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 * 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, so in practice we ignore values <= 1. This must be set at boot.
160 * We will create at most one thread per CPU.
162 static int netisr_maxthreads = -1; /* Max number of threads. */
163 TUNABLE_INT("net.isr.maxthreads", &netisr_maxthreads);
164 SYSCTL_INT(_net_isr, OID_AUTO, maxthreads, CTLFLAG_RDTUN,
165 &netisr_maxthreads, 0,
166 "Use at most this many CPUs for netisr processing");
168 static int netisr_bindthreads = 0; /* Bind threads to CPUs. */
169 TUNABLE_INT("net.isr.bindthreads", &netisr_bindthreads);
170 SYSCTL_INT(_net_isr, OID_AUTO, bindthreads, CTLFLAG_RDTUN,
171 &netisr_bindthreads, 0, "Bind netisr threads to CPUs.");
174 * Limit per-workstream mbuf queue limits s to at most net.isr.maxqlimit,
175 * both for initial configuration and later modification using
176 * netisr_setqlimit().
178 #define NETISR_DEFAULT_MAXQLIMIT 10240
179 static u_int netisr_maxqlimit = NETISR_DEFAULT_MAXQLIMIT;
180 TUNABLE_INT("net.isr.maxqlimit", &netisr_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 TUNABLE_INT("net.isr.defaultqlimit", &netisr_defaultqlimit);
193 SYSCTL_UINT(_net_isr, OID_AUTO, defaultqlimit, CTLFLAG_RDTUN,
194 &netisr_defaultqlimit, 0,
195 "Default netisr per-protocol, per-CPU queue limit if not set by protocol");
198 * Store and export the compile-time constant NETISR_MAXPROT limit on the
199 * number of protocols that can register with netisr at a time. This is
200 * required for crashdump analysis, as it sizes netisr_proto[].
202 static u_int netisr_maxprot = NETISR_MAXPROT;
203 SYSCTL_UINT(_net_isr, OID_AUTO, maxprot, CTLFLAG_RD,
205 "Compile-time limit on the number of protocols supported by netisr.");
208 * The netisr_proto array describes all registered protocols, indexed by
209 * protocol number. See netisr_internal.h for more details.
211 static struct netisr_proto netisr_proto[NETISR_MAXPROT];
214 * Per-CPU workstream data. See netisr_internal.h for more details.
216 DPCPU_DEFINE(struct netisr_workstream, nws);
219 * Map contiguous values between 0 and nws_count into CPU IDs appropriate for
220 * accessing workstreams. This allows constructions of the form
221 * DPCPU_ID_GET(nws_array[arbitraryvalue % nws_count], nws).
223 static u_int nws_array[MAXCPU];
226 * Number of registered workstreams. Will be at most the number of running
227 * CPUs once fully started.
229 static u_int nws_count;
230 SYSCTL_UINT(_net_isr, OID_AUTO, numthreads, CTLFLAG_RD,
231 &nws_count, 0, "Number of extant netisr threads.");
234 * Synchronization for each workstream: a mutex protects all mutable fields
235 * in each stream, including per-protocol state (mbuf queues). The SWI is
236 * woken up if asynchronous dispatch is required.
238 #define NWS_LOCK(s) mtx_lock(&(s)->nws_mtx)
239 #define NWS_LOCK_ASSERT(s) mtx_assert(&(s)->nws_mtx, MA_OWNED)
240 #define NWS_UNLOCK(s) mtx_unlock(&(s)->nws_mtx)
241 #define NWS_SIGNAL(s) swi_sched((s)->nws_swi_cookie, 0)
244 * Utility routines for protocols that implement their own mapping of flows
248 netisr_get_cpucount(void)
255 netisr_get_cpuid(u_int cpunumber)
258 KASSERT(cpunumber < nws_count, ("%s: %u > %u", __func__, cpunumber,
261 return (nws_array[cpunumber]);
265 * The default implementation of flow -> CPU ID mapping.
267 * Non-static so that protocols can use it to map their own work to specific
268 * CPUs in a manner consistent to netisr for affinity purposes.
271 netisr_default_flow2cpu(u_int flowid)
274 return (nws_array[flowid % nws_count]);
278 * Dispatch tunable and sysctl configuration.
280 struct netisr_dispatch_table_entry {
282 const char *ndte_policy_str;
284 static const struct netisr_dispatch_table_entry netisr_dispatch_table[] = {
285 { NETISR_DISPATCH_DEFAULT, "default" },
286 { NETISR_DISPATCH_DEFERRED, "deferred" },
287 { NETISR_DISPATCH_HYBRID, "hybrid" },
288 { NETISR_DISPATCH_DIRECT, "direct" },
290 static const u_int netisr_dispatch_table_len =
291 (sizeof(netisr_dispatch_table) / sizeof(netisr_dispatch_table[0]));
294 netisr_dispatch_policy_to_str(u_int dispatch_policy, char *buffer,
297 const struct netisr_dispatch_table_entry *ndtep;
302 for (i = 0; i < netisr_dispatch_table_len; i++) {
303 ndtep = &netisr_dispatch_table[i];
304 if (ndtep->ndte_policy == dispatch_policy) {
305 str = ndtep->ndte_policy_str;
309 snprintf(buffer, buflen, "%s", str);
313 netisr_dispatch_policy_from_str(const char *str, u_int *dispatch_policyp)
315 const struct netisr_dispatch_table_entry *ndtep;
318 for (i = 0; i < netisr_dispatch_table_len; i++) {
319 ndtep = &netisr_dispatch_table[i];
320 if (strcmp(ndtep->ndte_policy_str, str) == 0) {
321 *dispatch_policyp = ndtep->ndte_policy;
329 sysctl_netisr_dispatch_policy(SYSCTL_HANDLER_ARGS)
331 char tmp[NETISR_DISPATCH_POLICY_MAXSTR];
332 u_int dispatch_policy;
335 netisr_dispatch_policy_to_str(netisr_dispatch_policy, tmp,
337 error = sysctl_handle_string(oidp, tmp, sizeof(tmp), req);
338 if (error == 0 && req->newptr != NULL) {
339 error = netisr_dispatch_policy_from_str(tmp,
341 if (error == 0 && dispatch_policy == NETISR_DISPATCH_DEFAULT)
344 netisr_dispatch_policy = dispatch_policy;
350 * Register a new netisr handler, which requires initializing per-protocol
351 * fields for each workstream. All netisr work is briefly suspended while
352 * the protocol is installed.
355 netisr_register(const struct netisr_handler *nhp)
357 struct netisr_work *npwp;
361 proto = nhp->nh_proto;
365 * Test that the requested registration is valid.
367 KASSERT(nhp->nh_name != NULL,
368 ("%s: nh_name NULL for %u", __func__, proto));
369 KASSERT(nhp->nh_handler != NULL,
370 ("%s: nh_handler NULL for %s", __func__, name));
371 KASSERT(nhp->nh_policy == NETISR_POLICY_SOURCE ||
372 nhp->nh_policy == NETISR_POLICY_FLOW ||
373 nhp->nh_policy == NETISR_POLICY_CPU,
374 ("%s: unsupported nh_policy %u for %s", __func__,
375 nhp->nh_policy, name));
376 KASSERT(nhp->nh_policy == NETISR_POLICY_FLOW ||
377 nhp->nh_m2flow == NULL,
378 ("%s: nh_policy != FLOW but m2flow defined for %s", __func__,
380 KASSERT(nhp->nh_policy == NETISR_POLICY_CPU || nhp->nh_m2cpuid == NULL,
381 ("%s: nh_policy != CPU but m2cpuid defined for %s", __func__,
383 KASSERT(nhp->nh_policy != NETISR_POLICY_CPU || nhp->nh_m2cpuid != NULL,
384 ("%s: nh_policy == CPU but m2cpuid not defined for %s", __func__,
386 KASSERT(nhp->nh_dispatch == NETISR_DISPATCH_DEFAULT ||
387 nhp->nh_dispatch == NETISR_DISPATCH_DEFERRED ||
388 nhp->nh_dispatch == NETISR_DISPATCH_HYBRID ||
389 nhp->nh_dispatch == NETISR_DISPATCH_DIRECT,
390 ("%s: invalid nh_dispatch (%u)", __func__, nhp->nh_dispatch));
392 KASSERT(proto < NETISR_MAXPROT,
393 ("%s(%u, %s): protocol too big", __func__, proto, name));
396 * Test that no existing registration exists for this protocol.
399 KASSERT(netisr_proto[proto].np_name == NULL,
400 ("%s(%u, %s): name present", __func__, proto, name));
401 KASSERT(netisr_proto[proto].np_handler == NULL,
402 ("%s(%u, %s): handler present", __func__, proto, name));
404 netisr_proto[proto].np_name = name;
405 netisr_proto[proto].np_handler = nhp->nh_handler;
406 netisr_proto[proto].np_m2flow = nhp->nh_m2flow;
407 netisr_proto[proto].np_m2cpuid = nhp->nh_m2cpuid;
408 netisr_proto[proto].np_drainedcpu = nhp->nh_drainedcpu;
409 if (nhp->nh_qlimit == 0)
410 netisr_proto[proto].np_qlimit = netisr_defaultqlimit;
411 else if (nhp->nh_qlimit > netisr_maxqlimit) {
412 printf("%s: %s requested queue limit %u capped to "
413 "net.isr.maxqlimit %u\n", __func__, name, nhp->nh_qlimit,
415 netisr_proto[proto].np_qlimit = netisr_maxqlimit;
417 netisr_proto[proto].np_qlimit = nhp->nh_qlimit;
418 netisr_proto[proto].np_policy = nhp->nh_policy;
419 netisr_proto[proto].np_dispatch = nhp->nh_dispatch;
421 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
422 bzero(npwp, sizeof(*npwp));
423 npwp->nw_qlimit = netisr_proto[proto].np_qlimit;
429 * Clear drop counters across all workstreams for a protocol.
432 netisr_clearqdrops(const struct netisr_handler *nhp)
434 struct netisr_work *npwp;
440 proto = nhp->nh_proto;
444 KASSERT(proto < NETISR_MAXPROT,
445 ("%s(%u): protocol too big for %s", __func__, proto, name));
448 KASSERT(netisr_proto[proto].np_handler != NULL,
449 ("%s(%u): protocol not registered for %s", __func__, proto,
453 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
460 * Query current drop counters across all workstreams for a protocol.
463 netisr_getqdrops(const struct netisr_handler *nhp, u_int64_t *qdropp)
465 struct netisr_work *npwp;
466 struct rm_priotracker tracker;
473 proto = nhp->nh_proto;
477 KASSERT(proto < NETISR_MAXPROT,
478 ("%s(%u): protocol too big for %s", __func__, proto, name));
480 NETISR_RLOCK(&tracker);
481 KASSERT(netisr_proto[proto].np_handler != NULL,
482 ("%s(%u): protocol not registered for %s", __func__, proto,
486 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
487 *qdropp += npwp->nw_qdrops;
489 NETISR_RUNLOCK(&tracker);
493 * Query current per-workstream queue limit for a protocol.
496 netisr_getqlimit(const struct netisr_handler *nhp, u_int *qlimitp)
498 struct rm_priotracker tracker;
504 proto = nhp->nh_proto;
508 KASSERT(proto < NETISR_MAXPROT,
509 ("%s(%u): protocol too big for %s", __func__, proto, name));
511 NETISR_RLOCK(&tracker);
512 KASSERT(netisr_proto[proto].np_handler != NULL,
513 ("%s(%u): protocol not registered for %s", __func__, proto,
515 *qlimitp = netisr_proto[proto].np_qlimit;
516 NETISR_RUNLOCK(&tracker);
520 * Update the queue limit across per-workstream queues for a protocol. We
521 * simply change the limits, and don't drain overflowed packets as they will
522 * (hopefully) take care of themselves shortly.
525 netisr_setqlimit(const struct netisr_handler *nhp, u_int qlimit)
527 struct netisr_work *npwp;
533 if (qlimit > netisr_maxqlimit)
536 proto = nhp->nh_proto;
540 KASSERT(proto < NETISR_MAXPROT,
541 ("%s(%u): protocol too big for %s", __func__, proto, name));
544 KASSERT(netisr_proto[proto].np_handler != NULL,
545 ("%s(%u): protocol not registered for %s", __func__, proto,
548 netisr_proto[proto].np_qlimit = qlimit;
550 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
551 npwp->nw_qlimit = qlimit;
558 * Drain all packets currently held in a particular protocol work queue.
561 netisr_drain_proto(struct netisr_work *npwp)
566 * We would assert the lock on the workstream but it's not passed in.
568 while ((m = npwp->nw_head) != NULL) {
569 npwp->nw_head = m->m_nextpkt;
571 if (npwp->nw_head == NULL)
572 npwp->nw_tail = NULL;
576 KASSERT(npwp->nw_tail == NULL, ("%s: tail", __func__));
577 KASSERT(npwp->nw_len == 0, ("%s: len", __func__));
581 * Remove the registration of a network protocol, which requires clearing
582 * per-protocol fields across all workstreams, including freeing all mbufs in
583 * the queues at time of unregister. All work in netisr is briefly suspended
584 * while this takes place.
587 netisr_unregister(const struct netisr_handler *nhp)
589 struct netisr_work *npwp;
595 proto = nhp->nh_proto;
599 KASSERT(proto < NETISR_MAXPROT,
600 ("%s(%u): protocol too big for %s", __func__, proto, name));
603 KASSERT(netisr_proto[proto].np_handler != NULL,
604 ("%s(%u): protocol not registered for %s", __func__, proto,
607 netisr_proto[proto].np_name = NULL;
608 netisr_proto[proto].np_handler = NULL;
609 netisr_proto[proto].np_m2flow = NULL;
610 netisr_proto[proto].np_m2cpuid = NULL;
611 netisr_proto[proto].np_qlimit = 0;
612 netisr_proto[proto].np_policy = 0;
614 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
615 netisr_drain_proto(npwp);
616 bzero(npwp, sizeof(*npwp));
622 * Compose the global and per-protocol policies on dispatch, and return the
623 * dispatch policy to use.
626 netisr_get_dispatch(struct netisr_proto *npp)
630 * Protocol-specific configuration overrides the global default.
632 if (npp->np_dispatch != NETISR_DISPATCH_DEFAULT)
633 return (npp->np_dispatch);
634 return (netisr_dispatch_policy);
638 * Look up the workstream given a packet and source identifier. Do this by
639 * checking the protocol's policy, and optionally call out to the protocol
640 * for assistance if required.
643 netisr_select_cpuid(struct netisr_proto *npp, u_int dispatch_policy,
644 uintptr_t source, struct mbuf *m, u_int *cpuidp)
649 NETISR_LOCK_ASSERT();
652 * In the event we have only one worker, shortcut and deliver to it
653 * without further ado.
655 if (nws_count == 1) {
656 *cpuidp = nws_array[0];
661 * What happens next depends on the policy selected by the protocol.
662 * If we want to support per-interface policies, we should do that
665 policy = npp->np_policy;
666 if (policy == NETISR_POLICY_CPU) {
667 m = npp->np_m2cpuid(m, source, cpuidp);
672 * It's possible for a protocol not to have a good idea about
673 * where to process a packet, in which case we fall back on
674 * the netisr code to decide. In the hybrid case, return the
675 * current CPU ID, which will force an immediate direct
676 * dispatch. In the queued case, fall back on the SOURCE
679 if (*cpuidp != NETISR_CPUID_NONE)
681 if (dispatch_policy == NETISR_DISPATCH_HYBRID) {
685 policy = NETISR_POLICY_SOURCE;
688 if (policy == NETISR_POLICY_FLOW) {
689 if (M_HASHTYPE_GET(m) == M_HASHTYPE_NONE &&
690 npp->np_m2flow != NULL) {
691 m = npp->np_m2flow(m, source);
695 if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
697 netisr_default_flow2cpu(m->m_pkthdr.flowid);
700 policy = NETISR_POLICY_SOURCE;
703 KASSERT(policy == NETISR_POLICY_SOURCE,
704 ("%s: invalid policy %u for %s", __func__, npp->np_policy,
707 ifp = m->m_pkthdr.rcvif;
709 *cpuidp = nws_array[(ifp->if_index + source) % nws_count];
711 *cpuidp = nws_array[source % nws_count];
716 * Process packets associated with a workstream and protocol. For reasons of
717 * fairness, we process up to one complete netisr queue at a time, moving the
718 * queue to a stack-local queue for processing, but do not loop refreshing
719 * from the global queue. The caller is responsible for deciding whether to
720 * loop, and for setting the NWS_RUNNING flag. The passed workstream will be
721 * locked on entry and relocked before return, but will be released while
722 * processing. The number of packets processed is returned.
725 netisr_process_workstream_proto(struct netisr_workstream *nwsp, u_int proto)
727 struct netisr_work local_npw, *npwp;
731 NETISR_LOCK_ASSERT();
732 NWS_LOCK_ASSERT(nwsp);
734 KASSERT(nwsp->nws_flags & NWS_RUNNING,
735 ("%s(%u): not running", __func__, proto));
736 KASSERT(proto >= 0 && proto < NETISR_MAXPROT,
737 ("%s(%u): invalid proto\n", __func__, proto));
739 npwp = &nwsp->nws_work[proto];
740 if (npwp->nw_len == 0)
744 * Move the global work queue to a thread-local work queue.
746 * Notice that this means the effective maximum length of the queue
747 * is actually twice that of the maximum queue length specified in
748 * the protocol registration call.
750 handled = npwp->nw_len;
752 npwp->nw_head = NULL;
753 npwp->nw_tail = NULL;
755 nwsp->nws_pendingbits &= ~(1 << proto);
757 while ((m = local_npw.nw_head) != NULL) {
758 local_npw.nw_head = m->m_nextpkt;
760 if (local_npw.nw_head == NULL)
761 local_npw.nw_tail = NULL;
763 VNET_ASSERT(m->m_pkthdr.rcvif != NULL,
764 ("%s:%d rcvif == NULL: m=%p", __func__, __LINE__, m));
765 CURVNET_SET(m->m_pkthdr.rcvif->if_vnet);
766 netisr_proto[proto].np_handler(m);
769 KASSERT(local_npw.nw_len == 0,
770 ("%s(%u): len %u", __func__, proto, local_npw.nw_len));
771 if (netisr_proto[proto].np_drainedcpu)
772 netisr_proto[proto].np_drainedcpu(nwsp->nws_cpu);
774 npwp->nw_handled += handled;
779 * SWI handler for netisr -- processes packets in a set of workstreams that
780 * it owns, woken up by calls to NWS_SIGNAL(). If this workstream is already
781 * being direct dispatched, go back to sleep and wait for the dispatching
782 * thread to wake us up again.
787 #ifdef NETISR_LOCKING
788 struct rm_priotracker tracker;
790 struct netisr_workstream *nwsp;
795 #ifdef DEVICE_POLLING
796 KASSERT(nws_count == 1,
797 ("%s: device_polling but nws_count != 1", __func__));
800 #ifdef NETISR_LOCKING
801 NETISR_RLOCK(&tracker);
804 KASSERT(!(nwsp->nws_flags & NWS_RUNNING), ("swi_net: running"));
805 if (nwsp->nws_flags & NWS_DISPATCHING)
807 nwsp->nws_flags |= NWS_RUNNING;
808 nwsp->nws_flags &= ~NWS_SCHEDULED;
809 while ((bits = nwsp->nws_pendingbits) != 0) {
810 while ((prot = ffs(bits)) != 0) {
812 bits &= ~(1 << prot);
813 (void)netisr_process_workstream_proto(nwsp, prot);
816 nwsp->nws_flags &= ~NWS_RUNNING;
819 #ifdef NETISR_LOCKING
820 NETISR_RUNLOCK(&tracker);
822 #ifdef DEVICE_POLLING
828 netisr_queue_workstream(struct netisr_workstream *nwsp, u_int proto,
829 struct netisr_work *npwp, struct mbuf *m, int *dosignalp)
832 NWS_LOCK_ASSERT(nwsp);
835 if (npwp->nw_len < npwp->nw_qlimit) {
837 if (npwp->nw_head == NULL) {
841 npwp->nw_tail->m_nextpkt = m;
845 if (npwp->nw_len > npwp->nw_watermark)
846 npwp->nw_watermark = npwp->nw_len;
849 * We must set the bit regardless of NWS_RUNNING, so that
850 * swi_net() keeps calling netisr_process_workstream_proto().
852 nwsp->nws_pendingbits |= (1 << proto);
853 if (!(nwsp->nws_flags &
854 (NWS_RUNNING | NWS_DISPATCHING | NWS_SCHEDULED))) {
855 nwsp->nws_flags |= NWS_SCHEDULED;
856 *dosignalp = 1; /* Defer until unlocked. */
868 netisr_queue_internal(u_int proto, struct mbuf *m, u_int cpuid)
870 struct netisr_workstream *nwsp;
871 struct netisr_work *npwp;
874 #ifdef NETISR_LOCKING
875 NETISR_LOCK_ASSERT();
877 KASSERT(cpuid <= mp_maxid, ("%s: cpuid too big (%u, %u)", __func__,
879 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
883 nwsp = DPCPU_ID_PTR(cpuid, nws);
884 npwp = &nwsp->nws_work[proto];
886 error = netisr_queue_workstream(nwsp, proto, npwp, m, &dosignal);
894 netisr_queue_src(u_int proto, uintptr_t source, struct mbuf *m)
896 #ifdef NETISR_LOCKING
897 struct rm_priotracker tracker;
902 KASSERT(proto < NETISR_MAXPROT,
903 ("%s: invalid proto %u", __func__, proto));
905 #ifdef NETISR_LOCKING
906 NETISR_RLOCK(&tracker);
908 KASSERT(netisr_proto[proto].np_handler != NULL,
909 ("%s: invalid proto %u", __func__, proto));
911 m = netisr_select_cpuid(&netisr_proto[proto], NETISR_DISPATCH_DEFERRED,
914 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__,
916 error = netisr_queue_internal(proto, m, cpuid);
919 #ifdef NETISR_LOCKING
920 NETISR_RUNLOCK(&tracker);
926 netisr_queue(u_int proto, struct mbuf *m)
929 return (netisr_queue_src(proto, 0, m));
933 * Dispatch a packet for netisr processing; direct dispatch is permitted by
937 netisr_dispatch_src(u_int proto, uintptr_t source, struct mbuf *m)
939 #ifdef NETISR_LOCKING
940 struct rm_priotracker tracker;
942 struct netisr_workstream *nwsp;
943 struct netisr_proto *npp;
944 struct netisr_work *npwp;
946 u_int cpuid, dispatch_policy;
948 KASSERT(proto < NETISR_MAXPROT,
949 ("%s: invalid proto %u", __func__, proto));
950 #ifdef NETISR_LOCKING
951 NETISR_RLOCK(&tracker);
953 npp = &netisr_proto[proto];
954 KASSERT(npp->np_handler != NULL, ("%s: invalid proto %u", __func__,
957 dispatch_policy = netisr_get_dispatch(npp);
958 if (dispatch_policy == NETISR_DISPATCH_DEFERRED)
959 return (netisr_queue_src(proto, source, m));
962 * If direct dispatch is forced, then unconditionally dispatch
963 * without a formal CPU selection. Borrow the current CPU's stats,
964 * even if there's no worker on it. In this case we don't update
965 * nws_flags because all netisr processing will be source ordered due
966 * to always being forced to directly dispatch.
968 if (dispatch_policy == NETISR_DISPATCH_DIRECT) {
969 nwsp = DPCPU_PTR(nws);
970 npwp = &nwsp->nws_work[proto];
971 npwp->nw_dispatched++;
973 netisr_proto[proto].np_handler(m);
978 KASSERT(dispatch_policy == NETISR_DISPATCH_HYBRID,
979 ("%s: unknown dispatch policy (%u)", __func__, dispatch_policy));
982 * Otherwise, we execute in a hybrid mode where we will try to direct
983 * dispatch if we're on the right CPU and the netisr worker isn't
987 m = netisr_select_cpuid(&netisr_proto[proto], NETISR_DISPATCH_HYBRID,
993 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
996 nwsp = DPCPU_PTR(nws);
997 npwp = &nwsp->nws_work[proto];
1000 * We are willing to direct dispatch only if three conditions hold:
1002 * (1) The netisr worker isn't already running,
1003 * (2) Another thread isn't already directly dispatching, and
1004 * (3) The netisr hasn't already been woken up.
1007 if (nwsp->nws_flags & (NWS_RUNNING | NWS_DISPATCHING | NWS_SCHEDULED)) {
1008 error = netisr_queue_workstream(nwsp, proto, npwp, m,
1017 * The current thread is now effectively the netisr worker, so set
1018 * the dispatching flag to prevent concurrent processing of the
1019 * stream from another thread (even the netisr worker), which could
1020 * otherwise lead to effective misordering of the stream.
1022 nwsp->nws_flags |= NWS_DISPATCHING;
1024 netisr_proto[proto].np_handler(m);
1026 nwsp->nws_flags &= ~NWS_DISPATCHING;
1028 npwp->nw_hybrid_dispatched++;
1031 * If other work was enqueued by another thread while we were direct
1032 * dispatching, we need to signal the netisr worker to do that work.
1033 * In the future, we might want to do some of that work in the
1034 * current thread, rather than trigger further context switches. If
1035 * so, we'll want to establish a reasonable bound on the work done in
1036 * the "borrowed" context.
1038 if (nwsp->nws_pendingbits != 0) {
1039 nwsp->nws_flags |= NWS_SCHEDULED;
1050 error = netisr_queue_internal(proto, m, cpuid);
1054 #ifdef NETISR_LOCKING
1055 NETISR_RUNLOCK(&tracker);
1061 netisr_dispatch(u_int proto, struct mbuf *m)
1064 return (netisr_dispatch_src(proto, 0, m));
1067 #ifdef DEVICE_POLLING
1069 * Kernel polling borrows a netisr thread to run interface polling in; this
1070 * function allows kernel polling to request that the netisr thread be
1071 * scheduled even if no packets are pending for protocols.
1074 netisr_sched_poll(void)
1076 struct netisr_workstream *nwsp;
1078 nwsp = DPCPU_ID_PTR(nws_array[0], nws);
1084 netisr_start_swi(u_int cpuid, struct pcpu *pc)
1087 struct netisr_workstream *nwsp;
1090 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
1092 nwsp = DPCPU_ID_PTR(cpuid, nws);
1093 mtx_init(&nwsp->nws_mtx, "netisr_mtx", NULL, MTX_DEF);
1094 nwsp->nws_cpu = cpuid;
1095 snprintf(swiname, sizeof(swiname), "netisr %u", cpuid);
1096 error = swi_add(&nwsp->nws_intr_event, swiname, swi_net, nwsp,
1097 SWI_NET, INTR_MPSAFE, &nwsp->nws_swi_cookie);
1099 panic("%s: swi_add %d", __func__, error);
1100 pc->pc_netisr = nwsp->nws_intr_event;
1101 if (netisr_bindthreads) {
1102 error = intr_event_bind(nwsp->nws_intr_event, cpuid);
1104 printf("%s: cpu %u: intr_event_bind: %d", __func__,
1108 nws_array[nws_count] = nwsp->nws_cpu;
1114 * Initialize the netisr subsystem. We rely on BSS and static initialization
1115 * of most fields in global data structures.
1117 * Start a worker thread for the boot CPU so that we can support network
1118 * traffic immediately in case the network stack is used before additional
1119 * CPUs are started (for example, diskless boot).
1122 netisr_init(void *arg)
1124 char tmp[NETISR_DISPATCH_POLICY_MAXSTR];
1125 u_int dispatch_policy;
1128 KASSERT(curcpu == 0, ("%s: not on CPU 0", __func__));
1131 if (netisr_maxthreads < 1)
1132 netisr_maxthreads = 1;
1133 if (netisr_maxthreads > mp_ncpus) {
1134 printf("netisr_init: forcing maxthreads from %d to %d\n",
1135 netisr_maxthreads, mp_ncpus);
1136 netisr_maxthreads = mp_ncpus;
1138 if (netisr_defaultqlimit > netisr_maxqlimit) {
1139 printf("netisr_init: forcing defaultqlimit from %d to %d\n",
1140 netisr_defaultqlimit, netisr_maxqlimit);
1141 netisr_defaultqlimit = netisr_maxqlimit;
1143 #ifdef DEVICE_POLLING
1145 * The device polling code is not yet aware of how to deal with
1146 * multiple netisr threads, so for the time being compiling in device
1147 * polling disables parallel netisr workers.
1149 if (netisr_maxthreads != 1 || netisr_bindthreads != 0) {
1150 printf("netisr_init: forcing maxthreads to 1 and "
1151 "bindthreads to 0 for device polling\n");
1152 netisr_maxthreads = 1;
1153 netisr_bindthreads = 0;
1157 if (TUNABLE_STR_FETCH("net.isr.dispatch", tmp, sizeof(tmp))) {
1158 error = netisr_dispatch_policy_from_str(tmp,
1160 if (error == 0 && dispatch_policy == NETISR_DISPATCH_DEFAULT)
1163 netisr_dispatch_policy = dispatch_policy;
1166 "%s: invalid dispatch policy %s, using default\n",
1170 netisr_start_swi(curcpu, pcpu_find(curcpu));
1172 SYSINIT(netisr_init, SI_SUB_SOFTINTR, SI_ORDER_FIRST, netisr_init, NULL);
1175 * Start worker threads for additional CPUs. No attempt to gracefully handle
1176 * work reassignment, we don't yet support dynamic reconfiguration.
1179 netisr_start(void *arg)
1183 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1184 if (nws_count >= netisr_maxthreads)
1186 /* XXXRW: Is skipping absent CPUs still required here? */
1187 if (CPU_ABSENT(pc->pc_cpuid))
1189 /* Worker will already be present for boot CPU. */
1190 if (pc->pc_netisr != NULL)
1192 netisr_start_swi(pc->pc_cpuid, pc);
1195 SYSINIT(netisr_start, SI_SUB_SMP, SI_ORDER_MIDDLE, netisr_start, NULL);
1198 * Sysctl monitoring for netisr: query a list of registered protocols.
1201 sysctl_netisr_proto(SYSCTL_HANDLER_ARGS)
1203 struct rm_priotracker tracker;
1204 struct sysctl_netisr_proto *snpp, *snp_array;
1205 struct netisr_proto *npp;
1206 u_int counter, proto;
1209 if (req->newptr != NULL)
1211 snp_array = malloc(sizeof(*snp_array) * NETISR_MAXPROT, M_TEMP,
1214 NETISR_RLOCK(&tracker);
1215 for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1216 npp = &netisr_proto[proto];
1217 if (npp->np_name == NULL)
1219 snpp = &snp_array[counter];
1220 snpp->snp_version = sizeof(*snpp);
1221 strlcpy(snpp->snp_name, npp->np_name, NETISR_NAMEMAXLEN);
1222 snpp->snp_proto = proto;
1223 snpp->snp_qlimit = npp->np_qlimit;
1224 snpp->snp_policy = npp->np_policy;
1225 snpp->snp_dispatch = npp->np_dispatch;
1226 if (npp->np_m2flow != NULL)
1227 snpp->snp_flags |= NETISR_SNP_FLAGS_M2FLOW;
1228 if (npp->np_m2cpuid != NULL)
1229 snpp->snp_flags |= NETISR_SNP_FLAGS_M2CPUID;
1230 if (npp->np_drainedcpu != NULL)
1231 snpp->snp_flags |= NETISR_SNP_FLAGS_DRAINEDCPU;
1234 NETISR_RUNLOCK(&tracker);
1235 KASSERT(counter <= NETISR_MAXPROT,
1236 ("sysctl_netisr_proto: counter too big (%d)", counter));
1237 error = SYSCTL_OUT(req, snp_array, sizeof(*snp_array) * counter);
1238 free(snp_array, M_TEMP);
1242 SYSCTL_PROC(_net_isr, OID_AUTO, proto,
1243 CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_proto,
1244 "S,sysctl_netisr_proto",
1245 "Return list of protocols registered with netisr");
1248 * Sysctl monitoring for netisr: query a list of workstreams.
1251 sysctl_netisr_workstream(SYSCTL_HANDLER_ARGS)
1253 struct rm_priotracker tracker;
1254 struct sysctl_netisr_workstream *snwsp, *snws_array;
1255 struct netisr_workstream *nwsp;
1256 u_int counter, cpuid;
1259 if (req->newptr != NULL)
1261 snws_array = malloc(sizeof(*snws_array) * MAXCPU, M_TEMP,
1264 NETISR_RLOCK(&tracker);
1265 CPU_FOREACH(cpuid) {
1266 nwsp = DPCPU_ID_PTR(cpuid, nws);
1267 if (nwsp->nws_intr_event == NULL)
1270 snwsp = &snws_array[counter];
1271 snwsp->snws_version = sizeof(*snwsp);
1274 * For now, we equate workstream IDs and CPU IDs in the
1275 * kernel, but expose them independently to userspace in case
1276 * that assumption changes in the future.
1278 snwsp->snws_wsid = cpuid;
1279 snwsp->snws_cpu = cpuid;
1280 if (nwsp->nws_intr_event != NULL)
1281 snwsp->snws_flags |= NETISR_SNWS_FLAGS_INTR;
1285 NETISR_RUNLOCK(&tracker);
1286 KASSERT(counter <= MAXCPU,
1287 ("sysctl_netisr_workstream: counter too big (%d)", counter));
1288 error = SYSCTL_OUT(req, snws_array, sizeof(*snws_array) * counter);
1289 free(snws_array, M_TEMP);
1293 SYSCTL_PROC(_net_isr, OID_AUTO, workstream,
1294 CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_workstream,
1295 "S,sysctl_netisr_workstream",
1296 "Return list of workstreams implemented by netisr");
1299 * Sysctl monitoring for netisr: query per-protocol data across all
1303 sysctl_netisr_work(SYSCTL_HANDLER_ARGS)
1305 struct rm_priotracker tracker;
1306 struct sysctl_netisr_work *snwp, *snw_array;
1307 struct netisr_workstream *nwsp;
1308 struct netisr_proto *npp;
1309 struct netisr_work *nwp;
1310 u_int counter, cpuid, proto;
1313 if (req->newptr != NULL)
1315 snw_array = malloc(sizeof(*snw_array) * MAXCPU * NETISR_MAXPROT,
1316 M_TEMP, M_ZERO | M_WAITOK);
1318 NETISR_RLOCK(&tracker);
1319 CPU_FOREACH(cpuid) {
1320 nwsp = DPCPU_ID_PTR(cpuid, nws);
1321 if (nwsp->nws_intr_event == NULL)
1324 for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1325 npp = &netisr_proto[proto];
1326 if (npp->np_name == NULL)
1328 nwp = &nwsp->nws_work[proto];
1329 snwp = &snw_array[counter];
1330 snwp->snw_version = sizeof(*snwp);
1331 snwp->snw_wsid = cpuid; /* See comment above. */
1332 snwp->snw_proto = proto;
1333 snwp->snw_len = nwp->nw_len;
1334 snwp->snw_watermark = nwp->nw_watermark;
1335 snwp->snw_dispatched = nwp->nw_dispatched;
1336 snwp->snw_hybrid_dispatched =
1337 nwp->nw_hybrid_dispatched;
1338 snwp->snw_qdrops = nwp->nw_qdrops;
1339 snwp->snw_queued = nwp->nw_queued;
1340 snwp->snw_handled = nwp->nw_handled;
1345 KASSERT(counter <= MAXCPU * NETISR_MAXPROT,
1346 ("sysctl_netisr_work: counter too big (%d)", counter));
1347 NETISR_RUNLOCK(&tracker);
1348 error = SYSCTL_OUT(req, snw_array, sizeof(*snw_array) * counter);
1349 free(snw_array, M_TEMP);
1353 SYSCTL_PROC(_net_isr, OID_AUTO, work,
1354 CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_work,
1355 "S,sysctl_netisr_work",
1356 "Return list of per-workstream, per-protocol work in netisr");
1359 DB_SHOW_COMMAND(netisr, db_show_netisr)
1361 struct netisr_workstream *nwsp;
1362 struct netisr_work *nwp;
1366 db_printf("%3s %6s %5s %5s %5s %8s %8s %8s %8s\n", "CPU", "Proto",
1367 "Len", "WMark", "Max", "Disp", "HDisp", "Drop", "Queue");
1368 CPU_FOREACH(cpuid) {
1369 nwsp = DPCPU_ID_PTR(cpuid, nws);
1370 if (nwsp->nws_intr_event == NULL)
1373 for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1374 if (netisr_proto[proto].np_handler == NULL)
1376 nwp = &nwsp->nws_work[proto];
1378 db_printf("%3d ", cpuid);
1381 db_printf("%3s ", "");
1383 "%6s %5d %5d %5d %8ju %8ju %8ju %8ju\n",
1384 netisr_proto[proto].np_name, nwp->nw_len,
1385 nwp->nw_watermark, nwp->nw_qlimit,
1386 nwp->nw_dispatched, nwp->nw_hybrid_dispatched,
1387 nwp->nw_qdrops, nwp->nw_queued);