xref: /freebsd/sys/kern/kern_mbuf.c (revision d056fa046c6a91b90cd98165face0e42a33a5173)
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 /* XXX: These should be tuneables. Can't change UMA limits on the fly. */
114 static int
115 sysctl_nmbclusters(SYSCTL_HANDLER_ARGS)
116 {
117 	int error, newnmbclusters;
118 
119 	newnmbclusters = nmbclusters;
120 	error = sysctl_handle_int(oidp, &newnmbclusters, sizeof(int), req);
121 	if (error == 0 && req->newptr) {
122 		if (newnmbclusters > nmbclusters) {
123 			nmbclusters = newnmbclusters;
124 			uma_zone_set_max(zone_clust, nmbclusters);
125 			EVENTHANDLER_INVOKE(nmbclusters_change);
126 		} else
127 			error = EINVAL;
128 	}
129 	return (error);
130 }
131 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbclusters, CTLTYPE_INT|CTLFLAG_RW,
132 &nmbclusters, 0, sysctl_nmbclusters, "IU",
133     "Maximum number of mbuf clusters allowed");
134 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbjumbop, CTLFLAG_RW, &nmbjumbop, 0,
135     "Maximum number of mbuf page size jumbo clusters allowed");
136 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbjumbo9, CTLFLAG_RW, &nmbjumbo9, 0,
137     "Maximum number of mbuf 9k jumbo clusters allowed");
138 SYSCTL_INT(_kern_ipc, OID_AUTO, nmbjumbo16, CTLFLAG_RW, &nmbjumbo16, 0,
139     "Maximum number of mbuf 16k jumbo clusters allowed");
140 SYSCTL_STRUCT(_kern_ipc, OID_AUTO, mbstat, CTLFLAG_RD, &mbstat, mbstat,
141     "Mbuf general information and statistics");
142 
143 /*
144  * Zones from which we allocate.
145  */
146 uma_zone_t	zone_mbuf;
147 uma_zone_t	zone_clust;
148 uma_zone_t	zone_pack;
149 uma_zone_t	zone_jumbop;
150 uma_zone_t	zone_jumbo9;
151 uma_zone_t	zone_jumbo16;
152 uma_zone_t	zone_ext_refcnt;
153 uma_zone_t	zone_mtag_vlan;
154 
155 /*
156  * Local prototypes.
157  */
158 static int	mb_ctor_mbuf(void *, int, void *, int);
159 static int	mb_ctor_clust(void *, int, void *, int);
160 static int	mb_ctor_pack(void *, int, void *, int);
161 static void	mb_dtor_mbuf(void *, int, void *);
162 static void	mb_dtor_clust(void *, int, void *);
163 static void	mb_dtor_pack(void *, int, void *);
164 static int	mb_zinit_pack(void *, int, int);
165 static void	mb_zfini_pack(void *, int);
166 static int	mt_zinit_vlan(void *, int, int);
167 
168 static void	mb_reclaim(void *);
169 static void	mbuf_init(void *);
170 
171 /* Ensure that MSIZE doesn't break dtom() - it must be a power of 2 */
172 CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE);
173 
174 /*
175  * Initialize FreeBSD Network buffer allocation.
176  */
177 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL)
178 static void
179 mbuf_init(void *dummy)
180 {
181 
182 	/*
183 	 * Configure UMA zones for Mbufs, Clusters, and Packets.
184 	 */
185 	zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE,
186 	    mb_ctor_mbuf, mb_dtor_mbuf,
187 #ifdef INVARIANTS
188 	    trash_init, trash_fini,
189 #else
190 	    NULL, NULL,
191 #endif
192 	    MSIZE - 1, UMA_ZONE_MAXBUCKET);
193 
194 	zone_clust = uma_zcreate(MBUF_CLUSTER_MEM_NAME, MCLBYTES,
195 	    mb_ctor_clust, mb_dtor_clust,
196 #ifdef INVARIANTS
197 	    trash_init, trash_fini,
198 #else
199 	    NULL, NULL,
200 #endif
201 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
202 	if (nmbclusters > 0)
203 		uma_zone_set_max(zone_clust, nmbclusters);
204 
205 	zone_pack = uma_zsecond_create(MBUF_PACKET_MEM_NAME, mb_ctor_pack,
206 	    mb_dtor_pack, mb_zinit_pack, mb_zfini_pack, zone_mbuf);
207 
208 	/* Make jumbo frame zone too. Page size, 9k and 16k. */
209 	zone_jumbop = uma_zcreate(MBUF_JUMBOP_MEM_NAME, MJUMPAGESIZE,
210 	    mb_ctor_clust, mb_dtor_clust,
211 #ifdef INVARIANTS
212 	    trash_init, trash_fini,
213 #else
214 	    NULL, NULL,
215 #endif
216 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
217 	if (nmbjumbop > 0)
218 		uma_zone_set_max(zone_jumbop, nmbjumbop);
219 
220 	zone_jumbo9 = uma_zcreate(MBUF_JUMBO9_MEM_NAME, MJUM9BYTES,
221 	    mb_ctor_clust, mb_dtor_clust,
222 #ifdef INVARIANTS
223 	    trash_init, trash_fini,
224 #else
225 	    NULL, NULL,
226 #endif
227 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
228 	if (nmbjumbo9 > 0)
229 		uma_zone_set_max(zone_jumbo9, nmbjumbo9);
230 
231 	zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES,
232 	    mb_ctor_clust, mb_dtor_clust,
233 #ifdef INVARIANTS
234 	    trash_init, trash_fini,
235 #else
236 	    NULL, NULL,
237 #endif
238 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
239 	if (nmbjumbo16 > 0)
240 		uma_zone_set_max(zone_jumbo16, nmbjumbo16);
241 
242 	zone_ext_refcnt = uma_zcreate(MBUF_EXTREFCNT_MEM_NAME, sizeof(u_int),
243 	    NULL, NULL,
244 	    NULL, NULL,
245 	    UMA_ALIGN_PTR, UMA_ZONE_ZINIT);
246 
247 	zone_mtag_vlan = uma_zcreate("mtag_vlan",
248 	    sizeof(struct m_tag) + sizeof(u_int),
249 	    NULL, NULL,
250 	    mt_zinit_vlan, NULL,
251 	    UMA_ALIGN_INT, 0);
252 
253 	/* uma_prealloc() goes here... */
254 
255 	/*
256 	 * Hook event handler for low-memory situation, used to
257 	 * drain protocols and push data back to the caches (UMA
258 	 * later pushes it back to VM).
259 	 */
260 	EVENTHANDLER_REGISTER(vm_lowmem, mb_reclaim, NULL,
261 	    EVENTHANDLER_PRI_FIRST);
262 
263 	/*
264 	 * [Re]set counters and local statistics knobs.
265 	 * XXX Some of these should go and be replaced, but UMA stat
266 	 * gathering needs to be revised.
267 	 */
268 	mbstat.m_mbufs = 0;
269 	mbstat.m_mclusts = 0;
270 	mbstat.m_drain = 0;
271 	mbstat.m_msize = MSIZE;
272 	mbstat.m_mclbytes = MCLBYTES;
273 	mbstat.m_minclsize = MINCLSIZE;
274 	mbstat.m_mlen = MLEN;
275 	mbstat.m_mhlen = MHLEN;
276 	mbstat.m_numtypes = MT_NTYPES;
277 
278 	mbstat.m_mcfail = mbstat.m_mpfail = 0;
279 	mbstat.sf_iocnt = 0;
280 	mbstat.sf_allocwait = mbstat.sf_allocfail = 0;
281 }
282 
283 /*
284  * Constructor for Mbuf master zone.
285  *
286  * The 'arg' pointer points to a mb_args structure which
287  * contains call-specific information required to support the
288  * mbuf allocation API.  See mbuf.h.
289  */
290 static int
291 mb_ctor_mbuf(void *mem, int size, void *arg, int how)
292 {
293 	struct mbuf *m;
294 	struct mb_args *args;
295 #ifdef MAC
296 	int error;
297 #endif
298 	int flags;
299 	short type;
300 
301 #ifdef INVARIANTS
302 	trash_ctor(mem, size, arg, how);
303 #endif
304 	m = (struct mbuf *)mem;
305 	args = (struct mb_args *)arg;
306 	flags = args->flags;
307 	type = args->type;
308 
309 	/*
310 	 * The mbuf is initialized later.  The caller has the
311 	 * responsibility to set up any MAC labels too.
312 	 */
313 	if (type == MT_NOINIT)
314 		return (0);
315 
316 	m->m_next = NULL;
317 	m->m_nextpkt = NULL;
318 	m->m_len = 0;
319 	m->m_flags = flags;
320 	m->m_type = type;
321 	if (flags & M_PKTHDR) {
322 		m->m_data = m->m_pktdat;
323 		m->m_pkthdr.rcvif = NULL;
324 		m->m_pkthdr.len = 0;
325 		m->m_pkthdr.header = NULL;
326 		m->m_pkthdr.csum_flags = 0;
327 		m->m_pkthdr.csum_data = 0;
328 		SLIST_INIT(&m->m_pkthdr.tags);
329 #ifdef MAC
330 		/* If the label init fails, fail the alloc */
331 		error = mac_init_mbuf(m, how);
332 		if (error)
333 			return (error);
334 #endif
335 	} else
336 		m->m_data = m->m_dat;
337 	return (0);
338 }
339 
340 /*
341  * The Mbuf master zone destructor.
342  */
343 static void
344 mb_dtor_mbuf(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 	KASSERT((m->m_flags & M_EXT) == 0, ("%s: M_EXT set", __func__));
352 #ifdef INVARIANTS
353 	trash_dtor(mem, size, arg);
354 #endif
355 }
356 
357 /*
358  * The Mbuf Packet zone destructor.
359  */
360 static void
361 mb_dtor_pack(void *mem, int size, void *arg)
362 {
363 	struct mbuf *m;
364 
365 	m = (struct mbuf *)mem;
366 	if ((m->m_flags & M_PKTHDR) != 0)
367 		m_tag_delete_chain(m, NULL);
368 
369 	/* Make sure we've got a clean cluster back. */
370 	KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__));
371 	KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__));
372 	KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__));
373 	KASSERT(m->m_ext.ext_args == NULL, ("%s: ext_args != NULL", __func__));
374 	KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__));
375 	KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__));
376 	KASSERT(*m->m_ext.ref_cnt == 1, ("%s: ref_cnt != 1", __func__));
377 #ifdef INVARIANTS
378 	trash_dtor(m->m_ext.ext_buf, MCLBYTES, arg);
379 #endif
380 }
381 
382 /*
383  * The Cluster and Jumbo[PAGESIZE|9|16] zone constructor.
384  *
385  * Here the 'arg' pointer points to the Mbuf which we
386  * are configuring cluster storage for.  If 'arg' is
387  * empty we allocate just the cluster without setting
388  * the mbuf to it.  See mbuf.h.
389  */
390 static int
391 mb_ctor_clust(void *mem, int size, void *arg, int how)
392 {
393 	struct mbuf *m;
394 	u_int *refcnt;
395 	int type = 0;
396 
397 #ifdef INVARIANTS
398 	trash_ctor(mem, size, arg, how);
399 #endif
400 	m = (struct mbuf *)arg;
401 	if (m != NULL) {
402 		switch (size) {
403 		case MCLBYTES:
404 			type = EXT_CLUSTER;
405 			break;
406 #if MJUMPAGESIZE != MCLBYTES
407 		case MJUMPAGESIZE:
408 			type = EXT_JUMBOP;
409 			break;
410 #endif
411 		case MJUM9BYTES:
412 			type = EXT_JUMBO9;
413 			break;
414 		case MJUM16BYTES:
415 			type = EXT_JUMBO16;
416 			break;
417 		default:
418 			panic("unknown cluster size");
419 			break;
420 		}
421 		m->m_ext.ext_buf = (caddr_t)mem;
422 		m->m_data = m->m_ext.ext_buf;
423 		m->m_flags |= M_EXT;
424 		m->m_ext.ext_free = NULL;
425 		m->m_ext.ext_args = NULL;
426 		m->m_ext.ext_size = size;
427 		m->m_ext.ext_type = type;
428 		m->m_ext.ref_cnt = uma_find_refcnt(zone_clust, mem);
429 		*m->m_ext.ref_cnt = 1;
430 	} else {
431 		refcnt =  uma_find_refcnt(zone_clust, mem);
432 		*refcnt = 1;
433 	}
434 	return (0);
435 }
436 
437 /*
438  * The Mbuf Cluster zone destructor.
439  */
440 static void
441 mb_dtor_clust(void *mem, int size, void *arg)
442 {
443 
444 	KASSERT(*(uma_find_refcnt(zone_clust, mem)) <= 1,
445 		("%s: refcnt incorrect %u", __func__,
446 		 *(uma_find_refcnt(zone_clust, mem))) );
447 #ifdef INVARIANTS
448 	trash_dtor(mem, size, arg);
449 #endif
450 }
451 
452 /*
453  * The Packet secondary zone's init routine, executed on the
454  * object's transition from mbuf keg slab to zone cache.
455  */
456 static int
457 mb_zinit_pack(void *mem, int size, int how)
458 {
459 	struct mbuf *m;
460 
461 	m = (struct mbuf *)mem;		/* m is virgin. */
462 	if (uma_zalloc_arg(zone_clust, m, how) == NULL ||
463 	    m->m_ext.ext_buf == NULL)
464 		return (ENOMEM);
465 	m->m_ext.ext_type = EXT_PACKET;	/* Override. */
466 #ifdef INVARIANTS
467 	trash_init(m->m_ext.ext_buf, MCLBYTES, how);
468 #endif
469 	return (0);
470 }
471 
472 /*
473  * The Packet secondary zone's fini routine, executed on the
474  * object's transition from zone cache to keg slab.
475  */
476 static void
477 mb_zfini_pack(void *mem, int size)
478 {
479 	struct mbuf *m;
480 
481 	m = (struct mbuf *)mem;
482 #ifdef INVARIANTS
483 	trash_fini(m->m_ext.ext_buf, MCLBYTES);
484 #endif
485 	uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL);
486 #ifdef INVARIANTS
487 	trash_dtor(mem, size, NULL);
488 #endif
489 }
490 
491 /*
492  * The "packet" keg constructor.
493  */
494 static int
495 mb_ctor_pack(void *mem, int size, void *arg, int how)
496 {
497 	struct mbuf *m;
498 	struct mb_args *args;
499 #ifdef MAC
500 	int error;
501 #endif
502 	int flags;
503 	short type;
504 
505 	m = (struct mbuf *)mem;
506 	args = (struct mb_args *)arg;
507 	flags = args->flags;
508 	type = args->type;
509 
510 #ifdef INVARIANTS
511 	trash_ctor(m->m_ext.ext_buf, MCLBYTES, arg, how);
512 #endif
513 	m->m_next = NULL;
514 	m->m_nextpkt = NULL;
515 	m->m_data = m->m_ext.ext_buf;
516 	m->m_len = 0;
517 	m->m_flags = (flags | M_EXT);
518 	m->m_type = type;
519 
520 	if (flags & M_PKTHDR) {
521 		m->m_pkthdr.rcvif = NULL;
522 		m->m_pkthdr.len = 0;
523 		m->m_pkthdr.header = NULL;
524 		m->m_pkthdr.csum_flags = 0;
525 		m->m_pkthdr.csum_data = 0;
526 		SLIST_INIT(&m->m_pkthdr.tags);
527 #ifdef MAC
528 		/* If the label init fails, fail the alloc */
529 		error = mac_init_mbuf(m, how);
530 		if (error)
531 			return (error);
532 #endif
533 	}
534 	/* m_ext is already initialized. */
535 
536 	return (0);
537 }
538 
539 static void
540 mt_vlan_free(struct m_tag *mtag)
541 {
542 	uma_zfree(zone_mtag_vlan, mtag);
543 }
544 
545 static int
546 mt_zinit_vlan(void *mem, int size, int how)
547 {
548 	struct m_tag *mtag = (struct m_tag *)mem;
549 
550 	m_tag_setup(mtag, MTAG_VLAN, MTAG_VLAN_TAG, sizeof(u_int));
551 	mtag->m_tag_free = mt_vlan_free;
552 
553 	return (0);
554 }
555 
556 /*
557  * This is the protocol drain routine.
558  *
559  * No locks should be held when this is called.  The drain routines have to
560  * presently acquire some locks which raises the possibility of lock order
561  * reversal.
562  */
563 static void
564 mb_reclaim(void *junk)
565 {
566 	struct domain *dp;
567 	struct protosw *pr;
568 
569 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK | WARN_PANIC, NULL,
570 	    "mb_reclaim()");
571 
572 	for (dp = domains; dp != NULL; dp = dp->dom_next)
573 		for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++)
574 			if (pr->pr_drain != NULL)
575 				(*pr->pr_drain)();
576 }
577