xref: /freebsd/sys/kern/kern_mbuf.c (revision 595e514d0df2bac5b813d35f83e32875dbf16a83)
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_param.h"
32 
33 #include <sys/param.h>
34 #include <sys/malloc.h>
35 #include <sys/systm.h>
36 #include <sys/mbuf.h>
37 #include <sys/domain.h>
38 #include <sys/eventhandler.h>
39 #include <sys/kernel.h>
40 #include <sys/protosw.h>
41 #include <sys/smp.h>
42 #include <sys/sysctl.h>
43 
44 #include <security/mac/mac_framework.h>
45 
46 #include <vm/vm.h>
47 #include <vm/vm_extern.h>
48 #include <vm/vm_kern.h>
49 #include <vm/vm_page.h>
50 #include <vm/vm_map.h>
51 #include <vm/uma.h>
52 #include <vm/uma_int.h>
53 #include <vm/uma_dbg.h>
54 
55 /*
56  * In FreeBSD, Mbufs and Mbuf Clusters are allocated from UMA
57  * Zones.
58  *
59  * Mbuf Clusters (2K, contiguous) are allocated from the Cluster
60  * Zone.  The Zone can be capped at kern.ipc.nmbclusters, if the
61  * administrator so desires.
62  *
63  * Mbufs are allocated from a UMA Master Zone called the Mbuf
64  * Zone.
65  *
66  * Additionally, FreeBSD provides a Packet Zone, which it
67  * configures as a Secondary Zone to the Mbuf Master Zone,
68  * thus sharing backend Slab kegs with the Mbuf Master Zone.
69  *
70  * Thus common-case allocations and locking are simplified:
71  *
72  *  m_clget()                m_getcl()
73  *    |                         |
74  *    |   .------------>[(Packet Cache)]    m_get(), m_gethdr()
75  *    |   |             [     Packet   ]            |
76  *  [(Cluster Cache)]   [    Secondary ]   [ (Mbuf Cache)     ]
77  *  [ Cluster Zone  ]   [     Zone     ]   [ Mbuf Master Zone ]
78  *        |                       \________         |
79  *  [ Cluster Keg   ]                      \       /
80  *        |	                         [ Mbuf Keg   ]
81  *  [ Cluster Slabs ]                         |
82  *        |                              [ Mbuf Slabs ]
83  *         \____________(VM)_________________/
84  *
85  *
86  * Whenever an object is allocated with uma_zalloc() out of
87  * one of the Zones its _ctor_ function is executed.  The same
88  * for any deallocation through uma_zfree() the _dtor_ function
89  * is executed.
90  *
91  * Caches are per-CPU and are filled from the Master Zone.
92  *
93  * Whenever an object is allocated from the underlying global
94  * memory pool it gets pre-initialized with the _zinit_ functions.
95  * When the Keg's are overfull objects get decomissioned with
96  * _zfini_ functions and free'd back to the global memory pool.
97  *
98  */
99 
100 int nmbufs;			/* limits number of mbufs */
101 int nmbclusters;		/* limits number of mbuf clusters */
102 int nmbjumbop;			/* limits number of page size jumbo clusters */
103 int nmbjumbo9;			/* limits number of 9k jumbo clusters */
104 int nmbjumbo16;			/* limits number of 16k jumbo clusters */
105 struct mbstat mbstat;
106 
107 /*
108  * tunable_mbinit() has to be run before any mbuf allocations are done.
109  */
110 static void
111 tunable_mbinit(void *dummy)
112 {
113 	quad_t realmem, maxmbufmem;
114 
115 	/*
116 	 * The default limit for all mbuf related memory is 1/2 of all
117 	 * available kernel memory (physical or kmem).
118 	 * At most it can be 3/4 of available kernel memory.
119 	 */
120 	realmem = qmin((quad_t)physmem * PAGE_SIZE,
121 	    vm_map_max(kmem_map) - vm_map_min(kmem_map));
122 	maxmbufmem = realmem / 2;
123 	TUNABLE_QUAD_FETCH("kern.maxmbufmem", &maxmbufmem);
124 	if (maxmbufmem > realmem / 4 * 3)
125 		maxmbufmem = realmem / 4 * 3;
126 
127 	TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
128 	if (nmbclusters == 0)
129 		nmbclusters = maxmbufmem / MCLBYTES / 4;
130 
131 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbop", &nmbjumbop);
132 	if (nmbjumbop == 0)
133 		nmbjumbop = maxmbufmem / MJUMPAGESIZE / 4;
134 
135 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbo9", &nmbjumbo9);
136 	if (nmbjumbo9 == 0)
137 		nmbjumbo9 = maxmbufmem / MJUM9BYTES / 6;
138 
139 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbo16", &nmbjumbo16);
140 	if (nmbjumbo16 == 0)
141 		nmbjumbo16 = maxmbufmem / MJUM16BYTES / 6;
142 
143 	/*
144 	 * We need at least as many mbufs as we have clusters of
145 	 * the various types added together.
146 	 */
147 	TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
148 	if (nmbufs < nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16)
149 		nmbufs = lmax(maxmbufmem / MSIZE / 5,
150 		    nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16);
151 }
152 SYSINIT(tunable_mbinit, SI_SUB_KMEM, SI_ORDER_MIDDLE, tunable_mbinit, NULL);
153 
154 static int
155 sysctl_nmbclusters(SYSCTL_HANDLER_ARGS)
156 {
157 	int error, newnmbclusters;
158 
159 	newnmbclusters = nmbclusters;
160 	error = sysctl_handle_int(oidp, &newnmbclusters, 0, req);
161 	if (error == 0 && req->newptr) {
162 		if (newnmbclusters > nmbclusters &&
163 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
164 			nmbclusters = newnmbclusters;
165 			uma_zone_set_max(zone_clust, nmbclusters);
166 			nmbclusters = uma_zone_get_max(zone_clust);
167 			EVENTHANDLER_INVOKE(nmbclusters_change);
168 		} else
169 			error = EINVAL;
170 	}
171 	return (error);
172 }
173 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbclusters, CTLTYPE_INT|CTLFLAG_RW,
174 &nmbclusters, 0, sysctl_nmbclusters, "IU",
175     "Maximum number of mbuf clusters allowed");
176 
177 static int
178 sysctl_nmbjumbop(SYSCTL_HANDLER_ARGS)
179 {
180 	int error, newnmbjumbop;
181 
182 	newnmbjumbop = nmbjumbop;
183 	error = sysctl_handle_int(oidp, &newnmbjumbop, 0, req);
184 	if (error == 0 && req->newptr) {
185 		if (newnmbjumbop > nmbjumbop &&
186 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
187 			nmbjumbop = newnmbjumbop;
188 			uma_zone_set_max(zone_jumbop, nmbjumbop);
189 			nmbjumbop = uma_zone_get_max(zone_jumbop);
190 		} else
191 			error = EINVAL;
192 	}
193 	return (error);
194 }
195 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbop, CTLTYPE_INT|CTLFLAG_RW,
196 &nmbjumbop, 0, sysctl_nmbjumbop, "IU",
197     "Maximum number of mbuf page size jumbo clusters allowed");
198 
199 static int
200 sysctl_nmbjumbo9(SYSCTL_HANDLER_ARGS)
201 {
202 	int error, newnmbjumbo9;
203 
204 	newnmbjumbo9 = nmbjumbo9;
205 	error = sysctl_handle_int(oidp, &newnmbjumbo9, 0, req);
206 	if (error == 0 && req->newptr) {
207 		if (newnmbjumbo9 > nmbjumbo9&&
208 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
209 			nmbjumbo9 = newnmbjumbo9;
210 			uma_zone_set_max(zone_jumbo9, nmbjumbo9);
211 			nmbjumbo9 = uma_zone_get_max(zone_jumbo9);
212 		} else
213 			error = EINVAL;
214 	}
215 	return (error);
216 }
217 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo9, CTLTYPE_INT|CTLFLAG_RW,
218 &nmbjumbo9, 0, sysctl_nmbjumbo9, "IU",
219     "Maximum number of mbuf 9k jumbo clusters allowed");
220 
221 static int
222 sysctl_nmbjumbo16(SYSCTL_HANDLER_ARGS)
223 {
224 	int error, newnmbjumbo16;
225 
226 	newnmbjumbo16 = nmbjumbo16;
227 	error = sysctl_handle_int(oidp, &newnmbjumbo16, 0, req);
228 	if (error == 0 && req->newptr) {
229 		if (newnmbjumbo16 > nmbjumbo16 &&
230 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
231 			nmbjumbo16 = newnmbjumbo16;
232 			uma_zone_set_max(zone_jumbo16, nmbjumbo16);
233 			nmbjumbo16 = uma_zone_get_max(zone_jumbo16);
234 		} else
235 			error = EINVAL;
236 	}
237 	return (error);
238 }
239 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo16, CTLTYPE_INT|CTLFLAG_RW,
240 &nmbjumbo16, 0, sysctl_nmbjumbo16, "IU",
241     "Maximum number of mbuf 16k jumbo clusters allowed");
242 
243 static int
244 sysctl_nmbufs(SYSCTL_HANDLER_ARGS)
245 {
246 	int error, newnmbufs;
247 
248 	newnmbufs = nmbufs;
249 	error = sysctl_handle_int(oidp, &newnmbufs, 0, req);
250 	if (error == 0 && req->newptr) {
251 		if (newnmbufs > nmbufs) {
252 			nmbufs = newnmbufs;
253 			uma_zone_set_max(zone_mbuf, nmbufs);
254 			nmbufs = uma_zone_get_max(zone_mbuf);
255 			EVENTHANDLER_INVOKE(nmbufs_change);
256 		} else
257 			error = EINVAL;
258 	}
259 	return (error);
260 }
261 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbuf, CTLTYPE_INT|CTLFLAG_RW,
262 &nmbufs, 0, sysctl_nmbufs, "IU",
263     "Maximum number of mbufs allowed");
264 
265 SYSCTL_STRUCT(_kern_ipc, OID_AUTO, mbstat, CTLFLAG_RD, &mbstat, mbstat,
266     "Mbuf general information and statistics");
267 
268 /*
269  * Zones from which we allocate.
270  */
271 uma_zone_t	zone_mbuf;
272 uma_zone_t	zone_clust;
273 uma_zone_t	zone_pack;
274 uma_zone_t	zone_jumbop;
275 uma_zone_t	zone_jumbo9;
276 uma_zone_t	zone_jumbo16;
277 uma_zone_t	zone_ext_refcnt;
278 
279 /*
280  * Local prototypes.
281  */
282 static int	mb_ctor_mbuf(void *, int, void *, int);
283 static int	mb_ctor_clust(void *, int, void *, int);
284 static int	mb_ctor_pack(void *, int, void *, int);
285 static void	mb_dtor_mbuf(void *, int, void *);
286 static void	mb_dtor_clust(void *, int, void *);
287 static void	mb_dtor_pack(void *, int, void *);
288 static int	mb_zinit_pack(void *, int, int);
289 static void	mb_zfini_pack(void *, int);
290 
291 static void	mb_reclaim(void *);
292 static void    *mbuf_jumbo_alloc(uma_zone_t, int, uint8_t *, int);
293 
294 /* Ensure that MSIZE is a power of 2. */
295 CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE);
296 
297 /*
298  * Initialize FreeBSD Network buffer allocation.
299  */
300 static void
301 mbuf_init(void *dummy)
302 {
303 
304 	/*
305 	 * Configure UMA zones for Mbufs, Clusters, and Packets.
306 	 */
307 	zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE,
308 	    mb_ctor_mbuf, mb_dtor_mbuf,
309 #ifdef INVARIANTS
310 	    trash_init, trash_fini,
311 #else
312 	    NULL, NULL,
313 #endif
314 	    MSIZE - 1, UMA_ZONE_MAXBUCKET);
315 	if (nmbufs > 0)
316 		nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
317 	uma_zone_set_warning(zone_mbuf, "kern.ipc.nmbufs limit reached");
318 
319 	zone_clust = uma_zcreate(MBUF_CLUSTER_MEM_NAME, MCLBYTES,
320 	    mb_ctor_clust, mb_dtor_clust,
321 #ifdef INVARIANTS
322 	    trash_init, trash_fini,
323 #else
324 	    NULL, NULL,
325 #endif
326 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
327 	if (nmbclusters > 0)
328 		nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
329 	uma_zone_set_warning(zone_clust, "kern.ipc.nmbclusters limit reached");
330 
331 	zone_pack = uma_zsecond_create(MBUF_PACKET_MEM_NAME, mb_ctor_pack,
332 	    mb_dtor_pack, mb_zinit_pack, mb_zfini_pack, zone_mbuf);
333 
334 	/* Make jumbo frame zone too. Page size, 9k and 16k. */
335 	zone_jumbop = uma_zcreate(MBUF_JUMBOP_MEM_NAME, MJUMPAGESIZE,
336 	    mb_ctor_clust, mb_dtor_clust,
337 #ifdef INVARIANTS
338 	    trash_init, trash_fini,
339 #else
340 	    NULL, NULL,
341 #endif
342 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
343 	if (nmbjumbop > 0)
344 		nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
345 	uma_zone_set_warning(zone_jumbop, "kern.ipc.nmbjumbop limit reached");
346 
347 	zone_jumbo9 = uma_zcreate(MBUF_JUMBO9_MEM_NAME, MJUM9BYTES,
348 	    mb_ctor_clust, mb_dtor_clust,
349 #ifdef INVARIANTS
350 	    trash_init, trash_fini,
351 #else
352 	    NULL, NULL,
353 #endif
354 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
355 	uma_zone_set_allocf(zone_jumbo9, mbuf_jumbo_alloc);
356 	if (nmbjumbo9 > 0)
357 		nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
358 	uma_zone_set_warning(zone_jumbo9, "kern.ipc.nmbjumbo9 limit reached");
359 
360 	zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES,
361 	    mb_ctor_clust, mb_dtor_clust,
362 #ifdef INVARIANTS
363 	    trash_init, trash_fini,
364 #else
365 	    NULL, NULL,
366 #endif
367 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
368 	uma_zone_set_allocf(zone_jumbo16, mbuf_jumbo_alloc);
369 	if (nmbjumbo16 > 0)
370 		nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
371 	uma_zone_set_warning(zone_jumbo16, "kern.ipc.nmbjumbo16 limit reached");
372 
373 	zone_ext_refcnt = uma_zcreate(MBUF_EXTREFCNT_MEM_NAME, sizeof(u_int),
374 	    NULL, NULL,
375 	    NULL, NULL,
376 	    UMA_ALIGN_PTR, UMA_ZONE_ZINIT);
377 
378 	/* uma_prealloc() goes here... */
379 
380 	/*
381 	 * Hook event handler for low-memory situation, used to
382 	 * drain protocols and push data back to the caches (UMA
383 	 * later pushes it back to VM).
384 	 */
385 	EVENTHANDLER_REGISTER(vm_lowmem, mb_reclaim, NULL,
386 	    EVENTHANDLER_PRI_FIRST);
387 
388 	/*
389 	 * [Re]set counters and local statistics knobs.
390 	 * XXX Some of these should go and be replaced, but UMA stat
391 	 * gathering needs to be revised.
392 	 */
393 	mbstat.m_mbufs = 0;
394 	mbstat.m_mclusts = 0;
395 	mbstat.m_drain = 0;
396 	mbstat.m_msize = MSIZE;
397 	mbstat.m_mclbytes = MCLBYTES;
398 	mbstat.m_minclsize = MINCLSIZE;
399 	mbstat.m_mlen = MLEN;
400 	mbstat.m_mhlen = MHLEN;
401 	mbstat.m_numtypes = MT_NTYPES;
402 
403 	mbstat.m_mcfail = mbstat.m_mpfail = 0;
404 	mbstat.sf_iocnt = 0;
405 	mbstat.sf_allocwait = mbstat.sf_allocfail = 0;
406 }
407 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL);
408 
409 /*
410  * UMA backend page allocator for the jumbo frame zones.
411  *
412  * Allocates kernel virtual memory that is backed by contiguous physical
413  * pages.
414  */
415 static void *
416 mbuf_jumbo_alloc(uma_zone_t zone, int bytes, uint8_t *flags, int wait)
417 {
418 
419 	/* Inform UMA that this allocator uses kernel_map/object. */
420 	*flags = UMA_SLAB_KERNEL;
421 	return ((void *)kmem_alloc_contig(kernel_map, bytes, wait,
422 	    (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
423 }
424 
425 /*
426  * Constructor for Mbuf master zone.
427  *
428  * The 'arg' pointer points to a mb_args structure which
429  * contains call-specific information required to support the
430  * mbuf allocation API.  See mbuf.h.
431  */
432 static int
433 mb_ctor_mbuf(void *mem, int size, void *arg, int how)
434 {
435 	struct mbuf *m;
436 	struct mb_args *args;
437 #ifdef MAC
438 	int error;
439 #endif
440 	int flags;
441 	short type;
442 
443 #ifdef INVARIANTS
444 	trash_ctor(mem, size, arg, how);
445 #endif
446 	m = (struct mbuf *)mem;
447 	args = (struct mb_args *)arg;
448 	flags = args->flags;
449 	type = args->type;
450 
451 	/*
452 	 * The mbuf is initialized later.  The caller has the
453 	 * responsibility to set up any MAC labels too.
454 	 */
455 	if (type == MT_NOINIT)
456 		return (0);
457 
458 	m->m_next = NULL;
459 	m->m_nextpkt = NULL;
460 	m->m_len = 0;
461 	m->m_flags = flags;
462 	m->m_type = type;
463 	if (flags & M_PKTHDR) {
464 		m->m_data = m->m_pktdat;
465 		m->m_pkthdr.rcvif = NULL;
466 		m->m_pkthdr.header = NULL;
467 		m->m_pkthdr.len = 0;
468 		m->m_pkthdr.csum_flags = 0;
469 		m->m_pkthdr.csum_data = 0;
470 		m->m_pkthdr.tso_segsz = 0;
471 		m->m_pkthdr.ether_vtag = 0;
472 		m->m_pkthdr.flowid = 0;
473 		m->m_pkthdr.fibnum = 0;
474 		SLIST_INIT(&m->m_pkthdr.tags);
475 #ifdef MAC
476 		/* If the label init fails, fail the alloc */
477 		error = mac_mbuf_init(m, how);
478 		if (error)
479 			return (error);
480 #endif
481 	} else
482 		m->m_data = m->m_dat;
483 	return (0);
484 }
485 
486 /*
487  * The Mbuf master zone destructor.
488  */
489 static void
490 mb_dtor_mbuf(void *mem, int size, void *arg)
491 {
492 	struct mbuf *m;
493 	unsigned long flags;
494 
495 	m = (struct mbuf *)mem;
496 	flags = (unsigned long)arg;
497 
498 	if ((flags & MB_NOTAGS) == 0 && (m->m_flags & M_PKTHDR) != 0)
499 		m_tag_delete_chain(m, NULL);
500 	KASSERT((m->m_flags & M_EXT) == 0, ("%s: M_EXT set", __func__));
501 	KASSERT((m->m_flags & M_NOFREE) == 0, ("%s: M_NOFREE set", __func__));
502 #ifdef INVARIANTS
503 	trash_dtor(mem, size, arg);
504 #endif
505 }
506 
507 /*
508  * The Mbuf Packet zone destructor.
509  */
510 static void
511 mb_dtor_pack(void *mem, int size, void *arg)
512 {
513 	struct mbuf *m;
514 
515 	m = (struct mbuf *)mem;
516 	if ((m->m_flags & M_PKTHDR) != 0)
517 		m_tag_delete_chain(m, NULL);
518 
519 	/* Make sure we've got a clean cluster back. */
520 	KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__));
521 	KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__));
522 	KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__));
523 	KASSERT(m->m_ext.ext_arg1 == NULL, ("%s: ext_arg1 != NULL", __func__));
524 	KASSERT(m->m_ext.ext_arg2 == NULL, ("%s: ext_arg2 != NULL", __func__));
525 	KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__));
526 	KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__));
527 	KASSERT(*m->m_ext.ref_cnt == 1, ("%s: ref_cnt != 1", __func__));
528 #ifdef INVARIANTS
529 	trash_dtor(m->m_ext.ext_buf, MCLBYTES, arg);
530 #endif
531 	/*
532 	 * If there are processes blocked on zone_clust, waiting for pages
533 	 * to be freed up, * cause them to be woken up by draining the
534 	 * packet zone.  We are exposed to a race here * (in the check for
535 	 * the UMA_ZFLAG_FULL) where we might miss the flag set, but that
536 	 * is deliberate. We don't want to acquire the zone lock for every
537 	 * mbuf free.
538 	 */
539 	if (uma_zone_exhausted_nolock(zone_clust))
540 		zone_drain(zone_pack);
541 }
542 
543 /*
544  * The Cluster and Jumbo[PAGESIZE|9|16] zone constructor.
545  *
546  * Here the 'arg' pointer points to the Mbuf which we
547  * are configuring cluster storage for.  If 'arg' is
548  * empty we allocate just the cluster without setting
549  * the mbuf to it.  See mbuf.h.
550  */
551 static int
552 mb_ctor_clust(void *mem, int size, void *arg, int how)
553 {
554 	struct mbuf *m;
555 	u_int *refcnt;
556 	int type;
557 	uma_zone_t zone;
558 
559 #ifdef INVARIANTS
560 	trash_ctor(mem, size, arg, how);
561 #endif
562 	switch (size) {
563 	case MCLBYTES:
564 		type = EXT_CLUSTER;
565 		zone = zone_clust;
566 		break;
567 #if MJUMPAGESIZE != MCLBYTES
568 	case MJUMPAGESIZE:
569 		type = EXT_JUMBOP;
570 		zone = zone_jumbop;
571 		break;
572 #endif
573 	case MJUM9BYTES:
574 		type = EXT_JUMBO9;
575 		zone = zone_jumbo9;
576 		break;
577 	case MJUM16BYTES:
578 		type = EXT_JUMBO16;
579 		zone = zone_jumbo16;
580 		break;
581 	default:
582 		panic("unknown cluster size");
583 		break;
584 	}
585 
586 	m = (struct mbuf *)arg;
587 	refcnt = uma_find_refcnt(zone, mem);
588 	*refcnt = 1;
589 	if (m != NULL) {
590 		m->m_ext.ext_buf = (caddr_t)mem;
591 		m->m_data = m->m_ext.ext_buf;
592 		m->m_flags |= M_EXT;
593 		m->m_ext.ext_free = NULL;
594 		m->m_ext.ext_arg1 = NULL;
595 		m->m_ext.ext_arg2 = NULL;
596 		m->m_ext.ext_size = size;
597 		m->m_ext.ext_type = type;
598 		m->m_ext.ref_cnt = refcnt;
599 	}
600 
601 	return (0);
602 }
603 
604 /*
605  * The Mbuf Cluster zone destructor.
606  */
607 static void
608 mb_dtor_clust(void *mem, int size, void *arg)
609 {
610 #ifdef INVARIANTS
611 	uma_zone_t zone;
612 
613 	zone = m_getzone(size);
614 	KASSERT(*(uma_find_refcnt(zone, mem)) <= 1,
615 		("%s: refcnt incorrect %u", __func__,
616 		 *(uma_find_refcnt(zone, mem))) );
617 
618 	trash_dtor(mem, size, arg);
619 #endif
620 }
621 
622 /*
623  * The Packet secondary zone's init routine, executed on the
624  * object's transition from mbuf keg slab to zone cache.
625  */
626 static int
627 mb_zinit_pack(void *mem, int size, int how)
628 {
629 	struct mbuf *m;
630 
631 	m = (struct mbuf *)mem;		/* m is virgin. */
632 	if (uma_zalloc_arg(zone_clust, m, how) == NULL ||
633 	    m->m_ext.ext_buf == NULL)
634 		return (ENOMEM);
635 	m->m_ext.ext_type = EXT_PACKET;	/* Override. */
636 #ifdef INVARIANTS
637 	trash_init(m->m_ext.ext_buf, MCLBYTES, how);
638 #endif
639 	return (0);
640 }
641 
642 /*
643  * The Packet secondary zone's fini routine, executed on the
644  * object's transition from zone cache to keg slab.
645  */
646 static void
647 mb_zfini_pack(void *mem, int size)
648 {
649 	struct mbuf *m;
650 
651 	m = (struct mbuf *)mem;
652 #ifdef INVARIANTS
653 	trash_fini(m->m_ext.ext_buf, MCLBYTES);
654 #endif
655 	uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL);
656 #ifdef INVARIANTS
657 	trash_dtor(mem, size, NULL);
658 #endif
659 }
660 
661 /*
662  * The "packet" keg constructor.
663  */
664 static int
665 mb_ctor_pack(void *mem, int size, void *arg, int how)
666 {
667 	struct mbuf *m;
668 	struct mb_args *args;
669 #ifdef MAC
670 	int error;
671 #endif
672 	int flags;
673 	short type;
674 
675 	m = (struct mbuf *)mem;
676 	args = (struct mb_args *)arg;
677 	flags = args->flags;
678 	type = args->type;
679 
680 #ifdef INVARIANTS
681 	trash_ctor(m->m_ext.ext_buf, MCLBYTES, arg, how);
682 #endif
683 	m->m_next = NULL;
684 	m->m_nextpkt = NULL;
685 	m->m_data = m->m_ext.ext_buf;
686 	m->m_len = 0;
687 	m->m_flags = (flags | M_EXT);
688 	m->m_type = type;
689 
690 	if (flags & M_PKTHDR) {
691 		m->m_pkthdr.rcvif = NULL;
692 		m->m_pkthdr.len = 0;
693 		m->m_pkthdr.header = NULL;
694 		m->m_pkthdr.csum_flags = 0;
695 		m->m_pkthdr.csum_data = 0;
696 		m->m_pkthdr.tso_segsz = 0;
697 		m->m_pkthdr.ether_vtag = 0;
698 		m->m_pkthdr.flowid = 0;
699 		m->m_pkthdr.fibnum = 0;
700 		SLIST_INIT(&m->m_pkthdr.tags);
701 #ifdef MAC
702 		/* If the label init fails, fail the alloc */
703 		error = mac_mbuf_init(m, how);
704 		if (error)
705 			return (error);
706 #endif
707 	}
708 	/* m_ext is already initialized. */
709 
710 	return (0);
711 }
712 
713 int
714 m_pkthdr_init(struct mbuf *m, int how)
715 {
716 #ifdef MAC
717 	int error;
718 #endif
719 	m->m_data = m->m_pktdat;
720 	SLIST_INIT(&m->m_pkthdr.tags);
721 	m->m_pkthdr.rcvif = NULL;
722 	m->m_pkthdr.header = NULL;
723 	m->m_pkthdr.len = 0;
724 	m->m_pkthdr.flowid = 0;
725 	m->m_pkthdr.fibnum = 0;
726 	m->m_pkthdr.csum_flags = 0;
727 	m->m_pkthdr.csum_data = 0;
728 	m->m_pkthdr.tso_segsz = 0;
729 	m->m_pkthdr.ether_vtag = 0;
730 #ifdef MAC
731 	/* If the label init fails, fail the alloc */
732 	error = mac_mbuf_init(m, how);
733 	if (error)
734 		return (error);
735 #endif
736 
737 	return (0);
738 }
739 
740 /*
741  * This is the protocol drain routine.
742  *
743  * No locks should be held when this is called.  The drain routines have to
744  * presently acquire some locks which raises the possibility of lock order
745  * reversal.
746  */
747 static void
748 mb_reclaim(void *junk)
749 {
750 	struct domain *dp;
751 	struct protosw *pr;
752 
753 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK | WARN_PANIC, NULL,
754 	    "mb_reclaim()");
755 
756 	for (dp = domains; dp != NULL; dp = dp->dom_next)
757 		for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++)
758 			if (pr->pr_drain != NULL)
759 				(*pr->pr_drain)();
760 }
761