xref: /freebsd/sys/kern/kern_mbuf.c (revision d37eb51047221dc3322b34db1038ff3aa533883f)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2004, 2005,
5  *	Bosko Milekic <bmilekic@FreeBSD.org>.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice unmodified, this list of conditions and the following
12  *    disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 #include "opt_param.h"
32 #include "opt_kern_tls.h"
33 
34 #include <sys/param.h>
35 #include <sys/conf.h>
36 #include <sys/domainset.h>
37 #include <sys/malloc.h>
38 #include <sys/systm.h>
39 #include <sys/mbuf.h>
40 #include <sys/eventhandler.h>
41 #include <sys/kernel.h>
42 #include <sys/ktls.h>
43 #include <sys/limits.h>
44 #include <sys/lock.h>
45 #include <sys/mutex.h>
46 #include <sys/refcount.h>
47 #include <sys/sf_buf.h>
48 #include <sys/smp.h>
49 #include <sys/socket.h>
50 #include <sys/sysctl.h>
51 
52 #include <net/if.h>
53 #include <net/if_var.h>
54 
55 #include <vm/vm.h>
56 #include <vm/vm_extern.h>
57 #include <vm/vm_kern.h>
58 #include <vm/vm_page.h>
59 #include <vm/vm_pageout.h>
60 #include <vm/vm_map.h>
61 #include <vm/uma.h>
62 #include <vm/uma_dbg.h>
63 
64 _Static_assert(MJUMPAGESIZE > MCLBYTES,
65     "Cluster must be smaller than a jumbo page");
66 
67 /*
68  * In FreeBSD, Mbufs and Mbuf Clusters are allocated from UMA
69  * Zones.
70  *
71  * Mbuf Clusters (2K, contiguous) are allocated from the Cluster
72  * Zone.  The Zone can be capped at kern.ipc.nmbclusters, if the
73  * administrator so desires.
74  *
75  * Mbufs are allocated from a UMA Primary Zone called the Mbuf
76  * Zone.
77  *
78  * Additionally, FreeBSD provides a Packet Zone, which it
79  * configures as a Secondary Zone to the Mbuf Primary Zone,
80  * thus sharing backend Slab kegs with the Mbuf Primary Zone.
81  *
82  * Thus common-case allocations and locking are simplified:
83  *
84  *  m_clget()                m_getcl()
85  *    |                         |
86  *    |   .------------>[(Packet Cache)]    m_get(), m_gethdr()
87  *    |   |             [     Packet   ]            |
88  *  [(Cluster Cache)]   [    Secondary ]   [ (Mbuf Cache)     ]
89  *  [ Cluster Zone  ]   [     Zone     ]   [ Mbuf Primary Zone ]
90  *        |                       \________         |
91  *  [ Cluster Keg   ]                      \       /
92  *        |	                         [ Mbuf Keg   ]
93  *  [ Cluster Slabs ]                         |
94  *        |                              [ Mbuf Slabs ]
95  *         \____________(VM)_________________/
96  *
97  *
98  * Whenever an object is allocated with uma_zalloc() out of
99  * one of the Zones its _ctor_ function is executed.  The same
100  * for any deallocation through uma_zfree() the _dtor_ function
101  * is executed.
102  *
103  * Caches are per-CPU and are filled from the Primary Zone.
104  *
105  * Whenever an object is allocated from the underlying global
106  * memory pool it gets pre-initialized with the _zinit_ functions.
107  * When the Keg's are overfull objects get decommissioned with
108  * _zfini_ functions and free'd back to the global memory pool.
109  *
110  */
111 
112 int nmbufs;			/* limits number of mbufs */
113 int nmbclusters;		/* limits number of mbuf clusters */
114 int nmbjumbop;			/* limits number of page size jumbo clusters */
115 int nmbjumbo9;			/* limits number of 9k jumbo clusters */
116 int nmbjumbo16;			/* limits number of 16k jumbo clusters */
117 
118 bool mb_use_ext_pgs = false;	/* use M_EXTPG mbufs for sendfile & TLS */
119 
120 static int
121 sysctl_mb_use_ext_pgs(SYSCTL_HANDLER_ARGS)
122 {
123 	int error, extpg;
124 
125 	extpg = mb_use_ext_pgs;
126 	error = sysctl_handle_int(oidp, &extpg, 0, req);
127 	if (error == 0 && req->newptr != NULL) {
128 		if (extpg != 0 && !PMAP_HAS_DMAP)
129 			error = EOPNOTSUPP;
130 		else
131 			mb_use_ext_pgs = extpg != 0;
132 	}
133 	return (error);
134 }
135 SYSCTL_PROC(_kern_ipc, OID_AUTO, mb_use_ext_pgs,
136     CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
137     &mb_use_ext_pgs, 0, sysctl_mb_use_ext_pgs, "IU",
138     "Use unmapped mbufs for sendfile(2) and TLS offload");
139 
140 static quad_t maxmbufmem;	/* overall real memory limit for all mbufs */
141 
142 SYSCTL_QUAD(_kern_ipc, OID_AUTO, maxmbufmem, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &maxmbufmem, 0,
143     "Maximum real memory allocatable to various mbuf types");
144 
145 static counter_u64_t snd_tag_count;
146 SYSCTL_COUNTER_U64(_kern_ipc, OID_AUTO, num_snd_tags, CTLFLAG_RW,
147     &snd_tag_count, "# of active mbuf send tags");
148 
149 /*
150  * tunable_mbinit() has to be run before any mbuf allocations are done.
151  */
152 static void
153 tunable_mbinit(void *dummy)
154 {
155 	quad_t realmem;
156 	int extpg;
157 
158 	/*
159 	 * The default limit for all mbuf related memory is 1/2 of all
160 	 * available kernel memory (physical or kmem).
161 	 * At most it can be 3/4 of available kernel memory.
162 	 */
163 	realmem = qmin((quad_t)physmem * PAGE_SIZE, vm_kmem_size);
164 	maxmbufmem = realmem / 2;
165 	TUNABLE_QUAD_FETCH("kern.ipc.maxmbufmem", &maxmbufmem);
166 	if (maxmbufmem > realmem / 4 * 3)
167 		maxmbufmem = realmem / 4 * 3;
168 
169 	TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
170 	if (nmbclusters == 0)
171 		nmbclusters = maxmbufmem / MCLBYTES / 4;
172 
173 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbop", &nmbjumbop);
174 	if (nmbjumbop == 0)
175 		nmbjumbop = maxmbufmem / MJUMPAGESIZE / 4;
176 
177 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbo9", &nmbjumbo9);
178 	if (nmbjumbo9 == 0)
179 		nmbjumbo9 = maxmbufmem / MJUM9BYTES / 6;
180 
181 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbo16", &nmbjumbo16);
182 	if (nmbjumbo16 == 0)
183 		nmbjumbo16 = maxmbufmem / MJUM16BYTES / 6;
184 
185 	/*
186 	 * We need at least as many mbufs as we have clusters of
187 	 * the various types added together.
188 	 */
189 	TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
190 	if (nmbufs < nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16)
191 		nmbufs = lmax(maxmbufmem / MSIZE / 5,
192 		    nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16);
193 
194 	/*
195 	 * Unmapped mbufs can only safely be used on platforms with a direct
196 	 * map.
197 	 */
198 	if (PMAP_HAS_DMAP) {
199 		extpg = 1;
200 		TUNABLE_INT_FETCH("kern.ipc.mb_use_ext_pgs", &extpg);
201 		mb_use_ext_pgs = extpg != 0;
202 	}
203 }
204 SYSINIT(tunable_mbinit, SI_SUB_KMEM, SI_ORDER_MIDDLE, tunable_mbinit, NULL);
205 
206 static int
207 sysctl_nmbclusters(SYSCTL_HANDLER_ARGS)
208 {
209 	int error, newnmbclusters;
210 
211 	newnmbclusters = nmbclusters;
212 	error = sysctl_handle_int(oidp, &newnmbclusters, 0, req);
213 	if (error == 0 && req->newptr && newnmbclusters != nmbclusters) {
214 		if (newnmbclusters > nmbclusters &&
215 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
216 			nmbclusters = newnmbclusters;
217 			nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
218 			EVENTHANDLER_INVOKE(nmbclusters_change);
219 		} else
220 			error = EINVAL;
221 	}
222 	return (error);
223 }
224 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbclusters,
225     CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE,
226     &nmbclusters, 0, sysctl_nmbclusters, "IU",
227     "Maximum number of mbuf clusters allowed");
228 
229 static int
230 sysctl_nmbjumbop(SYSCTL_HANDLER_ARGS)
231 {
232 	int error, newnmbjumbop;
233 
234 	newnmbjumbop = nmbjumbop;
235 	error = sysctl_handle_int(oidp, &newnmbjumbop, 0, req);
236 	if (error == 0 && req->newptr && newnmbjumbop != nmbjumbop) {
237 		if (newnmbjumbop > nmbjumbop &&
238 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
239 			nmbjumbop = newnmbjumbop;
240 			nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
241 		} else
242 			error = EINVAL;
243 	}
244 	return (error);
245 }
246 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbop,
247     CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE,
248     &nmbjumbop, 0, sysctl_nmbjumbop, "IU",
249     "Maximum number of mbuf page size jumbo clusters allowed");
250 
251 static int
252 sysctl_nmbjumbo9(SYSCTL_HANDLER_ARGS)
253 {
254 	int error, newnmbjumbo9;
255 
256 	newnmbjumbo9 = nmbjumbo9;
257 	error = sysctl_handle_int(oidp, &newnmbjumbo9, 0, req);
258 	if (error == 0 && req->newptr && newnmbjumbo9 != nmbjumbo9) {
259 		if (newnmbjumbo9 > nmbjumbo9 &&
260 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
261 			nmbjumbo9 = newnmbjumbo9;
262 			nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
263 		} else
264 			error = EINVAL;
265 	}
266 	return (error);
267 }
268 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo9,
269     CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE,
270     &nmbjumbo9, 0, sysctl_nmbjumbo9, "IU",
271     "Maximum number of mbuf 9k jumbo clusters allowed");
272 
273 static int
274 sysctl_nmbjumbo16(SYSCTL_HANDLER_ARGS)
275 {
276 	int error, newnmbjumbo16;
277 
278 	newnmbjumbo16 = nmbjumbo16;
279 	error = sysctl_handle_int(oidp, &newnmbjumbo16, 0, req);
280 	if (error == 0 && req->newptr && newnmbjumbo16 != nmbjumbo16) {
281 		if (newnmbjumbo16 > nmbjumbo16 &&
282 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
283 			nmbjumbo16 = newnmbjumbo16;
284 			nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
285 		} else
286 			error = EINVAL;
287 	}
288 	return (error);
289 }
290 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo16,
291     CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE,
292     &nmbjumbo16, 0, sysctl_nmbjumbo16, "IU",
293     "Maximum number of mbuf 16k jumbo clusters allowed");
294 
295 static int
296 sysctl_nmbufs(SYSCTL_HANDLER_ARGS)
297 {
298 	int error, newnmbufs;
299 
300 	newnmbufs = nmbufs;
301 	error = sysctl_handle_int(oidp, &newnmbufs, 0, req);
302 	if (error == 0 && req->newptr && newnmbufs != nmbufs) {
303 		if (newnmbufs > nmbufs) {
304 			nmbufs = newnmbufs;
305 			nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
306 			EVENTHANDLER_INVOKE(nmbufs_change);
307 		} else
308 			error = EINVAL;
309 	}
310 	return (error);
311 }
312 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbufs,
313     CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE,
314     &nmbufs, 0, sysctl_nmbufs, "IU",
315     "Maximum number of mbufs allowed");
316 
317 /*
318  * Zones from which we allocate.
319  */
320 uma_zone_t	zone_mbuf;
321 uma_zone_t	zone_clust;
322 uma_zone_t	zone_pack;
323 uma_zone_t	zone_jumbop;
324 uma_zone_t	zone_jumbo9;
325 uma_zone_t	zone_jumbo16;
326 
327 /*
328  * Local prototypes.
329  */
330 static int	mb_ctor_mbuf(void *, int, void *, int);
331 static int	mb_ctor_clust(void *, int, void *, int);
332 static int	mb_ctor_pack(void *, int, void *, int);
333 static void	mb_dtor_mbuf(void *, int, void *);
334 static void	mb_dtor_pack(void *, int, void *);
335 static int	mb_zinit_pack(void *, int, int);
336 static void	mb_zfini_pack(void *, int);
337 static void	mb_reclaim(uma_zone_t, int);
338 
339 /* Ensure that MSIZE is a power of 2. */
340 CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE);
341 
342 _Static_assert(sizeof(struct mbuf) <= MSIZE,
343     "size of mbuf exceeds MSIZE");
344 /*
345  * Initialize FreeBSD Network buffer allocation.
346  */
347 static void
348 mbuf_init(void *dummy)
349 {
350 
351 	/*
352 	 * Configure UMA zones for Mbufs, Clusters, and Packets.
353 	 */
354 	zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE,
355 	    mb_ctor_mbuf, mb_dtor_mbuf, NULL, NULL,
356 	    MSIZE - 1, UMA_ZONE_CONTIG | UMA_ZONE_MAXBUCKET);
357 	if (nmbufs > 0)
358 		nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
359 	uma_zone_set_warning(zone_mbuf, "kern.ipc.nmbufs limit reached");
360 	uma_zone_set_maxaction(zone_mbuf, mb_reclaim);
361 
362 	zone_clust = uma_zcreate(MBUF_CLUSTER_MEM_NAME, MCLBYTES,
363 	    mb_ctor_clust, NULL, NULL, NULL,
364 	    UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
365 	if (nmbclusters > 0)
366 		nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
367 	uma_zone_set_warning(zone_clust, "kern.ipc.nmbclusters limit reached");
368 	uma_zone_set_maxaction(zone_clust, mb_reclaim);
369 
370 	zone_pack = uma_zsecond_create(MBUF_PACKET_MEM_NAME, mb_ctor_pack,
371 	    mb_dtor_pack, mb_zinit_pack, mb_zfini_pack, zone_mbuf);
372 
373 	/* Make jumbo frame zone too. Page size, 9k and 16k. */
374 	zone_jumbop = uma_zcreate(MBUF_JUMBOP_MEM_NAME, MJUMPAGESIZE,
375 	    mb_ctor_clust, NULL, NULL, NULL,
376 	    UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
377 	if (nmbjumbop > 0)
378 		nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
379 	uma_zone_set_warning(zone_jumbop, "kern.ipc.nmbjumbop limit reached");
380 	uma_zone_set_maxaction(zone_jumbop, mb_reclaim);
381 
382 	zone_jumbo9 = uma_zcreate(MBUF_JUMBO9_MEM_NAME, MJUM9BYTES,
383 	    mb_ctor_clust, NULL, NULL, NULL,
384 	    UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
385 	if (nmbjumbo9 > 0)
386 		nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
387 	uma_zone_set_warning(zone_jumbo9, "kern.ipc.nmbjumbo9 limit reached");
388 	uma_zone_set_maxaction(zone_jumbo9, mb_reclaim);
389 
390 	zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES,
391 	    mb_ctor_clust, NULL, NULL, NULL,
392 	    UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
393 	if (nmbjumbo16 > 0)
394 		nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
395 	uma_zone_set_warning(zone_jumbo16, "kern.ipc.nmbjumbo16 limit reached");
396 	uma_zone_set_maxaction(zone_jumbo16, mb_reclaim);
397 
398 	snd_tag_count = counter_u64_alloc(M_WAITOK);
399 }
400 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL);
401 
402 #ifdef DEBUGNET
403 /*
404  * debugnet makes use of a pre-allocated pool of mbufs and clusters.  When
405  * debugnet is configured, we initialize a set of UMA cache zones which return
406  * items from this pool.  At panic-time, the regular UMA zone pointers are
407  * overwritten with those of the cache zones so that drivers may allocate and
408  * free mbufs and clusters without attempting to allocate physical memory.
409  *
410  * We keep mbufs and clusters in a pair of mbuf queues.  In particular, for
411  * the purpose of caching clusters, we treat them as mbufs.
412  */
413 static struct mbufq dn_mbufq =
414     { STAILQ_HEAD_INITIALIZER(dn_mbufq.mq_head), 0, INT_MAX };
415 static struct mbufq dn_clustq =
416     { STAILQ_HEAD_INITIALIZER(dn_clustq.mq_head), 0, INT_MAX };
417 
418 static int dn_clsize;
419 static uma_zone_t dn_zone_mbuf;
420 static uma_zone_t dn_zone_clust;
421 static uma_zone_t dn_zone_pack;
422 
423 static struct debugnet_saved_zones {
424 	uma_zone_t dsz_mbuf;
425 	uma_zone_t dsz_clust;
426 	uma_zone_t dsz_pack;
427 	uma_zone_t dsz_jumbop;
428 	uma_zone_t dsz_jumbo9;
429 	uma_zone_t dsz_jumbo16;
430 	bool dsz_debugnet_zones_enabled;
431 } dn_saved_zones;
432 
433 static int
434 dn_buf_import(void *arg, void **store, int count, int domain __unused,
435     int flags)
436 {
437 	struct mbufq *q;
438 	struct mbuf *m;
439 	int i;
440 
441 	q = arg;
442 
443 	for (i = 0; i < count; i++) {
444 		m = mbufq_dequeue(q);
445 		if (m == NULL)
446 			break;
447 		trash_init(m, q == &dn_mbufq ? MSIZE : dn_clsize, flags);
448 		store[i] = m;
449 	}
450 	KASSERT((flags & M_WAITOK) == 0 || i == count,
451 	    ("%s: ran out of pre-allocated mbufs", __func__));
452 	return (i);
453 }
454 
455 static void
456 dn_buf_release(void *arg, void **store, int count)
457 {
458 	struct mbufq *q;
459 	struct mbuf *m;
460 	int i;
461 
462 	q = arg;
463 
464 	for (i = 0; i < count; i++) {
465 		m = store[i];
466 		(void)mbufq_enqueue(q, m);
467 	}
468 }
469 
470 static int
471 dn_pack_import(void *arg __unused, void **store, int count, int domain __unused,
472     int flags __unused)
473 {
474 	struct mbuf *m;
475 	void *clust;
476 	int i;
477 
478 	for (i = 0; i < count; i++) {
479 		m = m_get(M_NOWAIT, MT_DATA);
480 		if (m == NULL)
481 			break;
482 		clust = uma_zalloc(dn_zone_clust, M_NOWAIT);
483 		if (clust == NULL) {
484 			m_free(m);
485 			break;
486 		}
487 		mb_ctor_clust(clust, dn_clsize, m, 0);
488 		store[i] = m;
489 	}
490 	KASSERT((flags & M_WAITOK) == 0 || i == count,
491 	    ("%s: ran out of pre-allocated mbufs", __func__));
492 	return (i);
493 }
494 
495 static void
496 dn_pack_release(void *arg __unused, void **store, int count)
497 {
498 	struct mbuf *m;
499 	void *clust;
500 	int i;
501 
502 	for (i = 0; i < count; i++) {
503 		m = store[i];
504 		clust = m->m_ext.ext_buf;
505 		uma_zfree(dn_zone_clust, clust);
506 		uma_zfree(dn_zone_mbuf, m);
507 	}
508 }
509 
510 /*
511  * Free the pre-allocated mbufs and clusters reserved for debugnet, and destroy
512  * the corresponding UMA cache zones.
513  */
514 void
515 debugnet_mbuf_drain(void)
516 {
517 	struct mbuf *m;
518 	void *item;
519 
520 	if (dn_zone_mbuf != NULL) {
521 		uma_zdestroy(dn_zone_mbuf);
522 		dn_zone_mbuf = NULL;
523 	}
524 	if (dn_zone_clust != NULL) {
525 		uma_zdestroy(dn_zone_clust);
526 		dn_zone_clust = NULL;
527 	}
528 	if (dn_zone_pack != NULL) {
529 		uma_zdestroy(dn_zone_pack);
530 		dn_zone_pack = NULL;
531 	}
532 
533 	while ((m = mbufq_dequeue(&dn_mbufq)) != NULL)
534 		m_free(m);
535 	while ((item = mbufq_dequeue(&dn_clustq)) != NULL)
536 		uma_zfree(m_getzone(dn_clsize), item);
537 }
538 
539 /*
540  * Callback invoked immediately prior to starting a debugnet connection.
541  */
542 void
543 debugnet_mbuf_start(void)
544 {
545 
546 	MPASS(!dn_saved_zones.dsz_debugnet_zones_enabled);
547 
548 	/* Save the old zone pointers to restore when debugnet is closed. */
549 	dn_saved_zones = (struct debugnet_saved_zones) {
550 		.dsz_debugnet_zones_enabled = true,
551 		.dsz_mbuf = zone_mbuf,
552 		.dsz_clust = zone_clust,
553 		.dsz_pack = zone_pack,
554 		.dsz_jumbop = zone_jumbop,
555 		.dsz_jumbo9 = zone_jumbo9,
556 		.dsz_jumbo16 = zone_jumbo16,
557 	};
558 
559 	/*
560 	 * All cluster zones return buffers of the size requested by the
561 	 * drivers.  It's up to the driver to reinitialize the zones if the
562 	 * MTU of a debugnet-enabled interface changes.
563 	 */
564 	printf("debugnet: overwriting mbuf zone pointers\n");
565 	zone_mbuf = dn_zone_mbuf;
566 	zone_clust = dn_zone_clust;
567 	zone_pack = dn_zone_pack;
568 	zone_jumbop = dn_zone_clust;
569 	zone_jumbo9 = dn_zone_clust;
570 	zone_jumbo16 = dn_zone_clust;
571 }
572 
573 /*
574  * Callback invoked when a debugnet connection is closed/finished.
575  */
576 void
577 debugnet_mbuf_finish(void)
578 {
579 
580 	MPASS(dn_saved_zones.dsz_debugnet_zones_enabled);
581 
582 	printf("debugnet: restoring mbuf zone pointers\n");
583 	zone_mbuf = dn_saved_zones.dsz_mbuf;
584 	zone_clust = dn_saved_zones.dsz_clust;
585 	zone_pack = dn_saved_zones.dsz_pack;
586 	zone_jumbop = dn_saved_zones.dsz_jumbop;
587 	zone_jumbo9 = dn_saved_zones.dsz_jumbo9;
588 	zone_jumbo16 = dn_saved_zones.dsz_jumbo16;
589 
590 	memset(&dn_saved_zones, 0, sizeof(dn_saved_zones));
591 }
592 
593 /*
594  * Reinitialize the debugnet mbuf+cluster pool and cache zones.
595  */
596 void
597 debugnet_mbuf_reinit(int nmbuf, int nclust, int clsize)
598 {
599 	struct mbuf *m;
600 	void *item;
601 
602 	debugnet_mbuf_drain();
603 
604 	dn_clsize = clsize;
605 
606 	dn_zone_mbuf = uma_zcache_create("debugnet_" MBUF_MEM_NAME,
607 	    MSIZE, mb_ctor_mbuf, mb_dtor_mbuf, NULL, NULL,
608 	    dn_buf_import, dn_buf_release,
609 	    &dn_mbufq, UMA_ZONE_NOBUCKET);
610 
611 	dn_zone_clust = uma_zcache_create("debugnet_" MBUF_CLUSTER_MEM_NAME,
612 	    clsize, mb_ctor_clust, NULL, NULL, NULL,
613 	    dn_buf_import, dn_buf_release,
614 	    &dn_clustq, UMA_ZONE_NOBUCKET);
615 
616 	dn_zone_pack = uma_zcache_create("debugnet_" MBUF_PACKET_MEM_NAME,
617 	    MCLBYTES, mb_ctor_pack, mb_dtor_pack, NULL, NULL,
618 	    dn_pack_import, dn_pack_release,
619 	    NULL, UMA_ZONE_NOBUCKET);
620 
621 	while (nmbuf-- > 0) {
622 		m = m_get(M_WAITOK, MT_DATA);
623 		uma_zfree(dn_zone_mbuf, m);
624 	}
625 	while (nclust-- > 0) {
626 		item = uma_zalloc(m_getzone(dn_clsize), M_WAITOK);
627 		uma_zfree(dn_zone_clust, item);
628 	}
629 }
630 #endif /* DEBUGNET */
631 
632 /*
633  * Constructor for Mbuf primary zone.
634  *
635  * The 'arg' pointer points to a mb_args structure which
636  * contains call-specific information required to support the
637  * mbuf allocation API.  See mbuf.h.
638  */
639 static int
640 mb_ctor_mbuf(void *mem, int size, void *arg, int how)
641 {
642 	struct mbuf *m;
643 	struct mb_args *args;
644 	int error;
645 	int flags;
646 	short type;
647 
648 	args = (struct mb_args *)arg;
649 	type = args->type;
650 
651 	/*
652 	 * The mbuf is initialized later.  The caller has the
653 	 * responsibility to set up any MAC labels too.
654 	 */
655 	if (type == MT_NOINIT)
656 		return (0);
657 
658 	m = (struct mbuf *)mem;
659 	flags = args->flags;
660 	MPASS((flags & M_NOFREE) == 0);
661 
662 	error = m_init(m, how, type, flags);
663 
664 	return (error);
665 }
666 
667 /*
668  * The Mbuf primary zone destructor.
669  */
670 static void
671 mb_dtor_mbuf(void *mem, int size, void *arg)
672 {
673 	struct mbuf *m;
674 	unsigned long flags __diagused;
675 
676 	m = (struct mbuf *)mem;
677 	flags = (unsigned long)arg;
678 
679 	KASSERT((m->m_flags & M_NOFREE) == 0, ("%s: M_NOFREE set", __func__));
680 	KASSERT((flags & 0x1) == 0, ("%s: obsolete MB_DTOR_SKIP passed", __func__));
681 	if ((m->m_flags & M_PKTHDR) && !SLIST_EMPTY(&m->m_pkthdr.tags))
682 		m_tag_delete_chain(m, NULL);
683 }
684 
685 /*
686  * The Mbuf Packet zone destructor.
687  */
688 static void
689 mb_dtor_pack(void *mem, int size, void *arg)
690 {
691 	struct mbuf *m;
692 
693 	m = (struct mbuf *)mem;
694 	if ((m->m_flags & M_PKTHDR) != 0)
695 		m_tag_delete_chain(m, NULL);
696 
697 	/* Make sure we've got a clean cluster back. */
698 	KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__));
699 	KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__));
700 	KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__));
701 	KASSERT(m->m_ext.ext_arg1 == NULL, ("%s: ext_arg1 != NULL", __func__));
702 	KASSERT(m->m_ext.ext_arg2 == NULL, ("%s: ext_arg2 != NULL", __func__));
703 	KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__));
704 	KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__));
705 #if defined(INVARIANTS) && !defined(KMSAN)
706 	trash_dtor(m->m_ext.ext_buf, MCLBYTES, zone_clust);
707 #endif
708 	/*
709 	 * If there are processes blocked on zone_clust, waiting for pages
710 	 * to be freed up, cause them to be woken up by draining the
711 	 * packet zone.  We are exposed to a race here (in the check for
712 	 * the UMA_ZFLAG_FULL) where we might miss the flag set, but that
713 	 * is deliberate. We don't want to acquire the zone lock for every
714 	 * mbuf free.
715 	 */
716 	if (uma_zone_exhausted(zone_clust))
717 		uma_zone_reclaim(zone_pack, UMA_RECLAIM_DRAIN);
718 }
719 
720 /*
721  * The Cluster and Jumbo[PAGESIZE|9|16] zone constructor.
722  *
723  * Here the 'arg' pointer points to the Mbuf which we
724  * are configuring cluster storage for.  If 'arg' is
725  * empty we allocate just the cluster without setting
726  * the mbuf to it.  See mbuf.h.
727  */
728 static int
729 mb_ctor_clust(void *mem, int size, void *arg, int how)
730 {
731 	struct mbuf *m;
732 
733 	m = (struct mbuf *)arg;
734 	if (m != NULL) {
735 		m->m_ext.ext_buf = (char *)mem;
736 		m->m_data = m->m_ext.ext_buf;
737 		m->m_flags |= M_EXT;
738 		m->m_ext.ext_free = NULL;
739 		m->m_ext.ext_arg1 = NULL;
740 		m->m_ext.ext_arg2 = NULL;
741 		m->m_ext.ext_size = size;
742 		m->m_ext.ext_type = m_gettype(size);
743 		m->m_ext.ext_flags = EXT_FLAG_EMBREF;
744 		m->m_ext.ext_count = 1;
745 	}
746 
747 	return (0);
748 }
749 
750 /*
751  * The Packet secondary zone's init routine, executed on the
752  * object's transition from mbuf keg slab to zone cache.
753  */
754 static int
755 mb_zinit_pack(void *mem, int size, int how)
756 {
757 	struct mbuf *m;
758 
759 	m = (struct mbuf *)mem;		/* m is virgin. */
760 	if (uma_zalloc_arg(zone_clust, m, how) == NULL ||
761 	    m->m_ext.ext_buf == NULL)
762 		return (ENOMEM);
763 	m->m_ext.ext_type = EXT_PACKET;	/* Override. */
764 #if defined(INVARIANTS) && !defined(KMSAN)
765 	trash_init(m->m_ext.ext_buf, MCLBYTES, how);
766 #endif
767 	return (0);
768 }
769 
770 /*
771  * The Packet secondary zone's fini routine, executed on the
772  * object's transition from zone cache to keg slab.
773  */
774 static void
775 mb_zfini_pack(void *mem, int size)
776 {
777 	struct mbuf *m;
778 
779 	m = (struct mbuf *)mem;
780 #if defined(INVARIANTS) && !defined(KMSAN)
781 	trash_fini(m->m_ext.ext_buf, MCLBYTES);
782 #endif
783 	uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL);
784 #if defined(INVARIANTS) && !defined(KMSAN)
785 	trash_dtor(mem, size, zone_clust);
786 #endif
787 }
788 
789 /*
790  * The "packet" keg constructor.
791  */
792 static int
793 mb_ctor_pack(void *mem, int size, void *arg, int how)
794 {
795 	struct mbuf *m;
796 	struct mb_args *args;
797 	int error, flags;
798 	short type;
799 
800 	m = (struct mbuf *)mem;
801 	args = (struct mb_args *)arg;
802 	flags = args->flags;
803 	type = args->type;
804 	MPASS((flags & M_NOFREE) == 0);
805 
806 #if defined(INVARIANTS) && !defined(KMSAN)
807 	trash_ctor(m->m_ext.ext_buf, MCLBYTES, zone_clust, how);
808 #endif
809 
810 	error = m_init(m, how, type, flags);
811 
812 	/* m_ext is already initialized. */
813 	m->m_data = m->m_ext.ext_buf;
814  	m->m_flags = (flags | M_EXT);
815 
816 	return (error);
817 }
818 
819 /*
820  * This is the protocol drain routine.  Called by UMA whenever any of the
821  * mbuf zones is closed to its limit.
822  */
823 static void
824 mb_reclaim(uma_zone_t zone __unused, int pending __unused)
825 {
826 
827 	EVENTHANDLER_INVOKE(mbuf_lowmem, VM_LOW_MBUFS);
828 }
829 
830 /*
831  * Free "count" units of I/O from an mbuf chain.  They could be held
832  * in M_EXTPG or just as a normal mbuf.  This code is intended to be
833  * called in an error path (I/O error, closed connection, etc).
834  */
835 void
836 mb_free_notready(struct mbuf *m, int count)
837 {
838 	int i;
839 
840 	for (i = 0; i < count && m != NULL; i++) {
841 		if ((m->m_flags & M_EXTPG) != 0) {
842 			m->m_epg_nrdy--;
843 			if (m->m_epg_nrdy != 0)
844 				continue;
845 		}
846 		m = m_free(m);
847 	}
848 	KASSERT(i == count, ("Removed only %d items from %p", i, m));
849 }
850 
851 /*
852  * Compress an unmapped mbuf into a simple mbuf when it holds a small
853  * amount of data.  This is used as a DOS defense to avoid having
854  * small packets tie up wired pages, an ext_pgs structure, and an
855  * mbuf.  Since this converts the existing mbuf in place, it can only
856  * be used if there are no other references to 'm'.
857  */
858 int
859 mb_unmapped_compress(struct mbuf *m)
860 {
861 	volatile u_int *refcnt;
862 	char buf[MLEN];
863 
864 	/*
865 	 * Assert that 'm' does not have a packet header.  If 'm' had
866 	 * a packet header, it would only be able to hold MHLEN bytes
867 	 * and m_data would have to be initialized differently.
868 	 */
869 	KASSERT((m->m_flags & M_PKTHDR) == 0 && (m->m_flags & M_EXTPG),
870             ("%s: m %p !M_EXTPG or M_PKTHDR", __func__, m));
871 	KASSERT(m->m_len <= MLEN, ("m_len too large %p", m));
872 
873 	if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
874 		refcnt = &m->m_ext.ext_count;
875 	} else {
876 		KASSERT(m->m_ext.ext_cnt != NULL,
877 		    ("%s: no refcounting pointer on %p", __func__, m));
878 		refcnt = m->m_ext.ext_cnt;
879 	}
880 
881 	if (*refcnt != 1)
882 		return (EBUSY);
883 
884 	m_copydata(m, 0, m->m_len, buf);
885 
886 	/* Free the backing pages. */
887 	m->m_ext.ext_free(m);
888 
889 	/* Turn 'm' into a "normal" mbuf. */
890 	m->m_flags &= ~(M_EXT | M_RDONLY | M_EXTPG);
891 	m->m_data = m->m_dat;
892 
893 	/* Copy data back into m. */
894 	bcopy(buf, mtod(m, char *), m->m_len);
895 
896 	return (0);
897 }
898 
899 /*
900  * These next few routines are used to permit downgrading an unmapped
901  * mbuf to a chain of mapped mbufs.  This is used when an interface
902  * doesn't supported unmapped mbufs or if checksums need to be
903  * computed in software.
904  *
905  * Each unmapped mbuf is converted to a chain of mbufs.  First, any
906  * TLS header data is stored in a regular mbuf.  Second, each page of
907  * unmapped data is stored in an mbuf with an EXT_SFBUF external
908  * cluster.  These mbufs use an sf_buf to provide a valid KVA for the
909  * associated physical page.  They also hold a reference on the
910  * original M_EXTPG mbuf to ensure the physical page doesn't go away.
911  * Finally, any TLS trailer data is stored in a regular mbuf.
912  *
913  * mb_unmapped_free_mext() is the ext_free handler for the EXT_SFBUF
914  * mbufs.  It frees the associated sf_buf and releases its reference
915  * on the original M_EXTPG mbuf.
916  *
917  * _mb_unmapped_to_ext() is a helper function that converts a single
918  * unmapped mbuf into a chain of mbufs.
919  *
920  * mb_unmapped_to_ext() is the public function that walks an mbuf
921  * chain converting any unmapped mbufs to mapped mbufs.  It returns
922  * the new chain of unmapped mbufs on success.  On failure it frees
923  * the original mbuf chain and returns NULL.
924  */
925 static void
926 mb_unmapped_free_mext(struct mbuf *m)
927 {
928 	struct sf_buf *sf;
929 	struct mbuf *old_m;
930 
931 	sf = m->m_ext.ext_arg1;
932 	sf_buf_free(sf);
933 
934 	/* Drop the reference on the backing M_EXTPG mbuf. */
935 	old_m = m->m_ext.ext_arg2;
936 	mb_free_extpg(old_m);
937 }
938 
939 static struct mbuf *
940 _mb_unmapped_to_ext(struct mbuf *m)
941 {
942 	struct mbuf *m_new, *top, *prev, *mref;
943 	struct sf_buf *sf;
944 	vm_page_t pg;
945 	int i, len, off, pglen, pgoff, seglen, segoff;
946 	volatile u_int *refcnt;
947 	u_int ref_inc = 0;
948 
949 	M_ASSERTEXTPG(m);
950 	len = m->m_len;
951 	KASSERT(m->m_epg_tls == NULL, ("%s: can't convert TLS mbuf %p",
952 	    __func__, m));
953 
954 	/* See if this is the mbuf that holds the embedded refcount. */
955 	if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
956 		refcnt = &m->m_ext.ext_count;
957 		mref = m;
958 	} else {
959 		KASSERT(m->m_ext.ext_cnt != NULL,
960 		    ("%s: no refcounting pointer on %p", __func__, m));
961 		refcnt = m->m_ext.ext_cnt;
962 		mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
963 	}
964 
965 	/* Skip over any data removed from the front. */
966 	off = mtod(m, vm_offset_t);
967 
968 	top = NULL;
969 	if (m->m_epg_hdrlen != 0) {
970 		if (off >= m->m_epg_hdrlen) {
971 			off -= m->m_epg_hdrlen;
972 		} else {
973 			seglen = m->m_epg_hdrlen - off;
974 			segoff = off;
975 			seglen = min(seglen, len);
976 			off = 0;
977 			len -= seglen;
978 			m_new = m_get(M_NOWAIT, MT_DATA);
979 			if (m_new == NULL)
980 				goto fail;
981 			m_new->m_len = seglen;
982 			prev = top = m_new;
983 			memcpy(mtod(m_new, void *), &m->m_epg_hdr[segoff],
984 			    seglen);
985 		}
986 	}
987 	pgoff = m->m_epg_1st_off;
988 	for (i = 0; i < m->m_epg_npgs && len > 0; i++) {
989 		pglen = m_epg_pagelen(m, i, pgoff);
990 		if (off >= pglen) {
991 			off -= pglen;
992 			pgoff = 0;
993 			continue;
994 		}
995 		seglen = pglen - off;
996 		segoff = pgoff + off;
997 		off = 0;
998 		seglen = min(seglen, len);
999 		len -= seglen;
1000 
1001 		pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]);
1002 		m_new = m_get(M_NOWAIT, MT_DATA);
1003 		if (m_new == NULL)
1004 			goto fail;
1005 		if (top == NULL) {
1006 			top = prev = m_new;
1007 		} else {
1008 			prev->m_next = m_new;
1009 			prev = m_new;
1010 		}
1011 		sf = sf_buf_alloc(pg, SFB_NOWAIT);
1012 		if (sf == NULL)
1013 			goto fail;
1014 
1015 		ref_inc++;
1016 		m_extadd(m_new, (char *)sf_buf_kva(sf), PAGE_SIZE,
1017 		    mb_unmapped_free_mext, sf, mref, M_RDONLY, EXT_SFBUF);
1018 		m_new->m_data += segoff;
1019 		m_new->m_len = seglen;
1020 
1021 		pgoff = 0;
1022 	};
1023 	if (len != 0) {
1024 		KASSERT((off + len) <= m->m_epg_trllen,
1025 		    ("off + len > trail (%d + %d > %d)", off, len,
1026 		    m->m_epg_trllen));
1027 		m_new = m_get(M_NOWAIT, MT_DATA);
1028 		if (m_new == NULL)
1029 			goto fail;
1030 		if (top == NULL)
1031 			top = m_new;
1032 		else
1033 			prev->m_next = m_new;
1034 		m_new->m_len = len;
1035 		memcpy(mtod(m_new, void *), &m->m_epg_trail[off], len);
1036 	}
1037 
1038 	if (ref_inc != 0) {
1039 		/*
1040 		 * Obtain an additional reference on the old mbuf for
1041 		 * each created EXT_SFBUF mbuf.  They will be dropped
1042 		 * in mb_unmapped_free_mext().
1043 		 */
1044 		if (*refcnt == 1)
1045 			*refcnt += ref_inc;
1046 		else
1047 			atomic_add_int(refcnt, ref_inc);
1048 	}
1049 	m_free(m);
1050 	return (top);
1051 
1052 fail:
1053 	if (ref_inc != 0) {
1054 		/*
1055 		 * Obtain an additional reference on the old mbuf for
1056 		 * each created EXT_SFBUF mbuf.  They will be
1057 		 * immediately dropped when these mbufs are freed
1058 		 * below.
1059 		 */
1060 		if (*refcnt == 1)
1061 			*refcnt += ref_inc;
1062 		else
1063 			atomic_add_int(refcnt, ref_inc);
1064 	}
1065 	m_free(m);
1066 	m_freem(top);
1067 	return (NULL);
1068 }
1069 
1070 struct mbuf *
1071 mb_unmapped_to_ext(struct mbuf *top)
1072 {
1073 	struct mbuf *m, *next, *prev = NULL;
1074 
1075 	prev = NULL;
1076 	for (m = top; m != NULL; m = next) {
1077 		/* m might be freed, so cache the next pointer. */
1078 		next = m->m_next;
1079 		if (m->m_flags & M_EXTPG) {
1080 			if (prev != NULL) {
1081 				/*
1082 				 * Remove 'm' from the new chain so
1083 				 * that the 'top' chain terminates
1084 				 * before 'm' in case 'top' is freed
1085 				 * due to an error.
1086 				 */
1087 				prev->m_next = NULL;
1088 			}
1089 			m = _mb_unmapped_to_ext(m);
1090 			if (m == NULL) {
1091 				m_freem(top);
1092 				m_freem(next);
1093 				return (NULL);
1094 			}
1095 			if (prev == NULL) {
1096 				top = m;
1097 			} else {
1098 				prev->m_next = m;
1099 			}
1100 
1101 			/*
1102 			 * Replaced one mbuf with a chain, so we must
1103 			 * find the end of chain.
1104 			 */
1105 			prev = m_last(m);
1106 		} else {
1107 			if (prev != NULL) {
1108 				prev->m_next = m;
1109 			}
1110 			prev = m;
1111 		}
1112 	}
1113 	return (top);
1114 }
1115 
1116 /*
1117  * Allocate an empty M_EXTPG mbuf.  The ext_free routine is
1118  * responsible for freeing any pages backing this mbuf when it is
1119  * freed.
1120  */
1121 struct mbuf *
1122 mb_alloc_ext_pgs(int how, m_ext_free_t ext_free)
1123 {
1124 	struct mbuf *m;
1125 
1126 	m = m_get(how, MT_DATA);
1127 	if (m == NULL)
1128 		return (NULL);
1129 
1130 	m->m_epg_npgs = 0;
1131 	m->m_epg_nrdy = 0;
1132 	m->m_epg_1st_off = 0;
1133 	m->m_epg_last_len = 0;
1134 	m->m_epg_flags = 0;
1135 	m->m_epg_hdrlen = 0;
1136 	m->m_epg_trllen = 0;
1137 	m->m_epg_tls = NULL;
1138 	m->m_epg_so = NULL;
1139 	m->m_data = NULL;
1140 	m->m_flags |= (M_EXT | M_RDONLY | M_EXTPG);
1141 	m->m_ext.ext_flags = EXT_FLAG_EMBREF;
1142 	m->m_ext.ext_count = 1;
1143 	m->m_ext.ext_size = 0;
1144 	m->m_ext.ext_free = ext_free;
1145 	return (m);
1146 }
1147 
1148 /*
1149  * Clean up after mbufs with M_EXT storage attached to them if the
1150  * reference count hits 1.
1151  */
1152 void
1153 mb_free_ext(struct mbuf *m)
1154 {
1155 	volatile u_int *refcnt;
1156 	struct mbuf *mref;
1157 	int freembuf;
1158 
1159 	KASSERT(m->m_flags & M_EXT, ("%s: M_EXT not set on %p", __func__, m));
1160 
1161 	/* See if this is the mbuf that holds the embedded refcount. */
1162 	if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
1163 		refcnt = &m->m_ext.ext_count;
1164 		mref = m;
1165 	} else {
1166 		KASSERT(m->m_ext.ext_cnt != NULL,
1167 		    ("%s: no refcounting pointer on %p", __func__, m));
1168 		refcnt = m->m_ext.ext_cnt;
1169 		mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
1170 	}
1171 
1172 	/*
1173 	 * Check if the header is embedded in the cluster.  It is
1174 	 * important that we can't touch any of the mbuf fields
1175 	 * after we have freed the external storage, since mbuf
1176 	 * could have been embedded in it.  For now, the mbufs
1177 	 * embedded into the cluster are always of type EXT_EXTREF,
1178 	 * and for this type we won't free the mref.
1179 	 */
1180 	if (m->m_flags & M_NOFREE) {
1181 		freembuf = 0;
1182 		KASSERT(m->m_ext.ext_type == EXT_EXTREF ||
1183 		    m->m_ext.ext_type == EXT_RXRING,
1184 		    ("%s: no-free mbuf %p has wrong type", __func__, m));
1185 	} else
1186 		freembuf = 1;
1187 
1188 	/* Free attached storage if this mbuf is the only reference to it. */
1189 	if (*refcnt == 1 || atomic_fetchadd_int(refcnt, -1) == 1) {
1190 		switch (m->m_ext.ext_type) {
1191 		case EXT_PACKET:
1192 			/* The packet zone is special. */
1193 			if (*refcnt == 0)
1194 				*refcnt = 1;
1195 			uma_zfree(zone_pack, mref);
1196 			break;
1197 		case EXT_CLUSTER:
1198 			uma_zfree(zone_clust, m->m_ext.ext_buf);
1199 			m_free_raw(mref);
1200 			break;
1201 		case EXT_JUMBOP:
1202 			uma_zfree(zone_jumbop, m->m_ext.ext_buf);
1203 			m_free_raw(mref);
1204 			break;
1205 		case EXT_JUMBO9:
1206 			uma_zfree(zone_jumbo9, m->m_ext.ext_buf);
1207 			m_free_raw(mref);
1208 			break;
1209 		case EXT_JUMBO16:
1210 			uma_zfree(zone_jumbo16, m->m_ext.ext_buf);
1211 			m_free_raw(mref);
1212 			break;
1213 		case EXT_SFBUF:
1214 		case EXT_NET_DRV:
1215 		case EXT_CTL:
1216 		case EXT_MOD_TYPE:
1217 		case EXT_DISPOSABLE:
1218 			KASSERT(mref->m_ext.ext_free != NULL,
1219 			    ("%s: ext_free not set", __func__));
1220 			mref->m_ext.ext_free(mref);
1221 			m_free_raw(mref);
1222 			break;
1223 		case EXT_EXTREF:
1224 			KASSERT(m->m_ext.ext_free != NULL,
1225 			    ("%s: ext_free not set", __func__));
1226 			m->m_ext.ext_free(m);
1227 			break;
1228 		case EXT_RXRING:
1229 			KASSERT(m->m_ext.ext_free == NULL,
1230 			    ("%s: ext_free is set", __func__));
1231 			break;
1232 		default:
1233 			KASSERT(m->m_ext.ext_type == 0,
1234 			    ("%s: unknown ext_type", __func__));
1235 		}
1236 	}
1237 
1238 	if (freembuf && m != mref)
1239 		m_free_raw(m);
1240 }
1241 
1242 /*
1243  * Clean up after mbufs with M_EXTPG storage attached to them if the
1244  * reference count hits 1.
1245  */
1246 void
1247 mb_free_extpg(struct mbuf *m)
1248 {
1249 	volatile u_int *refcnt;
1250 	struct mbuf *mref;
1251 
1252 	M_ASSERTEXTPG(m);
1253 
1254 	/* See if this is the mbuf that holds the embedded refcount. */
1255 	if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
1256 		refcnt = &m->m_ext.ext_count;
1257 		mref = m;
1258 	} else {
1259 		KASSERT(m->m_ext.ext_cnt != NULL,
1260 		    ("%s: no refcounting pointer on %p", __func__, m));
1261 		refcnt = m->m_ext.ext_cnt;
1262 		mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
1263 	}
1264 
1265 	/* Free attached storage if this mbuf is the only reference to it. */
1266 	if (*refcnt == 1 || atomic_fetchadd_int(refcnt, -1) == 1) {
1267 		KASSERT(mref->m_ext.ext_free != NULL,
1268 		    ("%s: ext_free not set", __func__));
1269 
1270 		mref->m_ext.ext_free(mref);
1271 #ifdef KERN_TLS
1272 		if (mref->m_epg_tls != NULL &&
1273 		    !refcount_release_if_not_last(&mref->m_epg_tls->refcount))
1274 			ktls_enqueue_to_free(mref);
1275 		else
1276 #endif
1277 			m_free_raw(mref);
1278 	}
1279 
1280 	if (m != mref)
1281 		m_free_raw(m);
1282 }
1283 
1284 /*
1285  * Official mbuf(9) allocation KPI for stack and drivers:
1286  *
1287  * m_get()	- a single mbuf without any attachments, sys/mbuf.h.
1288  * m_gethdr()	- a single mbuf initialized as M_PKTHDR, sys/mbuf.h.
1289  * m_getcl()	- an mbuf + 2k cluster, sys/mbuf.h.
1290  * m_clget()	- attach cluster to already allocated mbuf.
1291  * m_cljget()	- attach jumbo cluster to already allocated mbuf.
1292  * m_get2()	- allocate minimum mbuf that would fit size argument.
1293  * m_getm2()	- allocate a chain of mbufs/clusters.
1294  * m_extadd()	- attach external cluster to mbuf.
1295  *
1296  * m_free()	- free single mbuf with its tags and ext, sys/mbuf.h.
1297  * m_freem()	- free chain of mbufs.
1298  */
1299 
1300 int
1301 m_clget(struct mbuf *m, int how)
1302 {
1303 
1304 	KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
1305 	    __func__, m));
1306 	m->m_ext.ext_buf = (char *)NULL;
1307 	uma_zalloc_arg(zone_clust, m, how);
1308 	/*
1309 	 * On a cluster allocation failure, drain the packet zone and retry,
1310 	 * we might be able to loosen a few clusters up on the drain.
1311 	 */
1312 	if ((how & M_NOWAIT) && (m->m_ext.ext_buf == NULL)) {
1313 		uma_zone_reclaim(zone_pack, UMA_RECLAIM_DRAIN);
1314 		uma_zalloc_arg(zone_clust, m, how);
1315 	}
1316 	MBUF_PROBE2(m__clget, m, how);
1317 	return (m->m_flags & M_EXT);
1318 }
1319 
1320 /*
1321  * m_cljget() is different from m_clget() as it can allocate clusters without
1322  * attaching them to an mbuf.  In that case the return value is the pointer
1323  * to the cluster of the requested size.  If an mbuf was specified, it gets
1324  * the cluster attached to it and the return value can be safely ignored.
1325  * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
1326  */
1327 void *
1328 m_cljget(struct mbuf *m, int how, int size)
1329 {
1330 	uma_zone_t zone;
1331 	void *retval;
1332 
1333 	if (m != NULL) {
1334 		KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
1335 		    __func__, m));
1336 		m->m_ext.ext_buf = NULL;
1337 	}
1338 
1339 	zone = m_getzone(size);
1340 	retval = uma_zalloc_arg(zone, m, how);
1341 
1342 	MBUF_PROBE4(m__cljget, m, how, size, retval);
1343 
1344 	return (retval);
1345 }
1346 
1347 /*
1348  * m_get2() allocates minimum mbuf that would fit "size" argument.
1349  */
1350 struct mbuf *
1351 m_get2(int size, int how, short type, int flags)
1352 {
1353 	struct mb_args args;
1354 	struct mbuf *m, *n;
1355 
1356 	args.flags = flags;
1357 	args.type = type;
1358 
1359 	if (size <= MHLEN || (size <= MLEN && (flags & M_PKTHDR) == 0))
1360 		return (uma_zalloc_arg(zone_mbuf, &args, how));
1361 	if (size <= MCLBYTES)
1362 		return (uma_zalloc_arg(zone_pack, &args, how));
1363 
1364 	if (size > MJUMPAGESIZE)
1365 		return (NULL);
1366 
1367 	m = uma_zalloc_arg(zone_mbuf, &args, how);
1368 	if (m == NULL)
1369 		return (NULL);
1370 
1371 	n = uma_zalloc_arg(zone_jumbop, m, how);
1372 	if (n == NULL) {
1373 		m_free_raw(m);
1374 		return (NULL);
1375 	}
1376 
1377 	return (m);
1378 }
1379 
1380 /*
1381  * m_get3() allocates minimum mbuf that would fit "size" argument.
1382  * Unlike m_get2() it can allocate clusters up to MJUM16BYTES.
1383  */
1384 struct mbuf *
1385 m_get3(int size, int how, short type, int flags)
1386 {
1387 	struct mb_args args;
1388 	struct mbuf *m, *n;
1389 	uma_zone_t zone;
1390 
1391 	if (size <= MJUMPAGESIZE)
1392 		return (m_get2(size, how, type, flags));
1393 
1394 	if (size > MJUM16BYTES)
1395 		return (NULL);
1396 
1397 	args.flags = flags;
1398 	args.type = type;
1399 
1400 	m = uma_zalloc_arg(zone_mbuf, &args, how);
1401 	if (m == NULL)
1402 		return (NULL);
1403 
1404 	if (size <= MJUM9BYTES)
1405 		zone = zone_jumbo9;
1406 	else
1407 		zone = zone_jumbo16;
1408 
1409 	n = uma_zalloc_arg(zone, m, how);
1410 	if (n == NULL) {
1411 		m_free_raw(m);
1412 		return (NULL);
1413 	}
1414 
1415 	return (m);
1416 }
1417 
1418 /*
1419  * m_getjcl() returns an mbuf with a cluster of the specified size attached.
1420  * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
1421  */
1422 struct mbuf *
1423 m_getjcl(int how, short type, int flags, int size)
1424 {
1425 	struct mb_args args;
1426 	struct mbuf *m, *n;
1427 	uma_zone_t zone;
1428 
1429 	if (size == MCLBYTES)
1430 		return m_getcl(how, type, flags);
1431 
1432 	args.flags = flags;
1433 	args.type = type;
1434 
1435 	m = uma_zalloc_arg(zone_mbuf, &args, how);
1436 	if (m == NULL)
1437 		return (NULL);
1438 
1439 	zone = m_getzone(size);
1440 	n = uma_zalloc_arg(zone, m, how);
1441 	if (n == NULL) {
1442 		m_free_raw(m);
1443 		return (NULL);
1444 	}
1445 	MBUF_PROBE5(m__getjcl, how, type, flags, size, m);
1446 	return (m);
1447 }
1448 
1449 /*
1450  * Allocate mchain of a given length of mbufs and/or clusters (whatever fits
1451  * best).  May fail due to ENOMEM.  In case of failure state of mchain is
1452  * inconsistent.
1453  */
1454 int
1455 mc_get(struct mchain *mc, u_int length, int how, short type, int flags)
1456 {
1457 	struct mbuf *mb;
1458 	u_int progress;
1459 
1460 	MPASS(length >= 0);
1461 
1462 	*mc = MCHAIN_INITIALIZER(mc);
1463 	flags &= (M_PKTHDR | M_EOR);
1464 	progress = 0;
1465 
1466 	/* Loop and append maximum sized mbufs to the chain tail. */
1467 	do {
1468 		if (length - progress > MCLBYTES) {
1469 			/*
1470 			 * M_NOWAIT here is intentional, it avoids blocking if
1471 			 * the jumbop zone is exhausted. See 796d4eb89e2c and
1472 			 * D26150 for more detail.
1473 			 */
1474 			mb = m_getjcl(M_NOWAIT, type, (flags & M_PKTHDR),
1475 			    MJUMPAGESIZE);
1476 		} else
1477 			mb = NULL;
1478 		if (mb == NULL) {
1479 			if (length - progress >= MINCLSIZE)
1480 				mb = m_getcl(how, type, (flags & M_PKTHDR));
1481 			else if (flags & M_PKTHDR)
1482 				mb = m_gethdr(how, type);
1483 			else
1484 				mb = m_get(how, type);
1485 
1486 			/*
1487 			 * Fail the whole operation if one mbuf can't be
1488 			 * allocated.
1489 			 */
1490 			if (mb == NULL) {
1491 				m_freem(mc_first(mc));
1492 				return (ENOMEM);
1493 			}
1494 		}
1495 
1496 		progress += M_SIZE(mb);
1497 		mc_append(mc, mb);
1498 		/* Only valid on the first mbuf. */
1499 		flags &= ~M_PKTHDR;
1500 	} while (progress < length);
1501 	if (flags & M_EOR)
1502 		/* Only valid on the last mbuf. */
1503 		mc_last(mc)->m_flags |= M_EOR;
1504 
1505 	return (0);
1506 }
1507 
1508 /*
1509  * Allocate a given length worth of mbufs and/or clusters (whatever fits
1510  * best) and return a pointer to the top of the allocated chain.  If an
1511  * existing mbuf chain is provided, then we will append the new chain
1512  * to the existing one and return a pointer to the provided mbuf.
1513  */
1514 struct mbuf *
1515 m_getm2(struct mbuf *m, int len, int how, short type, int flags)
1516 {
1517 	struct mchain mc;
1518 
1519 	/* Packet header mbuf must be first in chain. */
1520 	if (m != NULL && (flags & M_PKTHDR))
1521 		flags &= ~M_PKTHDR;
1522 
1523 	if (__predict_false(mc_get(&mc, len, how, type, flags) != 0))
1524 		return (NULL);
1525 
1526 	/* If mbuf was supplied, append new chain to the end of it. */
1527 	if (m != NULL) {
1528 		struct mbuf *mtail;
1529 
1530 		mtail = m_last(m);
1531 		mtail->m_next = mc_first(&mc);
1532 		mtail->m_flags &= ~M_EOR;
1533 	} else
1534 		m = mc_first(&mc);
1535 
1536 	return (m);
1537 }
1538 
1539 /*-
1540  * Configure a provided mbuf to refer to the provided external storage
1541  * buffer and setup a reference count for said buffer.
1542  *
1543  * Arguments:
1544  *    mb     The existing mbuf to which to attach the provided buffer.
1545  *    buf    The address of the provided external storage buffer.
1546  *    size   The size of the provided buffer.
1547  *    freef  A pointer to a routine that is responsible for freeing the
1548  *           provided external storage buffer.
1549  *    args   A pointer to an argument structure (of any type) to be passed
1550  *           to the provided freef routine (may be NULL).
1551  *    flags  Any other flags to be passed to the provided mbuf.
1552  *    type   The type that the external storage buffer should be
1553  *           labeled with.
1554  *
1555  * Returns:
1556  *    Nothing.
1557  */
1558 void
1559 m_extadd(struct mbuf *mb, char *buf, u_int size, m_ext_free_t freef,
1560     void *arg1, void *arg2, int flags, int type)
1561 {
1562 
1563 	KASSERT(type != EXT_CLUSTER, ("%s: EXT_CLUSTER not allowed", __func__));
1564 
1565 	mb->m_flags |= (M_EXT | flags);
1566 	mb->m_ext.ext_buf = buf;
1567 	mb->m_data = mb->m_ext.ext_buf;
1568 	mb->m_ext.ext_size = size;
1569 	mb->m_ext.ext_free = freef;
1570 	mb->m_ext.ext_arg1 = arg1;
1571 	mb->m_ext.ext_arg2 = arg2;
1572 	mb->m_ext.ext_type = type;
1573 
1574 	if (type != EXT_EXTREF) {
1575 		mb->m_ext.ext_count = 1;
1576 		mb->m_ext.ext_flags = EXT_FLAG_EMBREF;
1577 	} else
1578 		mb->m_ext.ext_flags = 0;
1579 }
1580 
1581 /*
1582  * Free an entire chain of mbufs and associated external buffers, if
1583  * applicable.
1584  */
1585 void
1586 m_freem(struct mbuf *mb)
1587 {
1588 
1589 	MBUF_PROBE1(m__freem, mb);
1590 	while (mb != NULL)
1591 		mb = m_free(mb);
1592 }
1593 
1594 /*
1595  * Temporary primitive to allow freeing without going through m_free.
1596  */
1597 void
1598 m_free_raw(struct mbuf *mb)
1599 {
1600 
1601 	uma_zfree(zone_mbuf, mb);
1602 }
1603 
1604 int
1605 m_snd_tag_alloc(struct ifnet *ifp, union if_snd_tag_alloc_params *params,
1606     struct m_snd_tag **mstp)
1607 {
1608 
1609 	return (if_snd_tag_alloc(ifp, params, mstp));
1610 }
1611 
1612 void
1613 m_snd_tag_init(struct m_snd_tag *mst, struct ifnet *ifp,
1614     const struct if_snd_tag_sw *sw)
1615 {
1616 
1617 	if_ref(ifp);
1618 	mst->ifp = ifp;
1619 	refcount_init(&mst->refcount, 1);
1620 	mst->sw = sw;
1621 	counter_u64_add(snd_tag_count, 1);
1622 }
1623 
1624 void
1625 m_snd_tag_destroy(struct m_snd_tag *mst)
1626 {
1627 	struct ifnet *ifp;
1628 
1629 	ifp = mst->ifp;
1630 	mst->sw->snd_tag_free(mst);
1631 	if_rele(ifp);
1632 	counter_u64_add(snd_tag_count, -1);
1633 }
1634 
1635 void
1636 m_rcvif_serialize(struct mbuf *m)
1637 {
1638 	u_short idx, gen;
1639 
1640 	M_ASSERTPKTHDR(m);
1641 	idx = if_getindex(m->m_pkthdr.rcvif);
1642 	gen = if_getidxgen(m->m_pkthdr.rcvif);
1643 	m->m_pkthdr.rcvidx = idx;
1644 	m->m_pkthdr.rcvgen = gen;
1645 	if (__predict_false(m->m_pkthdr.leaf_rcvif != NULL)) {
1646 		idx = if_getindex(m->m_pkthdr.leaf_rcvif);
1647 		gen = if_getidxgen(m->m_pkthdr.leaf_rcvif);
1648 	} else {
1649 		idx = -1;
1650 		gen = 0;
1651 	}
1652 	m->m_pkthdr.leaf_rcvidx = idx;
1653 	m->m_pkthdr.leaf_rcvgen = gen;
1654 }
1655 
1656 struct ifnet *
1657 m_rcvif_restore(struct mbuf *m)
1658 {
1659 	struct ifnet *ifp, *leaf_ifp;
1660 
1661 	M_ASSERTPKTHDR(m);
1662 	NET_EPOCH_ASSERT();
1663 
1664 	ifp = ifnet_byindexgen(m->m_pkthdr.rcvidx, m->m_pkthdr.rcvgen);
1665 	if (ifp == NULL || (if_getflags(ifp) & IFF_DYING))
1666 		return (NULL);
1667 
1668 	if (__predict_true(m->m_pkthdr.leaf_rcvidx == (u_short)-1)) {
1669 		leaf_ifp = NULL;
1670 	} else {
1671 		leaf_ifp = ifnet_byindexgen(m->m_pkthdr.leaf_rcvidx,
1672 		    m->m_pkthdr.leaf_rcvgen);
1673 		if (__predict_false(leaf_ifp != NULL && (if_getflags(leaf_ifp) & IFF_DYING)))
1674 			leaf_ifp = NULL;
1675 	}
1676 
1677 	m->m_pkthdr.leaf_rcvif = leaf_ifp;
1678 	m->m_pkthdr.rcvif = ifp;
1679 
1680 	return (ifp);
1681 }
1682 
1683 /*
1684  * Allocate an mbuf with anonymous external pages.
1685  */
1686 struct mbuf *
1687 mb_alloc_ext_plus_pages(int len, int how)
1688 {
1689 	struct mbuf *m;
1690 	vm_page_t pg;
1691 	int i, npgs;
1692 
1693 	m = mb_alloc_ext_pgs(how, mb_free_mext_pgs);
1694 	if (m == NULL)
1695 		return (NULL);
1696 	m->m_epg_flags |= EPG_FLAG_ANON;
1697 	npgs = howmany(len, PAGE_SIZE);
1698 	for (i = 0; i < npgs; i++) {
1699 		do {
1700 			pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
1701 			    VM_ALLOC_WIRED);
1702 			if (pg == NULL) {
1703 				if (how == M_NOWAIT) {
1704 					m->m_epg_npgs = i;
1705 					m_free(m);
1706 					return (NULL);
1707 				}
1708 				vm_wait(NULL);
1709 			}
1710 		} while (pg == NULL);
1711 		m->m_epg_pa[i] = VM_PAGE_TO_PHYS(pg);
1712 	}
1713 	m->m_epg_npgs = npgs;
1714 	return (m);
1715 }
1716 
1717 /*
1718  * Copy the data in the mbuf chain to a chain of mbufs with anonymous external
1719  * unmapped pages.
1720  * len is the length of data in the input mbuf chain.
1721  * mlen is the maximum number of bytes put into each ext_page mbuf.
1722  */
1723 struct mbuf *
1724 mb_mapped_to_unmapped(struct mbuf *mp, int len, int mlen, int how,
1725     struct mbuf **mlast)
1726 {
1727 	struct mbuf *m, *mout;
1728 	char *pgpos, *mbpos;
1729 	int i, mblen, mbufsiz, pglen, xfer;
1730 
1731 	if (len == 0)
1732 		return (NULL);
1733 	mbufsiz = min(mlen, len);
1734 	m = mout = mb_alloc_ext_plus_pages(mbufsiz, how);
1735 	if (m == NULL)
1736 		return (m);
1737 	pgpos = (char *)(void *)PHYS_TO_DMAP(m->m_epg_pa[0]);
1738 	pglen = PAGE_SIZE;
1739 	mblen = 0;
1740 	i = 0;
1741 	do {
1742 		if (pglen == 0) {
1743 			if (++i == m->m_epg_npgs) {
1744 				m->m_epg_last_len = PAGE_SIZE;
1745 				mbufsiz = min(mlen, len);
1746 				m->m_next = mb_alloc_ext_plus_pages(mbufsiz,
1747 				    how);
1748 				m = m->m_next;
1749 				if (m == NULL) {
1750 					m_freem(mout);
1751 					return (m);
1752 				}
1753 				i = 0;
1754 			}
1755 			pgpos = (char *)(void *)PHYS_TO_DMAP(m->m_epg_pa[i]);
1756 			pglen = PAGE_SIZE;
1757 		}
1758 		while (mblen == 0) {
1759 			if (mp == NULL) {
1760 				m_freem(mout);
1761 				return (NULL);
1762 			}
1763 			KASSERT((mp->m_flags & M_EXTPG) == 0,
1764 			    ("mb_copym_ext_pgs: ext_pgs input mbuf"));
1765 			mbpos = mtod(mp, char *);
1766 			mblen = mp->m_len;
1767 			mp = mp->m_next;
1768 		}
1769 		xfer = min(mblen, pglen);
1770 		memcpy(pgpos, mbpos, xfer);
1771 		pgpos += xfer;
1772 		mbpos += xfer;
1773 		pglen -= xfer;
1774 		mblen -= xfer;
1775 		len -= xfer;
1776 		m->m_len += xfer;
1777 	} while (len > 0);
1778 	m->m_epg_last_len = PAGE_SIZE - pglen;
1779 	if (mlast != NULL)
1780 		*mlast = m;
1781 	return (mout);
1782 }
1783