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_QUEUED: 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_RW |
153 CTLFLAG_TUN, 0, 0, sysctl_netisr_dispatch_policy, "A",
154 "netisr dispatch policy");
157 * These sysctls were used in previous versions to control and export
158 * dispatch policy state. Now, we provide read-only export via them so that
159 * older netstat binaries work. At some point they can be garbage collected.
161 static int netisr_direct_force;
162 SYSCTL_INT(_net_isr, OID_AUTO, direct_force, CTLFLAG_RD,
163 &netisr_direct_force, 0, "compat: force direct dispatch");
165 static int netisr_direct;
166 SYSCTL_INT(_net_isr, OID_AUTO, direct, CTLFLAG_RD, &netisr_direct, 0,
167 "compat: enable direct dispatch");
170 * Allow the administrator to limit the number of threads (CPUs) to use for
171 * netisr. We don't check netisr_maxthreads before creating the thread for
172 * CPU 0, so in practice we ignore values <= 1. This must be set at boot.
173 * We will create at most one thread per CPU.
175 static int netisr_maxthreads = -1; /* Max number of threads. */
176 TUNABLE_INT("net.isr.maxthreads", &netisr_maxthreads);
177 SYSCTL_INT(_net_isr, OID_AUTO, maxthreads, CTLFLAG_RDTUN,
178 &netisr_maxthreads, 0,
179 "Use at most this many CPUs for netisr processing");
181 static int netisr_bindthreads = 0; /* Bind threads to CPUs. */
182 TUNABLE_INT("net.isr.bindthreads", &netisr_bindthreads);
183 SYSCTL_INT(_net_isr, OID_AUTO, bindthreads, CTLFLAG_RDTUN,
184 &netisr_bindthreads, 0, "Bind netisr threads to CPUs.");
187 * Limit per-workstream mbuf queue limits s to at most net.isr.maxqlimit,
188 * both for initial configuration and later modification using
189 * netisr_setqlimit().
191 #define NETISR_DEFAULT_MAXQLIMIT 10240
192 static u_int netisr_maxqlimit = NETISR_DEFAULT_MAXQLIMIT;
193 TUNABLE_INT("net.isr.maxqlimit", &netisr_maxqlimit);
194 SYSCTL_UINT(_net_isr, OID_AUTO, maxqlimit, CTLFLAG_RDTUN,
195 &netisr_maxqlimit, 0,
196 "Maximum netisr per-protocol, per-CPU queue depth.");
199 * The default per-workstream mbuf queue limit for protocols that don't
200 * initialize the nh_qlimit field of their struct netisr_handler. If this is
201 * set above netisr_maxqlimit, we truncate it to the maximum during boot.
203 #define NETISR_DEFAULT_DEFAULTQLIMIT 256
204 static u_int netisr_defaultqlimit = NETISR_DEFAULT_DEFAULTQLIMIT;
205 TUNABLE_INT("net.isr.defaultqlimit", &netisr_defaultqlimit);
206 SYSCTL_UINT(_net_isr, OID_AUTO, defaultqlimit, CTLFLAG_RDTUN,
207 &netisr_defaultqlimit, 0,
208 "Default netisr per-protocol, per-CPU queue limit if not set by protocol");
211 * Store and export the compile-time constant NETISR_MAXPROT limit on the
212 * number of protocols that can register with netisr at a time. This is
213 * required for crashdump analysis, as it sizes netisr_proto[].
215 static u_int netisr_maxprot = NETISR_MAXPROT;
216 SYSCTL_UINT(_net_isr, OID_AUTO, maxprot, CTLFLAG_RD,
218 "Compile-time limit on the number of protocols supported by netisr.");
221 * The netisr_proto array describes all registered protocols, indexed by
222 * protocol number. See netisr_internal.h for more details.
224 static struct netisr_proto netisr_proto[NETISR_MAXPROT];
227 * Per-CPU workstream data. See netisr_internal.h for more details.
229 DPCPU_DEFINE(struct netisr_workstream, nws);
232 * Map contiguous values between 0 and nws_count into CPU IDs appropriate for
233 * accessing workstreams. This allows constructions of the form
234 * DPCPU_ID_GET(nws_array[arbitraryvalue % nws_count], nws).
236 static u_int nws_array[MAXCPU];
239 * Number of registered workstreams. Will be at most the number of running
240 * CPUs once fully started.
242 static u_int nws_count;
243 SYSCTL_UINT(_net_isr, OID_AUTO, numthreads, CTLFLAG_RD,
244 &nws_count, 0, "Number of extant netisr threads.");
247 * Synchronization for each workstream: a mutex protects all mutable fields
248 * in each stream, including per-protocol state (mbuf queues). The SWI is
249 * woken up if asynchronous dispatch is required.
251 #define NWS_LOCK(s) mtx_lock(&(s)->nws_mtx)
252 #define NWS_LOCK_ASSERT(s) mtx_assert(&(s)->nws_mtx, MA_OWNED)
253 #define NWS_UNLOCK(s) mtx_unlock(&(s)->nws_mtx)
254 #define NWS_SIGNAL(s) swi_sched((s)->nws_swi_cookie, 0)
257 * Utility routines for protocols that implement their own mapping of flows
261 netisr_get_cpucount(void)
268 netisr_get_cpuid(u_int cpunumber)
271 KASSERT(cpunumber < nws_count, ("%s: %u > %u", __func__, cpunumber,
274 return (nws_array[cpunumber]);
278 * The default implementation of flow -> CPU ID mapping.
280 * Non-static so that protocols can use it to map their own work to specific
281 * CPUs in a manner consistent to netisr for affinity purposes.
284 netisr_default_flow2cpu(u_int flowid)
287 return (nws_array[flowid % nws_count]);
291 * Dispatch tunable and sysctl configuration.
293 struct netisr_dispatch_table_entry {
295 const char *ndte_policy_str;
297 static const struct netisr_dispatch_table_entry netisr_dispatch_table[] = {
298 { NETISR_DISPATCH_DEFAULT, "default" },
299 { NETISR_DISPATCH_DEFERRED, "deferred" },
300 { NETISR_DISPATCH_HYBRID, "hybrid" },
301 { NETISR_DISPATCH_DIRECT, "direct" },
303 static const u_int netisr_dispatch_table_len =
304 (sizeof(netisr_dispatch_table) / sizeof(netisr_dispatch_table[0]));
307 netisr_dispatch_policy_to_str(u_int dispatch_policy, char *buffer,
310 const struct netisr_dispatch_table_entry *ndtep;
315 for (i = 0; i < netisr_dispatch_table_len; i++) {
316 ndtep = &netisr_dispatch_table[i];
317 if (ndtep->ndte_policy == dispatch_policy) {
318 str = ndtep->ndte_policy_str;
322 snprintf(buffer, buflen, "%s", str);
326 netisr_dispatch_policy_from_str(const char *str, u_int *dispatch_policyp)
328 const struct netisr_dispatch_table_entry *ndtep;
331 for (i = 0; i < netisr_dispatch_table_len; i++) {
332 ndtep = &netisr_dispatch_table[i];
333 if (strcmp(ndtep->ndte_policy_str, str) == 0) {
334 *dispatch_policyp = ndtep->ndte_policy;
342 netisr_dispatch_policy_compat(void)
345 switch (netisr_dispatch_policy) {
346 case NETISR_DISPATCH_DEFERRED:
347 netisr_direct_force = 0;
351 case NETISR_DISPATCH_HYBRID:
352 netisr_direct_force = 0;
356 case NETISR_DISPATCH_DIRECT:
357 netisr_direct_force = 1;
362 panic("%s: unknown policy %u", __func__,
363 netisr_dispatch_policy);
368 sysctl_netisr_dispatch_policy(SYSCTL_HANDLER_ARGS)
370 char tmp[NETISR_DISPATCH_POLICY_MAXSTR];
371 u_int dispatch_policy;
374 netisr_dispatch_policy_to_str(netisr_dispatch_policy, tmp,
376 error = sysctl_handle_string(oidp, tmp, sizeof(tmp), req);
377 if (error == 0 && req->newptr != NULL) {
378 error = netisr_dispatch_policy_from_str(tmp,
380 if (error == 0 && dispatch_policy == NETISR_DISPATCH_DEFAULT)
383 netisr_dispatch_policy = dispatch_policy;
384 netisr_dispatch_policy_compat();
391 * Register a new netisr handler, which requires initializing per-protocol
392 * fields for each workstream. All netisr work is briefly suspended while
393 * the protocol is installed.
396 netisr_register(const struct netisr_handler *nhp)
398 struct netisr_work *npwp;
402 proto = nhp->nh_proto;
406 * Test that the requested registration is valid.
408 KASSERT(nhp->nh_name != NULL,
409 ("%s: nh_name NULL for %u", __func__, proto));
410 KASSERT(nhp->nh_handler != NULL,
411 ("%s: nh_handler NULL for %s", __func__, name));
412 KASSERT(nhp->nh_policy == NETISR_POLICY_SOURCE ||
413 nhp->nh_policy == NETISR_POLICY_FLOW ||
414 nhp->nh_policy == NETISR_POLICY_CPU,
415 ("%s: unsupported nh_policy %u for %s", __func__,
416 nhp->nh_policy, name));
417 KASSERT(nhp->nh_policy == NETISR_POLICY_FLOW ||
418 nhp->nh_m2flow == NULL,
419 ("%s: nh_policy != FLOW but m2flow defined for %s", __func__,
421 KASSERT(nhp->nh_policy == NETISR_POLICY_CPU || nhp->nh_m2cpuid == NULL,
422 ("%s: nh_policy != CPU but m2cpuid defined for %s", __func__,
424 KASSERT(nhp->nh_policy != NETISR_POLICY_CPU || nhp->nh_m2cpuid != NULL,
425 ("%s: nh_policy == CPU but m2cpuid not defined for %s", __func__,
427 KASSERT(nhp->nh_dispatch == NETISR_DISPATCH_DEFAULT ||
428 nhp->nh_dispatch == NETISR_DISPATCH_DEFERRED ||
429 nhp->nh_dispatch == NETISR_DISPATCH_HYBRID ||
430 nhp->nh_dispatch == NETISR_DISPATCH_DIRECT,
431 ("%s: invalid nh_dispatch (%u)", __func__, nhp->nh_dispatch));
433 KASSERT(proto < NETISR_MAXPROT,
434 ("%s(%u, %s): protocol too big", __func__, proto, name));
437 * Test that no existing registration exists for this protocol.
440 KASSERT(netisr_proto[proto].np_name == NULL,
441 ("%s(%u, %s): name present", __func__, proto, name));
442 KASSERT(netisr_proto[proto].np_handler == NULL,
443 ("%s(%u, %s): handler present", __func__, proto, name));
445 netisr_proto[proto].np_name = name;
446 netisr_proto[proto].np_handler = nhp->nh_handler;
447 netisr_proto[proto].np_m2flow = nhp->nh_m2flow;
448 netisr_proto[proto].np_m2cpuid = nhp->nh_m2cpuid;
449 netisr_proto[proto].np_drainedcpu = nhp->nh_drainedcpu;
450 if (nhp->nh_qlimit == 0)
451 netisr_proto[proto].np_qlimit = netisr_defaultqlimit;
452 else if (nhp->nh_qlimit > netisr_maxqlimit) {
453 printf("%s: %s requested queue limit %u capped to "
454 "net.isr.maxqlimit %u\n", __func__, name, nhp->nh_qlimit,
456 netisr_proto[proto].np_qlimit = netisr_maxqlimit;
458 netisr_proto[proto].np_qlimit = nhp->nh_qlimit;
459 netisr_proto[proto].np_policy = nhp->nh_policy;
460 netisr_proto[proto].np_dispatch = nhp->nh_dispatch;
462 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
463 bzero(npwp, sizeof(*npwp));
464 npwp->nw_qlimit = netisr_proto[proto].np_qlimit;
470 * Clear drop counters across all workstreams for a protocol.
473 netisr_clearqdrops(const struct netisr_handler *nhp)
475 struct netisr_work *npwp;
481 proto = nhp->nh_proto;
485 KASSERT(proto < NETISR_MAXPROT,
486 ("%s(%u): protocol too big for %s", __func__, proto, name));
489 KASSERT(netisr_proto[proto].np_handler != NULL,
490 ("%s(%u): protocol not registered for %s", __func__, proto,
494 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
501 * Query current drop counters across all workstreams for a protocol.
504 netisr_getqdrops(const struct netisr_handler *nhp, u_int64_t *qdropp)
506 struct netisr_work *npwp;
507 struct rm_priotracker tracker;
514 proto = nhp->nh_proto;
518 KASSERT(proto < NETISR_MAXPROT,
519 ("%s(%u): protocol too big for %s", __func__, proto, name));
521 NETISR_RLOCK(&tracker);
522 KASSERT(netisr_proto[proto].np_handler != NULL,
523 ("%s(%u): protocol not registered for %s", __func__, proto,
527 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
528 *qdropp += npwp->nw_qdrops;
530 NETISR_RUNLOCK(&tracker);
534 * Query current per-workstream queue limit for a protocol.
537 netisr_getqlimit(const struct netisr_handler *nhp, u_int *qlimitp)
539 struct rm_priotracker tracker;
545 proto = nhp->nh_proto;
549 KASSERT(proto < NETISR_MAXPROT,
550 ("%s(%u): protocol too big for %s", __func__, proto, name));
552 NETISR_RLOCK(&tracker);
553 KASSERT(netisr_proto[proto].np_handler != NULL,
554 ("%s(%u): protocol not registered for %s", __func__, proto,
556 *qlimitp = netisr_proto[proto].np_qlimit;
557 NETISR_RUNLOCK(&tracker);
561 * Update the queue limit across per-workstream queues for a protocol. We
562 * simply change the limits, and don't drain overflowed packets as they will
563 * (hopefully) take care of themselves shortly.
566 netisr_setqlimit(const struct netisr_handler *nhp, u_int qlimit)
568 struct netisr_work *npwp;
574 if (qlimit > netisr_maxqlimit)
577 proto = nhp->nh_proto;
581 KASSERT(proto < NETISR_MAXPROT,
582 ("%s(%u): protocol too big for %s", __func__, proto, name));
585 KASSERT(netisr_proto[proto].np_handler != NULL,
586 ("%s(%u): protocol not registered for %s", __func__, proto,
589 netisr_proto[proto].np_qlimit = qlimit;
591 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
592 npwp->nw_qlimit = qlimit;
599 * Drain all packets currently held in a particular protocol work queue.
602 netisr_drain_proto(struct netisr_work *npwp)
607 * We would assert the lock on the workstream but it's not passed in.
609 while ((m = npwp->nw_head) != NULL) {
610 npwp->nw_head = m->m_nextpkt;
612 if (npwp->nw_head == NULL)
613 npwp->nw_tail = NULL;
617 KASSERT(npwp->nw_tail == NULL, ("%s: tail", __func__));
618 KASSERT(npwp->nw_len == 0, ("%s: len", __func__));
622 * Remove the registration of a network protocol, which requires clearing
623 * per-protocol fields across all workstreams, including freeing all mbufs in
624 * the queues at time of unregister. All work in netisr is briefly suspended
625 * while this takes place.
628 netisr_unregister(const struct netisr_handler *nhp)
630 struct netisr_work *npwp;
636 proto = nhp->nh_proto;
640 KASSERT(proto < NETISR_MAXPROT,
641 ("%s(%u): protocol too big for %s", __func__, proto, name));
644 KASSERT(netisr_proto[proto].np_handler != NULL,
645 ("%s(%u): protocol not registered for %s", __func__, proto,
648 netisr_proto[proto].np_name = NULL;
649 netisr_proto[proto].np_handler = NULL;
650 netisr_proto[proto].np_m2flow = NULL;
651 netisr_proto[proto].np_m2cpuid = NULL;
652 netisr_proto[proto].np_qlimit = 0;
653 netisr_proto[proto].np_policy = 0;
655 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
656 netisr_drain_proto(npwp);
657 bzero(npwp, sizeof(*npwp));
663 * Compose the global and per-protocol policies on dispatch, and return the
664 * dispatch policy to use.
667 netisr_get_dispatch(struct netisr_proto *npp)
671 * Protocol-specific configuration overrides the global default.
673 if (npp->np_dispatch != NETISR_DISPATCH_DEFAULT)
674 return (npp->np_dispatch);
675 return (netisr_dispatch_policy);
679 * Look up the workstream given a packet and source identifier. Do this by
680 * checking the protocol's policy, and optionally call out to the protocol
681 * for assistance if required.
684 netisr_select_cpuid(struct netisr_proto *npp, u_int dispatch_policy,
685 uintptr_t source, struct mbuf *m, u_int *cpuidp)
690 NETISR_LOCK_ASSERT();
693 * In the event we have only one worker, shortcut and deliver to it
694 * without further ado.
696 if (nws_count == 1) {
697 *cpuidp = nws_array[0];
702 * What happens next depends on the policy selected by the protocol.
703 * If we want to support per-interface policies, we should do that
706 policy = npp->np_policy;
707 if (policy == NETISR_POLICY_CPU) {
708 m = npp->np_m2cpuid(m, source, cpuidp);
713 * It's possible for a protocol not to have a good idea about
714 * where to process a packet, in which case we fall back on
715 * the netisr code to decide. In the hybrid case, return the
716 * current CPU ID, which will force an immediate direct
717 * dispatch. In the queued case, fall back on the SOURCE
720 if (*cpuidp != NETISR_CPUID_NONE)
722 if (dispatch_policy == NETISR_DISPATCH_HYBRID) {
726 policy = NETISR_POLICY_SOURCE;
729 if (policy == NETISR_POLICY_FLOW) {
730 if (!(m->m_flags & M_FLOWID) && npp->np_m2flow != NULL) {
731 m = npp->np_m2flow(m, source);
735 if (m->m_flags & M_FLOWID) {
737 netisr_default_flow2cpu(m->m_pkthdr.flowid);
740 policy = NETISR_POLICY_SOURCE;
743 KASSERT(policy == NETISR_POLICY_SOURCE,
744 ("%s: invalid policy %u for %s", __func__, npp->np_policy,
747 ifp = m->m_pkthdr.rcvif;
749 *cpuidp = nws_array[(ifp->if_index + source) % nws_count];
751 *cpuidp = nws_array[source % nws_count];
756 * Process packets associated with a workstream and protocol. For reasons of
757 * fairness, we process up to one complete netisr queue at a time, moving the
758 * queue to a stack-local queue for processing, but do not loop refreshing
759 * from the global queue. The caller is responsible for deciding whether to
760 * loop, and for setting the NWS_RUNNING flag. The passed workstream will be
761 * locked on entry and relocked before return, but will be released while
762 * processing. The number of packets processed is returned.
765 netisr_process_workstream_proto(struct netisr_workstream *nwsp, u_int proto)
767 struct netisr_work local_npw, *npwp;
771 NETISR_LOCK_ASSERT();
772 NWS_LOCK_ASSERT(nwsp);
774 KASSERT(nwsp->nws_flags & NWS_RUNNING,
775 ("%s(%u): not running", __func__, proto));
776 KASSERT(proto >= 0 && proto < NETISR_MAXPROT,
777 ("%s(%u): invalid proto\n", __func__, proto));
779 npwp = &nwsp->nws_work[proto];
780 if (npwp->nw_len == 0)
784 * Move the global work queue to a thread-local work queue.
786 * Notice that this means the effective maximum length of the queue
787 * is actually twice that of the maximum queue length specified in
788 * the protocol registration call.
790 handled = npwp->nw_len;
792 npwp->nw_head = NULL;
793 npwp->nw_tail = NULL;
795 nwsp->nws_pendingbits &= ~(1 << proto);
797 while ((m = local_npw.nw_head) != NULL) {
798 local_npw.nw_head = m->m_nextpkt;
800 if (local_npw.nw_head == NULL)
801 local_npw.nw_tail = NULL;
803 VNET_ASSERT(m->m_pkthdr.rcvif != NULL,
804 ("%s:%d rcvif == NULL: m=%p", __func__, __LINE__, m));
805 CURVNET_SET(m->m_pkthdr.rcvif->if_vnet);
806 netisr_proto[proto].np_handler(m);
809 KASSERT(local_npw.nw_len == 0,
810 ("%s(%u): len %u", __func__, proto, local_npw.nw_len));
811 if (netisr_proto[proto].np_drainedcpu)
812 netisr_proto[proto].np_drainedcpu(nwsp->nws_cpu);
814 npwp->nw_handled += handled;
819 * SWI handler for netisr -- processes packets in a set of workstreams that
820 * it owns, woken up by calls to NWS_SIGNAL(). If this workstream is already
821 * being direct dispatched, go back to sleep and wait for the dispatching
822 * thread to wake us up again.
827 #ifdef NETISR_LOCKING
828 struct rm_priotracker tracker;
830 struct netisr_workstream *nwsp;
835 #ifdef DEVICE_POLLING
836 KASSERT(nws_count == 1,
837 ("%s: device_polling but nws_count != 1", __func__));
840 #ifdef NETISR_LOCKING
841 NETISR_RLOCK(&tracker);
844 KASSERT(!(nwsp->nws_flags & NWS_RUNNING), ("swi_net: running"));
845 if (nwsp->nws_flags & NWS_DISPATCHING)
847 nwsp->nws_flags |= NWS_RUNNING;
848 nwsp->nws_flags &= ~NWS_SCHEDULED;
849 while ((bits = nwsp->nws_pendingbits) != 0) {
850 while ((prot = ffs(bits)) != 0) {
852 bits &= ~(1 << prot);
853 (void)netisr_process_workstream_proto(nwsp, prot);
856 nwsp->nws_flags &= ~NWS_RUNNING;
859 #ifdef NETISR_LOCKING
860 NETISR_RUNLOCK(&tracker);
862 #ifdef DEVICE_POLLING
868 netisr_queue_workstream(struct netisr_workstream *nwsp, u_int proto,
869 struct netisr_work *npwp, struct mbuf *m, int *dosignalp)
872 NWS_LOCK_ASSERT(nwsp);
875 if (npwp->nw_len < npwp->nw_qlimit) {
877 if (npwp->nw_head == NULL) {
881 npwp->nw_tail->m_nextpkt = m;
885 if (npwp->nw_len > npwp->nw_watermark)
886 npwp->nw_watermark = npwp->nw_len;
889 * We must set the bit regardless of NWS_RUNNING, so that
890 * swi_net() keeps calling netisr_process_workstream_proto().
892 nwsp->nws_pendingbits |= (1 << proto);
893 if (!(nwsp->nws_flags &
894 (NWS_RUNNING | NWS_DISPATCHING | NWS_SCHEDULED))) {
895 nwsp->nws_flags |= NWS_SCHEDULED;
896 *dosignalp = 1; /* Defer until unlocked. */
908 netisr_queue_internal(u_int proto, struct mbuf *m, u_int cpuid)
910 struct netisr_workstream *nwsp;
911 struct netisr_work *npwp;
914 #ifdef NETISR_LOCKING
915 NETISR_LOCK_ASSERT();
917 KASSERT(cpuid <= mp_maxid, ("%s: cpuid too big (%u, %u)", __func__,
919 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
923 nwsp = DPCPU_ID_PTR(cpuid, nws);
924 npwp = &nwsp->nws_work[proto];
926 error = netisr_queue_workstream(nwsp, proto, npwp, m, &dosignal);
934 netisr_queue_src(u_int proto, uintptr_t source, struct mbuf *m)
936 #ifdef NETISR_LOCKING
937 struct rm_priotracker tracker;
942 KASSERT(proto < NETISR_MAXPROT,
943 ("%s: invalid proto %u", __func__, proto));
945 #ifdef NETISR_LOCKING
946 NETISR_RLOCK(&tracker);
948 KASSERT(netisr_proto[proto].np_handler != NULL,
949 ("%s: invalid proto %u", __func__, proto));
951 m = netisr_select_cpuid(&netisr_proto[proto], NETISR_DISPATCH_DEFERRED,
954 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__,
956 error = netisr_queue_internal(proto, m, cpuid);
959 #ifdef NETISR_LOCKING
960 NETISR_RUNLOCK(&tracker);
966 netisr_queue(u_int proto, struct mbuf *m)
969 return (netisr_queue_src(proto, 0, m));
973 * Dispatch a packet for netisr processing; direct dispatch is permitted by
977 netisr_dispatch_src(u_int proto, uintptr_t source, struct mbuf *m)
979 #ifdef NETISR_LOCKING
980 struct rm_priotracker tracker;
982 struct netisr_workstream *nwsp;
983 struct netisr_proto *npp;
984 struct netisr_work *npwp;
986 u_int cpuid, dispatch_policy;
988 KASSERT(proto < NETISR_MAXPROT,
989 ("%s: invalid proto %u", __func__, proto));
990 #ifdef NETISR_LOCKING
991 NETISR_RLOCK(&tracker);
993 npp = &netisr_proto[proto];
994 KASSERT(npp->np_handler != NULL, ("%s: invalid proto %u", __func__,
997 dispatch_policy = netisr_get_dispatch(npp);
998 if (dispatch_policy == NETISR_DISPATCH_DEFERRED)
999 return (netisr_queue_src(proto, source, m));
1002 * If direct dispatch is forced, then unconditionally dispatch
1003 * without a formal CPU selection. Borrow the current CPU's stats,
1004 * even if there's no worker on it. In this case we don't update
1005 * nws_flags because all netisr processing will be source ordered due
1006 * to always being forced to directly dispatch.
1008 if (dispatch_policy == NETISR_DISPATCH_DIRECT) {
1009 nwsp = DPCPU_PTR(nws);
1010 npwp = &nwsp->nws_work[proto];
1011 npwp->nw_dispatched++;
1013 netisr_proto[proto].np_handler(m);
1018 KASSERT(dispatch_policy == NETISR_DISPATCH_HYBRID,
1019 ("%s: unknown dispatch policy (%u)", __func__, dispatch_policy));
1022 * Otherwise, we execute in a hybrid mode where we will try to direct
1023 * dispatch if we're on the right CPU and the netisr worker isn't
1027 m = netisr_select_cpuid(&netisr_proto[proto], NETISR_DISPATCH_HYBRID,
1033 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
1034 if (cpuid != curcpu)
1035 goto queue_fallback;
1036 nwsp = DPCPU_PTR(nws);
1037 npwp = &nwsp->nws_work[proto];
1040 * We are willing to direct dispatch only if three conditions hold:
1042 * (1) The netisr worker isn't already running,
1043 * (2) Another thread isn't already directly dispatching, and
1044 * (3) The netisr hasn't already been woken up.
1047 if (nwsp->nws_flags & (NWS_RUNNING | NWS_DISPATCHING | NWS_SCHEDULED)) {
1048 error = netisr_queue_workstream(nwsp, proto, npwp, m,
1057 * The current thread is now effectively the netisr worker, so set
1058 * the dispatching flag to prevent concurrent processing of the
1059 * stream from another thread (even the netisr worker), which could
1060 * otherwise lead to effective misordering of the stream.
1062 nwsp->nws_flags |= NWS_DISPATCHING;
1064 netisr_proto[proto].np_handler(m);
1066 nwsp->nws_flags &= ~NWS_DISPATCHING;
1068 npwp->nw_hybrid_dispatched++;
1071 * If other work was enqueued by another thread while we were direct
1072 * dispatching, we need to signal the netisr worker to do that work.
1073 * In the future, we might want to do some of that work in the
1074 * current thread, rather than trigger further context switches. If
1075 * so, we'll want to establish a reasonable bound on the work done in
1076 * the "borrowed" context.
1078 if (nwsp->nws_pendingbits != 0) {
1079 nwsp->nws_flags |= NWS_SCHEDULED;
1090 error = netisr_queue_internal(proto, m, cpuid);
1094 #ifdef NETISR_LOCKING
1095 NETISR_RUNLOCK(&tracker);
1101 netisr_dispatch(u_int proto, struct mbuf *m)
1104 return (netisr_dispatch_src(proto, 0, m));
1107 #ifdef DEVICE_POLLING
1109 * Kernel polling borrows a netisr thread to run interface polling in; this
1110 * function allows kernel polling to request that the netisr thread be
1111 * scheduled even if no packets are pending for protocols.
1114 netisr_sched_poll(void)
1116 struct netisr_workstream *nwsp;
1118 nwsp = DPCPU_ID_PTR(nws_array[0], nws);
1124 netisr_start_swi(u_int cpuid, struct pcpu *pc)
1127 struct netisr_workstream *nwsp;
1130 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
1132 nwsp = DPCPU_ID_PTR(cpuid, nws);
1133 mtx_init(&nwsp->nws_mtx, "netisr_mtx", NULL, MTX_DEF);
1134 nwsp->nws_cpu = cpuid;
1135 snprintf(swiname, sizeof(swiname), "netisr %u", cpuid);
1136 error = swi_add(&nwsp->nws_intr_event, swiname, swi_net, nwsp,
1137 SWI_NET, INTR_MPSAFE, &nwsp->nws_swi_cookie);
1139 panic("%s: swi_add %d", __func__, error);
1140 pc->pc_netisr = nwsp->nws_intr_event;
1141 if (netisr_bindthreads) {
1142 error = intr_event_bind(nwsp->nws_intr_event, cpuid);
1144 printf("%s: cpu %u: intr_event_bind: %d", __func__,
1148 nws_array[nws_count] = nwsp->nws_cpu;
1154 * Initialize the netisr subsystem. We rely on BSS and static initialization
1155 * of most fields in global data structures.
1157 * Start a worker thread for the boot CPU so that we can support network
1158 * traffic immediately in case the network stack is used before additional
1159 * CPUs are started (for example, diskless boot).
1162 netisr_init(void *arg)
1164 char tmp[NETISR_DISPATCH_POLICY_MAXSTR];
1165 u_int dispatch_policy;
1168 KASSERT(curcpu == 0, ("%s: not on CPU 0", __func__));
1171 if (netisr_maxthreads < 1)
1172 netisr_maxthreads = 1;
1173 if (netisr_maxthreads > mp_ncpus) {
1174 printf("netisr_init: forcing maxthreads from %d to %d\n",
1175 netisr_maxthreads, mp_ncpus);
1176 netisr_maxthreads = mp_ncpus;
1178 if (netisr_defaultqlimit > netisr_maxqlimit) {
1179 printf("netisr_init: forcing defaultqlimit from %d to %d\n",
1180 netisr_defaultqlimit, netisr_maxqlimit);
1181 netisr_defaultqlimit = netisr_maxqlimit;
1183 #ifdef DEVICE_POLLING
1185 * The device polling code is not yet aware of how to deal with
1186 * multiple netisr threads, so for the time being compiling in device
1187 * polling disables parallel netisr workers.
1189 if (netisr_maxthreads != 1 || netisr_bindthreads != 0) {
1190 printf("netisr_init: forcing maxthreads to 1 and "
1191 "bindthreads to 0 for device polling\n");
1192 netisr_maxthreads = 1;
1193 netisr_bindthreads = 0;
1197 if (TUNABLE_STR_FETCH("net.isr.dispatch", tmp, sizeof(tmp))) {
1198 error = netisr_dispatch_policy_from_str(tmp,
1200 if (error == 0 && dispatch_policy == NETISR_DISPATCH_DEFAULT)
1203 netisr_dispatch_policy = dispatch_policy;
1204 netisr_dispatch_policy_compat();
1207 "%s: invalid dispatch policy %s, using default\n",
1211 netisr_start_swi(curcpu, pcpu_find(curcpu));
1213 SYSINIT(netisr_init, SI_SUB_SOFTINTR, SI_ORDER_FIRST, netisr_init, NULL);
1216 * Start worker threads for additional CPUs. No attempt to gracefully handle
1217 * work reassignment, we don't yet support dynamic reconfiguration.
1220 netisr_start(void *arg)
1224 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1225 if (nws_count >= netisr_maxthreads)
1227 /* XXXRW: Is skipping absent CPUs still required here? */
1228 if (CPU_ABSENT(pc->pc_cpuid))
1230 /* Worker will already be present for boot CPU. */
1231 if (pc->pc_netisr != NULL)
1233 netisr_start_swi(pc->pc_cpuid, pc);
1236 SYSINIT(netisr_start, SI_SUB_SMP, SI_ORDER_MIDDLE, netisr_start, NULL);
1239 * Sysctl monitoring for netisr: query a list of registered protocols.
1242 sysctl_netisr_proto(SYSCTL_HANDLER_ARGS)
1244 struct rm_priotracker tracker;
1245 struct sysctl_netisr_proto *snpp, *snp_array;
1246 struct netisr_proto *npp;
1247 u_int counter, proto;
1250 if (req->newptr != NULL)
1252 snp_array = malloc(sizeof(*snp_array) * NETISR_MAXPROT, M_TEMP,
1255 NETISR_RLOCK(&tracker);
1256 for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1257 npp = &netisr_proto[proto];
1258 if (npp->np_name == NULL)
1260 snpp = &snp_array[counter];
1261 snpp->snp_version = sizeof(*snpp);
1262 strlcpy(snpp->snp_name, npp->np_name, NETISR_NAMEMAXLEN);
1263 snpp->snp_proto = proto;
1264 snpp->snp_qlimit = npp->np_qlimit;
1265 snpp->snp_policy = npp->np_policy;
1266 snpp->snp_dispatch = npp->np_dispatch;
1267 if (npp->np_m2flow != NULL)
1268 snpp->snp_flags |= NETISR_SNP_FLAGS_M2FLOW;
1269 if (npp->np_m2cpuid != NULL)
1270 snpp->snp_flags |= NETISR_SNP_FLAGS_M2CPUID;
1271 if (npp->np_drainedcpu != NULL)
1272 snpp->snp_flags |= NETISR_SNP_FLAGS_DRAINEDCPU;
1275 NETISR_RUNLOCK(&tracker);
1276 KASSERT(counter <= NETISR_MAXPROT,
1277 ("sysctl_netisr_proto: counter too big (%d)", counter));
1278 error = SYSCTL_OUT(req, snp_array, sizeof(*snp_array) * counter);
1279 free(snp_array, M_TEMP);
1283 SYSCTL_PROC(_net_isr, OID_AUTO, proto,
1284 CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_proto,
1285 "S,sysctl_netisr_proto",
1286 "Return list of protocols registered with netisr");
1289 * Sysctl monitoring for netisr: query a list of workstreams.
1292 sysctl_netisr_workstream(SYSCTL_HANDLER_ARGS)
1294 struct rm_priotracker tracker;
1295 struct sysctl_netisr_workstream *snwsp, *snws_array;
1296 struct netisr_workstream *nwsp;
1297 u_int counter, cpuid;
1300 if (req->newptr != NULL)
1302 snws_array = malloc(sizeof(*snws_array) * MAXCPU, M_TEMP,
1305 NETISR_RLOCK(&tracker);
1306 CPU_FOREACH(cpuid) {
1307 nwsp = DPCPU_ID_PTR(cpuid, nws);
1308 if (nwsp->nws_intr_event == NULL)
1311 snwsp = &snws_array[counter];
1312 snwsp->snws_version = sizeof(*snwsp);
1315 * For now, we equate workstream IDs and CPU IDs in the
1316 * kernel, but expose them independently to userspace in case
1317 * that assumption changes in the future.
1319 snwsp->snws_wsid = cpuid;
1320 snwsp->snws_cpu = cpuid;
1321 if (nwsp->nws_intr_event != NULL)
1322 snwsp->snws_flags |= NETISR_SNWS_FLAGS_INTR;
1326 NETISR_RUNLOCK(&tracker);
1327 KASSERT(counter <= MAXCPU,
1328 ("sysctl_netisr_workstream: counter too big (%d)", counter));
1329 error = SYSCTL_OUT(req, snws_array, sizeof(*snws_array) * counter);
1330 free(snws_array, M_TEMP);
1334 SYSCTL_PROC(_net_isr, OID_AUTO, workstream,
1335 CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_workstream,
1336 "S,sysctl_netisr_workstream",
1337 "Return list of workstreams implemented by netisr");
1340 * Sysctl monitoring for netisr: query per-protocol data across all
1344 sysctl_netisr_work(SYSCTL_HANDLER_ARGS)
1346 struct rm_priotracker tracker;
1347 struct sysctl_netisr_work *snwp, *snw_array;
1348 struct netisr_workstream *nwsp;
1349 struct netisr_proto *npp;
1350 struct netisr_work *nwp;
1351 u_int counter, cpuid, proto;
1354 if (req->newptr != NULL)
1356 snw_array = malloc(sizeof(*snw_array) * MAXCPU * NETISR_MAXPROT,
1357 M_TEMP, M_ZERO | M_WAITOK);
1359 NETISR_RLOCK(&tracker);
1360 CPU_FOREACH(cpuid) {
1361 nwsp = DPCPU_ID_PTR(cpuid, nws);
1362 if (nwsp->nws_intr_event == NULL)
1365 for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1366 npp = &netisr_proto[proto];
1367 if (npp->np_name == NULL)
1369 nwp = &nwsp->nws_work[proto];
1370 snwp = &snw_array[counter];
1371 snwp->snw_version = sizeof(*snwp);
1372 snwp->snw_wsid = cpuid; /* See comment above. */
1373 snwp->snw_proto = proto;
1374 snwp->snw_len = nwp->nw_len;
1375 snwp->snw_watermark = nwp->nw_watermark;
1376 snwp->snw_dispatched = nwp->nw_dispatched;
1377 snwp->snw_hybrid_dispatched =
1378 nwp->nw_hybrid_dispatched;
1379 snwp->snw_qdrops = nwp->nw_qdrops;
1380 snwp->snw_queued = nwp->nw_queued;
1381 snwp->snw_handled = nwp->nw_handled;
1386 KASSERT(counter <= MAXCPU * NETISR_MAXPROT,
1387 ("sysctl_netisr_work: counter too big (%d)", counter));
1388 NETISR_RUNLOCK(&tracker);
1389 error = SYSCTL_OUT(req, snw_array, sizeof(*snw_array) * counter);
1390 free(snw_array, M_TEMP);
1394 SYSCTL_PROC(_net_isr, OID_AUTO, work,
1395 CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_work,
1396 "S,sysctl_netisr_work",
1397 "Return list of per-workstream, per-protocol work in netisr");
1400 DB_SHOW_COMMAND(netisr, db_show_netisr)
1402 struct netisr_workstream *nwsp;
1403 struct netisr_work *nwp;
1407 db_printf("%3s %6s %5s %5s %5s %8s %8s %8s %8s\n", "CPU", "Proto",
1408 "Len", "WMark", "Max", "Disp", "HDisp", "Drop", "Queue");
1409 CPU_FOREACH(cpuid) {
1410 nwsp = DPCPU_ID_PTR(cpuid, nws);
1411 if (nwsp->nws_intr_event == NULL)
1414 for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1415 if (netisr_proto[proto].np_handler == NULL)
1417 nwp = &nwsp->nws_work[proto];
1419 db_printf("%3d ", cpuid);
1422 db_printf("%3s ", "");
1424 "%6s %5d %5d %5d %8ju %8ju %8ju %8ju\n",
1425 netisr_proto[proto].np_name, nwp->nw_len,
1426 nwp->nw_watermark, nwp->nw_qlimit,
1427 nwp->nw_dispatched, nwp->nw_hybrid_dispatched,
1428 nwp->nw_qdrops, nwp->nw_queued);