xref: /freebsd/sys/kern/kern_mbuf.c (revision 1f4bcc459a76b7aa664f3fd557684cd0ba6da352)
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/types.h>
36 #include <sys/systm.h>
37 #include <sys/mbuf.h>
38 #include <sys/domain.h>
39 #include <sys/eventhandler.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/protosw.h>
44 #include <sys/smp.h>
45 #include <sys/sysctl.h>
46 
47 #include <security/mac/mac_framework.h>
48 
49 #include <vm/vm.h>
50 #include <vm/vm_extern.h>
51 #include <vm/vm_kern.h>
52 #include <vm/vm_page.h>
53 #include <vm/vm_map.h>
54 #include <vm/uma.h>
55 #include <vm/uma_int.h>
56 #include <vm/uma_dbg.h>
57 
58 /*
59  * In FreeBSD, Mbufs and Mbuf Clusters are allocated from UMA
60  * Zones.
61  *
62  * Mbuf Clusters (2K, contiguous) are allocated from the Cluster
63  * Zone.  The Zone can be capped at kern.ipc.nmbclusters, if the
64  * administrator so desires.
65  *
66  * Mbufs are allocated from a UMA Master Zone called the Mbuf
67  * Zone.
68  *
69  * Additionally, FreeBSD provides a Packet Zone, which it
70  * configures as a Secondary Zone to the Mbuf Master Zone,
71  * thus sharing backend Slab kegs with the Mbuf Master Zone.
72  *
73  * Thus common-case allocations and locking are simplified:
74  *
75  *  m_clget()                m_getcl()
76  *    |                         |
77  *    |   .------------>[(Packet Cache)]    m_get(), m_gethdr()
78  *    |   |             [     Packet   ]            |
79  *  [(Cluster Cache)]   [    Secondary ]   [ (Mbuf Cache)     ]
80  *  [ Cluster Zone  ]   [     Zone     ]   [ Mbuf Master Zone ]
81  *        |                       \________         |
82  *  [ Cluster Keg   ]                      \       /
83  *        |	                         [ Mbuf Keg   ]
84  *  [ Cluster Slabs ]                         |
85  *        |                              [ Mbuf Slabs ]
86  *         \____________(VM)_________________/
87  *
88  *
89  * Whenever an object is allocated with uma_zalloc() out of
90  * one of the Zones its _ctor_ function is executed.  The same
91  * for any deallocation through uma_zfree() the _dtor_ function
92  * is executed.
93  *
94  * Caches are per-CPU and are filled from the Master Zone.
95  *
96  * Whenever an object is allocated from the underlying global
97  * memory pool it gets pre-initialized with the _zinit_ functions.
98  * When the Keg's are overfull objects get decomissioned with
99  * _zfini_ functions and free'd back to the global memory pool.
100  *
101  */
102 
103 int nmbufs;			/* limits number of mbufs */
104 int nmbclusters;		/* limits number of mbuf clusters */
105 int nmbjumbop;			/* limits number of page size jumbo clusters */
106 int nmbjumbo9;			/* limits number of 9k jumbo clusters */
107 int nmbjumbo16;			/* limits number of 16k jumbo clusters */
108 
109 static quad_t maxmbufmem;	/* overall real memory limit for all mbufs */
110 
111 SYSCTL_QUAD(_kern_ipc, OID_AUTO, maxmbufmem, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &maxmbufmem, 0,
112     "Maximum real memory allocatable to various mbuf types");
113 
114 /*
115  * tunable_mbinit() has to be run before any mbuf allocations are done.
116  */
117 static void
118 tunable_mbinit(void *dummy)
119 {
120 	quad_t realmem;
121 
122 	/*
123 	 * The default limit for all mbuf related memory is 1/2 of all
124 	 * available kernel memory (physical or kmem).
125 	 * At most it can be 3/4 of available kernel memory.
126 	 */
127 	realmem = qmin((quad_t)physmem * PAGE_SIZE, vm_kmem_size);
128 	maxmbufmem = realmem / 2;
129 	TUNABLE_QUAD_FETCH("kern.ipc.maxmbufmem", &maxmbufmem);
130 	if (maxmbufmem > realmem / 4 * 3)
131 		maxmbufmem = realmem / 4 * 3;
132 
133 	TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
134 	if (nmbclusters == 0)
135 		nmbclusters = maxmbufmem / MCLBYTES / 4;
136 
137 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbop", &nmbjumbop);
138 	if (nmbjumbop == 0)
139 		nmbjumbop = maxmbufmem / MJUMPAGESIZE / 4;
140 
141 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbo9", &nmbjumbo9);
142 	if (nmbjumbo9 == 0)
143 		nmbjumbo9 = maxmbufmem / MJUM9BYTES / 6;
144 
145 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbo16", &nmbjumbo16);
146 	if (nmbjumbo16 == 0)
147 		nmbjumbo16 = maxmbufmem / MJUM16BYTES / 6;
148 
149 	/*
150 	 * We need at least as many mbufs as we have clusters of
151 	 * the various types added together.
152 	 */
153 	TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
154 	if (nmbufs < nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16)
155 		nmbufs = lmax(maxmbufmem / MSIZE / 5,
156 		    nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16);
157 }
158 SYSINIT(tunable_mbinit, SI_SUB_KMEM, SI_ORDER_MIDDLE, tunable_mbinit, NULL);
159 
160 static int
161 sysctl_nmbclusters(SYSCTL_HANDLER_ARGS)
162 {
163 	int error, newnmbclusters;
164 
165 	newnmbclusters = nmbclusters;
166 	error = sysctl_handle_int(oidp, &newnmbclusters, 0, req);
167 	if (error == 0 && req->newptr && newnmbclusters != nmbclusters) {
168 		if (newnmbclusters > nmbclusters &&
169 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
170 			nmbclusters = newnmbclusters;
171 			nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
172 			EVENTHANDLER_INVOKE(nmbclusters_change);
173 		} else
174 			error = EINVAL;
175 	}
176 	return (error);
177 }
178 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbclusters, CTLTYPE_INT|CTLFLAG_RW,
179 &nmbclusters, 0, sysctl_nmbclusters, "IU",
180     "Maximum number of mbuf clusters allowed");
181 
182 static int
183 sysctl_nmbjumbop(SYSCTL_HANDLER_ARGS)
184 {
185 	int error, newnmbjumbop;
186 
187 	newnmbjumbop = nmbjumbop;
188 	error = sysctl_handle_int(oidp, &newnmbjumbop, 0, req);
189 	if (error == 0 && req->newptr && newnmbjumbop != nmbjumbop) {
190 		if (newnmbjumbop > nmbjumbop &&
191 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
192 			nmbjumbop = newnmbjumbop;
193 			nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
194 		} else
195 			error = EINVAL;
196 	}
197 	return (error);
198 }
199 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbop, CTLTYPE_INT|CTLFLAG_RW,
200 &nmbjumbop, 0, sysctl_nmbjumbop, "IU",
201     "Maximum number of mbuf page size jumbo clusters allowed");
202 
203 static int
204 sysctl_nmbjumbo9(SYSCTL_HANDLER_ARGS)
205 {
206 	int error, newnmbjumbo9;
207 
208 	newnmbjumbo9 = nmbjumbo9;
209 	error = sysctl_handle_int(oidp, &newnmbjumbo9, 0, req);
210 	if (error == 0 && req->newptr && newnmbjumbo9 != nmbjumbo9) {
211 		if (newnmbjumbo9 > nmbjumbo9 &&
212 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
213 			nmbjumbo9 = newnmbjumbo9;
214 			nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
215 		} else
216 			error = EINVAL;
217 	}
218 	return (error);
219 }
220 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo9, CTLTYPE_INT|CTLFLAG_RW,
221 &nmbjumbo9, 0, sysctl_nmbjumbo9, "IU",
222     "Maximum number of mbuf 9k jumbo clusters allowed");
223 
224 static int
225 sysctl_nmbjumbo16(SYSCTL_HANDLER_ARGS)
226 {
227 	int error, newnmbjumbo16;
228 
229 	newnmbjumbo16 = nmbjumbo16;
230 	error = sysctl_handle_int(oidp, &newnmbjumbo16, 0, req);
231 	if (error == 0 && req->newptr && newnmbjumbo16 != nmbjumbo16) {
232 		if (newnmbjumbo16 > nmbjumbo16 &&
233 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
234 			nmbjumbo16 = newnmbjumbo16;
235 			nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
236 		} else
237 			error = EINVAL;
238 	}
239 	return (error);
240 }
241 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo16, CTLTYPE_INT|CTLFLAG_RW,
242 &nmbjumbo16, 0, sysctl_nmbjumbo16, "IU",
243     "Maximum number of mbuf 16k jumbo clusters allowed");
244 
245 static int
246 sysctl_nmbufs(SYSCTL_HANDLER_ARGS)
247 {
248 	int error, newnmbufs;
249 
250 	newnmbufs = nmbufs;
251 	error = sysctl_handle_int(oidp, &newnmbufs, 0, req);
252 	if (error == 0 && req->newptr && newnmbufs != nmbufs) {
253 		if (newnmbufs > nmbufs) {
254 			nmbufs = newnmbufs;
255 			nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
256 			EVENTHANDLER_INVOKE(nmbufs_change);
257 		} else
258 			error = EINVAL;
259 	}
260 	return (error);
261 }
262 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbufs, CTLTYPE_INT|CTLFLAG_RW,
263 &nmbufs, 0, sysctl_nmbufs, "IU",
264     "Maximum number of mbufs allowed");
265 
266 /*
267  * Zones from which we allocate.
268  */
269 uma_zone_t	zone_mbuf;
270 uma_zone_t	zone_clust;
271 uma_zone_t	zone_pack;
272 uma_zone_t	zone_jumbop;
273 uma_zone_t	zone_jumbo9;
274 uma_zone_t	zone_jumbo16;
275 uma_zone_t	zone_ext_refcnt;
276 
277 /*
278  * Callout to assist us in freeing mbufs.
279  */
280 static struct callout	mb_reclaim_callout;
281 static struct mtx	mb_reclaim_callout_mtx;
282 
283 /*
284  * Local prototypes.
285  */
286 static int	mb_ctor_mbuf(void *, int, void *, int);
287 static int	mb_ctor_clust(void *, int, void *, int);
288 static int	mb_ctor_pack(void *, int, void *, int);
289 static void	mb_dtor_mbuf(void *, int, void *);
290 static void	mb_dtor_clust(void *, int, void *);
291 static void	mb_dtor_pack(void *, int, void *);
292 static int	mb_zinit_pack(void *, int, int);
293 static void	mb_zfini_pack(void *, int);
294 
295 static void	mb_reclaim(void *);
296 static void    *mbuf_jumbo_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
297 static void	mb_maxaction(uma_zone_t);
298 
299 /* Ensure that MSIZE is a power of 2. */
300 CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE);
301 
302 /*
303  * Initialize FreeBSD Network buffer allocation.
304  */
305 static void
306 mbuf_init(void *dummy)
307 {
308 
309 	/*
310 	 * Configure UMA zones for Mbufs, Clusters, and Packets.
311 	 */
312 	zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE,
313 	    mb_ctor_mbuf, mb_dtor_mbuf,
314 #ifdef INVARIANTS
315 	    trash_init, trash_fini,
316 #else
317 	    NULL, NULL,
318 #endif
319 	    MSIZE - 1, UMA_ZONE_MAXBUCKET);
320 	if (nmbufs > 0)
321 		nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
322 	uma_zone_set_warning(zone_mbuf, "kern.ipc.nmbufs limit reached");
323 	uma_zone_set_maxaction(zone_mbuf, mb_maxaction);
324 
325 	zone_clust = uma_zcreate(MBUF_CLUSTER_MEM_NAME, MCLBYTES,
326 	    mb_ctor_clust, mb_dtor_clust,
327 #ifdef INVARIANTS
328 	    trash_init, trash_fini,
329 #else
330 	    NULL, NULL,
331 #endif
332 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
333 	if (nmbclusters > 0)
334 		nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
335 	uma_zone_set_warning(zone_clust, "kern.ipc.nmbclusters limit reached");
336 	uma_zone_set_maxaction(zone_clust, mb_maxaction);
337 
338 	zone_pack = uma_zsecond_create(MBUF_PACKET_MEM_NAME, mb_ctor_pack,
339 	    mb_dtor_pack, mb_zinit_pack, mb_zfini_pack, zone_mbuf);
340 
341 	/* Make jumbo frame zone too. Page size, 9k and 16k. */
342 	zone_jumbop = uma_zcreate(MBUF_JUMBOP_MEM_NAME, MJUMPAGESIZE,
343 	    mb_ctor_clust, mb_dtor_clust,
344 #ifdef INVARIANTS
345 	    trash_init, trash_fini,
346 #else
347 	    NULL, NULL,
348 #endif
349 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
350 	if (nmbjumbop > 0)
351 		nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
352 	uma_zone_set_warning(zone_jumbop, "kern.ipc.nmbjumbop limit reached");
353 	uma_zone_set_maxaction(zone_jumbop, mb_maxaction);
354 
355 	zone_jumbo9 = uma_zcreate(MBUF_JUMBO9_MEM_NAME, MJUM9BYTES,
356 	    mb_ctor_clust, mb_dtor_clust,
357 #ifdef INVARIANTS
358 	    trash_init, trash_fini,
359 #else
360 	    NULL, NULL,
361 #endif
362 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
363 	uma_zone_set_allocf(zone_jumbo9, mbuf_jumbo_alloc);
364 	if (nmbjumbo9 > 0)
365 		nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
366 	uma_zone_set_warning(zone_jumbo9, "kern.ipc.nmbjumbo9 limit reached");
367 	uma_zone_set_maxaction(zone_jumbo9, mb_maxaction);
368 
369 	zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES,
370 	    mb_ctor_clust, mb_dtor_clust,
371 #ifdef INVARIANTS
372 	    trash_init, trash_fini,
373 #else
374 	    NULL, NULL,
375 #endif
376 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
377 	uma_zone_set_allocf(zone_jumbo16, mbuf_jumbo_alloc);
378 	if (nmbjumbo16 > 0)
379 		nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
380 	uma_zone_set_warning(zone_jumbo16, "kern.ipc.nmbjumbo16 limit reached");
381 	uma_zone_set_maxaction(zone_jumbo16, mb_maxaction);
382 
383 	zone_ext_refcnt = uma_zcreate(MBUF_EXTREFCNT_MEM_NAME, sizeof(u_int),
384 	    NULL, NULL,
385 	    NULL, NULL,
386 	    UMA_ALIGN_PTR, UMA_ZONE_ZINIT);
387 
388 	/* uma_prealloc() goes here... */
389 
390 	/* Initialize the mb_reclaim() callout. */
391 	mtx_init(&mb_reclaim_callout_mtx, "mb_reclaim_callout_mtx", NULL,
392 	    MTX_DEF);
393 	callout_init(&mb_reclaim_callout, 1);
394 
395 	/*
396 	 * Hook event handler for low-memory situation, used to
397 	 * drain protocols and push data back to the caches (UMA
398 	 * later pushes it back to VM).
399 	 */
400 	EVENTHANDLER_REGISTER(vm_lowmem, mb_reclaim, NULL,
401 	    EVENTHANDLER_PRI_FIRST);
402 }
403 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL);
404 
405 /*
406  * UMA backend page allocator for the jumbo frame zones.
407  *
408  * Allocates kernel virtual memory that is backed by contiguous physical
409  * pages.
410  */
411 static void *
412 mbuf_jumbo_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait)
413 {
414 
415 	/* Inform UMA that this allocator uses kernel_map/object. */
416 	*flags = UMA_SLAB_KERNEL;
417 	return ((void *)kmem_alloc_contig(kernel_arena, bytes, wait,
418 	    (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
419 }
420 
421 /*
422  * Constructor for Mbuf master zone.
423  *
424  * The 'arg' pointer points to a mb_args structure which
425  * contains call-specific information required to support the
426  * mbuf allocation API.  See mbuf.h.
427  */
428 static int
429 mb_ctor_mbuf(void *mem, int size, void *arg, int how)
430 {
431 	struct mbuf *m;
432 	struct mb_args *args;
433 	int error;
434 	int flags;
435 	short type;
436 
437 #ifdef INVARIANTS
438 	trash_ctor(mem, size, arg, how);
439 #endif
440 	args = (struct mb_args *)arg;
441 	type = args->type;
442 
443 	/*
444 	 * The mbuf is initialized later.  The caller has the
445 	 * responsibility to set up any MAC labels too.
446 	 */
447 	if (type == MT_NOINIT)
448 		return (0);
449 
450 	m = (struct mbuf *)mem;
451 	flags = args->flags;
452 
453 	error = m_init(m, NULL, size, how, type, flags);
454 
455 	return (error);
456 }
457 
458 /*
459  * The Mbuf master zone destructor.
460  */
461 static void
462 mb_dtor_mbuf(void *mem, int size, void *arg)
463 {
464 	struct mbuf *m;
465 	unsigned long flags;
466 
467 	m = (struct mbuf *)mem;
468 	flags = (unsigned long)arg;
469 
470 	KASSERT((m->m_flags & M_NOFREE) == 0, ("%s: M_NOFREE set", __func__));
471 	if ((m->m_flags & M_PKTHDR) && !SLIST_EMPTY(&m->m_pkthdr.tags))
472 		m_tag_delete_chain(m, NULL);
473 #ifdef INVARIANTS
474 	trash_dtor(mem, size, arg);
475 #endif
476 }
477 
478 /*
479  * The Mbuf Packet zone destructor.
480  */
481 static void
482 mb_dtor_pack(void *mem, int size, void *arg)
483 {
484 	struct mbuf *m;
485 
486 	m = (struct mbuf *)mem;
487 	if ((m->m_flags & M_PKTHDR) != 0)
488 		m_tag_delete_chain(m, NULL);
489 
490 	/* Make sure we've got a clean cluster back. */
491 	KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__));
492 	KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__));
493 	KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__));
494 	KASSERT(m->m_ext.ext_arg1 == NULL, ("%s: ext_arg1 != NULL", __func__));
495 	KASSERT(m->m_ext.ext_arg2 == NULL, ("%s: ext_arg2 != NULL", __func__));
496 	KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__));
497 	KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__));
498 	KASSERT(*m->m_ext.ext_cnt == 1, ("%s: ext_cnt != 1", __func__));
499 #ifdef INVARIANTS
500 	trash_dtor(m->m_ext.ext_buf, MCLBYTES, arg);
501 #endif
502 	/*
503 	 * If there are processes blocked on zone_clust, waiting for pages
504 	 * to be freed up, * cause them to be woken up by draining the
505 	 * packet zone.  We are exposed to a race here * (in the check for
506 	 * the UMA_ZFLAG_FULL) where we might miss the flag set, but that
507 	 * is deliberate. We don't want to acquire the zone lock for every
508 	 * mbuf free.
509 	 */
510 	if (uma_zone_exhausted_nolock(zone_clust))
511 		zone_drain(zone_pack);
512 }
513 
514 /*
515  * The Cluster and Jumbo[PAGESIZE|9|16] zone constructor.
516  *
517  * Here the 'arg' pointer points to the Mbuf which we
518  * are configuring cluster storage for.  If 'arg' is
519  * empty we allocate just the cluster without setting
520  * the mbuf to it.  See mbuf.h.
521  */
522 static int
523 mb_ctor_clust(void *mem, int size, void *arg, int how)
524 {
525 	struct mbuf *m;
526 	u_int *refcnt;
527 	int type;
528 	uma_zone_t zone;
529 
530 #ifdef INVARIANTS
531 	trash_ctor(mem, size, arg, how);
532 #endif
533 	switch (size) {
534 	case MCLBYTES:
535 		type = EXT_CLUSTER;
536 		zone = zone_clust;
537 		break;
538 #if MJUMPAGESIZE != MCLBYTES
539 	case MJUMPAGESIZE:
540 		type = EXT_JUMBOP;
541 		zone = zone_jumbop;
542 		break;
543 #endif
544 	case MJUM9BYTES:
545 		type = EXT_JUMBO9;
546 		zone = zone_jumbo9;
547 		break;
548 	case MJUM16BYTES:
549 		type = EXT_JUMBO16;
550 		zone = zone_jumbo16;
551 		break;
552 	default:
553 		panic("unknown cluster size");
554 		break;
555 	}
556 
557 	m = (struct mbuf *)arg;
558 	refcnt = uma_find_refcnt(zone, mem);
559 	*refcnt = 1;
560 	if (m != NULL) {
561 		m->m_ext.ext_buf = (caddr_t)mem;
562 		m->m_data = m->m_ext.ext_buf;
563 		m->m_flags |= M_EXT;
564 		m->m_ext.ext_free = NULL;
565 		m->m_ext.ext_arg1 = NULL;
566 		m->m_ext.ext_arg2 = NULL;
567 		m->m_ext.ext_size = size;
568 		m->m_ext.ext_type = type;
569 		m->m_ext.ext_flags = 0;
570 		m->m_ext.ext_cnt = refcnt;
571 	}
572 
573 	return (0);
574 }
575 
576 /*
577  * The Mbuf Cluster zone destructor.
578  */
579 static void
580 mb_dtor_clust(void *mem, int size, void *arg)
581 {
582 #ifdef INVARIANTS
583 	uma_zone_t zone;
584 
585 	zone = m_getzone(size);
586 	KASSERT(*(uma_find_refcnt(zone, mem)) <= 1,
587 		("%s: refcnt incorrect %u", __func__,
588 		 *(uma_find_refcnt(zone, mem))) );
589 
590 	trash_dtor(mem, size, arg);
591 #endif
592 }
593 
594 /*
595  * The Packet secondary zone's init routine, executed on the
596  * object's transition from mbuf keg slab to zone cache.
597  */
598 static int
599 mb_zinit_pack(void *mem, int size, int how)
600 {
601 	struct mbuf *m;
602 
603 	m = (struct mbuf *)mem;		/* m is virgin. */
604 	if (uma_zalloc_arg(zone_clust, m, how) == NULL ||
605 	    m->m_ext.ext_buf == NULL)
606 		return (ENOMEM);
607 	m->m_ext.ext_type = EXT_PACKET;	/* Override. */
608 #ifdef INVARIANTS
609 	trash_init(m->m_ext.ext_buf, MCLBYTES, how);
610 #endif
611 	return (0);
612 }
613 
614 /*
615  * The Packet secondary zone's fini routine, executed on the
616  * object's transition from zone cache to keg slab.
617  */
618 static void
619 mb_zfini_pack(void *mem, int size)
620 {
621 	struct mbuf *m;
622 
623 	m = (struct mbuf *)mem;
624 #ifdef INVARIANTS
625 	trash_fini(m->m_ext.ext_buf, MCLBYTES);
626 #endif
627 	uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL);
628 #ifdef INVARIANTS
629 	trash_dtor(mem, size, NULL);
630 #endif
631 }
632 
633 /*
634  * The "packet" keg constructor.
635  */
636 static int
637 mb_ctor_pack(void *mem, int size, void *arg, int how)
638 {
639 	struct mbuf *m;
640 	struct mb_args *args;
641 	int error, flags;
642 	short type;
643 
644 	m = (struct mbuf *)mem;
645 	args = (struct mb_args *)arg;
646 	flags = args->flags;
647 	type = args->type;
648 
649 #ifdef INVARIANTS
650 	trash_ctor(m->m_ext.ext_buf, MCLBYTES, arg, how);
651 #endif
652 
653 	error = m_init(m, NULL, size, how, type, flags);
654 
655 	/* m_ext is already initialized. */
656 	m->m_data = m->m_ext.ext_buf;
657  	m->m_flags = (flags | M_EXT);
658 
659 	return (error);
660 }
661 
662 int
663 m_pkthdr_init(struct mbuf *m, int how)
664 {
665 #ifdef MAC
666 	int error;
667 #endif
668 	m->m_data = m->m_pktdat;
669 	bzero(&m->m_pkthdr, sizeof(m->m_pkthdr));
670 #ifdef MAC
671 	/* If the label init fails, fail the alloc */
672 	error = mac_mbuf_init(m, how);
673 	if (error)
674 		return (error);
675 #endif
676 
677 	return (0);
678 }
679 
680 /*
681  * This is the protocol drain routine.
682  *
683  * No locks should be held when this is called.  The drain routines have to
684  * presently acquire some locks which raises the possibility of lock order
685  * reversal.
686  */
687 static void
688 mb_reclaim(void *junk)
689 {
690 	struct domain *dp;
691 	struct protosw *pr;
692 
693 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK | WARN_PANIC, NULL,
694 	    "mb_reclaim()");
695 
696 	for (dp = domains; dp != NULL; dp = dp->dom_next)
697 		for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++)
698 			if (pr->pr_drain != NULL)
699 				(*pr->pr_drain)();
700 }
701 
702 /*
703  * This is the function called by the mb_reclaim_callout, which is
704  * used when we hit the maximum for a zone.
705  *
706  * (See mb_maxaction() below.)
707  */
708 static void
709 mb_reclaim_timer(void *junk __unused)
710 {
711 
712 	mtx_lock(&mb_reclaim_callout_mtx);
713 
714 	/*
715 	 * Avoid running this function extra times by skipping this invocation
716 	 * if the callout has already been rescheduled.
717 	 */
718 	if (callout_pending(&mb_reclaim_callout) ||
719 	    !callout_active(&mb_reclaim_callout)) {
720 		mtx_unlock(&mb_reclaim_callout_mtx);
721 		return;
722 	}
723 	mtx_unlock(&mb_reclaim_callout_mtx);
724 
725 	mb_reclaim(NULL);
726 
727 	mtx_lock(&mb_reclaim_callout_mtx);
728 	callout_deactivate(&mb_reclaim_callout);
729 	mtx_unlock(&mb_reclaim_callout_mtx);
730 }
731 
732 /*
733  * This function is called when we hit the maximum for a zone.
734  *
735  * At that point, we want to call the protocol drain routine to free up some
736  * mbufs. However, we will use the callout routines to schedule this to
737  * occur in another thread. (The thread calling this function holds the
738  * zone lock.)
739  */
740 static void
741 mb_maxaction(uma_zone_t zone __unused)
742 {
743 
744 	/*
745 	 * If we can't immediately obtain the lock, either the callout
746 	 * is currently running, or another thread is scheduling the
747 	 * callout.
748 	 */
749 	if (!mtx_trylock(&mb_reclaim_callout_mtx))
750 		return;
751 
752 	/* If not already scheduled/running, schedule the callout. */
753 	if (!callout_active(&mb_reclaim_callout)) {
754 		callout_reset(&mb_reclaim_callout, 1, mb_reclaim_timer, NULL);
755 	}
756 
757 	mtx_unlock(&mb_reclaim_callout_mtx);
758 }
759