2 * Copyright (c) 2007-2009 Robert N. M. Watson
3 * Copyright (c) 2010-2011 Juniper Networks, Inc.
6 * This software was developed by Robert N. M. Watson under contract
7 * to Juniper Networks, Inc.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
35 * netisr is a packet dispatch service, allowing synchronous (directly
36 * dispatched) and asynchronous (deferred dispatch) processing of packets by
37 * registered protocol handlers. Callers pass a protocol identifier and
38 * packet to netisr, along with a direct dispatch hint, and work will either
39 * be immediately processed by the registered handler, or passed to a
40 * software interrupt (SWI) thread for deferred dispatch. Callers will
41 * generally select one or the other based on:
43 * - Whether directly dispatching a netisr handler lead to code reentrance or
44 * lock recursion, such as entering the socket code from the socket code.
45 * - Whether directly dispatching a netisr handler lead to recursive
46 * processing, such as when decapsulating several wrapped layers of tunnel
47 * information (IPSEC within IPSEC within ...).
49 * Maintaining ordering for protocol streams is a critical design concern.
50 * Enforcing ordering limits the opportunity for concurrency, but maintains
51 * the strong ordering requirements found in some protocols, such as TCP. Of
52 * related concern is CPU affinity--it is desirable to process all data
53 * associated with a particular stream on the same CPU over time in order to
54 * avoid acquiring locks associated with the connection on different CPUs,
55 * keep connection data in one cache, and to generally encourage associated
56 * user threads to live on the same CPU as the stream. It's also desirable
57 * to avoid lock migration and contention where locks are associated with
60 * netisr supports several policy variations, represented by the
61 * NETISR_POLICY_* constants, allowing protocols to play various roles in
62 * identifying flows, assigning work to CPUs, etc. These are described in
67 #include "opt_device_polling.h"
69 #include <sys/param.h>
71 #include <sys/kernel.h>
72 #include <sys/kthread.h>
73 #include <sys/interrupt.h>
76 #include <sys/mutex.h>
79 #include <sys/rmlock.h>
80 #include <sys/sched.h>
82 #include <sys/socket.h>
83 #include <sys/sysctl.h>
84 #include <sys/systm.h>
90 #define _WANT_NETISR_INTERNAL /* Enable definitions from netisr_internal.h */
92 #include <net/if_var.h>
93 #include <net/netisr.h>
94 #include <net/netisr_internal.h>
98 * Synchronize use and modification of the registered netisr data structures;
99 * acquire a read lock while modifying the set of registered protocols to
100 * prevent partially registered or unregistered protocols from being run.
102 * The following data structures and fields are protected by this lock:
104 * - The netisr_proto array, including all fields of struct netisr_proto.
105 * - The nws array, including all fields of struct netisr_worker.
106 * - The nws_array array.
108 * Note: the NETISR_LOCKING define controls whether read locks are acquired
109 * in packet processing paths requiring netisr registration stability. This
110 * is disabled by default as it can lead to measurable performance
111 * degradation even with rmlocks (3%-6% for loopback ping-pong traffic), and
112 * because netisr registration and unregistration is extremely rare at
113 * runtime. If it becomes more common, this decision should be revisited.
115 * XXXRW: rmlocks don't support assertions.
117 static struct rmlock netisr_rmlock;
118 #define NETISR_LOCK_INIT() rm_init_flags(&netisr_rmlock, "netisr", \
120 #define NETISR_LOCK_ASSERT()
121 #define NETISR_RLOCK(tracker) rm_rlock(&netisr_rmlock, (tracker))
122 #define NETISR_RUNLOCK(tracker) rm_runlock(&netisr_rmlock, (tracker))
123 #define NETISR_WLOCK() rm_wlock(&netisr_rmlock)
124 #define NETISR_WUNLOCK() rm_wunlock(&netisr_rmlock)
125 /* #define NETISR_LOCKING */
127 static SYSCTL_NODE(_net, OID_AUTO, isr, CTLFLAG_RW, 0, "netisr");
130 * Three global direct dispatch policies are supported:
132 * NETISR_DISPATCH_DEFERRED: All work is deferred for a netisr, regardless of
133 * context (may be overriden by protocols).
135 * NETISR_DISPATCH_HYBRID: If the executing context allows direct dispatch,
136 * and we're running on the CPU the work would be performed on, then direct
137 * dispatch it if it wouldn't violate ordering constraints on the workstream.
139 * NETISR_DISPATCH_DIRECT: If the executing context allows direct dispatch,
140 * always direct dispatch. (The default.)
142 * Notice that changing the global policy could lead to short periods of
143 * misordered processing, but this is considered acceptable as compared to
144 * the complexity of enforcing ordering during policy changes. Protocols can
145 * override the global policy (when they're not doing that, they select
146 * NETISR_DISPATCH_DEFAULT).
148 #define NETISR_DISPATCH_POLICY_DEFAULT NETISR_DISPATCH_DIRECT
149 #define NETISR_DISPATCH_POLICY_MAXSTR 20 /* Used for temporary buffers. */
150 static u_int netisr_dispatch_policy = NETISR_DISPATCH_POLICY_DEFAULT;
151 static int sysctl_netisr_dispatch_policy(SYSCTL_HANDLER_ARGS);
152 SYSCTL_PROC(_net_isr, OID_AUTO, dispatch, CTLTYPE_STRING | CTLFLAG_RWTUN,
153 0, 0, sysctl_netisr_dispatch_policy, "A",
154 "netisr dispatch policy");
157 * Allow the administrator to limit the number of threads (CPUs) to use for
158 * netisr. We don't check netisr_maxthreads before creating the thread for
159 * CPU 0. This must be set at boot. We will create at most one thread per CPU.
160 * By default we initialize this to 1 which would assign just 1 cpu (cpu0) and
161 * therefore only 1 workstream. If set to -1, netisr would use all cpus
162 * (mp_ncpus) and therefore would have those many workstreams. One workstream
165 static int netisr_maxthreads = 1; /* Max number of threads. */
166 SYSCTL_INT(_net_isr, OID_AUTO, maxthreads, CTLFLAG_RDTUN,
167 &netisr_maxthreads, 0,
168 "Use at most this many CPUs for netisr processing");
170 static int netisr_bindthreads = 0; /* Bind threads to CPUs. */
171 SYSCTL_INT(_net_isr, OID_AUTO, bindthreads, CTLFLAG_RDTUN,
172 &netisr_bindthreads, 0, "Bind netisr threads to CPUs.");
175 * Limit per-workstream mbuf queue limits s to at most net.isr.maxqlimit,
176 * both for initial configuration and later modification using
177 * netisr_setqlimit().
179 #define NETISR_DEFAULT_MAXQLIMIT 10240
180 static u_int netisr_maxqlimit = NETISR_DEFAULT_MAXQLIMIT;
181 SYSCTL_UINT(_net_isr, OID_AUTO, maxqlimit, CTLFLAG_RDTUN,
182 &netisr_maxqlimit, 0,
183 "Maximum netisr per-protocol, per-CPU queue depth.");
186 * The default per-workstream mbuf queue limit for protocols that don't
187 * initialize the nh_qlimit field of their struct netisr_handler. If this is
188 * set above netisr_maxqlimit, we truncate it to the maximum during boot.
190 #define NETISR_DEFAULT_DEFAULTQLIMIT 256
191 static u_int netisr_defaultqlimit = NETISR_DEFAULT_DEFAULTQLIMIT;
192 SYSCTL_UINT(_net_isr, OID_AUTO, defaultqlimit, CTLFLAG_RDTUN,
193 &netisr_defaultqlimit, 0,
194 "Default netisr per-protocol, per-CPU queue limit if not set by protocol");
197 * Store and export the compile-time constant NETISR_MAXPROT limit on the
198 * number of protocols that can register with netisr at a time. This is
199 * required for crashdump analysis, as it sizes netisr_proto[].
201 static u_int netisr_maxprot = NETISR_MAXPROT;
202 SYSCTL_UINT(_net_isr, OID_AUTO, maxprot, CTLFLAG_RD,
204 "Compile-time limit on the number of protocols supported by netisr.");
207 * The netisr_proto array describes all registered protocols, indexed by
208 * protocol number. See netisr_internal.h for more details.
210 static struct netisr_proto netisr_proto[NETISR_MAXPROT];
213 * Per-CPU workstream data. See netisr_internal.h for more details.
215 DPCPU_DEFINE(struct netisr_workstream, nws);
218 * Map contiguous values between 0 and nws_count into CPU IDs appropriate for
219 * accessing workstreams. This allows constructions of the form
220 * DPCPU_ID_GET(nws_array[arbitraryvalue % nws_count], nws).
222 static u_int nws_array[MAXCPU];
225 * Number of registered workstreams. Will be at most the number of running
226 * CPUs once fully started.
228 static u_int nws_count;
229 SYSCTL_UINT(_net_isr, OID_AUTO, numthreads, CTLFLAG_RD,
230 &nws_count, 0, "Number of extant netisr threads.");
233 * Synchronization for each workstream: a mutex protects all mutable fields
234 * in each stream, including per-protocol state (mbuf queues). The SWI is
235 * woken up if asynchronous dispatch is required.
237 #define NWS_LOCK(s) mtx_lock(&(s)->nws_mtx)
238 #define NWS_LOCK_ASSERT(s) mtx_assert(&(s)->nws_mtx, MA_OWNED)
239 #define NWS_UNLOCK(s) mtx_unlock(&(s)->nws_mtx)
240 #define NWS_SIGNAL(s) swi_sched((s)->nws_swi_cookie, 0)
243 * Utility routines for protocols that implement their own mapping of flows
247 netisr_get_cpucount(void)
254 netisr_get_cpuid(u_int cpunumber)
257 KASSERT(cpunumber < nws_count, ("%s: %u > %u", __func__, cpunumber,
260 return (nws_array[cpunumber]);
264 * The default implementation of flow -> CPU ID mapping.
266 * Non-static so that protocols can use it to map their own work to specific
267 * CPUs in a manner consistent to netisr for affinity purposes.
270 netisr_default_flow2cpu(u_int flowid)
273 return (nws_array[flowid % nws_count]);
277 * Dispatch tunable and sysctl configuration.
279 struct netisr_dispatch_table_entry {
281 const char *ndte_policy_str;
283 static const struct netisr_dispatch_table_entry netisr_dispatch_table[] = {
284 { NETISR_DISPATCH_DEFAULT, "default" },
285 { NETISR_DISPATCH_DEFERRED, "deferred" },
286 { NETISR_DISPATCH_HYBRID, "hybrid" },
287 { NETISR_DISPATCH_DIRECT, "direct" },
289 static const u_int netisr_dispatch_table_len =
290 (sizeof(netisr_dispatch_table) / sizeof(netisr_dispatch_table[0]));
293 netisr_dispatch_policy_to_str(u_int dispatch_policy, char *buffer,
296 const struct netisr_dispatch_table_entry *ndtep;
301 for (i = 0; i < netisr_dispatch_table_len; i++) {
302 ndtep = &netisr_dispatch_table[i];
303 if (ndtep->ndte_policy == dispatch_policy) {
304 str = ndtep->ndte_policy_str;
308 snprintf(buffer, buflen, "%s", str);
312 netisr_dispatch_policy_from_str(const char *str, u_int *dispatch_policyp)
314 const struct netisr_dispatch_table_entry *ndtep;
317 for (i = 0; i < netisr_dispatch_table_len; i++) {
318 ndtep = &netisr_dispatch_table[i];
319 if (strcmp(ndtep->ndte_policy_str, str) == 0) {
320 *dispatch_policyp = ndtep->ndte_policy;
328 sysctl_netisr_dispatch_policy(SYSCTL_HANDLER_ARGS)
330 char tmp[NETISR_DISPATCH_POLICY_MAXSTR];
331 u_int dispatch_policy;
334 netisr_dispatch_policy_to_str(netisr_dispatch_policy, tmp,
336 error = sysctl_handle_string(oidp, tmp, sizeof(tmp), req);
337 if (error == 0 && req->newptr != NULL) {
338 error = netisr_dispatch_policy_from_str(tmp,
340 if (error == 0 && dispatch_policy == NETISR_DISPATCH_DEFAULT)
343 netisr_dispatch_policy = dispatch_policy;
349 * Register a new netisr handler, which requires initializing per-protocol
350 * fields for each workstream. All netisr work is briefly suspended while
351 * the protocol is installed.
354 netisr_register(const struct netisr_handler *nhp)
356 struct netisr_work *npwp;
360 proto = nhp->nh_proto;
364 * Test that the requested registration is valid.
366 KASSERT(nhp->nh_name != NULL,
367 ("%s: nh_name NULL for %u", __func__, proto));
368 KASSERT(nhp->nh_handler != NULL,
369 ("%s: nh_handler NULL for %s", __func__, name));
370 KASSERT(nhp->nh_policy == NETISR_POLICY_SOURCE ||
371 nhp->nh_policy == NETISR_POLICY_FLOW ||
372 nhp->nh_policy == NETISR_POLICY_CPU,
373 ("%s: unsupported nh_policy %u for %s", __func__,
374 nhp->nh_policy, name));
375 KASSERT(nhp->nh_policy == NETISR_POLICY_FLOW ||
376 nhp->nh_m2flow == NULL,
377 ("%s: nh_policy != FLOW but m2flow defined for %s", __func__,
379 KASSERT(nhp->nh_policy == NETISR_POLICY_CPU || nhp->nh_m2cpuid == NULL,
380 ("%s: nh_policy != CPU but m2cpuid defined for %s", __func__,
382 KASSERT(nhp->nh_policy != NETISR_POLICY_CPU || nhp->nh_m2cpuid != NULL,
383 ("%s: nh_policy == CPU but m2cpuid not defined for %s", __func__,
385 KASSERT(nhp->nh_dispatch == NETISR_DISPATCH_DEFAULT ||
386 nhp->nh_dispatch == NETISR_DISPATCH_DEFERRED ||
387 nhp->nh_dispatch == NETISR_DISPATCH_HYBRID ||
388 nhp->nh_dispatch == NETISR_DISPATCH_DIRECT,
389 ("%s: invalid nh_dispatch (%u)", __func__, nhp->nh_dispatch));
391 KASSERT(proto < NETISR_MAXPROT,
392 ("%s(%u, %s): protocol too big", __func__, proto, name));
395 * Test that no existing registration exists for this protocol.
398 KASSERT(netisr_proto[proto].np_name == NULL,
399 ("%s(%u, %s): name present", __func__, proto, name));
400 KASSERT(netisr_proto[proto].np_handler == NULL,
401 ("%s(%u, %s): handler present", __func__, proto, name));
403 netisr_proto[proto].np_name = name;
404 netisr_proto[proto].np_handler = nhp->nh_handler;
405 netisr_proto[proto].np_m2flow = nhp->nh_m2flow;
406 netisr_proto[proto].np_m2cpuid = nhp->nh_m2cpuid;
407 netisr_proto[proto].np_drainedcpu = nhp->nh_drainedcpu;
408 if (nhp->nh_qlimit == 0)
409 netisr_proto[proto].np_qlimit = netisr_defaultqlimit;
410 else if (nhp->nh_qlimit > netisr_maxqlimit) {
411 printf("%s: %s requested queue limit %u capped to "
412 "net.isr.maxqlimit %u\n", __func__, name, nhp->nh_qlimit,
414 netisr_proto[proto].np_qlimit = netisr_maxqlimit;
416 netisr_proto[proto].np_qlimit = nhp->nh_qlimit;
417 netisr_proto[proto].np_policy = nhp->nh_policy;
418 netisr_proto[proto].np_dispatch = nhp->nh_dispatch;
420 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
421 bzero(npwp, sizeof(*npwp));
422 npwp->nw_qlimit = netisr_proto[proto].np_qlimit;
428 * Clear drop counters across all workstreams for a protocol.
431 netisr_clearqdrops(const struct netisr_handler *nhp)
433 struct netisr_work *npwp;
439 proto = nhp->nh_proto;
443 KASSERT(proto < NETISR_MAXPROT,
444 ("%s(%u): protocol too big for %s", __func__, proto, name));
447 KASSERT(netisr_proto[proto].np_handler != NULL,
448 ("%s(%u): protocol not registered for %s", __func__, proto,
452 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
459 * Query current drop counters across all workstreams for a protocol.
462 netisr_getqdrops(const struct netisr_handler *nhp, u_int64_t *qdropp)
464 struct netisr_work *npwp;
465 struct rm_priotracker tracker;
472 proto = nhp->nh_proto;
476 KASSERT(proto < NETISR_MAXPROT,
477 ("%s(%u): protocol too big for %s", __func__, proto, name));
479 NETISR_RLOCK(&tracker);
480 KASSERT(netisr_proto[proto].np_handler != NULL,
481 ("%s(%u): protocol not registered for %s", __func__, proto,
485 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
486 *qdropp += npwp->nw_qdrops;
488 NETISR_RUNLOCK(&tracker);
492 * Query current per-workstream queue limit for a protocol.
495 netisr_getqlimit(const struct netisr_handler *nhp, u_int *qlimitp)
497 struct rm_priotracker tracker;
503 proto = nhp->nh_proto;
507 KASSERT(proto < NETISR_MAXPROT,
508 ("%s(%u): protocol too big for %s", __func__, proto, name));
510 NETISR_RLOCK(&tracker);
511 KASSERT(netisr_proto[proto].np_handler != NULL,
512 ("%s(%u): protocol not registered for %s", __func__, proto,
514 *qlimitp = netisr_proto[proto].np_qlimit;
515 NETISR_RUNLOCK(&tracker);
519 * Update the queue limit across per-workstream queues for a protocol. We
520 * simply change the limits, and don't drain overflowed packets as they will
521 * (hopefully) take care of themselves shortly.
524 netisr_setqlimit(const struct netisr_handler *nhp, u_int qlimit)
526 struct netisr_work *npwp;
532 if (qlimit > netisr_maxqlimit)
535 proto = nhp->nh_proto;
539 KASSERT(proto < NETISR_MAXPROT,
540 ("%s(%u): protocol too big for %s", __func__, proto, name));
543 KASSERT(netisr_proto[proto].np_handler != NULL,
544 ("%s(%u): protocol not registered for %s", __func__, proto,
547 netisr_proto[proto].np_qlimit = qlimit;
549 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
550 npwp->nw_qlimit = qlimit;
557 * Drain all packets currently held in a particular protocol work queue.
560 netisr_drain_proto(struct netisr_work *npwp)
565 * We would assert the lock on the workstream but it's not passed in.
567 while ((m = npwp->nw_head) != NULL) {
568 npwp->nw_head = m->m_nextpkt;
570 if (npwp->nw_head == NULL)
571 npwp->nw_tail = NULL;
575 KASSERT(npwp->nw_tail == NULL, ("%s: tail", __func__));
576 KASSERT(npwp->nw_len == 0, ("%s: len", __func__));
580 * Remove the registration of a network protocol, which requires clearing
581 * per-protocol fields across all workstreams, including freeing all mbufs in
582 * the queues at time of unregister. All work in netisr is briefly suspended
583 * while this takes place.
586 netisr_unregister(const struct netisr_handler *nhp)
588 struct netisr_work *npwp;
594 proto = nhp->nh_proto;
598 KASSERT(proto < NETISR_MAXPROT,
599 ("%s(%u): protocol too big for %s", __func__, proto, name));
602 KASSERT(netisr_proto[proto].np_handler != NULL,
603 ("%s(%u): protocol not registered for %s", __func__, proto,
606 netisr_proto[proto].np_name = NULL;
607 netisr_proto[proto].np_handler = NULL;
608 netisr_proto[proto].np_m2flow = NULL;
609 netisr_proto[proto].np_m2cpuid = NULL;
610 netisr_proto[proto].np_qlimit = 0;
611 netisr_proto[proto].np_policy = 0;
613 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
614 netisr_drain_proto(npwp);
615 bzero(npwp, sizeof(*npwp));
621 * Compose the global and per-protocol policies on dispatch, and return the
622 * dispatch policy to use.
625 netisr_get_dispatch(struct netisr_proto *npp)
629 * Protocol-specific configuration overrides the global default.
631 if (npp->np_dispatch != NETISR_DISPATCH_DEFAULT)
632 return (npp->np_dispatch);
633 return (netisr_dispatch_policy);
637 * Look up the workstream given a packet and source identifier. Do this by
638 * checking the protocol's policy, and optionally call out to the protocol
639 * for assistance if required.
642 netisr_select_cpuid(struct netisr_proto *npp, u_int dispatch_policy,
643 uintptr_t source, struct mbuf *m, u_int *cpuidp)
648 NETISR_LOCK_ASSERT();
651 * In the event we have only one worker, shortcut and deliver to it
652 * without further ado.
654 if (nws_count == 1) {
655 *cpuidp = nws_array[0];
660 * What happens next depends on the policy selected by the protocol.
661 * If we want to support per-interface policies, we should do that
664 policy = npp->np_policy;
665 if (policy == NETISR_POLICY_CPU) {
666 m = npp->np_m2cpuid(m, source, cpuidp);
671 * It's possible for a protocol not to have a good idea about
672 * where to process a packet, in which case we fall back on
673 * the netisr code to decide. In the hybrid case, return the
674 * current CPU ID, which will force an immediate direct
675 * dispatch. In the queued case, fall back on the SOURCE
678 if (*cpuidp != NETISR_CPUID_NONE)
680 if (dispatch_policy == NETISR_DISPATCH_HYBRID) {
684 policy = NETISR_POLICY_SOURCE;
687 if (policy == NETISR_POLICY_FLOW) {
688 if (M_HASHTYPE_GET(m) == M_HASHTYPE_NONE &&
689 npp->np_m2flow != NULL) {
690 m = npp->np_m2flow(m, source);
694 if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
696 netisr_default_flow2cpu(m->m_pkthdr.flowid);
699 policy = NETISR_POLICY_SOURCE;
702 KASSERT(policy == NETISR_POLICY_SOURCE,
703 ("%s: invalid policy %u for %s", __func__, npp->np_policy,
706 ifp = m->m_pkthdr.rcvif;
708 *cpuidp = nws_array[(ifp->if_index + source) % nws_count];
710 *cpuidp = nws_array[source % nws_count];
715 * Process packets associated with a workstream and protocol. For reasons of
716 * fairness, we process up to one complete netisr queue at a time, moving the
717 * queue to a stack-local queue for processing, but do not loop refreshing
718 * from the global queue. The caller is responsible for deciding whether to
719 * loop, and for setting the NWS_RUNNING flag. The passed workstream will be
720 * locked on entry and relocked before return, but will be released while
721 * processing. The number of packets processed is returned.
724 netisr_process_workstream_proto(struct netisr_workstream *nwsp, u_int proto)
726 struct netisr_work local_npw, *npwp;
730 NETISR_LOCK_ASSERT();
731 NWS_LOCK_ASSERT(nwsp);
733 KASSERT(nwsp->nws_flags & NWS_RUNNING,
734 ("%s(%u): not running", __func__, proto));
735 KASSERT(proto >= 0 && proto < NETISR_MAXPROT,
736 ("%s(%u): invalid proto\n", __func__, proto));
738 npwp = &nwsp->nws_work[proto];
739 if (npwp->nw_len == 0)
743 * Move the global work queue to a thread-local work queue.
745 * Notice that this means the effective maximum length of the queue
746 * is actually twice that of the maximum queue length specified in
747 * the protocol registration call.
749 handled = npwp->nw_len;
751 npwp->nw_head = NULL;
752 npwp->nw_tail = NULL;
754 nwsp->nws_pendingbits &= ~(1 << proto);
756 while ((m = local_npw.nw_head) != NULL) {
757 local_npw.nw_head = m->m_nextpkt;
759 if (local_npw.nw_head == NULL)
760 local_npw.nw_tail = NULL;
762 VNET_ASSERT(m->m_pkthdr.rcvif != NULL,
763 ("%s:%d rcvif == NULL: m=%p", __func__, __LINE__, m));
764 CURVNET_SET(m->m_pkthdr.rcvif->if_vnet);
765 netisr_proto[proto].np_handler(m);
768 KASSERT(local_npw.nw_len == 0,
769 ("%s(%u): len %u", __func__, proto, local_npw.nw_len));
770 if (netisr_proto[proto].np_drainedcpu)
771 netisr_proto[proto].np_drainedcpu(nwsp->nws_cpu);
773 npwp->nw_handled += handled;
778 * SWI handler for netisr -- processes packets in a set of workstreams that
779 * it owns, woken up by calls to NWS_SIGNAL(). If this workstream is already
780 * being direct dispatched, go back to sleep and wait for the dispatching
781 * thread to wake us up again.
786 #ifdef NETISR_LOCKING
787 struct rm_priotracker tracker;
789 struct netisr_workstream *nwsp;
794 #ifdef DEVICE_POLLING
795 KASSERT(nws_count == 1,
796 ("%s: device_polling but nws_count != 1", __func__));
799 #ifdef NETISR_LOCKING
800 NETISR_RLOCK(&tracker);
803 KASSERT(!(nwsp->nws_flags & NWS_RUNNING), ("swi_net: running"));
804 if (nwsp->nws_flags & NWS_DISPATCHING)
806 nwsp->nws_flags |= NWS_RUNNING;
807 nwsp->nws_flags &= ~NWS_SCHEDULED;
808 while ((bits = nwsp->nws_pendingbits) != 0) {
809 while ((prot = ffs(bits)) != 0) {
811 bits &= ~(1 << prot);
812 (void)netisr_process_workstream_proto(nwsp, prot);
815 nwsp->nws_flags &= ~NWS_RUNNING;
818 #ifdef NETISR_LOCKING
819 NETISR_RUNLOCK(&tracker);
821 #ifdef DEVICE_POLLING
827 netisr_queue_workstream(struct netisr_workstream *nwsp, u_int proto,
828 struct netisr_work *npwp, struct mbuf *m, int *dosignalp)
831 NWS_LOCK_ASSERT(nwsp);
834 if (npwp->nw_len < npwp->nw_qlimit) {
836 if (npwp->nw_head == NULL) {
840 npwp->nw_tail->m_nextpkt = m;
844 if (npwp->nw_len > npwp->nw_watermark)
845 npwp->nw_watermark = npwp->nw_len;
848 * We must set the bit regardless of NWS_RUNNING, so that
849 * swi_net() keeps calling netisr_process_workstream_proto().
851 nwsp->nws_pendingbits |= (1 << proto);
852 if (!(nwsp->nws_flags &
853 (NWS_RUNNING | NWS_DISPATCHING | NWS_SCHEDULED))) {
854 nwsp->nws_flags |= NWS_SCHEDULED;
855 *dosignalp = 1; /* Defer until unlocked. */
867 netisr_queue_internal(u_int proto, struct mbuf *m, u_int cpuid)
869 struct netisr_workstream *nwsp;
870 struct netisr_work *npwp;
873 #ifdef NETISR_LOCKING
874 NETISR_LOCK_ASSERT();
876 KASSERT(cpuid <= mp_maxid, ("%s: cpuid too big (%u, %u)", __func__,
878 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
882 nwsp = DPCPU_ID_PTR(cpuid, nws);
883 npwp = &nwsp->nws_work[proto];
885 error = netisr_queue_workstream(nwsp, proto, npwp, m, &dosignal);
893 netisr_queue_src(u_int proto, uintptr_t source, struct mbuf *m)
895 #ifdef NETISR_LOCKING
896 struct rm_priotracker tracker;
901 KASSERT(proto < NETISR_MAXPROT,
902 ("%s: invalid proto %u", __func__, proto));
904 #ifdef NETISR_LOCKING
905 NETISR_RLOCK(&tracker);
907 KASSERT(netisr_proto[proto].np_handler != NULL,
908 ("%s: invalid proto %u", __func__, proto));
910 m = netisr_select_cpuid(&netisr_proto[proto], NETISR_DISPATCH_DEFERRED,
913 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__,
915 error = netisr_queue_internal(proto, m, cpuid);
918 #ifdef NETISR_LOCKING
919 NETISR_RUNLOCK(&tracker);
925 netisr_queue(u_int proto, struct mbuf *m)
928 return (netisr_queue_src(proto, 0, m));
932 * Dispatch a packet for netisr processing; direct dispatch is permitted by
936 netisr_dispatch_src(u_int proto, uintptr_t source, struct mbuf *m)
938 #ifdef NETISR_LOCKING
939 struct rm_priotracker tracker;
941 struct netisr_workstream *nwsp;
942 struct netisr_proto *npp;
943 struct netisr_work *npwp;
945 u_int cpuid, dispatch_policy;
947 KASSERT(proto < NETISR_MAXPROT,
948 ("%s: invalid proto %u", __func__, proto));
949 #ifdef NETISR_LOCKING
950 NETISR_RLOCK(&tracker);
952 npp = &netisr_proto[proto];
953 KASSERT(npp->np_handler != NULL, ("%s: invalid proto %u", __func__,
956 dispatch_policy = netisr_get_dispatch(npp);
957 if (dispatch_policy == NETISR_DISPATCH_DEFERRED)
958 return (netisr_queue_src(proto, source, m));
961 * If direct dispatch is forced, then unconditionally dispatch
962 * without a formal CPU selection. Borrow the current CPU's stats,
963 * even if there's no worker on it. In this case we don't update
964 * nws_flags because all netisr processing will be source ordered due
965 * to always being forced to directly dispatch.
967 if (dispatch_policy == NETISR_DISPATCH_DIRECT) {
968 nwsp = DPCPU_PTR(nws);
969 npwp = &nwsp->nws_work[proto];
970 npwp->nw_dispatched++;
972 netisr_proto[proto].np_handler(m);
977 KASSERT(dispatch_policy == NETISR_DISPATCH_HYBRID,
978 ("%s: unknown dispatch policy (%u)", __func__, dispatch_policy));
981 * Otherwise, we execute in a hybrid mode where we will try to direct
982 * dispatch if we're on the right CPU and the netisr worker isn't
986 m = netisr_select_cpuid(&netisr_proto[proto], NETISR_DISPATCH_HYBRID,
992 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
995 nwsp = DPCPU_PTR(nws);
996 npwp = &nwsp->nws_work[proto];
999 * We are willing to direct dispatch only if three conditions hold:
1001 * (1) The netisr worker isn't already running,
1002 * (2) Another thread isn't already directly dispatching, and
1003 * (3) The netisr hasn't already been woken up.
1006 if (nwsp->nws_flags & (NWS_RUNNING | NWS_DISPATCHING | NWS_SCHEDULED)) {
1007 error = netisr_queue_workstream(nwsp, proto, npwp, m,
1016 * The current thread is now effectively the netisr worker, so set
1017 * the dispatching flag to prevent concurrent processing of the
1018 * stream from another thread (even the netisr worker), which could
1019 * otherwise lead to effective misordering of the stream.
1021 nwsp->nws_flags |= NWS_DISPATCHING;
1023 netisr_proto[proto].np_handler(m);
1025 nwsp->nws_flags &= ~NWS_DISPATCHING;
1027 npwp->nw_hybrid_dispatched++;
1030 * If other work was enqueued by another thread while we were direct
1031 * dispatching, we need to signal the netisr worker to do that work.
1032 * In the future, we might want to do some of that work in the
1033 * current thread, rather than trigger further context switches. If
1034 * so, we'll want to establish a reasonable bound on the work done in
1035 * the "borrowed" context.
1037 if (nwsp->nws_pendingbits != 0) {
1038 nwsp->nws_flags |= NWS_SCHEDULED;
1049 error = netisr_queue_internal(proto, m, cpuid);
1053 #ifdef NETISR_LOCKING
1054 NETISR_RUNLOCK(&tracker);
1060 netisr_dispatch(u_int proto, struct mbuf *m)
1063 return (netisr_dispatch_src(proto, 0, m));
1066 #ifdef DEVICE_POLLING
1068 * Kernel polling borrows a netisr thread to run interface polling in; this
1069 * function allows kernel polling to request that the netisr thread be
1070 * scheduled even if no packets are pending for protocols.
1073 netisr_sched_poll(void)
1075 struct netisr_workstream *nwsp;
1077 nwsp = DPCPU_ID_PTR(nws_array[0], nws);
1083 netisr_start_swi(u_int cpuid, struct pcpu *pc)
1086 struct netisr_workstream *nwsp;
1089 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
1091 nwsp = DPCPU_ID_PTR(cpuid, nws);
1092 mtx_init(&nwsp->nws_mtx, "netisr_mtx", NULL, MTX_DEF);
1093 nwsp->nws_cpu = cpuid;
1094 snprintf(swiname, sizeof(swiname), "netisr %u", cpuid);
1095 error = swi_add(&nwsp->nws_intr_event, swiname, swi_net, nwsp,
1096 SWI_NET, INTR_MPSAFE, &nwsp->nws_swi_cookie);
1098 panic("%s: swi_add %d", __func__, error);
1099 pc->pc_netisr = nwsp->nws_intr_event;
1100 if (netisr_bindthreads) {
1101 error = intr_event_bind(nwsp->nws_intr_event, cpuid);
1103 printf("%s: cpu %u: intr_event_bind: %d", __func__,
1107 nws_array[nws_count] = nwsp->nws_cpu;
1113 * Initialize the netisr subsystem. We rely on BSS and static initialization
1114 * of most fields in global data structures.
1116 * Start a worker thread for the boot CPU so that we can support network
1117 * traffic immediately in case the network stack is used before additional
1118 * CPUs are started (for example, diskless boot).
1121 netisr_init(void *arg)
1123 KASSERT(curcpu == 0, ("%s: not on CPU 0", __func__));
1126 if (netisr_maxthreads == 0 || netisr_maxthreads < -1 )
1127 netisr_maxthreads = 1; /* default behavior */
1128 else if (netisr_maxthreads == -1)
1129 netisr_maxthreads = mp_ncpus; /* use max cpus */
1130 if (netisr_maxthreads > mp_ncpus) {
1131 printf("netisr_init: forcing maxthreads from %d to %d\n",
1132 netisr_maxthreads, mp_ncpus);
1133 netisr_maxthreads = mp_ncpus;
1135 if (netisr_defaultqlimit > netisr_maxqlimit) {
1136 printf("netisr_init: forcing defaultqlimit from %d to %d\n",
1137 netisr_defaultqlimit, netisr_maxqlimit);
1138 netisr_defaultqlimit = netisr_maxqlimit;
1140 #ifdef DEVICE_POLLING
1142 * The device polling code is not yet aware of how to deal with
1143 * multiple netisr threads, so for the time being compiling in device
1144 * polling disables parallel netisr workers.
1146 if (netisr_maxthreads != 1 || netisr_bindthreads != 0) {
1147 printf("netisr_init: forcing maxthreads to 1 and "
1148 "bindthreads to 0 for device polling\n");
1149 netisr_maxthreads = 1;
1150 netisr_bindthreads = 0;
1153 netisr_start_swi(curcpu, pcpu_find(curcpu));
1155 SYSINIT(netisr_init, SI_SUB_SOFTINTR, SI_ORDER_FIRST, netisr_init, NULL);
1158 * Start worker threads for additional CPUs. No attempt to gracefully handle
1159 * work reassignment, we don't yet support dynamic reconfiguration.
1162 netisr_start(void *arg)
1166 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1167 if (nws_count >= netisr_maxthreads)
1169 /* XXXRW: Is skipping absent CPUs still required here? */
1170 if (CPU_ABSENT(pc->pc_cpuid))
1172 /* Worker will already be present for boot CPU. */
1173 if (pc->pc_netisr != NULL)
1175 netisr_start_swi(pc->pc_cpuid, pc);
1178 SYSINIT(netisr_start, SI_SUB_SMP, SI_ORDER_MIDDLE, netisr_start, NULL);
1181 * Sysctl monitoring for netisr: query a list of registered protocols.
1184 sysctl_netisr_proto(SYSCTL_HANDLER_ARGS)
1186 struct rm_priotracker tracker;
1187 struct sysctl_netisr_proto *snpp, *snp_array;
1188 struct netisr_proto *npp;
1189 u_int counter, proto;
1192 if (req->newptr != NULL)
1194 snp_array = malloc(sizeof(*snp_array) * NETISR_MAXPROT, M_TEMP,
1197 NETISR_RLOCK(&tracker);
1198 for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1199 npp = &netisr_proto[proto];
1200 if (npp->np_name == NULL)
1202 snpp = &snp_array[counter];
1203 snpp->snp_version = sizeof(*snpp);
1204 strlcpy(snpp->snp_name, npp->np_name, NETISR_NAMEMAXLEN);
1205 snpp->snp_proto = proto;
1206 snpp->snp_qlimit = npp->np_qlimit;
1207 snpp->snp_policy = npp->np_policy;
1208 snpp->snp_dispatch = npp->np_dispatch;
1209 if (npp->np_m2flow != NULL)
1210 snpp->snp_flags |= NETISR_SNP_FLAGS_M2FLOW;
1211 if (npp->np_m2cpuid != NULL)
1212 snpp->snp_flags |= NETISR_SNP_FLAGS_M2CPUID;
1213 if (npp->np_drainedcpu != NULL)
1214 snpp->snp_flags |= NETISR_SNP_FLAGS_DRAINEDCPU;
1217 NETISR_RUNLOCK(&tracker);
1218 KASSERT(counter <= NETISR_MAXPROT,
1219 ("sysctl_netisr_proto: counter too big (%d)", counter));
1220 error = SYSCTL_OUT(req, snp_array, sizeof(*snp_array) * counter);
1221 free(snp_array, M_TEMP);
1225 SYSCTL_PROC(_net_isr, OID_AUTO, proto,
1226 CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_proto,
1227 "S,sysctl_netisr_proto",
1228 "Return list of protocols registered with netisr");
1231 * Sysctl monitoring for netisr: query a list of workstreams.
1234 sysctl_netisr_workstream(SYSCTL_HANDLER_ARGS)
1236 struct rm_priotracker tracker;
1237 struct sysctl_netisr_workstream *snwsp, *snws_array;
1238 struct netisr_workstream *nwsp;
1239 u_int counter, cpuid;
1242 if (req->newptr != NULL)
1244 snws_array = malloc(sizeof(*snws_array) * MAXCPU, M_TEMP,
1247 NETISR_RLOCK(&tracker);
1248 CPU_FOREACH(cpuid) {
1249 nwsp = DPCPU_ID_PTR(cpuid, nws);
1250 if (nwsp->nws_intr_event == NULL)
1253 snwsp = &snws_array[counter];
1254 snwsp->snws_version = sizeof(*snwsp);
1257 * For now, we equate workstream IDs and CPU IDs in the
1258 * kernel, but expose them independently to userspace in case
1259 * that assumption changes in the future.
1261 snwsp->snws_wsid = cpuid;
1262 snwsp->snws_cpu = cpuid;
1263 if (nwsp->nws_intr_event != NULL)
1264 snwsp->snws_flags |= NETISR_SNWS_FLAGS_INTR;
1268 NETISR_RUNLOCK(&tracker);
1269 KASSERT(counter <= MAXCPU,
1270 ("sysctl_netisr_workstream: counter too big (%d)", counter));
1271 error = SYSCTL_OUT(req, snws_array, sizeof(*snws_array) * counter);
1272 free(snws_array, M_TEMP);
1276 SYSCTL_PROC(_net_isr, OID_AUTO, workstream,
1277 CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_workstream,
1278 "S,sysctl_netisr_workstream",
1279 "Return list of workstreams implemented by netisr");
1282 * Sysctl monitoring for netisr: query per-protocol data across all
1286 sysctl_netisr_work(SYSCTL_HANDLER_ARGS)
1288 struct rm_priotracker tracker;
1289 struct sysctl_netisr_work *snwp, *snw_array;
1290 struct netisr_workstream *nwsp;
1291 struct netisr_proto *npp;
1292 struct netisr_work *nwp;
1293 u_int counter, cpuid, proto;
1296 if (req->newptr != NULL)
1298 snw_array = malloc(sizeof(*snw_array) * MAXCPU * NETISR_MAXPROT,
1299 M_TEMP, M_ZERO | M_WAITOK);
1301 NETISR_RLOCK(&tracker);
1302 CPU_FOREACH(cpuid) {
1303 nwsp = DPCPU_ID_PTR(cpuid, nws);
1304 if (nwsp->nws_intr_event == NULL)
1307 for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1308 npp = &netisr_proto[proto];
1309 if (npp->np_name == NULL)
1311 nwp = &nwsp->nws_work[proto];
1312 snwp = &snw_array[counter];
1313 snwp->snw_version = sizeof(*snwp);
1314 snwp->snw_wsid = cpuid; /* See comment above. */
1315 snwp->snw_proto = proto;
1316 snwp->snw_len = nwp->nw_len;
1317 snwp->snw_watermark = nwp->nw_watermark;
1318 snwp->snw_dispatched = nwp->nw_dispatched;
1319 snwp->snw_hybrid_dispatched =
1320 nwp->nw_hybrid_dispatched;
1321 snwp->snw_qdrops = nwp->nw_qdrops;
1322 snwp->snw_queued = nwp->nw_queued;
1323 snwp->snw_handled = nwp->nw_handled;
1328 KASSERT(counter <= MAXCPU * NETISR_MAXPROT,
1329 ("sysctl_netisr_work: counter too big (%d)", counter));
1330 NETISR_RUNLOCK(&tracker);
1331 error = SYSCTL_OUT(req, snw_array, sizeof(*snw_array) * counter);
1332 free(snw_array, M_TEMP);
1336 SYSCTL_PROC(_net_isr, OID_AUTO, work,
1337 CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_work,
1338 "S,sysctl_netisr_work",
1339 "Return list of per-workstream, per-protocol work in netisr");
1342 DB_SHOW_COMMAND(netisr, db_show_netisr)
1344 struct netisr_workstream *nwsp;
1345 struct netisr_work *nwp;
1349 db_printf("%3s %6s %5s %5s %5s %8s %8s %8s %8s\n", "CPU", "Proto",
1350 "Len", "WMark", "Max", "Disp", "HDisp", "Drop", "Queue");
1351 CPU_FOREACH(cpuid) {
1352 nwsp = DPCPU_ID_PTR(cpuid, nws);
1353 if (nwsp->nws_intr_event == NULL)
1356 for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1357 if (netisr_proto[proto].np_handler == NULL)
1359 nwp = &nwsp->nws_work[proto];
1361 db_printf("%3d ", cpuid);
1364 db_printf("%3s ", "");
1366 "%6s %5d %5d %5d %8ju %8ju %8ju %8ju\n",
1367 netisr_proto[proto].np_name, nwp->nw_len,
1368 nwp->nw_watermark, nwp->nw_qlimit,
1369 nwp->nw_dispatched, nwp->nw_hybrid_dispatched,
1370 nwp->nw_qdrops, nwp->nw_queued);