netisr.c (938448cd871ff41f72bef12c9bc43865260cd194) netisr.c (e3b6e33c07b11798e5e24baaa7d596902debdcba)
1/*-
2 * Copyright (c) 2007-2009 Robert N. M. Watson
3 * Copyright (c) 2010 Juniper Networks, Inc.
1/*
2 * Copyright (c) 1997, Stefan Esser <se@freebsd.org>
4 * All rights reserved.
5 *
3 * All rights reserved.
4 *
6 * This software was developed by Robert N. M. Watson under contract
7 * to Juniper Networks, Inc.
8 *
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
11 * are met:
12 * 1. Redistributions of source code must retain the above copyright
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
9 * notice unmodified, this list of conditions, and the following
10 * disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 *
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
14 *
18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 * SUCH DAMAGE.
29 */
30
31#include <sys/cdefs.h>
32__FBSDID("$FreeBSD$");
33
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:
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
42 *
25 *
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.
26 * $FreeBSD$
64 */
65
27 */
28
66#include "opt_ddb.h"
67#include "opt_device_polling.h"
68
69#include <sys/param.h>
29#include <sys/param.h>
30#include <sys/systm.h>
70#include <sys/bus.h>
31#include <sys/bus.h>
71#include <sys/kernel.h>
72#include <sys/kthread.h>
73#include <sys/interrupt.h>
74#include <sys/lock.h>
75#include <sys/mbuf.h>
76#include <sys/mutex.h>
77#include <sys/pcpu.h>
78#include <sys/proc.h>
32#include <sys/proc.h>
79#include <sys/rmlock.h>
80#include <sys/sched.h>
81#include <sys/smp.h>
82#include <sys/socket.h>
83#include <sys/sysctl.h>
84#include <sys/systm.h>
33#include <sys/interrupt.h>
34#include <sys/kernel.h>
85
35
86#ifdef DDB
87#include <ddb/ddb.h>
88#endif
89
90#define _WANT_NETISR_INTERNAL /* Enable definitions from netisr_internal.h */
91#include <net/if.h>
92#include <net/if_var.h>
93#include <net/netisr.h>
36#include <net/netisr.h>
94#include <net/netisr_internal.h>
95#include <net/vnet.h>
96
37
97/*-
98 * Synchronize use and modification of the registered netisr data structures;
99 * acquire a read lock while modifying the set of registered protocols to
100 * prevent partially registered or unregistered protocols from being run.
101 *
102 * The following data structures and fields are protected by this lock:
103 *
104 * - The netisr_proto array, including all fields of struct netisr_proto.
105 * - The nws array, including all fields of struct netisr_worker.
106 * - The nws_array array.
107 *
108 * Note: the NETISR_LOCKING define controls whether read locks are acquired
109 * in packet processing paths requiring netisr registration stability. This
110 * is disabled by default as it can lead to measurable performance
111 * degradation even with rmlocks (3%-6% for loopback ping-pong traffic), and
112 * because netisr registration and unregistration is extremely rare at
113 * runtime. If it becomes more common, this decision should be revisited.
114 *
115 * XXXRW: rmlocks don't support assertions.
116 */
117static struct rmlock netisr_rmlock;
118#define NETISR_LOCK_INIT() rm_init_flags(&netisr_rmlock, "netisr", \
119 RM_NOWITNESS)
120#define NETISR_LOCK_ASSERT()
121#define NETISR_RLOCK(tracker) rm_rlock(&netisr_rmlock, (tracker))
122#define NETISR_RUNLOCK(tracker) rm_runlock(&netisr_rmlock, (tracker))
123#define NETISR_WLOCK() rm_wlock(&netisr_rmlock)
124#define NETISR_WUNLOCK() rm_wunlock(&netisr_rmlock)
125/* #define NETISR_LOCKING */
38static void swi_net(void *);
126
39
127SYSCTL_NODE(_net, OID_AUTO, isr, CTLFLAG_RW, 0, "netisr");
40void *net_ih;
41volatile unsigned int netisr;
42void (*netisrs[32])(void);
128
43
129/*-
130 * Three direct dispatch policies are supported:
131 *
132 * - Always defer: all work is scheduled for a netisr, regardless of context.
133 * (!direct)
134 *
135 * - Hybrid: if the executing context allows direct dispatch, and we're
136 * running on the CPU the work would be done on, then direct dispatch if it
137 * wouldn't violate ordering constraints on the workstream.
138 * (direct && !direct_force)
139 *
140 * - Always direct: if the executing context allows direct dispatch, always
141 * direct dispatch. (direct && direct_force)
142 *
143 * Notice that changing the global policy could lead to short periods of
144 * misordered processing, but this is considered acceptable as compared to
145 * the complexity of enforcing ordering during policy changes.
146 */
147static int netisr_direct_force = 1; /* Always direct dispatch. */
148TUNABLE_INT("net.isr.direct_force", &netisr_direct_force);
149SYSCTL_INT(_net_isr, OID_AUTO, direct_force, CTLFLAG_RW,
150 &netisr_direct_force, 0, "Force direct dispatch");
151
152static int netisr_direct = 1; /* Enable direct dispatch. */
153TUNABLE_INT("net.isr.direct", &netisr_direct);
154SYSCTL_INT(_net_isr, OID_AUTO, direct, CTLFLAG_RW,
155 &netisr_direct, 0, "Enable direct dispatch");
156
157/*
158 * Allow the administrator to limit the number of threads (CPUs) to use for
159 * netisr. We don't check netisr_maxthreads before creating the thread for
160 * CPU 0, so in practice we ignore values <= 1. This must be set at boot.
161 * We will create at most one thread per CPU.
162 */
163static int netisr_maxthreads = -1; /* Max number of threads. */
164TUNABLE_INT("net.isr.maxthreads", &netisr_maxthreads);
165SYSCTL_INT(_net_isr, OID_AUTO, maxthreads, CTLFLAG_RDTUN,
166 &netisr_maxthreads, 0,
167 "Use at most this many CPUs for netisr processing");
168
169static int netisr_bindthreads = 0; /* Bind threads to CPUs. */
170TUNABLE_INT("net.isr.bindthreads", &netisr_bindthreads);
171SYSCTL_INT(_net_isr, OID_AUTO, bindthreads, CTLFLAG_RDTUN,
172 &netisr_bindthreads, 0, "Bind netisr threads to CPUs.");
173
174/*
175 * Limit per-workstream mbuf queue limits s to at most net.isr.maxqlimit,
176 * both for initial configuration and later modification using
177 * netisr_setqlimit().
178 */
179#define NETISR_DEFAULT_MAXQLIMIT 10240
180static u_int netisr_maxqlimit = NETISR_DEFAULT_MAXQLIMIT;
181TUNABLE_INT("net.isr.maxqlimit", &netisr_maxqlimit);
182SYSCTL_INT(_net_isr, OID_AUTO, maxqlimit, CTLFLAG_RDTUN,
183 &netisr_maxqlimit, 0,
184 "Maximum netisr per-protocol, per-CPU queue depth.");
185
186/*
187 * The default per-workstream mbuf queue limit for protocols that don't
188 * initialize the nh_qlimit field of their struct netisr_handler. If this is
189 * set above netisr_maxqlimit, we truncate it to the maximum during boot.
190 */
191#define NETISR_DEFAULT_DEFAULTQLIMIT 256
192static u_int netisr_defaultqlimit = NETISR_DEFAULT_DEFAULTQLIMIT;
193TUNABLE_INT("net.isr.defaultqlimit", &netisr_defaultqlimit);
194SYSCTL_INT(_net_isr, OID_AUTO, defaultqlimit, CTLFLAG_RDTUN,
195 &netisr_defaultqlimit, 0,
196 "Default netisr per-protocol, per-CPU queue limit if not set by protocol");
197
198/*
199 * Store and export the compile-time constant NETISR_MAXPROT limit on the
200 * number of protocols that can register with netisr at a time. This is
201 * required for crashdump analysis, as it sizes netisr_proto[].
202 */
203static u_int netisr_maxprot = NETISR_MAXPROT;
204SYSCTL_INT(_net_isr, OID_AUTO, maxprot, CTLFLAG_RD,
205 &netisr_maxprot, 0,
206 "Compile-time limit on the number of protocols supported by netisr.");
207
208/*
209 * The netisr_proto array describes all registered protocols, indexed by
210 * protocol number. See netisr_internal.h for more details.
211 */
212static struct netisr_proto netisr_proto[NETISR_MAXPROT];
213
214/*
215 * Per-CPU workstream data. See netisr_internal.h for more details.
216 */
217DPCPU_DEFINE(struct netisr_workstream, nws);
218
219/*
220 * Map contiguous values between 0 and nws_count into CPU IDs appropriate for
221 * accessing workstreams. This allows constructions of the form
222 * DPCPU_ID_GET(nws_array[arbitraryvalue % nws_count], nws).
223 */
224static u_int nws_array[MAXCPU];
225
226/*
227 * Number of registered workstreams. Will be at most the number of running
228 * CPUs once fully started.
229 */
230static u_int nws_count;
231SYSCTL_INT(_net_isr, OID_AUTO, numthreads, CTLFLAG_RD,
232 &nws_count, 0, "Number of extant netisr threads.");
233
234/*
235 * Synchronization for each workstream: a mutex protects all mutable fields
236 * in each stream, including per-protocol state (mbuf queues). The SWI is
237 * woken up if asynchronous dispatch is required.
238 */
239#define NWS_LOCK(s) mtx_lock(&(s)->nws_mtx)
240#define NWS_LOCK_ASSERT(s) mtx_assert(&(s)->nws_mtx, MA_OWNED)
241#define NWS_UNLOCK(s) mtx_unlock(&(s)->nws_mtx)
242#define NWS_SIGNAL(s) swi_sched((s)->nws_swi_cookie, 0)
243
244/*
245 * Utility routines for protocols that implement their own mapping of flows
246 * to CPUs.
247 */
248u_int
249netisr_get_cpucount(void)
250{
251
252 return (nws_count);
253}
254
255u_int
256netisr_get_cpuid(u_int cpunumber)
257{
258
259 KASSERT(cpunumber < nws_count, ("%s: %u > %u", __func__, cpunumber,
260 nws_count));
261
262 return (nws_array[cpunumber]);
263}
264
265/*
266 * The default implementation of flow -> CPU ID mapping.
267 *
268 * Non-static so that protocols can use it to map their own work to specific
269 * CPUs in a manner consistent to netisr for affinity purposes.
270 */
271u_int
272netisr_default_flow2cpu(u_int flowid)
273{
274
275 return (nws_array[flowid % nws_count]);
276}
277
278/*
279 * Register a new netisr handler, which requires initializing per-protocol
280 * fields for each workstream. All netisr work is briefly suspended while
281 * the protocol is installed.
282 */
283void
44void
284netisr_register(const struct netisr_handler *nhp)
45legacy_setsoftnet(void)
285{
46{
286 struct netisr_work *npwp;
287 const char *name;
288 u_int i, proto;
289
290 proto = nhp->nh_proto;
291 name = nhp->nh_name;
292
293 /*
294 * Test that the requested registration is valid.
295 */
296 KASSERT(nhp->nh_name != NULL,
297 ("%s: nh_name NULL for %u", __func__, proto));
298 KASSERT(nhp->nh_handler != NULL,
299 ("%s: nh_handler NULL for %s", __func__, name));
300 KASSERT(nhp->nh_policy == NETISR_POLICY_SOURCE ||
301 nhp->nh_policy == NETISR_POLICY_FLOW ||
302 nhp->nh_policy == NETISR_POLICY_CPU,
303 ("%s: unsupported nh_policy %u for %s", __func__,
304 nhp->nh_policy, name));
305 KASSERT(nhp->nh_policy == NETISR_POLICY_FLOW ||
306 nhp->nh_m2flow == NULL,
307 ("%s: nh_policy != FLOW but m2flow defined for %s", __func__,
308 name));
309 KASSERT(nhp->nh_policy == NETISR_POLICY_CPU || nhp->nh_m2cpuid == NULL,
310 ("%s: nh_policy != CPU but m2cpuid defined for %s", __func__,
311 name));
312 KASSERT(nhp->nh_policy != NETISR_POLICY_CPU || nhp->nh_m2cpuid != NULL,
313 ("%s: nh_policy == CPU but m2cpuid not defined for %s", __func__,
314 name));
315 KASSERT(proto < NETISR_MAXPROT,
316 ("%s(%u, %s): protocol too big", __func__, proto, name));
317
318 /*
319 * Test that no existing registration exists for this protocol.
320 */
321 NETISR_WLOCK();
322 KASSERT(netisr_proto[proto].np_name == NULL,
323 ("%s(%u, %s): name present", __func__, proto, name));
324 KASSERT(netisr_proto[proto].np_handler == NULL,
325 ("%s(%u, %s): handler present", __func__, proto, name));
326
327 netisr_proto[proto].np_name = name;
328 netisr_proto[proto].np_handler = nhp->nh_handler;
329 netisr_proto[proto].np_m2flow = nhp->nh_m2flow;
330 netisr_proto[proto].np_m2cpuid = nhp->nh_m2cpuid;
331 netisr_proto[proto].np_drainedcpu = nhp->nh_drainedcpu;
332 if (nhp->nh_qlimit == 0)
333 netisr_proto[proto].np_qlimit = netisr_defaultqlimit;
334 else if (nhp->nh_qlimit > netisr_maxqlimit) {
335 printf("%s: %s requested queue limit %u capped to "
336 "net.isr.maxqlimit %u\n", __func__, name, nhp->nh_qlimit,
337 netisr_maxqlimit);
338 netisr_proto[proto].np_qlimit = netisr_maxqlimit;
339 } else
340 netisr_proto[proto].np_qlimit = nhp->nh_qlimit;
341 netisr_proto[proto].np_policy = nhp->nh_policy;
342 for (i = 0; i <= mp_maxid; i++) {
343 if (CPU_ABSENT(i))
344 continue;
345 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
346 bzero(npwp, sizeof(*npwp));
347 npwp->nw_qlimit = netisr_proto[proto].np_qlimit;
348 }
349 NETISR_WUNLOCK();
47 swi_sched(net_ih, 0);
350}
351
48}
49
352/*
353 * Clear drop counters across all workstreams for a protocol.
354 */
355void
356netisr_clearqdrops(const struct netisr_handler *nhp)
50int
51register_netisr(int num, netisr_t *handler)
357{
52{
358 struct netisr_work *npwp;
359#ifdef INVARIANTS
360 const char *name;
361#endif
362 u_int i, proto;
363
364 proto = nhp->nh_proto;
365#ifdef INVARIANTS
366 name = nhp->nh_name;
367#endif
368 KASSERT(proto < NETISR_MAXPROT,
369 ("%s(%u): protocol too big for %s", __func__, proto, name));
370
371 NETISR_WLOCK();
372 KASSERT(netisr_proto[proto].np_handler != NULL,
373 ("%s(%u): protocol not registered for %s", __func__, proto,
374 name));
375
376 for (i = 0; i <= mp_maxid; i++) {
377 if (CPU_ABSENT(i))
378 continue;
379 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
380 npwp->nw_qdrops = 0;
53
54 if (num < 0 || num >= (sizeof(netisrs)/sizeof(*netisrs)) ) {
55 printf("register_netisr: bad isr number: %d\n", num);
56 return (EINVAL);
381 }
57 }
382 NETISR_WUNLOCK();
58 netisrs[num] = handler;
59 return (0);
383}
384
60}
61
385/*
386 * Query current drop counters across all workstreams for a protocol.
387 */
388void
389netisr_getqdrops(const struct netisr_handler *nhp, u_int64_t *qdropp)
390{
391 struct netisr_work *npwp;
392 struct rm_priotracker tracker;
393#ifdef INVARIANTS
394 const char *name;
395#endif
396 u_int i, proto;
397
398 *qdropp = 0;
399 proto = nhp->nh_proto;
400#ifdef INVARIANTS
401 name = nhp->nh_name;
402#endif
403 KASSERT(proto < NETISR_MAXPROT,
404 ("%s(%u): protocol too big for %s", __func__, proto, name));
405
406 NETISR_RLOCK(&tracker);
407 KASSERT(netisr_proto[proto].np_handler != NULL,
408 ("%s(%u): protocol not registered for %s", __func__, proto,
409 name));
410
411 for (i = 0; i <= mp_maxid; i++) {
412 if (CPU_ABSENT(i))
413 continue;
414 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
415 *qdropp += npwp->nw_qdrops;
416 }
417 NETISR_RUNLOCK(&tracker);
418}
419
420/*
421 * Query current per-workstream queue limit for a protocol.
422 */
423void
424netisr_getqlimit(const struct netisr_handler *nhp, u_int *qlimitp)
425{
426 struct rm_priotracker tracker;
427#ifdef INVARIANTS
428 const char *name;
429#endif
430 u_int proto;
431
432 proto = nhp->nh_proto;
433#ifdef INVARIANTS
434 name = nhp->nh_name;
435#endif
436 KASSERT(proto < NETISR_MAXPROT,
437 ("%s(%u): protocol too big for %s", __func__, proto, name));
438
439 NETISR_RLOCK(&tracker);
440 KASSERT(netisr_proto[proto].np_handler != NULL,
441 ("%s(%u): protocol not registered for %s", __func__, proto,
442 name));
443 *qlimitp = netisr_proto[proto].np_qlimit;
444 NETISR_RUNLOCK(&tracker);
445}
446
447/*
448 * Update the queue limit across per-workstream queues for a protocol. We
449 * simply change the limits, and don't drain overflowed packets as they will
450 * (hopefully) take care of themselves shortly.
451 */
452int
62int
453netisr_setqlimit(const struct netisr_handler *nhp, u_int qlimit)
63unregister_netisr(int num)
454{
64{
455 struct netisr_work *npwp;
456#ifdef INVARIANTS
457 const char *name;
458#endif
459 u_int i, proto;
460
461 if (qlimit > netisr_maxqlimit)
65
66 if (num < 0 || num >= (sizeof(netisrs)/sizeof(*netisrs)) ) {
67 printf("unregister_netisr: bad isr number: %d\n", num);
462 return (EINVAL);
68 return (EINVAL);
463
464 proto = nhp->nh_proto;
465#ifdef INVARIANTS
466 name = nhp->nh_name;
467#endif
468 KASSERT(proto < NETISR_MAXPROT,
469 ("%s(%u): protocol too big for %s", __func__, proto, name));
470
471 NETISR_WLOCK();
472 KASSERT(netisr_proto[proto].np_handler != NULL,
473 ("%s(%u): protocol not registered for %s", __func__, proto,
474 name));
475
476 netisr_proto[proto].np_qlimit = qlimit;
477 for (i = 0; i <= mp_maxid; i++) {
478 if (CPU_ABSENT(i))
479 continue;
480 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
481 npwp->nw_qlimit = qlimit;
482 }
69 }
483 NETISR_WUNLOCK();
70 netisrs[num] = NULL;
484 return (0);
485}
486
71 return (0);
72}
73
487/*
488 * Drain all packets currently held in a particular protocol work queue.
489 */
490static void
491netisr_drain_proto(struct netisr_work *npwp)
492{
493 struct mbuf *m;
494
495 /*
496 * We would assert the lock on the workstream but it's not passed in.
497 */
498 while ((m = npwp->nw_head) != NULL) {
499 npwp->nw_head = m->m_nextpkt;
500 m->m_nextpkt = NULL;
501 if (npwp->nw_head == NULL)
502 npwp->nw_tail = NULL;
503 npwp->nw_len--;
504 m_freem(m);
505 }
506 KASSERT(npwp->nw_tail == NULL, ("%s: tail", __func__));
507 KASSERT(npwp->nw_len == 0, ("%s: len", __func__));
508}
509
510/*
511 * Remove the registration of a network protocol, which requires clearing
512 * per-protocol fields across all workstreams, including freeing all mbufs in
513 * the queues at time of unregister. All work in netisr is briefly suspended
514 * while this takes place.
515 */
516void
517netisr_unregister(const struct netisr_handler *nhp)
518{
519 struct netisr_work *npwp;
520#ifdef INVARIANTS
521 const char *name;
74#ifdef DEVICE_POLLING
75void netisr_pollmore(void);
522#endif
76#endif
523 u_int i, proto;
524
77
525 proto = nhp->nh_proto;
526#ifdef INVARIANTS
527 name = nhp->nh_name;
528#endif
529 KASSERT(proto < NETISR_MAXPROT,
530 ("%s(%u): protocol too big for %s", __func__, proto, name));
531
532 NETISR_WLOCK();
533 KASSERT(netisr_proto[proto].np_handler != NULL,
534 ("%s(%u): protocol not registered for %s", __func__, proto,
535 name));
536
537 netisr_proto[proto].np_name = NULL;
538 netisr_proto[proto].np_handler = NULL;
539 netisr_proto[proto].np_m2flow = NULL;
540 netisr_proto[proto].np_m2cpuid = NULL;
541 netisr_proto[proto].np_qlimit = 0;
542 netisr_proto[proto].np_policy = 0;
543 for (i = 0; i <= mp_maxid; i++) {
544 if (CPU_ABSENT(i))
545 continue;
546 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
547 netisr_drain_proto(npwp);
548 bzero(npwp, sizeof(*npwp));
549 }
550 NETISR_WUNLOCK();
551}
552
553/*
554 * Look up the workstream given a packet and source identifier. Do this by
555 * checking the protocol's policy, and optionally call out to the protocol
556 * for assistance if required.
557 */
558static struct mbuf *
559netisr_select_cpuid(struct netisr_proto *npp, uintptr_t source,
560 struct mbuf *m, u_int *cpuidp)
561{
562 struct ifnet *ifp;
563
564 NETISR_LOCK_ASSERT();
565
566 /*
567 * In the event we have only one worker, shortcut and deliver to it
568 * without further ado.
569 */
570 if (nws_count == 1) {
571 *cpuidp = nws_array[0];
572 return (m);
573 }
574
575 /*
576 * What happens next depends on the policy selected by the protocol.
577 * If we want to support per-interface policies, we should do that
578 * here first.
579 */
580 switch (npp->np_policy) {
581 case NETISR_POLICY_CPU:
582 return (npp->np_m2cpuid(m, source, cpuidp));
583
584 case NETISR_POLICY_FLOW:
585 if (!(m->m_flags & M_FLOWID) && npp->np_m2flow != NULL) {
586 m = npp->np_m2flow(m, source);
587 if (m == NULL)
588 return (NULL);
589 }
590 if (m->m_flags & M_FLOWID) {
591 *cpuidp =
592 netisr_default_flow2cpu(m->m_pkthdr.flowid);
593 return (m);
594 }
595 /* FALLTHROUGH */
596
597 case NETISR_POLICY_SOURCE:
598 ifp = m->m_pkthdr.rcvif;
599 if (ifp != NULL)
600 *cpuidp = nws_array[(ifp->if_index + source) %
601 nws_count];
602 else
603 *cpuidp = nws_array[source % nws_count];
604 return (m);
605
606 default:
607 panic("%s: invalid policy %u for %s", __func__,
608 npp->np_policy, npp->np_name);
609 }
610}
611
612/*
613 * Process packets associated with a workstream and protocol. For reasons of
614 * fairness, we process up to one complete netisr queue at a time, moving the
615 * queue to a stack-local queue for processing, but do not loop refreshing
616 * from the global queue. The caller is responsible for deciding whether to
617 * loop, and for setting the NWS_RUNNING flag. The passed workstream will be
618 * locked on entry and relocked before return, but will be released while
619 * processing. The number of packets processed is returned.
620 */
621static u_int
622netisr_process_workstream_proto(struct netisr_workstream *nwsp, u_int proto)
623{
624 struct netisr_work local_npw, *npwp;
625 u_int handled;
626 struct mbuf *m;
627
628 NETISR_LOCK_ASSERT();
629 NWS_LOCK_ASSERT(nwsp);
630
631 KASSERT(nwsp->nws_flags & NWS_RUNNING,
632 ("%s(%u): not running", __func__, proto));
633 KASSERT(proto >= 0 && proto < NETISR_MAXPROT,
634 ("%s(%u): invalid proto\n", __func__, proto));
635
636 npwp = &nwsp->nws_work[proto];
637 if (npwp->nw_len == 0)
638 return (0);
639
640 /*
641 * Move the global work queue to a thread-local work queue.
642 *
643 * Notice that this means the effective maximum length of the queue
644 * is actually twice that of the maximum queue length specified in
645 * the protocol registration call.
646 */
647 handled = npwp->nw_len;
648 local_npw = *npwp;
649 npwp->nw_head = NULL;
650 npwp->nw_tail = NULL;
651 npwp->nw_len = 0;
652 nwsp->nws_pendingbits &= ~(1 << proto);
653 NWS_UNLOCK(nwsp);
654 while ((m = local_npw.nw_head) != NULL) {
655 local_npw.nw_head = m->m_nextpkt;
656 m->m_nextpkt = NULL;
657 if (local_npw.nw_head == NULL)
658 local_npw.nw_tail = NULL;
659 local_npw.nw_len--;
660 VNET_ASSERT(m->m_pkthdr.rcvif != NULL);
661 CURVNET_SET(m->m_pkthdr.rcvif->if_vnet);
662 netisr_proto[proto].np_handler(m);
663 CURVNET_RESTORE();
664 }
665 KASSERT(local_npw.nw_len == 0,
666 ("%s(%u): len %u", __func__, proto, local_npw.nw_len));
667 if (netisr_proto[proto].np_drainedcpu)
668 netisr_proto[proto].np_drainedcpu(nwsp->nws_cpu);
669 NWS_LOCK(nwsp);
670 npwp->nw_handled += handled;
671 return (handled);
672}
673
674/*
675 * SWI handler for netisr -- processes packets in a set of workstreams that
676 * it owns, woken up by calls to NWS_SIGNAL(). If this workstream is already
677 * being direct dispatched, go back to sleep and wait for the dispatching
678 * thread to wake us up again.
679 */
680static void
78static void
681swi_net(void *arg)
79swi_net(void *dummy)
682{
80{
683#ifdef NETISR_LOCKING
684 struct rm_priotracker tracker;
685#endif
686 struct netisr_workstream *nwsp;
687 u_int bits, prot;
81 u_int bits;
82 int i;
688
83
689 nwsp = arg;
690
691#ifdef DEVICE_POLLING
84#ifdef DEVICE_POLLING
692 KASSERT(nws_count == 1,
693 ("%s: device_polling but nws_count != 1", __func__));
694 netisr_poll();
85 for (;;) {
86 int pollmore;
695#endif
87#endif
696#ifdef NETISR_LOCKING
697 NETISR_RLOCK(&tracker);
698#endif
699 NWS_LOCK(nwsp);
700 KASSERT(!(nwsp->nws_flags & NWS_RUNNING), ("swi_net: running"));
701 if (nwsp->nws_flags & NWS_DISPATCHING)
702 goto out;
703 nwsp->nws_flags |= NWS_RUNNING;
704 nwsp->nws_flags &= ~NWS_SCHEDULED;
705 while ((bits = nwsp->nws_pendingbits) != 0) {
706 while ((prot = ffs(bits)) != 0) {
707 prot--;
708 bits &= ~(1 << prot);
709 (void)netisr_process_workstream_proto(nwsp, prot);
710 }
711 }
712 nwsp->nws_flags &= ~NWS_RUNNING;
713out:
714 NWS_UNLOCK(nwsp);
715#ifdef NETISR_LOCKING
716 NETISR_RUNLOCK(&tracker);
717#endif
88 bits = atomic_readandclear_int(&netisr);
718#ifdef DEVICE_POLLING
89#ifdef DEVICE_POLLING
719 netisr_pollmore();
90 if (bits == 0)
91 return;
92 pollmore = bits & (1 << NETISR_POLL);
720#endif
93#endif
721}
722
723static int
724netisr_queue_workstream(struct netisr_workstream *nwsp, u_int proto,
725 struct netisr_work *npwp, struct mbuf *m, int *dosignalp)
726{
727
728 NWS_LOCK_ASSERT(nwsp);
729
730 *dosignalp = 0;
731 if (npwp->nw_len < npwp->nw_qlimit) {
732 m->m_nextpkt = NULL;
733 if (npwp->nw_head == NULL) {
734 npwp->nw_head = m;
735 npwp->nw_tail = m;
736 } else {
737 npwp->nw_tail->m_nextpkt = m;
738 npwp->nw_tail = m;
739 }
740 npwp->nw_len++;
741 if (npwp->nw_len > npwp->nw_watermark)
742 npwp->nw_watermark = npwp->nw_len;
743
744 /*
745 * We must set the bit regardless of NWS_RUNNING, so that
746 * swi_net() keeps calling netisr_process_workstream_proto().
747 */
748 nwsp->nws_pendingbits |= (1 << proto);
749 if (!(nwsp->nws_flags &
750 (NWS_RUNNING | NWS_DISPATCHING | NWS_SCHEDULED))) {
751 nwsp->nws_flags |= NWS_SCHEDULED;
752 *dosignalp = 1; /* Defer until unlocked. */
753 }
754 npwp->nw_queued++;
755 return (0);
756 } else {
757 m_freem(m);
758 npwp->nw_qdrops++;
759 return (ENOBUFS);
94 while ((i = ffs(bits)) != 0) {
95 i--;
96 if (netisrs[i] != NULL)
97 netisrs[i]();
98 else
99 printf("swi_net: unregistered isr number: %d.\n", i);
100 bits &= ~(1 << i);
760 }
101 }
761}
762
763static int
764netisr_queue_internal(u_int proto, struct mbuf *m, u_int cpuid)
765{
766 struct netisr_workstream *nwsp;
767 struct netisr_work *npwp;
768 int dosignal, error;
769
770#ifdef NETISR_LOCKING
771 NETISR_LOCK_ASSERT();
772#endif
773 KASSERT(cpuid <= mp_maxid, ("%s: cpuid too big (%u, %u)", __func__,
774 cpuid, mp_maxid));
775 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
776
777 dosignal = 0;
778 error = 0;
779 nwsp = DPCPU_ID_PTR(cpuid, nws);
780 npwp = &nwsp->nws_work[proto];
781 NWS_LOCK(nwsp);
782 error = netisr_queue_workstream(nwsp, proto, npwp, m, &dosignal);
783 NWS_UNLOCK(nwsp);
784 if (dosignal)
785 NWS_SIGNAL(nwsp);
786 return (error);
787}
788
789int
790netisr_queue_src(u_int proto, uintptr_t source, struct mbuf *m)
791{
792#ifdef NETISR_LOCKING
793 struct rm_priotracker tracker;
794#endif
795 u_int cpuid;
796 int error;
797
798 KASSERT(proto < NETISR_MAXPROT,
799 ("%s: invalid proto %u", __func__, proto));
800
801#ifdef NETISR_LOCKING
802 NETISR_RLOCK(&tracker);
803#endif
804 KASSERT(netisr_proto[proto].np_handler != NULL,
805 ("%s: invalid proto %u", __func__, proto));
806
807 m = netisr_select_cpuid(&netisr_proto[proto], source, m, &cpuid);
808 if (m != NULL) {
809 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__,
810 cpuid));
811 error = netisr_queue_internal(proto, m, cpuid);
812 } else
813 error = ENOBUFS;
814#ifdef NETISR_LOCKING
815 NETISR_RUNLOCK(&tracker);
816#endif
817 return (error);
818}
819
820int
821netisr_queue(u_int proto, struct mbuf *m)
822{
823
824 return (netisr_queue_src(proto, 0, m));
825}
826
827/*
828 * Dispatch a packet for netisr processing; direct dispatch is permitted by
829 * calling context.
830 */
831int
832netisr_dispatch_src(u_int proto, uintptr_t source, struct mbuf *m)
833{
834#ifdef NETISR_LOCKING
835 struct rm_priotracker tracker;
836#endif
837 struct netisr_workstream *nwsp;
838 struct netisr_work *npwp;
839 int dosignal, error;
840 u_int cpuid;
841
842 /*
843 * If direct dispatch is entirely disabled, fall back on queueing.
844 */
845 if (!netisr_direct)
846 return (netisr_queue_src(proto, source, m));
847
848 KASSERT(proto < NETISR_MAXPROT,
849 ("%s: invalid proto %u", __func__, proto));
850#ifdef NETISR_LOCKING
851 NETISR_RLOCK(&tracker);
852#endif
853 KASSERT(netisr_proto[proto].np_handler != NULL,
854 ("%s: invalid proto %u", __func__, proto));
855
856 /*
857 * If direct dispatch is forced, then unconditionally dispatch
858 * without a formal CPU selection. Borrow the current CPU's stats,
859 * even if there's no worker on it. In this case we don't update
860 * nws_flags because all netisr processing will be source ordered due
861 * to always being forced to directly dispatch.
862 */
863 if (netisr_direct_force) {
864 nwsp = DPCPU_PTR(nws);
865 npwp = &nwsp->nws_work[proto];
866 npwp->nw_dispatched++;
867 npwp->nw_handled++;
868 netisr_proto[proto].np_handler(m);
869 error = 0;
870 goto out_unlock;
871 }
872
873 /*
874 * Otherwise, we execute in a hybrid mode where we will try to direct
875 * dispatch if we're on the right CPU and the netisr worker isn't
876 * already running.
877 */
878 m = netisr_select_cpuid(&netisr_proto[proto], source, m, &cpuid);
879 if (m == NULL) {
880 error = ENOBUFS;
881 goto out_unlock;
882 }
883 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
884 sched_pin();
885 if (cpuid != curcpu)
886 goto queue_fallback;
887 nwsp = DPCPU_PTR(nws);
888 npwp = &nwsp->nws_work[proto];
889
890 /*-
891 * We are willing to direct dispatch only if three conditions hold:
892 *
893 * (1) The netisr worker isn't already running,
894 * (2) Another thread isn't already directly dispatching, and
895 * (3) The netisr hasn't already been woken up.
896 */
897 NWS_LOCK(nwsp);
898 if (nwsp->nws_flags & (NWS_RUNNING | NWS_DISPATCHING | NWS_SCHEDULED)) {
899 error = netisr_queue_workstream(nwsp, proto, npwp, m,
900 &dosignal);
901 NWS_UNLOCK(nwsp);
902 if (dosignal)
903 NWS_SIGNAL(nwsp);
904 goto out_unpin;
905 }
906
907 /*
908 * The current thread is now effectively the netisr worker, so set
909 * the dispatching flag to prevent concurrent processing of the
910 * stream from another thread (even the netisr worker), which could
911 * otherwise lead to effective misordering of the stream.
912 */
913 nwsp->nws_flags |= NWS_DISPATCHING;
914 NWS_UNLOCK(nwsp);
915 netisr_proto[proto].np_handler(m);
916 NWS_LOCK(nwsp);
917 nwsp->nws_flags &= ~NWS_DISPATCHING;
918 npwp->nw_handled++;
919 npwp->nw_hybrid_dispatched++;
920
921 /*
922 * If other work was enqueued by another thread while we were direct
923 * dispatching, we need to signal the netisr worker to do that work.
924 * In the future, we might want to do some of that work in the
925 * current thread, rather than trigger further context switches. If
926 * so, we'll want to establish a reasonable bound on the work done in
927 * the "borrowed" context.
928 */
929 if (nwsp->nws_pendingbits != 0) {
930 nwsp->nws_flags |= NWS_SCHEDULED;
931 dosignal = 1;
932 } else
933 dosignal = 0;
934 NWS_UNLOCK(nwsp);
935 if (dosignal)
936 NWS_SIGNAL(nwsp);
937 error = 0;
938 goto out_unpin;
939
940queue_fallback:
941 error = netisr_queue_internal(proto, m, cpuid);
942out_unpin:
943 sched_unpin();
944out_unlock:
945#ifdef NETISR_LOCKING
946 NETISR_RUNLOCK(&tracker);
947#endif
948 return (error);
949}
950
951int
952netisr_dispatch(u_int proto, struct mbuf *m)
953{
954
955 return (netisr_dispatch_src(proto, 0, m));
956}
957
958#ifdef DEVICE_POLLING
102#ifdef DEVICE_POLLING
959/*
960 * Kernel polling borrows a netisr thread to run interface polling in; this
961 * function allows kernel polling to request that the netisr thread be
962 * scheduled even if no packets are pending for protocols.
963 */
964void
965netisr_sched_poll(void)
966{
967 struct netisr_workstream *nwsp;
968
969 nwsp = DPCPU_ID_PTR(nws_array[0], nws);
970 NWS_SIGNAL(nwsp);
971}
103 if (pollmore)
104 netisr_pollmore();
105 }
972#endif
106#endif
973
974static void
975netisr_start_swi(u_int cpuid, struct pcpu *pc)
976{
977 char swiname[12];
978 struct netisr_workstream *nwsp;
979 int error;
980
981 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
982
983 nwsp = DPCPU_ID_PTR(cpuid, nws);
984 mtx_init(&nwsp->nws_mtx, "netisr_mtx", NULL, MTX_DEF);
985 nwsp->nws_cpu = cpuid;
986 snprintf(swiname, sizeof(swiname), "netisr %u", cpuid);
987 error = swi_add(&nwsp->nws_intr_event, swiname, swi_net, nwsp,
988 SWI_NET, INTR_MPSAFE, &nwsp->nws_swi_cookie);
989 if (error)
990 panic("%s: swi_add %d", __func__, error);
991 pc->pc_netisr = nwsp->nws_intr_event;
992 if (netisr_bindthreads) {
993 error = intr_event_bind(nwsp->nws_intr_event, cpuid);
994 if (error != 0)
995 printf("%s: cpu %u: intr_event_bind: %d", __func__,
996 cpuid, error);
997 }
998 NETISR_WLOCK();
999 nws_array[nws_count] = nwsp->nws_cpu;
1000 nws_count++;
1001 NETISR_WUNLOCK();
1002}
1003
107}
108
1004/*
1005 * Initialize the netisr subsystem. We rely on BSS and static initialization
1006 * of most fields in global data structures.
1007 *
1008 * Start a worker thread for the boot CPU so that we can support network
1009 * traffic immediately in case the network stack is used before additional
1010 * CPUs are started (for example, diskless boot).
1011 */
1012static void
109static void
1013netisr_init(void *arg)
110start_netisr(void *dummy)
1014{
1015
111{
112
1016 KASSERT(curcpu == 0, ("%s: not on CPU 0", __func__));
1017
1018 NETISR_LOCK_INIT();
1019 if (netisr_maxthreads < 1)
1020 netisr_maxthreads = 1;
1021 if (netisr_maxthreads > mp_ncpus) {
1022 printf("netisr_init: forcing maxthreads from %d to %d\n",
1023 netisr_maxthreads, mp_ncpus);
1024 netisr_maxthreads = mp_ncpus;
1025 }
1026 if (netisr_defaultqlimit > netisr_maxqlimit) {
1027 printf("netisr_init: forcing defaultqlimit from %d to %d\n",
1028 netisr_defaultqlimit, netisr_maxqlimit);
1029 netisr_defaultqlimit = netisr_maxqlimit;
1030 }
1031#ifdef DEVICE_POLLING
1032 /*
1033 * The device polling code is not yet aware of how to deal with
1034 * multiple netisr threads, so for the time being compiling in device
1035 * polling disables parallel netisr workers.
1036 */
1037 if (netisr_maxthreads != 1 || netisr_bindthreads != 0) {
1038 printf("netisr_init: forcing maxthreads to 1 and "
1039 "bindthreads to 0 for device polling\n");
1040 netisr_maxthreads = 1;
1041 netisr_bindthreads = 0;
1042 }
1043#endif
1044
1045 netisr_start_swi(curcpu, pcpu_find(curcpu));
113 if (swi_add(NULL, "net", swi_net, NULL, SWI_NET, 0, &net_ih))
114 panic("start_netisr");
1046}
115}
1047SYSINIT(netisr_init, SI_SUB_SOFTINTR, SI_ORDER_FIRST, netisr_init, NULL);
1048
1049/*
1050 * Start worker threads for additional CPUs. No attempt to gracefully handle
1051 * work reassignment, we don't yet support dynamic reconfiguration.
1052 */
1053static void
1054netisr_start(void *arg)
1055{
1056 struct pcpu *pc;
1057
1058 SLIST_FOREACH(pc, &cpuhead, pc_allcpu) {
1059 if (nws_count >= netisr_maxthreads)
1060 break;
1061 /* XXXRW: Is skipping absent CPUs still required here? */
1062 if (CPU_ABSENT(pc->pc_cpuid))
1063 continue;
1064 /* Worker will already be present for boot CPU. */
1065 if (pc->pc_netisr != NULL)
1066 continue;
1067 netisr_start_swi(pc->pc_cpuid, pc);
1068 }
1069}
1070SYSINIT(netisr_start, SI_SUB_SMP, SI_ORDER_MIDDLE, netisr_start, NULL);
1071
1072/*
1073 * Sysctl monitoring for netisr: query a list of registered protocols.
1074 */
1075static int
1076sysctl_netisr_proto(SYSCTL_HANDLER_ARGS)
1077{
1078 struct rm_priotracker tracker;
1079 struct sysctl_netisr_proto *snpp, *snp_array;
1080 struct netisr_proto *npp;
1081 u_int counter, proto;
1082 int error;
1083
1084 if (req->newptr != NULL)
1085 return (EINVAL);
1086 snp_array = malloc(sizeof(*snp_array) * NETISR_MAXPROT, M_TEMP,
1087 M_ZERO | M_WAITOK);
1088 counter = 0;
1089 NETISR_RLOCK(&tracker);
1090 for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1091 npp = &netisr_proto[proto];
1092 if (npp->np_name == NULL)
1093 continue;
1094 snpp = &snp_array[counter];
1095 snpp->snp_version = sizeof(*snpp);
1096 strlcpy(snpp->snp_name, npp->np_name, NETISR_NAMEMAXLEN);
1097 snpp->snp_proto = proto;
1098 snpp->snp_qlimit = npp->np_qlimit;
1099 snpp->snp_policy = npp->np_policy;
1100 if (npp->np_m2flow != NULL)
1101 snpp->snp_flags |= NETISR_SNP_FLAGS_M2FLOW;
1102 if (npp->np_m2cpuid != NULL)
1103 snpp->snp_flags |= NETISR_SNP_FLAGS_M2CPUID;
1104 if (npp->np_drainedcpu != NULL)
1105 snpp->snp_flags |= NETISR_SNP_FLAGS_DRAINEDCPU;
1106 counter++;
1107 }
1108 NETISR_RUNLOCK(&tracker);
1109 KASSERT(counter <= NETISR_MAXPROT,
1110 ("sysctl_netisr_proto: counter too big (%d)", counter));
1111 error = SYSCTL_OUT(req, snp_array, sizeof(*snp_array) * counter);
1112 free(snp_array, M_TEMP);
1113 return (error);
1114}
1115
1116SYSCTL_PROC(_net_isr, OID_AUTO, proto,
1117 CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_proto,
1118 "S,sysctl_netisr_proto",
1119 "Return list of protocols registered with netisr");
1120
1121/*
1122 * Sysctl monitoring for netisr: query a list of workstreams.
1123 */
1124static int
1125sysctl_netisr_workstream(SYSCTL_HANDLER_ARGS)
1126{
1127 struct rm_priotracker tracker;
1128 struct sysctl_netisr_workstream *snwsp, *snws_array;
1129 struct netisr_workstream *nwsp;
1130 u_int counter, cpuid;
1131 int error;
1132
1133 if (req->newptr != NULL)
1134 return (EINVAL);
1135 snws_array = malloc(sizeof(*snws_array) * MAXCPU, M_TEMP,
1136 M_ZERO | M_WAITOK);
1137 counter = 0;
1138 NETISR_RLOCK(&tracker);
1139 for (cpuid = 0; cpuid < MAXCPU; cpuid++) {
1140 if (CPU_ABSENT(cpuid))
1141 continue;
1142 nwsp = DPCPU_ID_PTR(cpuid, nws);
1143 if (nwsp->nws_intr_event == NULL)
1144 continue;
1145 NWS_LOCK(nwsp);
1146 snwsp = &snws_array[counter];
1147 snwsp->snws_version = sizeof(*snwsp);
1148
1149 /*
1150 * For now, we equate workstream IDs and CPU IDs in the
1151 * kernel, but expose them independently to userspace in case
1152 * that assumption changes in the future.
1153 */
1154 snwsp->snws_wsid = cpuid;
1155 snwsp->snws_cpu = cpuid;
1156 if (nwsp->nws_intr_event != NULL)
1157 snwsp->snws_flags |= NETISR_SNWS_FLAGS_INTR;
1158 NWS_UNLOCK(nwsp);
1159 counter++;
1160 }
1161 NETISR_RUNLOCK(&tracker);
1162 KASSERT(counter <= MAXCPU,
1163 ("sysctl_netisr_workstream: counter too big (%d)", counter));
1164 error = SYSCTL_OUT(req, snws_array, sizeof(*snws_array) * counter);
1165 free(snws_array, M_TEMP);
1166 return (error);
1167}
1168
1169SYSCTL_PROC(_net_isr, OID_AUTO, workstream,
1170 CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_workstream,
1171 "S,sysctl_netisr_workstream",
1172 "Return list of workstreams implemented by netisr");
1173
1174/*
1175 * Sysctl monitoring for netisr: query per-protocol data across all
1176 * workstreams.
1177 */
1178static int
1179sysctl_netisr_work(SYSCTL_HANDLER_ARGS)
1180{
1181 struct rm_priotracker tracker;
1182 struct sysctl_netisr_work *snwp, *snw_array;
1183 struct netisr_workstream *nwsp;
1184 struct netisr_proto *npp;
1185 struct netisr_work *nwp;
1186 u_int counter, cpuid, proto;
1187 int error;
1188
1189 if (req->newptr != NULL)
1190 return (EINVAL);
1191 snw_array = malloc(sizeof(*snw_array) * MAXCPU * NETISR_MAXPROT,
1192 M_TEMP, M_ZERO | M_WAITOK);
1193 counter = 0;
1194 NETISR_RLOCK(&tracker);
1195 for (cpuid = 0; cpuid < MAXCPU; cpuid++) {
1196 if (CPU_ABSENT(cpuid))
1197 continue;
1198 nwsp = DPCPU_ID_PTR(cpuid, nws);
1199 if (nwsp->nws_intr_event == NULL)
1200 continue;
1201 NWS_LOCK(nwsp);
1202 for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1203 npp = &netisr_proto[proto];
1204 if (npp->np_name == NULL)
1205 continue;
1206 nwp = &nwsp->nws_work[proto];
1207 snwp = &snw_array[counter];
1208 snwp->snw_version = sizeof(*snwp);
1209 snwp->snw_wsid = cpuid; /* See comment above. */
1210 snwp->snw_proto = proto;
1211 snwp->snw_len = nwp->nw_len;
1212 snwp->snw_watermark = nwp->nw_watermark;
1213 snwp->snw_dispatched = nwp->nw_dispatched;
1214 snwp->snw_hybrid_dispatched =
1215 nwp->nw_hybrid_dispatched;
1216 snwp->snw_qdrops = nwp->nw_qdrops;
1217 snwp->snw_queued = nwp->nw_queued;
1218 snwp->snw_handled = nwp->nw_handled;
1219 counter++;
1220 }
1221 NWS_UNLOCK(nwsp);
1222 }
1223 KASSERT(counter <= MAXCPU * NETISR_MAXPROT,
1224 ("sysctl_netisr_work: counter too big (%d)", counter));
1225 NETISR_RUNLOCK(&tracker);
1226 error = SYSCTL_OUT(req, snw_array, sizeof(*snw_array) * counter);
1227 free(snw_array, M_TEMP);
1228 return (error);
1229}
1230
1231SYSCTL_PROC(_net_isr, OID_AUTO, work,
1232 CTLFLAG_RD|CTLTYPE_STRUCT|CTLFLAG_MPSAFE, 0, 0, sysctl_netisr_work,
1233 "S,sysctl_netisr_work",
1234 "Return list of per-workstream, per-protocol work in netisr");
1235
1236#ifdef DDB
1237DB_SHOW_COMMAND(netisr, db_show_netisr)
1238{
1239 struct netisr_workstream *nwsp;
1240 struct netisr_work *nwp;
1241 int first, proto;
1242 u_int cpuid;
1243
1244 db_printf("%3s %6s %5s %5s %5s %8s %8s %8s %8s\n", "CPU", "Proto",
1245 "Len", "WMark", "Max", "Disp", "HDisp", "Drop", "Queue");
1246 for (cpuid = 0; cpuid <= mp_maxid; cpuid++) {
1247 if (CPU_ABSENT(cpuid))
1248 continue;
1249 nwsp = DPCPU_ID_PTR(cpuid, nws);
1250 if (nwsp->nws_intr_event == NULL)
1251 continue;
1252 first = 1;
1253 for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1254 if (netisr_proto[proto].np_handler == NULL)
1255 continue;
1256 nwp = &nwsp->nws_work[proto];
1257 if (first) {
1258 db_printf("%3d ", cpuid);
1259 first = 0;
1260 } else
1261 db_printf("%3s ", "");
1262 db_printf(
1263 "%6s %5d %5d %5d %8ju %8ju %8ju %8ju\n",
1264 netisr_proto[proto].np_name, nwp->nw_len,
1265 nwp->nw_watermark, nwp->nw_qlimit,
1266 nwp->nw_dispatched, nwp->nw_hybrid_dispatched,
1267 nwp->nw_qdrops, nwp->nw_queued);
1268 }
1269 }
1270}
1271#endif
116SYSINIT(start_netisr, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_netisr, NULL)