmith)
enables direct dispatch of the network stack from the device driver
ithread, enabling input path parallelism by default when multiple
interfaces are present.
The strategy for network stack parallelism is something being actively
discussed, and this is just one of several possible (and perfectly
reasonable) strategies, but has the distinct advantage of reducing the
number of context switches and preemptions significantly, resulting in
higher efficiency in many cases. In some caes, this may reduce
network stack parallelism due to work not being deferred from the
ithread to the netisr. Therefore, the strategy may change in the
future, but this offers a reasonable first pass and enabling
parallelism while maintaining strong ordering.
Hopefully this will trigger lots of nice new bugs.
This change is not intended for MFC.