xref: /freebsd/sys/net/netisr.c (revision 1d897d19f31dd6d0411eb01ac4695e1c6aa8836a)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2007-2009 Robert N. M. Watson
5  * Copyright (c) 2010-2011 Juniper Networks, Inc.
6  * All rights reserved.
7  *
8  * This software was developed by Robert N. M. Watson under contract
9  * to Juniper Networks, Inc.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 #include <sys/cdefs.h>
34 /*
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:
42  *
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 ...).
48  *
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
58  * more than one flow.
59  *
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
63  * netisr.h.
64  */
65 
66 #include "opt_ddb.h"
67 #include "opt_device_polling.h"
68 
69 #include <sys/param.h>
70 #include <sys/bus.h>
71 #include <sys/kernel.h>
72 #include <sys/kthread.h>
73 #include <sys/malloc.h>
74 #include <sys/interrupt.h>
75 #include <sys/lock.h>
76 #include <sys/mbuf.h>
77 #include <sys/mutex.h>
78 #include <sys/pcpu.h>
79 #include <sys/proc.h>
80 #include <sys/rmlock.h>
81 #include <sys/sched.h>
82 #include <sys/smp.h>
83 #include <sys/socket.h>
84 #include <sys/sysctl.h>
85 #include <sys/systm.h>
86 
87 #ifdef DDB
88 #include <ddb/ddb.h>
89 #endif
90 
91 #define	_WANT_NETISR_INTERNAL	/* Enable definitions from netisr_internal.h */
92 #include <net/if.h>
93 #include <net/if_var.h>
94 #include <net/if_private.h>
95 #include <net/netisr.h>
96 #include <net/netisr_internal.h>
97 #include <net/vnet.h>
98 
99 /*-
100  * Synchronize use and modification of the registered netisr data structures;
101  * acquire a read lock while modifying the set of registered protocols to
102  * prevent partially registered or unregistered protocols from being run.
103  *
104  * The following data structures and fields are protected by this lock:
105  *
106  * - The netisr_proto array, including all fields of struct netisr_proto.
107  * - The nws array, including all fields of struct netisr_worker.
108  * - The nws_array array.
109  *
110  * Note: the NETISR_LOCKING define controls whether read locks are acquired
111  * in packet processing paths requiring netisr registration stability.  This
112  * is disabled by default as it can lead to measurable performance
113  * degradation even with rmlocks (3%-6% for loopback ping-pong traffic), and
114  * because netisr registration and unregistration is extremely rare at
115  * runtime.  If it becomes more common, this decision should be revisited.
116  *
117  * XXXRW: rmlocks don't support assertions.
118  */
119 static struct rmlock	netisr_rmlock;
120 #define	NETISR_LOCK_INIT()	rm_init_flags(&netisr_rmlock, "netisr", \
121 				    RM_NOWITNESS)
122 #define	NETISR_LOCK_ASSERT()
123 #define	NETISR_RLOCK(tracker)	rm_rlock(&netisr_rmlock, (tracker))
124 #define	NETISR_RUNLOCK(tracker)	rm_runlock(&netisr_rmlock, (tracker))
125 #define	NETISR_WLOCK()		rm_wlock(&netisr_rmlock)
126 #define	NETISR_WUNLOCK()	rm_wunlock(&netisr_rmlock)
127 /* #define	NETISR_LOCKING */
128 
129 static SYSCTL_NODE(_net, OID_AUTO, isr, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
130     "netisr");
131 
132 /*-
133  * Three global direct dispatch policies are supported:
134  *
135  * NETISR_DISPATCH_DEFERRED: All work is deferred for a netisr, regardless of
136  * context (may be overridden by protocols).
137  *
138  * NETISR_DISPATCH_HYBRID: If the executing context allows direct dispatch,
139  * and we're running on the CPU the work would be performed on, then direct
140  * dispatch it if it wouldn't violate ordering constraints on the workstream.
141  *
142  * NETISR_DISPATCH_DIRECT: If the executing context allows direct dispatch,
143  * always direct dispatch.  (The default.)
144  *
145  * Notice that changing the global policy could lead to short periods of
146  * misordered processing, but this is considered acceptable as compared to
147  * the complexity of enforcing ordering during policy changes.  Protocols can
148  * override the global policy (when they're not doing that, they select
149  * NETISR_DISPATCH_DEFAULT).
150  */
151 #define	NETISR_DISPATCH_POLICY_DEFAULT	NETISR_DISPATCH_DIRECT
152 #define	NETISR_DISPATCH_POLICY_MAXSTR	20 /* Used for temporary buffers. */
153 static u_int	netisr_dispatch_policy = NETISR_DISPATCH_POLICY_DEFAULT;
154 static int	sysctl_netisr_dispatch_policy(SYSCTL_HANDLER_ARGS);
155 SYSCTL_PROC(_net_isr, OID_AUTO, dispatch,
156     CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT,
157     0, 0, sysctl_netisr_dispatch_policy, "A",
158     "netisr dispatch policy");
159 
160 /*
161  * Allow the administrator to limit the number of threads (CPUs) to use for
162  * netisr.  We don't check netisr_maxthreads before creating the thread for
163  * CPU 0. This must be set at boot. We will create at most one thread per CPU.
164  * By default we initialize this to 1 which would assign just 1 cpu (cpu0) and
165  * therefore only 1 workstream. If set to -1, netisr would use all cpus
166  * (mp_ncpus) and therefore would have those many workstreams. One workstream
167  * per thread (CPU).
168  */
169 static int	netisr_maxthreads = 1;		/* Max number of threads. */
170 SYSCTL_INT(_net_isr, OID_AUTO, maxthreads, CTLFLAG_RDTUN,
171     &netisr_maxthreads, 0,
172     "Use at most this many CPUs for netisr processing");
173 
174 static int	netisr_bindthreads = 0;		/* Bind threads to CPUs. */
175 SYSCTL_INT(_net_isr, OID_AUTO, bindthreads, CTLFLAG_RDTUN,
176     &netisr_bindthreads, 0, "Bind netisr threads to CPUs.");
177 
178 /*
179  * Limit per-workstream mbuf queue limits s to at most net.isr.maxqlimit,
180  * both for initial configuration and later modification using
181  * netisr_setqlimit().
182  */
183 #define	NETISR_DEFAULT_MAXQLIMIT	10240
184 static u_int	netisr_maxqlimit = NETISR_DEFAULT_MAXQLIMIT;
185 SYSCTL_UINT(_net_isr, OID_AUTO, maxqlimit, CTLFLAG_RDTUN,
186     &netisr_maxqlimit, 0,
187     "Maximum netisr per-protocol, per-CPU queue depth.");
188 
189 /*
190  * The default per-workstream mbuf queue limit for protocols that don't
191  * initialize the nh_qlimit field of their struct netisr_handler.  If this is
192  * set above netisr_maxqlimit, we truncate it to the maximum during boot.
193  */
194 #define	NETISR_DEFAULT_DEFAULTQLIMIT	256
195 static u_int	netisr_defaultqlimit = NETISR_DEFAULT_DEFAULTQLIMIT;
196 SYSCTL_UINT(_net_isr, OID_AUTO, defaultqlimit, CTLFLAG_RDTUN,
197     &netisr_defaultqlimit, 0,
198     "Default netisr per-protocol, per-CPU queue limit if not set by protocol");
199 
200 /*
201  * Store and export the compile-time constant NETISR_MAXPROT limit on the
202  * number of protocols that can register with netisr at a time.  This is
203  * required for crashdump analysis, as it sizes netisr_proto[].
204  */
205 static u_int	netisr_maxprot = NETISR_MAXPROT;
206 SYSCTL_UINT(_net_isr, OID_AUTO, maxprot, CTLFLAG_RD,
207     &netisr_maxprot, 0,
208     "Compile-time limit on the number of protocols supported by netisr.");
209 
210 /*
211  * The netisr_proto array describes all registered protocols, indexed by
212  * protocol number.  See netisr_internal.h for more details.
213  */
214 static struct netisr_proto	netisr_proto[NETISR_MAXPROT];
215 
216 #ifdef VIMAGE
217 /*
218  * The netisr_enable array describes a per-VNET flag for registered
219  * protocols on whether this netisr is active in this VNET or not.
220  * netisr_register() will automatically enable the netisr for the
221  * default VNET and all currently active instances.
222  * netisr_unregister() will disable all active VNETs, including vnet0.
223  * Individual network stack instances can be enabled/disabled by the
224  * netisr_(un)register _vnet() functions.
225  * With this we keep the one netisr_proto per protocol but add a
226  * mechanism to stop netisr processing for vnet teardown.
227  * Apart from that we expect a VNET to always be enabled.
228  */
229 VNET_DEFINE_STATIC(u_int,	netisr_enable[NETISR_MAXPROT]);
230 #define	V_netisr_enable		VNET(netisr_enable)
231 #endif
232 
233 /*
234  * Per-CPU workstream data.  See netisr_internal.h for more details.
235  */
236 DPCPU_DEFINE(struct netisr_workstream, nws);
237 
238 /*
239  * Map contiguous values between 0 and nws_count into CPU IDs appropriate for
240  * accessing workstreams.  This allows constructions of the form
241  * DPCPU_ID_GET(nws_array[arbitraryvalue % nws_count], nws).
242  */
243 static u_int				 nws_array[MAXCPU];
244 
245 /*
246  * Number of registered workstreams.  Will be at most the number of running
247  * CPUs once fully started.
248  */
249 static u_int				 nws_count;
250 SYSCTL_UINT(_net_isr, OID_AUTO, numthreads, CTLFLAG_RD,
251     &nws_count, 0, "Number of extant netisr threads.");
252 
253 /*
254  * Synchronization for each workstream: a mutex protects all mutable fields
255  * in each stream, including per-protocol state (mbuf queues).  The SWI is
256  * woken up if asynchronous dispatch is required.
257  */
258 #define	NWS_LOCK(s)		mtx_lock(&(s)->nws_mtx)
259 #define	NWS_LOCK_ASSERT(s)	mtx_assert(&(s)->nws_mtx, MA_OWNED)
260 #define	NWS_UNLOCK(s)		mtx_unlock(&(s)->nws_mtx)
261 #define	NWS_SIGNAL(s)		swi_sched((s)->nws_swi_cookie, 0)
262 
263 /*
264  * Utility routines for protocols that implement their own mapping of flows
265  * to CPUs.
266  */
267 u_int
268 netisr_get_cpucount(void)
269 {
270 
271 	return (nws_count);
272 }
273 
274 u_int
275 netisr_get_cpuid(u_int cpunumber)
276 {
277 
278 	return (nws_array[cpunumber % nws_count]);
279 }
280 
281 /*
282  * The default implementation of flow -> CPU ID mapping.
283  *
284  * Non-static so that protocols can use it to map their own work to specific
285  * CPUs in a manner consistent to netisr for affinity purposes.
286  */
287 u_int
288 netisr_default_flow2cpu(u_int flowid)
289 {
290 
291 	return (nws_array[flowid % nws_count]);
292 }
293 
294 /*
295  * Dispatch tunable and sysctl configuration.
296  */
297 struct netisr_dispatch_table_entry {
298 	u_int		 ndte_policy;
299 	const char	*ndte_policy_str;
300 };
301 static const struct netisr_dispatch_table_entry netisr_dispatch_table[] = {
302 	{ NETISR_DISPATCH_DEFAULT, "default" },
303 	{ NETISR_DISPATCH_DEFERRED, "deferred" },
304 	{ NETISR_DISPATCH_HYBRID, "hybrid" },
305 	{ NETISR_DISPATCH_DIRECT, "direct" },
306 };
307 
308 static void
309 netisr_dispatch_policy_to_str(u_int dispatch_policy, char *buffer,
310     u_int buflen)
311 {
312 	const struct netisr_dispatch_table_entry *ndtep;
313 	const char *str;
314 	u_int i;
315 
316 	str = "unknown";
317 	for (i = 0; i < nitems(netisr_dispatch_table); i++) {
318 		ndtep = &netisr_dispatch_table[i];
319 		if (ndtep->ndte_policy == dispatch_policy) {
320 			str = ndtep->ndte_policy_str;
321 			break;
322 		}
323 	}
324 	snprintf(buffer, buflen, "%s", str);
325 }
326 
327 static int
328 netisr_dispatch_policy_from_str(const char *str, u_int *dispatch_policyp)
329 {
330 	const struct netisr_dispatch_table_entry *ndtep;
331 	u_int i;
332 
333 	for (i = 0; i < nitems(netisr_dispatch_table); i++) {
334 		ndtep = &netisr_dispatch_table[i];
335 		if (strcmp(ndtep->ndte_policy_str, str) == 0) {
336 			*dispatch_policyp = ndtep->ndte_policy;
337 			return (0);
338 		}
339 	}
340 	return (EINVAL);
341 }
342 
343 static int
344 sysctl_netisr_dispatch_policy(SYSCTL_HANDLER_ARGS)
345 {
346 	char tmp[NETISR_DISPATCH_POLICY_MAXSTR];
347 	size_t len;
348 	u_int dispatch_policy;
349 	int error;
350 
351 	netisr_dispatch_policy_to_str(netisr_dispatch_policy, tmp,
352 	    sizeof(tmp));
353 	/*
354 	 * netisr is initialised very early during the boot when malloc isn't
355 	 * available yet so we can't use sysctl_handle_string() to process
356 	 * any non-default value that was potentially set via loader.
357 	 */
358 	if (req->newptr != NULL) {
359 		len = req->newlen - req->newidx;
360 		if (len >= NETISR_DISPATCH_POLICY_MAXSTR)
361 			return (EINVAL);
362 		error = SYSCTL_IN(req, tmp, len);
363 		if (error == 0) {
364 			tmp[len] = '\0';
365 			error = netisr_dispatch_policy_from_str(tmp,
366 			    &dispatch_policy);
367 			if (error == 0 &&
368 			    dispatch_policy == NETISR_DISPATCH_DEFAULT)
369 				error = EINVAL;
370 			if (error == 0)
371 				netisr_dispatch_policy = dispatch_policy;
372 		}
373 	} else {
374 		error = sysctl_handle_string(oidp, tmp, sizeof(tmp), req);
375 	}
376 	return (error);
377 }
378 
379 /*
380  * Register a new netisr handler, which requires initializing per-protocol
381  * fields for each workstream.  All netisr work is briefly suspended while
382  * the protocol is installed.
383  */
384 void
385 netisr_register(const struct netisr_handler *nhp)
386 {
387 	VNET_ITERATOR_DECL(vnet_iter);
388 	struct netisr_work *npwp;
389 	const char *name;
390 	u_int i, proto;
391 
392 	proto = nhp->nh_proto;
393 	name = nhp->nh_name;
394 
395 	/*
396 	 * Test that the requested registration is valid.
397 	 */
398 	KASSERT(nhp->nh_name != NULL,
399 	    ("%s: nh_name NULL for %u", __func__, proto));
400 	KASSERT(nhp->nh_handler != NULL,
401 	    ("%s: nh_handler NULL for %s", __func__, name));
402 	KASSERT(nhp->nh_policy == NETISR_POLICY_SOURCE ||
403 	    nhp->nh_policy == NETISR_POLICY_FLOW ||
404 	    nhp->nh_policy == NETISR_POLICY_CPU,
405 	    ("%s: unsupported nh_policy %u for %s", __func__,
406 	    nhp->nh_policy, name));
407 	KASSERT(nhp->nh_policy == NETISR_POLICY_FLOW ||
408 	    nhp->nh_m2flow == NULL,
409 	    ("%s: nh_policy != FLOW but m2flow defined for %s", __func__,
410 	    name));
411 	KASSERT(nhp->nh_policy == NETISR_POLICY_CPU || nhp->nh_m2cpuid == NULL,
412 	    ("%s: nh_policy != CPU but m2cpuid defined for %s", __func__,
413 	    name));
414 	KASSERT(nhp->nh_policy != NETISR_POLICY_CPU || nhp->nh_m2cpuid != NULL,
415 	    ("%s: nh_policy == CPU but m2cpuid not defined for %s", __func__,
416 	    name));
417 	KASSERT(nhp->nh_dispatch == NETISR_DISPATCH_DEFAULT ||
418 	    nhp->nh_dispatch == NETISR_DISPATCH_DEFERRED ||
419 	    nhp->nh_dispatch == NETISR_DISPATCH_HYBRID ||
420 	    nhp->nh_dispatch == NETISR_DISPATCH_DIRECT,
421 	    ("%s: invalid nh_dispatch (%u)", __func__, nhp->nh_dispatch));
422 
423 	KASSERT(proto < NETISR_MAXPROT,
424 	    ("%s(%u, %s): protocol too big", __func__, proto, name));
425 
426 	/*
427 	 * Test that no existing registration exists for this protocol.
428 	 */
429 	NETISR_WLOCK();
430 	KASSERT(netisr_proto[proto].np_name == NULL,
431 	    ("%s(%u, %s): name present", __func__, proto, name));
432 	KASSERT(netisr_proto[proto].np_handler == NULL,
433 	    ("%s(%u, %s): handler present", __func__, proto, name));
434 
435 	netisr_proto[proto].np_name = name;
436 	netisr_proto[proto].np_handler = nhp->nh_handler;
437 	netisr_proto[proto].np_m2flow = nhp->nh_m2flow;
438 	netisr_proto[proto].np_m2cpuid = nhp->nh_m2cpuid;
439 	netisr_proto[proto].np_drainedcpu = nhp->nh_drainedcpu;
440 	if (nhp->nh_qlimit == 0)
441 		netisr_proto[proto].np_qlimit = netisr_defaultqlimit;
442 	else if (nhp->nh_qlimit > netisr_maxqlimit) {
443 		printf("%s: %s requested queue limit %u capped to "
444 		    "net.isr.maxqlimit %u\n", __func__, name, nhp->nh_qlimit,
445 		    netisr_maxqlimit);
446 		netisr_proto[proto].np_qlimit = netisr_maxqlimit;
447 	} else
448 		netisr_proto[proto].np_qlimit = nhp->nh_qlimit;
449 	netisr_proto[proto].np_policy = nhp->nh_policy;
450 	netisr_proto[proto].np_dispatch = nhp->nh_dispatch;
451 	CPU_FOREACH(i) {
452 		npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
453 		bzero(npwp, sizeof(*npwp));
454 		npwp->nw_qlimit = netisr_proto[proto].np_qlimit;
455 	}
456 
457 #ifdef VIMAGE
458 	/*
459 	 * Test that we are in vnet0 and have a curvnet set.
460 	 */
461 	KASSERT(curvnet != NULL, ("%s: curvnet is NULL", __func__));
462 	KASSERT(IS_DEFAULT_VNET(curvnet), ("%s: curvnet %p is not vnet0 %p",
463 	    __func__, curvnet, vnet0));
464 	VNET_LIST_RLOCK_NOSLEEP();
465 	VNET_FOREACH(vnet_iter) {
466 		CURVNET_SET(vnet_iter);
467 		V_netisr_enable[proto] = 1;
468 		CURVNET_RESTORE();
469 	}
470 	VNET_LIST_RUNLOCK_NOSLEEP();
471 #endif
472 	NETISR_WUNLOCK();
473 }
474 
475 /*
476  * Clear drop counters across all workstreams for a protocol.
477  */
478 void
479 netisr_clearqdrops(const struct netisr_handler *nhp)
480 {
481 	struct netisr_work *npwp;
482 #ifdef INVARIANTS
483 	const char *name;
484 #endif
485 	u_int i, proto;
486 
487 	proto = nhp->nh_proto;
488 #ifdef INVARIANTS
489 	name = nhp->nh_name;
490 #endif
491 	KASSERT(proto < NETISR_MAXPROT,
492 	    ("%s(%u): protocol too big for %s", __func__, proto, name));
493 
494 	NETISR_WLOCK();
495 	KASSERT(netisr_proto[proto].np_handler != NULL,
496 	    ("%s(%u): protocol not registered for %s", __func__, proto,
497 	    name));
498 
499 	CPU_FOREACH(i) {
500 		npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
501 		npwp->nw_qdrops = 0;
502 	}
503 	NETISR_WUNLOCK();
504 }
505 
506 /*
507  * Query current drop counters across all workstreams for a protocol.
508  */
509 void
510 netisr_getqdrops(const struct netisr_handler *nhp, u_int64_t *qdropp)
511 {
512 	struct netisr_work *npwp;
513 	struct rm_priotracker tracker;
514 #ifdef INVARIANTS
515 	const char *name;
516 #endif
517 	u_int i, proto;
518 
519 	*qdropp = 0;
520 	proto = nhp->nh_proto;
521 #ifdef INVARIANTS
522 	name = nhp->nh_name;
523 #endif
524 	KASSERT(proto < NETISR_MAXPROT,
525 	    ("%s(%u): protocol too big for %s", __func__, proto, name));
526 
527 	NETISR_RLOCK(&tracker);
528 	KASSERT(netisr_proto[proto].np_handler != NULL,
529 	    ("%s(%u): protocol not registered for %s", __func__, proto,
530 	    name));
531 
532 	CPU_FOREACH(i) {
533 		npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
534 		*qdropp += npwp->nw_qdrops;
535 	}
536 	NETISR_RUNLOCK(&tracker);
537 }
538 
539 /*
540  * Query current per-workstream queue limit for a protocol.
541  */
542 void
543 netisr_getqlimit(const struct netisr_handler *nhp, u_int *qlimitp)
544 {
545 	struct rm_priotracker tracker;
546 #ifdef INVARIANTS
547 	const char *name;
548 #endif
549 	u_int proto;
550 
551 	proto = nhp->nh_proto;
552 #ifdef INVARIANTS
553 	name = nhp->nh_name;
554 #endif
555 	KASSERT(proto < NETISR_MAXPROT,
556 	    ("%s(%u): protocol too big for %s", __func__, proto, name));
557 
558 	NETISR_RLOCK(&tracker);
559 	KASSERT(netisr_proto[proto].np_handler != NULL,
560 	    ("%s(%u): protocol not registered for %s", __func__, proto,
561 	    name));
562 	*qlimitp = netisr_proto[proto].np_qlimit;
563 	NETISR_RUNLOCK(&tracker);
564 }
565 
566 /*
567  * Update the queue limit across per-workstream queues for a protocol.  We
568  * simply change the limits, and don't drain overflowed packets as they will
569  * (hopefully) take care of themselves shortly.
570  */
571 int
572 netisr_setqlimit(const struct netisr_handler *nhp, u_int qlimit)
573 {
574 	struct netisr_work *npwp;
575 #ifdef INVARIANTS
576 	const char *name;
577 #endif
578 	u_int i, proto;
579 
580 	if (qlimit > netisr_maxqlimit)
581 		return (EINVAL);
582 
583 	proto = nhp->nh_proto;
584 #ifdef INVARIANTS
585 	name = nhp->nh_name;
586 #endif
587 	KASSERT(proto < NETISR_MAXPROT,
588 	    ("%s(%u): protocol too big for %s", __func__, proto, name));
589 
590 	NETISR_WLOCK();
591 	KASSERT(netisr_proto[proto].np_handler != NULL,
592 	    ("%s(%u): protocol not registered for %s", __func__, proto,
593 	    name));
594 
595 	netisr_proto[proto].np_qlimit = qlimit;
596 	CPU_FOREACH(i) {
597 		npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
598 		npwp->nw_qlimit = qlimit;
599 	}
600 	NETISR_WUNLOCK();
601 	return (0);
602 }
603 
604 /*
605  * Drain all packets currently held in a particular protocol work queue.
606  */
607 static void
608 netisr_drain_proto(struct netisr_work *npwp)
609 {
610 	struct mbuf *m;
611 
612 	/*
613 	 * We would assert the lock on the workstream but it's not passed in.
614 	 */
615 	while ((m = npwp->nw_head) != NULL) {
616 		npwp->nw_head = m->m_nextpkt;
617 		m->m_nextpkt = NULL;
618 		if (npwp->nw_head == NULL)
619 			npwp->nw_tail = NULL;
620 		npwp->nw_len--;
621 		m_freem(m);
622 	}
623 	KASSERT(npwp->nw_tail == NULL, ("%s: tail", __func__));
624 	KASSERT(npwp->nw_len == 0, ("%s: len", __func__));
625 }
626 
627 /*
628  * Remove the registration of a network protocol, which requires clearing
629  * per-protocol fields across all workstreams, including freeing all mbufs in
630  * the queues at time of unregister.  All work in netisr is briefly suspended
631  * while this takes place.
632  */
633 void
634 netisr_unregister(const struct netisr_handler *nhp)
635 {
636 	VNET_ITERATOR_DECL(vnet_iter);
637 	struct netisr_work *npwp;
638 #ifdef INVARIANTS
639 	const char *name;
640 #endif
641 	u_int i, proto;
642 
643 	proto = nhp->nh_proto;
644 #ifdef INVARIANTS
645 	name = nhp->nh_name;
646 #endif
647 	KASSERT(proto < NETISR_MAXPROT,
648 	    ("%s(%u): protocol too big for %s", __func__, proto, name));
649 
650 	NETISR_WLOCK();
651 	KASSERT(netisr_proto[proto].np_handler != NULL,
652 	    ("%s(%u): protocol not registered for %s", __func__, proto,
653 	    name));
654 
655 #ifdef VIMAGE
656 	VNET_LIST_RLOCK_NOSLEEP();
657 	VNET_FOREACH(vnet_iter) {
658 		CURVNET_SET(vnet_iter);
659 		V_netisr_enable[proto] = 0;
660 		CURVNET_RESTORE();
661 	}
662 	VNET_LIST_RUNLOCK_NOSLEEP();
663 #endif
664 
665 	netisr_proto[proto].np_name = NULL;
666 	netisr_proto[proto].np_handler = NULL;
667 	netisr_proto[proto].np_m2flow = NULL;
668 	netisr_proto[proto].np_m2cpuid = NULL;
669 	netisr_proto[proto].np_qlimit = 0;
670 	netisr_proto[proto].np_policy = 0;
671 	CPU_FOREACH(i) {
672 		npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
673 		netisr_drain_proto(npwp);
674 		bzero(npwp, sizeof(*npwp));
675 	}
676 	NETISR_WUNLOCK();
677 }
678 
679 #ifdef VIMAGE
680 void
681 netisr_register_vnet(const struct netisr_handler *nhp)
682 {
683 	u_int proto;
684 
685 	proto = nhp->nh_proto;
686 
687 	KASSERT(curvnet != NULL, ("%s: curvnet is NULL", __func__));
688 	KASSERT(proto < NETISR_MAXPROT,
689 	    ("%s(%u): protocol too big for %s", __func__, proto, nhp->nh_name));
690 	NETISR_WLOCK();
691 	KASSERT(netisr_proto[proto].np_handler != NULL,
692 	    ("%s(%u): protocol not registered for %s", __func__, proto,
693 	    nhp->nh_name));
694 
695 	V_netisr_enable[proto] = 1;
696 	NETISR_WUNLOCK();
697 }
698 
699 static void
700 netisr_drain_proto_vnet(struct vnet *vnet, u_int proto)
701 {
702 	struct epoch_tracker et;
703 	struct netisr_workstream *nwsp;
704 	struct netisr_work *npwp;
705 	struct mbuf *m, *mp, *n, *ne;
706 	struct ifnet *ifp;
707 	u_int i;
708 
709 	KASSERT(vnet != NULL, ("%s: vnet is NULL", __func__));
710 	NETISR_LOCK_ASSERT();
711 
712 	CPU_FOREACH(i) {
713 		nwsp = DPCPU_ID_PTR(i, nws);
714 		if (nwsp->nws_intr_event == NULL)
715 			continue;
716 		npwp = &nwsp->nws_work[proto];
717 		NWS_LOCK(nwsp);
718 
719 		/*
720 		 * Rather than dissecting and removing mbufs from the middle
721 		 * of the chain, we build a new chain if the packet stays and
722 		 * update the head and tail pointers at the end.  All packets
723 		 * matching the given vnet are freed.
724 		 */
725 		m = npwp->nw_head;
726 		n = ne = NULL;
727 		NET_EPOCH_ENTER(et);
728 		while (m != NULL) {
729 			mp = m;
730 			m = m->m_nextpkt;
731 			mp->m_nextpkt = NULL;
732 			if ((ifp = ifnet_byindexgen(mp->m_pkthdr.rcvidx,
733 			    mp->m_pkthdr.rcvgen)) != NULL &&
734 			    ifp->if_vnet != vnet) {
735 				if (n == NULL) {
736 					n = ne = mp;
737 				} else {
738 					ne->m_nextpkt = mp;
739 					ne = mp;
740 				}
741 				continue;
742 			}
743 			/* This is a packet in the selected vnet, or belongs
744 			   to destroyed interface. Free it. */
745 			npwp->nw_len--;
746 			m_freem(mp);
747 		}
748 		NET_EPOCH_EXIT(et);
749 		npwp->nw_head = n;
750 		npwp->nw_tail = ne;
751 		NWS_UNLOCK(nwsp);
752 	}
753 }
754 
755 void
756 netisr_unregister_vnet(const struct netisr_handler *nhp)
757 {
758 	u_int proto;
759 
760 	proto = nhp->nh_proto;
761 
762 	KASSERT(curvnet != NULL, ("%s: curvnet is NULL", __func__));
763 	KASSERT(proto < NETISR_MAXPROT,
764 	    ("%s(%u): protocol too big for %s", __func__, proto, nhp->nh_name));
765 	NETISR_WLOCK();
766 	KASSERT(netisr_proto[proto].np_handler != NULL,
767 	    ("%s(%u): protocol not registered for %s", __func__, proto,
768 	    nhp->nh_name));
769 
770 	V_netisr_enable[proto] = 0;
771 
772 	netisr_drain_proto_vnet(curvnet, proto);
773 	NETISR_WUNLOCK();
774 }
775 #endif
776 
777 /*
778  * Compose the global and per-protocol policies on dispatch, and return the
779  * dispatch policy to use.
780  */
781 static u_int
782 netisr_get_dispatch(struct netisr_proto *npp)
783 {
784 
785 	/*
786 	 * Protocol-specific configuration overrides the global default.
787 	 */
788 	if (npp->np_dispatch != NETISR_DISPATCH_DEFAULT)
789 		return (npp->np_dispatch);
790 	return (netisr_dispatch_policy);
791 }
792 
793 /*
794  * Look up the workstream given a packet and source identifier.  Do this by
795  * checking the protocol's policy, and optionally call out to the protocol
796  * for assistance if required.
797  */
798 static struct mbuf *
799 netisr_select_cpuid(struct netisr_proto *npp, u_int dispatch_policy,
800     uintptr_t source, struct mbuf *m, u_int *cpuidp)
801 {
802 	struct ifnet *ifp;
803 	u_int policy;
804 
805 	NETISR_LOCK_ASSERT();
806 
807 	/*
808 	 * In the event we have only one worker, shortcut and deliver to it
809 	 * without further ado.
810 	 */
811 	if (nws_count == 1) {
812 		*cpuidp = nws_array[0];
813 		return (m);
814 	}
815 
816 	/*
817 	 * What happens next depends on the policy selected by the protocol.
818 	 * If we want to support per-interface policies, we should do that
819 	 * here first.
820 	 */
821 	policy = npp->np_policy;
822 	if (policy == NETISR_POLICY_CPU) {
823 		m = npp->np_m2cpuid(m, source, cpuidp);
824 		if (m == NULL)
825 			return (NULL);
826 
827 		/*
828 		 * It's possible for a protocol not to have a good idea about
829 		 * where to process a packet, in which case we fall back on
830 		 * the netisr code to decide.  In the hybrid case, return the
831 		 * current CPU ID, which will force an immediate direct
832 		 * dispatch.  In the queued case, fall back on the SOURCE
833 		 * policy.
834 		 */
835 		if (*cpuidp != NETISR_CPUID_NONE) {
836 			*cpuidp = netisr_get_cpuid(*cpuidp);
837 			return (m);
838 		}
839 		if (dispatch_policy == NETISR_DISPATCH_HYBRID) {
840 			*cpuidp = netisr_get_cpuid(curcpu);
841 			return (m);
842 		}
843 		policy = NETISR_POLICY_SOURCE;
844 	}
845 
846 	if (policy == NETISR_POLICY_FLOW) {
847 		if (M_HASHTYPE_GET(m) == M_HASHTYPE_NONE &&
848 		    npp->np_m2flow != NULL) {
849 			m = npp->np_m2flow(m, source);
850 			if (m == NULL)
851 				return (NULL);
852 		}
853 		if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
854 			*cpuidp =
855 			    netisr_default_flow2cpu(m->m_pkthdr.flowid);
856 			return (m);
857 		}
858 		policy = NETISR_POLICY_SOURCE;
859 	}
860 
861 	KASSERT(policy == NETISR_POLICY_SOURCE,
862 	    ("%s: invalid policy %u for %s", __func__, npp->np_policy,
863 	    npp->np_name));
864 
865 	MPASS((m->m_pkthdr.csum_flags & CSUM_SND_TAG) == 0);
866 	ifp = m->m_pkthdr.rcvif;
867 	if (ifp != NULL)
868 		*cpuidp = nws_array[(ifp->if_index + source) % nws_count];
869 	else
870 		*cpuidp = nws_array[source % nws_count];
871 	return (m);
872 }
873 
874 /*
875  * Process packets associated with a workstream and protocol.  For reasons of
876  * fairness, we process up to one complete netisr queue at a time, moving the
877  * queue to a stack-local queue for processing, but do not loop refreshing
878  * from the global queue.  The caller is responsible for deciding whether to
879  * loop, and for setting the NWS_RUNNING flag.  The passed workstream will be
880  * locked on entry and relocked before return, but will be released while
881  * processing.  The number of packets processed is returned.
882  */
883 static u_int
884 netisr_process_workstream_proto(struct netisr_workstream *nwsp, u_int proto)
885 {
886 	struct netisr_work local_npw, *npwp;
887 	u_int handled;
888 	struct mbuf *m;
889 
890 	NETISR_LOCK_ASSERT();
891 	NWS_LOCK_ASSERT(nwsp);
892 
893 	KASSERT(nwsp->nws_flags & NWS_RUNNING,
894 	    ("%s(%u): not running", __func__, proto));
895 	KASSERT(proto >= 0 && proto < NETISR_MAXPROT,
896 	    ("%s(%u): invalid proto\n", __func__, proto));
897 
898 	npwp = &nwsp->nws_work[proto];
899 	if (npwp->nw_len == 0)
900 		return (0);
901 
902 	/*
903 	 * Move the global work queue to a thread-local work queue.
904 	 *
905 	 * Notice that this means the effective maximum length of the queue
906 	 * is actually twice that of the maximum queue length specified in
907 	 * the protocol registration call.
908 	 */
909 	handled = npwp->nw_len;
910 	local_npw = *npwp;
911 	npwp->nw_head = NULL;
912 	npwp->nw_tail = NULL;
913 	npwp->nw_len = 0;
914 	nwsp->nws_pendingbits &= ~(1 << proto);
915 	NWS_UNLOCK(nwsp);
916 	while ((m = local_npw.nw_head) != NULL) {
917 		local_npw.nw_head = m->m_nextpkt;
918 		m->m_nextpkt = NULL;
919 		if (local_npw.nw_head == NULL)
920 			local_npw.nw_tail = NULL;
921 		local_npw.nw_len--;
922 		if (__predict_false(m_rcvif_restore(m) == NULL)) {
923 			m_freem(m);
924 			continue;
925 		}
926 		CURVNET_SET(m->m_pkthdr.rcvif->if_vnet);
927 		netisr_proto[proto].np_handler(m);
928 		CURVNET_RESTORE();
929 	}
930 	KASSERT(local_npw.nw_len == 0,
931 	    ("%s(%u): len %u", __func__, proto, local_npw.nw_len));
932 	if (netisr_proto[proto].np_drainedcpu)
933 		netisr_proto[proto].np_drainedcpu(nwsp->nws_cpu);
934 	NWS_LOCK(nwsp);
935 	npwp->nw_handled += handled;
936 	return (handled);
937 }
938 
939 /*
940  * SWI handler for netisr -- processes packets in a set of workstreams that
941  * it owns, woken up by calls to NWS_SIGNAL().  If this workstream is already
942  * being direct dispatched, go back to sleep and wait for the dispatching
943  * thread to wake us up again.
944  */
945 static void
946 swi_net(void *arg)
947 {
948 #ifdef NETISR_LOCKING
949 	struct rm_priotracker tracker;
950 #endif
951 	struct netisr_workstream *nwsp;
952 	u_int bits, prot;
953 
954 	nwsp = arg;
955 
956 #ifdef DEVICE_POLLING
957 	KASSERT(nws_count == 1,
958 	    ("%s: device_polling but nws_count != 1", __func__));
959 	netisr_poll();
960 #endif
961 #ifdef NETISR_LOCKING
962 	NETISR_RLOCK(&tracker);
963 #endif
964 	NWS_LOCK(nwsp);
965 	KASSERT(!(nwsp->nws_flags & NWS_RUNNING), ("swi_net: running"));
966 	if (nwsp->nws_flags & NWS_DISPATCHING)
967 		goto out;
968 	nwsp->nws_flags |= NWS_RUNNING;
969 	nwsp->nws_flags &= ~NWS_SCHEDULED;
970 	while ((bits = nwsp->nws_pendingbits) != 0) {
971 		while (bits != 0) {
972 			prot = ffs(bits) - 1;
973 			bits &= ~(1 << prot);
974 			(void)netisr_process_workstream_proto(nwsp, prot);
975 		}
976 	}
977 	nwsp->nws_flags &= ~NWS_RUNNING;
978 out:
979 	NWS_UNLOCK(nwsp);
980 #ifdef NETISR_LOCKING
981 	NETISR_RUNLOCK(&tracker);
982 #endif
983 #ifdef DEVICE_POLLING
984 	netisr_pollmore();
985 #endif
986 }
987 
988 static int
989 netisr_queue_workstream(struct netisr_workstream *nwsp, u_int proto,
990     struct netisr_work *npwp, struct mbuf *m, int *dosignalp)
991 {
992 
993 	NWS_LOCK_ASSERT(nwsp);
994 
995 	*dosignalp = 0;
996 	if (npwp->nw_len < npwp->nw_qlimit) {
997 		m_rcvif_serialize(m);
998 		m->m_nextpkt = NULL;
999 		if (npwp->nw_head == NULL) {
1000 			npwp->nw_head = m;
1001 			npwp->nw_tail = m;
1002 		} else {
1003 			npwp->nw_tail->m_nextpkt = m;
1004 			npwp->nw_tail = m;
1005 		}
1006 		npwp->nw_len++;
1007 		if (npwp->nw_len > npwp->nw_watermark)
1008 			npwp->nw_watermark = npwp->nw_len;
1009 
1010 		/*
1011 		 * We must set the bit regardless of NWS_RUNNING, so that
1012 		 * swi_net() keeps calling netisr_process_workstream_proto().
1013 		 */
1014 		nwsp->nws_pendingbits |= (1 << proto);
1015 		if (!(nwsp->nws_flags &
1016 		    (NWS_RUNNING | NWS_DISPATCHING | NWS_SCHEDULED))) {
1017 			nwsp->nws_flags |= NWS_SCHEDULED;
1018 			*dosignalp = 1;	/* Defer until unlocked. */
1019 		}
1020 		npwp->nw_queued++;
1021 		return (0);
1022 	} else {
1023 		m_freem(m);
1024 		npwp->nw_qdrops++;
1025 		return (ENOBUFS);
1026 	}
1027 }
1028 
1029 static int
1030 netisr_queue_internal(u_int proto, struct mbuf *m, u_int cpuid)
1031 {
1032 	struct netisr_workstream *nwsp;
1033 	struct netisr_work *npwp;
1034 	int dosignal, error;
1035 
1036 #ifdef NETISR_LOCKING
1037 	NETISR_LOCK_ASSERT();
1038 #endif
1039 	KASSERT(cpuid <= mp_maxid, ("%s: cpuid too big (%u, %u)", __func__,
1040 	    cpuid, mp_maxid));
1041 	KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
1042 
1043 	dosignal = 0;
1044 	error = 0;
1045 	nwsp = DPCPU_ID_PTR(cpuid, nws);
1046 	npwp = &nwsp->nws_work[proto];
1047 	NWS_LOCK(nwsp);
1048 	error = netisr_queue_workstream(nwsp, proto, npwp, m, &dosignal);
1049 	NWS_UNLOCK(nwsp);
1050 	if (dosignal)
1051 		NWS_SIGNAL(nwsp);
1052 	return (error);
1053 }
1054 
1055 int
1056 netisr_queue_src(u_int proto, uintptr_t source, struct mbuf *m)
1057 {
1058 #ifdef NETISR_LOCKING
1059 	struct rm_priotracker tracker;
1060 #endif
1061 	u_int cpuid;
1062 	int error;
1063 
1064 	KASSERT(proto < NETISR_MAXPROT,
1065 	    ("%s: invalid proto %u", __func__, proto));
1066 
1067 #ifdef NETISR_LOCKING
1068 	NETISR_RLOCK(&tracker);
1069 #endif
1070 	KASSERT(netisr_proto[proto].np_handler != NULL,
1071 	    ("%s: invalid proto %u", __func__, proto));
1072 
1073 #ifdef VIMAGE
1074 	if (V_netisr_enable[proto] == 0) {
1075 		m_freem(m);
1076 		return (ENOPROTOOPT);
1077 	}
1078 #endif
1079 
1080 	m = netisr_select_cpuid(&netisr_proto[proto], NETISR_DISPATCH_DEFERRED,
1081 	    source, m, &cpuid);
1082 	if (m != NULL) {
1083 		KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__,
1084 		    cpuid));
1085 		VNET_ASSERT(m->m_pkthdr.rcvif != NULL,
1086 		    ("%s:%d rcvif == NULL: m=%p", __func__, __LINE__, m));
1087 		error = netisr_queue_internal(proto, m, cpuid);
1088 	} else
1089 		error = ENOBUFS;
1090 #ifdef NETISR_LOCKING
1091 	NETISR_RUNLOCK(&tracker);
1092 #endif
1093 	return (error);
1094 }
1095 
1096 int
1097 netisr_queue(u_int proto, struct mbuf *m)
1098 {
1099 
1100 	return (netisr_queue_src(proto, 0, m));
1101 }
1102 
1103 /*
1104  * Dispatch a packet for netisr processing; direct dispatch is permitted by
1105  * calling context.
1106  */
1107 int
1108 netisr_dispatch_src(u_int proto, uintptr_t source, struct mbuf *m)
1109 {
1110 #ifdef NETISR_LOCKING
1111 	struct rm_priotracker tracker;
1112 #endif
1113 	struct netisr_workstream *nwsp;
1114 	struct netisr_proto *npp;
1115 	struct netisr_work *npwp;
1116 	int dosignal, error;
1117 	u_int cpuid, dispatch_policy;
1118 
1119 	NET_EPOCH_ASSERT();
1120 	KASSERT(proto < NETISR_MAXPROT,
1121 	    ("%s: invalid proto %u", __func__, proto));
1122 #ifdef NETISR_LOCKING
1123 	NETISR_RLOCK(&tracker);
1124 #endif
1125 	npp = &netisr_proto[proto];
1126 	KASSERT(npp->np_handler != NULL, ("%s: invalid proto %u", __func__,
1127 	    proto));
1128 
1129 #ifdef VIMAGE
1130 	if (V_netisr_enable[proto] == 0) {
1131 		m_freem(m);
1132 		return (ENOPROTOOPT);
1133 	}
1134 #endif
1135 
1136 	dispatch_policy = netisr_get_dispatch(npp);
1137 	if (dispatch_policy == NETISR_DISPATCH_DEFERRED)
1138 		return (netisr_queue_src(proto, source, m));
1139 
1140 	/*
1141 	 * If direct dispatch is forced, then unconditionally dispatch
1142 	 * without a formal CPU selection.  Borrow the current CPU's stats,
1143 	 * even if there's no worker on it.  In this case we don't update
1144 	 * nws_flags because all netisr processing will be source ordered due
1145 	 * to always being forced to directly dispatch.
1146 	 */
1147 	if (dispatch_policy == NETISR_DISPATCH_DIRECT) {
1148 		nwsp = DPCPU_PTR(nws);
1149 		npwp = &nwsp->nws_work[proto];
1150 		npwp->nw_dispatched++;
1151 		npwp->nw_handled++;
1152 		netisr_proto[proto].np_handler(m);
1153 		error = 0;
1154 		goto out_unlock;
1155 	}
1156 
1157 	KASSERT(dispatch_policy == NETISR_DISPATCH_HYBRID,
1158 	    ("%s: unknown dispatch policy (%u)", __func__, dispatch_policy));
1159 
1160 	/*
1161 	 * Otherwise, we execute in a hybrid mode where we will try to direct
1162 	 * dispatch if we're on the right CPU and the netisr worker isn't
1163 	 * already running.
1164 	 */
1165 	sched_pin();
1166 	m = netisr_select_cpuid(&netisr_proto[proto], NETISR_DISPATCH_HYBRID,
1167 	    source, m, &cpuid);
1168 	if (m == NULL) {
1169 		error = ENOBUFS;
1170 		goto out_unpin;
1171 	}
1172 	KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
1173 	if (cpuid != curcpu)
1174 		goto queue_fallback;
1175 	nwsp = DPCPU_PTR(nws);
1176 	npwp = &nwsp->nws_work[proto];
1177 
1178 	/*-
1179 	 * We are willing to direct dispatch only if three conditions hold:
1180 	 *
1181 	 * (1) The netisr worker isn't already running,
1182 	 * (2) Another thread isn't already directly dispatching, and
1183 	 * (3) The netisr hasn't already been woken up.
1184 	 */
1185 	NWS_LOCK(nwsp);
1186 	if (nwsp->nws_flags & (NWS_RUNNING | NWS_DISPATCHING | NWS_SCHEDULED)) {
1187 		error = netisr_queue_workstream(nwsp, proto, npwp, m,
1188 		    &dosignal);
1189 		NWS_UNLOCK(nwsp);
1190 		if (dosignal)
1191 			NWS_SIGNAL(nwsp);
1192 		goto out_unpin;
1193 	}
1194 
1195 	/*
1196 	 * The current thread is now effectively the netisr worker, so set
1197 	 * the dispatching flag to prevent concurrent processing of the
1198 	 * stream from another thread (even the netisr worker), which could
1199 	 * otherwise lead to effective misordering of the stream.
1200 	 */
1201 	nwsp->nws_flags |= NWS_DISPATCHING;
1202 	NWS_UNLOCK(nwsp);
1203 	netisr_proto[proto].np_handler(m);
1204 	NWS_LOCK(nwsp);
1205 	nwsp->nws_flags &= ~NWS_DISPATCHING;
1206 	npwp->nw_handled++;
1207 	npwp->nw_hybrid_dispatched++;
1208 
1209 	/*
1210 	 * If other work was enqueued by another thread while we were direct
1211 	 * dispatching, we need to signal the netisr worker to do that work.
1212 	 * In the future, we might want to do some of that work in the
1213 	 * current thread, rather than trigger further context switches.  If
1214 	 * so, we'll want to establish a reasonable bound on the work done in
1215 	 * the "borrowed" context.
1216 	 */
1217 	if (nwsp->nws_pendingbits != 0) {
1218 		nwsp->nws_flags |= NWS_SCHEDULED;
1219 		dosignal = 1;
1220 	} else
1221 		dosignal = 0;
1222 	NWS_UNLOCK(nwsp);
1223 	if (dosignal)
1224 		NWS_SIGNAL(nwsp);
1225 	error = 0;
1226 	goto out_unpin;
1227 
1228 queue_fallback:
1229 	error = netisr_queue_internal(proto, m, cpuid);
1230 out_unpin:
1231 	sched_unpin();
1232 out_unlock:
1233 #ifdef NETISR_LOCKING
1234 	NETISR_RUNLOCK(&tracker);
1235 #endif
1236 	return (error);
1237 }
1238 
1239 int
1240 netisr_dispatch(u_int proto, struct mbuf *m)
1241 {
1242 
1243 	return (netisr_dispatch_src(proto, 0, m));
1244 }
1245 
1246 #ifdef DEVICE_POLLING
1247 /*
1248  * Kernel polling borrows a netisr thread to run interface polling in; this
1249  * function allows kernel polling to request that the netisr thread be
1250  * scheduled even if no packets are pending for protocols.
1251  */
1252 void
1253 netisr_sched_poll(void)
1254 {
1255 	struct netisr_workstream *nwsp;
1256 
1257 	nwsp = DPCPU_ID_PTR(nws_array[0], nws);
1258 	NWS_SIGNAL(nwsp);
1259 }
1260 #endif
1261 
1262 static void
1263 netisr_start_swi(u_int cpuid, struct pcpu *pc)
1264 {
1265 	char swiname[12];
1266 	struct netisr_workstream *nwsp;
1267 	int error;
1268 
1269 	KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
1270 
1271 	nwsp = DPCPU_ID_PTR(cpuid, nws);
1272 	mtx_init(&nwsp->nws_mtx, "netisr_mtx", NULL, MTX_DEF);
1273 	nwsp->nws_cpu = cpuid;
1274 	snprintf(swiname, sizeof(swiname), "netisr %u", cpuid);
1275 	error = swi_add(&nwsp->nws_intr_event, swiname, swi_net, nwsp,
1276 	    SWI_NET, INTR_TYPE_NET | INTR_MPSAFE, &nwsp->nws_swi_cookie);
1277 	if (error)
1278 		panic("%s: swi_add %d", __func__, error);
1279 	pc->pc_netisr = nwsp->nws_intr_event;
1280 	if (netisr_bindthreads) {
1281 		error = intr_event_bind(nwsp->nws_intr_event, cpuid);
1282 		if (error != 0)
1283 			printf("%s: cpu %u: intr_event_bind: %d", __func__,
1284 			    cpuid, error);
1285 	}
1286 	NETISR_WLOCK();
1287 	nws_array[nws_count] = nwsp->nws_cpu;
1288 	nws_count++;
1289 	NETISR_WUNLOCK();
1290 }
1291 
1292 /*
1293  * Initialize the netisr subsystem.  We rely on BSS and static initialization
1294  * of most fields in global data structures.
1295  *
1296  * Start a worker thread for the boot CPU so that we can support network
1297  * traffic immediately in case the network stack is used before additional
1298  * CPUs are started (for example, diskless boot).
1299  */
1300 static void
1301 netisr_init(void *arg)
1302 {
1303 	struct pcpu *pc;
1304 
1305 	NETISR_LOCK_INIT();
1306 	if (netisr_maxthreads == 0 || netisr_maxthreads < -1 )
1307 		netisr_maxthreads = 1;		/* default behavior */
1308 	else if (netisr_maxthreads == -1)
1309 		netisr_maxthreads = mp_ncpus;	/* use max cpus */
1310 	if (netisr_maxthreads > mp_ncpus) {
1311 		printf("netisr_init: forcing maxthreads from %d to %d\n",
1312 		    netisr_maxthreads, mp_ncpus);
1313 		netisr_maxthreads = mp_ncpus;
1314 	}
1315 	if (netisr_defaultqlimit > netisr_maxqlimit) {
1316 		printf("netisr_init: forcing defaultqlimit from %d to %d\n",
1317 		    netisr_defaultqlimit, netisr_maxqlimit);
1318 		netisr_defaultqlimit = netisr_maxqlimit;
1319 	}
1320 #ifdef DEVICE_POLLING
1321 	/*
1322 	 * The device polling code is not yet aware of how to deal with
1323 	 * multiple netisr threads, so for the time being compiling in device
1324 	 * polling disables parallel netisr workers.
1325 	 */
1326 	if (netisr_maxthreads != 1 || netisr_bindthreads != 0) {
1327 		printf("netisr_init: forcing maxthreads to 1 and "
1328 		    "bindthreads to 0 for device polling\n");
1329 		netisr_maxthreads = 1;
1330 		netisr_bindthreads = 0;
1331 	}
1332 #endif
1333 
1334 #ifdef EARLY_AP_STARTUP
1335 	STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1336 		if (nws_count >= netisr_maxthreads)
1337 			break;
1338 		netisr_start_swi(pc->pc_cpuid, pc);
1339 	}
1340 #else
1341 	pc = get_pcpu();
1342 	netisr_start_swi(pc->pc_cpuid, pc);
1343 #endif
1344 }
1345 SYSINIT(netisr_init, SI_SUB_SOFTINTR, SI_ORDER_FIRST, netisr_init, NULL);
1346 
1347 #ifndef EARLY_AP_STARTUP
1348 /*
1349  * Start worker threads for additional CPUs.  No attempt to gracefully handle
1350  * work reassignment, we don't yet support dynamic reconfiguration.
1351  */
1352 static void
1353 netisr_start(void *arg)
1354 {
1355 	struct pcpu *pc;
1356 
1357 	STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1358 		if (nws_count >= netisr_maxthreads)
1359 			break;
1360 		/* Worker will already be present for boot CPU. */
1361 		if (pc->pc_netisr != NULL)
1362 			continue;
1363 		netisr_start_swi(pc->pc_cpuid, pc);
1364 	}
1365 }
1366 SYSINIT(netisr_start, SI_SUB_SMP, SI_ORDER_MIDDLE, netisr_start, NULL);
1367 #endif
1368 
1369 /*
1370  * Sysctl monitoring for netisr: query a list of registered protocols.
1371  */
1372 static int
1373 sysctl_netisr_proto(SYSCTL_HANDLER_ARGS)
1374 {
1375 	struct rm_priotracker tracker;
1376 	struct sysctl_netisr_proto *snpp, *snp_array;
1377 	struct netisr_proto *npp;
1378 	u_int counter, proto;
1379 	int error;
1380 
1381 	if (req->newptr != NULL)
1382 		return (EINVAL);
1383 	snp_array = malloc(sizeof(*snp_array) * NETISR_MAXPROT, M_TEMP,
1384 	    M_ZERO | M_WAITOK);
1385 	counter = 0;
1386 	NETISR_RLOCK(&tracker);
1387 	for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1388 		npp = &netisr_proto[proto];
1389 		if (npp->np_name == NULL)
1390 			continue;
1391 		snpp = &snp_array[counter];
1392 		snpp->snp_version = sizeof(*snpp);
1393 		strlcpy(snpp->snp_name, npp->np_name, NETISR_NAMEMAXLEN);
1394 		snpp->snp_proto = proto;
1395 		snpp->snp_qlimit = npp->np_qlimit;
1396 		snpp->snp_policy = npp->np_policy;
1397 		snpp->snp_dispatch = npp->np_dispatch;
1398 		if (npp->np_m2flow != NULL)
1399 			snpp->snp_flags |= NETISR_SNP_FLAGS_M2FLOW;
1400 		if (npp->np_m2cpuid != NULL)
1401 			snpp->snp_flags |= NETISR_SNP_FLAGS_M2CPUID;
1402 		if (npp->np_drainedcpu != NULL)
1403 			snpp->snp_flags |= NETISR_SNP_FLAGS_DRAINEDCPU;
1404 		counter++;
1405 	}
1406 	NETISR_RUNLOCK(&tracker);
1407 	KASSERT(counter <= NETISR_MAXPROT,
1408 	    ("sysctl_netisr_proto: counter too big (%d)", counter));
1409 	error = SYSCTL_OUT(req, snp_array, sizeof(*snp_array) * counter);
1410 	free(snp_array, M_TEMP);
1411 	return (error);
1412 }
1413 
1414 SYSCTL_PROC(_net_isr, OID_AUTO, proto,
1415     CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_proto,
1416     "S,sysctl_netisr_proto",
1417     "Return list of protocols registered with netisr");
1418 
1419 /*
1420  * Sysctl monitoring for netisr: query a list of workstreams.
1421  */
1422 static int
1423 sysctl_netisr_workstream(SYSCTL_HANDLER_ARGS)
1424 {
1425 	struct rm_priotracker tracker;
1426 	struct sysctl_netisr_workstream *snwsp, *snws_array;
1427 	struct netisr_workstream *nwsp;
1428 	u_int counter, cpuid;
1429 	int error;
1430 
1431 	if (req->newptr != NULL)
1432 		return (EINVAL);
1433 	snws_array = malloc(sizeof(*snws_array) * MAXCPU, M_TEMP,
1434 	    M_ZERO | M_WAITOK);
1435 	counter = 0;
1436 	NETISR_RLOCK(&tracker);
1437 	CPU_FOREACH(cpuid) {
1438 		nwsp = DPCPU_ID_PTR(cpuid, nws);
1439 		if (nwsp->nws_intr_event == NULL)
1440 			continue;
1441 		NWS_LOCK(nwsp);
1442 		snwsp = &snws_array[counter];
1443 		snwsp->snws_version = sizeof(*snwsp);
1444 
1445 		/*
1446 		 * For now, we equate workstream IDs and CPU IDs in the
1447 		 * kernel, but expose them independently to userspace in case
1448 		 * that assumption changes in the future.
1449 		 */
1450 		snwsp->snws_wsid = cpuid;
1451 		snwsp->snws_cpu = cpuid;
1452 		if (nwsp->nws_intr_event != NULL)
1453 			snwsp->snws_flags |= NETISR_SNWS_FLAGS_INTR;
1454 		NWS_UNLOCK(nwsp);
1455 		counter++;
1456 	}
1457 	NETISR_RUNLOCK(&tracker);
1458 	KASSERT(counter <= MAXCPU,
1459 	    ("sysctl_netisr_workstream: counter too big (%d)", counter));
1460 	error = SYSCTL_OUT(req, snws_array, sizeof(*snws_array) * counter);
1461 	free(snws_array, M_TEMP);
1462 	return (error);
1463 }
1464 
1465 SYSCTL_PROC(_net_isr, OID_AUTO, workstream,
1466     CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_workstream,
1467     "S,sysctl_netisr_workstream",
1468     "Return list of workstreams implemented by netisr");
1469 
1470 /*
1471  * Sysctl monitoring for netisr: query per-protocol data across all
1472  * workstreams.
1473  */
1474 static int
1475 sysctl_netisr_work(SYSCTL_HANDLER_ARGS)
1476 {
1477 	struct rm_priotracker tracker;
1478 	struct sysctl_netisr_work *snwp, *snw_array;
1479 	struct netisr_workstream *nwsp;
1480 	struct netisr_proto *npp;
1481 	struct netisr_work *nwp;
1482 	u_int counter, cpuid, proto;
1483 	int error;
1484 
1485 	if (req->newptr != NULL)
1486 		return (EINVAL);
1487 	snw_array = malloc(sizeof(*snw_array) * MAXCPU * NETISR_MAXPROT,
1488 	    M_TEMP, M_ZERO | M_WAITOK);
1489 	counter = 0;
1490 	NETISR_RLOCK(&tracker);
1491 	CPU_FOREACH(cpuid) {
1492 		nwsp = DPCPU_ID_PTR(cpuid, nws);
1493 		if (nwsp->nws_intr_event == NULL)
1494 			continue;
1495 		NWS_LOCK(nwsp);
1496 		for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1497 			npp = &netisr_proto[proto];
1498 			if (npp->np_name == NULL)
1499 				continue;
1500 			nwp = &nwsp->nws_work[proto];
1501 			snwp = &snw_array[counter];
1502 			snwp->snw_version = sizeof(*snwp);
1503 			snwp->snw_wsid = cpuid;		/* See comment above. */
1504 			snwp->snw_proto = proto;
1505 			snwp->snw_len = nwp->nw_len;
1506 			snwp->snw_watermark = nwp->nw_watermark;
1507 			snwp->snw_dispatched = nwp->nw_dispatched;
1508 			snwp->snw_hybrid_dispatched =
1509 			    nwp->nw_hybrid_dispatched;
1510 			snwp->snw_qdrops = nwp->nw_qdrops;
1511 			snwp->snw_queued = nwp->nw_queued;
1512 			snwp->snw_handled = nwp->nw_handled;
1513 			counter++;
1514 		}
1515 		NWS_UNLOCK(nwsp);
1516 	}
1517 	KASSERT(counter <= MAXCPU * NETISR_MAXPROT,
1518 	    ("sysctl_netisr_work: counter too big (%d)", counter));
1519 	NETISR_RUNLOCK(&tracker);
1520 	error = SYSCTL_OUT(req, snw_array, sizeof(*snw_array) * counter);
1521 	free(snw_array, M_TEMP);
1522 	return (error);
1523 }
1524 
1525 SYSCTL_PROC(_net_isr, OID_AUTO, work,
1526     CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_work,
1527     "S,sysctl_netisr_work",
1528     "Return list of per-workstream, per-protocol work in netisr");
1529 
1530 #ifdef DDB
1531 DB_SHOW_COMMAND(netisr, db_show_netisr)
1532 {
1533 	struct netisr_workstream *nwsp;
1534 	struct netisr_work *nwp;
1535 	int first, proto;
1536 	u_int cpuid;
1537 
1538 	db_printf("%3s %6s %5s %5s %5s %8s %8s %8s %8s\n", "CPU", "Proto",
1539 	    "Len", "WMark", "Max", "Disp", "HDisp", "Drop", "Queue");
1540 	CPU_FOREACH(cpuid) {
1541 		nwsp = DPCPU_ID_PTR(cpuid, nws);
1542 		if (nwsp->nws_intr_event == NULL)
1543 			continue;
1544 		first = 1;
1545 		for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1546 			if (netisr_proto[proto].np_handler == NULL)
1547 				continue;
1548 			nwp = &nwsp->nws_work[proto];
1549 			if (first) {
1550 				db_printf("%3d ", cpuid);
1551 				first = 0;
1552 			} else
1553 				db_printf("%3s ", "");
1554 			db_printf(
1555 			    "%6s %5d %5d %5d %8ju %8ju %8ju %8ju\n",
1556 			    netisr_proto[proto].np_name, nwp->nw_len,
1557 			    nwp->nw_watermark, nwp->nw_qlimit,
1558 			    nwp->nw_dispatched, nwp->nw_hybrid_dispatched,
1559 			    nwp->nw_qdrops, nwp->nw_queued);
1560 		}
1561 	}
1562 }
1563 #endif
1564