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
35#include <sys/param.h>
36#include <sys/domainset.h>
37#include <sys/malloc.h>
38#include <sys/types.h>
39#include <sys/systm.h>
40#include <sys/mbuf.h>
41#include <sys/domain.h>
42#include <sys/eventhandler.h>
43#include <sys/kernel.h>
44#include <sys/limits.h>
45#include <sys/lock.h>
46#include <sys/mutex.h>
47#include <sys/protosw.h>
48#include <sys/smp.h>
49#include <sys/sysctl.h>
50
51#include <vm/vm.h>
52#include <vm/vm_extern.h>
53#include <vm/vm_kern.h>
54#include <vm/vm_page.h>
55#include <vm/vm_map.h>
56#include <vm/uma.h>
57#include <vm/uma_dbg.h>
58
59/*
60 * In FreeBSD, Mbufs and Mbuf Clusters are allocated from UMA
61 * Zones.
62 *
63 * Mbuf Clusters (2K, contiguous) are allocated from the Cluster
64 * Zone.  The Zone can be capped at kern.ipc.nmbclusters, if the
65 * administrator so desires.
66 *
67 * Mbufs are allocated from a UMA Master Zone called the Mbuf
68 * Zone.
69 *
70 * Additionally, FreeBSD provides a Packet Zone, which it
71 * configures as a Secondary Zone to the Mbuf Master Zone,
72 * thus sharing backend Slab kegs with the Mbuf Master Zone.
73 *
74 * Thus common-case allocations and locking are simplified:
75 *
76 *  m_clget()                m_getcl()
77 *    |                         |
78 *    |   .------------>[(Packet Cache)]    m_get(), m_gethdr()
79 *    |   |             [     Packet   ]            |
80 *  [(Cluster Cache)]   [    Secondary ]   [ (Mbuf Cache)     ]
81 *  [ Cluster Zone  ]   [     Zone     ]   [ Mbuf Master Zone ]
82 *        |                       \________         |
83 *  [ Cluster Keg   ]                      \       /
84 *        |	                         [ Mbuf Keg   ]
85 *  [ Cluster Slabs ]                         |
86 *        |                              [ Mbuf Slabs ]
87 *         \____________(VM)_________________/
88 *
89 *
90 * Whenever an object is allocated with uma_zalloc() out of
91 * one of the Zones its _ctor_ function is executed.  The same
92 * for any deallocation through uma_zfree() the _dtor_ function
93 * is executed.
94 *
95 * Caches are per-CPU and are filled from the Master Zone.
96 *
97 * Whenever an object is allocated from the underlying global
98 * memory pool it gets pre-initialized with the _zinit_ functions.
99 * When the Keg's are overfull objects get decommissioned with
100 * _zfini_ functions and free'd back to the global memory pool.
101 *
102 */
103
104int nmbufs;			/* limits number of mbufs */
105int nmbclusters;		/* limits number of mbuf clusters */
106int nmbjumbop;			/* limits number of page size jumbo clusters */
107int nmbjumbo9;			/* limits number of 9k jumbo clusters */
108int nmbjumbo16;			/* limits number of 16k jumbo clusters */
109
110static quad_t maxmbufmem;	/* overall real memory limit for all mbufs */
111
112SYSCTL_QUAD(_kern_ipc, OID_AUTO, maxmbufmem, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &maxmbufmem, 0,
113    "Maximum real memory allocatable to various mbuf types");
114
115/*
116 * tunable_mbinit() has to be run before any mbuf allocations are done.
117 */
118static void
119tunable_mbinit(void *dummy)
120{
121	quad_t realmem;
122
123	/*
124	 * The default limit for all mbuf related memory is 1/2 of all
125	 * available kernel memory (physical or kmem).
126	 * At most it can be 3/4 of available kernel memory.
127	 */
128	realmem = qmin((quad_t)physmem * PAGE_SIZE, vm_kmem_size);
129	maxmbufmem = realmem / 2;
130	TUNABLE_QUAD_FETCH("kern.ipc.maxmbufmem", &maxmbufmem);
131	if (maxmbufmem > realmem / 4 * 3)
132		maxmbufmem = realmem / 4 * 3;
133
134	TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
135	if (nmbclusters == 0)
136		nmbclusters = maxmbufmem / MCLBYTES / 4;
137
138	TUNABLE_INT_FETCH("kern.ipc.nmbjumbop", &nmbjumbop);
139	if (nmbjumbop == 0)
140		nmbjumbop = maxmbufmem / MJUMPAGESIZE / 4;
141
142	TUNABLE_INT_FETCH("kern.ipc.nmbjumbo9", &nmbjumbo9);
143	if (nmbjumbo9 == 0)
144		nmbjumbo9 = maxmbufmem / MJUM9BYTES / 6;
145
146	TUNABLE_INT_FETCH("kern.ipc.nmbjumbo16", &nmbjumbo16);
147	if (nmbjumbo16 == 0)
148		nmbjumbo16 = maxmbufmem / MJUM16BYTES / 6;
149
150	/*
151	 * We need at least as many mbufs as we have clusters of
152	 * the various types added together.
153	 */
154	TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
155	if (nmbufs < nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16)
156		nmbufs = lmax(maxmbufmem / MSIZE / 5,
157		    nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16);
158}
159SYSINIT(tunable_mbinit, SI_SUB_KMEM, SI_ORDER_MIDDLE, tunable_mbinit, NULL);
160
161static int
162sysctl_nmbclusters(SYSCTL_HANDLER_ARGS)
163{
164	int error, newnmbclusters;
165
166	newnmbclusters = nmbclusters;
167	error = sysctl_handle_int(oidp, &newnmbclusters, 0, req);
168	if (error == 0 && req->newptr && newnmbclusters != nmbclusters) {
169		if (newnmbclusters > nmbclusters &&
170		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
171			nmbclusters = newnmbclusters;
172			nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
173			EVENTHANDLER_INVOKE(nmbclusters_change);
174		} else
175			error = EINVAL;
176	}
177	return (error);
178}
179SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbclusters, CTLTYPE_INT|CTLFLAG_RW,
180&nmbclusters, 0, sysctl_nmbclusters, "IU",
181    "Maximum number of mbuf clusters allowed");
182
183static int
184sysctl_nmbjumbop(SYSCTL_HANDLER_ARGS)
185{
186	int error, newnmbjumbop;
187
188	newnmbjumbop = nmbjumbop;
189	error = sysctl_handle_int(oidp, &newnmbjumbop, 0, req);
190	if (error == 0 && req->newptr && newnmbjumbop != nmbjumbop) {
191		if (newnmbjumbop > nmbjumbop &&
192		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
193			nmbjumbop = newnmbjumbop;
194			nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
195		} else
196			error = EINVAL;
197	}
198	return (error);
199}
200SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbop, CTLTYPE_INT|CTLFLAG_RW,
201&nmbjumbop, 0, sysctl_nmbjumbop, "IU",
202    "Maximum number of mbuf page size jumbo clusters allowed");
203
204static int
205sysctl_nmbjumbo9(SYSCTL_HANDLER_ARGS)
206{
207	int error, newnmbjumbo9;
208
209	newnmbjumbo9 = nmbjumbo9;
210	error = sysctl_handle_int(oidp, &newnmbjumbo9, 0, req);
211	if (error == 0 && req->newptr && newnmbjumbo9 != nmbjumbo9) {
212		if (newnmbjumbo9 > nmbjumbo9 &&
213		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
214			nmbjumbo9 = newnmbjumbo9;
215			nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
216		} else
217			error = EINVAL;
218	}
219	return (error);
220}
221SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo9, CTLTYPE_INT|CTLFLAG_RW,
222&nmbjumbo9, 0, sysctl_nmbjumbo9, "IU",
223    "Maximum number of mbuf 9k jumbo clusters allowed");
224
225static int
226sysctl_nmbjumbo16(SYSCTL_HANDLER_ARGS)
227{
228	int error, newnmbjumbo16;
229
230	newnmbjumbo16 = nmbjumbo16;
231	error = sysctl_handle_int(oidp, &newnmbjumbo16, 0, req);
232	if (error == 0 && req->newptr && newnmbjumbo16 != nmbjumbo16) {
233		if (newnmbjumbo16 > nmbjumbo16 &&
234		    nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
235			nmbjumbo16 = newnmbjumbo16;
236			nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
237		} else
238			error = EINVAL;
239	}
240	return (error);
241}
242SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo16, CTLTYPE_INT|CTLFLAG_RW,
243&nmbjumbo16, 0, sysctl_nmbjumbo16, "IU",
244    "Maximum number of mbuf 16k jumbo clusters allowed");
245
246static int
247sysctl_nmbufs(SYSCTL_HANDLER_ARGS)
248{
249	int error, newnmbufs;
250
251	newnmbufs = nmbufs;
252	error = sysctl_handle_int(oidp, &newnmbufs, 0, req);
253	if (error == 0 && req->newptr && newnmbufs != nmbufs) {
254		if (newnmbufs > nmbufs) {
255			nmbufs = newnmbufs;
256			nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
257			EVENTHANDLER_INVOKE(nmbufs_change);
258		} else
259			error = EINVAL;
260	}
261	return (error);
262}
263SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbufs, CTLTYPE_INT|CTLFLAG_RW,
264&nmbufs, 0, sysctl_nmbufs, "IU",
265    "Maximum number of mbufs allowed");
266
267/*
268 * Zones from which we allocate.
269 */
270uma_zone_t	zone_mbuf;
271uma_zone_t	zone_clust;
272uma_zone_t	zone_pack;
273uma_zone_t	zone_jumbop;
274uma_zone_t	zone_jumbo9;
275uma_zone_t	zone_jumbo16;
276
277/*
278 * Local prototypes.
279 */
280static int	mb_ctor_mbuf(void *, int, void *, int);
281static int	mb_ctor_clust(void *, int, void *, int);
282static int	mb_ctor_pack(void *, int, void *, int);
283static void	mb_dtor_mbuf(void *, int, void *);
284static void	mb_dtor_pack(void *, int, void *);
285static int	mb_zinit_pack(void *, int, int);
286static void	mb_zfini_pack(void *, int);
287static void	mb_reclaim(uma_zone_t, int);
288static void    *mbuf_jumbo_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
289
290/* Ensure that MSIZE is a power of 2. */
291CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE);
292
293/*
294 * Initialize FreeBSD Network buffer allocation.
295 */
296static void
297mbuf_init(void *dummy)
298{
299
300	/*
301	 * Configure UMA zones for Mbufs, Clusters, and Packets.
302	 */
303	zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE,
304	    mb_ctor_mbuf, mb_dtor_mbuf,
305#ifdef INVARIANTS
306	    trash_init, trash_fini,
307#else
308	    NULL, NULL,
309#endif
310	    MSIZE - 1, UMA_ZONE_MAXBUCKET);
311	if (nmbufs > 0)
312		nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
313	uma_zone_set_warning(zone_mbuf, "kern.ipc.nmbufs limit reached");
314	uma_zone_set_maxaction(zone_mbuf, mb_reclaim);
315
316	zone_clust = uma_zcreate(MBUF_CLUSTER_MEM_NAME, MCLBYTES,
317	    mb_ctor_clust,
318#ifdef INVARIANTS
319	    trash_dtor, trash_init, trash_fini,
320#else
321	    NULL, NULL, NULL,
322#endif
323	    UMA_ALIGN_PTR, 0);
324	if (nmbclusters > 0)
325		nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
326	uma_zone_set_warning(zone_clust, "kern.ipc.nmbclusters limit reached");
327	uma_zone_set_maxaction(zone_clust, mb_reclaim);
328
329	zone_pack = uma_zsecond_create(MBUF_PACKET_MEM_NAME, mb_ctor_pack,
330	    mb_dtor_pack, mb_zinit_pack, mb_zfini_pack, zone_mbuf);
331
332	/* Make jumbo frame zone too. Page size, 9k and 16k. */
333	zone_jumbop = uma_zcreate(MBUF_JUMBOP_MEM_NAME, MJUMPAGESIZE,
334	    mb_ctor_clust,
335#ifdef INVARIANTS
336	    trash_dtor, trash_init, trash_fini,
337#else
338	    NULL, NULL, NULL,
339#endif
340	    UMA_ALIGN_PTR, 0);
341	if (nmbjumbop > 0)
342		nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
343	uma_zone_set_warning(zone_jumbop, "kern.ipc.nmbjumbop limit reached");
344	uma_zone_set_maxaction(zone_jumbop, mb_reclaim);
345
346	zone_jumbo9 = uma_zcreate(MBUF_JUMBO9_MEM_NAME, MJUM9BYTES,
347	    mb_ctor_clust,
348#ifdef INVARIANTS
349	    trash_dtor, trash_init, trash_fini,
350#else
351	    NULL, NULL, NULL,
352#endif
353	    UMA_ALIGN_PTR, 0);
354	uma_zone_set_allocf(zone_jumbo9, mbuf_jumbo_alloc);
355	if (nmbjumbo9 > 0)
356		nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
357	uma_zone_set_warning(zone_jumbo9, "kern.ipc.nmbjumbo9 limit reached");
358	uma_zone_set_maxaction(zone_jumbo9, mb_reclaim);
359
360	zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES,
361	    mb_ctor_clust,
362#ifdef INVARIANTS
363	    trash_dtor, trash_init, trash_fini,
364#else
365	    NULL, NULL, NULL,
366#endif
367	    UMA_ALIGN_PTR, 0);
368	uma_zone_set_allocf(zone_jumbo16, mbuf_jumbo_alloc);
369	if (nmbjumbo16 > 0)
370		nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
371	uma_zone_set_warning(zone_jumbo16, "kern.ipc.nmbjumbo16 limit reached");
372	uma_zone_set_maxaction(zone_jumbo16, mb_reclaim);
373
374	/*
375	 * Hook event handler for low-memory situation, used to
376	 * drain protocols and push data back to the caches (UMA
377	 * later pushes it back to VM).
378	 */
379	EVENTHANDLER_REGISTER(vm_lowmem, mb_reclaim, NULL,
380	    EVENTHANDLER_PRI_FIRST);
381}
382SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL);
383
384#ifdef NETDUMP
385/*
386 * netdump makes use of a pre-allocated pool of mbufs and clusters.  When
387 * netdump is configured, we initialize a set of UMA cache zones which return
388 * items from this pool.  At panic-time, the regular UMA zone pointers are
389 * overwritten with those of the cache zones so that drivers may allocate and
390 * free mbufs and clusters without attempting to allocate physical memory.
391 *
392 * We keep mbufs and clusters in a pair of mbuf queues.  In particular, for
393 * the purpose of caching clusters, we treat them as mbufs.
394 */
395static struct mbufq nd_mbufq =
396    { STAILQ_HEAD_INITIALIZER(nd_mbufq.mq_head), 0, INT_MAX };
397static struct mbufq nd_clustq =
398    { STAILQ_HEAD_INITIALIZER(nd_clustq.mq_head), 0, INT_MAX };
399
400static int nd_clsize;
401static uma_zone_t nd_zone_mbuf;
402static uma_zone_t nd_zone_clust;
403static uma_zone_t nd_zone_pack;
404
405static int
406nd_buf_import(void *arg, void **store, int count, int domain __unused,
407    int flags)
408{
409	struct mbufq *q;
410	struct mbuf *m;
411	int i;
412
413	q = arg;
414
415	for (i = 0; i < count; i++) {
416		m = mbufq_dequeue(q);
417		if (m == NULL)
418			break;
419		trash_init(m, q == &nd_mbufq ? MSIZE : nd_clsize, flags);
420		store[i] = m;
421	}
422	return (i);
423}
424
425static void
426nd_buf_release(void *arg, void **store, int count)
427{
428	struct mbufq *q;
429	struct mbuf *m;
430	int i;
431
432	q = arg;
433
434	for (i = 0; i < count; i++) {
435		m = store[i];
436		(void)mbufq_enqueue(q, m);
437	}
438}
439
440static int
441nd_pack_import(void *arg __unused, void **store, int count, int domain __unused,
442    int flags __unused)
443{
444	struct mbuf *m;
445	void *clust;
446	int i;
447
448	for (i = 0; i < count; i++) {
449		m = m_get(MT_DATA, M_NOWAIT);
450		if (m == NULL)
451			break;
452		clust = uma_zalloc(nd_zone_clust, M_NOWAIT);
453		if (clust == NULL) {
454			m_free(m);
455			break;
456		}
457		mb_ctor_clust(clust, nd_clsize, m, 0);
458		store[i] = m;
459	}
460	return (i);
461}
462
463static void
464nd_pack_release(void *arg __unused, void **store, int count)
465{
466	struct mbuf *m;
467	void *clust;
468	int i;
469
470	for (i = 0; i < count; i++) {
471		m = store[i];
472		clust = m->m_ext.ext_buf;
473		uma_zfree(nd_zone_clust, clust);
474		uma_zfree(nd_zone_mbuf, m);
475	}
476}
477
478/*
479 * Free the pre-allocated mbufs and clusters reserved for netdump, and destroy
480 * the corresponding UMA cache zones.
481 */
482void
483netdump_mbuf_drain(void)
484{
485	struct mbuf *m;
486	void *item;
487
488	if (nd_zone_mbuf != NULL) {
489		uma_zdestroy(nd_zone_mbuf);
490		nd_zone_mbuf = NULL;
491	}
492	if (nd_zone_clust != NULL) {
493		uma_zdestroy(nd_zone_clust);
494		nd_zone_clust = NULL;
495	}
496	if (nd_zone_pack != NULL) {
497		uma_zdestroy(nd_zone_pack);
498		nd_zone_pack = NULL;
499	}
500
501	while ((m = mbufq_dequeue(&nd_mbufq)) != NULL)
502		m_free(m);
503	while ((item = mbufq_dequeue(&nd_clustq)) != NULL)
504		uma_zfree(m_getzone(nd_clsize), item);
505}
506
507/*
508 * Callback invoked immediately prior to starting a netdump.
509 */
510void
511netdump_mbuf_dump(void)
512{
513
514	/*
515	 * All cluster zones return buffers of the size requested by the
516	 * drivers.  It's up to the driver to reinitialize the zones if the
517	 * MTU of a netdump-enabled interface changes.
518	 */
519	printf("netdump: overwriting mbuf zone pointers\n");
520	zone_mbuf = nd_zone_mbuf;
521	zone_clust = nd_zone_clust;
522	zone_pack = nd_zone_pack;
523	zone_jumbop = nd_zone_clust;
524	zone_jumbo9 = nd_zone_clust;
525	zone_jumbo16 = nd_zone_clust;
526}
527
528/*
529 * Reinitialize the netdump mbuf+cluster pool and cache zones.
530 */
531void
532netdump_mbuf_reinit(int nmbuf, int nclust, int clsize)
533{
534	struct mbuf *m;
535	void *item;
536
537	netdump_mbuf_drain();
538
539	nd_clsize = clsize;
540
541	nd_zone_mbuf = uma_zcache_create("netdump_" MBUF_MEM_NAME,
542	    MSIZE, mb_ctor_mbuf, mb_dtor_mbuf,
543#ifdef INVARIANTS
544	    trash_init, trash_fini,
545#else
546	    NULL, NULL,
547#endif
548	    nd_buf_import, nd_buf_release,
549	    &nd_mbufq, UMA_ZONE_NOBUCKET);
550
551	nd_zone_clust = uma_zcache_create("netdump_" MBUF_CLUSTER_MEM_NAME,
552	    clsize, mb_ctor_clust,
553#ifdef INVARIANTS
554	    trash_dtor, trash_init, trash_fini,
555#else
556	    NULL, NULL, NULL,
557#endif
558	    nd_buf_import, nd_buf_release,
559	    &nd_clustq, UMA_ZONE_NOBUCKET);
560
561	nd_zone_pack = uma_zcache_create("netdump_" MBUF_PACKET_MEM_NAME,
562	    MCLBYTES, mb_ctor_pack, mb_dtor_pack, NULL, NULL,
563	    nd_pack_import, nd_pack_release,
564	    NULL, UMA_ZONE_NOBUCKET);
565
566	while (nmbuf-- > 0) {
567		m = m_get(MT_DATA, M_WAITOK);
568		uma_zfree(nd_zone_mbuf, m);
569	}
570	while (nclust-- > 0) {
571		item = uma_zalloc(m_getzone(nd_clsize), M_WAITOK);
572		uma_zfree(nd_zone_clust, item);
573	}
574}
575#endif /* NETDUMP */
576
577/*
578 * UMA backend page allocator for the jumbo frame zones.
579 *
580 * Allocates kernel virtual memory that is backed by contiguous physical
581 * pages.
582 */
583static void *
584mbuf_jumbo_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
585    int wait)
586{
587
588	/* Inform UMA that this allocator uses kernel_map/object. */
589	*flags = UMA_SLAB_KERNEL;
590	return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
591	    bytes, wait, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
592	    VM_MEMATTR_DEFAULT));
593}
594
595/*
596 * Constructor for Mbuf master zone.
597 *
598 * The 'arg' pointer points to a mb_args structure which
599 * contains call-specific information required to support the
600 * mbuf allocation API.  See mbuf.h.
601 */
602static int
603mb_ctor_mbuf(void *mem, int size, void *arg, int how)
604{
605	struct mbuf *m;
606	struct mb_args *args;
607	int error;
608	int flags;
609	short type;
610
611#ifdef INVARIANTS
612	trash_ctor(mem, size, arg, how);
613#endif
614	args = (struct mb_args *)arg;
615	type = args->type;
616
617	/*
618	 * The mbuf is initialized later.  The caller has the
619	 * responsibility to set up any MAC labels too.
620	 */
621	if (type == MT_NOINIT)
622		return (0);
623
624	m = (struct mbuf *)mem;
625	flags = args->flags;
626	MPASS((flags & M_NOFREE) == 0);
627
628	error = m_init(m, how, type, flags);
629
630	return (error);
631}
632
633/*
634 * The Mbuf master zone destructor.
635 */
636static void
637mb_dtor_mbuf(void *mem, int size, void *arg)
638{
639	struct mbuf *m;
640	unsigned long flags;
641
642	m = (struct mbuf *)mem;
643	flags = (unsigned long)arg;
644
645	KASSERT((m->m_flags & M_NOFREE) == 0, ("%s: M_NOFREE set", __func__));
646	if (!(flags & MB_DTOR_SKIP) && (m->m_flags & M_PKTHDR) && !SLIST_EMPTY(&m->m_pkthdr.tags))
647		m_tag_delete_chain(m, NULL);
648#ifdef INVARIANTS
649	trash_dtor(mem, size, arg);
650#endif
651}
652
653/*
654 * The Mbuf Packet zone destructor.
655 */
656static void
657mb_dtor_pack(void *mem, int size, void *arg)
658{
659	struct mbuf *m;
660
661	m = (struct mbuf *)mem;
662	if ((m->m_flags & M_PKTHDR) != 0)
663		m_tag_delete_chain(m, NULL);
664
665	/* Make sure we've got a clean cluster back. */
666	KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__));
667	KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__));
668	KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__));
669	KASSERT(m->m_ext.ext_arg1 == NULL, ("%s: ext_arg1 != NULL", __func__));
670	KASSERT(m->m_ext.ext_arg2 == NULL, ("%s: ext_arg2 != NULL", __func__));
671	KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__));
672	KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__));
673#ifdef INVARIANTS
674	trash_dtor(m->m_ext.ext_buf, MCLBYTES, arg);
675#endif
676	/*
677	 * If there are processes blocked on zone_clust, waiting for pages
678	 * to be freed up, * cause them to be woken up by draining the
679	 * packet zone.  We are exposed to a race here * (in the check for
680	 * the UMA_ZFLAG_FULL) where we might miss the flag set, but that
681	 * is deliberate. We don't want to acquire the zone lock for every
682	 * mbuf free.
683	 */
684	if (uma_zone_exhausted_nolock(zone_clust))
685		zone_drain(zone_pack);
686}
687
688/*
689 * The Cluster and Jumbo[PAGESIZE|9|16] zone constructor.
690 *
691 * Here the 'arg' pointer points to the Mbuf which we
692 * are configuring cluster storage for.  If 'arg' is
693 * empty we allocate just the cluster without setting
694 * the mbuf to it.  See mbuf.h.
695 */
696static int
697mb_ctor_clust(void *mem, int size, void *arg, int how)
698{
699	struct mbuf *m;
700
701#ifdef INVARIANTS
702	trash_ctor(mem, size, arg, how);
703#endif
704	m = (struct mbuf *)arg;
705	if (m != NULL) {
706		m->m_ext.ext_buf = (char *)mem;
707		m->m_data = m->m_ext.ext_buf;
708		m->m_flags |= M_EXT;
709		m->m_ext.ext_free = NULL;
710		m->m_ext.ext_arg1 = NULL;
711		m->m_ext.ext_arg2 = NULL;
712		m->m_ext.ext_size = size;
713		m->m_ext.ext_type = m_gettype(size);
714		m->m_ext.ext_flags = EXT_FLAG_EMBREF;
715		m->m_ext.ext_count = 1;
716	}
717
718	return (0);
719}
720
721/*
722 * The Packet secondary zone's init routine, executed on the
723 * object's transition from mbuf keg slab to zone cache.
724 */
725static int
726mb_zinit_pack(void *mem, int size, int how)
727{
728	struct mbuf *m;
729
730	m = (struct mbuf *)mem;		/* m is virgin. */
731	if (uma_zalloc_arg(zone_clust, m, how) == NULL ||
732	    m->m_ext.ext_buf == NULL)
733		return (ENOMEM);
734	m->m_ext.ext_type = EXT_PACKET;	/* Override. */
735#ifdef INVARIANTS
736	trash_init(m->m_ext.ext_buf, MCLBYTES, how);
737#endif
738	return (0);
739}
740
741/*
742 * The Packet secondary zone's fini routine, executed on the
743 * object's transition from zone cache to keg slab.
744 */
745static void
746mb_zfini_pack(void *mem, int size)
747{
748	struct mbuf *m;
749
750	m = (struct mbuf *)mem;
751#ifdef INVARIANTS
752	trash_fini(m->m_ext.ext_buf, MCLBYTES);
753#endif
754	uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL);
755#ifdef INVARIANTS
756	trash_dtor(mem, size, NULL);
757#endif
758}
759
760/*
761 * The "packet" keg constructor.
762 */
763static int
764mb_ctor_pack(void *mem, int size, void *arg, int how)
765{
766	struct mbuf *m;
767	struct mb_args *args;
768	int error, flags;
769	short type;
770
771	m = (struct mbuf *)mem;
772	args = (struct mb_args *)arg;
773	flags = args->flags;
774	type = args->type;
775	MPASS((flags & M_NOFREE) == 0);
776
777#ifdef INVARIANTS
778	trash_ctor(m->m_ext.ext_buf, MCLBYTES, arg, how);
779#endif
780
781	error = m_init(m, how, type, flags);
782
783	/* m_ext is already initialized. */
784	m->m_data = m->m_ext.ext_buf;
785 	m->m_flags = (flags | M_EXT);
786
787	return (error);
788}
789
790/*
791 * This is the protocol drain routine.  Called by UMA whenever any of the
792 * mbuf zones is closed to its limit.
793 *
794 * No locks should be held when this is called.  The drain routines have to
795 * presently acquire some locks which raises the possibility of lock order
796 * reversal.
797 */
798static void
799mb_reclaim(uma_zone_t zone __unused, int pending __unused)
800{
801	struct domain *dp;
802	struct protosw *pr;
803
804	WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK | WARN_PANIC, NULL, __func__);
805
806	for (dp = domains; dp != NULL; dp = dp->dom_next)
807		for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++)
808			if (pr->pr_drain != NULL)
809				(*pr->pr_drain)();
810}
811
812/*
813 * Clean up after mbufs with M_EXT storage attached to them if the
814 * reference count hits 1.
815 */
816void
817mb_free_ext(struct mbuf *m)
818{
819	volatile u_int *refcnt;
820	struct mbuf *mref;
821	int freembuf;
822
823	KASSERT(m->m_flags & M_EXT, ("%s: M_EXT not set on %p", __func__, m));
824
825	/* See if this is the mbuf that holds the embedded refcount. */
826	if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
827		refcnt = &m->m_ext.ext_count;
828		mref = m;
829	} else {
830		KASSERT(m->m_ext.ext_cnt != NULL,
831		    ("%s: no refcounting pointer on %p", __func__, m));
832		refcnt = m->m_ext.ext_cnt;
833		mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
834	}
835
836	/*
837	 * Check if the header is embedded in the cluster.  It is
838	 * important that we can't touch any of the mbuf fields
839	 * after we have freed the external storage, since mbuf
840	 * could have been embedded in it.  For now, the mbufs
841	 * embedded into the cluster are always of type EXT_EXTREF,
842	 * and for this type we won't free the mref.
843	 */
844	if (m->m_flags & M_NOFREE) {
845		freembuf = 0;
846		KASSERT(m->m_ext.ext_type == EXT_EXTREF,
847		    ("%s: no-free mbuf %p has wrong type", __func__, m));
848	} else
849		freembuf = 1;
850
851	/* Free attached storage if this mbuf is the only reference to it. */
852	if (*refcnt == 1 || atomic_fetchadd_int(refcnt, -1) == 1) {
853		switch (m->m_ext.ext_type) {
854		case EXT_PACKET:
855			/* The packet zone is special. */
856			if (*refcnt == 0)
857				*refcnt = 1;
858			uma_zfree(zone_pack, mref);
859			break;
860		case EXT_CLUSTER:
861			uma_zfree(zone_clust, m->m_ext.ext_buf);
862			uma_zfree(zone_mbuf, mref);
863			break;
864		case EXT_JUMBOP:
865			uma_zfree(zone_jumbop, m->m_ext.ext_buf);
866			uma_zfree(zone_mbuf, mref);
867			break;
868		case EXT_JUMBO9:
869			uma_zfree(zone_jumbo9, m->m_ext.ext_buf);
870			uma_zfree(zone_mbuf, mref);
871			break;
872		case EXT_JUMBO16:
873			uma_zfree(zone_jumbo16, m->m_ext.ext_buf);
874			uma_zfree(zone_mbuf, mref);
875			break;
876		case EXT_SFBUF:
877		case EXT_NET_DRV:
878		case EXT_MOD_TYPE:
879		case EXT_DISPOSABLE:
880			KASSERT(mref->m_ext.ext_free != NULL,
881			    ("%s: ext_free not set", __func__));
882			mref->m_ext.ext_free(mref);
883			uma_zfree(zone_mbuf, mref);
884			break;
885		case EXT_EXTREF:
886			KASSERT(m->m_ext.ext_free != NULL,
887			    ("%s: ext_free not set", __func__));
888			m->m_ext.ext_free(m);
889			break;
890		default:
891			KASSERT(m->m_ext.ext_type == 0,
892			    ("%s: unknown ext_type", __func__));
893		}
894	}
895
896	if (freembuf && m != mref)
897		uma_zfree(zone_mbuf, m);
898}
899
900/*
901 * Official mbuf(9) allocation KPI for stack and drivers:
902 *
903 * m_get()	- a single mbuf without any attachments, sys/mbuf.h.
904 * m_gethdr()	- a single mbuf initialized as M_PKTHDR, sys/mbuf.h.
905 * m_getcl()	- an mbuf + 2k cluster, sys/mbuf.h.
906 * m_clget()	- attach cluster to already allocated mbuf.
907 * m_cljget()	- attach jumbo cluster to already allocated mbuf.
908 * m_get2()	- allocate minimum mbuf that would fit size argument.
909 * m_getm2()	- allocate a chain of mbufs/clusters.
910 * m_extadd()	- attach external cluster to mbuf.
911 *
912 * m_free()	- free single mbuf with its tags and ext, sys/mbuf.h.
913 * m_freem()	- free chain of mbufs.
914 */
915
916int
917m_clget(struct mbuf *m, int how)
918{
919
920	KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
921	    __func__, m));
922	m->m_ext.ext_buf = (char *)NULL;
923	uma_zalloc_arg(zone_clust, m, how);
924	/*
925	 * On a cluster allocation failure, drain the packet zone and retry,
926	 * we might be able to loosen a few clusters up on the drain.
927	 */
928	if ((how & M_NOWAIT) && (m->m_ext.ext_buf == NULL)) {
929		zone_drain(zone_pack);
930		uma_zalloc_arg(zone_clust, m, how);
931	}
932	MBUF_PROBE2(m__clget, m, how);
933	return (m->m_flags & M_EXT);
934}
935
936/*
937 * m_cljget() is different from m_clget() as it can allocate clusters without
938 * attaching them to an mbuf.  In that case the return value is the pointer
939 * to the cluster of the requested size.  If an mbuf was specified, it gets
940 * the cluster attached to it and the return value can be safely ignored.
941 * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
942 */
943void *
944m_cljget(struct mbuf *m, int how, int size)
945{
946	uma_zone_t zone;
947	void *retval;
948
949	if (m != NULL) {
950		KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
951		    __func__, m));
952		m->m_ext.ext_buf = NULL;
953	}
954
955	zone = m_getzone(size);
956	retval = uma_zalloc_arg(zone, m, how);
957
958	MBUF_PROBE4(m__cljget, m, how, size, retval);
959
960	return (retval);
961}
962
963/*
964 * m_get2() allocates minimum mbuf that would fit "size" argument.
965 */
966struct mbuf *
967m_get2(int size, int how, short type, int flags)
968{
969	struct mb_args args;
970	struct mbuf *m, *n;
971
972	args.flags = flags;
973	args.type = type;
974
975	if (size <= MHLEN || (size <= MLEN && (flags & M_PKTHDR) == 0))
976		return (uma_zalloc_arg(zone_mbuf, &args, how));
977	if (size <= MCLBYTES)
978		return (uma_zalloc_arg(zone_pack, &args, how));
979
980	if (size > MJUMPAGESIZE)
981		return (NULL);
982
983	m = uma_zalloc_arg(zone_mbuf, &args, how);
984	if (m == NULL)
985		return (NULL);
986
987	n = uma_zalloc_arg(zone_jumbop, m, how);
988	if (n == NULL) {
989		uma_zfree(zone_mbuf, m);
990		return (NULL);
991	}
992
993	return (m);
994}
995
996/*
997 * m_getjcl() returns an mbuf with a cluster of the specified size attached.
998 * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
999 */
1000struct mbuf *
1001m_getjcl(int how, short type, int flags, int size)
1002{
1003	struct mb_args args;
1004	struct mbuf *m, *n;
1005	uma_zone_t zone;
1006
1007	if (size == MCLBYTES)
1008		return m_getcl(how, type, flags);
1009
1010	args.flags = flags;
1011	args.type = type;
1012
1013	m = uma_zalloc_arg(zone_mbuf, &args, how);
1014	if (m == NULL)
1015		return (NULL);
1016
1017	zone = m_getzone(size);
1018	n = uma_zalloc_arg(zone, m, how);
1019	if (n == NULL) {
1020		uma_zfree(zone_mbuf, m);
1021		return (NULL);
1022	}
1023	MBUF_PROBE5(m__getjcl, how, type, flags, size, m);
1024	return (m);
1025}
1026
1027/*
1028 * Allocate a given length worth of mbufs and/or clusters (whatever fits
1029 * best) and return a pointer to the top of the allocated chain.  If an
1030 * existing mbuf chain is provided, then we will append the new chain
1031 * to the existing one but still return the top of the newly allocated
1032 * chain.
1033 */
1034struct mbuf *
1035m_getm2(struct mbuf *m, int len, int how, short type, int flags)
1036{
1037	struct mbuf *mb, *nm = NULL, *mtail = NULL;
1038
1039	KASSERT(len >= 0, ("%s: len is < 0", __func__));
1040
1041	/* Validate flags. */
1042	flags &= (M_PKTHDR | M_EOR);
1043
1044	/* Packet header mbuf must be first in chain. */
1045	if ((flags & M_PKTHDR) && m != NULL)
1046		flags &= ~M_PKTHDR;
1047
1048	/* Loop and append maximum sized mbufs to the chain tail. */
1049	while (len > 0) {
1050		if (len > MCLBYTES)
1051			mb = m_getjcl(how, type, (flags & M_PKTHDR),
1052			    MJUMPAGESIZE);
1053		else if (len >= MINCLSIZE)
1054			mb = m_getcl(how, type, (flags & M_PKTHDR));
1055		else if (flags & M_PKTHDR)
1056			mb = m_gethdr(how, type);
1057		else
1058			mb = m_get(how, type);
1059
1060		/* Fail the whole operation if one mbuf can't be allocated. */
1061		if (mb == NULL) {
1062			if (nm != NULL)
1063				m_freem(nm);
1064			return (NULL);
1065		}
1066
1067		/* Book keeping. */
1068		len -= M_SIZE(mb);
1069		if (mtail != NULL)
1070			mtail->m_next = mb;
1071		else
1072			nm = mb;
1073		mtail = mb;
1074		flags &= ~M_PKTHDR;	/* Only valid on the first mbuf. */
1075	}
1076	if (flags & M_EOR)
1077		mtail->m_flags |= M_EOR;  /* Only valid on the last mbuf. */
1078
1079	/* If mbuf was supplied, append new chain to the end of it. */
1080	if (m != NULL) {
1081		for (mtail = m; mtail->m_next != NULL; mtail = mtail->m_next)
1082			;
1083		mtail->m_next = nm;
1084		mtail->m_flags &= ~M_EOR;
1085	} else
1086		m = nm;
1087
1088	return (m);
1089}
1090
1091/*-
1092 * Configure a provided mbuf to refer to the provided external storage
1093 * buffer and setup a reference count for said buffer.
1094 *
1095 * Arguments:
1096 *    mb     The existing mbuf to which to attach the provided buffer.
1097 *    buf    The address of the provided external storage buffer.
1098 *    size   The size of the provided buffer.
1099 *    freef  A pointer to a routine that is responsible for freeing the
1100 *           provided external storage buffer.
1101 *    args   A pointer to an argument structure (of any type) to be passed
1102 *           to the provided freef routine (may be NULL).
1103 *    flags  Any other flags to be passed to the provided mbuf.
1104 *    type   The type that the external storage buffer should be
1105 *           labeled with.
1106 *
1107 * Returns:
1108 *    Nothing.
1109 */
1110void
1111m_extadd(struct mbuf *mb, char *buf, u_int size, m_ext_free_t freef,
1112    void *arg1, void *arg2, int flags, int type)
1113{
1114
1115	KASSERT(type != EXT_CLUSTER, ("%s: EXT_CLUSTER not allowed", __func__));
1116
1117	mb->m_flags |= (M_EXT | flags);
1118	mb->m_ext.ext_buf = buf;
1119	mb->m_data = mb->m_ext.ext_buf;
1120	mb->m_ext.ext_size = size;
1121	mb->m_ext.ext_free = freef;
1122	mb->m_ext.ext_arg1 = arg1;
1123	mb->m_ext.ext_arg2 = arg2;
1124	mb->m_ext.ext_type = type;
1125
1126	if (type != EXT_EXTREF) {
1127		mb->m_ext.ext_count = 1;
1128		mb->m_ext.ext_flags = EXT_FLAG_EMBREF;
1129	} else
1130		mb->m_ext.ext_flags = 0;
1131}
1132
1133/*
1134 * Free an entire chain of mbufs and associated external buffers, if
1135 * applicable.
1136 */
1137void
1138m_freem(struct mbuf *mb)
1139{
1140
1141	MBUF_PROBE1(m__freem, mb);
1142	while (mb != NULL)
1143		mb = m_free(mb);
1144}
1145