xref: /freebsd/sys/kern/kern_mbuf.c (revision d101cdd6edd782f6ec56eef63ed91abd77a8b317)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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 __FBSDID("$FreeBSD$");
32 
33 #include "opt_param.h"
34 #include "opt_kern_tls.h"
35 
36 #include <sys/param.h>
37 #include <sys/conf.h>
38 #include <sys/domainset.h>
39 #include <sys/malloc.h>
40 #include <sys/systm.h>
41 #include <sys/mbuf.h>
42 #include <sys/eventhandler.h>
43 #include <sys/kernel.h>
44 #include <sys/ktls.h>
45 #include <sys/limits.h>
46 #include <sys/lock.h>
47 #include <sys/mutex.h>
48 #include <sys/refcount.h>
49 #include <sys/sf_buf.h>
50 #include <sys/smp.h>
51 #include <sys/socket.h>
52 #include <sys/sysctl.h>
53 
54 #include <net/if.h>
55 #include <net/if_var.h>
56 
57 #include <vm/vm.h>
58 #include <vm/vm_extern.h>
59 #include <vm/vm_kern.h>
60 #include <vm/vm_page.h>
61 #include <vm/vm_pageout.h>
62 #include <vm/vm_map.h>
63 #include <vm/uma.h>
64 #include <vm/uma_dbg.h>
65 
66 _Static_assert(MJUMPAGESIZE > MCLBYTES,
67     "Cluster must be smaller than a jumbo page");
68 
69 /*
70  * In FreeBSD, Mbufs and Mbuf Clusters are allocated from UMA
71  * Zones.
72  *
73  * Mbuf Clusters (2K, contiguous) are allocated from the Cluster
74  * Zone.  The Zone can be capped at kern.ipc.nmbclusters, if the
75  * administrator so desires.
76  *
77  * Mbufs are allocated from a UMA Primary Zone called the Mbuf
78  * Zone.
79  *
80  * Additionally, FreeBSD provides a Packet Zone, which it
81  * configures as a Secondary Zone to the Mbuf Primary Zone,
82  * thus sharing backend Slab kegs with the Mbuf Primary Zone.
83  *
84  * Thus common-case allocations and locking are simplified:
85  *
86  *  m_clget()                m_getcl()
87  *    |                         |
88  *    |   .------------>[(Packet Cache)]    m_get(), m_gethdr()
89  *    |   |             [     Packet   ]            |
90  *  [(Cluster Cache)]   [    Secondary ]   [ (Mbuf Cache)     ]
91  *  [ Cluster Zone  ]   [     Zone     ]   [ Mbuf Primary Zone ]
92  *        |                       \________         |
93  *  [ Cluster Keg   ]                      \       /
94  *        |	                         [ Mbuf Keg   ]
95  *  [ Cluster Slabs ]                         |
96  *        |                              [ Mbuf Slabs ]
97  *         \____________(VM)_________________/
98  *
99  *
100  * Whenever an object is allocated with uma_zalloc() out of
101  * one of the Zones its _ctor_ function is executed.  The same
102  * for any deallocation through uma_zfree() the _dtor_ function
103  * is executed.
104  *
105  * Caches are per-CPU and are filled from the Primary Zone.
106  *
107  * Whenever an object is allocated from the underlying global
108  * memory pool it gets pre-initialized with the _zinit_ functions.
109  * When the Keg's are overfull objects get decommissioned with
110  * _zfini_ functions and free'd back to the global memory pool.
111  *
112  */
113 
114 int nmbufs;			/* limits number of mbufs */
115 int nmbclusters;		/* limits number of mbuf clusters */
116 int nmbjumbop;			/* limits number of page size jumbo clusters */
117 int nmbjumbo9;			/* limits number of 9k jumbo clusters */
118 int nmbjumbo16;			/* limits number of 16k jumbo clusters */
119 
120 bool mb_use_ext_pgs = false;	/* use M_EXTPG mbufs for sendfile & TLS */
121 
122 static int
123 sysctl_mb_use_ext_pgs(SYSCTL_HANDLER_ARGS)
124 {
125 	int error, extpg;
126 
127 	extpg = mb_use_ext_pgs;
128 	error = sysctl_handle_int(oidp, &extpg, 0, req);
129 	if (error == 0 && req->newptr != NULL) {
130 		if (extpg != 0 && !PMAP_HAS_DMAP)
131 			error = EOPNOTSUPP;
132 		else
133 			mb_use_ext_pgs = extpg != 0;
134 	}
135 	return (error);
136 }
137 SYSCTL_PROC(_kern_ipc, OID_AUTO, mb_use_ext_pgs, CTLTYPE_INT | CTLFLAG_RW,
138     &mb_use_ext_pgs, 0,
139     sysctl_mb_use_ext_pgs, "IU",
140     "Use unmapped mbufs for sendfile(2) and TLS offload");
141 
142 static quad_t maxmbufmem;	/* overall real memory limit for all mbufs */
143 
144 SYSCTL_QUAD(_kern_ipc, OID_AUTO, maxmbufmem, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &maxmbufmem, 0,
145     "Maximum real memory allocatable to various mbuf types");
146 
147 static counter_u64_t snd_tag_count;
148 SYSCTL_COUNTER_U64(_kern_ipc, OID_AUTO, num_snd_tags, CTLFLAG_RW,
149     &snd_tag_count, "# of active mbuf send tags");
150 
151 /*
152  * tunable_mbinit() has to be run before any mbuf allocations are done.
153  */
154 static void
155 tunable_mbinit(void *dummy)
156 {
157 	quad_t realmem;
158 	int extpg;
159 
160 	/*
161 	 * The default limit for all mbuf related memory is 1/2 of all
162 	 * available kernel memory (physical or kmem).
163 	 * At most it can be 3/4 of available kernel memory.
164 	 */
165 	realmem = qmin((quad_t)physmem * PAGE_SIZE, vm_kmem_size);
166 	maxmbufmem = realmem / 2;
167 	TUNABLE_QUAD_FETCH("kern.ipc.maxmbufmem", &maxmbufmem);
168 	if (maxmbufmem > realmem / 4 * 3)
169 		maxmbufmem = realmem / 4 * 3;
170 
171 	TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
172 	if (nmbclusters == 0)
173 		nmbclusters = maxmbufmem / MCLBYTES / 4;
174 
175 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbop", &nmbjumbop);
176 	if (nmbjumbop == 0)
177 		nmbjumbop = maxmbufmem / MJUMPAGESIZE / 4;
178 
179 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbo9", &nmbjumbo9);
180 	if (nmbjumbo9 == 0)
181 		nmbjumbo9 = maxmbufmem / MJUM9BYTES / 6;
182 
183 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbo16", &nmbjumbo16);
184 	if (nmbjumbo16 == 0)
185 		nmbjumbo16 = maxmbufmem / MJUM16BYTES / 6;
186 
187 	/*
188 	 * We need at least as many mbufs as we have clusters of
189 	 * the various types added together.
190 	 */
191 	TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
192 	if (nmbufs < nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16)
193 		nmbufs = lmax(maxmbufmem / MSIZE / 5,
194 		    nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16);
195 
196 	/*
197 	 * Unmapped mbufs can only safely be used on platforms with a direct
198 	 * map.
199 	 */
200 	if (PMAP_HAS_DMAP) {
201 		extpg = 1;
202 		TUNABLE_INT_FETCH("kern.ipc.mb_use_ext_pgs", &extpg);
203 		mb_use_ext_pgs = extpg != 0;
204 	}
205 }
206 SYSINIT(tunable_mbinit, SI_SUB_KMEM, SI_ORDER_MIDDLE, tunable_mbinit, NULL);
207 
208 static int
209 sysctl_nmbclusters(SYSCTL_HANDLER_ARGS)
210 {
211 	int error, newnmbclusters;
212 
213 	newnmbclusters = nmbclusters;
214 	error = sysctl_handle_int(oidp, &newnmbclusters, 0, req);
215 	if (error == 0 && req->newptr && newnmbclusters != nmbclusters) {
216 		if (newnmbclusters > nmbclusters &&
217 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
218 			nmbclusters = newnmbclusters;
219 			nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
220 			EVENTHANDLER_INVOKE(nmbclusters_change);
221 		} else
222 			error = EINVAL;
223 	}
224 	return (error);
225 }
226 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbclusters,
227     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, &nmbclusters, 0,
228     sysctl_nmbclusters, "IU",
229     "Maximum number of mbuf clusters allowed");
230 
231 static int
232 sysctl_nmbjumbop(SYSCTL_HANDLER_ARGS)
233 {
234 	int error, newnmbjumbop;
235 
236 	newnmbjumbop = nmbjumbop;
237 	error = sysctl_handle_int(oidp, &newnmbjumbop, 0, req);
238 	if (error == 0 && req->newptr && newnmbjumbop != nmbjumbop) {
239 		if (newnmbjumbop > nmbjumbop &&
240 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
241 			nmbjumbop = newnmbjumbop;
242 			nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
243 		} else
244 			error = EINVAL;
245 	}
246 	return (error);
247 }
248 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbop,
249     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, &nmbjumbop, 0,
250     sysctl_nmbjumbop, "IU",
251     "Maximum number of mbuf page size jumbo clusters allowed");
252 
253 static int
254 sysctl_nmbjumbo9(SYSCTL_HANDLER_ARGS)
255 {
256 	int error, newnmbjumbo9;
257 
258 	newnmbjumbo9 = nmbjumbo9;
259 	error = sysctl_handle_int(oidp, &newnmbjumbo9, 0, req);
260 	if (error == 0 && req->newptr && newnmbjumbo9 != nmbjumbo9) {
261 		if (newnmbjumbo9 > nmbjumbo9 &&
262 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
263 			nmbjumbo9 = newnmbjumbo9;
264 			nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
265 		} else
266 			error = EINVAL;
267 	}
268 	return (error);
269 }
270 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo9,
271     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, &nmbjumbo9, 0,
272     sysctl_nmbjumbo9, "IU",
273     "Maximum number of mbuf 9k jumbo clusters allowed");
274 
275 static int
276 sysctl_nmbjumbo16(SYSCTL_HANDLER_ARGS)
277 {
278 	int error, newnmbjumbo16;
279 
280 	newnmbjumbo16 = nmbjumbo16;
281 	error = sysctl_handle_int(oidp, &newnmbjumbo16, 0, req);
282 	if (error == 0 && req->newptr && newnmbjumbo16 != nmbjumbo16) {
283 		if (newnmbjumbo16 > nmbjumbo16 &&
284 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
285 			nmbjumbo16 = newnmbjumbo16;
286 			nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
287 		} else
288 			error = EINVAL;
289 	}
290 	return (error);
291 }
292 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo16,
293     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, &nmbjumbo16, 0,
294     sysctl_nmbjumbo16, "IU",
295     "Maximum number of mbuf 16k jumbo clusters allowed");
296 
297 static int
298 sysctl_nmbufs(SYSCTL_HANDLER_ARGS)
299 {
300 	int error, newnmbufs;
301 
302 	newnmbufs = nmbufs;
303 	error = sysctl_handle_int(oidp, &newnmbufs, 0, req);
304 	if (error == 0 && req->newptr && newnmbufs != nmbufs) {
305 		if (newnmbufs > nmbufs) {
306 			nmbufs = newnmbufs;
307 			nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
308 			EVENTHANDLER_INVOKE(nmbufs_change);
309 		} else
310 			error = EINVAL;
311 	}
312 	return (error);
313 }
314 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbufs,
315     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
316     &nmbufs, 0, sysctl_nmbufs, "IU",
317     "Maximum number of mbufs allowed");
318 
319 /*
320  * Zones from which we allocate.
321  */
322 uma_zone_t	zone_mbuf;
323 uma_zone_t	zone_clust;
324 uma_zone_t	zone_pack;
325 uma_zone_t	zone_jumbop;
326 uma_zone_t	zone_jumbo9;
327 uma_zone_t	zone_jumbo16;
328 
329 /*
330  * Local prototypes.
331  */
332 static int	mb_ctor_mbuf(void *, int, void *, int);
333 static int	mb_ctor_clust(void *, int, void *, int);
334 static int	mb_ctor_pack(void *, int, void *, int);
335 static void	mb_dtor_mbuf(void *, int, void *);
336 static void	mb_dtor_pack(void *, int, void *);
337 static int	mb_zinit_pack(void *, int, int);
338 static void	mb_zfini_pack(void *, int);
339 static void	mb_reclaim(uma_zone_t, int);
340 
341 /* Ensure that MSIZE is a power of 2. */
342 CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE);
343 
344 _Static_assert(sizeof(struct mbuf) <= MSIZE,
345     "size of mbuf exceeds MSIZE");
346 /*
347  * Initialize FreeBSD Network buffer allocation.
348  */
349 static void
350 mbuf_init(void *dummy)
351 {
352 
353 	/*
354 	 * Configure UMA zones for Mbufs, Clusters, and Packets.
355 	 */
356 	zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE,
357 	    mb_ctor_mbuf, mb_dtor_mbuf, NULL, NULL,
358 	    MSIZE - 1, UMA_ZONE_CONTIG | UMA_ZONE_MAXBUCKET);
359 	if (nmbufs > 0)
360 		nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
361 	uma_zone_set_warning(zone_mbuf, "kern.ipc.nmbufs limit reached");
362 	uma_zone_set_maxaction(zone_mbuf, mb_reclaim);
363 
364 	zone_clust = uma_zcreate(MBUF_CLUSTER_MEM_NAME, MCLBYTES,
365 	    mb_ctor_clust, NULL, NULL, NULL,
366 	    UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
367 	if (nmbclusters > 0)
368 		nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
369 	uma_zone_set_warning(zone_clust, "kern.ipc.nmbclusters limit reached");
370 	uma_zone_set_maxaction(zone_clust, mb_reclaim);
371 
372 	zone_pack = uma_zsecond_create(MBUF_PACKET_MEM_NAME, mb_ctor_pack,
373 	    mb_dtor_pack, mb_zinit_pack, mb_zfini_pack, zone_mbuf);
374 
375 	/* Make jumbo frame zone too. Page size, 9k and 16k. */
376 	zone_jumbop = uma_zcreate(MBUF_JUMBOP_MEM_NAME, MJUMPAGESIZE,
377 	    mb_ctor_clust, NULL, NULL, NULL,
378 	    UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
379 	if (nmbjumbop > 0)
380 		nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
381 	uma_zone_set_warning(zone_jumbop, "kern.ipc.nmbjumbop limit reached");
382 	uma_zone_set_maxaction(zone_jumbop, mb_reclaim);
383 
384 	zone_jumbo9 = uma_zcreate(MBUF_JUMBO9_MEM_NAME, MJUM9BYTES,
385 	    mb_ctor_clust, NULL, NULL, NULL,
386 	    UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
387 	if (nmbjumbo9 > 0)
388 		nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
389 	uma_zone_set_warning(zone_jumbo9, "kern.ipc.nmbjumbo9 limit reached");
390 	uma_zone_set_maxaction(zone_jumbo9, mb_reclaim);
391 
392 	zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES,
393 	    mb_ctor_clust, NULL, NULL, NULL,
394 	    UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
395 	if (nmbjumbo16 > 0)
396 		nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
397 	uma_zone_set_warning(zone_jumbo16, "kern.ipc.nmbjumbo16 limit reached");
398 	uma_zone_set_maxaction(zone_jumbo16, mb_reclaim);
399 
400 	snd_tag_count = counter_u64_alloc(M_WAITOK);
401 }
402 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL);
403 
404 #ifdef DEBUGNET
405 /*
406  * debugnet makes use of a pre-allocated pool of mbufs and clusters.  When
407  * debugnet is configured, we initialize a set of UMA cache zones which return
408  * items from this pool.  At panic-time, the regular UMA zone pointers are
409  * overwritten with those of the cache zones so that drivers may allocate and
410  * free mbufs and clusters without attempting to allocate physical memory.
411  *
412  * We keep mbufs and clusters in a pair of mbuf queues.  In particular, for
413  * the purpose of caching clusters, we treat them as mbufs.
414  */
415 static struct mbufq dn_mbufq =
416     { STAILQ_HEAD_INITIALIZER(dn_mbufq.mq_head), 0, INT_MAX };
417 static struct mbufq dn_clustq =
418     { STAILQ_HEAD_INITIALIZER(dn_clustq.mq_head), 0, INT_MAX };
419 
420 static int dn_clsize;
421 static uma_zone_t dn_zone_mbuf;
422 static uma_zone_t dn_zone_clust;
423 static uma_zone_t dn_zone_pack;
424 
425 static struct debugnet_saved_zones {
426 	uma_zone_t dsz_mbuf;
427 	uma_zone_t dsz_clust;
428 	uma_zone_t dsz_pack;
429 	uma_zone_t dsz_jumbop;
430 	uma_zone_t dsz_jumbo9;
431 	uma_zone_t dsz_jumbo16;
432 	bool dsz_debugnet_zones_enabled;
433 } dn_saved_zones;
434 
435 static int
436 dn_buf_import(void *arg, void **store, int count, int domain __unused,
437     int flags)
438 {
439 	struct mbufq *q;
440 	struct mbuf *m;
441 	int i;
442 
443 	q = arg;
444 
445 	for (i = 0; i < count; i++) {
446 		m = mbufq_dequeue(q);
447 		if (m == NULL)
448 			break;
449 		trash_init(m, q == &dn_mbufq ? MSIZE : dn_clsize, flags);
450 		store[i] = m;
451 	}
452 	KASSERT((flags & M_WAITOK) == 0 || i == count,
453 	    ("%s: ran out of pre-allocated mbufs", __func__));
454 	return (i);
455 }
456 
457 static void
458 dn_buf_release(void *arg, void **store, int count)
459 {
460 	struct mbufq *q;
461 	struct mbuf *m;
462 	int i;
463 
464 	q = arg;
465 
466 	for (i = 0; i < count; i++) {
467 		m = store[i];
468 		(void)mbufq_enqueue(q, m);
469 	}
470 }
471 
472 static int
473 dn_pack_import(void *arg __unused, void **store, int count, int domain __unused,
474     int flags __unused)
475 {
476 	struct mbuf *m;
477 	void *clust;
478 	int i;
479 
480 	for (i = 0; i < count; i++) {
481 		m = m_get(M_NOWAIT, MT_DATA);
482 		if (m == NULL)
483 			break;
484 		clust = uma_zalloc(dn_zone_clust, M_NOWAIT);
485 		if (clust == NULL) {
486 			m_free(m);
487 			break;
488 		}
489 		mb_ctor_clust(clust, dn_clsize, m, 0);
490 		store[i] = m;
491 	}
492 	KASSERT((flags & M_WAITOK) == 0 || i == count,
493 	    ("%s: ran out of pre-allocated mbufs", __func__));
494 	return (i);
495 }
496 
497 static void
498 dn_pack_release(void *arg __unused, void **store, int count)
499 {
500 	struct mbuf *m;
501 	void *clust;
502 	int i;
503 
504 	for (i = 0; i < count; i++) {
505 		m = store[i];
506 		clust = m->m_ext.ext_buf;
507 		uma_zfree(dn_zone_clust, clust);
508 		uma_zfree(dn_zone_mbuf, m);
509 	}
510 }
511 
512 /*
513  * Free the pre-allocated mbufs and clusters reserved for debugnet, and destroy
514  * the corresponding UMA cache zones.
515  */
516 void
517 debugnet_mbuf_drain(void)
518 {
519 	struct mbuf *m;
520 	void *item;
521 
522 	if (dn_zone_mbuf != NULL) {
523 		uma_zdestroy(dn_zone_mbuf);
524 		dn_zone_mbuf = NULL;
525 	}
526 	if (dn_zone_clust != NULL) {
527 		uma_zdestroy(dn_zone_clust);
528 		dn_zone_clust = NULL;
529 	}
530 	if (dn_zone_pack != NULL) {
531 		uma_zdestroy(dn_zone_pack);
532 		dn_zone_pack = NULL;
533 	}
534 
535 	while ((m = mbufq_dequeue(&dn_mbufq)) != NULL)
536 		m_free(m);
537 	while ((item = mbufq_dequeue(&dn_clustq)) != NULL)
538 		uma_zfree(m_getzone(dn_clsize), item);
539 }
540 
541 /*
542  * Callback invoked immediately prior to starting a debugnet connection.
543  */
544 void
545 debugnet_mbuf_start(void)
546 {
547 
548 	MPASS(!dn_saved_zones.dsz_debugnet_zones_enabled);
549 
550 	/* Save the old zone pointers to restore when debugnet is closed. */
551 	dn_saved_zones = (struct debugnet_saved_zones) {
552 		.dsz_debugnet_zones_enabled = true,
553 		.dsz_mbuf = zone_mbuf,
554 		.dsz_clust = zone_clust,
555 		.dsz_pack = zone_pack,
556 		.dsz_jumbop = zone_jumbop,
557 		.dsz_jumbo9 = zone_jumbo9,
558 		.dsz_jumbo16 = zone_jumbo16,
559 	};
560 
561 	/*
562 	 * All cluster zones return buffers of the size requested by the
563 	 * drivers.  It's up to the driver to reinitialize the zones if the
564 	 * MTU of a debugnet-enabled interface changes.
565 	 */
566 	printf("debugnet: overwriting mbuf zone pointers\n");
567 	zone_mbuf = dn_zone_mbuf;
568 	zone_clust = dn_zone_clust;
569 	zone_pack = dn_zone_pack;
570 	zone_jumbop = dn_zone_clust;
571 	zone_jumbo9 = dn_zone_clust;
572 	zone_jumbo16 = dn_zone_clust;
573 }
574 
575 /*
576  * Callback invoked when a debugnet connection is closed/finished.
577  */
578 void
579 debugnet_mbuf_finish(void)
580 {
581 
582 	MPASS(dn_saved_zones.dsz_debugnet_zones_enabled);
583 
584 	printf("debugnet: restoring mbuf zone pointers\n");
585 	zone_mbuf = dn_saved_zones.dsz_mbuf;
586 	zone_clust = dn_saved_zones.dsz_clust;
587 	zone_pack = dn_saved_zones.dsz_pack;
588 	zone_jumbop = dn_saved_zones.dsz_jumbop;
589 	zone_jumbo9 = dn_saved_zones.dsz_jumbo9;
590 	zone_jumbo16 = dn_saved_zones.dsz_jumbo16;
591 
592 	memset(&dn_saved_zones, 0, sizeof(dn_saved_zones));
593 }
594 
595 /*
596  * Reinitialize the debugnet mbuf+cluster pool and cache zones.
597  */
598 void
599 debugnet_mbuf_reinit(int nmbuf, int nclust, int clsize)
600 {
601 	struct mbuf *m;
602 	void *item;
603 
604 	debugnet_mbuf_drain();
605 
606 	dn_clsize = clsize;
607 
608 	dn_zone_mbuf = uma_zcache_create("debugnet_" MBUF_MEM_NAME,
609 	    MSIZE, mb_ctor_mbuf, mb_dtor_mbuf, NULL, NULL,
610 	    dn_buf_import, dn_buf_release,
611 	    &dn_mbufq, UMA_ZONE_NOBUCKET);
612 
613 	dn_zone_clust = uma_zcache_create("debugnet_" MBUF_CLUSTER_MEM_NAME,
614 	    clsize, mb_ctor_clust, NULL, NULL, NULL,
615 	    dn_buf_import, dn_buf_release,
616 	    &dn_clustq, UMA_ZONE_NOBUCKET);
617 
618 	dn_zone_pack = uma_zcache_create("debugnet_" MBUF_PACKET_MEM_NAME,
619 	    MCLBYTES, mb_ctor_pack, mb_dtor_pack, NULL, NULL,
620 	    dn_pack_import, dn_pack_release,
621 	    NULL, UMA_ZONE_NOBUCKET);
622 
623 	while (nmbuf-- > 0) {
624 		m = m_get(M_WAITOK, MT_DATA);
625 		uma_zfree(dn_zone_mbuf, m);
626 	}
627 	while (nclust-- > 0) {
628 		item = uma_zalloc(m_getzone(dn_clsize), M_WAITOK);
629 		uma_zfree(dn_zone_clust, item);
630 	}
631 }
632 #endif /* DEBUGNET */
633 
634 /*
635  * Constructor for Mbuf primary zone.
636  *
637  * The 'arg' pointer points to a mb_args structure which
638  * contains call-specific information required to support the
639  * mbuf allocation API.  See mbuf.h.
640  */
641 static int
642 mb_ctor_mbuf(void *mem, int size, void *arg, int how)
643 {
644 	struct mbuf *m;
645 	struct mb_args *args;
646 	int error;
647 	int flags;
648 	short type;
649 
650 	args = (struct mb_args *)arg;
651 	type = args->type;
652 
653 	/*
654 	 * The mbuf is initialized later.  The caller has the
655 	 * responsibility to set up any MAC labels too.
656 	 */
657 	if (type == MT_NOINIT)
658 		return (0);
659 
660 	m = (struct mbuf *)mem;
661 	flags = args->flags;
662 	MPASS((flags & M_NOFREE) == 0);
663 
664 	error = m_init(m, how, type, flags);
665 
666 	return (error);
667 }
668 
669 /*
670  * The Mbuf primary zone destructor.
671  */
672 static void
673 mb_dtor_mbuf(void *mem, int size, void *arg)
674 {
675 	struct mbuf *m;
676 	unsigned long flags __diagused;
677 
678 	m = (struct mbuf *)mem;
679 	flags = (unsigned long)arg;
680 
681 	KASSERT((m->m_flags & M_NOFREE) == 0, ("%s: M_NOFREE set", __func__));
682 	KASSERT((flags & 0x1) == 0, ("%s: obsolete MB_DTOR_SKIP passed", __func__));
683 	if ((m->m_flags & M_PKTHDR) && !SLIST_EMPTY(&m->m_pkthdr.tags))
684 		m_tag_delete_chain(m, NULL);
685 }
686 
687 /*
688  * The Mbuf Packet zone destructor.
689  */
690 static void
691 mb_dtor_pack(void *mem, int size, void *arg)
692 {
693 	struct mbuf *m;
694 
695 	m = (struct mbuf *)mem;
696 	if ((m->m_flags & M_PKTHDR) != 0)
697 		m_tag_delete_chain(m, NULL);
698 
699 	/* Make sure we've got a clean cluster back. */
700 	KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__));
701 	KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__));
702 	KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__));
703 	KASSERT(m->m_ext.ext_arg1 == NULL, ("%s: ext_arg1 != NULL", __func__));
704 	KASSERT(m->m_ext.ext_arg2 == NULL, ("%s: ext_arg2 != NULL", __func__));
705 	KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__));
706 	KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__));
707 #if defined(INVARIANTS) && !defined(KMSAN)
708 	trash_dtor(m->m_ext.ext_buf, MCLBYTES, arg);
709 #endif
710 	/*
711 	 * If there are processes blocked on zone_clust, waiting for pages
712 	 * to be freed up, cause them to be woken up by draining the
713 	 * packet zone.  We are exposed to a race here (in the check for
714 	 * the UMA_ZFLAG_FULL) where we might miss the flag set, but that
715 	 * is deliberate. We don't want to acquire the zone lock for every
716 	 * mbuf free.
717 	 */
718 	if (uma_zone_exhausted(zone_clust))
719 		uma_zone_reclaim(zone_pack, UMA_RECLAIM_DRAIN);
720 }
721 
722 /*
723  * The Cluster and Jumbo[PAGESIZE|9|16] zone constructor.
724  *
725  * Here the 'arg' pointer points to the Mbuf which we
726  * are configuring cluster storage for.  If 'arg' is
727  * empty we allocate just the cluster without setting
728  * the mbuf to it.  See mbuf.h.
729  */
730 static int
731 mb_ctor_clust(void *mem, int size, void *arg, int how)
732 {
733 	struct mbuf *m;
734 
735 	m = (struct mbuf *)arg;
736 	if (m != NULL) {
737 		m->m_ext.ext_buf = (char *)mem;
738 		m->m_data = m->m_ext.ext_buf;
739 		m->m_flags |= M_EXT;
740 		m->m_ext.ext_free = NULL;
741 		m->m_ext.ext_arg1 = NULL;
742 		m->m_ext.ext_arg2 = NULL;
743 		m->m_ext.ext_size = size;
744 		m->m_ext.ext_type = m_gettype(size);
745 		m->m_ext.ext_flags = EXT_FLAG_EMBREF;
746 		m->m_ext.ext_count = 1;
747 	}
748 
749 	return (0);
750 }
751 
752 /*
753  * The Packet secondary zone's init routine, executed on the
754  * object's transition from mbuf keg slab to zone cache.
755  */
756 static int
757 mb_zinit_pack(void *mem, int size, int how)
758 {
759 	struct mbuf *m;
760 
761 	m = (struct mbuf *)mem;		/* m is virgin. */
762 	if (uma_zalloc_arg(zone_clust, m, how) == NULL ||
763 	    m->m_ext.ext_buf == NULL)
764 		return (ENOMEM);
765 	m->m_ext.ext_type = EXT_PACKET;	/* Override. */
766 #if defined(INVARIANTS) && !defined(KMSAN)
767 	trash_init(m->m_ext.ext_buf, MCLBYTES, how);
768 #endif
769 	return (0);
770 }
771 
772 /*
773  * The Packet secondary zone's fini routine, executed on the
774  * object's transition from zone cache to keg slab.
775  */
776 static void
777 mb_zfini_pack(void *mem, int size)
778 {
779 	struct mbuf *m;
780 
781 	m = (struct mbuf *)mem;
782 #if defined(INVARIANTS) && !defined(KMSAN)
783 	trash_fini(m->m_ext.ext_buf, MCLBYTES);
784 #endif
785 	uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL);
786 #if defined(INVARIANTS) && !defined(KMSAN)
787 	trash_dtor(mem, size, NULL);
788 #endif
789 }
790 
791 /*
792  * The "packet" keg constructor.
793  */
794 static int
795 mb_ctor_pack(void *mem, int size, void *arg, int how)
796 {
797 	struct mbuf *m;
798 	struct mb_args *args;
799 	int error, flags;
800 	short type;
801 
802 	m = (struct mbuf *)mem;
803 	args = (struct mb_args *)arg;
804 	flags = args->flags;
805 	type = args->type;
806 	MPASS((flags & M_NOFREE) == 0);
807 
808 #if defined(INVARIANTS) && !defined(KMSAN)
809 	trash_ctor(m->m_ext.ext_buf, MCLBYTES, arg, how);
810 #endif
811 
812 	error = m_init(m, how, type, flags);
813 
814 	/* m_ext is already initialized. */
815 	m->m_data = m->m_ext.ext_buf;
816  	m->m_flags = (flags | M_EXT);
817 
818 	return (error);
819 }
820 
821 /*
822  * This is the protocol drain routine.  Called by UMA whenever any of the
823  * mbuf zones is closed to its limit.
824  */
825 static void
826 mb_reclaim(uma_zone_t zone __unused, int pending __unused)
827 {
828 
829 	EVENTHANDLER_INVOKE(mbuf_lowmem, VM_LOW_MBUFS);
830 }
831 
832 /*
833  * Free "count" units of I/O from an mbuf chain.  They could be held
834  * in M_EXTPG or just as a normal mbuf.  This code is intended to be
835  * called in an error path (I/O error, closed connection, etc).
836  */
837 void
838 mb_free_notready(struct mbuf *m, int count)
839 {
840 	int i;
841 
842 	for (i = 0; i < count && m != NULL; i++) {
843 		if ((m->m_flags & M_EXTPG) != 0) {
844 			m->m_epg_nrdy--;
845 			if (m->m_epg_nrdy != 0)
846 				continue;
847 		}
848 		m = m_free(m);
849 	}
850 	KASSERT(i == count, ("Removed only %d items from %p", i, m));
851 }
852 
853 /*
854  * Compress an unmapped mbuf into a simple mbuf when it holds a small
855  * amount of data.  This is used as a DOS defense to avoid having
856  * small packets tie up wired pages, an ext_pgs structure, and an
857  * mbuf.  Since this converts the existing mbuf in place, it can only
858  * be used if there are no other references to 'm'.
859  */
860 int
861 mb_unmapped_compress(struct mbuf *m)
862 {
863 	volatile u_int *refcnt;
864 	char buf[MLEN];
865 
866 	/*
867 	 * Assert that 'm' does not have a packet header.  If 'm' had
868 	 * a packet header, it would only be able to hold MHLEN bytes
869 	 * and m_data would have to be initialized differently.
870 	 */
871 	KASSERT((m->m_flags & M_PKTHDR) == 0 && (m->m_flags & M_EXTPG),
872             ("%s: m %p !M_EXTPG or M_PKTHDR", __func__, m));
873 	KASSERT(m->m_len <= MLEN, ("m_len too large %p", m));
874 
875 	if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
876 		refcnt = &m->m_ext.ext_count;
877 	} else {
878 		KASSERT(m->m_ext.ext_cnt != NULL,
879 		    ("%s: no refcounting pointer on %p", __func__, m));
880 		refcnt = m->m_ext.ext_cnt;
881 	}
882 
883 	if (*refcnt != 1)
884 		return (EBUSY);
885 
886 	m_copydata(m, 0, m->m_len, buf);
887 
888 	/* Free the backing pages. */
889 	m->m_ext.ext_free(m);
890 
891 	/* Turn 'm' into a "normal" mbuf. */
892 	m->m_flags &= ~(M_EXT | M_RDONLY | M_EXTPG);
893 	m->m_data = m->m_dat;
894 
895 	/* Copy data back into m. */
896 	bcopy(buf, mtod(m, char *), m->m_len);
897 
898 	return (0);
899 }
900 
901 /*
902  * These next few routines are used to permit downgrading an unmapped
903  * mbuf to a chain of mapped mbufs.  This is used when an interface
904  * doesn't supported unmapped mbufs or if checksums need to be
905  * computed in software.
906  *
907  * Each unmapped mbuf is converted to a chain of mbufs.  First, any
908  * TLS header data is stored in a regular mbuf.  Second, each page of
909  * unmapped data is stored in an mbuf with an EXT_SFBUF external
910  * cluster.  These mbufs use an sf_buf to provide a valid KVA for the
911  * associated physical page.  They also hold a reference on the
912  * original M_EXTPG mbuf to ensure the physical page doesn't go away.
913  * Finally, any TLS trailer data is stored in a regular mbuf.
914  *
915  * mb_unmapped_free_mext() is the ext_free handler for the EXT_SFBUF
916  * mbufs.  It frees the associated sf_buf and releases its reference
917  * on the original M_EXTPG mbuf.
918  *
919  * _mb_unmapped_to_ext() is a helper function that converts a single
920  * unmapped mbuf into a chain of mbufs.
921  *
922  * mb_unmapped_to_ext() is the public function that walks an mbuf
923  * chain converting any unmapped mbufs to mapped mbufs.  It returns
924  * the new chain of unmapped mbufs on success.  On failure it frees
925  * the original mbuf chain and returns NULL.
926  */
927 static void
928 mb_unmapped_free_mext(struct mbuf *m)
929 {
930 	struct sf_buf *sf;
931 	struct mbuf *old_m;
932 
933 	sf = m->m_ext.ext_arg1;
934 	sf_buf_free(sf);
935 
936 	/* Drop the reference on the backing M_EXTPG mbuf. */
937 	old_m = m->m_ext.ext_arg2;
938 	mb_free_extpg(old_m);
939 }
940 
941 static struct mbuf *
942 _mb_unmapped_to_ext(struct mbuf *m)
943 {
944 	struct mbuf *m_new, *top, *prev, *mref;
945 	struct sf_buf *sf;
946 	vm_page_t pg;
947 	int i, len, off, pglen, pgoff, seglen, segoff;
948 	volatile u_int *refcnt;
949 	u_int ref_inc = 0;
950 
951 	M_ASSERTEXTPG(m);
952 	len = m->m_len;
953 	KASSERT(m->m_epg_tls == NULL, ("%s: can't convert TLS mbuf %p",
954 	    __func__, m));
955 
956 	/* See if this is the mbuf that holds the embedded refcount. */
957 	if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
958 		refcnt = &m->m_ext.ext_count;
959 		mref = m;
960 	} else {
961 		KASSERT(m->m_ext.ext_cnt != NULL,
962 		    ("%s: no refcounting pointer on %p", __func__, m));
963 		refcnt = m->m_ext.ext_cnt;
964 		mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
965 	}
966 
967 	/* Skip over any data removed from the front. */
968 	off = mtod(m, vm_offset_t);
969 
970 	top = NULL;
971 	if (m->m_epg_hdrlen != 0) {
972 		if (off >= m->m_epg_hdrlen) {
973 			off -= m->m_epg_hdrlen;
974 		} else {
975 			seglen = m->m_epg_hdrlen - off;
976 			segoff = off;
977 			seglen = min(seglen, len);
978 			off = 0;
979 			len -= seglen;
980 			m_new = m_get(M_NOWAIT, MT_DATA);
981 			if (m_new == NULL)
982 				goto fail;
983 			m_new->m_len = seglen;
984 			prev = top = m_new;
985 			memcpy(mtod(m_new, void *), &m->m_epg_hdr[segoff],
986 			    seglen);
987 		}
988 	}
989 	pgoff = m->m_epg_1st_off;
990 	for (i = 0; i < m->m_epg_npgs && len > 0; i++) {
991 		pglen = m_epg_pagelen(m, i, pgoff);
992 		if (off >= pglen) {
993 			off -= pglen;
994 			pgoff = 0;
995 			continue;
996 		}
997 		seglen = pglen - off;
998 		segoff = pgoff + off;
999 		off = 0;
1000 		seglen = min(seglen, len);
1001 		len -= seglen;
1002 
1003 		pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]);
1004 		m_new = m_get(M_NOWAIT, MT_DATA);
1005 		if (m_new == NULL)
1006 			goto fail;
1007 		if (top == NULL) {
1008 			top = prev = m_new;
1009 		} else {
1010 			prev->m_next = m_new;
1011 			prev = m_new;
1012 		}
1013 		sf = sf_buf_alloc(pg, SFB_NOWAIT);
1014 		if (sf == NULL)
1015 			goto fail;
1016 
1017 		ref_inc++;
1018 		m_extadd(m_new, (char *)sf_buf_kva(sf), PAGE_SIZE,
1019 		    mb_unmapped_free_mext, sf, mref, M_RDONLY, EXT_SFBUF);
1020 		m_new->m_data += segoff;
1021 		m_new->m_len = seglen;
1022 
1023 		pgoff = 0;
1024 	};
1025 	if (len != 0) {
1026 		KASSERT((off + len) <= m->m_epg_trllen,
1027 		    ("off + len > trail (%d + %d > %d)", off, len,
1028 		    m->m_epg_trllen));
1029 		m_new = m_get(M_NOWAIT, MT_DATA);
1030 		if (m_new == NULL)
1031 			goto fail;
1032 		if (top == NULL)
1033 			top = m_new;
1034 		else
1035 			prev->m_next = m_new;
1036 		m_new->m_len = len;
1037 		memcpy(mtod(m_new, void *), &m->m_epg_trail[off], len);
1038 	}
1039 
1040 	if (ref_inc != 0) {
1041 		/*
1042 		 * Obtain an additional reference on the old mbuf for
1043 		 * each created EXT_SFBUF mbuf.  They will be dropped
1044 		 * in mb_unmapped_free_mext().
1045 		 */
1046 		if (*refcnt == 1)
1047 			*refcnt += ref_inc;
1048 		else
1049 			atomic_add_int(refcnt, ref_inc);
1050 	}
1051 	m_free(m);
1052 	return (top);
1053 
1054 fail:
1055 	if (ref_inc != 0) {
1056 		/*
1057 		 * Obtain an additional reference on the old mbuf for
1058 		 * each created EXT_SFBUF mbuf.  They will be
1059 		 * immediately dropped when these mbufs are freed
1060 		 * below.
1061 		 */
1062 		if (*refcnt == 1)
1063 			*refcnt += ref_inc;
1064 		else
1065 			atomic_add_int(refcnt, ref_inc);
1066 	}
1067 	m_free(m);
1068 	m_freem(top);
1069 	return (NULL);
1070 }
1071 
1072 struct mbuf *
1073 mb_unmapped_to_ext(struct mbuf *top)
1074 {
1075 	struct mbuf *m, *next, *prev = NULL;
1076 
1077 	prev = NULL;
1078 	for (m = top; m != NULL; m = next) {
1079 		/* m might be freed, so cache the next pointer. */
1080 		next = m->m_next;
1081 		if (m->m_flags & M_EXTPG) {
1082 			if (prev != NULL) {
1083 				/*
1084 				 * Remove 'm' from the new chain so
1085 				 * that the 'top' chain terminates
1086 				 * before 'm' in case 'top' is freed
1087 				 * due to an error.
1088 				 */
1089 				prev->m_next = NULL;
1090 			}
1091 			m = _mb_unmapped_to_ext(m);
1092 			if (m == NULL) {
1093 				m_freem(top);
1094 				m_freem(next);
1095 				return (NULL);
1096 			}
1097 			if (prev == NULL) {
1098 				top = m;
1099 			} else {
1100 				prev->m_next = m;
1101 			}
1102 
1103 			/*
1104 			 * Replaced one mbuf with a chain, so we must
1105 			 * find the end of chain.
1106 			 */
1107 			prev = m_last(m);
1108 		} else {
1109 			if (prev != NULL) {
1110 				prev->m_next = m;
1111 			}
1112 			prev = m;
1113 		}
1114 	}
1115 	return (top);
1116 }
1117 
1118 /*
1119  * Allocate an empty M_EXTPG mbuf.  The ext_free routine is
1120  * responsible for freeing any pages backing this mbuf when it is
1121  * freed.
1122  */
1123 struct mbuf *
1124 mb_alloc_ext_pgs(int how, m_ext_free_t ext_free)
1125 {
1126 	struct mbuf *m;
1127 
1128 	m = m_get(how, MT_DATA);
1129 	if (m == NULL)
1130 		return (NULL);
1131 
1132 	m->m_epg_npgs = 0;
1133 	m->m_epg_nrdy = 0;
1134 	m->m_epg_1st_off = 0;
1135 	m->m_epg_last_len = 0;
1136 	m->m_epg_flags = 0;
1137 	m->m_epg_hdrlen = 0;
1138 	m->m_epg_trllen = 0;
1139 	m->m_epg_tls = NULL;
1140 	m->m_epg_so = NULL;
1141 	m->m_data = NULL;
1142 	m->m_flags |= (M_EXT | M_RDONLY | M_EXTPG);
1143 	m->m_ext.ext_flags = EXT_FLAG_EMBREF;
1144 	m->m_ext.ext_count = 1;
1145 	m->m_ext.ext_size = 0;
1146 	m->m_ext.ext_free = ext_free;
1147 	return (m);
1148 }
1149 
1150 /*
1151  * Clean up after mbufs with M_EXT storage attached to them if the
1152  * reference count hits 1.
1153  */
1154 void
1155 mb_free_ext(struct mbuf *m)
1156 {
1157 	volatile u_int *refcnt;
1158 	struct mbuf *mref;
1159 	int freembuf;
1160 
1161 	KASSERT(m->m_flags & M_EXT, ("%s: M_EXT not set on %p", __func__, m));
1162 
1163 	/* See if this is the mbuf that holds the embedded refcount. */
1164 	if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
1165 		refcnt = &m->m_ext.ext_count;
1166 		mref = m;
1167 	} else {
1168 		KASSERT(m->m_ext.ext_cnt != NULL,
1169 		    ("%s: no refcounting pointer on %p", __func__, m));
1170 		refcnt = m->m_ext.ext_cnt;
1171 		mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
1172 	}
1173 
1174 	/*
1175 	 * Check if the header is embedded in the cluster.  It is
1176 	 * important that we can't touch any of the mbuf fields
1177 	 * after we have freed the external storage, since mbuf
1178 	 * could have been embedded in it.  For now, the mbufs
1179 	 * embedded into the cluster are always of type EXT_EXTREF,
1180 	 * and for this type we won't free the mref.
1181 	 */
1182 	if (m->m_flags & M_NOFREE) {
1183 		freembuf = 0;
1184 		KASSERT(m->m_ext.ext_type == EXT_EXTREF ||
1185 		    m->m_ext.ext_type == EXT_RXRING,
1186 		    ("%s: no-free mbuf %p has wrong type", __func__, m));
1187 	} else
1188 		freembuf = 1;
1189 
1190 	/* Free attached storage if this mbuf is the only reference to it. */
1191 	if (*refcnt == 1 || atomic_fetchadd_int(refcnt, -1) == 1) {
1192 		switch (m->m_ext.ext_type) {
1193 		case EXT_PACKET:
1194 			/* The packet zone is special. */
1195 			if (*refcnt == 0)
1196 				*refcnt = 1;
1197 			uma_zfree(zone_pack, mref);
1198 			break;
1199 		case EXT_CLUSTER:
1200 			uma_zfree(zone_clust, m->m_ext.ext_buf);
1201 			m_free_raw(mref);
1202 			break;
1203 		case EXT_JUMBOP:
1204 			uma_zfree(zone_jumbop, m->m_ext.ext_buf);
1205 			m_free_raw(mref);
1206 			break;
1207 		case EXT_JUMBO9:
1208 			uma_zfree(zone_jumbo9, m->m_ext.ext_buf);
1209 			m_free_raw(mref);
1210 			break;
1211 		case EXT_JUMBO16:
1212 			uma_zfree(zone_jumbo16, m->m_ext.ext_buf);
1213 			m_free_raw(mref);
1214 			break;
1215 		case EXT_SFBUF:
1216 		case EXT_NET_DRV:
1217 		case EXT_MOD_TYPE:
1218 		case EXT_DISPOSABLE:
1219 			KASSERT(mref->m_ext.ext_free != NULL,
1220 			    ("%s: ext_free not set", __func__));
1221 			mref->m_ext.ext_free(mref);
1222 			m_free_raw(mref);
1223 			break;
1224 		case EXT_EXTREF:
1225 			KASSERT(m->m_ext.ext_free != NULL,
1226 			    ("%s: ext_free not set", __func__));
1227 			m->m_ext.ext_free(m);
1228 			break;
1229 		case EXT_RXRING:
1230 			KASSERT(m->m_ext.ext_free == NULL,
1231 			    ("%s: ext_free is set", __func__));
1232 			break;
1233 		default:
1234 			KASSERT(m->m_ext.ext_type == 0,
1235 			    ("%s: unknown ext_type", __func__));
1236 		}
1237 	}
1238 
1239 	if (freembuf && m != mref)
1240 		m_free_raw(m);
1241 }
1242 
1243 /*
1244  * Clean up after mbufs with M_EXTPG storage attached to them if the
1245  * reference count hits 1.
1246  */
1247 void
1248 mb_free_extpg(struct mbuf *m)
1249 {
1250 	volatile u_int *refcnt;
1251 	struct mbuf *mref;
1252 
1253 	M_ASSERTEXTPG(m);
1254 
1255 	/* See if this is the mbuf that holds the embedded refcount. */
1256 	if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
1257 		refcnt = &m->m_ext.ext_count;
1258 		mref = m;
1259 	} else {
1260 		KASSERT(m->m_ext.ext_cnt != NULL,
1261 		    ("%s: no refcounting pointer on %p", __func__, m));
1262 		refcnt = m->m_ext.ext_cnt;
1263 		mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
1264 	}
1265 
1266 	/* Free attached storage if this mbuf is the only reference to it. */
1267 	if (*refcnt == 1 || atomic_fetchadd_int(refcnt, -1) == 1) {
1268 		KASSERT(mref->m_ext.ext_free != NULL,
1269 		    ("%s: ext_free not set", __func__));
1270 
1271 		mref->m_ext.ext_free(mref);
1272 #ifdef KERN_TLS
1273 		if (mref->m_epg_tls != NULL &&
1274 		    !refcount_release_if_not_last(&mref->m_epg_tls->refcount))
1275 			ktls_enqueue_to_free(mref);
1276 		else
1277 #endif
1278 			m_free_raw(mref);
1279 	}
1280 
1281 	if (m != mref)
1282 		m_free_raw(m);
1283 }
1284 
1285 /*
1286  * Official mbuf(9) allocation KPI for stack and drivers:
1287  *
1288  * m_get()	- a single mbuf without any attachments, sys/mbuf.h.
1289  * m_gethdr()	- a single mbuf initialized as M_PKTHDR, sys/mbuf.h.
1290  * m_getcl()	- an mbuf + 2k cluster, sys/mbuf.h.
1291  * m_clget()	- attach cluster to already allocated mbuf.
1292  * m_cljget()	- attach jumbo cluster to already allocated mbuf.
1293  * m_get2()	- allocate minimum mbuf that would fit size argument.
1294  * m_getm2()	- allocate a chain of mbufs/clusters.
1295  * m_extadd()	- attach external cluster to mbuf.
1296  *
1297  * m_free()	- free single mbuf with its tags and ext, sys/mbuf.h.
1298  * m_freem()	- free chain of mbufs.
1299  */
1300 
1301 int
1302 m_clget(struct mbuf *m, int how)
1303 {
1304 
1305 	KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
1306 	    __func__, m));
1307 	m->m_ext.ext_buf = (char *)NULL;
1308 	uma_zalloc_arg(zone_clust, m, how);
1309 	/*
1310 	 * On a cluster allocation failure, drain the packet zone and retry,
1311 	 * we might be able to loosen a few clusters up on the drain.
1312 	 */
1313 	if ((how & M_NOWAIT) && (m->m_ext.ext_buf == NULL)) {
1314 		uma_zone_reclaim(zone_pack, UMA_RECLAIM_DRAIN);
1315 		uma_zalloc_arg(zone_clust, m, how);
1316 	}
1317 	MBUF_PROBE2(m__clget, m, how);
1318 	return (m->m_flags & M_EXT);
1319 }
1320 
1321 /*
1322  * m_cljget() is different from m_clget() as it can allocate clusters without
1323  * attaching them to an mbuf.  In that case the return value is the pointer
1324  * to the cluster of the requested size.  If an mbuf was specified, it gets
1325  * the cluster attached to it and the return value can be safely ignored.
1326  * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
1327  */
1328 void *
1329 m_cljget(struct mbuf *m, int how, int size)
1330 {
1331 	uma_zone_t zone;
1332 	void *retval;
1333 
1334 	if (m != NULL) {
1335 		KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
1336 		    __func__, m));
1337 		m->m_ext.ext_buf = NULL;
1338 	}
1339 
1340 	zone = m_getzone(size);
1341 	retval = uma_zalloc_arg(zone, m, how);
1342 
1343 	MBUF_PROBE4(m__cljget, m, how, size, retval);
1344 
1345 	return (retval);
1346 }
1347 
1348 /*
1349  * m_get2() allocates minimum mbuf that would fit "size" argument.
1350  */
1351 struct mbuf *
1352 m_get2(int size, int how, short type, int flags)
1353 {
1354 	struct mb_args args;
1355 	struct mbuf *m, *n;
1356 
1357 	args.flags = flags;
1358 	args.type = type;
1359 
1360 	if (size <= MHLEN || (size <= MLEN && (flags & M_PKTHDR) == 0))
1361 		return (uma_zalloc_arg(zone_mbuf, &args, how));
1362 	if (size <= MCLBYTES)
1363 		return (uma_zalloc_arg(zone_pack, &args, how));
1364 
1365 	if (size > MJUMPAGESIZE)
1366 		return (NULL);
1367 
1368 	m = uma_zalloc_arg(zone_mbuf, &args, how);
1369 	if (m == NULL)
1370 		return (NULL);
1371 
1372 	n = uma_zalloc_arg(zone_jumbop, m, how);
1373 	if (n == NULL) {
1374 		m_free_raw(m);
1375 		return (NULL);
1376 	}
1377 
1378 	return (m);
1379 }
1380 
1381 /*
1382  * m_get3() allocates minimum mbuf that would fit "size" argument.
1383  * Unlike m_get2() it can allocate clusters up to MJUM16BYTES.
1384  */
1385 struct mbuf *
1386 m_get3(int size, int how, short type, int flags)
1387 {
1388 	struct mb_args args;
1389 	struct mbuf *m, *n;
1390 	uma_zone_t zone;
1391 
1392 	if (size <= MJUMPAGESIZE)
1393 		return (m_get2(size, how, type, flags));
1394 
1395 	if (size > MJUM16BYTES)
1396 		return (NULL);
1397 
1398 	args.flags = flags;
1399 	args.type = type;
1400 
1401 	m = uma_zalloc_arg(zone_mbuf, &args, how);
1402 	if (m == NULL)
1403 		return (NULL);
1404 
1405 	if (size <= MJUM9BYTES)
1406 		zone = zone_jumbo9;
1407 	else
1408 		zone = zone_jumbo16;
1409 
1410 	n = uma_zalloc_arg(zone, m, how);
1411 	if (n == NULL) {
1412 		m_free_raw(m);
1413 		return (NULL);
1414 	}
1415 
1416 	return (m);
1417 }
1418 
1419 /*
1420  * m_getjcl() returns an mbuf with a cluster of the specified size attached.
1421  * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
1422  */
1423 struct mbuf *
1424 m_getjcl(int how, short type, int flags, int size)
1425 {
1426 	struct mb_args args;
1427 	struct mbuf *m, *n;
1428 	uma_zone_t zone;
1429 
1430 	if (size == MCLBYTES)
1431 		return m_getcl(how, type, flags);
1432 
1433 	args.flags = flags;
1434 	args.type = type;
1435 
1436 	m = uma_zalloc_arg(zone_mbuf, &args, how);
1437 	if (m == NULL)
1438 		return (NULL);
1439 
1440 	zone = m_getzone(size);
1441 	n = uma_zalloc_arg(zone, m, how);
1442 	if (n == NULL) {
1443 		m_free_raw(m);
1444 		return (NULL);
1445 	}
1446 	MBUF_PROBE5(m__getjcl, how, type, flags, size, m);
1447 	return (m);
1448 }
1449 
1450 /*
1451  * Allocate a given length worth of mbufs and/or clusters (whatever fits
1452  * best) and return a pointer to the top of the allocated chain.  If an
1453  * existing mbuf chain is provided, then we will append the new chain
1454  * to the existing one and return a pointer to the provided mbuf.
1455  */
1456 struct mbuf *
1457 m_getm2(struct mbuf *m, int len, int how, short type, int flags)
1458 {
1459 	struct mbuf *mb, *nm = NULL, *mtail = NULL;
1460 
1461 	KASSERT(len >= 0, ("%s: len is < 0", __func__));
1462 
1463 	/* Validate flags. */
1464 	flags &= (M_PKTHDR | M_EOR);
1465 
1466 	/* Packet header mbuf must be first in chain. */
1467 	if ((flags & M_PKTHDR) && m != NULL)
1468 		flags &= ~M_PKTHDR;
1469 
1470 	/* Loop and append maximum sized mbufs to the chain tail. */
1471 	while (len > 0) {
1472 		mb = NULL;
1473 		if (len > MCLBYTES) {
1474 			mb = m_getjcl(M_NOWAIT, type, (flags & M_PKTHDR),
1475 			    MJUMPAGESIZE);
1476 		}
1477 		if (mb == NULL) {
1478 			if (len >= MINCLSIZE)
1479 				mb = m_getcl(how, type, (flags & M_PKTHDR));
1480 			else if (flags & M_PKTHDR)
1481 				mb = m_gethdr(how, type);
1482 			else
1483 				mb = m_get(how, type);
1484 
1485 			/*
1486 			 * Fail the whole operation if one mbuf can't be
1487 			 * allocated.
1488 			 */
1489 			if (mb == NULL) {
1490 				m_freem(nm);
1491 				return (NULL);
1492 			}
1493 		}
1494 
1495 		/* Book keeping. */
1496 		len -= M_SIZE(mb);
1497 		if (mtail != NULL)
1498 			mtail->m_next = mb;
1499 		else
1500 			nm = mb;
1501 		mtail = mb;
1502 		flags &= ~M_PKTHDR;	/* Only valid on the first mbuf. */
1503 	}
1504 	if (flags & M_EOR)
1505 		mtail->m_flags |= M_EOR;  /* Only valid on the last mbuf. */
1506 
1507 	/* If mbuf was supplied, append new chain to the end of it. */
1508 	if (m != NULL) {
1509 		for (mtail = m; mtail->m_next != NULL; mtail = mtail->m_next)
1510 			;
1511 		mtail->m_next = nm;
1512 		mtail->m_flags &= ~M_EOR;
1513 	} else
1514 		m = nm;
1515 
1516 	return (m);
1517 }
1518 
1519 /*-
1520  * Configure a provided mbuf to refer to the provided external storage
1521  * buffer and setup a reference count for said buffer.
1522  *
1523  * Arguments:
1524  *    mb     The existing mbuf to which to attach the provided buffer.
1525  *    buf    The address of the provided external storage buffer.
1526  *    size   The size of the provided buffer.
1527  *    freef  A pointer to a routine that is responsible for freeing the
1528  *           provided external storage buffer.
1529  *    args   A pointer to an argument structure (of any type) to be passed
1530  *           to the provided freef routine (may be NULL).
1531  *    flags  Any other flags to be passed to the provided mbuf.
1532  *    type   The type that the external storage buffer should be
1533  *           labeled with.
1534  *
1535  * Returns:
1536  *    Nothing.
1537  */
1538 void
1539 m_extadd(struct mbuf *mb, char *buf, u_int size, m_ext_free_t freef,
1540     void *arg1, void *arg2, int flags, int type)
1541 {
1542 
1543 	KASSERT(type != EXT_CLUSTER, ("%s: EXT_CLUSTER not allowed", __func__));
1544 
1545 	mb->m_flags |= (M_EXT | flags);
1546 	mb->m_ext.ext_buf = buf;
1547 	mb->m_data = mb->m_ext.ext_buf;
1548 	mb->m_ext.ext_size = size;
1549 	mb->m_ext.ext_free = freef;
1550 	mb->m_ext.ext_arg1 = arg1;
1551 	mb->m_ext.ext_arg2 = arg2;
1552 	mb->m_ext.ext_type = type;
1553 
1554 	if (type != EXT_EXTREF) {
1555 		mb->m_ext.ext_count = 1;
1556 		mb->m_ext.ext_flags = EXT_FLAG_EMBREF;
1557 	} else
1558 		mb->m_ext.ext_flags = 0;
1559 }
1560 
1561 /*
1562  * Free an entire chain of mbufs and associated external buffers, if
1563  * applicable.
1564  */
1565 void
1566 m_freem(struct mbuf *mb)
1567 {
1568 
1569 	MBUF_PROBE1(m__freem, mb);
1570 	while (mb != NULL)
1571 		mb = m_free(mb);
1572 }
1573 
1574 /*
1575  * Temporary primitive to allow freeing without going through m_free.
1576  */
1577 void
1578 m_free_raw(struct mbuf *mb)
1579 {
1580 
1581 	uma_zfree(zone_mbuf, mb);
1582 }
1583 
1584 int
1585 m_snd_tag_alloc(struct ifnet *ifp, union if_snd_tag_alloc_params *params,
1586     struct m_snd_tag **mstp)
1587 {
1588 
1589 	if (ifp->if_snd_tag_alloc == NULL)
1590 		return (EOPNOTSUPP);
1591 	return (ifp->if_snd_tag_alloc(ifp, params, mstp));
1592 }
1593 
1594 void
1595 m_snd_tag_init(struct m_snd_tag *mst, struct ifnet *ifp,
1596     const struct if_snd_tag_sw *sw)
1597 {
1598 
1599 	if_ref(ifp);
1600 	mst->ifp = ifp;
1601 	refcount_init(&mst->refcount, 1);
1602 	mst->sw = sw;
1603 	counter_u64_add(snd_tag_count, 1);
1604 }
1605 
1606 void
1607 m_snd_tag_destroy(struct m_snd_tag *mst)
1608 {
1609 	struct ifnet *ifp;
1610 
1611 	ifp = mst->ifp;
1612 	mst->sw->snd_tag_free(mst);
1613 	if_rele(ifp);
1614 	counter_u64_add(snd_tag_count, -1);
1615 }
1616 
1617 void
1618 m_rcvif_serialize(struct mbuf *m)
1619 {
1620 	u_short idx, gen;
1621 
1622 	M_ASSERTPKTHDR(m);
1623 	idx = m->m_pkthdr.rcvif->if_index;
1624 	gen = m->m_pkthdr.rcvif->if_idxgen;
1625 	m->m_pkthdr.rcvidx = idx;
1626 	m->m_pkthdr.rcvgen = gen;
1627 	if (__predict_false(m->m_pkthdr.leaf_rcvif != NULL)) {
1628 		idx = m->m_pkthdr.leaf_rcvif->if_index;
1629 		gen = m->m_pkthdr.leaf_rcvif->if_idxgen;
1630 	} else {
1631 		idx = -1;
1632 		gen = 0;
1633 	}
1634 	m->m_pkthdr.leaf_rcvidx = idx;
1635 	m->m_pkthdr.leaf_rcvgen = gen;
1636 }
1637 
1638 struct ifnet *
1639 m_rcvif_restore(struct mbuf *m)
1640 {
1641 	struct ifnet *ifp, *leaf_ifp;
1642 
1643 	M_ASSERTPKTHDR(m);
1644 	NET_EPOCH_ASSERT();
1645 
1646 	ifp = ifnet_byindexgen(m->m_pkthdr.rcvidx, m->m_pkthdr.rcvgen);
1647 	if (ifp == NULL || (ifp->if_flags & IFF_DYING))
1648 		return (NULL);
1649 
1650 	if (__predict_true(m->m_pkthdr.leaf_rcvidx == (u_short)-1)) {
1651 		leaf_ifp = NULL;
1652 	} else {
1653 		leaf_ifp = ifnet_byindexgen(m->m_pkthdr.leaf_rcvidx,
1654 		    m->m_pkthdr.leaf_rcvgen);
1655 		if (__predict_false(leaf_ifp != NULL && (leaf_ifp->if_flags & IFF_DYING)))
1656 			leaf_ifp = NULL;
1657 	}
1658 
1659 	m->m_pkthdr.leaf_rcvif = leaf_ifp;
1660 	m->m_pkthdr.rcvif = ifp;
1661 
1662 	return (ifp);
1663 }
1664 
1665 /*
1666  * Allocate an mbuf with anonymous external pages.
1667  */
1668 struct mbuf *
1669 mb_alloc_ext_plus_pages(int len, int how)
1670 {
1671 	struct mbuf *m;
1672 	vm_page_t pg;
1673 	int i, npgs;
1674 
1675 	m = mb_alloc_ext_pgs(how, mb_free_mext_pgs);
1676 	if (m == NULL)
1677 		return (NULL);
1678 	m->m_epg_flags |= EPG_FLAG_ANON;
1679 	npgs = howmany(len, PAGE_SIZE);
1680 	for (i = 0; i < npgs; i++) {
1681 		do {
1682 			pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
1683 			    VM_ALLOC_WIRED);
1684 			if (pg == NULL) {
1685 				if (how == M_NOWAIT) {
1686 					m->m_epg_npgs = i;
1687 					m_free(m);
1688 					return (NULL);
1689 				}
1690 				vm_wait(NULL);
1691 			}
1692 		} while (pg == NULL);
1693 		m->m_epg_pa[i] = VM_PAGE_TO_PHYS(pg);
1694 	}
1695 	m->m_epg_npgs = npgs;
1696 	return (m);
1697 }
1698 
1699 /*
1700  * Copy the data in the mbuf chain to a chain of mbufs with anonymous external
1701  * unmapped pages.
1702  * len is the length of data in the input mbuf chain.
1703  * mlen is the maximum number of bytes put into each ext_page mbuf.
1704  */
1705 struct mbuf *
1706 mb_mapped_to_unmapped(struct mbuf *mp, int len, int mlen, int how,
1707     struct mbuf **mlast)
1708 {
1709 	struct mbuf *m, *mout;
1710 	char *pgpos, *mbpos;
1711 	int i, mblen, mbufsiz, pglen, xfer;
1712 
1713 	if (len == 0)
1714 		return (NULL);
1715 	mbufsiz = min(mlen, len);
1716 	m = mout = mb_alloc_ext_plus_pages(mbufsiz, how);
1717 	if (m == NULL)
1718 		return (m);
1719 	pgpos = (char *)(void *)PHYS_TO_DMAP(m->m_epg_pa[0]);
1720 	pglen = PAGE_SIZE;
1721 	mblen = 0;
1722 	i = 0;
1723 	do {
1724 		if (pglen == 0) {
1725 			if (++i == m->m_epg_npgs) {
1726 				m->m_epg_last_len = PAGE_SIZE;
1727 				mbufsiz = min(mlen, len);
1728 				m->m_next = mb_alloc_ext_plus_pages(mbufsiz,
1729 				    how);
1730 				m = m->m_next;
1731 				if (m == NULL) {
1732 					m_freem(mout);
1733 					return (m);
1734 				}
1735 				i = 0;
1736 			}
1737 			pgpos = (char *)(void *)PHYS_TO_DMAP(m->m_epg_pa[i]);
1738 			pglen = PAGE_SIZE;
1739 		}
1740 		while (mblen == 0) {
1741 			if (mp == NULL) {
1742 				m_freem(mout);
1743 				return (NULL);
1744 			}
1745 			KASSERT((mp->m_flags & M_EXTPG) == 0,
1746 			    ("mb_copym_ext_pgs: ext_pgs input mbuf"));
1747 			mbpos = mtod(mp, char *);
1748 			mblen = mp->m_len;
1749 			mp = mp->m_next;
1750 		}
1751 		xfer = min(mblen, pglen);
1752 		memcpy(pgpos, mbpos, xfer);
1753 		pgpos += xfer;
1754 		mbpos += xfer;
1755 		pglen -= xfer;
1756 		mblen -= xfer;
1757 		len -= xfer;
1758 		m->m_len += xfer;
1759 	} while (len > 0);
1760 	m->m_epg_last_len = PAGE_SIZE - pglen;
1761 	if (mlast != NULL)
1762 		*mlast = m;
1763 	return (mout);
1764 }
1765