FreeBSD/Linux Kernel Cross Reference
sys/net/netisr.c
1 /*-
2 * Copyright (c) 2007-2009 Robert N. M. Watson
3 * All rights reserved.
4 *
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
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 *
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24 * SUCH DAMAGE.
25 */
26
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD: releng/8.0/sys/net/netisr.c 196019 2009-08-01 19:26:27Z rwatson $");
29
30 /*
31 * netisr is a packet dispatch service, allowing synchronous (directly
32 * dispatched) and asynchronous (deferred dispatch) processing of packets by
33 * registered protocol handlers. Callers pass a protocol identifier and
34 * packet to netisr, along with a direct dispatch hint, and work will either
35 * be immediately processed with the registered handler, or passed to a
36 * kernel software interrupt (SWI) thread for deferred dispatch. Callers
37 * will generally select one or the other based on:
38 *
39 * - Might directly dispatching a netisr handler lead to code reentrance or
40 * lock recursion, such as entering the socket code from the socket code.
41 * - Might directly dispatching a netisr handler lead to recursive
42 * processing, such as when decapsulating several wrapped layers of tunnel
43 * information (IPSEC within IPSEC within ...).
44 *
45 * Maintaining ordering for protocol streams is a critical design concern.
46 * Enforcing ordering limits the opportunity for concurrency, but maintains
47 * the strong ordering requirements found in some protocols, such as TCP. Of
48 * related concern is CPU affinity--it is desirable to process all data
49 * associated with a particular stream on the same CPU over time in order to
50 * avoid acquiring locks associated with the connection on different CPUs,
51 * keep connection data in one cache, and to generally encourage associated
52 * user threads to live on the same CPU as the stream. It's also desirable
53 * to avoid lock migration and contention where locks are associated with
54 * more than one flow.
55 *
56 * netisr supports several policy variations, represented by the
57 * NETISR_POLICY_* constants, allowing protocols to play a varying role in
58 * identifying flows, assigning work to CPUs, etc. These are described in
59 * detail in netisr.h.
60 */
61
62 #include "opt_ddb.h"
63 #include "opt_device_polling.h"
64
65 #include <sys/param.h>
66 #include <sys/bus.h>
67 #include <sys/kernel.h>
68 #include <sys/kthread.h>
69 #include <sys/interrupt.h>
70 #include <sys/lock.h>
71 #include <sys/mbuf.h>
72 #include <sys/mutex.h>
73 #include <sys/pcpu.h>
74 #include <sys/proc.h>
75 #include <sys/rmlock.h>
76 #include <sys/sched.h>
77 #include <sys/smp.h>
78 #include <sys/socket.h>
79 #include <sys/sysctl.h>
80 #include <sys/systm.h>
81
82 #ifdef DDB
83 #include <ddb/ddb.h>
84 #endif
85
86 #include <net/if.h>
87 #include <net/if_var.h>
88 #include <net/netisr.h>
89 #include <net/vnet.h>
90
91 /*-
92 * Synchronize use and modification of the registered netisr data structures;
93 * acquire a read lock while modifying the set of registered protocols to
94 * prevent partially registered or unregistered protocols from being run.
95 *
96 * The following data structures and fields are protected by this lock:
97 *
98 * - The np array, including all fields of struct netisr_proto.
99 * - The nws array, including all fields of struct netisr_worker.
100 * - The nws_array array.
101 *
102 * Note: the NETISR_LOCKING define controls whether read locks are acquired
103 * in packet processing paths requiring netisr registration stability. This
104 * is disabled by default as it can lead to a measurable performance
105 * degradation even with rmlocks (3%-6% for loopback ping-pong traffic), and
106 * because netisr registration and unregistration is extremely rare at
107 * runtime. If it becomes more common, this decision should be revisited.
108 *
109 * XXXRW: rmlocks don't support assertions.
110 */
111 static struct rmlock netisr_rmlock;
112 #define NETISR_LOCK_INIT() rm_init_flags(&netisr_rmlock, "netisr", \
113 RM_NOWITNESS)
114 #define NETISR_LOCK_ASSERT()
115 #define NETISR_RLOCK(tracker) rm_rlock(&netisr_rmlock, (tracker))
116 #define NETISR_RUNLOCK(tracker) rm_runlock(&netisr_rmlock, (tracker))
117 #define NETISR_WLOCK() rm_wlock(&netisr_rmlock)
118 #define NETISR_WUNLOCK() rm_wunlock(&netisr_rmlock)
119 /* #define NETISR_LOCKING */
120
121 SYSCTL_NODE(_net, OID_AUTO, isr, CTLFLAG_RW, 0, "netisr");
122
123 /*-
124 * Three direct dispatch policies are supported:
125 *
126 * - Always defer: all work is scheduled for a netisr, regardless of context.
127 * (!direct)
128 *
129 * - Hybrid: if the executing context allows direct dispatch, and we're
130 * running on the CPU the work would be done on, then direct dispatch if it
131 * wouldn't violate ordering constraints on the workstream.
132 * (direct && !direct_force)
133 *
134 * - Always direct: if the executing context allows direct dispatch, always
135 * direct dispatch. (direct && direct_force)
136 *
137 * Notice that changing the global policy could lead to short periods of
138 * misordered processing, but this is considered acceptable as compared to
139 * the complexity of enforcing ordering during policy changes.
140 */
141 static int netisr_direct_force = 1; /* Always direct dispatch. */
142 TUNABLE_INT("net.isr.direct_force", &netisr_direct_force);
143 SYSCTL_INT(_net_isr, OID_AUTO, direct_force, CTLFLAG_RW,
144 &netisr_direct_force, 0, "Force direct dispatch");
145
146 static int netisr_direct = 1; /* Enable direct dispatch. */
147 TUNABLE_INT("net.isr.direct", &netisr_direct);
148 SYSCTL_INT(_net_isr, OID_AUTO, direct, CTLFLAG_RW,
149 &netisr_direct, 0, "Enable direct dispatch");
150
151 /*
152 * Allow the administrator to limit the number of threads (CPUs) to use for
153 * netisr. We don't check netisr_maxthreads before creating the thread for
154 * CPU 0, so in practice we ignore values <= 1. This must be set at boot.
155 * We will create at most one thread per CPU.
156 */
157 static int netisr_maxthreads = -1; /* Max number of threads. */
158 TUNABLE_INT("net.isr.maxthreads", &netisr_maxthreads);
159 SYSCTL_INT(_net_isr, OID_AUTO, maxthreads, CTLFLAG_RD,
160 &netisr_maxthreads, 0,
161 "Use at most this many CPUs for netisr processing");
162
163 static int netisr_bindthreads = 0; /* Bind threads to CPUs. */
164 TUNABLE_INT("net.isr.bindthreads", &netisr_bindthreads);
165 SYSCTL_INT(_net_isr, OID_AUTO, bindthreads, CTLFLAG_RD,
166 &netisr_bindthreads, 0, "Bind netisr threads to CPUs.");
167
168 /*
169 * Limit per-workstream queues to at most net.isr.maxqlimit, both for initial
170 * configuration and later modification using netisr_setqlimit().
171 */
172 #define NETISR_DEFAULT_MAXQLIMIT 10240
173 static u_int netisr_maxqlimit = NETISR_DEFAULT_MAXQLIMIT;
174 TUNABLE_INT("net.isr.maxqlimit", &netisr_maxqlimit);
175 SYSCTL_INT(_net_isr, OID_AUTO, maxqlimit, CTLFLAG_RD,
176 &netisr_maxqlimit, 0,
177 "Maximum netisr per-protocol, per-CPU queue depth.");
178
179 /*
180 * The default per-workstream queue limit for protocols that don't initialize
181 * the nh_qlimit field of their struct netisr_handler. If this is set above
182 * netisr_maxqlimit, we truncate it to the maximum during boot.
183 */
184 #define NETISR_DEFAULT_DEFAULTQLIMIT 256
185 static u_int netisr_defaultqlimit = NETISR_DEFAULT_DEFAULTQLIMIT;
186 TUNABLE_INT("net.isr.defaultqlimit", &netisr_defaultqlimit);
187 SYSCTL_INT(_net_isr, OID_AUTO, defaultqlimit, CTLFLAG_RD,
188 &netisr_defaultqlimit, 0,
189 "Default netisr per-protocol, per-CPU queue limit if not set by protocol");
190
191 /*
192 * Each protocol is described by a struct netisr_proto, which holds all
193 * global per-protocol information. This data structure is set up by
194 * netisr_register(), and derived from the public struct netisr_handler.
195 */
196 struct netisr_proto {
197 const char *np_name; /* Character string protocol name. */
198 netisr_handler_t *np_handler; /* Protocol handler. */
199 netisr_m2flow_t *np_m2flow; /* Query flow for untagged packet. */
200 netisr_m2cpuid_t *np_m2cpuid; /* Query CPU to process packet on. */
201 netisr_drainedcpu_t *np_drainedcpu; /* Callback when drained a queue. */
202 u_int np_qlimit; /* Maximum per-CPU queue depth. */
203 u_int np_policy; /* Work placement policy. */
204 };
205
206 #define NETISR_MAXPROT 16 /* Compile-time limit. */
207
208 /*
209 * The np array describes all registered protocols, indexed by protocol
210 * number.
211 */
212 static struct netisr_proto np[NETISR_MAXPROT];
213
214 /*
215 * Protocol-specific work for each workstream is described by struct
216 * netisr_work. Each work descriptor consists of an mbuf queue and
217 * statistics.
218 */
219 struct netisr_work {
220 /*
221 * Packet queue, linked by m_nextpkt.
222 */
223 struct mbuf *nw_head;
224 struct mbuf *nw_tail;
225 u_int nw_len;
226 u_int nw_qlimit;
227 u_int nw_watermark;
228
229 /*
230 * Statistics -- written unlocked, but mostly from curcpu.
231 */
232 u_int64_t nw_dispatched; /* Number of direct dispatches. */
233 u_int64_t nw_hybrid_dispatched; /* "" hybrid dispatches. */
234 u_int64_t nw_qdrops; /* "" drops. */
235 u_int64_t nw_queued; /* "" enqueues. */
236 u_int64_t nw_handled; /* "" handled in worker. */
237 };
238
239 /*
240 * Workstreams hold a set of ordered work across each protocol, and are
241 * described by netisr_workstream. Each workstream is associated with a
242 * worker thread, which in turn is pinned to a CPU. Work associated with a
243 * workstream can be processd in other threads during direct dispatch;
244 * concurrent processing is prevented by the NWS_RUNNING flag, which
245 * indicates that a thread is already processing the work queue.
246 */
247 struct netisr_workstream {
248 struct intr_event *nws_intr_event; /* Handler for stream. */
249 void *nws_swi_cookie; /* swi(9) cookie for stream. */
250 struct mtx nws_mtx; /* Synchronize work. */
251 u_int nws_cpu; /* CPU pinning. */
252 u_int nws_flags; /* Wakeup flags. */
253 u_int nws_pendingbits; /* Scheduled protocols. */
254
255 /*
256 * Each protocol has per-workstream data.
257 */
258 struct netisr_work nws_work[NETISR_MAXPROT];
259 } __aligned(CACHE_LINE_SIZE);
260
261 /*
262 * Per-CPU workstream data.
263 */
264 DPCPU_DEFINE(struct netisr_workstream, nws);
265
266 /*
267 * Map contiguous values between 0 and nws_count into CPU IDs appropriate for
268 * accessing workstreams. This allows constructions of the form
269 * DPCPU_ID_GET(nws_array[arbitraryvalue % nws_count], nws).
270 */
271 static u_int nws_array[MAXCPU];
272
273 /*
274 * Number of registered workstreams. Will be at most the number of running
275 * CPUs once fully started.
276 */
277 static u_int nws_count;
278 SYSCTL_INT(_net_isr, OID_AUTO, numthreads, CTLFLAG_RD,
279 &nws_count, 0, "Number of extant netisr threads.");
280
281 /*
282 * Per-workstream flags.
283 */
284 #define NWS_RUNNING 0x00000001 /* Currently running in a thread. */
285 #define NWS_DISPATCHING 0x00000002 /* Currently being direct-dispatched. */
286 #define NWS_SCHEDULED 0x00000004 /* Signal issued. */
287
288 /*
289 * Synchronization for each workstream: a mutex protects all mutable fields
290 * in each stream, including per-protocol state (mbuf queues). The SWI is
291 * woken up if asynchronous dispatch is required.
292 */
293 #define NWS_LOCK(s) mtx_lock(&(s)->nws_mtx)
294 #define NWS_LOCK_ASSERT(s) mtx_assert(&(s)->nws_mtx, MA_OWNED)
295 #define NWS_UNLOCK(s) mtx_unlock(&(s)->nws_mtx)
296 #define NWS_SIGNAL(s) swi_sched((s)->nws_swi_cookie, 0)
297
298 /*
299 * Utility routines for protocols that implement their own mapping of flows
300 * to CPUs.
301 */
302 u_int
303 netisr_get_cpucount(void)
304 {
305
306 return (nws_count);
307 }
308
309 u_int
310 netisr_get_cpuid(u_int cpunumber)
311 {
312
313 KASSERT(cpunumber < nws_count, ("%s: %u > %u", __func__, cpunumber,
314 nws_count));
315
316 return (nws_array[cpunumber]);
317 }
318
319 /*
320 * The default implementation of -> CPU ID mapping.
321 *
322 * Non-static so that protocols can use it to map their own work to specific
323 * CPUs in a manner consistent to netisr for affinity purposes.
324 */
325 u_int
326 netisr_default_flow2cpu(u_int flowid)
327 {
328
329 return (nws_array[flowid % nws_count]);
330 }
331
332 /*
333 * Register a new netisr handler, which requires initializing per-protocol
334 * fields for each workstream. All netisr work is briefly suspended while
335 * the protocol is installed.
336 */
337 void
338 netisr_register(const struct netisr_handler *nhp)
339 {
340 struct netisr_work *npwp;
341 const char *name;
342 u_int i, proto;
343
344 proto = nhp->nh_proto;
345 name = nhp->nh_name;
346
347 /*
348 * Test that the requested registration is valid.
349 */
350 KASSERT(nhp->nh_name != NULL,
351 ("%s: nh_name NULL for %u", __func__, proto));
352 KASSERT(nhp->nh_handler != NULL,
353 ("%s: nh_handler NULL for %s", __func__, name));
354 KASSERT(nhp->nh_policy == NETISR_POLICY_SOURCE ||
355 nhp->nh_policy == NETISR_POLICY_FLOW ||
356 nhp->nh_policy == NETISR_POLICY_CPU,
357 ("%s: unsupported nh_policy %u for %s", __func__,
358 nhp->nh_policy, name));
359 KASSERT(nhp->nh_policy == NETISR_POLICY_FLOW ||
360 nhp->nh_m2flow == NULL,
361 ("%s: nh_policy != FLOW but m2flow defined for %s", __func__,
362 name));
363 KASSERT(nhp->nh_policy == NETISR_POLICY_CPU || nhp->nh_m2cpuid == NULL,
364 ("%s: nh_policy != CPU but m2cpuid defined for %s", __func__,
365 name));
366 KASSERT(nhp->nh_policy != NETISR_POLICY_CPU || nhp->nh_m2cpuid != NULL,
367 ("%s: nh_policy == CPU but m2cpuid not defined for %s", __func__,
368 name));
369 KASSERT(proto < NETISR_MAXPROT,
370 ("%s(%u, %s): protocol too big", __func__, proto, name));
371
372 /*
373 * Test that no existing registration exists for this protocol.
374 */
375 NETISR_WLOCK();
376 KASSERT(np[proto].np_name == NULL,
377 ("%s(%u, %s): name present", __func__, proto, name));
378 KASSERT(np[proto].np_handler == NULL,
379 ("%s(%u, %s): handler present", __func__, proto, name));
380
381 np[proto].np_name = name;
382 np[proto].np_handler = nhp->nh_handler;
383 np[proto].np_m2flow = nhp->nh_m2flow;
384 np[proto].np_m2cpuid = nhp->nh_m2cpuid;
385 np[proto].np_drainedcpu = nhp->nh_drainedcpu;
386 if (nhp->nh_qlimit == 0)
387 np[proto].np_qlimit = netisr_defaultqlimit;
388 else if (nhp->nh_qlimit > netisr_maxqlimit) {
389 printf("%s: %s requested queue limit %u capped to "
390 "net.isr.maxqlimit %u\n", __func__, name, nhp->nh_qlimit,
391 netisr_maxqlimit);
392 np[proto].np_qlimit = netisr_maxqlimit;
393 } else
394 np[proto].np_qlimit = nhp->nh_qlimit;
395 np[proto].np_policy = nhp->nh_policy;
396 for (i = 0; i <= mp_maxid; i++) {
397 if (CPU_ABSENT(i))
398 continue;
399 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
400 bzero(npwp, sizeof(*npwp));
401 npwp->nw_qlimit = np[proto].np_qlimit;
402 }
403 NETISR_WUNLOCK();
404 }
405
406 /*
407 * Clear drop counters across all workstreams for a protocol.
408 */
409 void
410 netisr_clearqdrops(const struct netisr_handler *nhp)
411 {
412 struct netisr_work *npwp;
413 #ifdef INVARIANTS
414 const char *name;
415 #endif
416 u_int i, proto;
417
418 proto = nhp->nh_proto;
419 #ifdef INVARIANTS
420 name = nhp->nh_name;
421 #endif
422 KASSERT(proto < NETISR_MAXPROT,
423 ("%s(%u): protocol too big for %s", __func__, proto, name));
424
425 NETISR_WLOCK();
426 KASSERT(np[proto].np_handler != NULL,
427 ("%s(%u): protocol not registered for %s", __func__, proto,
428 name));
429
430 for (i = 0; i <= mp_maxid; i++) {
431 if (CPU_ABSENT(i))
432 continue;
433 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
434 npwp->nw_qdrops = 0;
435 }
436 NETISR_WUNLOCK();
437 }
438
439 /*
440 * Query the current drop counters across all workstreams for a protocol.
441 */
442 void
443 netisr_getqdrops(const struct netisr_handler *nhp, u_int64_t *qdropp)
444 {
445 struct netisr_work *npwp;
446 struct rm_priotracker tracker;
447 #ifdef INVARIANTS
448 const char *name;
449 #endif
450 u_int i, proto;
451
452 *qdropp = 0;
453 proto = nhp->nh_proto;
454 #ifdef INVARIANTS
455 name = nhp->nh_name;
456 #endif
457 KASSERT(proto < NETISR_MAXPROT,
458 ("%s(%u): protocol too big for %s", __func__, proto, name));
459
460 NETISR_RLOCK(&tracker);
461 KASSERT(np[proto].np_handler != NULL,
462 ("%s(%u): protocol not registered for %s", __func__, proto,
463 name));
464
465 for (i = 0; i <= mp_maxid; i++) {
466 if (CPU_ABSENT(i))
467 continue;
468 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
469 *qdropp += npwp->nw_qdrops;
470 }
471 NETISR_RUNLOCK(&tracker);
472 }
473
474 /*
475 * Query the current queue limit for per-workstream queues for a protocol.
476 */
477 void
478 netisr_getqlimit(const struct netisr_handler *nhp, u_int *qlimitp)
479 {
480 struct rm_priotracker tracker;
481 #ifdef INVARIANTS
482 const char *name;
483 #endif
484 u_int proto;
485
486 proto = nhp->nh_proto;
487 #ifdef INVARIANTS
488 name = nhp->nh_name;
489 #endif
490 KASSERT(proto < NETISR_MAXPROT,
491 ("%s(%u): protocol too big for %s", __func__, proto, name));
492
493 NETISR_RLOCK(&tracker);
494 KASSERT(np[proto].np_handler != NULL,
495 ("%s(%u): protocol not registered for %s", __func__, proto,
496 name));
497 *qlimitp = np[proto].np_qlimit;
498 NETISR_RUNLOCK(&tracker);
499 }
500
501 /*
502 * Update the queue limit across per-workstream queues for a protocol. We
503 * simply change the limits, and don't drain overflowed packets as they will
504 * (hopefully) take care of themselves shortly.
505 */
506 int
507 netisr_setqlimit(const struct netisr_handler *nhp, u_int qlimit)
508 {
509 struct netisr_work *npwp;
510 #ifdef INVARIANTS
511 const char *name;
512 #endif
513 u_int i, proto;
514
515 if (qlimit > netisr_maxqlimit)
516 return (EINVAL);
517
518 proto = nhp->nh_proto;
519 #ifdef INVARIANTS
520 name = nhp->nh_name;
521 #endif
522 KASSERT(proto < NETISR_MAXPROT,
523 ("%s(%u): protocol too big for %s", __func__, proto, name));
524
525 NETISR_WLOCK();
526 KASSERT(np[proto].np_handler != NULL,
527 ("%s(%u): protocol not registered for %s", __func__, proto,
528 name));
529
530 np[proto].np_qlimit = qlimit;
531 for (i = 0; i <= mp_maxid; i++) {
532 if (CPU_ABSENT(i))
533 continue;
534 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
535 npwp->nw_qlimit = qlimit;
536 }
537 NETISR_WUNLOCK();
538 return (0);
539 }
540
541 /*
542 * Drain all packets currently held in a particular protocol work queue.
543 */
544 static void
545 netisr_drain_proto(struct netisr_work *npwp)
546 {
547 struct mbuf *m;
548
549 /*
550 * We would assert the lock on the workstream but it's not passed in.
551 */
552 while ((m = npwp->nw_head) != NULL) {
553 npwp->nw_head = m->m_nextpkt;
554 m->m_nextpkt = NULL;
555 if (npwp->nw_head == NULL)
556 npwp->nw_tail = NULL;
557 npwp->nw_len--;
558 m_freem(m);
559 }
560 KASSERT(npwp->nw_tail == NULL, ("%s: tail", __func__));
561 KASSERT(npwp->nw_len == 0, ("%s: len", __func__));
562 }
563
564 /*
565 * Remove the registration of a network protocol, which requires clearing
566 * per-protocol fields across all workstreams, including freeing all mbufs in
567 * the queues at time of unregister. All work in netisr is briefly suspended
568 * while this takes place.
569 */
570 void
571 netisr_unregister(const struct netisr_handler *nhp)
572 {
573 struct netisr_work *npwp;
574 #ifdef INVARIANTS
575 const char *name;
576 #endif
577 u_int i, proto;
578
579 proto = nhp->nh_proto;
580 #ifdef INVARIANTS
581 name = nhp->nh_name;
582 #endif
583 KASSERT(proto < NETISR_MAXPROT,
584 ("%s(%u): protocol too big for %s", __func__, proto, name));
585
586 NETISR_WLOCK();
587 KASSERT(np[proto].np_handler != NULL,
588 ("%s(%u): protocol not registered for %s", __func__, proto,
589 name));
590
591 np[proto].np_name = NULL;
592 np[proto].np_handler = NULL;
593 np[proto].np_m2flow = NULL;
594 np[proto].np_m2cpuid = NULL;
595 np[proto].np_qlimit = 0;
596 np[proto].np_policy = 0;
597 for (i = 0; i <= mp_maxid; i++) {
598 if (CPU_ABSENT(i))
599 continue;
600 npwp = &(DPCPU_ID_PTR(i, nws))->nws_work[proto];
601 netisr_drain_proto(npwp);
602 bzero(npwp, sizeof(*npwp));
603 }
604 NETISR_WUNLOCK();
605 }
606
607 /*
608 * Look up the workstream given a packet and source identifier. Do this by
609 * checking the protocol's policy, and optionally call out to the protocol
610 * for assistance if required.
611 */
612 static struct mbuf *
613 netisr_select_cpuid(struct netisr_proto *npp, uintptr_t source,
614 struct mbuf *m, u_int *cpuidp)
615 {
616 struct ifnet *ifp;
617
618 NETISR_LOCK_ASSERT();
619
620 /*
621 * In the event we have only one worker, shortcut and deliver to it
622 * without further ado.
623 */
624 if (nws_count == 1) {
625 *cpuidp = nws_array[0];
626 return (m);
627 }
628
629 /*
630 * What happens next depends on the policy selected by the protocol.
631 * If we want to support per-interface policies, we should do that
632 * here first.
633 */
634 switch (npp->np_policy) {
635 case NETISR_POLICY_CPU:
636 return (npp->np_m2cpuid(m, source, cpuidp));
637
638 case NETISR_POLICY_FLOW:
639 if (!(m->m_flags & M_FLOWID) && npp->np_m2flow != NULL) {
640 m = npp->np_m2flow(m, source);
641 if (m == NULL)
642 return (NULL);
643 }
644 if (m->m_flags & M_FLOWID) {
645 *cpuidp =
646 netisr_default_flow2cpu(m->m_pkthdr.flowid);
647 return (m);
648 }
649 /* FALLTHROUGH */
650
651 case NETISR_POLICY_SOURCE:
652 ifp = m->m_pkthdr.rcvif;
653 if (ifp != NULL)
654 *cpuidp = nws_array[(ifp->if_index + source) %
655 nws_count];
656 else
657 *cpuidp = nws_array[source % nws_count];
658 return (m);
659
660 default:
661 panic("%s: invalid policy %u for %s", __func__,
662 npp->np_policy, npp->np_name);
663 }
664 }
665
666 /*
667 * Process packets associated with a workstream and protocol. For reasons of
668 * fairness, we process up to one complete netisr queue at a time, moving the
669 * queue to a stack-local queue for processing, but do not loop refreshing
670 * from the global queue. The caller is responsible for deciding whether to
671 * loop, and for setting the NWS_RUNNING flag. The passed workstream will be
672 * locked on entry and relocked before return, but will be released while
673 * processing. The number of packets processed is returned.
674 */
675 static u_int
676 netisr_process_workstream_proto(struct netisr_workstream *nwsp, u_int proto)
677 {
678 struct netisr_work local_npw, *npwp;
679 u_int handled;
680 struct mbuf *m;
681
682 NETISR_LOCK_ASSERT();
683 NWS_LOCK_ASSERT(nwsp);
684
685 KASSERT(nwsp->nws_flags & NWS_RUNNING,
686 ("%s(%u): not running", __func__, proto));
687 KASSERT(proto >= 0 && proto < NETISR_MAXPROT,
688 ("%s(%u): invalid proto\n", __func__, proto));
689
690 npwp = &nwsp->nws_work[proto];
691 if (npwp->nw_len == 0)
692 return (0);
693
694 /*
695 * Move the global work queue to a thread-local work queue.
696 *
697 * Notice that this means the effective maximum length of the queue
698 * is actually twice that of the maximum queue length specified in
699 * the protocol registration call.
700 */
701 handled = npwp->nw_len;
702 local_npw = *npwp;
703 npwp->nw_head = NULL;
704 npwp->nw_tail = NULL;
705 npwp->nw_len = 0;
706 nwsp->nws_pendingbits &= ~(1 << proto);
707 NWS_UNLOCK(nwsp);
708 while ((m = local_npw.nw_head) != NULL) {
709 local_npw.nw_head = m->m_nextpkt;
710 m->m_nextpkt = NULL;
711 if (local_npw.nw_head == NULL)
712 local_npw.nw_tail = NULL;
713 local_npw.nw_len--;
714 VNET_ASSERT(m->m_pkthdr.rcvif != NULL);
715 CURVNET_SET(m->m_pkthdr.rcvif->if_vnet);
716 np[proto].np_handler(m);
717 CURVNET_RESTORE();
718 }
719 KASSERT(local_npw.nw_len == 0,
720 ("%s(%u): len %u", __func__, proto, local_npw.nw_len));
721 if (np[proto].np_drainedcpu)
722 np[proto].np_drainedcpu(nwsp->nws_cpu);
723 NWS_LOCK(nwsp);
724 npwp->nw_handled += handled;
725 return (handled);
726 }
727
728 /*
729 * SWI handler for netisr -- processes prackets in a set of workstreams that
730 * it owns, woken up by calls to NWS_SIGNAL(). If this workstream is already
731 * being direct dispatched, go back to sleep and wait for the dispatching
732 * thread to wake us up again.
733 */
734 static void
735 swi_net(void *arg)
736 {
737 #ifdef NETISR_LOCKING
738 struct rm_priotracker tracker;
739 #endif
740 struct netisr_workstream *nwsp;
741 u_int bits, prot;
742
743 nwsp = arg;
744
745 #ifdef DEVICE_POLLING
746 KASSERT(nws_count == 1,
747 ("%s: device_polling but nws_count != 1", __func__));
748 netisr_poll();
749 #endif
750 #ifdef NETISR_LOCKING
751 NETISR_RLOCK(&tracker);
752 #endif
753 NWS_LOCK(nwsp);
754 KASSERT(!(nwsp->nws_flags & NWS_RUNNING), ("swi_net: running"));
755 if (nwsp->nws_flags & NWS_DISPATCHING)
756 goto out;
757 nwsp->nws_flags |= NWS_RUNNING;
758 nwsp->nws_flags &= ~NWS_SCHEDULED;
759 while ((bits = nwsp->nws_pendingbits) != 0) {
760 while ((prot = ffs(bits)) != 0) {
761 prot--;
762 bits &= ~(1 << prot);
763 (void)netisr_process_workstream_proto(nwsp, prot);
764 }
765 }
766 nwsp->nws_flags &= ~NWS_RUNNING;
767 out:
768 NWS_UNLOCK(nwsp);
769 #ifdef NETISR_LOCKING
770 NETISR_RUNLOCK(&tracker);
771 #endif
772 #ifdef DEVICE_POLLING
773 netisr_pollmore();
774 #endif
775 }
776
777 static int
778 netisr_queue_workstream(struct netisr_workstream *nwsp, u_int proto,
779 struct netisr_work *npwp, struct mbuf *m, int *dosignalp)
780 {
781
782 NWS_LOCK_ASSERT(nwsp);
783
784 *dosignalp = 0;
785 if (npwp->nw_len < npwp->nw_qlimit) {
786 m->m_nextpkt = NULL;
787 if (npwp->nw_head == NULL) {
788 npwp->nw_head = m;
789 npwp->nw_tail = m;
790 } else {
791 npwp->nw_tail->m_nextpkt = m;
792 npwp->nw_tail = m;
793 }
794 npwp->nw_len++;
795 if (npwp->nw_len > npwp->nw_watermark)
796 npwp->nw_watermark = npwp->nw_len;
797 nwsp->nws_pendingbits |= (1 << proto);
798 if (!(nwsp->nws_flags &
799 (NWS_RUNNING | NWS_DISPATCHING | NWS_SCHEDULED))) {
800 nwsp->nws_flags |= NWS_SCHEDULED;
801 *dosignalp = 1; /* Defer until unlocked. */
802 }
803 npwp->nw_queued++;
804 return (0);
805 } else {
806 m_freem(m);
807 npwp->nw_qdrops++;
808 return (ENOBUFS);
809 }
810 }
811
812 static int
813 netisr_queue_internal(u_int proto, struct mbuf *m, u_int cpuid)
814 {
815 struct netisr_workstream *nwsp;
816 struct netisr_work *npwp;
817 int dosignal, error;
818
819 #ifdef NETISR_LOCKING
820 NETISR_LOCK_ASSERT();
821 #endif
822 KASSERT(cpuid <= mp_maxid, ("%s: cpuid too big (%u, %u)", __func__,
823 cpuid, mp_maxid));
824 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
825
826 dosignal = 0;
827 error = 0;
828 nwsp = DPCPU_ID_PTR(cpuid, nws);
829 npwp = &nwsp->nws_work[proto];
830 NWS_LOCK(nwsp);
831 error = netisr_queue_workstream(nwsp, proto, npwp, m, &dosignal);
832 NWS_UNLOCK(nwsp);
833 if (dosignal)
834 NWS_SIGNAL(nwsp);
835 return (error);
836 }
837
838 int
839 netisr_queue_src(u_int proto, uintptr_t source, struct mbuf *m)
840 {
841 #ifdef NETISR_LOCKING
842 struct rm_priotracker tracker;
843 #endif
844 u_int cpuid;
845 int error;
846
847 KASSERT(proto < NETISR_MAXPROT,
848 ("%s: invalid proto %u", __func__, proto));
849
850 #ifdef NETISR_LOCKING
851 NETISR_RLOCK(&tracker);
852 #endif
853 KASSERT(np[proto].np_handler != NULL,
854 ("%s: invalid proto %u", __func__, proto));
855
856 m = netisr_select_cpuid(&np[proto], source, m, &cpuid);
857 if (m != NULL) {
858 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__,
859 cpuid));
860 error = netisr_queue_internal(proto, m, cpuid);
861 } else
862 error = ENOBUFS;
863 #ifdef NETISR_LOCKING
864 NETISR_RUNLOCK(&tracker);
865 #endif
866 return (error);
867 }
868
869 int
870 netisr_queue(u_int proto, struct mbuf *m)
871 {
872
873 return (netisr_queue_src(proto, 0, m));
874 }
875
876 /*
877 * Dispatch a packet for netisr processing, direct dispatch permitted by
878 * calling context.
879 */
880 int
881 netisr_dispatch_src(u_int proto, uintptr_t source, struct mbuf *m)
882 {
883 #ifdef NETISR_LOCKING
884 struct rm_priotracker tracker;
885 #endif
886 struct netisr_workstream *nwsp;
887 struct netisr_work *npwp;
888 int dosignal, error;
889 u_int cpuid;
890
891 /*
892 * If direct dispatch is entirely disabled, fall back on queueing.
893 */
894 if (!netisr_direct)
895 return (netisr_queue_src(proto, source, m));
896
897 KASSERT(proto < NETISR_MAXPROT,
898 ("%s: invalid proto %u", __func__, proto));
899 #ifdef NETISR_LOCKING
900 NETISR_RLOCK(&tracker);
901 #endif
902 KASSERT(np[proto].np_handler != NULL,
903 ("%s: invalid proto %u", __func__, proto));
904
905 /*
906 * If direct dispatch is forced, then unconditionally dispatch
907 * without a formal CPU selection. Borrow the current CPU's stats,
908 * even if there's no worker on it. In this case we don't update
909 * nws_flags because all netisr processing will be source ordered due
910 * to always being forced to directly dispatch.
911 */
912 if (netisr_direct_force) {
913 nwsp = DPCPU_PTR(nws);
914 npwp = &nwsp->nws_work[proto];
915 npwp->nw_dispatched++;
916 npwp->nw_handled++;
917 np[proto].np_handler(m);
918 error = 0;
919 goto out_unlock;
920 }
921
922 /*
923 * Otherwise, we execute in a hybrid mode where we will try to direct
924 * dispatch if we're on the right CPU and the netisr worker isn't
925 * already running.
926 */
927 m = netisr_select_cpuid(&np[proto], source, m, &cpuid);
928 if (m == NULL) {
929 error = ENOBUFS;
930 goto out_unlock;
931 }
932 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
933 sched_pin();
934 if (cpuid != curcpu)
935 goto queue_fallback;
936 nwsp = DPCPU_PTR(nws);
937 npwp = &nwsp->nws_work[proto];
938
939 /*-
940 * We are willing to direct dispatch only if three conditions hold:
941 *
942 * (1) The netisr worker isn't already running,
943 * (2) Another thread isn't already directly dispatching, and
944 * (3) The netisr hasn't already been woken up.
945 */
946 NWS_LOCK(nwsp);
947 if (nwsp->nws_flags & (NWS_RUNNING | NWS_DISPATCHING | NWS_SCHEDULED)) {
948 error = netisr_queue_workstream(nwsp, proto, npwp, m,
949 &dosignal);
950 NWS_UNLOCK(nwsp);
951 if (dosignal)
952 NWS_SIGNAL(nwsp);
953 goto out_unpin;
954 }
955
956 /*
957 * The current thread is now effectively the netisr worker, so set
958 * the dispatching flag to prevent concurrent processing of the
959 * stream from another thread (even the netisr worker), which could
960 * otherwise lead to effective misordering of the stream.
961 */
962 nwsp->nws_flags |= NWS_DISPATCHING;
963 NWS_UNLOCK(nwsp);
964 np[proto].np_handler(m);
965 NWS_LOCK(nwsp);
966 nwsp->nws_flags &= ~NWS_DISPATCHING;
967 npwp->nw_handled++;
968 npwp->nw_hybrid_dispatched++;
969
970 /*
971 * If other work was enqueued by another thread while we were direct
972 * dispatching, we need to signal the netisr worker to do that work.
973 * In the future, we might want to do some of that work in the
974 * current thread, rather than trigger further context switches. If
975 * so, we'll want to establish a reasonable bound on the work done in
976 * the "borrowed" context.
977 */
978 if (nwsp->nws_pendingbits != 0) {
979 nwsp->nws_flags |= NWS_SCHEDULED;
980 dosignal = 1;
981 } else
982 dosignal = 0;
983 NWS_UNLOCK(nwsp);
984 if (dosignal)
985 NWS_SIGNAL(nwsp);
986 error = 0;
987 goto out_unpin;
988
989 queue_fallback:
990 error = netisr_queue_internal(proto, m, cpuid);
991 out_unpin:
992 sched_unpin();
993 out_unlock:
994 #ifdef NETISR_LOCKING
995 NETISR_RUNLOCK(&tracker);
996 #endif
997 return (error);
998 }
999
1000 int
1001 netisr_dispatch(u_int proto, struct mbuf *m)
1002 {
1003
1004 return (netisr_dispatch_src(proto, 0, m));
1005 }
1006
1007 #ifdef DEVICE_POLLING
1008 /*
1009 * Kernel polling borrows a netisr thread to run interface polling in; this
1010 * function allows kernel polling to request that the netisr thread be
1011 * scheduled even if no packets are pending for protocols.
1012 */
1013 void
1014 netisr_sched_poll(void)
1015 {
1016 struct netisr_workstream *nwsp;
1017
1018 nwsp = DPCPU_ID_PTR(nws_array[0], nws);
1019 NWS_SIGNAL(nwsp);
1020 }
1021 #endif
1022
1023 static void
1024 netisr_start_swi(u_int cpuid, struct pcpu *pc)
1025 {
1026 char swiname[12];
1027 struct netisr_workstream *nwsp;
1028 int error;
1029
1030 KASSERT(!CPU_ABSENT(cpuid), ("%s: CPU %u absent", __func__, cpuid));
1031
1032 nwsp = DPCPU_ID_PTR(cpuid, nws);
1033 mtx_init(&nwsp->nws_mtx, "netisr_mtx", NULL, MTX_DEF);
1034 nwsp->nws_cpu = cpuid;
1035 snprintf(swiname, sizeof(swiname), "netisr %u", cpuid);
1036 error = swi_add(&nwsp->nws_intr_event, swiname, swi_net, nwsp,
1037 SWI_NET, INTR_MPSAFE, &nwsp->nws_swi_cookie);
1038 if (error)
1039 panic("%s: swi_add %d", __func__, error);
1040 pc->pc_netisr = nwsp->nws_intr_event;
1041 if (netisr_bindthreads) {
1042 error = intr_event_bind(nwsp->nws_intr_event, cpuid);
1043 if (error != 0)
1044 printf("%s: cpu %u: intr_event_bind: %d", __func__,
1045 cpuid, error);
1046 }
1047 NETISR_WLOCK();
1048 nws_array[nws_count] = nwsp->nws_cpu;
1049 nws_count++;
1050 NETISR_WUNLOCK();
1051 }
1052
1053 /*
1054 * Initialize the netisr subsystem. We rely on BSS and static initialization
1055 * of most fields in global data structures.
1056 *
1057 * Start a worker thread for the boot CPU so that we can support network
1058 * traffic immediately in case the network stack is used before additional
1059 * CPUs are started (for example, diskless boot).
1060 */
1061 static void
1062 netisr_init(void *arg)
1063 {
1064
1065 KASSERT(curcpu == 0, ("%s: not on CPU 0", __func__));
1066
1067 NETISR_LOCK_INIT();
1068 if (netisr_maxthreads < 1)
1069 netisr_maxthreads = 1;
1070 if (netisr_maxthreads > mp_ncpus) {
1071 printf("netisr2: forcing maxthreads from %d to %d\n",
1072 netisr_maxthreads, mp_ncpus);
1073 netisr_maxthreads = mp_ncpus;
1074 }
1075 if (netisr_defaultqlimit > netisr_maxqlimit) {
1076 printf("netisr2: forcing defaultqlimit from %d to %d\n",
1077 netisr_defaultqlimit, netisr_maxqlimit);
1078 netisr_defaultqlimit = netisr_maxqlimit;
1079 }
1080 #ifdef DEVICE_POLLING
1081 /*
1082 * The device polling code is not yet aware of how to deal with
1083 * multiple netisr threads, so for the time being compiling in device
1084 * polling disables parallel netisr workers.
1085 */
1086 if (netisr_maxthreads != 1 || netisr_bindthreads != 0) {
1087 printf("netisr2: forcing maxthreads to 1 and bindthreads to "
1088 "0 for device polling\n");
1089 netisr_maxthreads = 1;
1090 netisr_bindthreads = 0;
1091 }
1092 #endif
1093
1094 netisr_start_swi(curcpu, pcpu_find(curcpu));
1095 }
1096 SYSINIT(netisr_init, SI_SUB_SOFTINTR, SI_ORDER_FIRST, netisr_init, NULL);
1097
1098 /*
1099 * Start worker threads for additional CPUs. No attempt to gracefully handle
1100 * work reassignment, we don't yet support dynamic reconfiguration.
1101 */
1102 static void
1103 netisr_start(void *arg)
1104 {
1105 struct pcpu *pc;
1106
1107 SLIST_FOREACH(pc, &cpuhead, pc_allcpu) {
1108 if (nws_count >= netisr_maxthreads)
1109 break;
1110 /* XXXRW: Is skipping absent CPUs still required here? */
1111 if (CPU_ABSENT(pc->pc_cpuid))
1112 continue;
1113 /* Worker will already be present for boot CPU. */
1114 if (pc->pc_netisr != NULL)
1115 continue;
1116 netisr_start_swi(pc->pc_cpuid, pc);
1117 }
1118 }
1119 SYSINIT(netisr_start, SI_SUB_SMP, SI_ORDER_MIDDLE, netisr_start, NULL);
1120
1121 #ifdef DDB
1122 DB_SHOW_COMMAND(netisr, db_show_netisr)
1123 {
1124 struct netisr_workstream *nwsp;
1125 struct netisr_work *nwp;
1126 int first, proto;
1127 u_int cpuid;
1128
1129 db_printf("%3s %6s %5s %5s %5s %8s %8s %8s %8s\n", "CPU", "Proto",
1130 "Len", "WMark", "Max", "Disp", "HDisp", "Drop", "Queue");
1131 for (cpuid = 0; cpuid <= mp_maxid; cpuid++) {
1132 if (CPU_ABSENT(cpuid))
1133 continue;
1134 nwsp = DPCPU_ID_PTR(cpuid, nws);
1135 if (nwsp->nws_intr_event == NULL)
1136 continue;
1137 first = 1;
1138 for (proto = 0; proto < NETISR_MAXPROT; proto++) {
1139 if (np[proto].np_handler == NULL)
1140 continue;
1141 nwp = &nwsp->nws_work[proto];
1142 if (first) {
1143 db_printf("%3d ", cpuid);
1144 first = 0;
1145 } else
1146 db_printf("%3s ", "");
1147 db_printf(
1148 "%6s %5d %5d %5d %8ju %8ju %8ju %8ju\n",
1149 np[proto].np_name, nwp->nw_len,
1150 nwp->nw_watermark, nwp->nw_qlimit,
1151 nwp->nw_dispatched, nwp->nw_hybrid_dispatched,
1152 nwp->nw_qdrops, nwp->nw_queued);
1153 }
1154 }
1155 }
1156 #endif
Cache object: 105dd751233b5dcf173c86725832e8c5
|