xref: /freebsd/sys/net/netisr.c (revision 10dd04d39f4ac13cca68ed83ba17a93791190565)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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 __FBSDID("$FreeBSD$");
35 
36 /*
37  * netisr is a packet dispatch service, allowing synchronous (directly
38  * dispatched) and asynchronous (deferred dispatch) processing of packets by
39  * registered protocol handlers.  Callers pass a protocol identifier and
40  * packet to netisr, along with a direct dispatch hint, and work will either
41  * be immediately processed by the registered handler, or passed to a
42  * software interrupt (SWI) thread for deferred dispatch.  Callers will
43  * generally select one or the other based on:
44  *
45  * - Whether directly dispatching a netisr handler lead to code reentrance or
46  *   lock recursion, such as entering the socket code from the socket code.
47  * - Whether directly dispatching a netisr handler lead to recursive
48  *   processing, such as when decapsulating several wrapped layers of tunnel
49  *   information (IPSEC within IPSEC within ...).
50  *
51  * Maintaining ordering for protocol streams is a critical design concern.
52  * Enforcing ordering limits the opportunity for concurrency, but maintains
53  * the strong ordering requirements found in some protocols, such as TCP.  Of
54  * related concern is CPU affinity--it is desirable to process all data
55  * associated with a particular stream on the same CPU over time in order to
56  * avoid acquiring locks associated with the connection on different CPUs,
57  * keep connection data in one cache, and to generally encourage associated
58  * user threads to live on the same CPU as the stream.  It's also desirable
59  * to avoid lock migration and contention where locks are associated with
60  * more than one flow.
61  *
62  * netisr supports several policy variations, represented by the
63  * NETISR_POLICY_* constants, allowing protocols to play various roles in
64  * identifying flows, assigning work to CPUs, etc.  These are described in
65  * netisr.h.
66  */
67 
68 #include "opt_ddb.h"
69 #include "opt_device_polling.h"
70 
71 #include <sys/param.h>
72 #include <sys/bus.h>
73 #include <sys/kernel.h>
74 #include <sys/kthread.h>
75 #include <sys/malloc.h>
76 #include <sys/interrupt.h>
77 #include <sys/lock.h>
78 #include <sys/mbuf.h>
79 #include <sys/mutex.h>
80 #include <sys/pcpu.h>
81 #include <sys/proc.h>
82 #include <sys/rmlock.h>
83 #include <sys/sched.h>
84 #include <sys/smp.h>
85 #include <sys/socket.h>
86 #include <sys/sysctl.h>
87 #include <sys/systm.h>
88 
89 #ifdef DDB
90 #include <ddb/ddb.h>
91 #endif
92 
93 #define	_WANT_NETISR_INTERNAL	/* Enable definitions from netisr_internal.h */
94 #include <net/if.h>
95 #include <net/if_var.h>
96 #include <net/netisr.h>
97 #include <net/netisr_internal.h>
98 #include <net/vnet.h>
99 
100 /*-
101  * Synchronize use and modification of the registered netisr data structures;
102  * acquire a read lock while modifying the set of registered protocols to
103  * prevent partially registered or unregistered protocols from being run.
104  *
105  * The following data structures and fields are protected by this lock:
106  *
107  * - The netisr_proto array, including all fields of struct netisr_proto.
108  * - The nws array, including all fields of struct netisr_worker.
109  * - The nws_array array.
110  *
111  * Note: the NETISR_LOCKING define controls whether read locks are acquired
112  * in packet processing paths requiring netisr registration stability.  This
113  * is disabled by default as it can lead to measurable performance
114  * degradation even with rmlocks (3%-6% for loopback ping-pong traffic), and
115  * because netisr registration and unregistration is extremely rare at
116  * runtime.  If it becomes more common, this decision should be revisited.
117  *
118  * XXXRW: rmlocks don't support assertions.
119  */
120 static struct rmlock	netisr_rmlock;
121 #define	NETISR_LOCK_INIT()	rm_init_flags(&netisr_rmlock, "netisr", \
122 				    RM_NOWITNESS)
123 #define	NETISR_LOCK_ASSERT()
124 #define	NETISR_RLOCK(tracker)	rm_rlock(&netisr_rmlock, (tracker))
125 #define	NETISR_RUNLOCK(tracker)	rm_runlock(&netisr_rmlock, (tracker))
126 #define	NETISR_WLOCK()		rm_wlock(&netisr_rmlock)
127 #define	NETISR_WUNLOCK()	rm_wunlock(&netisr_rmlock)
128 /* #define	NETISR_LOCKING */
129 
130 static SYSCTL_NODE(_net, OID_AUTO, isr, CTLFLAG_RW, 0, "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 overriden 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, CTLTYPE_STRING | CTLFLAG_RWTUN,
156     0, 0, sysctl_netisr_dispatch_policy, "A",
157     "netisr dispatch policy");
158 
159 /*
160  * Allow the administrator to limit the number of threads (CPUs) to use for
161  * netisr.  We don't check netisr_maxthreads before creating the thread for
162  * CPU 0. This must be set at boot. We will create at most one thread per CPU.
163  * By default we initialize this to 1 which would assign just 1 cpu (cpu0) and
164  * therefore only 1 workstream. If set to -1, netisr would use all cpus
165  * (mp_ncpus) and therefore would have those many workstreams. One workstream
166  * per thread (CPU).
167  */
168 static int	netisr_maxthreads = 1;		/* Max number of threads. */
169 SYSCTL_INT(_net_isr, OID_AUTO, maxthreads, CTLFLAG_RDTUN,
170     &netisr_maxthreads, 0,
171     "Use at most this many CPUs for netisr processing");
172 
173 static int	netisr_bindthreads = 0;		/* Bind threads to CPUs. */
174 SYSCTL_INT(_net_isr, OID_AUTO, bindthreads, CTLFLAG_RDTUN,
175     &netisr_bindthreads, 0, "Bind netisr threads to CPUs.");
176 
177 /*
178  * Limit per-workstream mbuf queue limits s to at most net.isr.maxqlimit,
179  * both for initial configuration and later modification using
180  * netisr_setqlimit().
181  */
182 #define	NETISR_DEFAULT_MAXQLIMIT	10240
183 static u_int	netisr_maxqlimit = NETISR_DEFAULT_MAXQLIMIT;
184 SYSCTL_UINT(_net_isr, OID_AUTO, maxqlimit, CTLFLAG_RDTUN,
185     &netisr_maxqlimit, 0,
186     "Maximum netisr per-protocol, per-CPU queue depth.");
187 
188 /*
189  * The default per-workstream mbuf queue limit for protocols that don't
190  * initialize the nh_qlimit field of their struct netisr_handler.  If this is
191  * set above netisr_maxqlimit, we truncate it to the maximum during boot.
192  */
193 #define	NETISR_DEFAULT_DEFAULTQLIMIT	256
194 static u_int	netisr_defaultqlimit = NETISR_DEFAULT_DEFAULTQLIMIT;
195 SYSCTL_UINT(_net_isr, OID_AUTO, defaultqlimit, CTLFLAG_RDTUN,
196     &netisr_defaultqlimit, 0,
197     "Default netisr per-protocol, per-CPU queue limit if not set by protocol");
198 
199 /*
200  * Store and export the compile-time constant NETISR_MAXPROT limit on the
201  * number of protocols that can register with netisr at a time.  This is
202  * required for crashdump analysis, as it sizes netisr_proto[].
203  */
204 static u_int	netisr_maxprot = NETISR_MAXPROT;
205 SYSCTL_UINT(_net_isr, OID_AUTO, maxprot, CTLFLAG_RD,
206     &netisr_maxprot, 0,
207     "Compile-time limit on the number of protocols supported by netisr.");
208 
209 /*
210  * The netisr_proto array describes all registered protocols, indexed by
211  * protocol number.  See netisr_internal.h for more details.
212  */
213 static struct netisr_proto	netisr_proto[NETISR_MAXPROT];
214 
215 #ifdef VIMAGE
216 /*
217  * The netisr_enable array describes a per-VNET flag for registered
218  * protocols on whether this netisr is active in this VNET or not.
219  * netisr_register() will automatically enable the netisr for the
220  * default VNET and all currently active instances.
221  * netisr_unregister() will disable all active VNETs, including vnet0.
222  * Individual network stack instances can be enabled/disabled by the
223  * netisr_(un)register _vnet() functions.
224  * With this we keep the one netisr_proto per protocol but add a
225  * mechanism to stop netisr processing for vnet teardown.
226  * Apart from that we expect a VNET to always be enabled.
227  */
228 VNET_DEFINE_STATIC(u_int,	netisr_enable[NETISR_MAXPROT]);
229 #define	V_netisr_enable		VNET(netisr_enable)
230 #endif
231 
232 /*
233  * Per-CPU workstream data.  See netisr_internal.h for more details.
234  */
235 DPCPU_DEFINE(struct netisr_workstream, nws);
236 
237 /*
238  * Map contiguous values between 0 and nws_count into CPU IDs appropriate for
239  * accessing workstreams.  This allows constructions of the form
240  * DPCPU_ID_GET(nws_array[arbitraryvalue % nws_count], nws).
241  */
242 static u_int				 nws_array[MAXCPU];
243 
244 /*
245  * Number of registered workstreams.  Will be at most the number of running
246  * CPUs once fully started.
247  */
248 static u_int				 nws_count;
249 SYSCTL_UINT(_net_isr, OID_AUTO, numthreads, CTLFLAG_RD,
250     &nws_count, 0, "Number of extant netisr threads.");
251 
252 /*
253  * Synchronization for each workstream: a mutex protects all mutable fields
254  * in each stream, including per-protocol state (mbuf queues).  The SWI is
255  * woken up if asynchronous dispatch is required.
256  */
257 #define	NWS_LOCK(s)		mtx_lock(&(s)->nws_mtx)
258 #define	NWS_LOCK_ASSERT(s)	mtx_assert(&(s)->nws_mtx, MA_OWNED)
259 #define	NWS_UNLOCK(s)		mtx_unlock(&(s)->nws_mtx)
260 #define	NWS_SIGNAL(s)		swi_sched((s)->nws_swi_cookie, 0)
261 
262 /*
263  * Utility routines for protocols that implement their own mapping of flows
264  * to CPUs.
265  */
266 u_int
267 netisr_get_cpucount(void)
268 {
269 
270 	return (nws_count);
271 }
272 
273 u_int
274 netisr_get_cpuid(u_int cpunumber)
275 {
276 
277 	return (nws_array[cpunumber % nws_count]);
278 }
279 
280 /*
281  * The default implementation of flow -> CPU ID mapping.
282  *
283  * Non-static so that protocols can use it to map their own work to specific
284  * CPUs in a manner consistent to netisr for affinity purposes.
285  */
286 u_int
287 netisr_default_flow2cpu(u_int flowid)
288 {
289 
290 	return (nws_array[flowid % nws_count]);
291 }
292 
293 /*
294  * Dispatch tunable and sysctl configuration.
295  */
296 struct netisr_dispatch_table_entry {
297 	u_int		 ndte_policy;
298 	const char	*ndte_policy_str;
299 };
300 static const struct netisr_dispatch_table_entry netisr_dispatch_table[] = {
301 	{ NETISR_DISPATCH_DEFAULT, "default" },
302 	{ NETISR_DISPATCH_DEFERRED, "deferred" },
303 	{ NETISR_DISPATCH_HYBRID, "hybrid" },
304 	{ NETISR_DISPATCH_DIRECT, "direct" },
305 };
306 
307 static void
308 netisr_dispatch_policy_to_str(u_int dispatch_policy, char *buffer,
309     u_int buflen)
310 {
311 	const struct netisr_dispatch_table_entry *ndtep;
312 	const char *str;
313 	u_int i;
314 
315 	str = "unknown";
316 	for (i = 0; i < nitems(netisr_dispatch_table); i++) {
317 		ndtep = &netisr_dispatch_table[i];
318 		if (ndtep->ndte_policy == dispatch_policy) {
319 			str = ndtep->ndte_policy_str;
320 			break;
321 		}
322 	}
323 	snprintf(buffer, buflen, "%s", str);
324 }
325 
326 static int
327 netisr_dispatch_policy_from_str(const char *str, u_int *dispatch_policyp)
328 {
329 	const struct netisr_dispatch_table_entry *ndtep;
330 	u_int i;
331 
332 	for (i = 0; i < nitems(netisr_dispatch_table); i++) {
333 		ndtep = &netisr_dispatch_table[i];
334 		if (strcmp(ndtep->ndte_policy_str, str) == 0) {
335 			*dispatch_policyp = ndtep->ndte_policy;
336 			return (0);
337 		}
338 	}
339 	return (EINVAL);
340 }
341 
342 static int
343 sysctl_netisr_dispatch_policy(SYSCTL_HANDLER_ARGS)
344 {
345 	char tmp[NETISR_DISPATCH_POLICY_MAXSTR];
346 	u_int dispatch_policy;
347 	int error;
348 
349 	netisr_dispatch_policy_to_str(netisr_dispatch_policy, tmp,
350 	    sizeof(tmp));
351 	error = sysctl_handle_string(oidp, tmp, sizeof(tmp), req);
352 	if (error == 0 && req->newptr != NULL) {
353 		error = netisr_dispatch_policy_from_str(tmp,
354 		    &dispatch_policy);
355 		if (error == 0 && dispatch_policy == NETISR_DISPATCH_DEFAULT)
356 			error = EINVAL;
357 		if (error == 0)
358 			netisr_dispatch_policy = dispatch_policy;
359 	}
360 	return (error);
361 }
362 
363 /*
364  * Register a new netisr handler, which requires initializing per-protocol
365  * fields for each workstream.  All netisr work is briefly suspended while
366  * the protocol is installed.
367  */
368 void
369 netisr_register(const struct netisr_handler *nhp)
370 {
371 	VNET_ITERATOR_DECL(vnet_iter);
372 	struct netisr_work *npwp;
373 	const char *name;
374 	u_int i, proto;
375 
376 	proto = nhp->nh_proto;
377 	name = nhp->nh_name;
378 
379 	/*
380 	 * Test that the requested registration is valid.
381 	 */
382 	KASSERT(nhp->nh_name != NULL,
383 	    ("%s: nh_name NULL for %u", __func__, proto));
384 	KASSERT(nhp->nh_handler != NULL,
385 	    ("%s: nh_handler NULL for %s", __func__, name));
386 	KASSERT(nhp->nh_policy == NETISR_POLICY_SOURCE ||
387 	    nhp->nh_policy == NETISR_POLICY_FLOW ||
388 	    nhp->nh_policy == NETISR_POLICY_CPU,
389 	    ("%s: unsupported nh_policy %u for %s", __func__,
390 	    nhp->nh_policy, name));
391 	KASSERT(nhp->nh_policy == NETISR_POLICY_FLOW ||
392 	    nhp->nh_m2flow == NULL,
393 	    ("%s: nh_policy != FLOW but m2flow defined for %s", __func__,
394 	    name));
395 	KASSERT(nhp->nh_policy == NETISR_POLICY_CPU || nhp->nh_m2cpuid == NULL,
396 	    ("%s: nh_policy != CPU but m2cpuid defined for %s", __func__,
397 	    name));
398 	KASSERT(nhp->nh_policy != NETISR_POLICY_CPU || nhp->nh_m2cpuid != NULL,
399 	    ("%s: nh_policy == CPU but m2cpuid not defined for %s", __func__,
400 	    name));
401 	KASSERT(nhp->nh_dispatch == NETISR_DISPATCH_DEFAULT ||
402 	    nhp->nh_dispatch == NETISR_DISPATCH_DEFERRED ||
403 	    nhp->nh_dispatch == NETISR_DISPATCH_HYBRID ||
404 	    nhp->nh_dispatch == NETISR_DISPATCH_DIRECT,
405 	    ("%s: invalid nh_dispatch (%u)", __func__, nhp->nh_dispatch));
406 
407 	KASSERT(proto < NETISR_MAXPROT,
408 	    ("%s(%u, %s): protocol too big", __func__, proto, name));
409 
410 	/*
411 	 * Test that no existing registration exists for this protocol.
412 	 */
413 	NETISR_WLOCK();
414 	KASSERT(netisr_proto[proto].np_name == NULL,
415 	    ("%s(%u, %s): name present", __func__, proto, name));
416 	KASSERT(netisr_proto[proto].np_handler == NULL,
417 	    ("%s(%u, %s): handler present", __func__, proto, name));
418 
419 	netisr_proto[proto].np_name = name;
420 	netisr_proto[proto].np_handler = nhp->nh_handler;
421 	netisr_proto[proto].np_m2flow = nhp->nh_m2flow;
422 	netisr_proto[proto].np_m2cpuid = nhp->nh_m2cpuid;
423 	netisr_proto[proto].np_drainedcpu = nhp->nh_drainedcpu;
424 	if (nhp->nh_qlimit == 0)
425 		netisr_proto[proto].np_qlimit = netisr_defaultqlimit;
426 	else if (nhp->nh_qlimit > netisr_maxqlimit) {
427 		printf("%s: %s requested queue limit %u capped to "
428 		    "net.isr.maxqlimit %u\n", __func__, name, nhp->nh_qlimit,
429 		    netisr_maxqlimit);
430 		netisr_proto[proto].np_qlimit = netisr_maxqlimit;
431 	} else
432 		netisr_proto[proto].np_qlimit = nhp->nh_qlimit;
433 	netisr_proto[proto].np_policy = nhp->nh_policy;
434 	netisr_proto[proto].np_dispatch = nhp->nh_dispatch;
435 	CPU_FOREACH(i) {
436 		npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
437 		bzero(npwp, sizeof(*npwp));
438 		npwp->nw_qlimit = netisr_proto[proto].np_qlimit;
439 	}
440 
441 #ifdef VIMAGE
442 	/*
443 	 * Test that we are in vnet0 and have a curvnet set.
444 	 */
445 	KASSERT(curvnet != NULL, ("%s: curvnet is NULL", __func__));
446 	KASSERT(IS_DEFAULT_VNET(curvnet), ("%s: curvnet %p is not vnet0 %p",
447 	    __func__, curvnet, vnet0));
448 	VNET_LIST_RLOCK_NOSLEEP();
449 	VNET_FOREACH(vnet_iter) {
450 		CURVNET_SET(vnet_iter);
451 		V_netisr_enable[proto] = 1;
452 		CURVNET_RESTORE();
453 	}
454 	VNET_LIST_RUNLOCK_NOSLEEP();
455 #endif
456 	NETISR_WUNLOCK();
457 }
458 
459 /*
460  * Clear drop counters across all workstreams for a protocol.
461  */
462 void
463 netisr_clearqdrops(const struct netisr_handler *nhp)
464 {
465 	struct netisr_work *npwp;
466 #ifdef INVARIANTS
467 	const char *name;
468 #endif
469 	u_int i, proto;
470 
471 	proto = nhp->nh_proto;
472 #ifdef INVARIANTS
473 	name = nhp->nh_name;
474 #endif
475 	KASSERT(proto < NETISR_MAXPROT,
476 	    ("%s(%u): protocol too big for %s", __func__, proto, name));
477 
478 	NETISR_WLOCK();
479 	KASSERT(netisr_proto[proto].np_handler != NULL,
480 	    ("%s(%u): protocol not registered for %s", __func__, proto,
481 	    name));
482 
483 	CPU_FOREACH(i) {
484 		npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
485 		npwp->nw_qdrops = 0;
486 	}
487 	NETISR_WUNLOCK();
488 }
489 
490 /*
491  * Query current drop counters across all workstreams for a protocol.
492  */
493 void
494 netisr_getqdrops(const struct netisr_handler *nhp, u_int64_t *qdropp)
495 {
496 	struct netisr_work *npwp;
497 	struct rm_priotracker tracker;
498 #ifdef INVARIANTS
499 	const char *name;
500 #endif
501 	u_int i, proto;
502 
503 	*qdropp = 0;
504 	proto = nhp->nh_proto;
505 #ifdef INVARIANTS
506 	name = nhp->nh_name;
507 #endif
508 	KASSERT(proto < NETISR_MAXPROT,
509 	    ("%s(%u): protocol too big for %s", __func__, proto, name));
510 
511 	NETISR_RLOCK(&tracker);
512 	KASSERT(netisr_proto[proto].np_handler != NULL,
513 	    ("%s(%u): protocol not registered for %s", __func__, proto,
514 	    name));
515 
516 	CPU_FOREACH(i) {
517 		npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
518 		*qdropp += npwp->nw_qdrops;
519 	}
520 	NETISR_RUNLOCK(&tracker);
521 }
522 
523 /*
524  * Query current per-workstream queue limit for a protocol.
525  */
526 void
527 netisr_getqlimit(const struct netisr_handler *nhp, u_int *qlimitp)
528 {
529 	struct rm_priotracker tracker;
530 #ifdef INVARIANTS
531 	const char *name;
532 #endif
533 	u_int proto;
534 
535 	proto = nhp->nh_proto;
536 #ifdef INVARIANTS
537 	name = nhp->nh_name;
538 #endif
539 	KASSERT(proto < NETISR_MAXPROT,
540 	    ("%s(%u): protocol too big for %s", __func__, proto, name));
541 
542 	NETISR_RLOCK(&tracker);
543 	KASSERT(netisr_proto[proto].np_handler != NULL,
544 	    ("%s(%u): protocol not registered for %s", __func__, proto,
545 	    name));
546 	*qlimitp = netisr_proto[proto].np_qlimit;
547 	NETISR_RUNLOCK(&tracker);
548 }
549 
550 /*
551  * Update the queue limit across per-workstream queues for a protocol.  We
552  * simply change the limits, and don't drain overflowed packets as they will
553  * (hopefully) take care of themselves shortly.
554  */
555 int
556 netisr_setqlimit(const struct netisr_handler *nhp, u_int qlimit)
557 {
558 	struct netisr_work *npwp;
559 #ifdef INVARIANTS
560 	const char *name;
561 #endif
562 	u_int i, proto;
563 
564 	if (qlimit > netisr_maxqlimit)
565 		return (EINVAL);
566 
567 	proto = nhp->nh_proto;
568 #ifdef INVARIANTS
569 	name = nhp->nh_name;
570 #endif
571 	KASSERT(proto < NETISR_MAXPROT,
572 	    ("%s(%u): protocol too big for %s", __func__, proto, name));
573 
574 	NETISR_WLOCK();
575 	KASSERT(netisr_proto[proto].np_handler != NULL,
576 	    ("%s(%u): protocol not registered for %s", __func__, proto,
577 	    name));
578 
579 	netisr_proto[proto].np_qlimit = qlimit;
580 	CPU_FOREACH(i) {
581 		npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
582 		npwp->nw_qlimit = qlimit;
583 	}
584 	NETISR_WUNLOCK();
585 	return (0);
586 }
587 
588 /*
589  * Drain all packets currently held in a particular protocol work queue.
590  */
591 static void
592 netisr_drain_proto(struct netisr_work *npwp)
593 {
594 	struct mbuf *m;
595 
596 	/*
597 	 * We would assert the lock on the workstream but it's not passed in.
598 	 */
599 	while ((m = npwp->nw_head) != NULL) {
600 		npwp->nw_head = m->m_nextpkt;
601 		m->m_nextpkt = NULL;
602 		if (npwp->nw_head == NULL)
603 			npwp->nw_tail = NULL;
604 		npwp->nw_len--;
605 		m_freem(m);
606 	}
607 	KASSERT(npwp->nw_tail == NULL, ("%s: tail", __func__));
608 	KASSERT(npwp->nw_len == 0, ("%s: len", __func__));
609 }
610 
611 /*
612  * Remove the registration of a network protocol, which requires clearing
613  * per-protocol fields across all workstreams, including freeing all mbufs in
614  * the queues at time of unregister.  All work in netisr is briefly suspended
615  * while this takes place.
616  */
617 void
618 netisr_unregister(const struct netisr_handler *nhp)
619 {
620 	VNET_ITERATOR_DECL(vnet_iter);
621 	struct netisr_work *npwp;
622 #ifdef INVARIANTS
623 	const char *name;
624 #endif
625 	u_int i, proto;
626 
627 	proto = nhp->nh_proto;
628 #ifdef INVARIANTS
629 	name = nhp->nh_name;
630 #endif
631 	KASSERT(proto < NETISR_MAXPROT,
632 	    ("%s(%u): protocol too big for %s", __func__, proto, name));
633 
634 	NETISR_WLOCK();
635 	KASSERT(netisr_proto[proto].np_handler != NULL,
636 	    ("%s(%u): protocol not registered for %s", __func__, proto,
637 	    name));
638 
639 #ifdef VIMAGE
640 	VNET_LIST_RLOCK_NOSLEEP();
641 	VNET_FOREACH(vnet_iter) {
642 		CURVNET_SET(vnet_iter);
643 		V_netisr_enable[proto] = 0;
644 		CURVNET_RESTORE();
645 	}
646 	VNET_LIST_RUNLOCK_NOSLEEP();
647 #endif
648 
649 	netisr_proto[proto].np_name = NULL;
650 	netisr_proto[proto].np_handler = NULL;
651 	netisr_proto[proto].np_m2flow = NULL;
652 	netisr_proto[proto].np_m2cpuid = NULL;
653 	netisr_proto[proto].np_qlimit = 0;
654 	netisr_proto[proto].np_policy = 0;
655 	CPU_FOREACH(i) {
656 		npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
657 		netisr_drain_proto(npwp);
658 		bzero(npwp, sizeof(*npwp));
659 	}
660 	NETISR_WUNLOCK();
661 }
662 
663 #ifdef VIMAGE
664 void
665 netisr_register_vnet(const struct netisr_handler *nhp)
666 {
667 	u_int proto;
668 
669 	proto = nhp->nh_proto;
670 
671 	KASSERT(curvnet != NULL, ("%s: curvnet is NULL", __func__));
672 	KASSERT(proto < NETISR_MAXPROT,
673 	    ("%s(%u): protocol too big for %s", __func__, proto, nhp->nh_name));
674 	NETISR_WLOCK();
675 	KASSERT(netisr_proto[proto].np_handler != NULL,
676 	    ("%s(%u): protocol not registered for %s", __func__, proto,
677 	    nhp->nh_name));
678 
679 	V_netisr_enable[proto] = 1;
680 	NETISR_WUNLOCK();
681 }
682 
683 static void
684 netisr_drain_proto_vnet(struct vnet *vnet, u_int proto)
685 {
686 	struct netisr_workstream *nwsp;
687 	struct netisr_work *npwp;
688 	struct mbuf *m, *mp, *n, *ne;
689 	u_int i;
690 
691 	KASSERT(vnet != NULL, ("%s: vnet is NULL", __func__));
692 	NETISR_LOCK_ASSERT();
693 
694 	CPU_FOREACH(i) {
695 		nwsp = DPCPU_ID_PTR(i, nws);
696 		if (nwsp->nws_intr_event == NULL)
697 			continue;
698 		npwp = &nwsp->nws_work[proto];
699 		NWS_LOCK(nwsp);
700 
701 		/*
702 		 * Rather than dissecting and removing mbufs from the middle
703 		 * of the chain, we build a new chain if the packet stays and
704 		 * update the head and tail pointers at the end.  All packets
705 		 * matching the given vnet are freed.
706 		 */
707 		m = npwp->nw_head;
708 		n = ne = NULL;
709 		while (m != NULL) {
710 			mp = m;
711 			m = m->m_nextpkt;
712 			mp->m_nextpkt = NULL;
713 			if (mp->m_pkthdr.rcvif->if_vnet != vnet) {
714 				if (n == NULL) {
715 					n = ne = mp;
716 				} else {
717 					ne->m_nextpkt = mp;
718 					ne = mp;
719 				}
720 				continue;
721 			}
722 			/* This is a packet in the selected vnet. Free it. */
723 			npwp->nw_len--;
724 			m_freem(mp);
725 		}
726 		npwp->nw_head = n;
727 		npwp->nw_tail = ne;
728 		NWS_UNLOCK(nwsp);
729 	}
730 }
731 
732 void
733 netisr_unregister_vnet(const struct netisr_handler *nhp)
734 {
735 	u_int proto;
736 
737 	proto = nhp->nh_proto;
738 
739 	KASSERT(curvnet != NULL, ("%s: curvnet is NULL", __func__));
740 	KASSERT(proto < NETISR_MAXPROT,
741 	    ("%s(%u): protocol too big for %s", __func__, proto, nhp->nh_name));
742 	NETISR_WLOCK();
743 	KASSERT(netisr_proto[proto].np_handler != NULL,
744 	    ("%s(%u): protocol not registered for %s", __func__, proto,
745 	    nhp->nh_name));
746 
747 	V_netisr_enable[proto] = 0;
748 
749 	netisr_drain_proto_vnet(curvnet, proto);
750 	NETISR_WUNLOCK();
751 }
752 #endif
753 
754 /*
755  * Compose the global and per-protocol policies on dispatch, and return the
756  * dispatch policy to use.
757  */
758 static u_int
759 netisr_get_dispatch(struct netisr_proto *npp)
760 {
761 
762 	/*
763 	 * Protocol-specific configuration overrides the global default.
764 	 */
765 	if (npp->np_dispatch != NETISR_DISPATCH_DEFAULT)
766 		return (npp->np_dispatch);
767 	return (netisr_dispatch_policy);
768 }
769 
770 /*
771  * Look up the workstream given a packet and source identifier.  Do this by
772  * checking the protocol's policy, and optionally call out to the protocol
773  * for assistance if required.
774  */
775 static struct mbuf *
776 netisr_select_cpuid(struct netisr_proto *npp, u_int dispatch_policy,
777     uintptr_t source, struct mbuf *m, u_int *cpuidp)
778 {
779 	struct ifnet *ifp;
780 	u_int policy;
781 
782 	NETISR_LOCK_ASSERT();
783 
784 	/*
785 	 * In the event we have only one worker, shortcut and deliver to it
786 	 * without further ado.
787 	 */
788 	if (nws_count == 1) {
789 		*cpuidp = nws_array[0];
790 		return (m);
791 	}
792 
793 	/*
794 	 * What happens next depends on the policy selected by the protocol.
795 	 * If we want to support per-interface policies, we should do that
796 	 * here first.
797 	 */
798 	policy = npp->np_policy;
799 	if (policy == NETISR_POLICY_CPU) {
800 		m = npp->np_m2cpuid(m, source, cpuidp);
801 		if (m == NULL)
802 			return (NULL);
803 
804 		/*
805 		 * It's possible for a protocol not to have a good idea about
806 		 * where to process a packet, in which case we fall back on
807 		 * the netisr code to decide.  In the hybrid case, return the
808 		 * current CPU ID, which will force an immediate direct
809 		 * dispatch.  In the queued case, fall back on the SOURCE
810 		 * policy.
811 		 */
812 		if (*cpuidp != NETISR_CPUID_NONE) {
813 			*cpuidp = netisr_get_cpuid(*cpuidp);
814 			return (m);
815 		}
816 		if (dispatch_policy == NETISR_DISPATCH_HYBRID) {
817 			*cpuidp = netisr_get_cpuid(curcpu);
818 			return (m);
819 		}
820 		policy = NETISR_POLICY_SOURCE;
821 	}
822 
823 	if (policy == NETISR_POLICY_FLOW) {
824 		if (M_HASHTYPE_GET(m) == M_HASHTYPE_NONE &&
825 		    npp->np_m2flow != NULL) {
826 			m = npp->np_m2flow(m, source);
827 			if (m == NULL)
828 				return (NULL);
829 		}
830 		if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
831 			*cpuidp =
832 			    netisr_default_flow2cpu(m->m_pkthdr.flowid);
833 			return (m);
834 		}
835 		policy = NETISR_POLICY_SOURCE;
836 	}
837 
838 	KASSERT(policy == NETISR_POLICY_SOURCE,
839 	    ("%s: invalid policy %u for %s", __func__, npp->np_policy,
840 	    npp->np_name));
841 
842 	MPASS((m->m_pkthdr.csum_flags & CSUM_SND_TAG) == 0);
843 	ifp = m->m_pkthdr.rcvif;
844 	if (ifp != NULL)
845 		*cpuidp = nws_array[(ifp->if_index + source) % nws_count];
846 	else
847 		*cpuidp = nws_array[source % nws_count];
848 	return (m);
849 }
850 
851 /*
852  * Process packets associated with a workstream and protocol.  For reasons of
853  * fairness, we process up to one complete netisr queue at a time, moving the
854  * queue to a stack-local queue for processing, but do not loop refreshing
855  * from the global queue.  The caller is responsible for deciding whether to
856  * loop, and for setting the NWS_RUNNING flag.  The passed workstream will be
857  * locked on entry and relocked before return, but will be released while
858  * processing.  The number of packets processed is returned.
859  */
860 static u_int
861 netisr_process_workstream_proto(struct netisr_workstream *nwsp, u_int proto)
862 {
863 	struct netisr_work local_npw, *npwp;
864 	u_int handled;
865 	struct mbuf *m;
866 
867 	NETISR_LOCK_ASSERT();
868 	NWS_LOCK_ASSERT(nwsp);
869 
870 	KASSERT(nwsp->nws_flags & NWS_RUNNING,
871 	    ("%s(%u): not running", __func__, proto));
872 	KASSERT(proto >= 0 && proto < NETISR_MAXPROT,
873 	    ("%s(%u): invalid proto\n", __func__, proto));
874 
875 	npwp = &nwsp->nws_work[proto];
876 	if (npwp->nw_len == 0)
877 		return (0);
878 
879 	/*
880 	 * Move the global work queue to a thread-local work queue.
881 	 *
882 	 * Notice that this means the effective maximum length of the queue
883 	 * is actually twice that of the maximum queue length specified in
884 	 * the protocol registration call.
885 	 */
886 	handled = npwp->nw_len;
887 	local_npw = *npwp;
888 	npwp->nw_head = NULL;
889 	npwp->nw_tail = NULL;
890 	npwp->nw_len = 0;
891 	nwsp->nws_pendingbits &= ~(1 << proto);
892 	NWS_UNLOCK(nwsp);
893 	while ((m = local_npw.nw_head) != NULL) {
894 		local_npw.nw_head = m->m_nextpkt;
895 		m->m_nextpkt = NULL;
896 		if (local_npw.nw_head == NULL)
897 			local_npw.nw_tail = NULL;
898 		local_npw.nw_len--;
899 		VNET_ASSERT(m->m_pkthdr.rcvif != NULL,
900 		    ("%s:%d rcvif == NULL: m=%p", __func__, __LINE__, m));
901 		CURVNET_SET(m->m_pkthdr.rcvif->if_vnet);
902 		netisr_proto[proto].np_handler(m);
903 		CURVNET_RESTORE();
904 	}
905 	KASSERT(local_npw.nw_len == 0,
906 	    ("%s(%u): len %u", __func__, proto, local_npw.nw_len));
907 	if (netisr_proto[proto].np_drainedcpu)
908 		netisr_proto[proto].np_drainedcpu(nwsp->nws_cpu);
909 	NWS_LOCK(nwsp);
910 	npwp->nw_handled += handled;
911 	return (handled);
912 }
913 
914 /*
915  * SWI handler for netisr -- processes packets in a set of workstreams that
916  * it owns, woken up by calls to NWS_SIGNAL().  If this workstream is already
917  * being direct dispatched, go back to sleep and wait for the dispatching
918  * thread to wake us up again.
919  */
920 static void
921 swi_net(void *arg)
922 {
923 	struct epoch_tracker et;
924 #ifdef NETISR_LOCKING
925 	struct rm_priotracker tracker;
926 #endif
927 	struct netisr_workstream *nwsp;
928 	u_int bits, prot;
929 
930 	nwsp = arg;
931 
932 #ifdef DEVICE_POLLING
933 	KASSERT(nws_count == 1,
934 	    ("%s: device_polling but nws_count != 1", __func__));
935 	NET_EPOCH_ENTER(et);
936 	netisr_poll();
937 	NET_EPOCH_EXIT(et);
938 #endif
939 #ifdef NETISR_LOCKING
940 	NETISR_RLOCK(&tracker);
941 #endif
942 	NWS_LOCK(nwsp);
943 	KASSERT(!(nwsp->nws_flags & NWS_RUNNING), ("swi_net: running"));
944 	if (nwsp->nws_flags & NWS_DISPATCHING)
945 		goto out;
946 	NET_EPOCH_ENTER(et);
947 	nwsp->nws_flags |= NWS_RUNNING;
948 	nwsp->nws_flags &= ~NWS_SCHEDULED;
949 	while ((bits = nwsp->nws_pendingbits) != 0) {
950 		while ((prot = ffs(bits)) != 0) {
951 			prot--;
952 			bits &= ~(1 << prot);
953 			(void)netisr_process_workstream_proto(nwsp, prot);
954 		}
955 	}
956 	nwsp->nws_flags &= ~NWS_RUNNING;
957 	NET_EPOCH_EXIT(et);
958 out:
959 	NWS_UNLOCK(nwsp);
960 #ifdef NETISR_LOCKING
961 	NETISR_RUNLOCK(&tracker);
962 #endif
963 #ifdef DEVICE_POLLING
964 	netisr_pollmore();
965 #endif
966 }
967 
968 static int
969 netisr_queue_workstream(struct netisr_workstream *nwsp, u_int proto,
970     struct netisr_work *npwp, struct mbuf *m, int *dosignalp)
971 {
972 
973 	NWS_LOCK_ASSERT(nwsp);
974 
975 	*dosignalp = 0;
976 	if (npwp->nw_len < npwp->nw_qlimit) {
977 		m->m_nextpkt = NULL;
978 		if (npwp->nw_head == NULL) {
979 			npwp->nw_head = m;
980 			npwp->nw_tail = m;
981 		} else {
982 			npwp->nw_tail->m_nextpkt = m;
983 			npwp->nw_tail = m;
984 		}
985 		npwp->nw_len++;
986 		if (npwp->nw_len > npwp->nw_watermark)
987 			npwp->nw_watermark = npwp->nw_len;
988 
989 		/*
990 		 * We must set the bit regardless of NWS_RUNNING, so that
991 		 * swi_net() keeps calling netisr_process_workstream_proto().
992 		 */
993 		nwsp->nws_pendingbits |= (1 << proto);
994 		if (!(nwsp->nws_flags &
995 		    (NWS_RUNNING | NWS_DISPATCHING | NWS_SCHEDULED))) {
996 			nwsp->nws_flags |= NWS_SCHEDULED;
997 			*dosignalp = 1;	/* Defer until unlocked. */
998 		}
999 		npwp->nw_queued++;
1000 		return (0);
1001 	} else {
1002 		m_freem(m);
1003 		npwp->nw_qdrops++;
1004 		return (ENOBUFS);
1005 	}
1006 }
1007 
1008 static int
1009 netisr_queue_internal(u_int proto, struct mbuf *m, u_int cpuid)
1010 {
1011 	struct netisr_workstream *nwsp;
1012 	struct netisr_work *npwp;
1013 	int dosignal, error;
1014 
1015 #ifdef NETISR_LOCKING
1016 	NETISR_LOCK_ASSERT();
1017 #endif
1018 	KASSERT(cpuid <= mp_maxid, ("%s: cpuid too big (%u, %u)", __func__,
1019 	    cpuid, mp_maxid));
1020 	KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
1021 
1022 	dosignal = 0;
1023 	error = 0;
1024 	nwsp = DPCPU_ID_PTR(cpuid, nws);
1025 	npwp = &nwsp->nws_work[proto];
1026 	NWS_LOCK(nwsp);
1027 	error = netisr_queue_workstream(nwsp, proto, npwp, m, &dosignal);
1028 	NWS_UNLOCK(nwsp);
1029 	if (dosignal)
1030 		NWS_SIGNAL(nwsp);
1031 	return (error);
1032 }
1033 
1034 int
1035 netisr_queue_src(u_int proto, uintptr_t source, struct mbuf *m)
1036 {
1037 #ifdef NETISR_LOCKING
1038 	struct rm_priotracker tracker;
1039 #endif
1040 	u_int cpuid;
1041 	int error;
1042 
1043 	KASSERT(proto < NETISR_MAXPROT,
1044 	    ("%s: invalid proto %u", __func__, proto));
1045 
1046 #ifdef NETISR_LOCKING
1047 	NETISR_RLOCK(&tracker);
1048 #endif
1049 	KASSERT(netisr_proto[proto].np_handler != NULL,
1050 	    ("%s: invalid proto %u", __func__, proto));
1051 
1052 #ifdef VIMAGE
1053 	if (V_netisr_enable[proto] == 0) {
1054 		m_freem(m);
1055 		return (ENOPROTOOPT);
1056 	}
1057 #endif
1058 
1059 	m = netisr_select_cpuid(&netisr_proto[proto], NETISR_DISPATCH_DEFERRED,
1060 	    source, m, &cpuid);
1061 	if (m != NULL) {
1062 		KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__,
1063 		    cpuid));
1064 		error = netisr_queue_internal(proto, m, cpuid);
1065 	} else
1066 		error = ENOBUFS;
1067 #ifdef NETISR_LOCKING
1068 	NETISR_RUNLOCK(&tracker);
1069 #endif
1070 	return (error);
1071 }
1072 
1073 int
1074 netisr_queue(u_int proto, struct mbuf *m)
1075 {
1076 
1077 	return (netisr_queue_src(proto, 0, m));
1078 }
1079 
1080 /*
1081  * Dispatch a packet for netisr processing; direct dispatch is permitted by
1082  * calling context.
1083  */
1084 int
1085 netisr_dispatch_src(u_int proto, uintptr_t source, struct mbuf *m)
1086 {
1087 #ifdef NETISR_LOCKING
1088 	struct rm_priotracker tracker;
1089 #endif
1090 	struct netisr_workstream *nwsp;
1091 	struct netisr_proto *npp;
1092 	struct netisr_work *npwp;
1093 	int dosignal, error;
1094 	u_int cpuid, dispatch_policy;
1095 
1096 	NET_EPOCH_ASSERT();
1097 	KASSERT(proto < NETISR_MAXPROT,
1098 	    ("%s: invalid proto %u", __func__, proto));
1099 #ifdef NETISR_LOCKING
1100 	NETISR_RLOCK(&tracker);
1101 #endif
1102 	npp = &netisr_proto[proto];
1103 	KASSERT(npp->np_handler != NULL, ("%s: invalid proto %u", __func__,
1104 	    proto));
1105 
1106 #ifdef VIMAGE
1107 	if (V_netisr_enable[proto] == 0) {
1108 		m_freem(m);
1109 		return (ENOPROTOOPT);
1110 	}
1111 #endif
1112 
1113 	dispatch_policy = netisr_get_dispatch(npp);
1114 	if (dispatch_policy == NETISR_DISPATCH_DEFERRED)
1115 		return (netisr_queue_src(proto, source, m));
1116 
1117 	/*
1118 	 * If direct dispatch is forced, then unconditionally dispatch
1119 	 * without a formal CPU selection.  Borrow the current CPU's stats,
1120 	 * even if there's no worker on it.  In this case we don't update
1121 	 * nws_flags because all netisr processing will be source ordered due
1122 	 * to always being forced to directly dispatch.
1123 	 */
1124 	if (dispatch_policy == NETISR_DISPATCH_DIRECT) {
1125 		nwsp = DPCPU_PTR(nws);
1126 		npwp = &nwsp->nws_work[proto];
1127 		npwp->nw_dispatched++;
1128 		npwp->nw_handled++;
1129 		netisr_proto[proto].np_handler(m);
1130 		error = 0;
1131 		goto out_unlock;
1132 	}
1133 
1134 	KASSERT(dispatch_policy == NETISR_DISPATCH_HYBRID,
1135 	    ("%s: unknown dispatch policy (%u)", __func__, dispatch_policy));
1136 
1137 	/*
1138 	 * Otherwise, we execute in a hybrid mode where we will try to direct
1139 	 * dispatch if we're on the right CPU and the netisr worker isn't
1140 	 * already running.
1141 	 */
1142 	sched_pin();
1143 	m = netisr_select_cpuid(&netisr_proto[proto], NETISR_DISPATCH_HYBRID,
1144 	    source, m, &cpuid);
1145 	if (m == NULL) {
1146 		error = ENOBUFS;
1147 		goto out_unpin;
1148 	}
1149 	KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
1150 	if (cpuid != curcpu)
1151 		goto queue_fallback;
1152 	nwsp = DPCPU_PTR(nws);
1153 	npwp = &nwsp->nws_work[proto];
1154 
1155 	/*-
1156 	 * We are willing to direct dispatch only if three conditions hold:
1157 	 *
1158 	 * (1) The netisr worker isn't already running,
1159 	 * (2) Another thread isn't already directly dispatching, and
1160 	 * (3) The netisr hasn't already been woken up.
1161 	 */
1162 	NWS_LOCK(nwsp);
1163 	if (nwsp->nws_flags & (NWS_RUNNING | NWS_DISPATCHING | NWS_SCHEDULED)) {
1164 		error = netisr_queue_workstream(nwsp, proto, npwp, m,
1165 		    &dosignal);
1166 		NWS_UNLOCK(nwsp);
1167 		if (dosignal)
1168 			NWS_SIGNAL(nwsp);
1169 		goto out_unpin;
1170 	}
1171 
1172 	/*
1173 	 * The current thread is now effectively the netisr worker, so set
1174 	 * the dispatching flag to prevent concurrent processing of the
1175 	 * stream from another thread (even the netisr worker), which could
1176 	 * otherwise lead to effective misordering of the stream.
1177 	 */
1178 	nwsp->nws_flags |= NWS_DISPATCHING;
1179 	NWS_UNLOCK(nwsp);
1180 	netisr_proto[proto].np_handler(m);
1181 	NWS_LOCK(nwsp);
1182 	nwsp->nws_flags &= ~NWS_DISPATCHING;
1183 	npwp->nw_handled++;
1184 	npwp->nw_hybrid_dispatched++;
1185 
1186 	/*
1187 	 * If other work was enqueued by another thread while we were direct
1188 	 * dispatching, we need to signal the netisr worker to do that work.
1189 	 * In the future, we might want to do some of that work in the
1190 	 * current thread, rather than trigger further context switches.  If
1191 	 * so, we'll want to establish a reasonable bound on the work done in
1192 	 * the "borrowed" context.
1193 	 */
1194 	if (nwsp->nws_pendingbits != 0) {
1195 		nwsp->nws_flags |= NWS_SCHEDULED;
1196 		dosignal = 1;
1197 	} else
1198 		dosignal = 0;
1199 	NWS_UNLOCK(nwsp);
1200 	if (dosignal)
1201 		NWS_SIGNAL(nwsp);
1202 	error = 0;
1203 	goto out_unpin;
1204 
1205 queue_fallback:
1206 	error = netisr_queue_internal(proto, m, cpuid);
1207 out_unpin:
1208 	sched_unpin();
1209 out_unlock:
1210 #ifdef NETISR_LOCKING
1211 	NETISR_RUNLOCK(&tracker);
1212 #endif
1213 	return (error);
1214 }
1215 
1216 int
1217 netisr_dispatch(u_int proto, struct mbuf *m)
1218 {
1219 
1220 	return (netisr_dispatch_src(proto, 0, m));
1221 }
1222 
1223 #ifdef DEVICE_POLLING
1224 /*
1225  * Kernel polling borrows a netisr thread to run interface polling in; this
1226  * function allows kernel polling to request that the netisr thread be
1227  * scheduled even if no packets are pending for protocols.
1228  */
1229 void
1230 netisr_sched_poll(void)
1231 {
1232 	struct netisr_workstream *nwsp;
1233 
1234 	nwsp = DPCPU_ID_PTR(nws_array[0], nws);
1235 	NWS_SIGNAL(nwsp);
1236 }
1237 #endif
1238 
1239 static void
1240 netisr_start_swi(u_int cpuid, struct pcpu *pc)
1241 {
1242 	char swiname[12];
1243 	struct netisr_workstream *nwsp;
1244 	int error;
1245 
1246 	KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
1247 
1248 	nwsp = DPCPU_ID_PTR(cpuid, nws);
1249 	mtx_init(&nwsp->nws_mtx, "netisr_mtx", NULL, MTX_DEF);
1250 	nwsp->nws_cpu = cpuid;
1251 	snprintf(swiname, sizeof(swiname), "netisr %u", cpuid);
1252 	error = swi_add(&nwsp->nws_intr_event, swiname, swi_net, nwsp,
1253 	    SWI_NET, INTR_TYPE_NET | INTR_MPSAFE, &nwsp->nws_swi_cookie);
1254 	if (error)
1255 		panic("%s: swi_add %d", __func__, error);
1256 	pc->pc_netisr = nwsp->nws_intr_event;
1257 	if (netisr_bindthreads) {
1258 		error = intr_event_bind(nwsp->nws_intr_event, cpuid);
1259 		if (error != 0)
1260 			printf("%s: cpu %u: intr_event_bind: %d", __func__,
1261 			    cpuid, error);
1262 	}
1263 	NETISR_WLOCK();
1264 	nws_array[nws_count] = nwsp->nws_cpu;
1265 	nws_count++;
1266 	NETISR_WUNLOCK();
1267 }
1268 
1269 /*
1270  * Initialize the netisr subsystem.  We rely on BSS and static initialization
1271  * of most fields in global data structures.
1272  *
1273  * Start a worker thread for the boot CPU so that we can support network
1274  * traffic immediately in case the network stack is used before additional
1275  * CPUs are started (for example, diskless boot).
1276  */
1277 static void
1278 netisr_init(void *arg)
1279 {
1280 	struct pcpu *pc;
1281 
1282 	NETISR_LOCK_INIT();
1283 	if (netisr_maxthreads == 0 || netisr_maxthreads < -1 )
1284 		netisr_maxthreads = 1;		/* default behavior */
1285 	else if (netisr_maxthreads == -1)
1286 		netisr_maxthreads = mp_ncpus;	/* use max cpus */
1287 	if (netisr_maxthreads > mp_ncpus) {
1288 		printf("netisr_init: forcing maxthreads from %d to %d\n",
1289 		    netisr_maxthreads, mp_ncpus);
1290 		netisr_maxthreads = mp_ncpus;
1291 	}
1292 	if (netisr_defaultqlimit > netisr_maxqlimit) {
1293 		printf("netisr_init: forcing defaultqlimit from %d to %d\n",
1294 		    netisr_defaultqlimit, netisr_maxqlimit);
1295 		netisr_defaultqlimit = netisr_maxqlimit;
1296 	}
1297 #ifdef DEVICE_POLLING
1298 	/*
1299 	 * The device polling code is not yet aware of how to deal with
1300 	 * multiple netisr threads, so for the time being compiling in device
1301 	 * polling disables parallel netisr workers.
1302 	 */
1303 	if (netisr_maxthreads != 1 || netisr_bindthreads != 0) {
1304 		printf("netisr_init: forcing maxthreads to 1 and "
1305 		    "bindthreads to 0 for device polling\n");
1306 		netisr_maxthreads = 1;
1307 		netisr_bindthreads = 0;
1308 	}
1309 #endif
1310 
1311 #ifdef EARLY_AP_STARTUP
1312 	STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1313 		if (nws_count >= netisr_maxthreads)
1314 			break;
1315 		netisr_start_swi(pc->pc_cpuid, pc);
1316 	}
1317 #else
1318 	pc = get_pcpu();
1319 	netisr_start_swi(pc->pc_cpuid, pc);
1320 #endif
1321 }
1322 SYSINIT(netisr_init, SI_SUB_SOFTINTR, SI_ORDER_FIRST, netisr_init, NULL);
1323 
1324 #ifndef EARLY_AP_STARTUP
1325 /*
1326  * Start worker threads for additional CPUs.  No attempt to gracefully handle
1327  * work reassignment, we don't yet support dynamic reconfiguration.
1328  */
1329 static void
1330 netisr_start(void *arg)
1331 {
1332 	struct pcpu *pc;
1333 
1334 	STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) {
1335 		if (nws_count >= netisr_maxthreads)
1336 			break;
1337 		/* Worker will already be present for boot CPU. */
1338 		if (pc->pc_netisr != NULL)
1339 			continue;
1340 		netisr_start_swi(pc->pc_cpuid, pc);
1341 	}
1342 }
1343 SYSINIT(netisr_start, SI_SUB_SMP, SI_ORDER_MIDDLE, netisr_start, NULL);
1344 #endif
1345 
1346 /*
1347  * Sysctl monitoring for netisr: query a list of registered protocols.
1348  */
1349 static int
1350 sysctl_netisr_proto(SYSCTL_HANDLER_ARGS)
1351 {
1352 	struct rm_priotracker tracker;
1353 	struct sysctl_netisr_proto *snpp, *snp_array;
1354 	struct netisr_proto *npp;
1355 	u_int counter, proto;
1356 	int error;
1357 
1358 	if (req->newptr != NULL)
1359 		return (EINVAL);
1360 	snp_array = malloc(sizeof(*snp_array) * NETISR_MAXPROT, M_TEMP,
1361 	    M_ZERO | M_WAITOK);
1362 	counter = 0;
1363 	NETISR_RLOCK(&tracker);
1364 	for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1365 		npp = &netisr_proto[proto];
1366 		if (npp->np_name == NULL)
1367 			continue;
1368 		snpp = &snp_array[counter];
1369 		snpp->snp_version = sizeof(*snpp);
1370 		strlcpy(snpp->snp_name, npp->np_name, NETISR_NAMEMAXLEN);
1371 		snpp->snp_proto = proto;
1372 		snpp->snp_qlimit = npp->np_qlimit;
1373 		snpp->snp_policy = npp->np_policy;
1374 		snpp->snp_dispatch = npp->np_dispatch;
1375 		if (npp->np_m2flow != NULL)
1376 			snpp->snp_flags |= NETISR_SNP_FLAGS_M2FLOW;
1377 		if (npp->np_m2cpuid != NULL)
1378 			snpp->snp_flags |= NETISR_SNP_FLAGS_M2CPUID;
1379 		if (npp->np_drainedcpu != NULL)
1380 			snpp->snp_flags |= NETISR_SNP_FLAGS_DRAINEDCPU;
1381 		counter++;
1382 	}
1383 	NETISR_RUNLOCK(&tracker);
1384 	KASSERT(counter <= NETISR_MAXPROT,
1385 	    ("sysctl_netisr_proto: counter too big (%d)", counter));
1386 	error = SYSCTL_OUT(req, snp_array, sizeof(*snp_array) * counter);
1387 	free(snp_array, M_TEMP);
1388 	return (error);
1389 }
1390 
1391 SYSCTL_PROC(_net_isr, OID_AUTO, proto,
1392     CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_proto,
1393     "S,sysctl_netisr_proto",
1394     "Return list of protocols registered with netisr");
1395 
1396 /*
1397  * Sysctl monitoring for netisr: query a list of workstreams.
1398  */
1399 static int
1400 sysctl_netisr_workstream(SYSCTL_HANDLER_ARGS)
1401 {
1402 	struct rm_priotracker tracker;
1403 	struct sysctl_netisr_workstream *snwsp, *snws_array;
1404 	struct netisr_workstream *nwsp;
1405 	u_int counter, cpuid;
1406 	int error;
1407 
1408 	if (req->newptr != NULL)
1409 		return (EINVAL);
1410 	snws_array = malloc(sizeof(*snws_array) * MAXCPU, M_TEMP,
1411 	    M_ZERO | M_WAITOK);
1412 	counter = 0;
1413 	NETISR_RLOCK(&tracker);
1414 	CPU_FOREACH(cpuid) {
1415 		nwsp = DPCPU_ID_PTR(cpuid, nws);
1416 		if (nwsp->nws_intr_event == NULL)
1417 			continue;
1418 		NWS_LOCK(nwsp);
1419 		snwsp = &snws_array[counter];
1420 		snwsp->snws_version = sizeof(*snwsp);
1421 
1422 		/*
1423 		 * For now, we equate workstream IDs and CPU IDs in the
1424 		 * kernel, but expose them independently to userspace in case
1425 		 * that assumption changes in the future.
1426 		 */
1427 		snwsp->snws_wsid = cpuid;
1428 		snwsp->snws_cpu = cpuid;
1429 		if (nwsp->nws_intr_event != NULL)
1430 			snwsp->snws_flags |= NETISR_SNWS_FLAGS_INTR;
1431 		NWS_UNLOCK(nwsp);
1432 		counter++;
1433 	}
1434 	NETISR_RUNLOCK(&tracker);
1435 	KASSERT(counter <= MAXCPU,
1436 	    ("sysctl_netisr_workstream: counter too big (%d)", counter));
1437 	error = SYSCTL_OUT(req, snws_array, sizeof(*snws_array) * counter);
1438 	free(snws_array, M_TEMP);
1439 	return (error);
1440 }
1441 
1442 SYSCTL_PROC(_net_isr, OID_AUTO, workstream,
1443     CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_workstream,
1444     "S,sysctl_netisr_workstream",
1445     "Return list of workstreams implemented by netisr");
1446 
1447 /*
1448  * Sysctl monitoring for netisr: query per-protocol data across all
1449  * workstreams.
1450  */
1451 static int
1452 sysctl_netisr_work(SYSCTL_HANDLER_ARGS)
1453 {
1454 	struct rm_priotracker tracker;
1455 	struct sysctl_netisr_work *snwp, *snw_array;
1456 	struct netisr_workstream *nwsp;
1457 	struct netisr_proto *npp;
1458 	struct netisr_work *nwp;
1459 	u_int counter, cpuid, proto;
1460 	int error;
1461 
1462 	if (req->newptr != NULL)
1463 		return (EINVAL);
1464 	snw_array = malloc(sizeof(*snw_array) * MAXCPU * NETISR_MAXPROT,
1465 	    M_TEMP, M_ZERO | M_WAITOK);
1466 	counter = 0;
1467 	NETISR_RLOCK(&tracker);
1468 	CPU_FOREACH(cpuid) {
1469 		nwsp = DPCPU_ID_PTR(cpuid, nws);
1470 		if (nwsp->nws_intr_event == NULL)
1471 			continue;
1472 		NWS_LOCK(nwsp);
1473 		for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1474 			npp = &netisr_proto[proto];
1475 			if (npp->np_name == NULL)
1476 				continue;
1477 			nwp = &nwsp->nws_work[proto];
1478 			snwp = &snw_array[counter];
1479 			snwp->snw_version = sizeof(*snwp);
1480 			snwp->snw_wsid = cpuid;		/* See comment above. */
1481 			snwp->snw_proto = proto;
1482 			snwp->snw_len = nwp->nw_len;
1483 			snwp->snw_watermark = nwp->nw_watermark;
1484 			snwp->snw_dispatched = nwp->nw_dispatched;
1485 			snwp->snw_hybrid_dispatched =
1486 			    nwp->nw_hybrid_dispatched;
1487 			snwp->snw_qdrops = nwp->nw_qdrops;
1488 			snwp->snw_queued = nwp->nw_queued;
1489 			snwp->snw_handled = nwp->nw_handled;
1490 			counter++;
1491 		}
1492 		NWS_UNLOCK(nwsp);
1493 	}
1494 	KASSERT(counter <= MAXCPU * NETISR_MAXPROT,
1495 	    ("sysctl_netisr_work: counter too big (%d)", counter));
1496 	NETISR_RUNLOCK(&tracker);
1497 	error = SYSCTL_OUT(req, snw_array, sizeof(*snw_array) * counter);
1498 	free(snw_array, M_TEMP);
1499 	return (error);
1500 }
1501 
1502 SYSCTL_PROC(_net_isr, OID_AUTO, work,
1503     CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_work,
1504     "S,sysctl_netisr_work",
1505     "Return list of per-workstream, per-protocol work in netisr");
1506 
1507 #ifdef DDB
1508 DB_SHOW_COMMAND(netisr, db_show_netisr)
1509 {
1510 	struct netisr_workstream *nwsp;
1511 	struct netisr_work *nwp;
1512 	int first, proto;
1513 	u_int cpuid;
1514 
1515 	db_printf("%3s %6s %5s %5s %5s %8s %8s %8s %8s\n", "CPU", "Proto",
1516 	    "Len", "WMark", "Max", "Disp", "HDisp", "Drop", "Queue");
1517 	CPU_FOREACH(cpuid) {
1518 		nwsp = DPCPU_ID_PTR(cpuid, nws);
1519 		if (nwsp->nws_intr_event == NULL)
1520 			continue;
1521 		first = 1;
1522 		for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1523 			if (netisr_proto[proto].np_handler == NULL)
1524 				continue;
1525 			nwp = &nwsp->nws_work[proto];
1526 			if (first) {
1527 				db_printf("%3d ", cpuid);
1528 				first = 0;
1529 			} else
1530 				db_printf("%3s ", "");
1531 			db_printf(
1532 			    "%6s %5d %5d %5d %8ju %8ju %8ju %8ju\n",
1533 			    netisr_proto[proto].np_name, nwp->nw_len,
1534 			    nwp->nw_watermark, nwp->nw_qlimit,
1535 			    nwp->nw_dispatched, nwp->nw_hybrid_dispatched,
1536 			    nwp->nw_qdrops, nwp->nw_queued);
1537 		}
1538 	}
1539 }
1540 #endif
1541