2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2004, 2005,
5 * Bosko Milekic <bmilekic@FreeBSD.org>. All rights reserved.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice unmodified, this list of conditions and the following
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.
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
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
33 #include "opt_param.h"
35 #include <sys/param.h>
37 #include <sys/domainset.h>
38 #include <sys/malloc.h>
39 #include <sys/systm.h>
41 #include <sys/domain.h>
42 #include <sys/eventhandler.h>
43 #include <sys/kernel.h>
44 #include <sys/limits.h>
46 #include <sys/mutex.h>
47 #include <sys/protosw.h>
49 #include <sys/socket.h>
50 #include <sys/sysctl.h>
53 #include <net/if_var.h>
56 #include <vm/vm_extern.h>
57 #include <vm/vm_kern.h>
58 #include <vm/vm_page.h>
59 #include <vm/vm_map.h>
61 #include <vm/uma_dbg.h>
64 * In FreeBSD, Mbufs and Mbuf Clusters are allocated from UMA
67 * Mbuf Clusters (2K, contiguous) are allocated from the Cluster
68 * Zone. The Zone can be capped at kern.ipc.nmbclusters, if the
69 * administrator so desires.
71 * Mbufs are allocated from a UMA Master Zone called the Mbuf
74 * Additionally, FreeBSD provides a Packet Zone, which it
75 * configures as a Secondary Zone to the Mbuf Master Zone,
76 * thus sharing backend Slab kegs with the Mbuf Master Zone.
78 * Thus common-case allocations and locking are simplified:
82 * | .------------>[(Packet Cache)] m_get(), m_gethdr()
84 * [(Cluster Cache)] [ Secondary ] [ (Mbuf Cache) ]
85 * [ Cluster Zone ] [ Zone ] [ Mbuf Master Zone ]
91 * \____________(VM)_________________/
94 * Whenever an object is allocated with uma_zalloc() out of
95 * one of the Zones its _ctor_ function is executed. The same
96 * for any deallocation through uma_zfree() the _dtor_ function
99 * Caches are per-CPU and are filled from the Master Zone.
101 * Whenever an object is allocated from the underlying global
102 * memory pool it gets pre-initialized with the _zinit_ functions.
103 * When the Keg's are overfull objects get decommissioned with
104 * _zfini_ functions and free'd back to the global memory pool.
108 int nmbufs; /* limits number of mbufs */
109 int nmbclusters; /* limits number of mbuf clusters */
110 int nmbjumbop; /* limits number of page size jumbo clusters */
111 int nmbjumbo9; /* limits number of 9k jumbo clusters */
112 int nmbjumbo16; /* limits number of 16k jumbo clusters */
114 static quad_t maxmbufmem; /* overall real memory limit for all mbufs */
116 SYSCTL_QUAD(_kern_ipc, OID_AUTO, maxmbufmem, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &maxmbufmem, 0,
117 "Maximum real memory allocatable to various mbuf types");
119 static counter_u64_t snd_tag_count;
120 SYSCTL_COUNTER_U64(_kern_ipc, OID_AUTO, num_snd_tags, CTLFLAG_RW,
121 &snd_tag_count, "# of active mbuf send tags");
124 * tunable_mbinit() has to be run before any mbuf allocations are done.
127 tunable_mbinit(void *dummy)
132 * The default limit for all mbuf related memory is 1/2 of all
133 * available kernel memory (physical or kmem).
134 * At most it can be 3/4 of available kernel memory.
136 realmem = qmin((quad_t)physmem * PAGE_SIZE, vm_kmem_size);
137 maxmbufmem = realmem / 2;
138 TUNABLE_QUAD_FETCH("kern.ipc.maxmbufmem", &maxmbufmem);
139 if (maxmbufmem > realmem / 4 * 3)
140 maxmbufmem = realmem / 4 * 3;
142 TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
143 if (nmbclusters == 0)
144 nmbclusters = maxmbufmem / MCLBYTES / 4;
146 TUNABLE_INT_FETCH("kern.ipc.nmbjumbop", &nmbjumbop);
148 nmbjumbop = maxmbufmem / MJUMPAGESIZE / 4;
150 TUNABLE_INT_FETCH("kern.ipc.nmbjumbo9", &nmbjumbo9);
152 nmbjumbo9 = maxmbufmem / MJUM9BYTES / 6;
154 TUNABLE_INT_FETCH("kern.ipc.nmbjumbo16", &nmbjumbo16);
156 nmbjumbo16 = maxmbufmem / MJUM16BYTES / 6;
159 * We need at least as many mbufs as we have clusters of
160 * the various types added together.
162 TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
163 if (nmbufs < nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16)
164 nmbufs = lmax(maxmbufmem / MSIZE / 5,
165 nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16);
167 SYSINIT(tunable_mbinit, SI_SUB_KMEM, SI_ORDER_MIDDLE, tunable_mbinit, NULL);
170 sysctl_nmbclusters(SYSCTL_HANDLER_ARGS)
172 int error, newnmbclusters;
174 newnmbclusters = nmbclusters;
175 error = sysctl_handle_int(oidp, &newnmbclusters, 0, req);
176 if (error == 0 && req->newptr && newnmbclusters != nmbclusters) {
177 if (newnmbclusters > nmbclusters &&
178 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
179 nmbclusters = newnmbclusters;
180 nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
181 EVENTHANDLER_INVOKE(nmbclusters_change);
187 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbclusters, CTLTYPE_INT|CTLFLAG_RW,
188 &nmbclusters, 0, sysctl_nmbclusters, "IU",
189 "Maximum number of mbuf clusters allowed");
192 sysctl_nmbjumbop(SYSCTL_HANDLER_ARGS)
194 int error, newnmbjumbop;
196 newnmbjumbop = nmbjumbop;
197 error = sysctl_handle_int(oidp, &newnmbjumbop, 0, req);
198 if (error == 0 && req->newptr && newnmbjumbop != nmbjumbop) {
199 if (newnmbjumbop > nmbjumbop &&
200 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
201 nmbjumbop = newnmbjumbop;
202 nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
208 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbop, CTLTYPE_INT|CTLFLAG_RW,
209 &nmbjumbop, 0, sysctl_nmbjumbop, "IU",
210 "Maximum number of mbuf page size jumbo clusters allowed");
213 sysctl_nmbjumbo9(SYSCTL_HANDLER_ARGS)
215 int error, newnmbjumbo9;
217 newnmbjumbo9 = nmbjumbo9;
218 error = sysctl_handle_int(oidp, &newnmbjumbo9, 0, req);
219 if (error == 0 && req->newptr && newnmbjumbo9 != nmbjumbo9) {
220 if (newnmbjumbo9 > nmbjumbo9 &&
221 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
222 nmbjumbo9 = newnmbjumbo9;
223 nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
229 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo9, CTLTYPE_INT|CTLFLAG_RW,
230 &nmbjumbo9, 0, sysctl_nmbjumbo9, "IU",
231 "Maximum number of mbuf 9k jumbo clusters allowed");
234 sysctl_nmbjumbo16(SYSCTL_HANDLER_ARGS)
236 int error, newnmbjumbo16;
238 newnmbjumbo16 = nmbjumbo16;
239 error = sysctl_handle_int(oidp, &newnmbjumbo16, 0, req);
240 if (error == 0 && req->newptr && newnmbjumbo16 != nmbjumbo16) {
241 if (newnmbjumbo16 > nmbjumbo16 &&
242 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
243 nmbjumbo16 = newnmbjumbo16;
244 nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
250 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo16, CTLTYPE_INT|CTLFLAG_RW,
251 &nmbjumbo16, 0, sysctl_nmbjumbo16, "IU",
252 "Maximum number of mbuf 16k jumbo clusters allowed");
255 sysctl_nmbufs(SYSCTL_HANDLER_ARGS)
257 int error, newnmbufs;
260 error = sysctl_handle_int(oidp, &newnmbufs, 0, req);
261 if (error == 0 && req->newptr && newnmbufs != nmbufs) {
262 if (newnmbufs > nmbufs) {
264 nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
265 EVENTHANDLER_INVOKE(nmbufs_change);
271 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbufs, CTLTYPE_INT|CTLFLAG_RW,
272 &nmbufs, 0, sysctl_nmbufs, "IU",
273 "Maximum number of mbufs allowed");
276 * Zones from which we allocate.
278 uma_zone_t zone_mbuf;
279 uma_zone_t zone_clust;
280 uma_zone_t zone_pack;
281 uma_zone_t zone_jumbop;
282 uma_zone_t zone_jumbo9;
283 uma_zone_t zone_jumbo16;
288 static int mb_ctor_mbuf(void *, int, void *, int);
289 static int mb_ctor_clust(void *, int, void *, int);
290 static int mb_ctor_pack(void *, int, void *, int);
291 static void mb_dtor_mbuf(void *, int, void *);
292 static void mb_dtor_pack(void *, int, void *);
293 static int mb_zinit_pack(void *, int, int);
294 static void mb_zfini_pack(void *, int);
295 static void mb_reclaim(uma_zone_t, int);
296 static void *mbuf_jumbo_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
298 /* Ensure that MSIZE is a power of 2. */
299 CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE);
302 * Initialize FreeBSD Network buffer allocation.
305 mbuf_init(void *dummy)
309 * Configure UMA zones for Mbufs, Clusters, and Packets.
311 zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE,
312 mb_ctor_mbuf, mb_dtor_mbuf,
314 trash_init, trash_fini,
318 MSIZE - 1, UMA_ZONE_MAXBUCKET);
320 nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
321 uma_zone_set_warning(zone_mbuf, "kern.ipc.nmbufs limit reached");
322 uma_zone_set_maxaction(zone_mbuf, mb_reclaim);
324 zone_clust = uma_zcreate(MBUF_CLUSTER_MEM_NAME, MCLBYTES,
327 trash_dtor, trash_init, trash_fini,
333 nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
334 uma_zone_set_warning(zone_clust, "kern.ipc.nmbclusters limit reached");
335 uma_zone_set_maxaction(zone_clust, mb_reclaim);
337 zone_pack = uma_zsecond_create(MBUF_PACKET_MEM_NAME, mb_ctor_pack,
338 mb_dtor_pack, mb_zinit_pack, mb_zfini_pack, zone_mbuf);
340 /* Make jumbo frame zone too. Page size, 9k and 16k. */
341 zone_jumbop = uma_zcreate(MBUF_JUMBOP_MEM_NAME, MJUMPAGESIZE,
344 trash_dtor, trash_init, trash_fini,
350 nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
351 uma_zone_set_warning(zone_jumbop, "kern.ipc.nmbjumbop limit reached");
352 uma_zone_set_maxaction(zone_jumbop, mb_reclaim);
354 zone_jumbo9 = uma_zcreate(MBUF_JUMBO9_MEM_NAME, MJUM9BYTES,
357 trash_dtor, trash_init, trash_fini,
362 uma_zone_set_allocf(zone_jumbo9, mbuf_jumbo_alloc);
364 nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
365 uma_zone_set_warning(zone_jumbo9, "kern.ipc.nmbjumbo9 limit reached");
366 uma_zone_set_maxaction(zone_jumbo9, mb_reclaim);
368 zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES,
371 trash_dtor, trash_init, trash_fini,
376 uma_zone_set_allocf(zone_jumbo16, mbuf_jumbo_alloc);
378 nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
379 uma_zone_set_warning(zone_jumbo16, "kern.ipc.nmbjumbo16 limit reached");
380 uma_zone_set_maxaction(zone_jumbo16, mb_reclaim);
383 * Hook event handler for low-memory situation, used to
384 * drain protocols and push data back to the caches (UMA
385 * later pushes it back to VM).
387 EVENTHANDLER_REGISTER(vm_lowmem, mb_reclaim, NULL,
388 EVENTHANDLER_PRI_FIRST);
390 snd_tag_count = counter_u64_alloc(M_WAITOK);
392 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL);
396 * netdump makes use of a pre-allocated pool of mbufs and clusters. When
397 * netdump is configured, we initialize a set of UMA cache zones which return
398 * items from this pool. At panic-time, the regular UMA zone pointers are
399 * overwritten with those of the cache zones so that drivers may allocate and
400 * free mbufs and clusters without attempting to allocate physical memory.
402 * We keep mbufs and clusters in a pair of mbuf queues. In particular, for
403 * the purpose of caching clusters, we treat them as mbufs.
405 static struct mbufq nd_mbufq =
406 { STAILQ_HEAD_INITIALIZER(nd_mbufq.mq_head), 0, INT_MAX };
407 static struct mbufq nd_clustq =
408 { STAILQ_HEAD_INITIALIZER(nd_clustq.mq_head), 0, INT_MAX };
410 static int nd_clsize;
411 static uma_zone_t nd_zone_mbuf;
412 static uma_zone_t nd_zone_clust;
413 static uma_zone_t nd_zone_pack;
416 nd_buf_import(void *arg, void **store, int count, int domain __unused,
425 for (i = 0; i < count; i++) {
426 m = mbufq_dequeue(q);
429 trash_init(m, q == &nd_mbufq ? MSIZE : nd_clsize, flags);
432 KASSERT((flags & M_WAITOK) == 0 || i == count,
433 ("%s: ran out of pre-allocated mbufs", __func__));
438 nd_buf_release(void *arg, void **store, int count)
446 for (i = 0; i < count; i++) {
448 (void)mbufq_enqueue(q, m);
453 nd_pack_import(void *arg __unused, void **store, int count, int domain __unused,
460 for (i = 0; i < count; i++) {
461 m = m_get(MT_DATA, M_NOWAIT);
464 clust = uma_zalloc(nd_zone_clust, M_NOWAIT);
469 mb_ctor_clust(clust, nd_clsize, m, 0);
472 KASSERT((flags & M_WAITOK) == 0 || i == count,
473 ("%s: ran out of pre-allocated mbufs", __func__));
478 nd_pack_release(void *arg __unused, void **store, int count)
484 for (i = 0; i < count; i++) {
486 clust = m->m_ext.ext_buf;
487 uma_zfree(nd_zone_clust, clust);
488 uma_zfree(nd_zone_mbuf, m);
493 * Free the pre-allocated mbufs and clusters reserved for netdump, and destroy
494 * the corresponding UMA cache zones.
497 netdump_mbuf_drain(void)
502 if (nd_zone_mbuf != NULL) {
503 uma_zdestroy(nd_zone_mbuf);
506 if (nd_zone_clust != NULL) {
507 uma_zdestroy(nd_zone_clust);
508 nd_zone_clust = NULL;
510 if (nd_zone_pack != NULL) {
511 uma_zdestroy(nd_zone_pack);
515 while ((m = mbufq_dequeue(&nd_mbufq)) != NULL)
517 while ((item = mbufq_dequeue(&nd_clustq)) != NULL)
518 uma_zfree(m_getzone(nd_clsize), item);
522 * Callback invoked immediately prior to starting a netdump.
525 netdump_mbuf_dump(void)
529 * All cluster zones return buffers of the size requested by the
530 * drivers. It's up to the driver to reinitialize the zones if the
531 * MTU of a netdump-enabled interface changes.
533 printf("netdump: overwriting mbuf zone pointers\n");
534 zone_mbuf = nd_zone_mbuf;
535 zone_clust = nd_zone_clust;
536 zone_pack = nd_zone_pack;
537 zone_jumbop = nd_zone_clust;
538 zone_jumbo9 = nd_zone_clust;
539 zone_jumbo16 = nd_zone_clust;
543 * Reinitialize the netdump mbuf+cluster pool and cache zones.
546 netdump_mbuf_reinit(int nmbuf, int nclust, int clsize)
551 netdump_mbuf_drain();
555 nd_zone_mbuf = uma_zcache_create("netdump_" MBUF_MEM_NAME,
556 MSIZE, mb_ctor_mbuf, mb_dtor_mbuf,
558 trash_init, trash_fini,
562 nd_buf_import, nd_buf_release,
563 &nd_mbufq, UMA_ZONE_NOBUCKET);
565 nd_zone_clust = uma_zcache_create("netdump_" MBUF_CLUSTER_MEM_NAME,
566 clsize, mb_ctor_clust,
568 trash_dtor, trash_init, trash_fini,
572 nd_buf_import, nd_buf_release,
573 &nd_clustq, UMA_ZONE_NOBUCKET);
575 nd_zone_pack = uma_zcache_create("netdump_" MBUF_PACKET_MEM_NAME,
576 MCLBYTES, mb_ctor_pack, mb_dtor_pack, NULL, NULL,
577 nd_pack_import, nd_pack_release,
578 NULL, UMA_ZONE_NOBUCKET);
580 while (nmbuf-- > 0) {
581 m = m_get(MT_DATA, M_WAITOK);
582 uma_zfree(nd_zone_mbuf, m);
584 while (nclust-- > 0) {
585 item = uma_zalloc(m_getzone(nd_clsize), M_WAITOK);
586 uma_zfree(nd_zone_clust, item);
592 * UMA backend page allocator for the jumbo frame zones.
594 * Allocates kernel virtual memory that is backed by contiguous physical
598 mbuf_jumbo_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
602 /* Inform UMA that this allocator uses kernel_map/object. */
603 *flags = UMA_SLAB_KERNEL;
604 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
605 bytes, wait, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
606 VM_MEMATTR_DEFAULT));
610 * Constructor for Mbuf master zone.
612 * The 'arg' pointer points to a mb_args structure which
613 * contains call-specific information required to support the
614 * mbuf allocation API. See mbuf.h.
617 mb_ctor_mbuf(void *mem, int size, void *arg, int how)
620 struct mb_args *args;
626 trash_ctor(mem, size, arg, how);
628 args = (struct mb_args *)arg;
632 * The mbuf is initialized later. The caller has the
633 * responsibility to set up any MAC labels too.
635 if (type == MT_NOINIT)
638 m = (struct mbuf *)mem;
640 MPASS((flags & M_NOFREE) == 0);
642 error = m_init(m, how, type, flags);
648 * The Mbuf master zone destructor.
651 mb_dtor_mbuf(void *mem, int size, void *arg)
656 m = (struct mbuf *)mem;
657 flags = (unsigned long)arg;
659 KASSERT((m->m_flags & M_NOFREE) == 0, ("%s: M_NOFREE set", __func__));
660 if (!(flags & MB_DTOR_SKIP) && (m->m_flags & M_PKTHDR) && !SLIST_EMPTY(&m->m_pkthdr.tags))
661 m_tag_delete_chain(m, NULL);
663 trash_dtor(mem, size, arg);
668 * The Mbuf Packet zone destructor.
671 mb_dtor_pack(void *mem, int size, void *arg)
675 m = (struct mbuf *)mem;
676 if ((m->m_flags & M_PKTHDR) != 0)
677 m_tag_delete_chain(m, NULL);
679 /* Make sure we've got a clean cluster back. */
680 KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__));
681 KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__));
682 KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__));
683 KASSERT(m->m_ext.ext_arg1 == NULL, ("%s: ext_arg1 != NULL", __func__));
684 KASSERT(m->m_ext.ext_arg2 == NULL, ("%s: ext_arg2 != NULL", __func__));
685 KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__));
686 KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__));
688 trash_dtor(m->m_ext.ext_buf, MCLBYTES, arg);
691 * If there are processes blocked on zone_clust, waiting for pages
692 * to be freed up, * cause them to be woken up by draining the
693 * packet zone. We are exposed to a race here * (in the check for
694 * the UMA_ZFLAG_FULL) where we might miss the flag set, but that
695 * is deliberate. We don't want to acquire the zone lock for every
698 if (uma_zone_exhausted_nolock(zone_clust))
699 zone_drain(zone_pack);
703 * The Cluster and Jumbo[PAGESIZE|9|16] zone constructor.
705 * Here the 'arg' pointer points to the Mbuf which we
706 * are configuring cluster storage for. If 'arg' is
707 * empty we allocate just the cluster without setting
708 * the mbuf to it. See mbuf.h.
711 mb_ctor_clust(void *mem, int size, void *arg, int how)
716 trash_ctor(mem, size, arg, how);
718 m = (struct mbuf *)arg;
720 m->m_ext.ext_buf = (char *)mem;
721 m->m_data = m->m_ext.ext_buf;
723 m->m_ext.ext_free = NULL;
724 m->m_ext.ext_arg1 = NULL;
725 m->m_ext.ext_arg2 = NULL;
726 m->m_ext.ext_size = size;
727 m->m_ext.ext_type = m_gettype(size);
728 m->m_ext.ext_flags = EXT_FLAG_EMBREF;
729 m->m_ext.ext_count = 1;
736 * The Packet secondary zone's init routine, executed on the
737 * object's transition from mbuf keg slab to zone cache.
740 mb_zinit_pack(void *mem, int size, int how)
744 m = (struct mbuf *)mem; /* m is virgin. */
745 if (uma_zalloc_arg(zone_clust, m, how) == NULL ||
746 m->m_ext.ext_buf == NULL)
748 m->m_ext.ext_type = EXT_PACKET; /* Override. */
750 trash_init(m->m_ext.ext_buf, MCLBYTES, how);
756 * The Packet secondary zone's fini routine, executed on the
757 * object's transition from zone cache to keg slab.
760 mb_zfini_pack(void *mem, int size)
764 m = (struct mbuf *)mem;
766 trash_fini(m->m_ext.ext_buf, MCLBYTES);
768 uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL);
770 trash_dtor(mem, size, NULL);
775 * The "packet" keg constructor.
778 mb_ctor_pack(void *mem, int size, void *arg, int how)
781 struct mb_args *args;
785 m = (struct mbuf *)mem;
786 args = (struct mb_args *)arg;
789 MPASS((flags & M_NOFREE) == 0);
792 trash_ctor(m->m_ext.ext_buf, MCLBYTES, arg, how);
795 error = m_init(m, how, type, flags);
797 /* m_ext is already initialized. */
798 m->m_data = m->m_ext.ext_buf;
799 m->m_flags = (flags | M_EXT);
805 * This is the protocol drain routine. Called by UMA whenever any of the
806 * mbuf zones is closed to its limit.
808 * No locks should be held when this is called. The drain routines have to
809 * presently acquire some locks which raises the possibility of lock order
813 mb_reclaim(uma_zone_t zone __unused, int pending __unused)
818 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK | WARN_PANIC, NULL, __func__);
820 for (dp = domains; dp != NULL; dp = dp->dom_next)
821 for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++)
822 if (pr->pr_drain != NULL)
827 * Clean up after mbufs with M_EXT storage attached to them if the
828 * reference count hits 1.
831 mb_free_ext(struct mbuf *m)
833 volatile u_int *refcnt;
837 KASSERT(m->m_flags & M_EXT, ("%s: M_EXT not set on %p", __func__, m));
839 /* See if this is the mbuf that holds the embedded refcount. */
840 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
841 refcnt = &m->m_ext.ext_count;
844 KASSERT(m->m_ext.ext_cnt != NULL,
845 ("%s: no refcounting pointer on %p", __func__, m));
846 refcnt = m->m_ext.ext_cnt;
847 mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
851 * Check if the header is embedded in the cluster. It is
852 * important that we can't touch any of the mbuf fields
853 * after we have freed the external storage, since mbuf
854 * could have been embedded in it. For now, the mbufs
855 * embedded into the cluster are always of type EXT_EXTREF,
856 * and for this type we won't free the mref.
858 if (m->m_flags & M_NOFREE) {
860 KASSERT(m->m_ext.ext_type == EXT_EXTREF ||
861 m->m_ext.ext_type == EXT_RXRING,
862 ("%s: no-free mbuf %p has wrong type", __func__, m));
866 /* Free attached storage if this mbuf is the only reference to it. */
867 if (*refcnt == 1 || atomic_fetchadd_int(refcnt, -1) == 1) {
868 switch (m->m_ext.ext_type) {
870 /* The packet zone is special. */
873 uma_zfree(zone_pack, mref);
876 uma_zfree(zone_clust, m->m_ext.ext_buf);
877 uma_zfree(zone_mbuf, mref);
880 uma_zfree(zone_jumbop, m->m_ext.ext_buf);
881 uma_zfree(zone_mbuf, mref);
884 uma_zfree(zone_jumbo9, m->m_ext.ext_buf);
885 uma_zfree(zone_mbuf, mref);
888 uma_zfree(zone_jumbo16, m->m_ext.ext_buf);
889 uma_zfree(zone_mbuf, mref);
895 KASSERT(mref->m_ext.ext_free != NULL,
896 ("%s: ext_free not set", __func__));
897 mref->m_ext.ext_free(mref);
898 uma_zfree(zone_mbuf, mref);
901 KASSERT(m->m_ext.ext_free != NULL,
902 ("%s: ext_free not set", __func__));
903 m->m_ext.ext_free(m);
906 KASSERT(m->m_ext.ext_free == NULL,
907 ("%s: ext_free is set", __func__));
910 KASSERT(m->m_ext.ext_type == 0,
911 ("%s: unknown ext_type", __func__));
915 if (freembuf && m != mref)
916 uma_zfree(zone_mbuf, m);
920 * Official mbuf(9) allocation KPI for stack and drivers:
922 * m_get() - a single mbuf without any attachments, sys/mbuf.h.
923 * m_gethdr() - a single mbuf initialized as M_PKTHDR, sys/mbuf.h.
924 * m_getcl() - an mbuf + 2k cluster, sys/mbuf.h.
925 * m_clget() - attach cluster to already allocated mbuf.
926 * m_cljget() - attach jumbo cluster to already allocated mbuf.
927 * m_get2() - allocate minimum mbuf that would fit size argument.
928 * m_getm2() - allocate a chain of mbufs/clusters.
929 * m_extadd() - attach external cluster to mbuf.
931 * m_free() - free single mbuf with its tags and ext, sys/mbuf.h.
932 * m_freem() - free chain of mbufs.
936 m_clget(struct mbuf *m, int how)
939 KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
941 m->m_ext.ext_buf = (char *)NULL;
942 uma_zalloc_arg(zone_clust, m, how);
944 * On a cluster allocation failure, drain the packet zone and retry,
945 * we might be able to loosen a few clusters up on the drain.
947 if ((how & M_NOWAIT) && (m->m_ext.ext_buf == NULL)) {
948 zone_drain(zone_pack);
949 uma_zalloc_arg(zone_clust, m, how);
951 MBUF_PROBE2(m__clget, m, how);
952 return (m->m_flags & M_EXT);
956 * m_cljget() is different from m_clget() as it can allocate clusters without
957 * attaching them to an mbuf. In that case the return value is the pointer
958 * to the cluster of the requested size. If an mbuf was specified, it gets
959 * the cluster attached to it and the return value can be safely ignored.
960 * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
963 m_cljget(struct mbuf *m, int how, int size)
969 KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
971 m->m_ext.ext_buf = NULL;
974 zone = m_getzone(size);
975 retval = uma_zalloc_arg(zone, m, how);
977 MBUF_PROBE4(m__cljget, m, how, size, retval);
983 * m_get2() allocates minimum mbuf that would fit "size" argument.
986 m_get2(int size, int how, short type, int flags)
994 if (size <= MHLEN || (size <= MLEN && (flags & M_PKTHDR) == 0))
995 return (uma_zalloc_arg(zone_mbuf, &args, how));
996 if (size <= MCLBYTES)
997 return (uma_zalloc_arg(zone_pack, &args, how));
999 if (size > MJUMPAGESIZE)
1002 m = uma_zalloc_arg(zone_mbuf, &args, how);
1006 n = uma_zalloc_arg(zone_jumbop, m, how);
1008 uma_zfree(zone_mbuf, m);
1016 * m_getjcl() returns an mbuf with a cluster of the specified size attached.
1017 * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
1020 m_getjcl(int how, short type, int flags, int size)
1022 struct mb_args args;
1026 if (size == MCLBYTES)
1027 return m_getcl(how, type, flags);
1032 m = uma_zalloc_arg(zone_mbuf, &args, how);
1036 zone = m_getzone(size);
1037 n = uma_zalloc_arg(zone, m, how);
1039 uma_zfree(zone_mbuf, m);
1046 * Allocate a given length worth of mbufs and/or clusters (whatever fits
1047 * best) and return a pointer to the top of the allocated chain. If an
1048 * existing mbuf chain is provided, then we will append the new chain
1049 * to the existing one and return a pointer to the provided mbuf.
1052 m_getm2(struct mbuf *m, int len, int how, short type, int flags)
1054 struct mbuf *mb, *nm = NULL, *mtail = NULL;
1056 KASSERT(len >= 0, ("%s: len is < 0", __func__));
1058 /* Validate flags. */
1059 flags &= (M_PKTHDR | M_EOR);
1061 /* Packet header mbuf must be first in chain. */
1062 if ((flags & M_PKTHDR) && m != NULL)
1065 /* Loop and append maximum sized mbufs to the chain tail. */
1068 mb = m_getjcl(how, type, (flags & M_PKTHDR),
1070 else if (len >= MINCLSIZE)
1071 mb = m_getcl(how, type, (flags & M_PKTHDR));
1072 else if (flags & M_PKTHDR)
1073 mb = m_gethdr(how, type);
1075 mb = m_get(how, type);
1077 /* Fail the whole operation if one mbuf can't be allocated. */
1091 flags &= ~M_PKTHDR; /* Only valid on the first mbuf. */
1094 mtail->m_flags |= M_EOR; /* Only valid on the last mbuf. */
1096 /* If mbuf was supplied, append new chain to the end of it. */
1098 for (mtail = m; mtail->m_next != NULL; mtail = mtail->m_next)
1101 mtail->m_flags &= ~M_EOR;
1109 * Configure a provided mbuf to refer to the provided external storage
1110 * buffer and setup a reference count for said buffer.
1113 * mb The existing mbuf to which to attach the provided buffer.
1114 * buf The address of the provided external storage buffer.
1115 * size The size of the provided buffer.
1116 * freef A pointer to a routine that is responsible for freeing the
1117 * provided external storage buffer.
1118 * args A pointer to an argument structure (of any type) to be passed
1119 * to the provided freef routine (may be NULL).
1120 * flags Any other flags to be passed to the provided mbuf.
1121 * type The type that the external storage buffer should be
1128 m_extadd(struct mbuf *mb, char *buf, u_int size, m_ext_free_t freef,
1129 void *arg1, void *arg2, int flags, int type)
1132 KASSERT(type != EXT_CLUSTER, ("%s: EXT_CLUSTER not allowed", __func__));
1134 mb->m_flags |= (M_EXT | flags);
1135 mb->m_ext.ext_buf = buf;
1136 mb->m_data = mb->m_ext.ext_buf;
1137 mb->m_ext.ext_size = size;
1138 mb->m_ext.ext_free = freef;
1139 mb->m_ext.ext_arg1 = arg1;
1140 mb->m_ext.ext_arg2 = arg2;
1141 mb->m_ext.ext_type = type;
1143 if (type != EXT_EXTREF) {
1144 mb->m_ext.ext_count = 1;
1145 mb->m_ext.ext_flags = EXT_FLAG_EMBREF;
1147 mb->m_ext.ext_flags = 0;
1151 * Free an entire chain of mbufs and associated external buffers, if
1155 m_freem(struct mbuf *mb)
1158 MBUF_PROBE1(m__freem, mb);
1164 m_snd_tag_init(struct m_snd_tag *mst, struct ifnet *ifp)
1169 refcount_init(&mst->refcount, 1);
1170 counter_u64_add(snd_tag_count, 1);
1174 m_snd_tag_destroy(struct m_snd_tag *mst)
1179 ifp->if_snd_tag_free(mst);
1181 counter_u64_add(snd_tag_count, -1);