xref: /freebsd/sys/kern/kern_mbuf.c (revision dd41de95a84d979615a2ef11df6850622bf6184e)
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/domain.h>
43 #include <sys/eventhandler.h>
44 #include <sys/kernel.h>
45 #include <sys/ktls.h>
46 #include <sys/limits.h>
47 #include <sys/lock.h>
48 #include <sys/mutex.h>
49 #include <sys/protosw.h>
50 #include <sys/refcount.h>
51 #include <sys/sf_buf.h>
52 #include <sys/smp.h>
53 #include <sys/socket.h>
54 #include <sys/sysctl.h>
55 
56 #include <net/if.h>
57 #include <net/if_var.h>
58 
59 #include <vm/vm.h>
60 #include <vm/vm_extern.h>
61 #include <vm/vm_kern.h>
62 #include <vm/vm_page.h>
63 #include <vm/vm_pageout.h>
64 #include <vm/vm_map.h>
65 #include <vm/uma.h>
66 #include <vm/uma_dbg.h>
67 
68 /*
69  * In FreeBSD, Mbufs and Mbuf Clusters are allocated from UMA
70  * Zones.
71  *
72  * Mbuf Clusters (2K, contiguous) are allocated from the Cluster
73  * Zone.  The Zone can be capped at kern.ipc.nmbclusters, if the
74  * administrator so desires.
75  *
76  * Mbufs are allocated from a UMA Primary Zone called the Mbuf
77  * Zone.
78  *
79  * Additionally, FreeBSD provides a Packet Zone, which it
80  * configures as a Secondary Zone to the Mbuf Primary Zone,
81  * thus sharing backend Slab kegs with the Mbuf Primary Zone.
82  *
83  * Thus common-case allocations and locking are simplified:
84  *
85  *  m_clget()                m_getcl()
86  *    |                         |
87  *    |   .------------>[(Packet Cache)]    m_get(), m_gethdr()
88  *    |   |             [     Packet   ]            |
89  *  [(Cluster Cache)]   [    Secondary ]   [ (Mbuf Cache)     ]
90  *  [ Cluster Zone  ]   [     Zone     ]   [ Mbuf Primary Zone ]
91  *        |                       \________         |
92  *  [ Cluster Keg   ]                      \       /
93  *        |	                         [ Mbuf Keg   ]
94  *  [ Cluster Slabs ]                         |
95  *        |                              [ Mbuf Slabs ]
96  *         \____________(VM)_________________/
97  *
98  *
99  * Whenever an object is allocated with uma_zalloc() out of
100  * one of the Zones its _ctor_ function is executed.  The same
101  * for any deallocation through uma_zfree() the _dtor_ function
102  * is executed.
103  *
104  * Caches are per-CPU and are filled from the Primary Zone.
105  *
106  * Whenever an object is allocated from the underlying global
107  * memory pool it gets pre-initialized with the _zinit_ functions.
108  * When the Keg's are overfull objects get decommissioned with
109  * _zfini_ functions and free'd back to the global memory pool.
110  *
111  */
112 
113 int nmbufs;			/* limits number of mbufs */
114 int nmbclusters;		/* limits number of mbuf clusters */
115 int nmbjumbop;			/* limits number of page size jumbo clusters */
116 int nmbjumbo9;			/* limits number of 9k jumbo clusters */
117 int nmbjumbo16;			/* limits number of 16k jumbo clusters */
118 
119 bool mb_use_ext_pgs = true;	/* use M_EXTPG mbufs for sendfile & TLS */
120 SYSCTL_BOOL(_kern_ipc, OID_AUTO, mb_use_ext_pgs, CTLFLAG_RWTUN,
121     &mb_use_ext_pgs, 0,
122     "Use unmapped mbufs for sendfile(2) and TLS offload");
123 
124 static quad_t maxmbufmem;	/* overall real memory limit for all mbufs */
125 
126 SYSCTL_QUAD(_kern_ipc, OID_AUTO, maxmbufmem, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &maxmbufmem, 0,
127     "Maximum real memory allocatable to various mbuf types");
128 
129 static counter_u64_t snd_tag_count;
130 SYSCTL_COUNTER_U64(_kern_ipc, OID_AUTO, num_snd_tags, CTLFLAG_RW,
131     &snd_tag_count, "# of active mbuf send tags");
132 
133 /*
134  * tunable_mbinit() has to be run before any mbuf allocations are done.
135  */
136 static void
137 tunable_mbinit(void *dummy)
138 {
139 	quad_t realmem;
140 
141 	/*
142 	 * The default limit for all mbuf related memory is 1/2 of all
143 	 * available kernel memory (physical or kmem).
144 	 * At most it can be 3/4 of available kernel memory.
145 	 */
146 	realmem = qmin((quad_t)physmem * PAGE_SIZE, vm_kmem_size);
147 	maxmbufmem = realmem / 2;
148 	TUNABLE_QUAD_FETCH("kern.ipc.maxmbufmem", &maxmbufmem);
149 	if (maxmbufmem > realmem / 4 * 3)
150 		maxmbufmem = realmem / 4 * 3;
151 
152 	TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
153 	if (nmbclusters == 0)
154 		nmbclusters = maxmbufmem / MCLBYTES / 4;
155 
156 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbop", &nmbjumbop);
157 	if (nmbjumbop == 0)
158 		nmbjumbop = maxmbufmem / MJUMPAGESIZE / 4;
159 
160 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbo9", &nmbjumbo9);
161 	if (nmbjumbo9 == 0)
162 		nmbjumbo9 = maxmbufmem / MJUM9BYTES / 6;
163 
164 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbo16", &nmbjumbo16);
165 	if (nmbjumbo16 == 0)
166 		nmbjumbo16 = maxmbufmem / MJUM16BYTES / 6;
167 
168 	/*
169 	 * We need at least as many mbufs as we have clusters of
170 	 * the various types added together.
171 	 */
172 	TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
173 	if (nmbufs < nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16)
174 		nmbufs = lmax(maxmbufmem / MSIZE / 5,
175 		    nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16);
176 }
177 SYSINIT(tunable_mbinit, SI_SUB_KMEM, SI_ORDER_MIDDLE, tunable_mbinit, NULL);
178 
179 static int
180 sysctl_nmbclusters(SYSCTL_HANDLER_ARGS)
181 {
182 	int error, newnmbclusters;
183 
184 	newnmbclusters = nmbclusters;
185 	error = sysctl_handle_int(oidp, &newnmbclusters, 0, req);
186 	if (error == 0 && req->newptr && newnmbclusters != nmbclusters) {
187 		if (newnmbclusters > nmbclusters &&
188 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
189 			nmbclusters = newnmbclusters;
190 			nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
191 			EVENTHANDLER_INVOKE(nmbclusters_change);
192 		} else
193 			error = EINVAL;
194 	}
195 	return (error);
196 }
197 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbclusters,
198     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &nmbclusters, 0,
199     sysctl_nmbclusters, "IU",
200     "Maximum number of mbuf clusters allowed");
201 
202 static int
203 sysctl_nmbjumbop(SYSCTL_HANDLER_ARGS)
204 {
205 	int error, newnmbjumbop;
206 
207 	newnmbjumbop = nmbjumbop;
208 	error = sysctl_handle_int(oidp, &newnmbjumbop, 0, req);
209 	if (error == 0 && req->newptr && newnmbjumbop != nmbjumbop) {
210 		if (newnmbjumbop > nmbjumbop &&
211 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
212 			nmbjumbop = newnmbjumbop;
213 			nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
214 		} else
215 			error = EINVAL;
216 	}
217 	return (error);
218 }
219 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbop,
220     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &nmbjumbop, 0,
221     sysctl_nmbjumbop, "IU",
222     "Maximum number of mbuf page size jumbo clusters allowed");
223 
224 static int
225 sysctl_nmbjumbo9(SYSCTL_HANDLER_ARGS)
226 {
227 	int error, newnmbjumbo9;
228 
229 	newnmbjumbo9 = nmbjumbo9;
230 	error = sysctl_handle_int(oidp, &newnmbjumbo9, 0, req);
231 	if (error == 0 && req->newptr && newnmbjumbo9 != nmbjumbo9) {
232 		if (newnmbjumbo9 > nmbjumbo9 &&
233 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
234 			nmbjumbo9 = newnmbjumbo9;
235 			nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
236 		} else
237 			error = EINVAL;
238 	}
239 	return (error);
240 }
241 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo9,
242     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &nmbjumbo9, 0,
243     sysctl_nmbjumbo9, "IU",
244     "Maximum number of mbuf 9k jumbo clusters allowed");
245 
246 static int
247 sysctl_nmbjumbo16(SYSCTL_HANDLER_ARGS)
248 {
249 	int error, newnmbjumbo16;
250 
251 	newnmbjumbo16 = nmbjumbo16;
252 	error = sysctl_handle_int(oidp, &newnmbjumbo16, 0, req);
253 	if (error == 0 && req->newptr && newnmbjumbo16 != nmbjumbo16) {
254 		if (newnmbjumbo16 > nmbjumbo16 &&
255 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
256 			nmbjumbo16 = newnmbjumbo16;
257 			nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
258 		} else
259 			error = EINVAL;
260 	}
261 	return (error);
262 }
263 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo16,
264     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &nmbjumbo16, 0,
265     sysctl_nmbjumbo16, "IU",
266     "Maximum number of mbuf 16k jumbo clusters allowed");
267 
268 static int
269 sysctl_nmbufs(SYSCTL_HANDLER_ARGS)
270 {
271 	int error, newnmbufs;
272 
273 	newnmbufs = nmbufs;
274 	error = sysctl_handle_int(oidp, &newnmbufs, 0, req);
275 	if (error == 0 && req->newptr && newnmbufs != nmbufs) {
276 		if (newnmbufs > nmbufs) {
277 			nmbufs = newnmbufs;
278 			nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
279 			EVENTHANDLER_INVOKE(nmbufs_change);
280 		} else
281 			error = EINVAL;
282 	}
283 	return (error);
284 }
285 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbufs,
286     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
287     &nmbufs, 0, sysctl_nmbufs, "IU",
288     "Maximum number of mbufs allowed");
289 
290 /*
291  * Zones from which we allocate.
292  */
293 uma_zone_t	zone_mbuf;
294 uma_zone_t	zone_clust;
295 uma_zone_t	zone_pack;
296 uma_zone_t	zone_jumbop;
297 uma_zone_t	zone_jumbo9;
298 uma_zone_t	zone_jumbo16;
299 
300 /*
301  * Local prototypes.
302  */
303 static int	mb_ctor_mbuf(void *, int, void *, int);
304 static int	mb_ctor_clust(void *, int, void *, int);
305 static int	mb_ctor_pack(void *, int, void *, int);
306 static void	mb_dtor_mbuf(void *, int, void *);
307 static void	mb_dtor_pack(void *, int, void *);
308 static int	mb_zinit_pack(void *, int, int);
309 static void	mb_zfini_pack(void *, int);
310 static void	mb_reclaim(uma_zone_t, int);
311 
312 /* Ensure that MSIZE is a power of 2. */
313 CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE);
314 
315 _Static_assert(sizeof(struct mbuf) <= MSIZE,
316     "size of mbuf exceeds MSIZE");
317 /*
318  * Initialize FreeBSD Network buffer allocation.
319  */
320 static void
321 mbuf_init(void *dummy)
322 {
323 
324 	/*
325 	 * Configure UMA zones for Mbufs, Clusters, and Packets.
326 	 */
327 	zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE,
328 	    mb_ctor_mbuf, mb_dtor_mbuf, NULL, NULL,
329 	    MSIZE - 1, UMA_ZONE_CONTIG | UMA_ZONE_MAXBUCKET);
330 	if (nmbufs > 0)
331 		nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
332 	uma_zone_set_warning(zone_mbuf, "kern.ipc.nmbufs limit reached");
333 	uma_zone_set_maxaction(zone_mbuf, mb_reclaim);
334 
335 	zone_clust = uma_zcreate(MBUF_CLUSTER_MEM_NAME, MCLBYTES,
336 	    mb_ctor_clust, NULL, NULL, NULL,
337 	    UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
338 	if (nmbclusters > 0)
339 		nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
340 	uma_zone_set_warning(zone_clust, "kern.ipc.nmbclusters limit reached");
341 	uma_zone_set_maxaction(zone_clust, mb_reclaim);
342 
343 	zone_pack = uma_zsecond_create(MBUF_PACKET_MEM_NAME, mb_ctor_pack,
344 	    mb_dtor_pack, mb_zinit_pack, mb_zfini_pack, zone_mbuf);
345 
346 	/* Make jumbo frame zone too. Page size, 9k and 16k. */
347 	zone_jumbop = uma_zcreate(MBUF_JUMBOP_MEM_NAME, MJUMPAGESIZE,
348 	    mb_ctor_clust, NULL, NULL, NULL,
349 	    UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
350 	if (nmbjumbop > 0)
351 		nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
352 	uma_zone_set_warning(zone_jumbop, "kern.ipc.nmbjumbop limit reached");
353 	uma_zone_set_maxaction(zone_jumbop, mb_reclaim);
354 
355 	zone_jumbo9 = uma_zcreate(MBUF_JUMBO9_MEM_NAME, MJUM9BYTES,
356 	    mb_ctor_clust, NULL, NULL, NULL,
357 	    UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
358 	if (nmbjumbo9 > 0)
359 		nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
360 	uma_zone_set_warning(zone_jumbo9, "kern.ipc.nmbjumbo9 limit reached");
361 	uma_zone_set_maxaction(zone_jumbo9, mb_reclaim);
362 
363 	zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES,
364 	    mb_ctor_clust, NULL, NULL, NULL,
365 	    UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
366 	if (nmbjumbo16 > 0)
367 		nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
368 	uma_zone_set_warning(zone_jumbo16, "kern.ipc.nmbjumbo16 limit reached");
369 	uma_zone_set_maxaction(zone_jumbo16, mb_reclaim);
370 
371 	/*
372 	 * Hook event handler for low-memory situation, used to
373 	 * drain protocols and push data back to the caches (UMA
374 	 * later pushes it back to VM).
375 	 */
376 	EVENTHANDLER_REGISTER(vm_lowmem, mb_reclaim, NULL,
377 	    EVENTHANDLER_PRI_FIRST);
378 
379 	snd_tag_count = counter_u64_alloc(M_WAITOK);
380 }
381 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL);
382 
383 #ifdef DEBUGNET
384 /*
385  * debugnet makes use of a pre-allocated pool of mbufs and clusters.  When
386  * debugnet is configured, we initialize a set of UMA cache zones which return
387  * items from this pool.  At panic-time, the regular UMA zone pointers are
388  * overwritten with those of the cache zones so that drivers may allocate and
389  * free mbufs and clusters without attempting to allocate physical memory.
390  *
391  * We keep mbufs and clusters in a pair of mbuf queues.  In particular, for
392  * the purpose of caching clusters, we treat them as mbufs.
393  */
394 static struct mbufq dn_mbufq =
395     { STAILQ_HEAD_INITIALIZER(dn_mbufq.mq_head), 0, INT_MAX };
396 static struct mbufq dn_clustq =
397     { STAILQ_HEAD_INITIALIZER(dn_clustq.mq_head), 0, INT_MAX };
398 
399 static int dn_clsize;
400 static uma_zone_t dn_zone_mbuf;
401 static uma_zone_t dn_zone_clust;
402 static uma_zone_t dn_zone_pack;
403 
404 static struct debugnet_saved_zones {
405 	uma_zone_t dsz_mbuf;
406 	uma_zone_t dsz_clust;
407 	uma_zone_t dsz_pack;
408 	uma_zone_t dsz_jumbop;
409 	uma_zone_t dsz_jumbo9;
410 	uma_zone_t dsz_jumbo16;
411 	bool dsz_debugnet_zones_enabled;
412 } dn_saved_zones;
413 
414 static int
415 dn_buf_import(void *arg, void **store, int count, int domain __unused,
416     int flags)
417 {
418 	struct mbufq *q;
419 	struct mbuf *m;
420 	int i;
421 
422 	q = arg;
423 
424 	for (i = 0; i < count; i++) {
425 		m = mbufq_dequeue(q);
426 		if (m == NULL)
427 			break;
428 		trash_init(m, q == &dn_mbufq ? MSIZE : dn_clsize, flags);
429 		store[i] = m;
430 	}
431 	KASSERT((flags & M_WAITOK) == 0 || i == count,
432 	    ("%s: ran out of pre-allocated mbufs", __func__));
433 	return (i);
434 }
435 
436 static void
437 dn_buf_release(void *arg, void **store, int count)
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 = store[i];
447 		(void)mbufq_enqueue(q, m);
448 	}
449 }
450 
451 static int
452 dn_pack_import(void *arg __unused, void **store, int count, int domain __unused,
453     int flags __unused)
454 {
455 	struct mbuf *m;
456 	void *clust;
457 	int i;
458 
459 	for (i = 0; i < count; i++) {
460 		m = m_get(MT_DATA, M_NOWAIT);
461 		if (m == NULL)
462 			break;
463 		clust = uma_zalloc(dn_zone_clust, M_NOWAIT);
464 		if (clust == NULL) {
465 			m_free(m);
466 			break;
467 		}
468 		mb_ctor_clust(clust, dn_clsize, m, 0);
469 		store[i] = m;
470 	}
471 	KASSERT((flags & M_WAITOK) == 0 || i == count,
472 	    ("%s: ran out of pre-allocated mbufs", __func__));
473 	return (i);
474 }
475 
476 static void
477 dn_pack_release(void *arg __unused, void **store, int count)
478 {
479 	struct mbuf *m;
480 	void *clust;
481 	int i;
482 
483 	for (i = 0; i < count; i++) {
484 		m = store[i];
485 		clust = m->m_ext.ext_buf;
486 		uma_zfree(dn_zone_clust, clust);
487 		uma_zfree(dn_zone_mbuf, m);
488 	}
489 }
490 
491 /*
492  * Free the pre-allocated mbufs and clusters reserved for debugnet, and destroy
493  * the corresponding UMA cache zones.
494  */
495 void
496 debugnet_mbuf_drain(void)
497 {
498 	struct mbuf *m;
499 	void *item;
500 
501 	if (dn_zone_mbuf != NULL) {
502 		uma_zdestroy(dn_zone_mbuf);
503 		dn_zone_mbuf = NULL;
504 	}
505 	if (dn_zone_clust != NULL) {
506 		uma_zdestroy(dn_zone_clust);
507 		dn_zone_clust = NULL;
508 	}
509 	if (dn_zone_pack != NULL) {
510 		uma_zdestroy(dn_zone_pack);
511 		dn_zone_pack = NULL;
512 	}
513 
514 	while ((m = mbufq_dequeue(&dn_mbufq)) != NULL)
515 		m_free(m);
516 	while ((item = mbufq_dequeue(&dn_clustq)) != NULL)
517 		uma_zfree(m_getzone(dn_clsize), item);
518 }
519 
520 /*
521  * Callback invoked immediately prior to starting a debugnet connection.
522  */
523 void
524 debugnet_mbuf_start(void)
525 {
526 
527 	MPASS(!dn_saved_zones.dsz_debugnet_zones_enabled);
528 
529 	/* Save the old zone pointers to restore when debugnet is closed. */
530 	dn_saved_zones = (struct debugnet_saved_zones) {
531 		.dsz_debugnet_zones_enabled = true,
532 		.dsz_mbuf = zone_mbuf,
533 		.dsz_clust = zone_clust,
534 		.dsz_pack = zone_pack,
535 		.dsz_jumbop = zone_jumbop,
536 		.dsz_jumbo9 = zone_jumbo9,
537 		.dsz_jumbo16 = zone_jumbo16,
538 	};
539 
540 	/*
541 	 * All cluster zones return buffers of the size requested by the
542 	 * drivers.  It's up to the driver to reinitialize the zones if the
543 	 * MTU of a debugnet-enabled interface changes.
544 	 */
545 	printf("debugnet: overwriting mbuf zone pointers\n");
546 	zone_mbuf = dn_zone_mbuf;
547 	zone_clust = dn_zone_clust;
548 	zone_pack = dn_zone_pack;
549 	zone_jumbop = dn_zone_clust;
550 	zone_jumbo9 = dn_zone_clust;
551 	zone_jumbo16 = dn_zone_clust;
552 }
553 
554 /*
555  * Callback invoked when a debugnet connection is closed/finished.
556  */
557 void
558 debugnet_mbuf_finish(void)
559 {
560 
561 	MPASS(dn_saved_zones.dsz_debugnet_zones_enabled);
562 
563 	printf("debugnet: restoring mbuf zone pointers\n");
564 	zone_mbuf = dn_saved_zones.dsz_mbuf;
565 	zone_clust = dn_saved_zones.dsz_clust;
566 	zone_pack = dn_saved_zones.dsz_pack;
567 	zone_jumbop = dn_saved_zones.dsz_jumbop;
568 	zone_jumbo9 = dn_saved_zones.dsz_jumbo9;
569 	zone_jumbo16 = dn_saved_zones.dsz_jumbo16;
570 
571 	memset(&dn_saved_zones, 0, sizeof(dn_saved_zones));
572 }
573 
574 /*
575  * Reinitialize the debugnet mbuf+cluster pool and cache zones.
576  */
577 void
578 debugnet_mbuf_reinit(int nmbuf, int nclust, int clsize)
579 {
580 	struct mbuf *m;
581 	void *item;
582 
583 	debugnet_mbuf_drain();
584 
585 	dn_clsize = clsize;
586 
587 	dn_zone_mbuf = uma_zcache_create("debugnet_" MBUF_MEM_NAME,
588 	    MSIZE, mb_ctor_mbuf, mb_dtor_mbuf, NULL, NULL,
589 	    dn_buf_import, dn_buf_release,
590 	    &dn_mbufq, UMA_ZONE_NOBUCKET);
591 
592 	dn_zone_clust = uma_zcache_create("debugnet_" MBUF_CLUSTER_MEM_NAME,
593 	    clsize, mb_ctor_clust, NULL, NULL, NULL,
594 	    dn_buf_import, dn_buf_release,
595 	    &dn_clustq, UMA_ZONE_NOBUCKET);
596 
597 	dn_zone_pack = uma_zcache_create("debugnet_" MBUF_PACKET_MEM_NAME,
598 	    MCLBYTES, mb_ctor_pack, mb_dtor_pack, NULL, NULL,
599 	    dn_pack_import, dn_pack_release,
600 	    NULL, UMA_ZONE_NOBUCKET);
601 
602 	while (nmbuf-- > 0) {
603 		m = m_get(MT_DATA, M_WAITOK);
604 		uma_zfree(dn_zone_mbuf, m);
605 	}
606 	while (nclust-- > 0) {
607 		item = uma_zalloc(m_getzone(dn_clsize), M_WAITOK);
608 		uma_zfree(dn_zone_clust, item);
609 	}
610 }
611 #endif /* DEBUGNET */
612 
613 /*
614  * Constructor for Mbuf primary zone.
615  *
616  * The 'arg' pointer points to a mb_args structure which
617  * contains call-specific information required to support the
618  * mbuf allocation API.  See mbuf.h.
619  */
620 static int
621 mb_ctor_mbuf(void *mem, int size, void *arg, int how)
622 {
623 	struct mbuf *m;
624 	struct mb_args *args;
625 	int error;
626 	int flags;
627 	short type;
628 
629 	args = (struct mb_args *)arg;
630 	type = args->type;
631 
632 	/*
633 	 * The mbuf is initialized later.  The caller has the
634 	 * responsibility to set up any MAC labels too.
635 	 */
636 	if (type == MT_NOINIT)
637 		return (0);
638 
639 	m = (struct mbuf *)mem;
640 	flags = args->flags;
641 	MPASS((flags & M_NOFREE) == 0);
642 
643 	error = m_init(m, how, type, flags);
644 
645 	return (error);
646 }
647 
648 /*
649  * The Mbuf primary zone destructor.
650  */
651 static void
652 mb_dtor_mbuf(void *mem, int size, void *arg)
653 {
654 	struct mbuf *m;
655 	unsigned long flags;
656 
657 	m = (struct mbuf *)mem;
658 	flags = (unsigned long)arg;
659 
660 	KASSERT((m->m_flags & M_NOFREE) == 0, ("%s: M_NOFREE set", __func__));
661 	if (!(flags & MB_DTOR_SKIP) && (m->m_flags & M_PKTHDR) && !SLIST_EMPTY(&m->m_pkthdr.tags))
662 		m_tag_delete_chain(m, NULL);
663 }
664 
665 /*
666  * The Mbuf Packet zone destructor.
667  */
668 static void
669 mb_dtor_pack(void *mem, int size, void *arg)
670 {
671 	struct mbuf *m;
672 
673 	m = (struct mbuf *)mem;
674 	if ((m->m_flags & M_PKTHDR) != 0)
675 		m_tag_delete_chain(m, NULL);
676 
677 	/* Make sure we've got a clean cluster back. */
678 	KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__));
679 	KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__));
680 	KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__));
681 	KASSERT(m->m_ext.ext_arg1 == NULL, ("%s: ext_arg1 != NULL", __func__));
682 	KASSERT(m->m_ext.ext_arg2 == NULL, ("%s: ext_arg2 != NULL", __func__));
683 	KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__));
684 	KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__));
685 #ifdef INVARIANTS
686 	trash_dtor(m->m_ext.ext_buf, MCLBYTES, arg);
687 #endif
688 	/*
689 	 * If there are processes blocked on zone_clust, waiting for pages
690 	 * to be freed up, cause them to be woken up by draining the
691 	 * packet zone.  We are exposed to a race here (in the check for
692 	 * the UMA_ZFLAG_FULL) where we might miss the flag set, but that
693 	 * is deliberate. We don't want to acquire the zone lock for every
694 	 * mbuf free.
695 	 */
696 	if (uma_zone_exhausted(zone_clust))
697 		uma_zone_reclaim(zone_pack, UMA_RECLAIM_DRAIN);
698 }
699 
700 /*
701  * The Cluster and Jumbo[PAGESIZE|9|16] zone constructor.
702  *
703  * Here the 'arg' pointer points to the Mbuf which we
704  * are configuring cluster storage for.  If 'arg' is
705  * empty we allocate just the cluster without setting
706  * the mbuf to it.  See mbuf.h.
707  */
708 static int
709 mb_ctor_clust(void *mem, int size, void *arg, int how)
710 {
711 	struct mbuf *m;
712 
713 	m = (struct mbuf *)arg;
714 	if (m != NULL) {
715 		m->m_ext.ext_buf = (char *)mem;
716 		m->m_data = m->m_ext.ext_buf;
717 		m->m_flags |= M_EXT;
718 		m->m_ext.ext_free = NULL;
719 		m->m_ext.ext_arg1 = NULL;
720 		m->m_ext.ext_arg2 = NULL;
721 		m->m_ext.ext_size = size;
722 		m->m_ext.ext_type = m_gettype(size);
723 		m->m_ext.ext_flags = EXT_FLAG_EMBREF;
724 		m->m_ext.ext_count = 1;
725 	}
726 
727 	return (0);
728 }
729 
730 /*
731  * The Packet secondary zone's init routine, executed on the
732  * object's transition from mbuf keg slab to zone cache.
733  */
734 static int
735 mb_zinit_pack(void *mem, int size, int how)
736 {
737 	struct mbuf *m;
738 
739 	m = (struct mbuf *)mem;		/* m is virgin. */
740 	if (uma_zalloc_arg(zone_clust, m, how) == NULL ||
741 	    m->m_ext.ext_buf == NULL)
742 		return (ENOMEM);
743 	m->m_ext.ext_type = EXT_PACKET;	/* Override. */
744 #ifdef INVARIANTS
745 	trash_init(m->m_ext.ext_buf, MCLBYTES, how);
746 #endif
747 	return (0);
748 }
749 
750 /*
751  * The Packet secondary zone's fini routine, executed on the
752  * object's transition from zone cache to keg slab.
753  */
754 static void
755 mb_zfini_pack(void *mem, int size)
756 {
757 	struct mbuf *m;
758 
759 	m = (struct mbuf *)mem;
760 #ifdef INVARIANTS
761 	trash_fini(m->m_ext.ext_buf, MCLBYTES);
762 #endif
763 	uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL);
764 #ifdef INVARIANTS
765 	trash_dtor(mem, size, NULL);
766 #endif
767 }
768 
769 /*
770  * The "packet" keg constructor.
771  */
772 static int
773 mb_ctor_pack(void *mem, int size, void *arg, int how)
774 {
775 	struct mbuf *m;
776 	struct mb_args *args;
777 	int error, flags;
778 	short type;
779 
780 	m = (struct mbuf *)mem;
781 	args = (struct mb_args *)arg;
782 	flags = args->flags;
783 	type = args->type;
784 	MPASS((flags & M_NOFREE) == 0);
785 
786 #ifdef INVARIANTS
787 	trash_ctor(m->m_ext.ext_buf, MCLBYTES, arg, how);
788 #endif
789 
790 	error = m_init(m, how, type, flags);
791 
792 	/* m_ext is already initialized. */
793 	m->m_data = m->m_ext.ext_buf;
794  	m->m_flags = (flags | M_EXT);
795 
796 	return (error);
797 }
798 
799 /*
800  * This is the protocol drain routine.  Called by UMA whenever any of the
801  * mbuf zones is closed to its limit.
802  *
803  * No locks should be held when this is called.  The drain routines have to
804  * presently acquire some locks which raises the possibility of lock order
805  * reversal.
806  */
807 static void
808 mb_reclaim(uma_zone_t zone __unused, int pending __unused)
809 {
810 	struct epoch_tracker et;
811 	struct domain *dp;
812 	struct protosw *pr;
813 
814 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK | WARN_PANIC, NULL, __func__);
815 
816 	NET_EPOCH_ENTER(et);
817 	for (dp = domains; dp != NULL; dp = dp->dom_next)
818 		for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++)
819 			if (pr->pr_drain != NULL)
820 				(*pr->pr_drain)();
821 	NET_EPOCH_EXIT(et);
822 }
823 
824 /*
825  * Free "count" units of I/O from an mbuf chain.  They could be held
826  * in M_EXTPG or just as a normal mbuf.  This code is intended to be
827  * called in an error path (I/O error, closed connection, etc).
828  */
829 void
830 mb_free_notready(struct mbuf *m, int count)
831 {
832 	int i;
833 
834 	for (i = 0; i < count && m != NULL; i++) {
835 		if ((m->m_flags & M_EXTPG) != 0) {
836 			m->m_epg_nrdy--;
837 			if (m->m_epg_nrdy != 0)
838 				continue;
839 		}
840 		m = m_free(m);
841 	}
842 	KASSERT(i == count, ("Removed only %d items from %p", i, m));
843 }
844 
845 /*
846  * Compress an unmapped mbuf into a simple mbuf when it holds a small
847  * amount of data.  This is used as a DOS defense to avoid having
848  * small packets tie up wired pages, an ext_pgs structure, and an
849  * mbuf.  Since this converts the existing mbuf in place, it can only
850  * be used if there are no other references to 'm'.
851  */
852 int
853 mb_unmapped_compress(struct mbuf *m)
854 {
855 	volatile u_int *refcnt;
856 	char buf[MLEN];
857 
858 	/*
859 	 * Assert that 'm' does not have a packet header.  If 'm' had
860 	 * a packet header, it would only be able to hold MHLEN bytes
861 	 * and m_data would have to be initialized differently.
862 	 */
863 	KASSERT((m->m_flags & M_PKTHDR) == 0 && (m->m_flags & M_EXTPG),
864             ("%s: m %p !M_EXTPG or M_PKTHDR", __func__, m));
865 	KASSERT(m->m_len <= MLEN, ("m_len too large %p", m));
866 
867 	if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
868 		refcnt = &m->m_ext.ext_count;
869 	} else {
870 		KASSERT(m->m_ext.ext_cnt != NULL,
871 		    ("%s: no refcounting pointer on %p", __func__, m));
872 		refcnt = m->m_ext.ext_cnt;
873 	}
874 
875 	if (*refcnt != 1)
876 		return (EBUSY);
877 
878 	m_copydata(m, 0, m->m_len, buf);
879 
880 	/* Free the backing pages. */
881 	m->m_ext.ext_free(m);
882 
883 	/* Turn 'm' into a "normal" mbuf. */
884 	m->m_flags &= ~(M_EXT | M_RDONLY | M_EXTPG);
885 	m->m_data = m->m_dat;
886 
887 	/* Copy data back into m. */
888 	bcopy(buf, mtod(m, char *), m->m_len);
889 
890 	return (0);
891 }
892 
893 /*
894  * These next few routines are used to permit downgrading an unmapped
895  * mbuf to a chain of mapped mbufs.  This is used when an interface
896  * doesn't supported unmapped mbufs or if checksums need to be
897  * computed in software.
898  *
899  * Each unmapped mbuf is converted to a chain of mbufs.  First, any
900  * TLS header data is stored in a regular mbuf.  Second, each page of
901  * unmapped data is stored in an mbuf with an EXT_SFBUF external
902  * cluster.  These mbufs use an sf_buf to provide a valid KVA for the
903  * associated physical page.  They also hold a reference on the
904  * original M_EXTPG mbuf to ensure the physical page doesn't go away.
905  * Finally, any TLS trailer data is stored in a regular mbuf.
906  *
907  * mb_unmapped_free_mext() is the ext_free handler for the EXT_SFBUF
908  * mbufs.  It frees the associated sf_buf and releases its reference
909  * on the original M_EXTPG mbuf.
910  *
911  * _mb_unmapped_to_ext() is a helper function that converts a single
912  * unmapped mbuf into a chain of mbufs.
913  *
914  * mb_unmapped_to_ext() is the public function that walks an mbuf
915  * chain converting any unmapped mbufs to mapped mbufs.  It returns
916  * the new chain of unmapped mbufs on success.  On failure it frees
917  * the original mbuf chain and returns NULL.
918  */
919 static void
920 mb_unmapped_free_mext(struct mbuf *m)
921 {
922 	struct sf_buf *sf;
923 	struct mbuf *old_m;
924 
925 	sf = m->m_ext.ext_arg1;
926 	sf_buf_free(sf);
927 
928 	/* Drop the reference on the backing M_EXTPG mbuf. */
929 	old_m = m->m_ext.ext_arg2;
930 	mb_free_extpg(old_m);
931 }
932 
933 static struct mbuf *
934 _mb_unmapped_to_ext(struct mbuf *m)
935 {
936 	struct mbuf *m_new, *top, *prev, *mref;
937 	struct sf_buf *sf;
938 	vm_page_t pg;
939 	int i, len, off, pglen, pgoff, seglen, segoff;
940 	volatile u_int *refcnt;
941 	u_int ref_inc = 0;
942 
943 	M_ASSERTEXTPG(m);
944 	len = m->m_len;
945 	KASSERT(m->m_epg_tls == NULL, ("%s: can't convert TLS mbuf %p",
946 	    __func__, m));
947 
948 	/* See if this is the mbuf that holds the embedded refcount. */
949 	if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
950 		refcnt = &m->m_ext.ext_count;
951 		mref = m;
952 	} else {
953 		KASSERT(m->m_ext.ext_cnt != NULL,
954 		    ("%s: no refcounting pointer on %p", __func__, m));
955 		refcnt = m->m_ext.ext_cnt;
956 		mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
957 	}
958 
959 	/* Skip over any data removed from the front. */
960 	off = mtod(m, vm_offset_t);
961 
962 	top = NULL;
963 	if (m->m_epg_hdrlen != 0) {
964 		if (off >= m->m_epg_hdrlen) {
965 			off -= m->m_epg_hdrlen;
966 		} else {
967 			seglen = m->m_epg_hdrlen - off;
968 			segoff = off;
969 			seglen = min(seglen, len);
970 			off = 0;
971 			len -= seglen;
972 			m_new = m_get(M_NOWAIT, MT_DATA);
973 			if (m_new == NULL)
974 				goto fail;
975 			m_new->m_len = seglen;
976 			prev = top = m_new;
977 			memcpy(mtod(m_new, void *), &m->m_epg_hdr[segoff],
978 			    seglen);
979 		}
980 	}
981 	pgoff = m->m_epg_1st_off;
982 	for (i = 0; i < m->m_epg_npgs && len > 0; i++) {
983 		pglen = m_epg_pagelen(m, i, pgoff);
984 		if (off >= pglen) {
985 			off -= pglen;
986 			pgoff = 0;
987 			continue;
988 		}
989 		seglen = pglen - off;
990 		segoff = pgoff + off;
991 		off = 0;
992 		seglen = min(seglen, len);
993 		len -= seglen;
994 
995 		pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]);
996 		m_new = m_get(M_NOWAIT, MT_DATA);
997 		if (m_new == NULL)
998 			goto fail;
999 		if (top == NULL) {
1000 			top = prev = m_new;
1001 		} else {
1002 			prev->m_next = m_new;
1003 			prev = m_new;
1004 		}
1005 		sf = sf_buf_alloc(pg, SFB_NOWAIT);
1006 		if (sf == NULL)
1007 			goto fail;
1008 
1009 		ref_inc++;
1010 		m_extadd(m_new, (char *)sf_buf_kva(sf), PAGE_SIZE,
1011 		    mb_unmapped_free_mext, sf, mref, M_RDONLY, EXT_SFBUF);
1012 		m_new->m_data += segoff;
1013 		m_new->m_len = seglen;
1014 
1015 		pgoff = 0;
1016 	};
1017 	if (len != 0) {
1018 		KASSERT((off + len) <= m->m_epg_trllen,
1019 		    ("off + len > trail (%d + %d > %d)", off, len,
1020 		    m->m_epg_trllen));
1021 		m_new = m_get(M_NOWAIT, MT_DATA);
1022 		if (m_new == NULL)
1023 			goto fail;
1024 		if (top == NULL)
1025 			top = m_new;
1026 		else
1027 			prev->m_next = m_new;
1028 		m_new->m_len = len;
1029 		memcpy(mtod(m_new, void *), &m->m_epg_trail[off], len);
1030 	}
1031 
1032 	if (ref_inc != 0) {
1033 		/*
1034 		 * Obtain an additional reference on the old mbuf for
1035 		 * each created EXT_SFBUF mbuf.  They will be dropped
1036 		 * in mb_unmapped_free_mext().
1037 		 */
1038 		if (*refcnt == 1)
1039 			*refcnt += ref_inc;
1040 		else
1041 			atomic_add_int(refcnt, ref_inc);
1042 	}
1043 	m_free(m);
1044 	return (top);
1045 
1046 fail:
1047 	if (ref_inc != 0) {
1048 		/*
1049 		 * Obtain an additional reference on the old mbuf for
1050 		 * each created EXT_SFBUF mbuf.  They will be
1051 		 * immediately dropped when these mbufs are freed
1052 		 * below.
1053 		 */
1054 		if (*refcnt == 1)
1055 			*refcnt += ref_inc;
1056 		else
1057 			atomic_add_int(refcnt, ref_inc);
1058 	}
1059 	m_free(m);
1060 	m_freem(top);
1061 	return (NULL);
1062 }
1063 
1064 struct mbuf *
1065 mb_unmapped_to_ext(struct mbuf *top)
1066 {
1067 	struct mbuf *m, *next, *prev = NULL;
1068 
1069 	prev = NULL;
1070 	for (m = top; m != NULL; m = next) {
1071 		/* m might be freed, so cache the next pointer. */
1072 		next = m->m_next;
1073 		if (m->m_flags & M_EXTPG) {
1074 			if (prev != NULL) {
1075 				/*
1076 				 * Remove 'm' from the new chain so
1077 				 * that the 'top' chain terminates
1078 				 * before 'm' in case 'top' is freed
1079 				 * due to an error.
1080 				 */
1081 				prev->m_next = NULL;
1082 			}
1083 			m = _mb_unmapped_to_ext(m);
1084 			if (m == NULL) {
1085 				m_freem(top);
1086 				m_freem(next);
1087 				return (NULL);
1088 			}
1089 			if (prev == NULL) {
1090 				top = m;
1091 			} else {
1092 				prev->m_next = m;
1093 			}
1094 
1095 			/*
1096 			 * Replaced one mbuf with a chain, so we must
1097 			 * find the end of chain.
1098 			 */
1099 			prev = m_last(m);
1100 		} else {
1101 			if (prev != NULL) {
1102 				prev->m_next = m;
1103 			}
1104 			prev = m;
1105 		}
1106 	}
1107 	return (top);
1108 }
1109 
1110 /*
1111  * Allocate an empty M_EXTPG mbuf.  The ext_free routine is
1112  * responsible for freeing any pages backing this mbuf when it is
1113  * freed.
1114  */
1115 struct mbuf *
1116 mb_alloc_ext_pgs(int how, m_ext_free_t ext_free)
1117 {
1118 	struct mbuf *m;
1119 
1120 	m = m_get(how, MT_DATA);
1121 	if (m == NULL)
1122 		return (NULL);
1123 
1124 	m->m_epg_npgs = 0;
1125 	m->m_epg_nrdy = 0;
1126 	m->m_epg_1st_off = 0;
1127 	m->m_epg_last_len = 0;
1128 	m->m_epg_flags = 0;
1129 	m->m_epg_hdrlen = 0;
1130 	m->m_epg_trllen = 0;
1131 	m->m_epg_tls = NULL;
1132 	m->m_epg_so = NULL;
1133 	m->m_data = NULL;
1134 	m->m_flags |= (M_EXT | M_RDONLY | M_EXTPG);
1135 	m->m_ext.ext_flags = EXT_FLAG_EMBREF;
1136 	m->m_ext.ext_count = 1;
1137 	m->m_ext.ext_size = 0;
1138 	m->m_ext.ext_free = ext_free;
1139 	return (m);
1140 }
1141 
1142 /*
1143  * Clean up after mbufs with M_EXT storage attached to them if the
1144  * reference count hits 1.
1145  */
1146 void
1147 mb_free_ext(struct mbuf *m)
1148 {
1149 	volatile u_int *refcnt;
1150 	struct mbuf *mref;
1151 	int freembuf;
1152 
1153 	KASSERT(m->m_flags & M_EXT, ("%s: M_EXT not set on %p", __func__, m));
1154 
1155 	/* See if this is the mbuf that holds the embedded refcount. */
1156 	if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
1157 		refcnt = &m->m_ext.ext_count;
1158 		mref = m;
1159 	} else {
1160 		KASSERT(m->m_ext.ext_cnt != NULL,
1161 		    ("%s: no refcounting pointer on %p", __func__, m));
1162 		refcnt = m->m_ext.ext_cnt;
1163 		mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
1164 	}
1165 
1166 	/*
1167 	 * Check if the header is embedded in the cluster.  It is
1168 	 * important that we can't touch any of the mbuf fields
1169 	 * after we have freed the external storage, since mbuf
1170 	 * could have been embedded in it.  For now, the mbufs
1171 	 * embedded into the cluster are always of type EXT_EXTREF,
1172 	 * and for this type we won't free the mref.
1173 	 */
1174 	if (m->m_flags & M_NOFREE) {
1175 		freembuf = 0;
1176 		KASSERT(m->m_ext.ext_type == EXT_EXTREF ||
1177 		    m->m_ext.ext_type == EXT_RXRING,
1178 		    ("%s: no-free mbuf %p has wrong type", __func__, m));
1179 	} else
1180 		freembuf = 1;
1181 
1182 	/* Free attached storage if this mbuf is the only reference to it. */
1183 	if (*refcnt == 1 || atomic_fetchadd_int(refcnt, -1) == 1) {
1184 		switch (m->m_ext.ext_type) {
1185 		case EXT_PACKET:
1186 			/* The packet zone is special. */
1187 			if (*refcnt == 0)
1188 				*refcnt = 1;
1189 			uma_zfree(zone_pack, mref);
1190 			break;
1191 		case EXT_CLUSTER:
1192 			uma_zfree(zone_clust, m->m_ext.ext_buf);
1193 			uma_zfree(zone_mbuf, mref);
1194 			break;
1195 		case EXT_JUMBOP:
1196 			uma_zfree(zone_jumbop, m->m_ext.ext_buf);
1197 			uma_zfree(zone_mbuf, mref);
1198 			break;
1199 		case EXT_JUMBO9:
1200 			uma_zfree(zone_jumbo9, m->m_ext.ext_buf);
1201 			uma_zfree(zone_mbuf, mref);
1202 			break;
1203 		case EXT_JUMBO16:
1204 			uma_zfree(zone_jumbo16, m->m_ext.ext_buf);
1205 			uma_zfree(zone_mbuf, mref);
1206 			break;
1207 		case EXT_SFBUF:
1208 		case EXT_NET_DRV:
1209 		case EXT_MOD_TYPE:
1210 		case EXT_DISPOSABLE:
1211 			KASSERT(mref->m_ext.ext_free != NULL,
1212 			    ("%s: ext_free not set", __func__));
1213 			mref->m_ext.ext_free(mref);
1214 			uma_zfree(zone_mbuf, mref);
1215 			break;
1216 		case EXT_EXTREF:
1217 			KASSERT(m->m_ext.ext_free != NULL,
1218 			    ("%s: ext_free not set", __func__));
1219 			m->m_ext.ext_free(m);
1220 			break;
1221 		case EXT_RXRING:
1222 			KASSERT(m->m_ext.ext_free == NULL,
1223 			    ("%s: ext_free is set", __func__));
1224 			break;
1225 		default:
1226 			KASSERT(m->m_ext.ext_type == 0,
1227 			    ("%s: unknown ext_type", __func__));
1228 		}
1229 	}
1230 
1231 	if (freembuf && m != mref)
1232 		uma_zfree(zone_mbuf, m);
1233 }
1234 
1235 /*
1236  * Clean up after mbufs with M_EXTPG storage attached to them if the
1237  * reference count hits 1.
1238  */
1239 void
1240 mb_free_extpg(struct mbuf *m)
1241 {
1242 	volatile u_int *refcnt;
1243 	struct mbuf *mref;
1244 
1245 	M_ASSERTEXTPG(m);
1246 
1247 	/* See if this is the mbuf that holds the embedded refcount. */
1248 	if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
1249 		refcnt = &m->m_ext.ext_count;
1250 		mref = m;
1251 	} else {
1252 		KASSERT(m->m_ext.ext_cnt != NULL,
1253 		    ("%s: no refcounting pointer on %p", __func__, m));
1254 		refcnt = m->m_ext.ext_cnt;
1255 		mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
1256 	}
1257 
1258 	/* Free attached storage if this mbuf is the only reference to it. */
1259 	if (*refcnt == 1 || atomic_fetchadd_int(refcnt, -1) == 1) {
1260 		KASSERT(mref->m_ext.ext_free != NULL,
1261 		    ("%s: ext_free not set", __func__));
1262 
1263 		mref->m_ext.ext_free(mref);
1264 #ifdef KERN_TLS
1265 		if (mref->m_epg_tls != NULL &&
1266 		    !refcount_release_if_not_last(&mref->m_epg_tls->refcount))
1267 			ktls_enqueue_to_free(mref);
1268 		else
1269 #endif
1270 			uma_zfree(zone_mbuf, mref);
1271 	}
1272 
1273 	if (m != mref)
1274 		uma_zfree(zone_mbuf, m);
1275 }
1276 
1277 /*
1278  * Official mbuf(9) allocation KPI for stack and drivers:
1279  *
1280  * m_get()	- a single mbuf without any attachments, sys/mbuf.h.
1281  * m_gethdr()	- a single mbuf initialized as M_PKTHDR, sys/mbuf.h.
1282  * m_getcl()	- an mbuf + 2k cluster, sys/mbuf.h.
1283  * m_clget()	- attach cluster to already allocated mbuf.
1284  * m_cljget()	- attach jumbo cluster to already allocated mbuf.
1285  * m_get2()	- allocate minimum mbuf that would fit size argument.
1286  * m_getm2()	- allocate a chain of mbufs/clusters.
1287  * m_extadd()	- attach external cluster to mbuf.
1288  *
1289  * m_free()	- free single mbuf with its tags and ext, sys/mbuf.h.
1290  * m_freem()	- free chain of mbufs.
1291  */
1292 
1293 int
1294 m_clget(struct mbuf *m, int how)
1295 {
1296 
1297 	KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
1298 	    __func__, m));
1299 	m->m_ext.ext_buf = (char *)NULL;
1300 	uma_zalloc_arg(zone_clust, m, how);
1301 	/*
1302 	 * On a cluster allocation failure, drain the packet zone and retry,
1303 	 * we might be able to loosen a few clusters up on the drain.
1304 	 */
1305 	if ((how & M_NOWAIT) && (m->m_ext.ext_buf == NULL)) {
1306 		uma_zone_reclaim(zone_pack, UMA_RECLAIM_DRAIN);
1307 		uma_zalloc_arg(zone_clust, m, how);
1308 	}
1309 	MBUF_PROBE2(m__clget, m, how);
1310 	return (m->m_flags & M_EXT);
1311 }
1312 
1313 /*
1314  * m_cljget() is different from m_clget() as it can allocate clusters without
1315  * attaching them to an mbuf.  In that case the return value is the pointer
1316  * to the cluster of the requested size.  If an mbuf was specified, it gets
1317  * the cluster attached to it and the return value can be safely ignored.
1318  * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
1319  */
1320 void *
1321 m_cljget(struct mbuf *m, int how, int size)
1322 {
1323 	uma_zone_t zone;
1324 	void *retval;
1325 
1326 	if (m != NULL) {
1327 		KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
1328 		    __func__, m));
1329 		m->m_ext.ext_buf = NULL;
1330 	}
1331 
1332 	zone = m_getzone(size);
1333 	retval = uma_zalloc_arg(zone, m, how);
1334 
1335 	MBUF_PROBE4(m__cljget, m, how, size, retval);
1336 
1337 	return (retval);
1338 }
1339 
1340 /*
1341  * m_get2() allocates minimum mbuf that would fit "size" argument.
1342  */
1343 struct mbuf *
1344 m_get2(int size, int how, short type, int flags)
1345 {
1346 	struct mb_args args;
1347 	struct mbuf *m, *n;
1348 
1349 	args.flags = flags;
1350 	args.type = type;
1351 
1352 	if (size <= MHLEN || (size <= MLEN && (flags & M_PKTHDR) == 0))
1353 		return (uma_zalloc_arg(zone_mbuf, &args, how));
1354 	if (size <= MCLBYTES)
1355 		return (uma_zalloc_arg(zone_pack, &args, how));
1356 
1357 	if (size > MJUMPAGESIZE)
1358 		return (NULL);
1359 
1360 	m = uma_zalloc_arg(zone_mbuf, &args, how);
1361 	if (m == NULL)
1362 		return (NULL);
1363 
1364 	n = uma_zalloc_arg(zone_jumbop, m, how);
1365 	if (n == NULL) {
1366 		uma_zfree(zone_mbuf, m);
1367 		return (NULL);
1368 	}
1369 
1370 	return (m);
1371 }
1372 
1373 /*
1374  * m_getjcl() returns an mbuf with a cluster of the specified size attached.
1375  * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
1376  */
1377 struct mbuf *
1378 m_getjcl(int how, short type, int flags, int size)
1379 {
1380 	struct mb_args args;
1381 	struct mbuf *m, *n;
1382 	uma_zone_t zone;
1383 
1384 	if (size == MCLBYTES)
1385 		return m_getcl(how, type, flags);
1386 
1387 	args.flags = flags;
1388 	args.type = type;
1389 
1390 	m = uma_zalloc_arg(zone_mbuf, &args, how);
1391 	if (m == NULL)
1392 		return (NULL);
1393 
1394 	zone = m_getzone(size);
1395 	n = uma_zalloc_arg(zone, m, how);
1396 	if (n == NULL) {
1397 		uma_zfree(zone_mbuf, m);
1398 		return (NULL);
1399 	}
1400 	MBUF_PROBE5(m__getjcl, how, type, flags, size, m);
1401 	return (m);
1402 }
1403 
1404 /*
1405  * Allocate a given length worth of mbufs and/or clusters (whatever fits
1406  * best) and return a pointer to the top of the allocated chain.  If an
1407  * existing mbuf chain is provided, then we will append the new chain
1408  * to the existing one and return a pointer to the provided mbuf.
1409  */
1410 struct mbuf *
1411 m_getm2(struct mbuf *m, int len, int how, short type, int flags)
1412 {
1413 	struct mbuf *mb, *nm = NULL, *mtail = NULL;
1414 
1415 	KASSERT(len >= 0, ("%s: len is < 0", __func__));
1416 
1417 	/* Validate flags. */
1418 	flags &= (M_PKTHDR | M_EOR);
1419 
1420 	/* Packet header mbuf must be first in chain. */
1421 	if ((flags & M_PKTHDR) && m != NULL)
1422 		flags &= ~M_PKTHDR;
1423 
1424 	/* Loop and append maximum sized mbufs to the chain tail. */
1425 	while (len > 0) {
1426 		mb = NULL;
1427 		if (len > MCLBYTES) {
1428 			mb = m_getjcl(M_NOWAIT, type, (flags & M_PKTHDR),
1429 			    MJUMPAGESIZE);
1430 		}
1431 		if (mb == NULL) {
1432 			if (len >= MINCLSIZE)
1433 				mb = m_getcl(how, type, (flags & M_PKTHDR));
1434 			else if (flags & M_PKTHDR)
1435 				mb = m_gethdr(how, type);
1436 			else
1437 				mb = m_get(how, type);
1438 
1439 			/*
1440 			 * Fail the whole operation if one mbuf can't be
1441 			 * allocated.
1442 			 */
1443 			if (mb == NULL) {
1444 				m_freem(nm);
1445 				return (NULL);
1446 			}
1447 		}
1448 
1449 		/* Book keeping. */
1450 		len -= M_SIZE(mb);
1451 		if (mtail != NULL)
1452 			mtail->m_next = mb;
1453 		else
1454 			nm = mb;
1455 		mtail = mb;
1456 		flags &= ~M_PKTHDR;	/* Only valid on the first mbuf. */
1457 	}
1458 	if (flags & M_EOR)
1459 		mtail->m_flags |= M_EOR;  /* Only valid on the last mbuf. */
1460 
1461 	/* If mbuf was supplied, append new chain to the end of it. */
1462 	if (m != NULL) {
1463 		for (mtail = m; mtail->m_next != NULL; mtail = mtail->m_next)
1464 			;
1465 		mtail->m_next = nm;
1466 		mtail->m_flags &= ~M_EOR;
1467 	} else
1468 		m = nm;
1469 
1470 	return (m);
1471 }
1472 
1473 /*-
1474  * Configure a provided mbuf to refer to the provided external storage
1475  * buffer and setup a reference count for said buffer.
1476  *
1477  * Arguments:
1478  *    mb     The existing mbuf to which to attach the provided buffer.
1479  *    buf    The address of the provided external storage buffer.
1480  *    size   The size of the provided buffer.
1481  *    freef  A pointer to a routine that is responsible for freeing the
1482  *           provided external storage buffer.
1483  *    args   A pointer to an argument structure (of any type) to be passed
1484  *           to the provided freef routine (may be NULL).
1485  *    flags  Any other flags to be passed to the provided mbuf.
1486  *    type   The type that the external storage buffer should be
1487  *           labeled with.
1488  *
1489  * Returns:
1490  *    Nothing.
1491  */
1492 void
1493 m_extadd(struct mbuf *mb, char *buf, u_int size, m_ext_free_t freef,
1494     void *arg1, void *arg2, int flags, int type)
1495 {
1496 
1497 	KASSERT(type != EXT_CLUSTER, ("%s: EXT_CLUSTER not allowed", __func__));
1498 
1499 	mb->m_flags |= (M_EXT | flags);
1500 	mb->m_ext.ext_buf = buf;
1501 	mb->m_data = mb->m_ext.ext_buf;
1502 	mb->m_ext.ext_size = size;
1503 	mb->m_ext.ext_free = freef;
1504 	mb->m_ext.ext_arg1 = arg1;
1505 	mb->m_ext.ext_arg2 = arg2;
1506 	mb->m_ext.ext_type = type;
1507 
1508 	if (type != EXT_EXTREF) {
1509 		mb->m_ext.ext_count = 1;
1510 		mb->m_ext.ext_flags = EXT_FLAG_EMBREF;
1511 	} else
1512 		mb->m_ext.ext_flags = 0;
1513 }
1514 
1515 /*
1516  * Free an entire chain of mbufs and associated external buffers, if
1517  * applicable.
1518  */
1519 void
1520 m_freem(struct mbuf *mb)
1521 {
1522 
1523 	MBUF_PROBE1(m__freem, mb);
1524 	while (mb != NULL)
1525 		mb = m_free(mb);
1526 }
1527 
1528 int
1529 m_snd_tag_alloc(struct ifnet *ifp, union if_snd_tag_alloc_params *params,
1530     struct m_snd_tag **mstp)
1531 {
1532 
1533 	if (ifp->if_snd_tag_alloc == NULL)
1534 		return (EOPNOTSUPP);
1535 	return (ifp->if_snd_tag_alloc(ifp, params, mstp));
1536 }
1537 
1538 void
1539 m_snd_tag_init(struct m_snd_tag *mst, struct ifnet *ifp, u_int type)
1540 {
1541 
1542 	if_ref(ifp);
1543 	mst->ifp = ifp;
1544 	refcount_init(&mst->refcount, 1);
1545 	mst->type = type;
1546 	counter_u64_add(snd_tag_count, 1);
1547 }
1548 
1549 void
1550 m_snd_tag_destroy(struct m_snd_tag *mst)
1551 {
1552 	struct ifnet *ifp;
1553 
1554 	ifp = mst->ifp;
1555 	ifp->if_snd_tag_free(mst);
1556 	if_rele(ifp);
1557 	counter_u64_add(snd_tag_count, -1);
1558 }
1559 
1560 /*
1561  * Allocate an mbuf with anonymous external pages.
1562  */
1563 struct mbuf *
1564 mb_alloc_ext_plus_pages(int len, int how)
1565 {
1566 	struct mbuf *m;
1567 	vm_page_t pg;
1568 	int i, npgs;
1569 
1570 	m = mb_alloc_ext_pgs(how, mb_free_mext_pgs);
1571 	if (m == NULL)
1572 		return (NULL);
1573 	m->m_epg_flags |= EPG_FLAG_ANON;
1574 	npgs = howmany(len, PAGE_SIZE);
1575 	for (i = 0; i < npgs; i++) {
1576 		do {
1577 			pg = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
1578 			    VM_ALLOC_NOOBJ | VM_ALLOC_NODUMP | VM_ALLOC_WIRED);
1579 			if (pg == NULL) {
1580 				if (how == M_NOWAIT) {
1581 					m->m_epg_npgs = i;
1582 					m_free(m);
1583 					return (NULL);
1584 				}
1585 				vm_wait(NULL);
1586 			}
1587 		} while (pg == NULL);
1588 		m->m_epg_pa[i] = VM_PAGE_TO_PHYS(pg);
1589 	}
1590 	m->m_epg_npgs = npgs;
1591 	return (m);
1592 }
1593 
1594 /*
1595  * Copy the data in the mbuf chain to a chain of mbufs with anonymous external
1596  * unmapped pages.
1597  * len is the length of data in the input mbuf chain.
1598  * mlen is the maximum number of bytes put into each ext_page mbuf.
1599  */
1600 struct mbuf *
1601 mb_mapped_to_unmapped(struct mbuf *mp, int len, int mlen, int how,
1602     struct mbuf **mlast)
1603 {
1604 	struct mbuf *m, *mout;
1605 	char *pgpos, *mbpos;
1606 	int i, mblen, mbufsiz, pglen, xfer;
1607 
1608 	if (len == 0)
1609 		return (NULL);
1610 	mbufsiz = min(mlen, len);
1611 	m = mout = mb_alloc_ext_plus_pages(mbufsiz, how);
1612 	if (m == NULL)
1613 		return (m);
1614 	pgpos = (char *)(void *)PHYS_TO_DMAP(m->m_epg_pa[0]);
1615 	pglen = PAGE_SIZE;
1616 	mblen = 0;
1617 	i = 0;
1618 	do {
1619 		if (pglen == 0) {
1620 			if (++i == m->m_epg_npgs) {
1621 				m->m_epg_last_len = PAGE_SIZE;
1622 				mbufsiz = min(mlen, len);
1623 				m->m_next = mb_alloc_ext_plus_pages(mbufsiz,
1624 				    how);
1625 				m = m->m_next;
1626 				if (m == NULL) {
1627 					m_freem(mout);
1628 					return (m);
1629 				}
1630 				i = 0;
1631 			}
1632 			pgpos = (char *)(void *)PHYS_TO_DMAP(m->m_epg_pa[i]);
1633 			pglen = PAGE_SIZE;
1634 		}
1635 		while (mblen == 0) {
1636 			if (mp == NULL) {
1637 				m_freem(mout);
1638 				return (NULL);
1639 			}
1640 			KASSERT((mp->m_flags & M_EXTPG) == 0,
1641 			    ("mb_copym_ext_pgs: ext_pgs input mbuf"));
1642 			mbpos = mtod(mp, char *);
1643 			mblen = mp->m_len;
1644 			mp = mp->m_next;
1645 		}
1646 		xfer = min(mblen, pglen);
1647 		memcpy(pgpos, mbpos, xfer);
1648 		pgpos += xfer;
1649 		mbpos += xfer;
1650 		pglen -= xfer;
1651 		mblen -= xfer;
1652 		len -= xfer;
1653 		m->m_len += xfer;
1654 	} while (len > 0);
1655 	m->m_epg_last_len = PAGE_SIZE - pglen;
1656 	if (mlast != NULL)
1657 		*mlast = m;
1658 	return (mout);
1659 }
1660