xref: /freebsd/sys/kern/kern_mbuf.c (revision 0b3105a37d7adcadcb720112fed4dc4e8040be99)
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 <vm/vm.h>
48 #include <vm/vm_extern.h>
49 #include <vm/vm_kern.h>
50 #include <vm/vm_page.h>
51 #include <vm/vm_map.h>
52 #include <vm/uma.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 
106 static quad_t maxmbufmem;	/* overall real memory limit for all mbufs */
107 
108 SYSCTL_QUAD(_kern_ipc, OID_AUTO, maxmbufmem, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &maxmbufmem, 0,
109     "Maximum real memory allocatable to various mbuf types");
110 
111 /*
112  * tunable_mbinit() has to be run before any mbuf allocations are done.
113  */
114 static void
115 tunable_mbinit(void *dummy)
116 {
117 	quad_t realmem;
118 
119 	/*
120 	 * The default limit for all mbuf related memory is 1/2 of all
121 	 * available kernel memory (physical or kmem).
122 	 * At most it can be 3/4 of available kernel memory.
123 	 */
124 	realmem = qmin((quad_t)physmem * PAGE_SIZE, vm_kmem_size);
125 	maxmbufmem = realmem / 2;
126 	TUNABLE_QUAD_FETCH("kern.ipc.maxmbufmem", &maxmbufmem);
127 	if (maxmbufmem > realmem / 4 * 3)
128 		maxmbufmem = realmem / 4 * 3;
129 
130 	TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
131 	if (nmbclusters == 0)
132 		nmbclusters = maxmbufmem / MCLBYTES / 4;
133 
134 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbop", &nmbjumbop);
135 	if (nmbjumbop == 0)
136 		nmbjumbop = maxmbufmem / MJUMPAGESIZE / 4;
137 
138 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbo9", &nmbjumbo9);
139 	if (nmbjumbo9 == 0)
140 		nmbjumbo9 = maxmbufmem / MJUM9BYTES / 6;
141 
142 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbo16", &nmbjumbo16);
143 	if (nmbjumbo16 == 0)
144 		nmbjumbo16 = maxmbufmem / MJUM16BYTES / 6;
145 
146 	/*
147 	 * We need at least as many mbufs as we have clusters of
148 	 * the various types added together.
149 	 */
150 	TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
151 	if (nmbufs < nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16)
152 		nmbufs = lmax(maxmbufmem / MSIZE / 5,
153 		    nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16);
154 }
155 SYSINIT(tunable_mbinit, SI_SUB_KMEM, SI_ORDER_MIDDLE, tunable_mbinit, NULL);
156 
157 static int
158 sysctl_nmbclusters(SYSCTL_HANDLER_ARGS)
159 {
160 	int error, newnmbclusters;
161 
162 	newnmbclusters = nmbclusters;
163 	error = sysctl_handle_int(oidp, &newnmbclusters, 0, req);
164 	if (error == 0 && req->newptr && newnmbclusters != nmbclusters) {
165 		if (newnmbclusters > nmbclusters &&
166 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
167 			nmbclusters = newnmbclusters;
168 			nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
169 			EVENTHANDLER_INVOKE(nmbclusters_change);
170 		} else
171 			error = EINVAL;
172 	}
173 	return (error);
174 }
175 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbclusters, CTLTYPE_INT|CTLFLAG_RW,
176 &nmbclusters, 0, sysctl_nmbclusters, "IU",
177     "Maximum number of mbuf clusters allowed");
178 
179 static int
180 sysctl_nmbjumbop(SYSCTL_HANDLER_ARGS)
181 {
182 	int error, newnmbjumbop;
183 
184 	newnmbjumbop = nmbjumbop;
185 	error = sysctl_handle_int(oidp, &newnmbjumbop, 0, req);
186 	if (error == 0 && req->newptr && newnmbjumbop != nmbjumbop) {
187 		if (newnmbjumbop > nmbjumbop &&
188 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
189 			nmbjumbop = newnmbjumbop;
190 			nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
191 		} else
192 			error = EINVAL;
193 	}
194 	return (error);
195 }
196 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbop, CTLTYPE_INT|CTLFLAG_RW,
197 &nmbjumbop, 0, sysctl_nmbjumbop, "IU",
198     "Maximum number of mbuf page size jumbo clusters allowed");
199 
200 static int
201 sysctl_nmbjumbo9(SYSCTL_HANDLER_ARGS)
202 {
203 	int error, newnmbjumbo9;
204 
205 	newnmbjumbo9 = nmbjumbo9;
206 	error = sysctl_handle_int(oidp, &newnmbjumbo9, 0, req);
207 	if (error == 0 && req->newptr && newnmbjumbo9 != nmbjumbo9) {
208 		if (newnmbjumbo9 > nmbjumbo9 &&
209 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
210 			nmbjumbo9 = newnmbjumbo9;
211 			nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
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 && newnmbjumbo16 != nmbjumbo16) {
229 		if (newnmbjumbo16 > nmbjumbo16 &&
230 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
231 			nmbjumbo16 = newnmbjumbo16;
232 			nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
233 		} else
234 			error = EINVAL;
235 	}
236 	return (error);
237 }
238 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo16, CTLTYPE_INT|CTLFLAG_RW,
239 &nmbjumbo16, 0, sysctl_nmbjumbo16, "IU",
240     "Maximum number of mbuf 16k jumbo clusters allowed");
241 
242 static int
243 sysctl_nmbufs(SYSCTL_HANDLER_ARGS)
244 {
245 	int error, newnmbufs;
246 
247 	newnmbufs = nmbufs;
248 	error = sysctl_handle_int(oidp, &newnmbufs, 0, req);
249 	if (error == 0 && req->newptr && newnmbufs != nmbufs) {
250 		if (newnmbufs > nmbufs) {
251 			nmbufs = newnmbufs;
252 			nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
253 			EVENTHANDLER_INVOKE(nmbufs_change);
254 		} else
255 			error = EINVAL;
256 	}
257 	return (error);
258 }
259 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbufs, CTLTYPE_INT|CTLFLAG_RW,
260 &nmbufs, 0, sysctl_nmbufs, "IU",
261     "Maximum number of mbufs allowed");
262 
263 /*
264  * Zones from which we allocate.
265  */
266 uma_zone_t	zone_mbuf;
267 uma_zone_t	zone_clust;
268 uma_zone_t	zone_pack;
269 uma_zone_t	zone_jumbop;
270 uma_zone_t	zone_jumbo9;
271 uma_zone_t	zone_jumbo16;
272 uma_zone_t	zone_ext_refcnt;
273 
274 /*
275  * Local prototypes.
276  */
277 static int	mb_ctor_mbuf(void *, int, void *, int);
278 static int	mb_ctor_clust(void *, int, void *, int);
279 static int	mb_ctor_pack(void *, int, void *, int);
280 static void	mb_dtor_mbuf(void *, int, void *);
281 static void	mb_dtor_clust(void *, int, void *);
282 static void	mb_dtor_pack(void *, int, void *);
283 static int	mb_zinit_pack(void *, int, int);
284 static void	mb_zfini_pack(void *, int);
285 static void	mb_reclaim(uma_zone_t, int);
286 static void    *mbuf_jumbo_alloc(uma_zone_t, vm_size_t, uint8_t *, int);
287 
288 /* Ensure that MSIZE is a power of 2. */
289 CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE);
290 
291 /*
292  * Initialize FreeBSD Network buffer allocation.
293  */
294 static void
295 mbuf_init(void *dummy)
296 {
297 
298 	/*
299 	 * Configure UMA zones for Mbufs, Clusters, and Packets.
300 	 */
301 	zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE,
302 	    mb_ctor_mbuf, mb_dtor_mbuf,
303 #ifdef INVARIANTS
304 	    trash_init, trash_fini,
305 #else
306 	    NULL, NULL,
307 #endif
308 	    MSIZE - 1, UMA_ZONE_MAXBUCKET);
309 	if (nmbufs > 0)
310 		nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
311 	uma_zone_set_warning(zone_mbuf, "kern.ipc.nmbufs limit reached");
312 	uma_zone_set_maxaction(zone_mbuf, mb_reclaim);
313 
314 	zone_clust = uma_zcreate(MBUF_CLUSTER_MEM_NAME, MCLBYTES,
315 	    mb_ctor_clust, mb_dtor_clust,
316 #ifdef INVARIANTS
317 	    trash_init, trash_fini,
318 #else
319 	    NULL, NULL,
320 #endif
321 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
322 	if (nmbclusters > 0)
323 		nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
324 	uma_zone_set_warning(zone_clust, "kern.ipc.nmbclusters limit reached");
325 	uma_zone_set_maxaction(zone_clust, mb_reclaim);
326 
327 	zone_pack = uma_zsecond_create(MBUF_PACKET_MEM_NAME, mb_ctor_pack,
328 	    mb_dtor_pack, mb_zinit_pack, mb_zfini_pack, zone_mbuf);
329 
330 	/* Make jumbo frame zone too. Page size, 9k and 16k. */
331 	zone_jumbop = uma_zcreate(MBUF_JUMBOP_MEM_NAME, MJUMPAGESIZE,
332 	    mb_ctor_clust, mb_dtor_clust,
333 #ifdef INVARIANTS
334 	    trash_init, trash_fini,
335 #else
336 	    NULL, NULL,
337 #endif
338 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
339 	if (nmbjumbop > 0)
340 		nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
341 	uma_zone_set_warning(zone_jumbop, "kern.ipc.nmbjumbop limit reached");
342 	uma_zone_set_maxaction(zone_jumbop, mb_reclaim);
343 
344 	zone_jumbo9 = uma_zcreate(MBUF_JUMBO9_MEM_NAME, MJUM9BYTES,
345 	    mb_ctor_clust, mb_dtor_clust,
346 #ifdef INVARIANTS
347 	    trash_init, trash_fini,
348 #else
349 	    NULL, NULL,
350 #endif
351 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
352 	uma_zone_set_allocf(zone_jumbo9, mbuf_jumbo_alloc);
353 	if (nmbjumbo9 > 0)
354 		nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
355 	uma_zone_set_warning(zone_jumbo9, "kern.ipc.nmbjumbo9 limit reached");
356 	uma_zone_set_maxaction(zone_jumbo9, mb_reclaim);
357 
358 	zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES,
359 	    mb_ctor_clust, mb_dtor_clust,
360 #ifdef INVARIANTS
361 	    trash_init, trash_fini,
362 #else
363 	    NULL, NULL,
364 #endif
365 	    UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
366 	uma_zone_set_allocf(zone_jumbo16, mbuf_jumbo_alloc);
367 	if (nmbjumbo16 > 0)
368 		nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
369 	uma_zone_set_warning(zone_jumbo16, "kern.ipc.nmbjumbo16 limit reached");
370 	uma_zone_set_maxaction(zone_jumbo16, mb_reclaim);
371 
372 	zone_ext_refcnt = uma_zcreate(MBUF_EXTREFCNT_MEM_NAME, sizeof(u_int),
373 	    NULL, NULL,
374 	    NULL, NULL,
375 	    UMA_ALIGN_PTR, UMA_ZONE_ZINIT);
376 
377 	/*
378 	 * Hook event handler for low-memory situation, used to
379 	 * drain protocols and push data back to the caches (UMA
380 	 * later pushes it back to VM).
381 	 */
382 	EVENTHANDLER_REGISTER(vm_lowmem, mb_reclaim, NULL,
383 	    EVENTHANDLER_PRI_FIRST);
384 }
385 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL);
386 
387 /*
388  * UMA backend page allocator for the jumbo frame zones.
389  *
390  * Allocates kernel virtual memory that is backed by contiguous physical
391  * pages.
392  */
393 static void *
394 mbuf_jumbo_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait)
395 {
396 
397 	/* Inform UMA that this allocator uses kernel_map/object. */
398 	*flags = UMA_SLAB_KERNEL;
399 	return ((void *)kmem_alloc_contig(kernel_arena, bytes, wait,
400 	    (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
401 }
402 
403 /*
404  * Constructor for Mbuf master zone.
405  *
406  * The 'arg' pointer points to a mb_args structure which
407  * contains call-specific information required to support the
408  * mbuf allocation API.  See mbuf.h.
409  */
410 static int
411 mb_ctor_mbuf(void *mem, int size, void *arg, int how)
412 {
413 	struct mbuf *m;
414 	struct mb_args *args;
415 	int error;
416 	int flags;
417 	short type;
418 
419 #ifdef INVARIANTS
420 	trash_ctor(mem, size, arg, how);
421 #endif
422 	args = (struct mb_args *)arg;
423 	type = args->type;
424 
425 	/*
426 	 * The mbuf is initialized later.  The caller has the
427 	 * responsibility to set up any MAC labels too.
428 	 */
429 	if (type == MT_NOINIT)
430 		return (0);
431 
432 	m = (struct mbuf *)mem;
433 	flags = args->flags;
434 
435 	error = m_init(m, how, type, flags);
436 
437 	return (error);
438 }
439 
440 /*
441  * The Mbuf master zone destructor.
442  */
443 static void
444 mb_dtor_mbuf(void *mem, int size, void *arg)
445 {
446 	struct mbuf *m;
447 	unsigned long flags;
448 
449 	m = (struct mbuf *)mem;
450 	flags = (unsigned long)arg;
451 
452 	KASSERT((m->m_flags & M_NOFREE) == 0, ("%s: M_NOFREE set", __func__));
453 	if ((m->m_flags & M_PKTHDR) && !SLIST_EMPTY(&m->m_pkthdr.tags))
454 		m_tag_delete_chain(m, NULL);
455 #ifdef INVARIANTS
456 	trash_dtor(mem, size, arg);
457 #endif
458 }
459 
460 /*
461  * The Mbuf Packet zone destructor.
462  */
463 static void
464 mb_dtor_pack(void *mem, int size, void *arg)
465 {
466 	struct mbuf *m;
467 
468 	m = (struct mbuf *)mem;
469 	if ((m->m_flags & M_PKTHDR) != 0)
470 		m_tag_delete_chain(m, NULL);
471 
472 	/* Make sure we've got a clean cluster back. */
473 	KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__));
474 	KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__));
475 	KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__));
476 	KASSERT(m->m_ext.ext_arg1 == NULL, ("%s: ext_arg1 != NULL", __func__));
477 	KASSERT(m->m_ext.ext_arg2 == NULL, ("%s: ext_arg2 != NULL", __func__));
478 	KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__));
479 	KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__));
480 	KASSERT(*m->m_ext.ext_cnt == 1, ("%s: ext_cnt != 1", __func__));
481 #ifdef INVARIANTS
482 	trash_dtor(m->m_ext.ext_buf, MCLBYTES, arg);
483 #endif
484 	/*
485 	 * If there are processes blocked on zone_clust, waiting for pages
486 	 * to be freed up, * cause them to be woken up by draining the
487 	 * packet zone.  We are exposed to a race here * (in the check for
488 	 * the UMA_ZFLAG_FULL) where we might miss the flag set, but that
489 	 * is deliberate. We don't want to acquire the zone lock for every
490 	 * mbuf free.
491 	 */
492 	if (uma_zone_exhausted_nolock(zone_clust))
493 		zone_drain(zone_pack);
494 }
495 
496 /*
497  * The Cluster and Jumbo[PAGESIZE|9|16] zone constructor.
498  *
499  * Here the 'arg' pointer points to the Mbuf which we
500  * are configuring cluster storage for.  If 'arg' is
501  * empty we allocate just the cluster without setting
502  * the mbuf to it.  See mbuf.h.
503  */
504 static int
505 mb_ctor_clust(void *mem, int size, void *arg, int how)
506 {
507 	struct mbuf *m;
508 	u_int *refcnt;
509 	int type;
510 	uma_zone_t zone;
511 
512 #ifdef INVARIANTS
513 	trash_ctor(mem, size, arg, how);
514 #endif
515 	switch (size) {
516 	case MCLBYTES:
517 		type = EXT_CLUSTER;
518 		zone = zone_clust;
519 		break;
520 #if MJUMPAGESIZE != MCLBYTES
521 	case MJUMPAGESIZE:
522 		type = EXT_JUMBOP;
523 		zone = zone_jumbop;
524 		break;
525 #endif
526 	case MJUM9BYTES:
527 		type = EXT_JUMBO9;
528 		zone = zone_jumbo9;
529 		break;
530 	case MJUM16BYTES:
531 		type = EXT_JUMBO16;
532 		zone = zone_jumbo16;
533 		break;
534 	default:
535 		panic("unknown cluster size");
536 		break;
537 	}
538 
539 	m = (struct mbuf *)arg;
540 	refcnt = uma_find_refcnt(zone, mem);
541 	*refcnt = 1;
542 	if (m != NULL) {
543 		m->m_ext.ext_buf = (caddr_t)mem;
544 		m->m_data = m->m_ext.ext_buf;
545 		m->m_flags |= M_EXT;
546 		m->m_ext.ext_free = NULL;
547 		m->m_ext.ext_arg1 = NULL;
548 		m->m_ext.ext_arg2 = NULL;
549 		m->m_ext.ext_size = size;
550 		m->m_ext.ext_type = type;
551 		m->m_ext.ext_flags = 0;
552 		m->m_ext.ext_cnt = refcnt;
553 	}
554 
555 	return (0);
556 }
557 
558 /*
559  * The Mbuf Cluster zone destructor.
560  */
561 static void
562 mb_dtor_clust(void *mem, int size, void *arg)
563 {
564 #ifdef INVARIANTS
565 	uma_zone_t zone;
566 
567 	zone = m_getzone(size);
568 	KASSERT(*(uma_find_refcnt(zone, mem)) <= 1,
569 		("%s: refcnt incorrect %u", __func__,
570 		 *(uma_find_refcnt(zone, mem))) );
571 
572 	trash_dtor(mem, size, arg);
573 #endif
574 }
575 
576 /*
577  * The Packet secondary zone's init routine, executed on the
578  * object's transition from mbuf keg slab to zone cache.
579  */
580 static int
581 mb_zinit_pack(void *mem, int size, int how)
582 {
583 	struct mbuf *m;
584 
585 	m = (struct mbuf *)mem;		/* m is virgin. */
586 	if (uma_zalloc_arg(zone_clust, m, how) == NULL ||
587 	    m->m_ext.ext_buf == NULL)
588 		return (ENOMEM);
589 	m->m_ext.ext_type = EXT_PACKET;	/* Override. */
590 #ifdef INVARIANTS
591 	trash_init(m->m_ext.ext_buf, MCLBYTES, how);
592 #endif
593 	return (0);
594 }
595 
596 /*
597  * The Packet secondary zone's fini routine, executed on the
598  * object's transition from zone cache to keg slab.
599  */
600 static void
601 mb_zfini_pack(void *mem, int size)
602 {
603 	struct mbuf *m;
604 
605 	m = (struct mbuf *)mem;
606 #ifdef INVARIANTS
607 	trash_fini(m->m_ext.ext_buf, MCLBYTES);
608 #endif
609 	uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL);
610 #ifdef INVARIANTS
611 	trash_dtor(mem, size, NULL);
612 #endif
613 }
614 
615 /*
616  * The "packet" keg constructor.
617  */
618 static int
619 mb_ctor_pack(void *mem, int size, void *arg, int how)
620 {
621 	struct mbuf *m;
622 	struct mb_args *args;
623 	int error, flags;
624 	short type;
625 
626 	m = (struct mbuf *)mem;
627 	args = (struct mb_args *)arg;
628 	flags = args->flags;
629 	type = args->type;
630 
631 #ifdef INVARIANTS
632 	trash_ctor(m->m_ext.ext_buf, MCLBYTES, arg, how);
633 #endif
634 
635 	error = m_init(m, how, type, flags);
636 
637 	/* m_ext is already initialized. */
638 	m->m_data = m->m_ext.ext_buf;
639  	m->m_flags = (flags | M_EXT);
640 
641 	return (error);
642 }
643 
644 /*
645  * This is the protocol drain routine.  Called by UMA whenever any of the
646  * mbuf zones is closed to its limit.
647  *
648  * No locks should be held when this is called.  The drain routines have to
649  * presently acquire some locks which raises the possibility of lock order
650  * reversal.
651  */
652 static void
653 mb_reclaim(uma_zone_t zone __unused, int pending __unused)
654 {
655 	struct domain *dp;
656 	struct protosw *pr;
657 
658 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK | WARN_PANIC, NULL, __func__);
659 
660 	for (dp = domains; dp != NULL; dp = dp->dom_next)
661 		for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++)
662 			if (pr->pr_drain != NULL)
663 				(*pr->pr_drain)();
664 }
665 
666 /*
667  * Clean up after mbufs with M_EXT storage attached to them if the
668  * reference count hits 1.
669  */
670 void
671 mb_free_ext(struct mbuf *m)
672 {
673 	int freembuf;
674 
675 	KASSERT(m->m_flags & M_EXT, ("%s: M_EXT not set on %p", __func__, m));
676 
677 	/*
678 	 * Check if the header is embedded in the cluster.
679 	 */
680 	freembuf = (m->m_flags & M_NOFREE) ? 0 : 1;
681 
682 	switch (m->m_ext.ext_type) {
683 	case EXT_SFBUF:
684 		sf_ext_free(m->m_ext.ext_arg1, m->m_ext.ext_arg2);
685 		break;
686 	case EXT_SFBUF_NOCACHE:
687 		sf_ext_free_nocache(m->m_ext.ext_arg1, m->m_ext.ext_arg2);
688 		break;
689 	default:
690 		KASSERT(m->m_ext.ext_cnt != NULL,
691 		    ("%s: no refcounting pointer on %p", __func__, m));
692 		/*
693 		 * Free attached storage if this mbuf is the only
694 		 * reference to it.
695 		 */
696 		if (*(m->m_ext.ext_cnt) != 1) {
697 			if (atomic_fetchadd_int(m->m_ext.ext_cnt, -1) != 1)
698 				break;
699 		}
700 
701 		switch (m->m_ext.ext_type) {
702 		case EXT_PACKET:	/* The packet zone is special. */
703 			if (*(m->m_ext.ext_cnt) == 0)
704 				*(m->m_ext.ext_cnt) = 1;
705 			uma_zfree(zone_pack, m);
706 			return;		/* Job done. */
707 		case EXT_CLUSTER:
708 			uma_zfree(zone_clust, m->m_ext.ext_buf);
709 			break;
710 		case EXT_JUMBOP:
711 			uma_zfree(zone_jumbop, m->m_ext.ext_buf);
712 			break;
713 		case EXT_JUMBO9:
714 			uma_zfree(zone_jumbo9, m->m_ext.ext_buf);
715 			break;
716 		case EXT_JUMBO16:
717 			uma_zfree(zone_jumbo16, m->m_ext.ext_buf);
718 			break;
719 		case EXT_NET_DRV:
720 		case EXT_MOD_TYPE:
721 		case EXT_DISPOSABLE:
722 			*(m->m_ext.ext_cnt) = 0;
723 			uma_zfree(zone_ext_refcnt, __DEVOLATILE(u_int *,
724 				m->m_ext.ext_cnt));
725 			/* FALLTHROUGH */
726 		case EXT_EXTREF:
727 			KASSERT(m->m_ext.ext_free != NULL,
728 				("%s: ext_free not set", __func__));
729 			(*(m->m_ext.ext_free))(m, m->m_ext.ext_arg1,
730 			    m->m_ext.ext_arg2);
731 			break;
732 		default:
733 			KASSERT(m->m_ext.ext_type == 0,
734 				("%s: unknown ext_type", __func__));
735 		}
736 	}
737 
738 	if (freembuf)
739 		uma_zfree(zone_mbuf, m);
740 }
741 
742 /*
743  * Official mbuf(9) allocation KPI for stack and drivers:
744  *
745  * m_get()	- a single mbuf without any attachments, sys/mbuf.h.
746  * m_gethdr()	- a single mbuf initialized as M_PKTHDR, sys/mbuf.h.
747  * m_getcl()	- an mbuf + 2k cluster, sys/mbuf.h.
748  * m_clget()	- attach cluster to already allocated mbuf.
749  * m_cljget()	- attach jumbo cluster to already allocated mbuf.
750  * m_get2()	- allocate minimum mbuf that would fit size argument.
751  * m_getm2()	- allocate a chain of mbufs/clusters.
752  * m_extadd()	- attach external cluster to mbuf.
753  *
754  * m_free()	- free single mbuf with its tags and ext, sys/mbuf.h.
755  * m_freem()	- free chain of mbufs.
756  */
757 
758 int
759 m_clget(struct mbuf *m, int how)
760 {
761 
762 	KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
763 	    __func__, m));
764 	m->m_ext.ext_buf = (char *)NULL;
765 	uma_zalloc_arg(zone_clust, m, how);
766 	/*
767 	 * On a cluster allocation failure, drain the packet zone and retry,
768 	 * we might be able to loosen a few clusters up on the drain.
769 	 */
770 	if ((how & M_NOWAIT) && (m->m_ext.ext_buf == NULL)) {
771 		zone_drain(zone_pack);
772 		uma_zalloc_arg(zone_clust, m, how);
773 	}
774 	return (m->m_flags & M_EXT);
775 }
776 
777 /*
778  * m_cljget() is different from m_clget() as it can allocate clusters without
779  * attaching them to an mbuf.  In that case the return value is the pointer
780  * to the cluster of the requested size.  If an mbuf was specified, it gets
781  * the cluster attached to it and the return value can be safely ignored.
782  * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
783  */
784 void *
785 m_cljget(struct mbuf *m, int how, int size)
786 {
787 	uma_zone_t zone;
788 
789 	if (m != NULL) {
790 		KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
791 		    __func__, m));
792 		m->m_ext.ext_buf = NULL;
793 	}
794 
795 	zone = m_getzone(size);
796 	return (uma_zalloc_arg(zone, m, how));
797 }
798 
799 /*
800  * m_get2() allocates minimum mbuf that would fit "size" argument.
801  */
802 struct mbuf *
803 m_get2(int size, int how, short type, int flags)
804 {
805 	struct mb_args args;
806 	struct mbuf *m, *n;
807 
808 	args.flags = flags;
809 	args.type = type;
810 
811 	if (size <= MHLEN || (size <= MLEN && (flags & M_PKTHDR) == 0))
812 		return (uma_zalloc_arg(zone_mbuf, &args, how));
813 	if (size <= MCLBYTES)
814 		return (uma_zalloc_arg(zone_pack, &args, how));
815 
816 	if (size > MJUMPAGESIZE)
817 		return (NULL);
818 
819 	m = uma_zalloc_arg(zone_mbuf, &args, how);
820 	if (m == NULL)
821 		return (NULL);
822 
823 	n = uma_zalloc_arg(zone_jumbop, m, how);
824 	if (n == NULL) {
825 		uma_zfree(zone_mbuf, m);
826 		return (NULL);
827 	}
828 
829 	return (m);
830 }
831 
832 /*
833  * m_getjcl() returns an mbuf with a cluster of the specified size attached.
834  * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
835  */
836 struct mbuf *
837 m_getjcl(int how, short type, int flags, int size)
838 {
839 	struct mb_args args;
840 	struct mbuf *m, *n;
841 	uma_zone_t zone;
842 
843 	if (size == MCLBYTES)
844 		return m_getcl(how, type, flags);
845 
846 	args.flags = flags;
847 	args.type = type;
848 
849 	m = uma_zalloc_arg(zone_mbuf, &args, how);
850 	if (m == NULL)
851 		return (NULL);
852 
853 	zone = m_getzone(size);
854 	n = uma_zalloc_arg(zone, m, how);
855 	if (n == NULL) {
856 		uma_zfree(zone_mbuf, m);
857 		return (NULL);
858 	}
859 	return (m);
860 }
861 
862 /*
863  * Allocate a given length worth of mbufs and/or clusters (whatever fits
864  * best) and return a pointer to the top of the allocated chain.  If an
865  * existing mbuf chain is provided, then we will append the new chain
866  * to the existing one but still return the top of the newly allocated
867  * chain.
868  */
869 struct mbuf *
870 m_getm2(struct mbuf *m, int len, int how, short type, int flags)
871 {
872 	struct mbuf *mb, *nm = NULL, *mtail = NULL;
873 
874 	KASSERT(len >= 0, ("%s: len is < 0", __func__));
875 
876 	/* Validate flags. */
877 	flags &= (M_PKTHDR | M_EOR);
878 
879 	/* Packet header mbuf must be first in chain. */
880 	if ((flags & M_PKTHDR) && m != NULL)
881 		flags &= ~M_PKTHDR;
882 
883 	/* Loop and append maximum sized mbufs to the chain tail. */
884 	while (len > 0) {
885 		if (len > MCLBYTES)
886 			mb = m_getjcl(how, type, (flags & M_PKTHDR),
887 			    MJUMPAGESIZE);
888 		else if (len >= MINCLSIZE)
889 			mb = m_getcl(how, type, (flags & M_PKTHDR));
890 		else if (flags & M_PKTHDR)
891 			mb = m_gethdr(how, type);
892 		else
893 			mb = m_get(how, type);
894 
895 		/* Fail the whole operation if one mbuf can't be allocated. */
896 		if (mb == NULL) {
897 			if (nm != NULL)
898 				m_freem(nm);
899 			return (NULL);
900 		}
901 
902 		/* Book keeping. */
903 		len -= M_SIZE(mb);
904 		if (mtail != NULL)
905 			mtail->m_next = mb;
906 		else
907 			nm = mb;
908 		mtail = mb;
909 		flags &= ~M_PKTHDR;	/* Only valid on the first mbuf. */
910 	}
911 	if (flags & M_EOR)
912 		mtail->m_flags |= M_EOR;  /* Only valid on the last mbuf. */
913 
914 	/* If mbuf was supplied, append new chain to the end of it. */
915 	if (m != NULL) {
916 		for (mtail = m; mtail->m_next != NULL; mtail = mtail->m_next)
917 			;
918 		mtail->m_next = nm;
919 		mtail->m_flags &= ~M_EOR;
920 	} else
921 		m = nm;
922 
923 	return (m);
924 }
925 
926 /*-
927  * Configure a provided mbuf to refer to the provided external storage
928  * buffer and setup a reference count for said buffer.  If the setting
929  * up of the reference count fails, the M_EXT bit will not be set.  If
930  * successfull, the M_EXT bit is set in the mbuf's flags.
931  *
932  * Arguments:
933  *    mb     The existing mbuf to which to attach the provided buffer.
934  *    buf    The address of the provided external storage buffer.
935  *    size   The size of the provided buffer.
936  *    freef  A pointer to a routine that is responsible for freeing the
937  *           provided external storage buffer.
938  *    args   A pointer to an argument structure (of any type) to be passed
939  *           to the provided freef routine (may be NULL).
940  *    flags  Any other flags to be passed to the provided mbuf.
941  *    type   The type that the external storage buffer should be
942  *           labeled with.
943  *
944  * Returns:
945  *    Nothing.
946  */
947 int
948 m_extadd(struct mbuf *mb, caddr_t buf, u_int size,
949     void (*freef)(struct mbuf *, void *, void *), void *arg1, void *arg2,
950     int flags, int type, int wait)
951 {
952 	KASSERT(type != EXT_CLUSTER, ("%s: EXT_CLUSTER not allowed", __func__));
953 
954 	if (type != EXT_EXTREF)
955 		mb->m_ext.ext_cnt = uma_zalloc(zone_ext_refcnt, wait);
956 
957 	if (mb->m_ext.ext_cnt == NULL)
958 		return (ENOMEM);
959 
960 	*(mb->m_ext.ext_cnt) = 1;
961 	mb->m_flags |= (M_EXT | flags);
962 	mb->m_ext.ext_buf = buf;
963 	mb->m_data = mb->m_ext.ext_buf;
964 	mb->m_ext.ext_size = size;
965 	mb->m_ext.ext_free = freef;
966 	mb->m_ext.ext_arg1 = arg1;
967 	mb->m_ext.ext_arg2 = arg2;
968 	mb->m_ext.ext_type = type;
969 	mb->m_ext.ext_flags = 0;
970 
971 	return (0);
972 }
973 
974 /*
975  * Free an entire chain of mbufs and associated external buffers, if
976  * applicable.
977  */
978 void
979 m_freem(struct mbuf *mb)
980 {
981 
982 	while (mb != NULL)
983 		mb = m_free(mb);
984 }
985