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/malloc.h>
74 #include <sys/interrupt.h>
77 #include <sys/mutex.h>
80 #include <sys/rmlock.h>
81 #include <sys/sched.h>
83 #include <sys/socket.h>
84 #include <sys/sysctl.h>
85 #include <sys/systm.h>
91 #define _WANT_NETISR_INTERNAL /* Enable definitions from netisr_internal.h */
93 #include <net/if_var.h>
94 #include <net/netisr.h>
95 #include <net/netisr_internal.h>
99 * Synchronize use and modification of the registered netisr data structures;
100 * acquire a read lock while modifying the set of registered protocols to
101 * prevent partially registered or unregistered protocols from being run.
103 * The following data structures and fields are protected by this lock:
105 * - The netisr_proto array, including all fields of struct netisr_proto.
106 * - The nws array, including all fields of struct netisr_worker.
107 * - The nws_array array.
109 * Note: the NETISR_LOCKING define controls whether read locks are acquired
110 * in packet processing paths requiring netisr registration stability. This
111 * is disabled by default as it can lead to measurable performance
112 * degradation even with rmlocks (3%-6% for loopback ping-pong traffic), and
113 * because netisr registration and unregistration is extremely rare at
114 * runtime. If it becomes more common, this decision should be revisited.
116 * XXXRW: rmlocks don't support assertions.
118 static struct rmlock netisr_rmlock;
119 #define NETISR_LOCK_INIT() rm_init_flags(&netisr_rmlock, "netisr", \
121 #define NETISR_LOCK_ASSERT()
122 #define NETISR_RLOCK(tracker) rm_rlock(&netisr_rmlock, (tracker))
123 #define NETISR_RUNLOCK(tracker) rm_runlock(&netisr_rmlock, (tracker))
124 #define NETISR_WLOCK() rm_wlock(&netisr_rmlock)
125 #define NETISR_WUNLOCK() rm_wunlock(&netisr_rmlock)
126 /* #define NETISR_LOCKING */
128 static SYSCTL_NODE(_net, OID_AUTO, isr, CTLFLAG_RW, 0, "netisr");
131 * Three global direct dispatch policies are supported:
133 * NETISR_DISPATCH_DEFERRED: All work is deferred for a netisr, regardless of
134 * context (may be overriden by protocols).
136 * NETISR_DISPATCH_HYBRID: If the executing context allows direct dispatch,
137 * and we're running on the CPU the work would be performed on, then direct
138 * dispatch it if it wouldn't violate ordering constraints on the workstream.
140 * NETISR_DISPATCH_DIRECT: If the executing context allows direct dispatch,
141 * always direct dispatch. (The default.)
143 * Notice that changing the global policy could lead to short periods of
144 * misordered processing, but this is considered acceptable as compared to
145 * the complexity of enforcing ordering during policy changes. Protocols can
146 * override the global policy (when they're not doing that, they select
147 * NETISR_DISPATCH_DEFAULT).
149 #define NETISR_DISPATCH_POLICY_DEFAULT NETISR_DISPATCH_DIRECT
150 #define NETISR_DISPATCH_POLICY_MAXSTR 20 /* Used for temporary buffers. */
151 static u_int netisr_dispatch_policy = NETISR_DISPATCH_POLICY_DEFAULT;
152 static int sysctl_netisr_dispatch_policy(SYSCTL_HANDLER_ARGS);
153 SYSCTL_PROC(_net_isr, OID_AUTO, dispatch, CTLTYPE_STRING | CTLFLAG_RWTUN,
154 0, 0, sysctl_netisr_dispatch_policy, "A",
155 "netisr dispatch policy");
158 * Allow the administrator to limit the number of threads (CPUs) to use for
159 * netisr. We don't check netisr_maxthreads before creating the thread for
160 * CPU 0. This must be set at boot. We will create at most one thread per CPU.
161 * By default we initialize this to 1 which would assign just 1 cpu (cpu0) and
162 * therefore only 1 workstream. If set to -1, netisr would use all cpus
163 * (mp_ncpus) and therefore would have those many workstreams. One workstream
166 static int netisr_maxthreads = 1; /* Max number of threads. */
167 SYSCTL_INT(_net_isr, OID_AUTO, maxthreads, CTLFLAG_RDTUN,
168 &netisr_maxthreads, 0,
169 "Use at most this many CPUs for netisr processing");
171 static int netisr_bindthreads = 0; /* Bind threads to CPUs. */
172 SYSCTL_INT(_net_isr, OID_AUTO, bindthreads, CTLFLAG_RDTUN,
173 &netisr_bindthreads, 0, "Bind netisr threads to CPUs.");
176 * Limit per-workstream mbuf queue limits s to at most net.isr.maxqlimit,
177 * both for initial configuration and later modification using
178 * netisr_setqlimit().
180 #define NETISR_DEFAULT_MAXQLIMIT 10240
181 static u_int netisr_maxqlimit = NETISR_DEFAULT_MAXQLIMIT;
182 SYSCTL_UINT(_net_isr, OID_AUTO, maxqlimit, CTLFLAG_RDTUN,
183 &netisr_maxqlimit, 0,
184 "Maximum netisr per-protocol, per-CPU queue depth.");
187 * The default per-workstream mbuf queue limit for protocols that don't
188 * initialize the nh_qlimit field of their struct netisr_handler. If this is
189 * set above netisr_maxqlimit, we truncate it to the maximum during boot.
191 #define NETISR_DEFAULT_DEFAULTQLIMIT 256
192 static u_int netisr_defaultqlimit = NETISR_DEFAULT_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" },
292 netisr_dispatch_policy_to_str(u_int dispatch_policy, char *buffer,
295 const struct netisr_dispatch_table_entry *ndtep;
300 for (i = 0; i < nitems(netisr_dispatch_table); i++) {
301 ndtep = &netisr_dispatch_table[i];
302 if (ndtep->ndte_policy == dispatch_policy) {
303 str = ndtep->ndte_policy_str;
307 snprintf(buffer, buflen, "%s", str);
311 netisr_dispatch_policy_from_str(const char *str, u_int *dispatch_policyp)
313 const struct netisr_dispatch_table_entry *ndtep;
316 for (i = 0; i < nitems(netisr_dispatch_table); i++) {
317 ndtep = &netisr_dispatch_table[i];
318 if (strcmp(ndtep->ndte_policy_str, str) == 0) {
319 *dispatch_policyp = ndtep->ndte_policy;
327 sysctl_netisr_dispatch_policy(SYSCTL_HANDLER_ARGS)
329 char tmp[NETISR_DISPATCH_POLICY_MAXSTR];
330 u_int dispatch_policy;
333 netisr_dispatch_policy_to_str(netisr_dispatch_policy, tmp,
335 error = sysctl_handle_string(oidp, tmp, sizeof(tmp), req);
336 if (error == 0 && req->newptr != NULL) {
337 error = netisr_dispatch_policy_from_str(tmp,
339 if (error == 0 && dispatch_policy == NETISR_DISPATCH_DEFAULT)
342 netisr_dispatch_policy = dispatch_policy;
348 * Register a new netisr handler, which requires initializing per-protocol
349 * fields for each workstream. All netisr work is briefly suspended while
350 * the protocol is installed.
353 netisr_register(const struct netisr_handler *nhp)
355 struct netisr_work *npwp;
359 proto = nhp->nh_proto;
363 * Test that the requested registration is valid.
365 KASSERT(nhp->nh_name != NULL,
366 ("%s: nh_name NULL for %u", __func__, proto));
367 KASSERT(nhp->nh_handler != NULL,
368 ("%s: nh_handler NULL for %s", __func__, name));
369 KASSERT(nhp->nh_policy == NETISR_POLICY_SOURCE ||
370 nhp->nh_policy == NETISR_POLICY_FLOW ||
371 nhp->nh_policy == NETISR_POLICY_CPU,
372 ("%s: unsupported nh_policy %u for %s", __func__,
373 nhp->nh_policy, name));
374 KASSERT(nhp->nh_policy == NETISR_POLICY_FLOW ||
375 nhp->nh_m2flow == NULL,
376 ("%s: nh_policy != FLOW but m2flow defined for %s", __func__,
378 KASSERT(nhp->nh_policy == NETISR_POLICY_CPU || nhp->nh_m2cpuid == NULL,
379 ("%s: nh_policy != CPU but m2cpuid defined for %s", __func__,
381 KASSERT(nhp->nh_policy != NETISR_POLICY_CPU || nhp->nh_m2cpuid != NULL,
382 ("%s: nh_policy == CPU but m2cpuid not defined for %s", __func__,
384 KASSERT(nhp->nh_dispatch == NETISR_DISPATCH_DEFAULT ||
385 nhp->nh_dispatch == NETISR_DISPATCH_DEFERRED ||
386 nhp->nh_dispatch == NETISR_DISPATCH_HYBRID ||
387 nhp->nh_dispatch == NETISR_DISPATCH_DIRECT,
388 ("%s: invalid nh_dispatch (%u)", __func__, nhp->nh_dispatch));
390 KASSERT(proto < NETISR_MAXPROT,
391 ("%s(%u, %s): protocol too big", __func__, proto, name));
394 * Test that no existing registration exists for this protocol.
397 KASSERT(netisr_proto[proto].np_name == NULL,
398 ("%s(%u, %s): name present", __func__, proto, name));
399 KASSERT(netisr_proto[proto].np_handler == NULL,
400 ("%s(%u, %s): handler present", __func__, proto, name));
402 netisr_proto[proto].np_name = name;
403 netisr_proto[proto].np_handler = nhp->nh_handler;
404 netisr_proto[proto].np_m2flow = nhp->nh_m2flow;
405 netisr_proto[proto].np_m2cpuid = nhp->nh_m2cpuid;
406 netisr_proto[proto].np_drainedcpu = nhp->nh_drainedcpu;
407 if (nhp->nh_qlimit == 0)
408 netisr_proto[proto].np_qlimit = netisr_defaultqlimit;
409 else if (nhp->nh_qlimit > netisr_maxqlimit) {
410 printf("%s: %s requested queue limit %u capped to "
411 "net.isr.maxqlimit %u\n", __func__, name, nhp->nh_qlimit,
413 netisr_proto[proto].np_qlimit = netisr_maxqlimit;
415 netisr_proto[proto].np_qlimit = nhp->nh_qlimit;
416 netisr_proto[proto].np_policy = nhp->nh_policy;
417 netisr_proto[proto].np_dispatch = nhp->nh_dispatch;
419 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
420 bzero(npwp, sizeof(*npwp));
421 npwp->nw_qlimit = netisr_proto[proto].np_qlimit;
427 * Clear drop counters across all workstreams for a protocol.
430 netisr_clearqdrops(const struct netisr_handler *nhp)
432 struct netisr_work *npwp;
438 proto = nhp->nh_proto;
442 KASSERT(proto < NETISR_MAXPROT,
443 ("%s(%u): protocol too big for %s", __func__, proto, name));
446 KASSERT(netisr_proto[proto].np_handler != NULL,
447 ("%s(%u): protocol not registered for %s", __func__, proto,
451 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
458 * Query current drop counters across all workstreams for a protocol.
461 netisr_getqdrops(const struct netisr_handler *nhp, u_int64_t *qdropp)
463 struct netisr_work *npwp;
464 struct rm_priotracker tracker;
471 proto = nhp->nh_proto;
475 KASSERT(proto < NETISR_MAXPROT,
476 ("%s(%u): protocol too big for %s", __func__, proto, name));
478 NETISR_RLOCK(&tracker);
479 KASSERT(netisr_proto[proto].np_handler != NULL,
480 ("%s(%u): protocol not registered for %s", __func__, proto,
484 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
485 *qdropp += npwp->nw_qdrops;
487 NETISR_RUNLOCK(&tracker);
491 * Query current per-workstream queue limit for a protocol.
494 netisr_getqlimit(const struct netisr_handler *nhp, u_int *qlimitp)
496 struct rm_priotracker tracker;
502 proto = nhp->nh_proto;
506 KASSERT(proto < NETISR_MAXPROT,
507 ("%s(%u): protocol too big for %s", __func__, proto, name));
509 NETISR_RLOCK(&tracker);
510 KASSERT(netisr_proto[proto].np_handler != NULL,
511 ("%s(%u): protocol not registered for %s", __func__, proto,
513 *qlimitp = netisr_proto[proto].np_qlimit;
514 NETISR_RUNLOCK(&tracker);
518 * Update the queue limit across per-workstream queues for a protocol. We
519 * simply change the limits, and don't drain overflowed packets as they will
520 * (hopefully) take care of themselves shortly.
523 netisr_setqlimit(const struct netisr_handler *nhp, u_int qlimit)
525 struct netisr_work *npwp;
531 if (qlimit > netisr_maxqlimit)
534 proto = nhp->nh_proto;
538 KASSERT(proto < NETISR_MAXPROT,
539 ("%s(%u): protocol too big for %s", __func__, proto, name));
542 KASSERT(netisr_proto[proto].np_handler != NULL,
543 ("%s(%u): protocol not registered for %s", __func__, proto,
546 netisr_proto[proto].np_qlimit = qlimit;
548 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
549 npwp->nw_qlimit = qlimit;
556 * Drain all packets currently held in a particular protocol work queue.
559 netisr_drain_proto(struct netisr_work *npwp)
564 * We would assert the lock on the workstream but it's not passed in.
566 while ((m = npwp->nw_head) != NULL) {
567 npwp->nw_head = m->m_nextpkt;
569 if (npwp->nw_head == NULL)
570 npwp->nw_tail = NULL;
574 KASSERT(npwp->nw_tail == NULL, ("%s: tail", __func__));
575 KASSERT(npwp->nw_len == 0, ("%s: len", __func__));
579 * Remove the registration of a network protocol, which requires clearing
580 * per-protocol fields across all workstreams, including freeing all mbufs in
581 * the queues at time of unregister. All work in netisr is briefly suspended
582 * while this takes place.
585 netisr_unregister(const struct netisr_handler *nhp)
587 struct netisr_work *npwp;
593 proto = nhp->nh_proto;
597 KASSERT(proto < NETISR_MAXPROT,
598 ("%s(%u): protocol too big for %s", __func__, proto, name));
601 KASSERT(netisr_proto[proto].np_handler != NULL,
602 ("%s(%u): protocol not registered for %s", __func__, proto,
605 netisr_proto[proto].np_name = NULL;
606 netisr_proto[proto].np_handler = NULL;
607 netisr_proto[proto].np_m2flow = NULL;
608 netisr_proto[proto].np_m2cpuid = NULL;
609 netisr_proto[proto].np_qlimit = 0;
610 netisr_proto[proto].np_policy = 0;
612 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
613 netisr_drain_proto(npwp);
614 bzero(npwp, sizeof(*npwp));
620 * Compose the global and per-protocol policies on dispatch, and return the
621 * dispatch policy to use.
624 netisr_get_dispatch(struct netisr_proto *npp)
628 * Protocol-specific configuration overrides the global default.
630 if (npp->np_dispatch != NETISR_DISPATCH_DEFAULT)
631 return (npp->np_dispatch);
632 return (netisr_dispatch_policy);
636 * Look up the workstream given a packet and source identifier. Do this by
637 * checking the protocol's policy, and optionally call out to the protocol
638 * for assistance if required.
641 netisr_select_cpuid(struct netisr_proto *npp, u_int dispatch_policy,
642 uintptr_t source, struct mbuf *m, u_int *cpuidp)
647 NETISR_LOCK_ASSERT();
650 * In the event we have only one worker, shortcut and deliver to it
651 * without further ado.
653 if (nws_count == 1) {
654 *cpuidp = nws_array[0];
659 * What happens next depends on the policy selected by the protocol.
660 * If we want to support per-interface policies, we should do that
663 policy = npp->np_policy;
664 if (policy == NETISR_POLICY_CPU) {
665 m = npp->np_m2cpuid(m, source, cpuidp);
670 * It's possible for a protocol not to have a good idea about
671 * where to process a packet, in which case we fall back on
672 * the netisr code to decide. In the hybrid case, return the
673 * current CPU ID, which will force an immediate direct
674 * dispatch. In the queued case, fall back on the SOURCE
677 if (*cpuidp != NETISR_CPUID_NONE)
679 if (dispatch_policy == NETISR_DISPATCH_HYBRID) {
683 policy = NETISR_POLICY_SOURCE;
686 if (policy == NETISR_POLICY_FLOW) {
687 if (M_HASHTYPE_GET(m) == M_HASHTYPE_NONE &&
688 npp->np_m2flow != NULL) {
689 m = npp->np_m2flow(m, source);
693 if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
695 netisr_default_flow2cpu(m->m_pkthdr.flowid);
698 policy = NETISR_POLICY_SOURCE;
701 KASSERT(policy == NETISR_POLICY_SOURCE,
702 ("%s: invalid policy %u for %s", __func__, npp->np_policy,
705 ifp = m->m_pkthdr.rcvif;
707 *cpuidp = nws_array[(ifp->if_index + source) % nws_count];
709 *cpuidp = nws_array[source % nws_count];
714 * Process packets associated with a workstream and protocol. For reasons of
715 * fairness, we process up to one complete netisr queue at a time, moving the
716 * queue to a stack-local queue for processing, but do not loop refreshing
717 * from the global queue. The caller is responsible for deciding whether to
718 * loop, and for setting the NWS_RUNNING flag. The passed workstream will be
719 * locked on entry and relocked before return, but will be released while
720 * processing. The number of packets processed is returned.
723 netisr_process_workstream_proto(struct netisr_workstream *nwsp, u_int proto)
725 struct netisr_work local_npw, *npwp;
729 NETISR_LOCK_ASSERT();
730 NWS_LOCK_ASSERT(nwsp);
732 KASSERT(nwsp->nws_flags & NWS_RUNNING,
733 ("%s(%u): not running", __func__, proto));
734 KASSERT(proto >= 0 && proto < NETISR_MAXPROT,
735 ("%s(%u): invalid proto\n", __func__, proto));
737 npwp = &nwsp->nws_work[proto];
738 if (npwp->nw_len == 0)
742 * Move the global work queue to a thread-local work queue.
744 * Notice that this means the effective maximum length of the queue
745 * is actually twice that of the maximum queue length specified in
746 * the protocol registration call.
748 handled = npwp->nw_len;
750 npwp->nw_head = NULL;
751 npwp->nw_tail = NULL;
753 nwsp->nws_pendingbits &= ~(1 << proto);
755 while ((m = local_npw.nw_head) != NULL) {
756 local_npw.nw_head = m->m_nextpkt;
758 if (local_npw.nw_head == NULL)
759 local_npw.nw_tail = NULL;
761 VNET_ASSERT(m->m_pkthdr.rcvif != NULL,
762 ("%s:%d rcvif == NULL: m=%p", __func__, __LINE__, m));
763 CURVNET_SET(m->m_pkthdr.rcvif->if_vnet);
764 netisr_proto[proto].np_handler(m);
767 KASSERT(local_npw.nw_len == 0,
768 ("%s(%u): len %u", __func__, proto, local_npw.nw_len));
769 if (netisr_proto[proto].np_drainedcpu)
770 netisr_proto[proto].np_drainedcpu(nwsp->nws_cpu);
772 npwp->nw_handled += handled;
777 * SWI handler for netisr -- processes packets in a set of workstreams that
778 * it owns, woken up by calls to NWS_SIGNAL(). If this workstream is already
779 * being direct dispatched, go back to sleep and wait for the dispatching
780 * thread to wake us up again.
785 #ifdef NETISR_LOCKING
786 struct rm_priotracker tracker;
788 struct netisr_workstream *nwsp;
793 #ifdef DEVICE_POLLING
794 KASSERT(nws_count == 1,
795 ("%s: device_polling but nws_count != 1", __func__));
798 #ifdef NETISR_LOCKING
799 NETISR_RLOCK(&tracker);
802 KASSERT(!(nwsp->nws_flags & NWS_RUNNING), ("swi_net: running"));
803 if (nwsp->nws_flags & NWS_DISPATCHING)
805 nwsp->nws_flags |= NWS_RUNNING;
806 nwsp->nws_flags &= ~NWS_SCHEDULED;
807 while ((bits = nwsp->nws_pendingbits) != 0) {
808 while ((prot = ffs(bits)) != 0) {
810 bits &= ~(1 << prot);
811 (void)netisr_process_workstream_proto(nwsp, prot);
814 nwsp->nws_flags &= ~NWS_RUNNING;
817 #ifdef NETISR_LOCKING
818 NETISR_RUNLOCK(&tracker);
820 #ifdef DEVICE_POLLING
826 netisr_queue_workstream(struct netisr_workstream *nwsp, u_int proto,
827 struct netisr_work *npwp, struct mbuf *m, int *dosignalp)
830 NWS_LOCK_ASSERT(nwsp);
833 if (npwp->nw_len < npwp->nw_qlimit) {
835 if (npwp->nw_head == NULL) {
839 npwp->nw_tail->m_nextpkt = m;
843 if (npwp->nw_len > npwp->nw_watermark)
844 npwp->nw_watermark = npwp->nw_len;
847 * We must set the bit regardless of NWS_RUNNING, so that
848 * swi_net() keeps calling netisr_process_workstream_proto().
850 nwsp->nws_pendingbits |= (1 << proto);
851 if (!(nwsp->nws_flags &
852 (NWS_RUNNING | NWS_DISPATCHING | NWS_SCHEDULED))) {
853 nwsp->nws_flags |= NWS_SCHEDULED;
854 *dosignalp = 1; /* Defer until unlocked. */
866 netisr_queue_internal(u_int proto, struct mbuf *m, u_int cpuid)
868 struct netisr_workstream *nwsp;
869 struct netisr_work *npwp;
872 #ifdef NETISR_LOCKING
873 NETISR_LOCK_ASSERT();
875 KASSERT(cpuid <= mp_maxid, ("%s: cpuid too big (%u, %u)", __func__,
877 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
881 nwsp = DPCPU_ID_PTR(cpuid, nws);
882 npwp = &nwsp->nws_work[proto];
884 error = netisr_queue_workstream(nwsp, proto, npwp, m, &dosignal);
892 netisr_queue_src(u_int proto, uintptr_t source, struct mbuf *m)
894 #ifdef NETISR_LOCKING
895 struct rm_priotracker tracker;
900 KASSERT(proto < NETISR_MAXPROT,
901 ("%s: invalid proto %u", __func__, proto));
903 #ifdef NETISR_LOCKING
904 NETISR_RLOCK(&tracker);
906 KASSERT(netisr_proto[proto].np_handler != NULL,
907 ("%s: invalid proto %u", __func__, proto));
909 m = netisr_select_cpuid(&netisr_proto[proto], NETISR_DISPATCH_DEFERRED,
912 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__,
914 error = netisr_queue_internal(proto, m, cpuid);
917 #ifdef NETISR_LOCKING
918 NETISR_RUNLOCK(&tracker);
924 netisr_queue(u_int proto, struct mbuf *m)
927 return (netisr_queue_src(proto, 0, m));
931 * Dispatch a packet for netisr processing; direct dispatch is permitted by
935 netisr_dispatch_src(u_int proto, uintptr_t source, struct mbuf *m)
937 #ifdef NETISR_LOCKING
938 struct rm_priotracker tracker;
940 struct netisr_workstream *nwsp;
941 struct netisr_proto *npp;
942 struct netisr_work *npwp;
944 u_int cpuid, dispatch_policy;
946 KASSERT(proto < NETISR_MAXPROT,
947 ("%s: invalid proto %u", __func__, proto));
948 #ifdef NETISR_LOCKING
949 NETISR_RLOCK(&tracker);
951 npp = &netisr_proto[proto];
952 KASSERT(npp->np_handler != NULL, ("%s: invalid proto %u", __func__,
955 dispatch_policy = netisr_get_dispatch(npp);
956 if (dispatch_policy == NETISR_DISPATCH_DEFERRED)
957 return (netisr_queue_src(proto, source, m));
960 * If direct dispatch is forced, then unconditionally dispatch
961 * without a formal CPU selection. Borrow the current CPU's stats,
962 * even if there's no worker on it. In this case we don't update
963 * nws_flags because all netisr processing will be source ordered due
964 * to always being forced to directly dispatch.
966 if (dispatch_policy == NETISR_DISPATCH_DIRECT) {
967 nwsp = DPCPU_PTR(nws);
968 npwp = &nwsp->nws_work[proto];
969 npwp->nw_dispatched++;
971 netisr_proto[proto].np_handler(m);
976 KASSERT(dispatch_policy == NETISR_DISPATCH_HYBRID,
977 ("%s: unknown dispatch policy (%u)", __func__, dispatch_policy));
980 * Otherwise, we execute in a hybrid mode where we will try to direct
981 * dispatch if we're on the right CPU and the netisr worker isn't
985 m = netisr_select_cpuid(&netisr_proto[proto], NETISR_DISPATCH_HYBRID,
991 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
994 nwsp = DPCPU_PTR(nws);
995 npwp = &nwsp->nws_work[proto];
998 * We are willing to direct dispatch only if three conditions hold:
1000 * (1) The netisr worker isn't already running,
1001 * (2) Another thread isn't already directly dispatching, and
1002 * (3) The netisr hasn't already been woken up.
1005 if (nwsp->nws_flags & (NWS_RUNNING | NWS_DISPATCHING | NWS_SCHEDULED)) {
1006 error = netisr_queue_workstream(nwsp, proto, npwp, m,
1015 * The current thread is now effectively the netisr worker, so set
1016 * the dispatching flag to prevent concurrent processing of the
1017 * stream from another thread (even the netisr worker), which could
1018 * otherwise lead to effective misordering of the stream.
1020 nwsp->nws_flags |= NWS_DISPATCHING;
1022 netisr_proto[proto].np_handler(m);
1024 nwsp->nws_flags &= ~NWS_DISPATCHING;
1026 npwp->nw_hybrid_dispatched++;
1029 * If other work was enqueued by another thread while we were direct
1030 * dispatching, we need to signal the netisr worker to do that work.
1031 * In the future, we might want to do some of that work in the
1032 * current thread, rather than trigger further context switches. If
1033 * so, we'll want to establish a reasonable bound on the work done in
1034 * the "borrowed" context.
1036 if (nwsp->nws_pendingbits != 0) {
1037 nwsp->nws_flags |= NWS_SCHEDULED;
1048 error = netisr_queue_internal(proto, m, cpuid);
1052 #ifdef NETISR_LOCKING
1053 NETISR_RUNLOCK(&tracker);
1059 netisr_dispatch(u_int proto, struct mbuf *m)
1062 return (netisr_dispatch_src(proto, 0, m));
1065 #ifdef DEVICE_POLLING
1067 * Kernel polling borrows a netisr thread to run interface polling in; this
1068 * function allows kernel polling to request that the netisr thread be
1069 * scheduled even if no packets are pending for protocols.
1072 netisr_sched_poll(void)
1074 struct netisr_workstream *nwsp;
1076 nwsp = DPCPU_ID_PTR(nws_array[0], nws);
1082 netisr_start_swi(u_int cpuid, struct pcpu *pc)
1085 struct netisr_workstream *nwsp;
1088 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
1090 nwsp = DPCPU_ID_PTR(cpuid, nws);
1091 mtx_init(&nwsp->nws_mtx, "netisr_mtx", NULL, MTX_DEF);
1092 nwsp->nws_cpu = cpuid;
1093 snprintf(swiname, sizeof(swiname), "netisr %u", cpuid);
1094 error = swi_add(&nwsp->nws_intr_event, swiname, swi_net, nwsp,
1095 SWI_NET, INTR_MPSAFE, &nwsp->nws_swi_cookie);
1097 panic("%s: swi_add %d", __func__, error);
1098 pc->pc_netisr = nwsp->nws_intr_event;
1099 if (netisr_bindthreads) {
1100 error = intr_event_bind(nwsp->nws_intr_event, cpuid);
1102 printf("%s: cpu %u: intr_event_bind: %d", __func__,
1106 nws_array[nws_count] = nwsp->nws_cpu;
1112 * Initialize the netisr subsystem. We rely on BSS and static initialization
1113 * of most fields in global data structures.
1115 * Start a worker thread for the boot CPU so that we can support network
1116 * traffic immediately in case the network stack is used before additional
1117 * CPUs are started (for example, diskless boot).
1120 netisr_init(void *arg)
1122 #ifdef EARLY_AP_STARTUP
1126 KASSERT(curcpu == 0, ("%s: not on CPU 0", __func__));
1129 if (netisr_maxthreads == 0 || netisr_maxthreads < -1 )
1130 netisr_maxthreads = 1; /* default behavior */
1131 else if (netisr_maxthreads == -1)
1132 netisr_maxthreads = mp_ncpus; /* use max cpus */
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 #ifdef EARLY_AP_STARTUP
1158 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1159 if (nws_count >= netisr_maxthreads)
1161 netisr_start_swi(pc->pc_cpuid, pc);
1164 netisr_start_swi(curcpu, pcpu_find(curcpu));
1167 SYSINIT(netisr_init, SI_SUB_SOFTINTR, SI_ORDER_FIRST, netisr_init, NULL);
1169 #ifndef EARLY_AP_STARTUP
1171 * Start worker threads for additional CPUs. No attempt to gracefully handle
1172 * work reassignment, we don't yet support dynamic reconfiguration.
1175 netisr_start(void *arg)
1179 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1180 if (nws_count >= netisr_maxthreads)
1182 /* Worker will already be present for boot CPU. */
1183 if (pc->pc_netisr != NULL)
1185 netisr_start_swi(pc->pc_cpuid, pc);
1188 SYSINIT(netisr_start, SI_SUB_SMP, SI_ORDER_MIDDLE, netisr_start, NULL);
1192 * Sysctl monitoring for netisr: query a list of registered protocols.
1195 sysctl_netisr_proto(SYSCTL_HANDLER_ARGS)
1197 struct rm_priotracker tracker;
1198 struct sysctl_netisr_proto *snpp, *snp_array;
1199 struct netisr_proto *npp;
1200 u_int counter, proto;
1203 if (req->newptr != NULL)
1205 snp_array = malloc(sizeof(*snp_array) * NETISR_MAXPROT, M_TEMP,
1208 NETISR_RLOCK(&tracker);
1209 for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1210 npp = &netisr_proto[proto];
1211 if (npp->np_name == NULL)
1213 snpp = &snp_array[counter];
1214 snpp->snp_version = sizeof(*snpp);
1215 strlcpy(snpp->snp_name, npp->np_name, NETISR_NAMEMAXLEN);
1216 snpp->snp_proto = proto;
1217 snpp->snp_qlimit = npp->np_qlimit;
1218 snpp->snp_policy = npp->np_policy;
1219 snpp->snp_dispatch = npp->np_dispatch;
1220 if (npp->np_m2flow != NULL)
1221 snpp->snp_flags |= NETISR_SNP_FLAGS_M2FLOW;
1222 if (npp->np_m2cpuid != NULL)
1223 snpp->snp_flags |= NETISR_SNP_FLAGS_M2CPUID;
1224 if (npp->np_drainedcpu != NULL)
1225 snpp->snp_flags |= NETISR_SNP_FLAGS_DRAINEDCPU;
1228 NETISR_RUNLOCK(&tracker);
1229 KASSERT(counter <= NETISR_MAXPROT,
1230 ("sysctl_netisr_proto: counter too big (%d)", counter));
1231 error = SYSCTL_OUT(req, snp_array, sizeof(*snp_array) * counter);
1232 free(snp_array, M_TEMP);
1236 SYSCTL_PROC(_net_isr, OID_AUTO, proto,
1237 CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_proto,
1238 "S,sysctl_netisr_proto",
1239 "Return list of protocols registered with netisr");
1242 * Sysctl monitoring for netisr: query a list of workstreams.
1245 sysctl_netisr_workstream(SYSCTL_HANDLER_ARGS)
1247 struct rm_priotracker tracker;
1248 struct sysctl_netisr_workstream *snwsp, *snws_array;
1249 struct netisr_workstream *nwsp;
1250 u_int counter, cpuid;
1253 if (req->newptr != NULL)
1255 snws_array = malloc(sizeof(*snws_array) * MAXCPU, M_TEMP,
1258 NETISR_RLOCK(&tracker);
1259 CPU_FOREACH(cpuid) {
1260 nwsp = DPCPU_ID_PTR(cpuid, nws);
1261 if (nwsp->nws_intr_event == NULL)
1264 snwsp = &snws_array[counter];
1265 snwsp->snws_version = sizeof(*snwsp);
1268 * For now, we equate workstream IDs and CPU IDs in the
1269 * kernel, but expose them independently to userspace in case
1270 * that assumption changes in the future.
1272 snwsp->snws_wsid = cpuid;
1273 snwsp->snws_cpu = cpuid;
1274 if (nwsp->nws_intr_event != NULL)
1275 snwsp->snws_flags |= NETISR_SNWS_FLAGS_INTR;
1279 NETISR_RUNLOCK(&tracker);
1280 KASSERT(counter <= MAXCPU,
1281 ("sysctl_netisr_workstream: counter too big (%d)", counter));
1282 error = SYSCTL_OUT(req, snws_array, sizeof(*snws_array) * counter);
1283 free(snws_array, M_TEMP);
1287 SYSCTL_PROC(_net_isr, OID_AUTO, workstream,
1288 CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_workstream,
1289 "S,sysctl_netisr_workstream",
1290 "Return list of workstreams implemented by netisr");
1293 * Sysctl monitoring for netisr: query per-protocol data across all
1297 sysctl_netisr_work(SYSCTL_HANDLER_ARGS)
1299 struct rm_priotracker tracker;
1300 struct sysctl_netisr_work *snwp, *snw_array;
1301 struct netisr_workstream *nwsp;
1302 struct netisr_proto *npp;
1303 struct netisr_work *nwp;
1304 u_int counter, cpuid, proto;
1307 if (req->newptr != NULL)
1309 snw_array = malloc(sizeof(*snw_array) * MAXCPU * NETISR_MAXPROT,
1310 M_TEMP, M_ZERO | M_WAITOK);
1312 NETISR_RLOCK(&tracker);
1313 CPU_FOREACH(cpuid) {
1314 nwsp = DPCPU_ID_PTR(cpuid, nws);
1315 if (nwsp->nws_intr_event == NULL)
1318 for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1319 npp = &netisr_proto[proto];
1320 if (npp->np_name == NULL)
1322 nwp = &nwsp->nws_work[proto];
1323 snwp = &snw_array[counter];
1324 snwp->snw_version = sizeof(*snwp);
1325 snwp->snw_wsid = cpuid; /* See comment above. */
1326 snwp->snw_proto = proto;
1327 snwp->snw_len = nwp->nw_len;
1328 snwp->snw_watermark = nwp->nw_watermark;
1329 snwp->snw_dispatched = nwp->nw_dispatched;
1330 snwp->snw_hybrid_dispatched =
1331 nwp->nw_hybrid_dispatched;
1332 snwp->snw_qdrops = nwp->nw_qdrops;
1333 snwp->snw_queued = nwp->nw_queued;
1334 snwp->snw_handled = nwp->nw_handled;
1339 KASSERT(counter <= MAXCPU * NETISR_MAXPROT,
1340 ("sysctl_netisr_work: counter too big (%d)", counter));
1341 NETISR_RUNLOCK(&tracker);
1342 error = SYSCTL_OUT(req, snw_array, sizeof(*snw_array) * counter);
1343 free(snw_array, M_TEMP);
1347 SYSCTL_PROC(_net_isr, OID_AUTO, work,
1348 CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_work,
1349 "S,sysctl_netisr_work",
1350 "Return list of per-workstream, per-protocol work in netisr");
1353 DB_SHOW_COMMAND(netisr, db_show_netisr)
1355 struct netisr_workstream *nwsp;
1356 struct netisr_work *nwp;
1360 db_printf("%3s %6s %5s %5s %5s %8s %8s %8s %8s\n", "CPU", "Proto",
1361 "Len", "WMark", "Max", "Disp", "HDisp", "Drop", "Queue");
1362 CPU_FOREACH(cpuid) {
1363 nwsp = DPCPU_ID_PTR(cpuid, nws);
1364 if (nwsp->nws_intr_event == NULL)
1367 for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1368 if (netisr_proto[proto].np_handler == NULL)
1370 nwp = &nwsp->nws_work[proto];
1372 db_printf("%3d ", cpuid);
1375 db_printf("%3s ", "");
1377 "%6s %5d %5d %5d %8ju %8ju %8ju %8ju\n",
1378 netisr_proto[proto].np_name, nwp->nw_len,
1379 nwp->nw_watermark, nwp->nw_qlimit,
1380 nwp->nw_dispatched, nwp->nw_hybrid_dispatched,
1381 nwp->nw_qdrops, nwp->nw_queued);