xref: /freebsd/sys/kern/kern_mbuf.c (revision 87569f75a91f298c52a71823c04d41cf53c88889)
1 /*-
2  * Copyright (c) 2004, 2005,
3  * 	Bosko Milekic <bmilekic@FreeBSD.org>.  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 unmodified, this list of conditions and the following
10  *    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 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_mac.h"
32 #include "opt_param.h"
33 
34 #include <sys/param.h>
35 #include <sys/mac.h>
36 #include <sys/malloc.h>
37 #include <sys/systm.h>
38 #include <sys/mbuf.h>
39 #include <sys/domain.h>
40 #include <sys/eventhandler.h>
41 #include <sys/kernel.h>
42 #include <sys/protosw.h>
43 #include <sys/smp.h>
44 #include <sys/sysctl.h>
45 
46 #include <vm/vm.h>
47 #include <vm/vm_page.h>
48 #include <vm/uma.h>
49 #include <vm/uma_int.h>
50 #include <vm/uma_dbg.h>
51 
52 /*
53  * In FreeBSD, Mbufs and Mbuf Clusters are allocated from UMA
54  * Zones.
55  *
56  * Mbuf Clusters (2K, contiguous) are allocated from the Cluster
57  * Zone.  The Zone can be capped at kern.ipc.nmbclusters, if the
58  * administrator so desires.
59  *
60  * Mbufs are allocated from a UMA Master Zone called the Mbuf
61  * Zone.
62  *
63  * Additionally, FreeBSD provides a Packet Zone, which it
64  * configures as a Secondary Zone to the Mbuf Master Zone,
65  * thus sharing backend Slab kegs with the Mbuf Master Zone.
66  *
67  * Thus common-case allocations and locking are simplified:
68  *
69  *  m_clget()                m_getcl()
70  *    |                         |
71  *    |   .------------>[(Packet Cache)]    m_get(), m_gethdr()
72  *    |   |             [     Packet   ]            |
73  *  [(Cluster Cache)]   [    Secondary ]   [ (Mbuf Cache)     ]
74  *  [ Cluster Zone  ]   [     Zone     ]   [ Mbuf Master Zone ]
75  *        |                       \________         |
76  *  [ Cluster Keg   ]                      \       /
77  *        |    	                         [ Mbuf Keg   ]
78  *  [ Cluster Slabs ]                         |
79  *        |                              [ Mbuf Slabs ]
80  *         \____________(VM)_________________/
81  *
82  *
83  * Whenever an object is allocated with uma_zalloc() out of
84  * one of the Zones its _ctor_ function is executed.  The same
85  * for any deallocation through uma_zfree() the _dtor_ function
86  * is executed.
87  *
88  * Caches are per-CPU and are filled from the Master Zone.
89  *
90  * Whenever an object is allocated from the underlying global
91  * memory pool it gets pre-initialized with the _zinit_ functions.
92  * When the Keg's are overfull objects get decomissioned with
93  * _zfini_ functions and free'd back to the global memory pool.
94  *
95  */
96 
97 int nmbclusters;		/* limits number of mbuf clusters */
98 int nmbjumbop;			/* limits number of page size jumbo clusters */
99 int nmbjumbo9;			/* limits number of 9k jumbo clusters */
100 int nmbjumbo16;			/* limits number of 16k jumbo clusters */
101 struct mbstat mbstat;
102 
103 static void
104 tunable_mbinit(void *dummy)
105 {
106 
107 	/* This has to be done before VM init. */
108 	nmbclusters = 1024 + maxusers * 64;
109 	TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
110 }
111 SYSINIT(tunable_mbinit, SI_SUB_TUNABLES, SI_ORDER_ANY, tunable_mbinit, NULL);
112 
113 SYSCTL_DECL(_kern_ipc);
114 /* XXX: These should be tuneables. Can't change UMA limits on the fly. */
115 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbclusters, CTLFLAG_RW, &nmbclusters, 0,
116     "Maximum number of mbuf clusters allowed");
117 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbjumbop, CTLFLAG_RW, &nmbjumbop, 0,
118     "Maximum number of mbuf page size jumbo clusters allowed");
119 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbjumbo9, CTLFLAG_RW, &nmbjumbo9, 0,
120     "Maximum number of mbuf 9k jumbo clusters allowed");
121 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbjumbo16, CTLFLAG_RW, &nmbjumbo16, 0,
122     "Maximum number of mbuf 16k jumbo clusters allowed");
123 SYSCTL_STRUCT(_kern_ipc, OID_AUTO, mbstat, CTLFLAG_RD, &mbstat, mbstat,
124     "Mbuf general information and statistics");
125 
126 /*
127  * Zones from which we allocate.
128  */
129 uma_zone_t	zone_mbuf;
130 uma_zone_t	zone_clust;
131 uma_zone_t	zone_pack;
132 uma_zone_t	zone_jumbop;
133 uma_zone_t	zone_jumbo9;
134 uma_zone_t	zone_jumbo16;
135 uma_zone_t	zone_ext_refcnt;
136 uma_zone_t	zone_mtag_vlan;
137 
138 /*
139  * Local prototypes.
140  */
141 static int	mb_ctor_mbuf(void *, int, void *, int);
142 static int	mb_ctor_clust(void *, int, void *, int);
143 static int	mb_ctor_pack(void *, int, void *, int);
144 static void	mb_dtor_mbuf(void *, int, void *);
145 static void	mb_dtor_clust(void *, int, void *);
146 static void	mb_dtor_pack(void *, int, void *);
147 static int	mb_zinit_pack(void *, int, int);
148 static void	mb_zfini_pack(void *, int);
149 static int	mt_zinit_vlan(void *, int, int);
150 
151 static void	mb_reclaim(void *);
152 static void	mbuf_init(void *);
153 
154 /* Ensure that MSIZE doesn't break dtom() - it must be a power of 2 */
155 CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE);
156 
157 /*
158  * Initialize FreeBSD Network buffer allocation.
159  */
160 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL)
161 static void
162 mbuf_init(void *dummy)
163 {
164 
165 	/*
166 	 * Configure UMA zones for Mbufs, Clusters, and Packets.
167 	 */
168 	zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE,
169 	    mb_ctor_mbuf, mb_dtor_mbuf,
170 #ifdef INVARIANTS
171 	    trash_init, trash_fini,
172 #else
173 	    NULL, NULL,
174 #endif
175 	    MSIZE - 1, UMA_ZONE_MAXBUCKET);
176 
177 	zone_clust = uma_zcreate(MBUF_CLUSTER_MEM_NAME, MCLBYTES,
178 	    mb_ctor_clust, mb_dtor_clust,
179 #ifdef INVARIANTS
180 	    trash_init, trash_fini,
181 #else
182 	    NULL, NULL,
183 #endif
184 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
185 	if (nmbclusters > 0)
186 		uma_zone_set_max(zone_clust, nmbclusters);
187 
188 	zone_pack = uma_zsecond_create(MBUF_PACKET_MEM_NAME, mb_ctor_pack,
189 	    mb_dtor_pack, mb_zinit_pack, mb_zfini_pack, zone_mbuf);
190 
191 	/* Make jumbo frame zone too. Page size, 9k and 16k. */
192 	zone_jumbop = uma_zcreate(MBUF_JUMBOP_MEM_NAME, MJUMPAGESIZE,
193 	    mb_ctor_clust, mb_dtor_clust,
194 #ifdef INVARIANTS
195 	    trash_init, trash_fini,
196 #else
197 	    NULL, NULL,
198 #endif
199 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
200 	if (nmbjumbop > 0)
201 		uma_zone_set_max(zone_jumbop, nmbjumbop);
202 
203 	zone_jumbo9 = uma_zcreate(MBUF_JUMBO9_MEM_NAME, MJUM9BYTES,
204 	    mb_ctor_clust, mb_dtor_clust,
205 #ifdef INVARIANTS
206 	    trash_init, trash_fini,
207 #else
208 	    NULL, NULL,
209 #endif
210 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
211 	if (nmbjumbo9 > 0)
212 		uma_zone_set_max(zone_jumbo9, nmbjumbo9);
213 
214 	zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES,
215 	    mb_ctor_clust, mb_dtor_clust,
216 #ifdef INVARIANTS
217 	    trash_init, trash_fini,
218 #else
219 	    NULL, NULL,
220 #endif
221 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
222 	if (nmbjumbo16 > 0)
223 		uma_zone_set_max(zone_jumbo16, nmbjumbo16);
224 
225 	zone_ext_refcnt = uma_zcreate(MBUF_EXTREFCNT_MEM_NAME, sizeof(u_int),
226 	    NULL, NULL,
227 	    NULL, NULL,
228 	    UMA_ALIGN_PTR, UMA_ZONE_ZINIT);
229 
230 	zone_mtag_vlan = uma_zcreate("mtag_vlan",
231 	    sizeof(struct m_tag) + sizeof(u_int),
232 	    NULL, NULL,
233 	    mt_zinit_vlan, NULL,
234 	    UMA_ALIGN_INT, 0);
235 
236 	/* uma_prealloc() goes here... */
237 
238 	/*
239 	 * Hook event handler for low-memory situation, used to
240 	 * drain protocols and push data back to the caches (UMA
241 	 * later pushes it back to VM).
242 	 */
243 	EVENTHANDLER_REGISTER(vm_lowmem, mb_reclaim, NULL,
244 	    EVENTHANDLER_PRI_FIRST);
245 
246 	/*
247 	 * [Re]set counters and local statistics knobs.
248 	 * XXX Some of these should go and be replaced, but UMA stat
249 	 * gathering needs to be revised.
250 	 */
251 	mbstat.m_mbufs = 0;
252 	mbstat.m_mclusts = 0;
253 	mbstat.m_drain = 0;
254 	mbstat.m_msize = MSIZE;
255 	mbstat.m_mclbytes = MCLBYTES;
256 	mbstat.m_minclsize = MINCLSIZE;
257 	mbstat.m_mlen = MLEN;
258 	mbstat.m_mhlen = MHLEN;
259 	mbstat.m_numtypes = MT_NTYPES;
260 
261 	mbstat.m_mcfail = mbstat.m_mpfail = 0;
262 	mbstat.sf_iocnt = 0;
263 	mbstat.sf_allocwait = mbstat.sf_allocfail = 0;
264 }
265 
266 /*
267  * Constructor for Mbuf master zone.
268  *
269  * The 'arg' pointer points to a mb_args structure which
270  * contains call-specific information required to support the
271  * mbuf allocation API.  See mbuf.h.
272  */
273 static int
274 mb_ctor_mbuf(void *mem, int size, void *arg, int how)
275 {
276 	struct mbuf *m;
277 	struct mb_args *args;
278 #ifdef MAC
279 	int error;
280 #endif
281 	int flags;
282 	short type;
283 
284 #ifdef INVARIANTS
285 	trash_ctor(mem, size, arg, how);
286 #endif
287 	m = (struct mbuf *)mem;
288 	args = (struct mb_args *)arg;
289 	flags = args->flags;
290 	type = args->type;
291 
292 	/*
293 	 * The mbuf is initialized later.  The caller has the
294 	 * responsibility to set up any MAC labels too.
295 	 */
296 	if (type == MT_NOINIT)
297 		return (0);
298 
299 	m->m_next = NULL;
300 	m->m_nextpkt = NULL;
301 	m->m_len = 0;
302 	m->m_flags = flags;
303 	m->m_type = type;
304 	if (flags & M_PKTHDR) {
305 		m->m_data = m->m_pktdat;
306 		m->m_pkthdr.rcvif = NULL;
307 		m->m_pkthdr.len = 0;
308 		m->m_pkthdr.header = NULL;
309 		m->m_pkthdr.csum_flags = 0;
310 		m->m_pkthdr.csum_data = 0;
311 		SLIST_INIT(&m->m_pkthdr.tags);
312 #ifdef MAC
313 		/* If the label init fails, fail the alloc */
314 		error = mac_init_mbuf(m, how);
315 		if (error)
316 			return (error);
317 #endif
318 	} else
319 		m->m_data = m->m_dat;
320 	return (0);
321 }
322 
323 /*
324  * The Mbuf master zone destructor.
325  */
326 static void
327 mb_dtor_mbuf(void *mem, int size, void *arg)
328 {
329 	struct mbuf *m;
330 
331 	m = (struct mbuf *)mem;
332 	if ((m->m_flags & M_PKTHDR) != 0)
333 		m_tag_delete_chain(m, NULL);
334 	KASSERT((m->m_flags & M_EXT) == 0, ("%s: M_EXT set", __func__));
335 #ifdef INVARIANTS
336 	trash_dtor(mem, size, arg);
337 #endif
338 }
339 
340 /*
341  * The Mbuf Packet zone destructor.
342  */
343 static void
344 mb_dtor_pack(void *mem, int size, void *arg)
345 {
346 	struct mbuf *m;
347 
348 	m = (struct mbuf *)mem;
349 	if ((m->m_flags & M_PKTHDR) != 0)
350 		m_tag_delete_chain(m, NULL);
351 
352 	/* Make sure we've got a clean cluster back. */
353 	KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__));
354 	KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__));
355 	KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__));
356 	KASSERT(m->m_ext.ext_args == NULL, ("%s: ext_args != NULL", __func__));
357 	KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__));
358 	KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__));
359 	KASSERT(*m->m_ext.ref_cnt == 1, ("%s: ref_cnt != 1", __func__));
360 #ifdef INVARIANTS
361 	trash_dtor(m->m_ext.ext_buf, MCLBYTES, arg);
362 #endif
363 }
364 
365 /*
366  * The Cluster and Jumbo[PAGESIZE|9|16] zone constructor.
367  *
368  * Here the 'arg' pointer points to the Mbuf which we
369  * are configuring cluster storage for.  If 'arg' is
370  * empty we allocate just the cluster without setting
371  * the mbuf to it.  See mbuf.h.
372  */
373 static int
374 mb_ctor_clust(void *mem, int size, void *arg, int how)
375 {
376 	struct mbuf *m;
377 	u_int *refcnt;
378 	int type = 0;
379 
380 #ifdef INVARIANTS
381 	trash_ctor(mem, size, arg, how);
382 #endif
383 	m = (struct mbuf *)arg;
384 	if (m != NULL) {
385 		switch (size) {
386 		case MCLBYTES:
387 			type = EXT_CLUSTER;
388 			break;
389 #if MJUMPAGESIZE != MCLBYTES
390 		case MJUMPAGESIZE:
391 			type = EXT_JUMBOP;
392 			break;
393 #endif
394 		case MJUM9BYTES:
395 			type = EXT_JUMBO9;
396 			break;
397 		case MJUM16BYTES:
398 			type = EXT_JUMBO16;
399 			break;
400 		default:
401 			panic("unknown cluster size");
402 			break;
403 		}
404 		m->m_ext.ext_buf = (caddr_t)mem;
405 		m->m_data = m->m_ext.ext_buf;
406 		m->m_flags |= M_EXT;
407 		m->m_ext.ext_free = NULL;
408 		m->m_ext.ext_args = NULL;
409 		m->m_ext.ext_size = size;
410 		m->m_ext.ext_type = type;
411 		m->m_ext.ref_cnt = uma_find_refcnt(zone_clust, mem);
412 		*m->m_ext.ref_cnt = 1;
413 	} else {
414 		refcnt =  uma_find_refcnt(zone_clust, mem);
415 		*refcnt = 1;
416 	}
417 	return (0);
418 }
419 
420 /*
421  * The Mbuf Cluster zone destructor.
422  */
423 static void
424 mb_dtor_clust(void *mem, int size, void *arg)
425 {
426 
427 	KASSERT(*(uma_find_refcnt(zone_clust, mem)) <= 1,
428 		("%s: refcnt incorrect %u", __func__,
429 		 *(uma_find_refcnt(zone_clust, mem))) );
430 #ifdef INVARIANTS
431 	trash_dtor(mem, size, arg);
432 #endif
433 }
434 
435 /*
436  * The Packet secondary zone's init routine, executed on the
437  * object's transition from mbuf keg slab to zone cache.
438  */
439 static int
440 mb_zinit_pack(void *mem, int size, int how)
441 {
442 	struct mbuf *m;
443 
444 	m = (struct mbuf *)mem;		/* m is virgin. */
445 	if (uma_zalloc_arg(zone_clust, m, how) == NULL ||
446 	    m->m_ext.ext_buf == NULL)
447 		return (ENOMEM);
448 	m->m_ext.ext_type = EXT_PACKET;	/* Override. */
449 #ifdef INVARIANTS
450 	trash_init(m->m_ext.ext_buf, MCLBYTES, how);
451 #endif
452 	return (0);
453 }
454 
455 /*
456  * The Packet secondary zone's fini routine, executed on the
457  * object's transition from zone cache to keg slab.
458  */
459 static void
460 mb_zfini_pack(void *mem, int size)
461 {
462 	struct mbuf *m;
463 
464 	m = (struct mbuf *)mem;
465 #ifdef INVARIANTS
466 	trash_fini(m->m_ext.ext_buf, MCLBYTES);
467 #endif
468 	uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL);
469 #ifdef INVARIANTS
470 	trash_dtor(mem, size, NULL);
471 #endif
472 }
473 
474 /*
475  * The "packet" keg constructor.
476  */
477 static int
478 mb_ctor_pack(void *mem, int size, void *arg, int how)
479 {
480 	struct mbuf *m;
481 	struct mb_args *args;
482 #ifdef MAC
483 	int error;
484 #endif
485 	int flags;
486 	short type;
487 
488 	m = (struct mbuf *)mem;
489 	args = (struct mb_args *)arg;
490 	flags = args->flags;
491 	type = args->type;
492 
493 #ifdef INVARIANTS
494 	trash_ctor(m->m_ext.ext_buf, MCLBYTES, arg, how);
495 #endif
496 	m->m_next = NULL;
497 	m->m_nextpkt = NULL;
498 	m->m_data = m->m_ext.ext_buf;
499 	m->m_len = 0;
500 	m->m_flags = (flags | M_EXT);
501 	m->m_type = type;
502 
503 	if (flags & M_PKTHDR) {
504 		m->m_pkthdr.rcvif = NULL;
505 		m->m_pkthdr.len = 0;
506 		m->m_pkthdr.header = NULL;
507 		m->m_pkthdr.csum_flags = 0;
508 		m->m_pkthdr.csum_data = 0;
509 		SLIST_INIT(&m->m_pkthdr.tags);
510 #ifdef MAC
511 		/* If the label init fails, fail the alloc */
512 		error = mac_init_mbuf(m, how);
513 		if (error)
514 			return (error);
515 #endif
516 	}
517 	/* m_ext is already initialized. */
518 
519 	return (0);
520 }
521 
522 static void
523 mt_vlan_free(struct m_tag *mtag)
524 {
525 	uma_zfree(zone_mtag_vlan, mtag);
526 }
527 
528 static int
529 mt_zinit_vlan(void *mem, int size, int how)
530 {
531 	struct m_tag *mtag = (struct m_tag *)mem;
532 
533 	m_tag_setup(mtag, MTAG_VLAN, MTAG_VLAN_TAG, sizeof(u_int));
534 	mtag->m_tag_free = mt_vlan_free;
535 
536 	return (0);
537 }
538 
539 /*
540  * This is the protocol drain routine.
541  *
542  * No locks should be held when this is called.  The drain routines have to
543  * presently acquire some locks which raises the possibility of lock order
544  * reversal.
545  */
546 static void
547 mb_reclaim(void *junk)
548 {
549 	struct domain *dp;
550 	struct protosw *pr;
551 
552 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK | WARN_PANIC, NULL,
553 	    "mb_reclaim()");
554 
555 	for (dp = domains; dp != NULL; dp = dp->dom_next)
556 		for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++)
557 			if (pr->pr_drain != NULL)
558 				(*pr->pr_drain)();
559 }
560