xref: /freebsd/sys/net/netisr.c (revision 6af83ee0d2941d18880b6aaa2b4facd1d30c6106)
1 /*-
2  * Copyright (c) 2001,2002,2003 Jonathan Lemon <jlemon@FreeBSD.org>
3  * Copyright (c) 1997, Stefan Esser <se@freebsd.org>
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
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  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  *
27  * $FreeBSD$
28  */
29 
30 #include "opt_net.h"
31 
32 #include <sys/param.h>
33 #include <sys/bus.h>
34 #include <sys/rtprio.h>
35 #include <sys/systm.h>
36 #include <sys/interrupt.h>
37 #include <sys/kernel.h>
38 #include <sys/kthread.h>
39 #include <sys/lock.h>
40 #include <sys/malloc.h>
41 #include <sys/proc.h>
42 #include <sys/random.h>
43 #include <sys/resourcevar.h>
44 #include <sys/sysctl.h>
45 #include <sys/unistd.h>
46 #include <machine/atomic.h>
47 #include <machine/cpu.h>
48 #include <machine/stdarg.h>
49 
50 #include <sys/mbuf.h>
51 #include <sys/socket.h>
52 
53 #include <net/if.h>
54 #include <net/if_types.h>
55 #include <net/if_var.h>
56 #include <net/netisr.h>
57 
58 /*
59  * debug_mpsafenet controls network subsystem-wide use of the Giant lock,
60  * from system calls down to interrupt handlers.  It can be changed only via
61  * a tunable at boot, not at run-time, due to the complexity of unwinding.
62  * The compiled default is set via a kernel option; right now, the default
63  * unless otherwise specified is to run the network stack without Giant.
64  */
65 #ifdef NET_WITH_GIANT
66 int	debug_mpsafenet = 0;
67 #else
68 int	debug_mpsafenet = 1;
69 #endif
70 int	debug_mpsafenet_toolatetotwiddle = 0;
71 
72 TUNABLE_INT("debug.mpsafenet", &debug_mpsafenet);
73 SYSCTL_INT(_debug, OID_AUTO, mpsafenet, CTLFLAG_RD, &debug_mpsafenet, 0,
74     "Enable/disable MPSAFE network support");
75 
76 volatile unsigned int	netisr;	/* scheduling bits for network */
77 
78 struct netisr {
79 	netisr_t	*ni_handler;
80 	struct ifqueue	*ni_queue;
81 	int		ni_flags;
82 } netisrs[32];
83 
84 static void *net_ih;
85 
86 /*
87  * Not all network code is currently capable of running MPSAFE; however,
88  * most of it is.  Since those sections that are not are generally optional
89  * components not shipped with default kernels, we provide a basic way to
90  * determine whether MPSAFE operation is permitted: based on a default of
91  * yes, we permit non-MPSAFE components to use a registration call to
92  * identify that they require Giant.  If the system is early in the boot
93  * process still, then we change the debug_mpsafenet setting to choose a
94  * non-MPSAFE execution mode (degraded).  If it's too late for that (since
95  * the setting cannot be changed at run time), we generate a console warning
96  * that the configuration may be unsafe.
97  */
98 static int mpsafe_warn_count;
99 
100 /*
101  * Function call implementing registration of a non-MPSAFE network component.
102  */
103 void
104 net_warn_not_mpsafe(const char *component)
105 {
106 
107 	/*
108 	 * If we're running with Giant over the network stack, there is no
109 	 * problem.
110 	 */
111 	if (!debug_mpsafenet)
112 		return;
113 
114 	/*
115 	 * If it's not too late to change the MPSAFE setting for the network
116 	 * stack, do so now.  This effectively suppresses warnings by
117 	 * components registering later.
118 	 */
119 	if (!debug_mpsafenet_toolatetotwiddle) {
120 		debug_mpsafenet = 0;
121 		printf("WARNING: debug.mpsafenet forced to 0 as %s requires "
122 		    "Giant\n", component);
123 		return;
124 	}
125 
126 	/*
127 	 * We must run without Giant, so generate a console warning with some
128 	 * information with what to do about it.  The system may be operating
129 	 * unsafely, however.
130 	 */
131 	printf("WARNING: Network stack Giant-free, but %s requires Giant.\n",
132 	    component);
133 	if (mpsafe_warn_count == 0)
134 		printf("    Consider adding 'options NET_WITH_GIANT' or "
135 		    "setting debug.mpsafenet=0\n");
136 	mpsafe_warn_count++;
137 }
138 
139 /*
140  * This sysinit is run after any pre-loaded or compiled-in components have
141  * announced that they require Giant, but before any modules loaded at
142  * run-time.
143  */
144 static void
145 net_mpsafe_toolate(void *arg)
146 {
147 
148 	debug_mpsafenet_toolatetotwiddle = 1;
149 
150 	if (!debug_mpsafenet)
151 		printf("WARNING: MPSAFE network stack disabled, expect "
152 		    "reduced performance.\n");
153 }
154 
155 SYSINIT(net_mpsafe_toolate, SI_SUB_SETTINGS, SI_ORDER_ANY, net_mpsafe_toolate,
156     NULL);
157 
158 void
159 legacy_setsoftnet(void)
160 {
161 	swi_sched(net_ih, 0);
162 }
163 
164 void
165 netisr_register(int num, netisr_t *handler, struct ifqueue *inq, int flags)
166 {
167 
168 	KASSERT(!(num < 0 || num >= (sizeof(netisrs)/sizeof(*netisrs))),
169 	    ("bad isr %d", num));
170 	netisrs[num].ni_handler = handler;
171 	netisrs[num].ni_queue = inq;
172 	if ((flags & NETISR_MPSAFE) && !debug_mpsafenet)
173 		flags &= ~NETISR_MPSAFE;
174 	netisrs[num].ni_flags = flags;
175 }
176 
177 void
178 netisr_unregister(int num)
179 {
180 	struct netisr *ni;
181 
182 	KASSERT(!(num < 0 || num >= (sizeof(netisrs)/sizeof(*netisrs))),
183 	    ("bad isr %d", num));
184 	ni = &netisrs[num];
185 	ni->ni_handler = NULL;
186 	if (ni->ni_queue != NULL)
187 		IF_DRAIN(ni->ni_queue);
188 	ni->ni_queue = NULL;
189 }
190 
191 struct isrstat {
192 	int	isrs_count;			/* dispatch count */
193 	int	isrs_directed;			/* ...directly dispatched */
194 	int	isrs_deferred;			/* ...queued instead */
195 	int	isrs_queued;			/* intentionally queueued */
196 	int	isrs_drop;			/* dropped 'cuz no handler */
197 	int	isrs_swi_count;			/* swi_net handlers called */
198 };
199 static struct isrstat isrstat;
200 
201 SYSCTL_NODE(_net, OID_AUTO, isr, CTLFLAG_RW, 0, "netisr counters");
202 
203 static int	netisr_enable = 0;
204 SYSCTL_INT(_net_isr, OID_AUTO, enable, CTLFLAG_RW,
205     &netisr_enable, 0, "enable direct dispatch");
206 TUNABLE_INT("net.isr.enable", &netisr_enable);
207 
208 SYSCTL_INT(_net_isr, OID_AUTO, count, CTLFLAG_RD,
209     &isrstat.isrs_count, 0, "");
210 SYSCTL_INT(_net_isr, OID_AUTO, directed, CTLFLAG_RD,
211     &isrstat.isrs_directed, 0, "");
212 SYSCTL_INT(_net_isr, OID_AUTO, deferred, CTLFLAG_RD,
213     &isrstat.isrs_deferred, 0, "");
214 SYSCTL_INT(_net_isr, OID_AUTO, queued, CTLFLAG_RD,
215     &isrstat.isrs_queued, 0, "");
216 SYSCTL_INT(_net_isr, OID_AUTO, drop, CTLFLAG_RD,
217     &isrstat.isrs_drop, 0, "");
218 SYSCTL_INT(_net_isr, OID_AUTO, swi_count, CTLFLAG_RD,
219     &isrstat.isrs_swi_count, 0, "");
220 
221 /*
222  * Process all packets currently present in a netisr queue.  Used to
223  * drain an existing set of packets waiting for processing when we
224  * begin direct dispatch, to avoid processing packets out of order.
225  */
226 static void
227 netisr_processqueue(struct netisr *ni)
228 {
229 	struct mbuf *m;
230 
231 	for (;;) {
232 		IF_DEQUEUE(ni->ni_queue, m);
233 		if (m == NULL)
234 			break;
235 		ni->ni_handler(m);
236 	}
237 }
238 
239 /*
240  * Call the netisr directly instead of queueing the packet, if possible.
241  */
242 void
243 netisr_dispatch(int num, struct mbuf *m)
244 {
245 	struct netisr *ni;
246 
247 	isrstat.isrs_count++;		/* XXX redundant */
248 	KASSERT(!(num < 0 || num >= (sizeof(netisrs)/sizeof(*netisrs))),
249 	    ("bad isr %d", num));
250 	ni = &netisrs[num];
251 	if (ni->ni_queue == NULL) {
252 		isrstat.isrs_drop++;
253 		m_freem(m);
254 		return;
255 	}
256 	/*
257 	 * Do direct dispatch only for MPSAFE netisrs (and
258 	 * only when enabled).  Note that when a netisr is
259 	 * marked MPSAFE we permit multiple concurrent instances
260 	 * to run.  We guarantee only the order in which
261 	 * packets are processed for each "dispatch point" in
262 	 * the system (i.e. call to netisr_dispatch or
263 	 * netisr_queue).  This insures ordering of packets
264 	 * from an interface but does not guarantee ordering
265 	 * between multiple places in the system (e.g. IP
266 	 * dispatched from interfaces vs. IP queued from IPSec).
267 	 */
268 	if (netisr_enable && (ni->ni_flags & NETISR_MPSAFE)) {
269 		isrstat.isrs_directed++;
270 		/*
271 		 * NB: We used to drain the queue before handling
272 		 * the packet but now do not.  Doing so here will
273 		 * not preserve ordering so instead we fallback to
274 		 * guaranteeing order only from dispatch points
275 		 * in the system (see above).
276 		 */
277 		ni->ni_handler(m);
278 	} else {
279 		isrstat.isrs_deferred++;
280 		if (IF_HANDOFF(ni->ni_queue, m, NULL))
281 			schednetisr(num);
282 	}
283 }
284 
285 /*
286  * Same as above, but always queue.
287  * This is either used in places where we are not confident that
288  * direct dispatch is possible, or where queueing is required.
289  * It returns (0) on success and ERRNO on failure.  On failure the
290  * mbuf has been free'd.
291  */
292 int
293 netisr_queue(int num, struct mbuf *m)
294 {
295 	struct netisr *ni;
296 
297 	KASSERT(!(num < 0 || num >= (sizeof(netisrs)/sizeof(*netisrs))),
298 	    ("bad isr %d", num));
299 	ni = &netisrs[num];
300 	if (ni->ni_queue == NULL) {
301 		isrstat.isrs_drop++;
302 		m_freem(m);
303 		return (ENXIO);
304 	}
305 	isrstat.isrs_queued++;
306 	if (!IF_HANDOFF(ni->ni_queue, m, NULL))
307 		return (ENOBUFS);	/* IF_HANDOFF has free'd the mbuf */
308 	schednetisr(num);
309 	return (0);
310 }
311 
312 static void
313 swi_net(void *dummy)
314 {
315 	struct netisr *ni;
316 	u_int bits;
317 	int i;
318 #ifdef DEVICE_POLLING
319 	const int polling = 1;
320 #else
321 	const int polling = 0;
322 #endif
323 
324 	do {
325 		bits = atomic_readandclear_int(&netisr);
326 		if (bits == 0)
327 			break;
328 		while ((i = ffs(bits)) != 0) {
329 			isrstat.isrs_swi_count++;
330 			i--;
331 			bits &= ~(1 << i);
332 			ni = &netisrs[i];
333 			if (ni->ni_handler == NULL) {
334 				printf("swi_net: unregistered isr %d.\n", i);
335 				continue;
336 			}
337 			if ((ni->ni_flags & NETISR_MPSAFE) == 0) {
338 				mtx_lock(&Giant);
339 				if (ni->ni_queue == NULL)
340 					ni->ni_handler(NULL);
341 				else
342 					netisr_processqueue(ni);
343 				mtx_unlock(&Giant);
344 			} else {
345 				if (ni->ni_queue == NULL)
346 					ni->ni_handler(NULL);
347 				else
348 					netisr_processqueue(ni);
349 			}
350 		}
351 	} while (polling);
352 }
353 
354 static void
355 start_netisr(void *dummy)
356 {
357 
358 	if (swi_add(NULL, "net", swi_net, NULL, SWI_NET, INTR_MPSAFE, &net_ih))
359 		panic("start_netisr");
360 }
361 SYSINIT(start_netisr, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_netisr, NULL)
362