xref: /freebsd/sys/kern/kern_mbuf.c (revision 8ddb146abcdf061be9f2c0db7e391697dafad85c)
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 = false;	/* use M_EXTPG mbufs for sendfile & TLS */
120 
121 static int
122 sysctl_mb_use_ext_pgs(SYSCTL_HANDLER_ARGS)
123 {
124 	int error, extpg;
125 
126 	extpg = mb_use_ext_pgs;
127 	error = sysctl_handle_int(oidp, &extpg, 0, req);
128 	if (error == 0 && req->newptr != NULL) {
129 		if (extpg != 0 && !PMAP_HAS_DMAP)
130 			error = EOPNOTSUPP;
131 		else
132 			mb_use_ext_pgs = extpg != 0;
133 	}
134 	return (error);
135 }
136 SYSCTL_PROC(_kern_ipc, OID_AUTO, mb_use_ext_pgs, CTLTYPE_INT | CTLFLAG_RW,
137     &mb_use_ext_pgs, 0,
138     sysctl_mb_use_ext_pgs, "IU",
139     "Use unmapped mbufs for sendfile(2) and TLS offload");
140 
141 static quad_t maxmbufmem;	/* overall real memory limit for all mbufs */
142 
143 SYSCTL_QUAD(_kern_ipc, OID_AUTO, maxmbufmem, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &maxmbufmem, 0,
144     "Maximum real memory allocatable to various mbuf types");
145 
146 static counter_u64_t snd_tag_count;
147 SYSCTL_COUNTER_U64(_kern_ipc, OID_AUTO, num_snd_tags, CTLFLAG_RW,
148     &snd_tag_count, "# of active mbuf send tags");
149 
150 /*
151  * tunable_mbinit() has to be run before any mbuf allocations are done.
152  */
153 static void
154 tunable_mbinit(void *dummy)
155 {
156 	quad_t realmem;
157 	int extpg;
158 
159 	/*
160 	 * The default limit for all mbuf related memory is 1/2 of all
161 	 * available kernel memory (physical or kmem).
162 	 * At most it can be 3/4 of available kernel memory.
163 	 */
164 	realmem = qmin((quad_t)physmem * PAGE_SIZE, vm_kmem_size);
165 	maxmbufmem = realmem / 2;
166 	TUNABLE_QUAD_FETCH("kern.ipc.maxmbufmem", &maxmbufmem);
167 	if (maxmbufmem > realmem / 4 * 3)
168 		maxmbufmem = realmem / 4 * 3;
169 
170 	TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
171 	if (nmbclusters == 0)
172 		nmbclusters = maxmbufmem / MCLBYTES / 4;
173 
174 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbop", &nmbjumbop);
175 	if (nmbjumbop == 0)
176 		nmbjumbop = maxmbufmem / MJUMPAGESIZE / 4;
177 
178 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbo9", &nmbjumbo9);
179 	if (nmbjumbo9 == 0)
180 		nmbjumbo9 = maxmbufmem / MJUM9BYTES / 6;
181 
182 	TUNABLE_INT_FETCH("kern.ipc.nmbjumbo16", &nmbjumbo16);
183 	if (nmbjumbo16 == 0)
184 		nmbjumbo16 = maxmbufmem / MJUM16BYTES / 6;
185 
186 	/*
187 	 * We need at least as many mbufs as we have clusters of
188 	 * the various types added together.
189 	 */
190 	TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
191 	if (nmbufs < nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16)
192 		nmbufs = lmax(maxmbufmem / MSIZE / 5,
193 		    nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16);
194 
195 	/*
196 	 * Unmapped mbufs can only safely be used on platforms with a direct
197 	 * map.
198 	 */
199 	if (PMAP_HAS_DMAP) {
200 		extpg = 1;
201 		TUNABLE_INT_FETCH("kern.ipc.mb_use_ext_pgs", &extpg);
202 		mb_use_ext_pgs = extpg != 0;
203 	}
204 }
205 SYSINIT(tunable_mbinit, SI_SUB_KMEM, SI_ORDER_MIDDLE, tunable_mbinit, NULL);
206 
207 static int
208 sysctl_nmbclusters(SYSCTL_HANDLER_ARGS)
209 {
210 	int error, newnmbclusters;
211 
212 	newnmbclusters = nmbclusters;
213 	error = sysctl_handle_int(oidp, &newnmbclusters, 0, req);
214 	if (error == 0 && req->newptr && newnmbclusters != nmbclusters) {
215 		if (newnmbclusters > nmbclusters &&
216 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
217 			nmbclusters = newnmbclusters;
218 			nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
219 			EVENTHANDLER_INVOKE(nmbclusters_change);
220 		} else
221 			error = EINVAL;
222 	}
223 	return (error);
224 }
225 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbclusters,
226     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, &nmbclusters, 0,
227     sysctl_nmbclusters, "IU",
228     "Maximum number of mbuf clusters allowed");
229 
230 static int
231 sysctl_nmbjumbop(SYSCTL_HANDLER_ARGS)
232 {
233 	int error, newnmbjumbop;
234 
235 	newnmbjumbop = nmbjumbop;
236 	error = sysctl_handle_int(oidp, &newnmbjumbop, 0, req);
237 	if (error == 0 && req->newptr && newnmbjumbop != nmbjumbop) {
238 		if (newnmbjumbop > nmbjumbop &&
239 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
240 			nmbjumbop = newnmbjumbop;
241 			nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
242 		} else
243 			error = EINVAL;
244 	}
245 	return (error);
246 }
247 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbop,
248     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, &nmbjumbop, 0,
249     sysctl_nmbjumbop, "IU",
250     "Maximum number of mbuf page size jumbo clusters allowed");
251 
252 static int
253 sysctl_nmbjumbo9(SYSCTL_HANDLER_ARGS)
254 {
255 	int error, newnmbjumbo9;
256 
257 	newnmbjumbo9 = nmbjumbo9;
258 	error = sysctl_handle_int(oidp, &newnmbjumbo9, 0, req);
259 	if (error == 0 && req->newptr && newnmbjumbo9 != nmbjumbo9) {
260 		if (newnmbjumbo9 > nmbjumbo9 &&
261 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
262 			nmbjumbo9 = newnmbjumbo9;
263 			nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
264 		} else
265 			error = EINVAL;
266 	}
267 	return (error);
268 }
269 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo9,
270     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, &nmbjumbo9, 0,
271     sysctl_nmbjumbo9, "IU",
272     "Maximum number of mbuf 9k jumbo clusters allowed");
273 
274 static int
275 sysctl_nmbjumbo16(SYSCTL_HANDLER_ARGS)
276 {
277 	int error, newnmbjumbo16;
278 
279 	newnmbjumbo16 = nmbjumbo16;
280 	error = sysctl_handle_int(oidp, &newnmbjumbo16, 0, req);
281 	if (error == 0 && req->newptr && newnmbjumbo16 != nmbjumbo16) {
282 		if (newnmbjumbo16 > nmbjumbo16 &&
283 		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
284 			nmbjumbo16 = newnmbjumbo16;
285 			nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
286 		} else
287 			error = EINVAL;
288 	}
289 	return (error);
290 }
291 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo16,
292     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, &nmbjumbo16, 0,
293     sysctl_nmbjumbo16, "IU",
294     "Maximum number of mbuf 16k jumbo clusters allowed");
295 
296 static int
297 sysctl_nmbufs(SYSCTL_HANDLER_ARGS)
298 {
299 	int error, newnmbufs;
300 
301 	newnmbufs = nmbufs;
302 	error = sysctl_handle_int(oidp, &newnmbufs, 0, req);
303 	if (error == 0 && req->newptr && newnmbufs != nmbufs) {
304 		if (newnmbufs > nmbufs) {
305 			nmbufs = newnmbufs;
306 			nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
307 			EVENTHANDLER_INVOKE(nmbufs_change);
308 		} else
309 			error = EINVAL;
310 	}
311 	return (error);
312 }
313 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbufs,
314     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
315     &nmbufs, 0, sysctl_nmbufs, "IU",
316     "Maximum number of mbufs allowed");
317 
318 /*
319  * Zones from which we allocate.
320  */
321 uma_zone_t	zone_mbuf;
322 uma_zone_t	zone_clust;
323 uma_zone_t	zone_pack;
324 uma_zone_t	zone_jumbop;
325 uma_zone_t	zone_jumbo9;
326 uma_zone_t	zone_jumbo16;
327 
328 /*
329  * Local prototypes.
330  */
331 static int	mb_ctor_mbuf(void *, int, void *, int);
332 static int	mb_ctor_clust(void *, int, void *, int);
333 static int	mb_ctor_pack(void *, int, void *, int);
334 static void	mb_dtor_mbuf(void *, int, void *);
335 static void	mb_dtor_pack(void *, int, void *);
336 static int	mb_zinit_pack(void *, int, int);
337 static void	mb_zfini_pack(void *, int);
338 static void	mb_reclaim(uma_zone_t, int);
339 
340 /* Ensure that MSIZE is a power of 2. */
341 CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE);
342 
343 _Static_assert(sizeof(struct mbuf) <= MSIZE,
344     "size of mbuf exceeds MSIZE");
345 /*
346  * Initialize FreeBSD Network buffer allocation.
347  */
348 static void
349 mbuf_init(void *dummy)
350 {
351 
352 	/*
353 	 * Configure UMA zones for Mbufs, Clusters, and Packets.
354 	 */
355 	zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE,
356 	    mb_ctor_mbuf, mb_dtor_mbuf, NULL, NULL,
357 	    MSIZE - 1, UMA_ZONE_CONTIG | UMA_ZONE_MAXBUCKET);
358 	if (nmbufs > 0)
359 		nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
360 	uma_zone_set_warning(zone_mbuf, "kern.ipc.nmbufs limit reached");
361 	uma_zone_set_maxaction(zone_mbuf, mb_reclaim);
362 
363 	zone_clust = uma_zcreate(MBUF_CLUSTER_MEM_NAME, MCLBYTES,
364 	    mb_ctor_clust, NULL, NULL, NULL,
365 	    UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
366 	if (nmbclusters > 0)
367 		nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
368 	uma_zone_set_warning(zone_clust, "kern.ipc.nmbclusters limit reached");
369 	uma_zone_set_maxaction(zone_clust, mb_reclaim);
370 
371 	zone_pack = uma_zsecond_create(MBUF_PACKET_MEM_NAME, mb_ctor_pack,
372 	    mb_dtor_pack, mb_zinit_pack, mb_zfini_pack, zone_mbuf);
373 
374 	/* Make jumbo frame zone too. Page size, 9k and 16k. */
375 	zone_jumbop = uma_zcreate(MBUF_JUMBOP_MEM_NAME, MJUMPAGESIZE,
376 	    mb_ctor_clust, NULL, NULL, NULL,
377 	    UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
378 	if (nmbjumbop > 0)
379 		nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
380 	uma_zone_set_warning(zone_jumbop, "kern.ipc.nmbjumbop limit reached");
381 	uma_zone_set_maxaction(zone_jumbop, mb_reclaim);
382 
383 	zone_jumbo9 = uma_zcreate(MBUF_JUMBO9_MEM_NAME, MJUM9BYTES,
384 	    mb_ctor_clust, NULL, NULL, NULL,
385 	    UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
386 	if (nmbjumbo9 > 0)
387 		nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
388 	uma_zone_set_warning(zone_jumbo9, "kern.ipc.nmbjumbo9 limit reached");
389 	uma_zone_set_maxaction(zone_jumbo9, mb_reclaim);
390 
391 	zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES,
392 	    mb_ctor_clust, NULL, NULL, NULL,
393 	    UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
394 	if (nmbjumbo16 > 0)
395 		nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
396 	uma_zone_set_warning(zone_jumbo16, "kern.ipc.nmbjumbo16 limit reached");
397 	uma_zone_set_maxaction(zone_jumbo16, mb_reclaim);
398 
399 	/*
400 	 * Hook event handler for low-memory situation, used to
401 	 * drain protocols and push data back to the caches (UMA
402 	 * later pushes it back to VM).
403 	 */
404 	EVENTHANDLER_REGISTER(vm_lowmem, mb_reclaim, NULL,
405 	    EVENTHANDLER_PRI_FIRST);
406 
407 	snd_tag_count = counter_u64_alloc(M_WAITOK);
408 }
409 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL);
410 
411 #ifdef DEBUGNET
412 /*
413  * debugnet makes use of a pre-allocated pool of mbufs and clusters.  When
414  * debugnet is configured, we initialize a set of UMA cache zones which return
415  * items from this pool.  At panic-time, the regular UMA zone pointers are
416  * overwritten with those of the cache zones so that drivers may allocate and
417  * free mbufs and clusters without attempting to allocate physical memory.
418  *
419  * We keep mbufs and clusters in a pair of mbuf queues.  In particular, for
420  * the purpose of caching clusters, we treat them as mbufs.
421  */
422 static struct mbufq dn_mbufq =
423     { STAILQ_HEAD_INITIALIZER(dn_mbufq.mq_head), 0, INT_MAX };
424 static struct mbufq dn_clustq =
425     { STAILQ_HEAD_INITIALIZER(dn_clustq.mq_head), 0, INT_MAX };
426 
427 static int dn_clsize;
428 static uma_zone_t dn_zone_mbuf;
429 static uma_zone_t dn_zone_clust;
430 static uma_zone_t dn_zone_pack;
431 
432 static struct debugnet_saved_zones {
433 	uma_zone_t dsz_mbuf;
434 	uma_zone_t dsz_clust;
435 	uma_zone_t dsz_pack;
436 	uma_zone_t dsz_jumbop;
437 	uma_zone_t dsz_jumbo9;
438 	uma_zone_t dsz_jumbo16;
439 	bool dsz_debugnet_zones_enabled;
440 } dn_saved_zones;
441 
442 static int
443 dn_buf_import(void *arg, void **store, int count, int domain __unused,
444     int flags)
445 {
446 	struct mbufq *q;
447 	struct mbuf *m;
448 	int i;
449 
450 	q = arg;
451 
452 	for (i = 0; i < count; i++) {
453 		m = mbufq_dequeue(q);
454 		if (m == NULL)
455 			break;
456 		trash_init(m, q == &dn_mbufq ? MSIZE : dn_clsize, flags);
457 		store[i] = m;
458 	}
459 	KASSERT((flags & M_WAITOK) == 0 || i == count,
460 	    ("%s: ran out of pre-allocated mbufs", __func__));
461 	return (i);
462 }
463 
464 static void
465 dn_buf_release(void *arg, void **store, int count)
466 {
467 	struct mbufq *q;
468 	struct mbuf *m;
469 	int i;
470 
471 	q = arg;
472 
473 	for (i = 0; i < count; i++) {
474 		m = store[i];
475 		(void)mbufq_enqueue(q, m);
476 	}
477 }
478 
479 static int
480 dn_pack_import(void *arg __unused, void **store, int count, int domain __unused,
481     int flags __unused)
482 {
483 	struct mbuf *m;
484 	void *clust;
485 	int i;
486 
487 	for (i = 0; i < count; i++) {
488 		m = m_get(MT_DATA, M_NOWAIT);
489 		if (m == NULL)
490 			break;
491 		clust = uma_zalloc(dn_zone_clust, M_NOWAIT);
492 		if (clust == NULL) {
493 			m_free(m);
494 			break;
495 		}
496 		mb_ctor_clust(clust, dn_clsize, m, 0);
497 		store[i] = m;
498 	}
499 	KASSERT((flags & M_WAITOK) == 0 || i == count,
500 	    ("%s: ran out of pre-allocated mbufs", __func__));
501 	return (i);
502 }
503 
504 static void
505 dn_pack_release(void *arg __unused, void **store, int count)
506 {
507 	struct mbuf *m;
508 	void *clust;
509 	int i;
510 
511 	for (i = 0; i < count; i++) {
512 		m = store[i];
513 		clust = m->m_ext.ext_buf;
514 		uma_zfree(dn_zone_clust, clust);
515 		uma_zfree(dn_zone_mbuf, m);
516 	}
517 }
518 
519 /*
520  * Free the pre-allocated mbufs and clusters reserved for debugnet, and destroy
521  * the corresponding UMA cache zones.
522  */
523 void
524 debugnet_mbuf_drain(void)
525 {
526 	struct mbuf *m;
527 	void *item;
528 
529 	if (dn_zone_mbuf != NULL) {
530 		uma_zdestroy(dn_zone_mbuf);
531 		dn_zone_mbuf = NULL;
532 	}
533 	if (dn_zone_clust != NULL) {
534 		uma_zdestroy(dn_zone_clust);
535 		dn_zone_clust = NULL;
536 	}
537 	if (dn_zone_pack != NULL) {
538 		uma_zdestroy(dn_zone_pack);
539 		dn_zone_pack = NULL;
540 	}
541 
542 	while ((m = mbufq_dequeue(&dn_mbufq)) != NULL)
543 		m_free(m);
544 	while ((item = mbufq_dequeue(&dn_clustq)) != NULL)
545 		uma_zfree(m_getzone(dn_clsize), item);
546 }
547 
548 /*
549  * Callback invoked immediately prior to starting a debugnet connection.
550  */
551 void
552 debugnet_mbuf_start(void)
553 {
554 
555 	MPASS(!dn_saved_zones.dsz_debugnet_zones_enabled);
556 
557 	/* Save the old zone pointers to restore when debugnet is closed. */
558 	dn_saved_zones = (struct debugnet_saved_zones) {
559 		.dsz_debugnet_zones_enabled = true,
560 		.dsz_mbuf = zone_mbuf,
561 		.dsz_clust = zone_clust,
562 		.dsz_pack = zone_pack,
563 		.dsz_jumbop = zone_jumbop,
564 		.dsz_jumbo9 = zone_jumbo9,
565 		.dsz_jumbo16 = zone_jumbo16,
566 	};
567 
568 	/*
569 	 * All cluster zones return buffers of the size requested by the
570 	 * drivers.  It's up to the driver to reinitialize the zones if the
571 	 * MTU of a debugnet-enabled interface changes.
572 	 */
573 	printf("debugnet: overwriting mbuf zone pointers\n");
574 	zone_mbuf = dn_zone_mbuf;
575 	zone_clust = dn_zone_clust;
576 	zone_pack = dn_zone_pack;
577 	zone_jumbop = dn_zone_clust;
578 	zone_jumbo9 = dn_zone_clust;
579 	zone_jumbo16 = dn_zone_clust;
580 }
581 
582 /*
583  * Callback invoked when a debugnet connection is closed/finished.
584  */
585 void
586 debugnet_mbuf_finish(void)
587 {
588 
589 	MPASS(dn_saved_zones.dsz_debugnet_zones_enabled);
590 
591 	printf("debugnet: restoring mbuf zone pointers\n");
592 	zone_mbuf = dn_saved_zones.dsz_mbuf;
593 	zone_clust = dn_saved_zones.dsz_clust;
594 	zone_pack = dn_saved_zones.dsz_pack;
595 	zone_jumbop = dn_saved_zones.dsz_jumbop;
596 	zone_jumbo9 = dn_saved_zones.dsz_jumbo9;
597 	zone_jumbo16 = dn_saved_zones.dsz_jumbo16;
598 
599 	memset(&dn_saved_zones, 0, sizeof(dn_saved_zones));
600 }
601 
602 /*
603  * Reinitialize the debugnet mbuf+cluster pool and cache zones.
604  */
605 void
606 debugnet_mbuf_reinit(int nmbuf, int nclust, int clsize)
607 {
608 	struct mbuf *m;
609 	void *item;
610 
611 	debugnet_mbuf_drain();
612 
613 	dn_clsize = clsize;
614 
615 	dn_zone_mbuf = uma_zcache_create("debugnet_" MBUF_MEM_NAME,
616 	    MSIZE, mb_ctor_mbuf, mb_dtor_mbuf, NULL, NULL,
617 	    dn_buf_import, dn_buf_release,
618 	    &dn_mbufq, UMA_ZONE_NOBUCKET);
619 
620 	dn_zone_clust = uma_zcache_create("debugnet_" MBUF_CLUSTER_MEM_NAME,
621 	    clsize, mb_ctor_clust, NULL, NULL, NULL,
622 	    dn_buf_import, dn_buf_release,
623 	    &dn_clustq, UMA_ZONE_NOBUCKET);
624 
625 	dn_zone_pack = uma_zcache_create("debugnet_" MBUF_PACKET_MEM_NAME,
626 	    MCLBYTES, mb_ctor_pack, mb_dtor_pack, NULL, NULL,
627 	    dn_pack_import, dn_pack_release,
628 	    NULL, UMA_ZONE_NOBUCKET);
629 
630 	while (nmbuf-- > 0) {
631 		m = m_get(MT_DATA, M_WAITOK);
632 		uma_zfree(dn_zone_mbuf, m);
633 	}
634 	while (nclust-- > 0) {
635 		item = uma_zalloc(m_getzone(dn_clsize), M_WAITOK);
636 		uma_zfree(dn_zone_clust, item);
637 	}
638 }
639 #endif /* DEBUGNET */
640 
641 /*
642  * Constructor for Mbuf primary zone.
643  *
644  * The 'arg' pointer points to a mb_args structure which
645  * contains call-specific information required to support the
646  * mbuf allocation API.  See mbuf.h.
647  */
648 static int
649 mb_ctor_mbuf(void *mem, int size, void *arg, int how)
650 {
651 	struct mbuf *m;
652 	struct mb_args *args;
653 	int error;
654 	int flags;
655 	short type;
656 
657 	args = (struct mb_args *)arg;
658 	type = args->type;
659 
660 	/*
661 	 * The mbuf is initialized later.  The caller has the
662 	 * responsibility to set up any MAC labels too.
663 	 */
664 	if (type == MT_NOINIT)
665 		return (0);
666 
667 	m = (struct mbuf *)mem;
668 	flags = args->flags;
669 	MPASS((flags & M_NOFREE) == 0);
670 
671 	error = m_init(m, how, type, flags);
672 
673 	return (error);
674 }
675 
676 /*
677  * The Mbuf primary zone destructor.
678  */
679 static void
680 mb_dtor_mbuf(void *mem, int size, void *arg)
681 {
682 	struct mbuf *m;
683 	unsigned long flags __diagused;
684 
685 	m = (struct mbuf *)mem;
686 	flags = (unsigned long)arg;
687 
688 	KASSERT((m->m_flags & M_NOFREE) == 0, ("%s: M_NOFREE set", __func__));
689 	KASSERT((flags & 0x1) == 0, ("%s: obsolete MB_DTOR_SKIP passed", __func__));
690 	if ((m->m_flags & M_PKTHDR) && !SLIST_EMPTY(&m->m_pkthdr.tags))
691 		m_tag_delete_chain(m, NULL);
692 }
693 
694 /*
695  * The Mbuf Packet zone destructor.
696  */
697 static void
698 mb_dtor_pack(void *mem, int size, void *arg)
699 {
700 	struct mbuf *m;
701 
702 	m = (struct mbuf *)mem;
703 	if ((m->m_flags & M_PKTHDR) != 0)
704 		m_tag_delete_chain(m, NULL);
705 
706 	/* Make sure we've got a clean cluster back. */
707 	KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__));
708 	KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__));
709 	KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__));
710 	KASSERT(m->m_ext.ext_arg1 == NULL, ("%s: ext_arg1 != NULL", __func__));
711 	KASSERT(m->m_ext.ext_arg2 == NULL, ("%s: ext_arg2 != NULL", __func__));
712 	KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__));
713 	KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__));
714 #if defined(INVARIANTS) && !defined(KMSAN)
715 	trash_dtor(m->m_ext.ext_buf, MCLBYTES, arg);
716 #endif
717 	/*
718 	 * If there are processes blocked on zone_clust, waiting for pages
719 	 * to be freed up, cause them to be woken up by draining the
720 	 * packet zone.  We are exposed to a race here (in the check for
721 	 * the UMA_ZFLAG_FULL) where we might miss the flag set, but that
722 	 * is deliberate. We don't want to acquire the zone lock for every
723 	 * mbuf free.
724 	 */
725 	if (uma_zone_exhausted(zone_clust))
726 		uma_zone_reclaim(zone_pack, UMA_RECLAIM_DRAIN);
727 }
728 
729 /*
730  * The Cluster and Jumbo[PAGESIZE|9|16] zone constructor.
731  *
732  * Here the 'arg' pointer points to the Mbuf which we
733  * are configuring cluster storage for.  If 'arg' is
734  * empty we allocate just the cluster without setting
735  * the mbuf to it.  See mbuf.h.
736  */
737 static int
738 mb_ctor_clust(void *mem, int size, void *arg, int how)
739 {
740 	struct mbuf *m;
741 
742 	m = (struct mbuf *)arg;
743 	if (m != NULL) {
744 		m->m_ext.ext_buf = (char *)mem;
745 		m->m_data = m->m_ext.ext_buf;
746 		m->m_flags |= M_EXT;
747 		m->m_ext.ext_free = NULL;
748 		m->m_ext.ext_arg1 = NULL;
749 		m->m_ext.ext_arg2 = NULL;
750 		m->m_ext.ext_size = size;
751 		m->m_ext.ext_type = m_gettype(size);
752 		m->m_ext.ext_flags = EXT_FLAG_EMBREF;
753 		m->m_ext.ext_count = 1;
754 	}
755 
756 	return (0);
757 }
758 
759 /*
760  * The Packet secondary zone's init routine, executed on the
761  * object's transition from mbuf keg slab to zone cache.
762  */
763 static int
764 mb_zinit_pack(void *mem, int size, int how)
765 {
766 	struct mbuf *m;
767 
768 	m = (struct mbuf *)mem;		/* m is virgin. */
769 	if (uma_zalloc_arg(zone_clust, m, how) == NULL ||
770 	    m->m_ext.ext_buf == NULL)
771 		return (ENOMEM);
772 	m->m_ext.ext_type = EXT_PACKET;	/* Override. */
773 #if defined(INVARIANTS) && !defined(KMSAN)
774 	trash_init(m->m_ext.ext_buf, MCLBYTES, how);
775 #endif
776 	return (0);
777 }
778 
779 /*
780  * The Packet secondary zone's fini routine, executed on the
781  * object's transition from zone cache to keg slab.
782  */
783 static void
784 mb_zfini_pack(void *mem, int size)
785 {
786 	struct mbuf *m;
787 
788 	m = (struct mbuf *)mem;
789 #if defined(INVARIANTS) && !defined(KMSAN)
790 	trash_fini(m->m_ext.ext_buf, MCLBYTES);
791 #endif
792 	uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL);
793 #if defined(INVARIANTS) && !defined(KMSAN)
794 	trash_dtor(mem, size, NULL);
795 #endif
796 }
797 
798 /*
799  * The "packet" keg constructor.
800  */
801 static int
802 mb_ctor_pack(void *mem, int size, void *arg, int how)
803 {
804 	struct mbuf *m;
805 	struct mb_args *args;
806 	int error, flags;
807 	short type;
808 
809 	m = (struct mbuf *)mem;
810 	args = (struct mb_args *)arg;
811 	flags = args->flags;
812 	type = args->type;
813 	MPASS((flags & M_NOFREE) == 0);
814 
815 #if defined(INVARIANTS) && !defined(KMSAN)
816 	trash_ctor(m->m_ext.ext_buf, MCLBYTES, arg, how);
817 #endif
818 
819 	error = m_init(m, how, type, flags);
820 
821 	/* m_ext is already initialized. */
822 	m->m_data = m->m_ext.ext_buf;
823  	m->m_flags = (flags | M_EXT);
824 
825 	return (error);
826 }
827 
828 /*
829  * This is the protocol drain routine.  Called by UMA whenever any of the
830  * mbuf zones is closed to its limit.
831  *
832  * No locks should be held when this is called.  The drain routines have to
833  * presently acquire some locks which raises the possibility of lock order
834  * reversal.
835  */
836 static void
837 mb_reclaim(uma_zone_t zone __unused, int pending __unused)
838 {
839 	struct epoch_tracker et;
840 	struct domain *dp;
841 	struct protosw *pr;
842 
843 	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK | WARN_PANIC, NULL, __func__);
844 
845 	NET_EPOCH_ENTER(et);
846 	for (dp = domains; dp != NULL; dp = dp->dom_next)
847 		for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++)
848 			if (pr->pr_drain != NULL)
849 				(*pr->pr_drain)();
850 	NET_EPOCH_EXIT(et);
851 }
852 
853 /*
854  * Free "count" units of I/O from an mbuf chain.  They could be held
855  * in M_EXTPG or just as a normal mbuf.  This code is intended to be
856  * called in an error path (I/O error, closed connection, etc).
857  */
858 void
859 mb_free_notready(struct mbuf *m, int count)
860 {
861 	int i;
862 
863 	for (i = 0; i < count && m != NULL; i++) {
864 		if ((m->m_flags & M_EXTPG) != 0) {
865 			m->m_epg_nrdy--;
866 			if (m->m_epg_nrdy != 0)
867 				continue;
868 		}
869 		m = m_free(m);
870 	}
871 	KASSERT(i == count, ("Removed only %d items from %p", i, m));
872 }
873 
874 /*
875  * Compress an unmapped mbuf into a simple mbuf when it holds a small
876  * amount of data.  This is used as a DOS defense to avoid having
877  * small packets tie up wired pages, an ext_pgs structure, and an
878  * mbuf.  Since this converts the existing mbuf in place, it can only
879  * be used if there are no other references to 'm'.
880  */
881 int
882 mb_unmapped_compress(struct mbuf *m)
883 {
884 	volatile u_int *refcnt;
885 	char buf[MLEN];
886 
887 	/*
888 	 * Assert that 'm' does not have a packet header.  If 'm' had
889 	 * a packet header, it would only be able to hold MHLEN bytes
890 	 * and m_data would have to be initialized differently.
891 	 */
892 	KASSERT((m->m_flags & M_PKTHDR) == 0 && (m->m_flags & M_EXTPG),
893             ("%s: m %p !M_EXTPG or M_PKTHDR", __func__, m));
894 	KASSERT(m->m_len <= MLEN, ("m_len too large %p", m));
895 
896 	if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
897 		refcnt = &m->m_ext.ext_count;
898 	} else {
899 		KASSERT(m->m_ext.ext_cnt != NULL,
900 		    ("%s: no refcounting pointer on %p", __func__, m));
901 		refcnt = m->m_ext.ext_cnt;
902 	}
903 
904 	if (*refcnt != 1)
905 		return (EBUSY);
906 
907 	m_copydata(m, 0, m->m_len, buf);
908 
909 	/* Free the backing pages. */
910 	m->m_ext.ext_free(m);
911 
912 	/* Turn 'm' into a "normal" mbuf. */
913 	m->m_flags &= ~(M_EXT | M_RDONLY | M_EXTPG);
914 	m->m_data = m->m_dat;
915 
916 	/* Copy data back into m. */
917 	bcopy(buf, mtod(m, char *), m->m_len);
918 
919 	return (0);
920 }
921 
922 /*
923  * These next few routines are used to permit downgrading an unmapped
924  * mbuf to a chain of mapped mbufs.  This is used when an interface
925  * doesn't supported unmapped mbufs or if checksums need to be
926  * computed in software.
927  *
928  * Each unmapped mbuf is converted to a chain of mbufs.  First, any
929  * TLS header data is stored in a regular mbuf.  Second, each page of
930  * unmapped data is stored in an mbuf with an EXT_SFBUF external
931  * cluster.  These mbufs use an sf_buf to provide a valid KVA for the
932  * associated physical page.  They also hold a reference on the
933  * original M_EXTPG mbuf to ensure the physical page doesn't go away.
934  * Finally, any TLS trailer data is stored in a regular mbuf.
935  *
936  * mb_unmapped_free_mext() is the ext_free handler for the EXT_SFBUF
937  * mbufs.  It frees the associated sf_buf and releases its reference
938  * on the original M_EXTPG mbuf.
939  *
940  * _mb_unmapped_to_ext() is a helper function that converts a single
941  * unmapped mbuf into a chain of mbufs.
942  *
943  * mb_unmapped_to_ext() is the public function that walks an mbuf
944  * chain converting any unmapped mbufs to mapped mbufs.  It returns
945  * the new chain of unmapped mbufs on success.  On failure it frees
946  * the original mbuf chain and returns NULL.
947  */
948 static void
949 mb_unmapped_free_mext(struct mbuf *m)
950 {
951 	struct sf_buf *sf;
952 	struct mbuf *old_m;
953 
954 	sf = m->m_ext.ext_arg1;
955 	sf_buf_free(sf);
956 
957 	/* Drop the reference on the backing M_EXTPG mbuf. */
958 	old_m = m->m_ext.ext_arg2;
959 	mb_free_extpg(old_m);
960 }
961 
962 static struct mbuf *
963 _mb_unmapped_to_ext(struct mbuf *m)
964 {
965 	struct mbuf *m_new, *top, *prev, *mref;
966 	struct sf_buf *sf;
967 	vm_page_t pg;
968 	int i, len, off, pglen, pgoff, seglen, segoff;
969 	volatile u_int *refcnt;
970 	u_int ref_inc = 0;
971 
972 	M_ASSERTEXTPG(m);
973 	len = m->m_len;
974 	KASSERT(m->m_epg_tls == NULL, ("%s: can't convert TLS mbuf %p",
975 	    __func__, m));
976 
977 	/* See if this is the mbuf that holds the embedded refcount. */
978 	if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
979 		refcnt = &m->m_ext.ext_count;
980 		mref = m;
981 	} else {
982 		KASSERT(m->m_ext.ext_cnt != NULL,
983 		    ("%s: no refcounting pointer on %p", __func__, m));
984 		refcnt = m->m_ext.ext_cnt;
985 		mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
986 	}
987 
988 	/* Skip over any data removed from the front. */
989 	off = mtod(m, vm_offset_t);
990 
991 	top = NULL;
992 	if (m->m_epg_hdrlen != 0) {
993 		if (off >= m->m_epg_hdrlen) {
994 			off -= m->m_epg_hdrlen;
995 		} else {
996 			seglen = m->m_epg_hdrlen - off;
997 			segoff = off;
998 			seglen = min(seglen, len);
999 			off = 0;
1000 			len -= seglen;
1001 			m_new = m_get(M_NOWAIT, MT_DATA);
1002 			if (m_new == NULL)
1003 				goto fail;
1004 			m_new->m_len = seglen;
1005 			prev = top = m_new;
1006 			memcpy(mtod(m_new, void *), &m->m_epg_hdr[segoff],
1007 			    seglen);
1008 		}
1009 	}
1010 	pgoff = m->m_epg_1st_off;
1011 	for (i = 0; i < m->m_epg_npgs && len > 0; i++) {
1012 		pglen = m_epg_pagelen(m, i, pgoff);
1013 		if (off >= pglen) {
1014 			off -= pglen;
1015 			pgoff = 0;
1016 			continue;
1017 		}
1018 		seglen = pglen - off;
1019 		segoff = pgoff + off;
1020 		off = 0;
1021 		seglen = min(seglen, len);
1022 		len -= seglen;
1023 
1024 		pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]);
1025 		m_new = m_get(M_NOWAIT, MT_DATA);
1026 		if (m_new == NULL)
1027 			goto fail;
1028 		if (top == NULL) {
1029 			top = prev = m_new;
1030 		} else {
1031 			prev->m_next = m_new;
1032 			prev = m_new;
1033 		}
1034 		sf = sf_buf_alloc(pg, SFB_NOWAIT);
1035 		if (sf == NULL)
1036 			goto fail;
1037 
1038 		ref_inc++;
1039 		m_extadd(m_new, (char *)sf_buf_kva(sf), PAGE_SIZE,
1040 		    mb_unmapped_free_mext, sf, mref, M_RDONLY, EXT_SFBUF);
1041 		m_new->m_data += segoff;
1042 		m_new->m_len = seglen;
1043 
1044 		pgoff = 0;
1045 	};
1046 	if (len != 0) {
1047 		KASSERT((off + len) <= m->m_epg_trllen,
1048 		    ("off + len > trail (%d + %d > %d)", off, len,
1049 		    m->m_epg_trllen));
1050 		m_new = m_get(M_NOWAIT, MT_DATA);
1051 		if (m_new == NULL)
1052 			goto fail;
1053 		if (top == NULL)
1054 			top = m_new;
1055 		else
1056 			prev->m_next = m_new;
1057 		m_new->m_len = len;
1058 		memcpy(mtod(m_new, void *), &m->m_epg_trail[off], len);
1059 	}
1060 
1061 	if (ref_inc != 0) {
1062 		/*
1063 		 * Obtain an additional reference on the old mbuf for
1064 		 * each created EXT_SFBUF mbuf.  They will be dropped
1065 		 * in mb_unmapped_free_mext().
1066 		 */
1067 		if (*refcnt == 1)
1068 			*refcnt += ref_inc;
1069 		else
1070 			atomic_add_int(refcnt, ref_inc);
1071 	}
1072 	m_free(m);
1073 	return (top);
1074 
1075 fail:
1076 	if (ref_inc != 0) {
1077 		/*
1078 		 * Obtain an additional reference on the old mbuf for
1079 		 * each created EXT_SFBUF mbuf.  They will be
1080 		 * immediately dropped when these mbufs are freed
1081 		 * below.
1082 		 */
1083 		if (*refcnt == 1)
1084 			*refcnt += ref_inc;
1085 		else
1086 			atomic_add_int(refcnt, ref_inc);
1087 	}
1088 	m_free(m);
1089 	m_freem(top);
1090 	return (NULL);
1091 }
1092 
1093 struct mbuf *
1094 mb_unmapped_to_ext(struct mbuf *top)
1095 {
1096 	struct mbuf *m, *next, *prev = NULL;
1097 
1098 	prev = NULL;
1099 	for (m = top; m != NULL; m = next) {
1100 		/* m might be freed, so cache the next pointer. */
1101 		next = m->m_next;
1102 		if (m->m_flags & M_EXTPG) {
1103 			if (prev != NULL) {
1104 				/*
1105 				 * Remove 'm' from the new chain so
1106 				 * that the 'top' chain terminates
1107 				 * before 'm' in case 'top' is freed
1108 				 * due to an error.
1109 				 */
1110 				prev->m_next = NULL;
1111 			}
1112 			m = _mb_unmapped_to_ext(m);
1113 			if (m == NULL) {
1114 				m_freem(top);
1115 				m_freem(next);
1116 				return (NULL);
1117 			}
1118 			if (prev == NULL) {
1119 				top = m;
1120 			} else {
1121 				prev->m_next = m;
1122 			}
1123 
1124 			/*
1125 			 * Replaced one mbuf with a chain, so we must
1126 			 * find the end of chain.
1127 			 */
1128 			prev = m_last(m);
1129 		} else {
1130 			if (prev != NULL) {
1131 				prev->m_next = m;
1132 			}
1133 			prev = m;
1134 		}
1135 	}
1136 	return (top);
1137 }
1138 
1139 /*
1140  * Allocate an empty M_EXTPG mbuf.  The ext_free routine is
1141  * responsible for freeing any pages backing this mbuf when it is
1142  * freed.
1143  */
1144 struct mbuf *
1145 mb_alloc_ext_pgs(int how, m_ext_free_t ext_free)
1146 {
1147 	struct mbuf *m;
1148 
1149 	m = m_get(how, MT_DATA);
1150 	if (m == NULL)
1151 		return (NULL);
1152 
1153 	m->m_epg_npgs = 0;
1154 	m->m_epg_nrdy = 0;
1155 	m->m_epg_1st_off = 0;
1156 	m->m_epg_last_len = 0;
1157 	m->m_epg_flags = 0;
1158 	m->m_epg_hdrlen = 0;
1159 	m->m_epg_trllen = 0;
1160 	m->m_epg_tls = NULL;
1161 	m->m_epg_so = NULL;
1162 	m->m_data = NULL;
1163 	m->m_flags |= (M_EXT | M_RDONLY | M_EXTPG);
1164 	m->m_ext.ext_flags = EXT_FLAG_EMBREF;
1165 	m->m_ext.ext_count = 1;
1166 	m->m_ext.ext_size = 0;
1167 	m->m_ext.ext_free = ext_free;
1168 	return (m);
1169 }
1170 
1171 /*
1172  * Clean up after mbufs with M_EXT storage attached to them if the
1173  * reference count hits 1.
1174  */
1175 void
1176 mb_free_ext(struct mbuf *m)
1177 {
1178 	volatile u_int *refcnt;
1179 	struct mbuf *mref;
1180 	int freembuf;
1181 
1182 	KASSERT(m->m_flags & M_EXT, ("%s: M_EXT not set on %p", __func__, m));
1183 
1184 	/* See if this is the mbuf that holds the embedded refcount. */
1185 	if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
1186 		refcnt = &m->m_ext.ext_count;
1187 		mref = m;
1188 	} else {
1189 		KASSERT(m->m_ext.ext_cnt != NULL,
1190 		    ("%s: no refcounting pointer on %p", __func__, m));
1191 		refcnt = m->m_ext.ext_cnt;
1192 		mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
1193 	}
1194 
1195 	/*
1196 	 * Check if the header is embedded in the cluster.  It is
1197 	 * important that we can't touch any of the mbuf fields
1198 	 * after we have freed the external storage, since mbuf
1199 	 * could have been embedded in it.  For now, the mbufs
1200 	 * embedded into the cluster are always of type EXT_EXTREF,
1201 	 * and for this type we won't free the mref.
1202 	 */
1203 	if (m->m_flags & M_NOFREE) {
1204 		freembuf = 0;
1205 		KASSERT(m->m_ext.ext_type == EXT_EXTREF ||
1206 		    m->m_ext.ext_type == EXT_RXRING,
1207 		    ("%s: no-free mbuf %p has wrong type", __func__, m));
1208 	} else
1209 		freembuf = 1;
1210 
1211 	/* Free attached storage if this mbuf is the only reference to it. */
1212 	if (*refcnt == 1 || atomic_fetchadd_int(refcnt, -1) == 1) {
1213 		switch (m->m_ext.ext_type) {
1214 		case EXT_PACKET:
1215 			/* The packet zone is special. */
1216 			if (*refcnt == 0)
1217 				*refcnt = 1;
1218 			uma_zfree(zone_pack, mref);
1219 			break;
1220 		case EXT_CLUSTER:
1221 			uma_zfree(zone_clust, m->m_ext.ext_buf);
1222 			m_free_raw(mref);
1223 			break;
1224 		case EXT_JUMBOP:
1225 			uma_zfree(zone_jumbop, m->m_ext.ext_buf);
1226 			m_free_raw(mref);
1227 			break;
1228 		case EXT_JUMBO9:
1229 			uma_zfree(zone_jumbo9, m->m_ext.ext_buf);
1230 			m_free_raw(mref);
1231 			break;
1232 		case EXT_JUMBO16:
1233 			uma_zfree(zone_jumbo16, m->m_ext.ext_buf);
1234 			m_free_raw(mref);
1235 			break;
1236 		case EXT_SFBUF:
1237 		case EXT_NET_DRV:
1238 		case EXT_MOD_TYPE:
1239 		case EXT_DISPOSABLE:
1240 			KASSERT(mref->m_ext.ext_free != NULL,
1241 			    ("%s: ext_free not set", __func__));
1242 			mref->m_ext.ext_free(mref);
1243 			m_free_raw(mref);
1244 			break;
1245 		case EXT_EXTREF:
1246 			KASSERT(m->m_ext.ext_free != NULL,
1247 			    ("%s: ext_free not set", __func__));
1248 			m->m_ext.ext_free(m);
1249 			break;
1250 		case EXT_RXRING:
1251 			KASSERT(m->m_ext.ext_free == NULL,
1252 			    ("%s: ext_free is set", __func__));
1253 			break;
1254 		default:
1255 			KASSERT(m->m_ext.ext_type == 0,
1256 			    ("%s: unknown ext_type", __func__));
1257 		}
1258 	}
1259 
1260 	if (freembuf && m != mref)
1261 		m_free_raw(m);
1262 }
1263 
1264 /*
1265  * Clean up after mbufs with M_EXTPG storage attached to them if the
1266  * reference count hits 1.
1267  */
1268 void
1269 mb_free_extpg(struct mbuf *m)
1270 {
1271 	volatile u_int *refcnt;
1272 	struct mbuf *mref;
1273 
1274 	M_ASSERTEXTPG(m);
1275 
1276 	/* See if this is the mbuf that holds the embedded refcount. */
1277 	if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
1278 		refcnt = &m->m_ext.ext_count;
1279 		mref = m;
1280 	} else {
1281 		KASSERT(m->m_ext.ext_cnt != NULL,
1282 		    ("%s: no refcounting pointer on %p", __func__, m));
1283 		refcnt = m->m_ext.ext_cnt;
1284 		mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
1285 	}
1286 
1287 	/* Free attached storage if this mbuf is the only reference to it. */
1288 	if (*refcnt == 1 || atomic_fetchadd_int(refcnt, -1) == 1) {
1289 		KASSERT(mref->m_ext.ext_free != NULL,
1290 		    ("%s: ext_free not set", __func__));
1291 
1292 		mref->m_ext.ext_free(mref);
1293 #ifdef KERN_TLS
1294 		if (mref->m_epg_tls != NULL &&
1295 		    !refcount_release_if_not_last(&mref->m_epg_tls->refcount))
1296 			ktls_enqueue_to_free(mref);
1297 		else
1298 #endif
1299 			m_free_raw(mref);
1300 	}
1301 
1302 	if (m != mref)
1303 		m_free_raw(m);
1304 }
1305 
1306 /*
1307  * Official mbuf(9) allocation KPI for stack and drivers:
1308  *
1309  * m_get()	- a single mbuf without any attachments, sys/mbuf.h.
1310  * m_gethdr()	- a single mbuf initialized as M_PKTHDR, sys/mbuf.h.
1311  * m_getcl()	- an mbuf + 2k cluster, sys/mbuf.h.
1312  * m_clget()	- attach cluster to already allocated mbuf.
1313  * m_cljget()	- attach jumbo cluster to already allocated mbuf.
1314  * m_get2()	- allocate minimum mbuf that would fit size argument.
1315  * m_getm2()	- allocate a chain of mbufs/clusters.
1316  * m_extadd()	- attach external cluster to mbuf.
1317  *
1318  * m_free()	- free single mbuf with its tags and ext, sys/mbuf.h.
1319  * m_freem()	- free chain of mbufs.
1320  */
1321 
1322 int
1323 m_clget(struct mbuf *m, int how)
1324 {
1325 
1326 	KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
1327 	    __func__, m));
1328 	m->m_ext.ext_buf = (char *)NULL;
1329 	uma_zalloc_arg(zone_clust, m, how);
1330 	/*
1331 	 * On a cluster allocation failure, drain the packet zone and retry,
1332 	 * we might be able to loosen a few clusters up on the drain.
1333 	 */
1334 	if ((how & M_NOWAIT) && (m->m_ext.ext_buf == NULL)) {
1335 		uma_zone_reclaim(zone_pack, UMA_RECLAIM_DRAIN);
1336 		uma_zalloc_arg(zone_clust, m, how);
1337 	}
1338 	MBUF_PROBE2(m__clget, m, how);
1339 	return (m->m_flags & M_EXT);
1340 }
1341 
1342 /*
1343  * m_cljget() is different from m_clget() as it can allocate clusters without
1344  * attaching them to an mbuf.  In that case the return value is the pointer
1345  * to the cluster of the requested size.  If an mbuf was specified, it gets
1346  * the cluster attached to it and the return value can be safely ignored.
1347  * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
1348  */
1349 void *
1350 m_cljget(struct mbuf *m, int how, int size)
1351 {
1352 	uma_zone_t zone;
1353 	void *retval;
1354 
1355 	if (m != NULL) {
1356 		KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
1357 		    __func__, m));
1358 		m->m_ext.ext_buf = NULL;
1359 	}
1360 
1361 	zone = m_getzone(size);
1362 	retval = uma_zalloc_arg(zone, m, how);
1363 
1364 	MBUF_PROBE4(m__cljget, m, how, size, retval);
1365 
1366 	return (retval);
1367 }
1368 
1369 /*
1370  * m_get2() allocates minimum mbuf that would fit "size" argument.
1371  */
1372 struct mbuf *
1373 m_get2(int size, int how, short type, int flags)
1374 {
1375 	struct mb_args args;
1376 	struct mbuf *m, *n;
1377 
1378 	args.flags = flags;
1379 	args.type = type;
1380 
1381 	if (size <= MHLEN || (size <= MLEN && (flags & M_PKTHDR) == 0))
1382 		return (uma_zalloc_arg(zone_mbuf, &args, how));
1383 	if (size <= MCLBYTES)
1384 		return (uma_zalloc_arg(zone_pack, &args, how));
1385 
1386 	if (size > MJUMPAGESIZE)
1387 		return (NULL);
1388 
1389 	m = uma_zalloc_arg(zone_mbuf, &args, how);
1390 	if (m == NULL)
1391 		return (NULL);
1392 
1393 	n = uma_zalloc_arg(zone_jumbop, m, how);
1394 	if (n == NULL) {
1395 		m_free_raw(m);
1396 		return (NULL);
1397 	}
1398 
1399 	return (m);
1400 }
1401 
1402 /*
1403  * m_get3() allocates minimum mbuf that would fit "size" argument.
1404  * Unlike m_get2() it can allocate clusters up to MJUM16BYTES.
1405  */
1406 struct mbuf *
1407 m_get3(int size, int how, short type, int flags)
1408 {
1409 	struct mb_args args;
1410 	struct mbuf *m, *n;
1411 	uma_zone_t zone;
1412 
1413 	if (size <= MJUMPAGESIZE)
1414 		return (m_get2(size, how, type, flags));
1415 
1416 	if (size > MJUM16BYTES)
1417 		return (NULL);
1418 
1419 	args.flags = flags;
1420 	args.type = type;
1421 
1422 	m = uma_zalloc_arg(zone_mbuf, &args, how);
1423 	if (m == NULL)
1424 		return (NULL);
1425 
1426 	if (size <= MJUM9BYTES)
1427 		zone = zone_jumbo9;
1428 	else
1429 		zone = zone_jumbo16;
1430 
1431 	n = uma_zalloc_arg(zone, m, how);
1432 	if (n == NULL) {
1433 		m_free_raw(m);
1434 		return (NULL);
1435 	}
1436 
1437 	return (m);
1438 }
1439 
1440 /*
1441  * m_getjcl() returns an mbuf with a cluster of the specified size attached.
1442  * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
1443  */
1444 struct mbuf *
1445 m_getjcl(int how, short type, int flags, int size)
1446 {
1447 	struct mb_args args;
1448 	struct mbuf *m, *n;
1449 	uma_zone_t zone;
1450 
1451 	if (size == MCLBYTES)
1452 		return m_getcl(how, type, flags);
1453 
1454 	args.flags = flags;
1455 	args.type = type;
1456 
1457 	m = uma_zalloc_arg(zone_mbuf, &args, how);
1458 	if (m == NULL)
1459 		return (NULL);
1460 
1461 	zone = m_getzone(size);
1462 	n = uma_zalloc_arg(zone, m, how);
1463 	if (n == NULL) {
1464 		m_free_raw(m);
1465 		return (NULL);
1466 	}
1467 	MBUF_PROBE5(m__getjcl, how, type, flags, size, m);
1468 	return (m);
1469 }
1470 
1471 /*
1472  * Allocate a given length worth of mbufs and/or clusters (whatever fits
1473  * best) and return a pointer to the top of the allocated chain.  If an
1474  * existing mbuf chain is provided, then we will append the new chain
1475  * to the existing one and return a pointer to the provided mbuf.
1476  */
1477 struct mbuf *
1478 m_getm2(struct mbuf *m, int len, int how, short type, int flags)
1479 {
1480 	struct mbuf *mb, *nm = NULL, *mtail = NULL;
1481 
1482 	KASSERT(len >= 0, ("%s: len is < 0", __func__));
1483 
1484 	/* Validate flags. */
1485 	flags &= (M_PKTHDR | M_EOR);
1486 
1487 	/* Packet header mbuf must be first in chain. */
1488 	if ((flags & M_PKTHDR) && m != NULL)
1489 		flags &= ~M_PKTHDR;
1490 
1491 	/* Loop and append maximum sized mbufs to the chain tail. */
1492 	while (len > 0) {
1493 		mb = NULL;
1494 		if (len > MCLBYTES) {
1495 			mb = m_getjcl(M_NOWAIT, type, (flags & M_PKTHDR),
1496 			    MJUMPAGESIZE);
1497 		}
1498 		if (mb == NULL) {
1499 			if (len >= MINCLSIZE)
1500 				mb = m_getcl(how, type, (flags & M_PKTHDR));
1501 			else if (flags & M_PKTHDR)
1502 				mb = m_gethdr(how, type);
1503 			else
1504 				mb = m_get(how, type);
1505 
1506 			/*
1507 			 * Fail the whole operation if one mbuf can't be
1508 			 * allocated.
1509 			 */
1510 			if (mb == NULL) {
1511 				m_freem(nm);
1512 				return (NULL);
1513 			}
1514 		}
1515 
1516 		/* Book keeping. */
1517 		len -= M_SIZE(mb);
1518 		if (mtail != NULL)
1519 			mtail->m_next = mb;
1520 		else
1521 			nm = mb;
1522 		mtail = mb;
1523 		flags &= ~M_PKTHDR;	/* Only valid on the first mbuf. */
1524 	}
1525 	if (flags & M_EOR)
1526 		mtail->m_flags |= M_EOR;  /* Only valid on the last mbuf. */
1527 
1528 	/* If mbuf was supplied, append new chain to the end of it. */
1529 	if (m != NULL) {
1530 		for (mtail = m; mtail->m_next != NULL; mtail = mtail->m_next)
1531 			;
1532 		mtail->m_next = nm;
1533 		mtail->m_flags &= ~M_EOR;
1534 	} else
1535 		m = nm;
1536 
1537 	return (m);
1538 }
1539 
1540 /*-
1541  * Configure a provided mbuf to refer to the provided external storage
1542  * buffer and setup a reference count for said buffer.
1543  *
1544  * Arguments:
1545  *    mb     The existing mbuf to which to attach the provided buffer.
1546  *    buf    The address of the provided external storage buffer.
1547  *    size   The size of the provided buffer.
1548  *    freef  A pointer to a routine that is responsible for freeing the
1549  *           provided external storage buffer.
1550  *    args   A pointer to an argument structure (of any type) to be passed
1551  *           to the provided freef routine (may be NULL).
1552  *    flags  Any other flags to be passed to the provided mbuf.
1553  *    type   The type that the external storage buffer should be
1554  *           labeled with.
1555  *
1556  * Returns:
1557  *    Nothing.
1558  */
1559 void
1560 m_extadd(struct mbuf *mb, char *buf, u_int size, m_ext_free_t freef,
1561     void *arg1, void *arg2, int flags, int type)
1562 {
1563 
1564 	KASSERT(type != EXT_CLUSTER, ("%s: EXT_CLUSTER not allowed", __func__));
1565 
1566 	mb->m_flags |= (M_EXT | flags);
1567 	mb->m_ext.ext_buf = buf;
1568 	mb->m_data = mb->m_ext.ext_buf;
1569 	mb->m_ext.ext_size = size;
1570 	mb->m_ext.ext_free = freef;
1571 	mb->m_ext.ext_arg1 = arg1;
1572 	mb->m_ext.ext_arg2 = arg2;
1573 	mb->m_ext.ext_type = type;
1574 
1575 	if (type != EXT_EXTREF) {
1576 		mb->m_ext.ext_count = 1;
1577 		mb->m_ext.ext_flags = EXT_FLAG_EMBREF;
1578 	} else
1579 		mb->m_ext.ext_flags = 0;
1580 }
1581 
1582 /*
1583  * Free an entire chain of mbufs and associated external buffers, if
1584  * applicable.
1585  */
1586 void
1587 m_freem(struct mbuf *mb)
1588 {
1589 
1590 	MBUF_PROBE1(m__freem, mb);
1591 	while (mb != NULL)
1592 		mb = m_free(mb);
1593 }
1594 
1595 /*
1596  * Temporary primitive to allow freeing without going through m_free.
1597  */
1598 void
1599 m_free_raw(struct mbuf *mb)
1600 {
1601 
1602 	uma_zfree(zone_mbuf, mb);
1603 }
1604 
1605 int
1606 m_snd_tag_alloc(struct ifnet *ifp, union if_snd_tag_alloc_params *params,
1607     struct m_snd_tag **mstp)
1608 {
1609 
1610 	if (ifp->if_snd_tag_alloc == NULL)
1611 		return (EOPNOTSUPP);
1612 	return (ifp->if_snd_tag_alloc(ifp, params, mstp));
1613 }
1614 
1615 void
1616 m_snd_tag_init(struct m_snd_tag *mst, struct ifnet *ifp,
1617     const struct if_snd_tag_sw *sw)
1618 {
1619 
1620 	if_ref(ifp);
1621 	mst->ifp = ifp;
1622 	refcount_init(&mst->refcount, 1);
1623 	mst->sw = sw;
1624 	counter_u64_add(snd_tag_count, 1);
1625 }
1626 
1627 void
1628 m_snd_tag_destroy(struct m_snd_tag *mst)
1629 {
1630 	struct ifnet *ifp;
1631 
1632 	ifp = mst->ifp;
1633 	mst->sw->snd_tag_free(mst);
1634 	if_rele(ifp);
1635 	counter_u64_add(snd_tag_count, -1);
1636 }
1637 
1638 void
1639 m_rcvif_serialize(struct mbuf *m)
1640 {
1641 	u_short idx, gen;
1642 
1643 	M_ASSERTPKTHDR(m);
1644 	idx = m->m_pkthdr.rcvif->if_index;
1645 	gen = m->m_pkthdr.rcvif->if_idxgen;
1646 	m->m_pkthdr.rcvidx = idx;
1647 	m->m_pkthdr.rcvgen = gen;
1648 	if (__predict_false(m->m_pkthdr.leaf_rcvif != NULL)) {
1649 		idx = m->m_pkthdr.leaf_rcvif->if_index;
1650 		gen = m->m_pkthdr.leaf_rcvif->if_idxgen;
1651 	} else {
1652 		idx = -1;
1653 		gen = 0;
1654 	}
1655 	m->m_pkthdr.leaf_rcvidx = idx;
1656 	m->m_pkthdr.leaf_rcvgen = gen;
1657 }
1658 
1659 struct ifnet *
1660 m_rcvif_restore(struct mbuf *m)
1661 {
1662 	struct ifnet *ifp, *leaf_ifp;
1663 
1664 	M_ASSERTPKTHDR(m);
1665 	NET_EPOCH_ASSERT();
1666 
1667 	ifp = ifnet_byindexgen(m->m_pkthdr.rcvidx, m->m_pkthdr.rcvgen);
1668 	if (ifp == NULL || (ifp->if_flags & IFF_DYING))
1669 		return (NULL);
1670 
1671 	if (__predict_true(m->m_pkthdr.leaf_rcvidx == (u_short)-1)) {
1672 		leaf_ifp = NULL;
1673 	} else {
1674 		leaf_ifp = ifnet_byindexgen(m->m_pkthdr.leaf_rcvidx,
1675 		    m->m_pkthdr.leaf_rcvgen);
1676 		if (__predict_false(leaf_ifp != NULL && (leaf_ifp->if_flags & IFF_DYING)))
1677 			leaf_ifp = NULL;
1678 	}
1679 
1680 	m->m_pkthdr.leaf_rcvif = leaf_ifp;
1681 	m->m_pkthdr.rcvif = ifp;
1682 
1683 	return (ifp);
1684 }
1685 
1686 /*
1687  * Allocate an mbuf with anonymous external pages.
1688  */
1689 struct mbuf *
1690 mb_alloc_ext_plus_pages(int len, int how)
1691 {
1692 	struct mbuf *m;
1693 	vm_page_t pg;
1694 	int i, npgs;
1695 
1696 	m = mb_alloc_ext_pgs(how, mb_free_mext_pgs);
1697 	if (m == NULL)
1698 		return (NULL);
1699 	m->m_epg_flags |= EPG_FLAG_ANON;
1700 	npgs = howmany(len, PAGE_SIZE);
1701 	for (i = 0; i < npgs; i++) {
1702 		do {
1703 			pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP |
1704 			    VM_ALLOC_WIRED);
1705 			if (pg == NULL) {
1706 				if (how == M_NOWAIT) {
1707 					m->m_epg_npgs = i;
1708 					m_free(m);
1709 					return (NULL);
1710 				}
1711 				vm_wait(NULL);
1712 			}
1713 		} while (pg == NULL);
1714 		m->m_epg_pa[i] = VM_PAGE_TO_PHYS(pg);
1715 	}
1716 	m->m_epg_npgs = npgs;
1717 	return (m);
1718 }
1719 
1720 /*
1721  * Copy the data in the mbuf chain to a chain of mbufs with anonymous external
1722  * unmapped pages.
1723  * len is the length of data in the input mbuf chain.
1724  * mlen is the maximum number of bytes put into each ext_page mbuf.
1725  */
1726 struct mbuf *
1727 mb_mapped_to_unmapped(struct mbuf *mp, int len, int mlen, int how,
1728     struct mbuf **mlast)
1729 {
1730 	struct mbuf *m, *mout;
1731 	char *pgpos, *mbpos;
1732 	int i, mblen, mbufsiz, pglen, xfer;
1733 
1734 	if (len == 0)
1735 		return (NULL);
1736 	mbufsiz = min(mlen, len);
1737 	m = mout = mb_alloc_ext_plus_pages(mbufsiz, how);
1738 	if (m == NULL)
1739 		return (m);
1740 	pgpos = (char *)(void *)PHYS_TO_DMAP(m->m_epg_pa[0]);
1741 	pglen = PAGE_SIZE;
1742 	mblen = 0;
1743 	i = 0;
1744 	do {
1745 		if (pglen == 0) {
1746 			if (++i == m->m_epg_npgs) {
1747 				m->m_epg_last_len = PAGE_SIZE;
1748 				mbufsiz = min(mlen, len);
1749 				m->m_next = mb_alloc_ext_plus_pages(mbufsiz,
1750 				    how);
1751 				m = m->m_next;
1752 				if (m == NULL) {
1753 					m_freem(mout);
1754 					return (m);
1755 				}
1756 				i = 0;
1757 			}
1758 			pgpos = (char *)(void *)PHYS_TO_DMAP(m->m_epg_pa[i]);
1759 			pglen = PAGE_SIZE;
1760 		}
1761 		while (mblen == 0) {
1762 			if (mp == NULL) {
1763 				m_freem(mout);
1764 				return (NULL);
1765 			}
1766 			KASSERT((mp->m_flags & M_EXTPG) == 0,
1767 			    ("mb_copym_ext_pgs: ext_pgs input mbuf"));
1768 			mbpos = mtod(mp, char *);
1769 			mblen = mp->m_len;
1770 			mp = mp->m_next;
1771 		}
1772 		xfer = min(mblen, pglen);
1773 		memcpy(pgpos, mbpos, xfer);
1774 		pgpos += xfer;
1775 		mbpos += xfer;
1776 		pglen -= xfer;
1777 		mblen -= xfer;
1778 		len -= xfer;
1779 		m->m_len += xfer;
1780 	} while (len > 0);
1781 	m->m_epg_last_len = PAGE_SIZE - pglen;
1782 	if (mlast != NULL)
1783 		*mlast = m;
1784 	return (mout);
1785 }
1786