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>
48 #include <sys/sf_buf.h>
50 #include <sys/socket.h>
51 #include <sys/sysctl.h>
54 #include <net/if_var.h>
57 #include <vm/vm_extern.h>
58 #include <vm/vm_kern.h>
59 #include <vm/vm_page.h>
60 #include <vm/vm_map.h>
62 #include <vm/uma_dbg.h>
65 * In FreeBSD, Mbufs and Mbuf Clusters are allocated from UMA
68 * Mbuf Clusters (2K, contiguous) are allocated from the Cluster
69 * Zone. The Zone can be capped at kern.ipc.nmbclusters, if the
70 * administrator so desires.
72 * Mbufs are allocated from a UMA Master Zone called the Mbuf
75 * Additionally, FreeBSD provides a Packet Zone, which it
76 * configures as a Secondary Zone to the Mbuf Master Zone,
77 * thus sharing backend Slab kegs with the Mbuf Master Zone.
79 * Thus common-case allocations and locking are simplified:
83 * | .------------>[(Packet Cache)] m_get(), m_gethdr()
85 * [(Cluster Cache)] [ Secondary ] [ (Mbuf Cache) ]
86 * [ Cluster Zone ] [ Zone ] [ Mbuf Master Zone ]
92 * \____________(VM)_________________/
95 * Whenever an object is allocated with uma_zalloc() out of
96 * one of the Zones its _ctor_ function is executed. The same
97 * for any deallocation through uma_zfree() the _dtor_ function
100 * Caches are per-CPU and are filled from the Master Zone.
102 * Whenever an object is allocated from the underlying global
103 * memory pool it gets pre-initialized with the _zinit_ functions.
104 * When the Keg's are overfull objects get decommissioned with
105 * _zfini_ functions and free'd back to the global memory pool.
109 int nmbufs; /* limits number of mbufs */
110 int nmbclusters; /* limits number of mbuf clusters */
111 int nmbjumbop; /* limits number of page size jumbo clusters */
112 int nmbjumbo9; /* limits number of 9k jumbo clusters */
113 int nmbjumbo16; /* limits number of 16k jumbo clusters */
115 bool mb_use_ext_pgs; /* use EXT_PGS mbufs for sendfile */
116 SYSCTL_BOOL(_kern_ipc, OID_AUTO, mb_use_ext_pgs, CTLFLAG_RWTUN,
118 "Use unmapped mbufs for sendfile(2)");
120 static quad_t maxmbufmem; /* overall real memory limit for all mbufs */
122 SYSCTL_QUAD(_kern_ipc, OID_AUTO, maxmbufmem, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &maxmbufmem, 0,
123 "Maximum real memory allocatable to various mbuf types");
125 static counter_u64_t snd_tag_count;
126 SYSCTL_COUNTER_U64(_kern_ipc, OID_AUTO, num_snd_tags, CTLFLAG_RW,
127 &snd_tag_count, "# of active mbuf send tags");
130 * tunable_mbinit() has to be run before any mbuf allocations are done.
133 tunable_mbinit(void *dummy)
138 * The default limit for all mbuf related memory is 1/2 of all
139 * available kernel memory (physical or kmem).
140 * At most it can be 3/4 of available kernel memory.
142 realmem = qmin((quad_t)physmem * PAGE_SIZE, vm_kmem_size);
143 maxmbufmem = realmem / 2;
144 TUNABLE_QUAD_FETCH("kern.ipc.maxmbufmem", &maxmbufmem);
145 if (maxmbufmem > realmem / 4 * 3)
146 maxmbufmem = realmem / 4 * 3;
148 TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
149 if (nmbclusters == 0)
150 nmbclusters = maxmbufmem / MCLBYTES / 4;
152 TUNABLE_INT_FETCH("kern.ipc.nmbjumbop", &nmbjumbop);
154 nmbjumbop = maxmbufmem / MJUMPAGESIZE / 4;
156 TUNABLE_INT_FETCH("kern.ipc.nmbjumbo9", &nmbjumbo9);
158 nmbjumbo9 = maxmbufmem / MJUM9BYTES / 6;
160 TUNABLE_INT_FETCH("kern.ipc.nmbjumbo16", &nmbjumbo16);
162 nmbjumbo16 = maxmbufmem / MJUM16BYTES / 6;
165 * We need at least as many mbufs as we have clusters of
166 * the various types added together.
168 TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
169 if (nmbufs < nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16)
170 nmbufs = lmax(maxmbufmem / MSIZE / 5,
171 nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16);
173 SYSINIT(tunable_mbinit, SI_SUB_KMEM, SI_ORDER_MIDDLE, tunable_mbinit, NULL);
176 sysctl_nmbclusters(SYSCTL_HANDLER_ARGS)
178 int error, newnmbclusters;
180 newnmbclusters = nmbclusters;
181 error = sysctl_handle_int(oidp, &newnmbclusters, 0, req);
182 if (error == 0 && req->newptr && newnmbclusters != nmbclusters) {
183 if (newnmbclusters > nmbclusters &&
184 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
185 nmbclusters = newnmbclusters;
186 nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
187 EVENTHANDLER_INVOKE(nmbclusters_change);
193 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbclusters, CTLTYPE_INT|CTLFLAG_RW,
194 &nmbclusters, 0, sysctl_nmbclusters, "IU",
195 "Maximum number of mbuf clusters allowed");
198 sysctl_nmbjumbop(SYSCTL_HANDLER_ARGS)
200 int error, newnmbjumbop;
202 newnmbjumbop = nmbjumbop;
203 error = sysctl_handle_int(oidp, &newnmbjumbop, 0, req);
204 if (error == 0 && req->newptr && newnmbjumbop != nmbjumbop) {
205 if (newnmbjumbop > nmbjumbop &&
206 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
207 nmbjumbop = newnmbjumbop;
208 nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
214 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbop, CTLTYPE_INT|CTLFLAG_RW,
215 &nmbjumbop, 0, sysctl_nmbjumbop, "IU",
216 "Maximum number of mbuf page size jumbo clusters allowed");
219 sysctl_nmbjumbo9(SYSCTL_HANDLER_ARGS)
221 int error, newnmbjumbo9;
223 newnmbjumbo9 = nmbjumbo9;
224 error = sysctl_handle_int(oidp, &newnmbjumbo9, 0, req);
225 if (error == 0 && req->newptr && newnmbjumbo9 != nmbjumbo9) {
226 if (newnmbjumbo9 > nmbjumbo9 &&
227 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
228 nmbjumbo9 = newnmbjumbo9;
229 nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
235 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo9, CTLTYPE_INT|CTLFLAG_RW,
236 &nmbjumbo9, 0, sysctl_nmbjumbo9, "IU",
237 "Maximum number of mbuf 9k jumbo clusters allowed");
240 sysctl_nmbjumbo16(SYSCTL_HANDLER_ARGS)
242 int error, newnmbjumbo16;
244 newnmbjumbo16 = nmbjumbo16;
245 error = sysctl_handle_int(oidp, &newnmbjumbo16, 0, req);
246 if (error == 0 && req->newptr && newnmbjumbo16 != nmbjumbo16) {
247 if (newnmbjumbo16 > nmbjumbo16 &&
248 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
249 nmbjumbo16 = newnmbjumbo16;
250 nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
256 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo16, CTLTYPE_INT|CTLFLAG_RW,
257 &nmbjumbo16, 0, sysctl_nmbjumbo16, "IU",
258 "Maximum number of mbuf 16k jumbo clusters allowed");
261 sysctl_nmbufs(SYSCTL_HANDLER_ARGS)
263 int error, newnmbufs;
266 error = sysctl_handle_int(oidp, &newnmbufs, 0, req);
267 if (error == 0 && req->newptr && newnmbufs != nmbufs) {
268 if (newnmbufs > nmbufs) {
270 nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
271 EVENTHANDLER_INVOKE(nmbufs_change);
277 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbufs, CTLTYPE_INT|CTLFLAG_RW,
278 &nmbufs, 0, sysctl_nmbufs, "IU",
279 "Maximum number of mbufs allowed");
282 * Zones from which we allocate.
284 uma_zone_t zone_mbuf;
285 uma_zone_t zone_clust;
286 uma_zone_t zone_pack;
287 uma_zone_t zone_jumbop;
288 uma_zone_t zone_jumbo9;
289 uma_zone_t zone_jumbo16;
290 uma_zone_t zone_extpgs;
295 static int mb_ctor_mbuf(void *, int, void *, int);
296 static int mb_ctor_clust(void *, int, void *, int);
297 static int mb_ctor_pack(void *, int, void *, int);
298 static void mb_dtor_mbuf(void *, int, void *);
299 static void mb_dtor_pack(void *, int, void *);
300 static int mb_zinit_pack(void *, int, int);
301 static void mb_zfini_pack(void *, int);
302 static void mb_reclaim(uma_zone_t, int);
303 static void *mbuf_jumbo_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
305 /* Ensure that MSIZE is a power of 2. */
306 CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE);
308 _Static_assert(sizeof(struct mbuf_ext_pgs) == 256,
309 "mbuf_ext_pgs size mismatch");
312 * Initialize FreeBSD Network buffer allocation.
315 mbuf_init(void *dummy)
319 * Configure UMA zones for Mbufs, Clusters, and Packets.
321 zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE,
322 mb_ctor_mbuf, mb_dtor_mbuf,
324 trash_init, trash_fini,
328 MSIZE - 1, UMA_ZONE_MAXBUCKET);
330 nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
331 uma_zone_set_warning(zone_mbuf, "kern.ipc.nmbufs limit reached");
332 uma_zone_set_maxaction(zone_mbuf, mb_reclaim);
334 zone_clust = uma_zcreate(MBUF_CLUSTER_MEM_NAME, MCLBYTES,
337 trash_dtor, trash_init, trash_fini,
343 nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
344 uma_zone_set_warning(zone_clust, "kern.ipc.nmbclusters limit reached");
345 uma_zone_set_maxaction(zone_clust, mb_reclaim);
347 zone_pack = uma_zsecond_create(MBUF_PACKET_MEM_NAME, mb_ctor_pack,
348 mb_dtor_pack, mb_zinit_pack, mb_zfini_pack, zone_mbuf);
350 /* Make jumbo frame zone too. Page size, 9k and 16k. */
351 zone_jumbop = uma_zcreate(MBUF_JUMBOP_MEM_NAME, MJUMPAGESIZE,
354 trash_dtor, trash_init, trash_fini,
360 nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
361 uma_zone_set_warning(zone_jumbop, "kern.ipc.nmbjumbop limit reached");
362 uma_zone_set_maxaction(zone_jumbop, mb_reclaim);
364 zone_jumbo9 = uma_zcreate(MBUF_JUMBO9_MEM_NAME, MJUM9BYTES,
367 trash_dtor, trash_init, trash_fini,
372 uma_zone_set_allocf(zone_jumbo9, mbuf_jumbo_alloc);
374 nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
375 uma_zone_set_warning(zone_jumbo9, "kern.ipc.nmbjumbo9 limit reached");
376 uma_zone_set_maxaction(zone_jumbo9, mb_reclaim);
378 zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES,
381 trash_dtor, trash_init, trash_fini,
386 uma_zone_set_allocf(zone_jumbo16, mbuf_jumbo_alloc);
388 nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
389 uma_zone_set_warning(zone_jumbo16, "kern.ipc.nmbjumbo16 limit reached");
390 uma_zone_set_maxaction(zone_jumbo16, mb_reclaim);
392 zone_extpgs = uma_zcreate(MBUF_EXTPGS_MEM_NAME,
393 sizeof(struct mbuf_ext_pgs),
395 trash_ctor, trash_dtor, trash_init, trash_fini,
397 NULL, NULL, NULL, NULL,
402 * Hook event handler for low-memory situation, used to
403 * drain protocols and push data back to the caches (UMA
404 * later pushes it back to VM).
406 EVENTHANDLER_REGISTER(vm_lowmem, mb_reclaim, NULL,
407 EVENTHANDLER_PRI_FIRST);
409 snd_tag_count = counter_u64_alloc(M_WAITOK);
411 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL);
415 * netdump makes use of a pre-allocated pool of mbufs and clusters. When
416 * netdump is configured, we initialize a set of UMA cache zones which return
417 * items from this pool. At panic-time, the regular UMA zone pointers are
418 * overwritten with those of the cache zones so that drivers may allocate and
419 * free mbufs and clusters without attempting to allocate physical memory.
421 * We keep mbufs and clusters in a pair of mbuf queues. In particular, for
422 * the purpose of caching clusters, we treat them as mbufs.
424 static struct mbufq nd_mbufq =
425 { STAILQ_HEAD_INITIALIZER(nd_mbufq.mq_head), 0, INT_MAX };
426 static struct mbufq nd_clustq =
427 { STAILQ_HEAD_INITIALIZER(nd_clustq.mq_head), 0, INT_MAX };
429 static int nd_clsize;
430 static uma_zone_t nd_zone_mbuf;
431 static uma_zone_t nd_zone_clust;
432 static uma_zone_t nd_zone_pack;
435 nd_buf_import(void *arg, void **store, int count, int domain __unused,
444 for (i = 0; i < count; i++) {
445 m = mbufq_dequeue(q);
448 trash_init(m, q == &nd_mbufq ? MSIZE : nd_clsize, flags);
451 KASSERT((flags & M_WAITOK) == 0 || i == count,
452 ("%s: ran out of pre-allocated mbufs", __func__));
457 nd_buf_release(void *arg, void **store, int count)
465 for (i = 0; i < count; i++) {
467 (void)mbufq_enqueue(q, m);
472 nd_pack_import(void *arg __unused, void **store, int count, int domain __unused,
479 for (i = 0; i < count; i++) {
480 m = m_get(MT_DATA, M_NOWAIT);
483 clust = uma_zalloc(nd_zone_clust, M_NOWAIT);
488 mb_ctor_clust(clust, nd_clsize, m, 0);
491 KASSERT((flags & M_WAITOK) == 0 || i == count,
492 ("%s: ran out of pre-allocated mbufs", __func__));
497 nd_pack_release(void *arg __unused, void **store, int count)
503 for (i = 0; i < count; i++) {
505 clust = m->m_ext.ext_buf;
506 uma_zfree(nd_zone_clust, clust);
507 uma_zfree(nd_zone_mbuf, m);
512 * Free the pre-allocated mbufs and clusters reserved for netdump, and destroy
513 * the corresponding UMA cache zones.
516 netdump_mbuf_drain(void)
521 if (nd_zone_mbuf != NULL) {
522 uma_zdestroy(nd_zone_mbuf);
525 if (nd_zone_clust != NULL) {
526 uma_zdestroy(nd_zone_clust);
527 nd_zone_clust = NULL;
529 if (nd_zone_pack != NULL) {
530 uma_zdestroy(nd_zone_pack);
534 while ((m = mbufq_dequeue(&nd_mbufq)) != NULL)
536 while ((item = mbufq_dequeue(&nd_clustq)) != NULL)
537 uma_zfree(m_getzone(nd_clsize), item);
541 * Callback invoked immediately prior to starting a netdump.
544 netdump_mbuf_dump(void)
548 * All cluster zones return buffers of the size requested by the
549 * drivers. It's up to the driver to reinitialize the zones if the
550 * MTU of a netdump-enabled interface changes.
552 printf("netdump: overwriting mbuf zone pointers\n");
553 zone_mbuf = nd_zone_mbuf;
554 zone_clust = nd_zone_clust;
555 zone_pack = nd_zone_pack;
556 zone_jumbop = nd_zone_clust;
557 zone_jumbo9 = nd_zone_clust;
558 zone_jumbo16 = nd_zone_clust;
562 * Reinitialize the netdump mbuf+cluster pool and cache zones.
565 netdump_mbuf_reinit(int nmbuf, int nclust, int clsize)
570 netdump_mbuf_drain();
574 nd_zone_mbuf = uma_zcache_create("netdump_" MBUF_MEM_NAME,
575 MSIZE, mb_ctor_mbuf, mb_dtor_mbuf,
577 trash_init, trash_fini,
581 nd_buf_import, nd_buf_release,
582 &nd_mbufq, UMA_ZONE_NOBUCKET);
584 nd_zone_clust = uma_zcache_create("netdump_" MBUF_CLUSTER_MEM_NAME,
585 clsize, mb_ctor_clust,
587 trash_dtor, trash_init, trash_fini,
591 nd_buf_import, nd_buf_release,
592 &nd_clustq, UMA_ZONE_NOBUCKET);
594 nd_zone_pack = uma_zcache_create("netdump_" MBUF_PACKET_MEM_NAME,
595 MCLBYTES, mb_ctor_pack, mb_dtor_pack, NULL, NULL,
596 nd_pack_import, nd_pack_release,
597 NULL, UMA_ZONE_NOBUCKET);
599 while (nmbuf-- > 0) {
600 m = m_get(MT_DATA, M_WAITOK);
601 uma_zfree(nd_zone_mbuf, m);
603 while (nclust-- > 0) {
604 item = uma_zalloc(m_getzone(nd_clsize), M_WAITOK);
605 uma_zfree(nd_zone_clust, item);
611 * UMA backend page allocator for the jumbo frame zones.
613 * Allocates kernel virtual memory that is backed by contiguous physical
617 mbuf_jumbo_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
621 /* Inform UMA that this allocator uses kernel_map/object. */
622 *flags = UMA_SLAB_KERNEL;
623 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
624 bytes, wait, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
625 VM_MEMATTR_DEFAULT));
629 * Constructor for Mbuf master zone.
631 * The 'arg' pointer points to a mb_args structure which
632 * contains call-specific information required to support the
633 * mbuf allocation API. See mbuf.h.
636 mb_ctor_mbuf(void *mem, int size, void *arg, int how)
639 struct mb_args *args;
645 trash_ctor(mem, size, arg, how);
647 args = (struct mb_args *)arg;
651 * The mbuf is initialized later. The caller has the
652 * responsibility to set up any MAC labels too.
654 if (type == MT_NOINIT)
657 m = (struct mbuf *)mem;
659 MPASS((flags & M_NOFREE) == 0);
661 error = m_init(m, how, type, flags);
667 * The Mbuf master zone destructor.
670 mb_dtor_mbuf(void *mem, int size, void *arg)
675 m = (struct mbuf *)mem;
676 flags = (unsigned long)arg;
678 KASSERT((m->m_flags & M_NOFREE) == 0, ("%s: M_NOFREE set", __func__));
679 if (!(flags & MB_DTOR_SKIP) && (m->m_flags & M_PKTHDR) && !SLIST_EMPTY(&m->m_pkthdr.tags))
680 m_tag_delete_chain(m, NULL);
682 trash_dtor(mem, size, arg);
687 * The Mbuf Packet zone destructor.
690 mb_dtor_pack(void *mem, int size, void *arg)
694 m = (struct mbuf *)mem;
695 if ((m->m_flags & M_PKTHDR) != 0)
696 m_tag_delete_chain(m, NULL);
698 /* Make sure we've got a clean cluster back. */
699 KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__));
700 KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__));
701 KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__));
702 KASSERT(m->m_ext.ext_arg1 == NULL, ("%s: ext_arg1 != NULL", __func__));
703 KASSERT(m->m_ext.ext_arg2 == NULL, ("%s: ext_arg2 != NULL", __func__));
704 KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__));
705 KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__));
707 trash_dtor(m->m_ext.ext_buf, MCLBYTES, arg);
710 * If there are processes blocked on zone_clust, waiting for pages
711 * to be freed up, * cause them to be woken up by draining the
712 * packet zone. We are exposed to a race here * (in the check for
713 * the UMA_ZFLAG_FULL) where we might miss the flag set, but that
714 * is deliberate. We don't want to acquire the zone lock for every
717 if (uma_zone_exhausted_nolock(zone_clust))
718 zone_drain(zone_pack);
722 * The Cluster and Jumbo[PAGESIZE|9|16] zone constructor.
724 * Here the 'arg' pointer points to the Mbuf which we
725 * are configuring cluster storage for. If 'arg' is
726 * empty we allocate just the cluster without setting
727 * the mbuf to it. See mbuf.h.
730 mb_ctor_clust(void *mem, int size, void *arg, int how)
735 trash_ctor(mem, size, arg, how);
737 m = (struct mbuf *)arg;
739 m->m_ext.ext_buf = (char *)mem;
740 m->m_data = m->m_ext.ext_buf;
742 m->m_ext.ext_free = NULL;
743 m->m_ext.ext_arg1 = NULL;
744 m->m_ext.ext_arg2 = NULL;
745 m->m_ext.ext_size = size;
746 m->m_ext.ext_type = m_gettype(size);
747 m->m_ext.ext_flags = EXT_FLAG_EMBREF;
748 m->m_ext.ext_count = 1;
755 * The Packet secondary zone's init routine, executed on the
756 * object's transition from mbuf keg slab to zone cache.
759 mb_zinit_pack(void *mem, int size, int how)
763 m = (struct mbuf *)mem; /* m is virgin. */
764 if (uma_zalloc_arg(zone_clust, m, how) == NULL ||
765 m->m_ext.ext_buf == NULL)
767 m->m_ext.ext_type = EXT_PACKET; /* Override. */
769 trash_init(m->m_ext.ext_buf, MCLBYTES, how);
775 * The Packet secondary zone's fini routine, executed on the
776 * object's transition from zone cache to keg slab.
779 mb_zfini_pack(void *mem, int size)
783 m = (struct mbuf *)mem;
785 trash_fini(m->m_ext.ext_buf, MCLBYTES);
787 uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL);
789 trash_dtor(mem, size, NULL);
794 * The "packet" keg constructor.
797 mb_ctor_pack(void *mem, int size, void *arg, int how)
800 struct mb_args *args;
804 m = (struct mbuf *)mem;
805 args = (struct mb_args *)arg;
808 MPASS((flags & M_NOFREE) == 0);
811 trash_ctor(m->m_ext.ext_buf, MCLBYTES, arg, how);
814 error = m_init(m, how, type, flags);
816 /* m_ext is already initialized. */
817 m->m_data = m->m_ext.ext_buf;
818 m->m_flags = (flags | M_EXT);
824 * This is the protocol drain routine. Called by UMA whenever any of the
825 * mbuf zones is closed to its limit.
827 * No locks should be held when this is called. The drain routines have to
828 * presently acquire some locks which raises the possibility of lock order
832 mb_reclaim(uma_zone_t zone __unused, int pending __unused)
837 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK | WARN_PANIC, NULL, __func__);
839 for (dp = domains; dp != NULL; dp = dp->dom_next)
840 for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++)
841 if (pr->pr_drain != NULL)
846 * Free "count" units of I/O from an mbuf chain. They could be held
847 * in EXT_PGS or just as a normal mbuf. This code is intended to be
848 * called in an error path (I/O error, closed connection, etc).
851 mb_free_notready(struct mbuf *m, int count)
855 for (i = 0; i < count && m != NULL; i++) {
856 if ((m->m_flags & M_EXT) != 0 &&
857 m->m_ext.ext_type == EXT_PGS) {
858 m->m_ext.ext_pgs->nrdy--;
859 if (m->m_ext.ext_pgs->nrdy != 0)
864 KASSERT(i == count, ("Removed only %d items from %p", i, m));
868 * Compress an unmapped mbuf into a simple mbuf when it holds a small
869 * amount of data. This is used as a DOS defense to avoid having
870 * small packets tie up wired pages, an ext_pgs structure, and an
871 * mbuf. Since this converts the existing mbuf in place, it can only
872 * be used if there are no other references to 'm'.
875 mb_unmapped_compress(struct mbuf *m)
877 volatile u_int *refcnt;
881 * Assert that 'm' does not have a packet header. If 'm' had
882 * a packet header, it would only be able to hold MHLEN bytes
883 * and m_data would have to be initialized differently.
885 KASSERT((m->m_flags & M_PKTHDR) == 0 && (m->m_flags & M_EXT) &&
886 m->m_ext.ext_type == EXT_PGS,
887 ("%s: m %p !M_EXT or !EXT_PGS or M_PKTHDR", __func__, m));
888 KASSERT(m->m_len <= MLEN, ("m_len too large %p", m));
890 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
891 refcnt = &m->m_ext.ext_count;
893 KASSERT(m->m_ext.ext_cnt != NULL,
894 ("%s: no refcounting pointer on %p", __func__, m));
895 refcnt = m->m_ext.ext_cnt;
902 * Copy mbuf header and m_ext portion of 'm' to 'm_temp' to
903 * create a "fake" EXT_PGS mbuf that can be used with
904 * m_copydata() as well as the ext_free callback.
906 memcpy(&m_temp, m, offsetof(struct mbuf, m_ext) + sizeof (m->m_ext));
907 m_temp.m_next = NULL;
908 m_temp.m_nextpkt = NULL;
910 /* Turn 'm' into a "normal" mbuf. */
911 m->m_flags &= ~(M_EXT | M_RDONLY | M_NOMAP);
912 m->m_data = m->m_dat;
914 /* Copy data from template's ext_pgs. */
915 m_copydata(&m_temp, 0, m_temp.m_len, mtod(m, caddr_t));
917 /* Free the backing pages. */
918 m_temp.m_ext.ext_free(&m_temp);
920 /* Finally, free the ext_pgs struct. */
921 uma_zfree(zone_extpgs, m_temp.m_ext.ext_pgs);
926 * These next few routines are used to permit downgrading an unmapped
927 * mbuf to a chain of mapped mbufs. This is used when an interface
928 * doesn't supported unmapped mbufs or if checksums need to be
929 * computed in software.
931 * Each unmapped mbuf is converted to a chain of mbufs. First, any
932 * TLS header data is stored in a regular mbuf. Second, each page of
933 * unmapped data is stored in an mbuf with an EXT_SFBUF external
934 * cluster. These mbufs use an sf_buf to provide a valid KVA for the
935 * associated physical page. They also hold a reference on the
936 * original EXT_PGS mbuf to ensure the physical page doesn't go away.
937 * Finally, any TLS trailer data is stored in a regular mbuf.
939 * mb_unmapped_free_mext() is the ext_free handler for the EXT_SFBUF
940 * mbufs. It frees the associated sf_buf and releases its reference
941 * on the original EXT_PGS mbuf.
943 * _mb_unmapped_to_ext() is a helper function that converts a single
944 * unmapped mbuf into a chain of mbufs.
946 * mb_unmapped_to_ext() is the public function that walks an mbuf
947 * chain converting any unmapped mbufs to mapped mbufs. It returns
948 * the new chain of unmapped mbufs on success. On failure it frees
949 * the original mbuf chain and returns NULL.
952 mb_unmapped_free_mext(struct mbuf *m)
957 sf = m->m_ext.ext_arg1;
960 /* Drop the reference on the backing EXT_PGS mbuf. */
961 old_m = m->m_ext.ext_arg2;
966 _mb_unmapped_to_ext(struct mbuf *m)
968 struct mbuf_ext_pgs *ext_pgs;
969 struct mbuf *m_new, *top, *prev, *mref;
972 int i, len, off, pglen, pgoff, seglen, segoff;
973 volatile u_int *refcnt;
976 MBUF_EXT_PGS_ASSERT(m);
977 ext_pgs = m->m_ext.ext_pgs;
979 KASSERT(ext_pgs->tls == NULL, ("%s: can't convert TLS mbuf %p",
982 /* See if this is the mbuf that holds the embedded refcount. */
983 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
984 refcnt = &m->m_ext.ext_count;
987 KASSERT(m->m_ext.ext_cnt != NULL,
988 ("%s: no refcounting pointer on %p", __func__, m));
989 refcnt = m->m_ext.ext_cnt;
990 mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
993 /* Skip over any data removed from the front. */
994 off = mtod(m, vm_offset_t);
997 if (ext_pgs->hdr_len != 0) {
998 if (off >= ext_pgs->hdr_len) {
999 off -= ext_pgs->hdr_len;
1001 seglen = ext_pgs->hdr_len - off;
1003 seglen = min(seglen, len);
1006 m_new = m_get(M_NOWAIT, MT_DATA);
1009 m_new->m_len = seglen;
1011 memcpy(mtod(m_new, void *), &ext_pgs->hdr[segoff],
1015 pgoff = ext_pgs->first_pg_off;
1016 for (i = 0; i < ext_pgs->npgs && len > 0; i++) {
1017 pglen = mbuf_ext_pg_len(ext_pgs, i, pgoff);
1023 seglen = pglen - off;
1024 segoff = pgoff + off;
1026 seglen = min(seglen, len);
1029 pg = PHYS_TO_VM_PAGE(ext_pgs->pa[i]);
1030 m_new = m_get(M_NOWAIT, MT_DATA);
1036 prev->m_next = m_new;
1039 sf = sf_buf_alloc(pg, SFB_NOWAIT);
1044 m_extadd(m_new, (char *)sf_buf_kva(sf), PAGE_SIZE,
1045 mb_unmapped_free_mext, sf, mref, M_RDONLY, EXT_SFBUF);
1046 m_new->m_data += segoff;
1047 m_new->m_len = seglen;
1052 KASSERT((off + len) <= ext_pgs->trail_len,
1053 ("off + len > trail (%d + %d > %d)", off, len,
1054 ext_pgs->trail_len));
1055 m_new = m_get(M_NOWAIT, MT_DATA);
1061 prev->m_next = m_new;
1063 memcpy(mtod(m_new, void *), &ext_pgs->trail[off], len);
1068 * Obtain an additional reference on the old mbuf for
1069 * each created EXT_SFBUF mbuf. They will be dropped
1070 * in mb_unmapped_free_mext().
1075 atomic_add_int(refcnt, ref_inc);
1083 * Obtain an additional reference on the old mbuf for
1084 * each created EXT_SFBUF mbuf. They will be
1085 * immediately dropped when these mbufs are freed
1091 atomic_add_int(refcnt, ref_inc);
1099 mb_unmapped_to_ext(struct mbuf *top)
1101 struct mbuf *m, *next, *prev = NULL;
1104 for (m = top; m != NULL; m = next) {
1105 /* m might be freed, so cache the next pointer. */
1107 if (m->m_flags & M_NOMAP) {
1110 * Remove 'm' from the new chain so
1111 * that the 'top' chain terminates
1112 * before 'm' in case 'top' is freed
1115 prev->m_next = NULL;
1117 m = _mb_unmapped_to_ext(m);
1130 * Replaced one mbuf with a chain, so we must
1131 * find the end of chain.
1145 * Allocate an empty EXT_PGS mbuf. The ext_free routine is
1146 * responsible for freeing any pages backing this mbuf when it is
1150 mb_alloc_ext_pgs(int how, bool pkthdr, m_ext_free_t ext_free)
1153 struct mbuf_ext_pgs *ext_pgs;
1156 m = m_gethdr(how, MT_DATA);
1158 m = m_get(how, MT_DATA);
1162 ext_pgs = uma_zalloc(zone_extpgs, how);
1163 if (ext_pgs == NULL) {
1169 ext_pgs->first_pg_off = 0;
1170 ext_pgs->last_pg_len = 0;
1171 ext_pgs->hdr_len = 0;
1172 ext_pgs->trail_len = 0;
1173 ext_pgs->tls = NULL;
1176 m->m_flags |= (M_EXT | M_RDONLY | M_NOMAP);
1177 m->m_ext.ext_type = EXT_PGS;
1178 m->m_ext.ext_flags = EXT_FLAG_EMBREF;
1179 m->m_ext.ext_count = 1;
1180 m->m_ext.ext_pgs = ext_pgs;
1181 m->m_ext.ext_size = 0;
1182 m->m_ext.ext_free = ext_free;
1186 #ifdef INVARIANT_SUPPORT
1188 mb_ext_pgs_check(struct mbuf_ext_pgs *ext_pgs)
1192 * NB: This expects a non-empty buffer (npgs > 0 and
1195 KASSERT(ext_pgs->npgs > 0,
1196 ("ext_pgs with no valid pages: %p", ext_pgs));
1197 KASSERT(ext_pgs->npgs <= nitems(ext_pgs->pa),
1198 ("ext_pgs with too many pages: %p", ext_pgs));
1199 KASSERT(ext_pgs->nrdy <= ext_pgs->npgs,
1200 ("ext_pgs with too many ready pages: %p", ext_pgs));
1201 KASSERT(ext_pgs->first_pg_off < PAGE_SIZE,
1202 ("ext_pgs with too large page offset: %p", ext_pgs));
1203 KASSERT(ext_pgs->last_pg_len > 0,
1204 ("ext_pgs with zero last page length: %p", ext_pgs));
1205 KASSERT(ext_pgs->last_pg_len <= PAGE_SIZE,
1206 ("ext_pgs with too large last page length: %p", ext_pgs));
1207 if (ext_pgs->npgs == 1) {
1208 KASSERT(ext_pgs->first_pg_off + ext_pgs->last_pg_len <=
1209 PAGE_SIZE, ("ext_pgs with single page too large: %p",
1212 KASSERT(ext_pgs->hdr_len <= sizeof(ext_pgs->hdr),
1213 ("ext_pgs with too large header length: %p", ext_pgs));
1214 KASSERT(ext_pgs->trail_len <= sizeof(ext_pgs->trail),
1215 ("ext_pgs with too large header length: %p", ext_pgs));
1220 * Clean up after mbufs with M_EXT storage attached to them if the
1221 * reference count hits 1.
1224 mb_free_ext(struct mbuf *m)
1226 volatile u_int *refcnt;
1230 KASSERT(m->m_flags & M_EXT, ("%s: M_EXT not set on %p", __func__, m));
1232 /* See if this is the mbuf that holds the embedded refcount. */
1233 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
1234 refcnt = &m->m_ext.ext_count;
1237 KASSERT(m->m_ext.ext_cnt != NULL,
1238 ("%s: no refcounting pointer on %p", __func__, m));
1239 refcnt = m->m_ext.ext_cnt;
1240 mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
1244 * Check if the header is embedded in the cluster. It is
1245 * important that we can't touch any of the mbuf fields
1246 * after we have freed the external storage, since mbuf
1247 * could have been embedded in it. For now, the mbufs
1248 * embedded into the cluster are always of type EXT_EXTREF,
1249 * and for this type we won't free the mref.
1251 if (m->m_flags & M_NOFREE) {
1253 KASSERT(m->m_ext.ext_type == EXT_EXTREF ||
1254 m->m_ext.ext_type == EXT_RXRING,
1255 ("%s: no-free mbuf %p has wrong type", __func__, m));
1259 /* Free attached storage if this mbuf is the only reference to it. */
1260 if (*refcnt == 1 || atomic_fetchadd_int(refcnt, -1) == 1) {
1261 switch (m->m_ext.ext_type) {
1263 /* The packet zone is special. */
1266 uma_zfree(zone_pack, mref);
1269 uma_zfree(zone_clust, m->m_ext.ext_buf);
1270 uma_zfree(zone_mbuf, mref);
1273 uma_zfree(zone_jumbop, m->m_ext.ext_buf);
1274 uma_zfree(zone_mbuf, mref);
1277 uma_zfree(zone_jumbo9, m->m_ext.ext_buf);
1278 uma_zfree(zone_mbuf, mref);
1281 uma_zfree(zone_jumbo16, m->m_ext.ext_buf);
1282 uma_zfree(zone_mbuf, mref);
1285 KASSERT(mref->m_ext.ext_free != NULL,
1286 ("%s: ext_free not set", __func__));
1287 mref->m_ext.ext_free(mref);
1288 uma_zfree(zone_extpgs, mref->m_ext.ext_pgs);
1289 uma_zfree(zone_mbuf, mref);
1294 case EXT_DISPOSABLE:
1295 KASSERT(mref->m_ext.ext_free != NULL,
1296 ("%s: ext_free not set", __func__));
1297 mref->m_ext.ext_free(mref);
1298 uma_zfree(zone_mbuf, mref);
1301 KASSERT(m->m_ext.ext_free != NULL,
1302 ("%s: ext_free not set", __func__));
1303 m->m_ext.ext_free(m);
1306 KASSERT(m->m_ext.ext_free == NULL,
1307 ("%s: ext_free is set", __func__));
1310 KASSERT(m->m_ext.ext_type == 0,
1311 ("%s: unknown ext_type", __func__));
1315 if (freembuf && m != mref)
1316 uma_zfree(zone_mbuf, m);
1320 * Official mbuf(9) allocation KPI for stack and drivers:
1322 * m_get() - a single mbuf without any attachments, sys/mbuf.h.
1323 * m_gethdr() - a single mbuf initialized as M_PKTHDR, sys/mbuf.h.
1324 * m_getcl() - an mbuf + 2k cluster, sys/mbuf.h.
1325 * m_clget() - attach cluster to already allocated mbuf.
1326 * m_cljget() - attach jumbo cluster to already allocated mbuf.
1327 * m_get2() - allocate minimum mbuf that would fit size argument.
1328 * m_getm2() - allocate a chain of mbufs/clusters.
1329 * m_extadd() - attach external cluster to mbuf.
1331 * m_free() - free single mbuf with its tags and ext, sys/mbuf.h.
1332 * m_freem() - free chain of mbufs.
1336 m_clget(struct mbuf *m, int how)
1339 KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
1341 m->m_ext.ext_buf = (char *)NULL;
1342 uma_zalloc_arg(zone_clust, m, how);
1344 * On a cluster allocation failure, drain the packet zone and retry,
1345 * we might be able to loosen a few clusters up on the drain.
1347 if ((how & M_NOWAIT) && (m->m_ext.ext_buf == NULL)) {
1348 zone_drain(zone_pack);
1349 uma_zalloc_arg(zone_clust, m, how);
1351 MBUF_PROBE2(m__clget, m, how);
1352 return (m->m_flags & M_EXT);
1356 * m_cljget() is different from m_clget() as it can allocate clusters without
1357 * attaching them to an mbuf. In that case the return value is the pointer
1358 * to the cluster of the requested size. If an mbuf was specified, it gets
1359 * the cluster attached to it and the return value can be safely ignored.
1360 * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
1363 m_cljget(struct mbuf *m, int how, int size)
1369 KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
1371 m->m_ext.ext_buf = NULL;
1374 zone = m_getzone(size);
1375 retval = uma_zalloc_arg(zone, m, how);
1377 MBUF_PROBE4(m__cljget, m, how, size, retval);
1383 * m_get2() allocates minimum mbuf that would fit "size" argument.
1386 m_get2(int size, int how, short type, int flags)
1388 struct mb_args args;
1394 if (size <= MHLEN || (size <= MLEN && (flags & M_PKTHDR) == 0))
1395 return (uma_zalloc_arg(zone_mbuf, &args, how));
1396 if (size <= MCLBYTES)
1397 return (uma_zalloc_arg(zone_pack, &args, how));
1399 if (size > MJUMPAGESIZE)
1402 m = uma_zalloc_arg(zone_mbuf, &args, how);
1406 n = uma_zalloc_arg(zone_jumbop, m, how);
1408 uma_zfree(zone_mbuf, m);
1416 * m_getjcl() returns an mbuf with a cluster of the specified size attached.
1417 * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
1420 m_getjcl(int how, short type, int flags, int size)
1422 struct mb_args args;
1426 if (size == MCLBYTES)
1427 return m_getcl(how, type, flags);
1432 m = uma_zalloc_arg(zone_mbuf, &args, how);
1436 zone = m_getzone(size);
1437 n = uma_zalloc_arg(zone, m, how);
1439 uma_zfree(zone_mbuf, m);
1446 * Allocate a given length worth of mbufs and/or clusters (whatever fits
1447 * best) and return a pointer to the top of the allocated chain. If an
1448 * existing mbuf chain is provided, then we will append the new chain
1449 * to the existing one and return a pointer to the provided mbuf.
1452 m_getm2(struct mbuf *m, int len, int how, short type, int flags)
1454 struct mbuf *mb, *nm = NULL, *mtail = NULL;
1456 KASSERT(len >= 0, ("%s: len is < 0", __func__));
1458 /* Validate flags. */
1459 flags &= (M_PKTHDR | M_EOR);
1461 /* Packet header mbuf must be first in chain. */
1462 if ((flags & M_PKTHDR) && m != NULL)
1465 /* Loop and append maximum sized mbufs to the chain tail. */
1468 mb = m_getjcl(how, type, (flags & M_PKTHDR),
1470 else if (len >= MINCLSIZE)
1471 mb = m_getcl(how, type, (flags & M_PKTHDR));
1472 else if (flags & M_PKTHDR)
1473 mb = m_gethdr(how, type);
1475 mb = m_get(how, type);
1477 /* Fail the whole operation if one mbuf can't be allocated. */
1491 flags &= ~M_PKTHDR; /* Only valid on the first mbuf. */
1494 mtail->m_flags |= M_EOR; /* Only valid on the last mbuf. */
1496 /* If mbuf was supplied, append new chain to the end of it. */
1498 for (mtail = m; mtail->m_next != NULL; mtail = mtail->m_next)
1501 mtail->m_flags &= ~M_EOR;
1509 * Configure a provided mbuf to refer to the provided external storage
1510 * buffer and setup a reference count for said buffer.
1513 * mb The existing mbuf to which to attach the provided buffer.
1514 * buf The address of the provided external storage buffer.
1515 * size The size of the provided buffer.
1516 * freef A pointer to a routine that is responsible for freeing the
1517 * provided external storage buffer.
1518 * args A pointer to an argument structure (of any type) to be passed
1519 * to the provided freef routine (may be NULL).
1520 * flags Any other flags to be passed to the provided mbuf.
1521 * type The type that the external storage buffer should be
1528 m_extadd(struct mbuf *mb, char *buf, u_int size, m_ext_free_t freef,
1529 void *arg1, void *arg2, int flags, int type)
1532 KASSERT(type != EXT_CLUSTER, ("%s: EXT_CLUSTER not allowed", __func__));
1534 mb->m_flags |= (M_EXT | flags);
1535 mb->m_ext.ext_buf = buf;
1536 mb->m_data = mb->m_ext.ext_buf;
1537 mb->m_ext.ext_size = size;
1538 mb->m_ext.ext_free = freef;
1539 mb->m_ext.ext_arg1 = arg1;
1540 mb->m_ext.ext_arg2 = arg2;
1541 mb->m_ext.ext_type = type;
1543 if (type != EXT_EXTREF) {
1544 mb->m_ext.ext_count = 1;
1545 mb->m_ext.ext_flags = EXT_FLAG_EMBREF;
1547 mb->m_ext.ext_flags = 0;
1551 * Free an entire chain of mbufs and associated external buffers, if
1555 m_freem(struct mbuf *mb)
1558 MBUF_PROBE1(m__freem, mb);
1564 m_snd_tag_init(struct m_snd_tag *mst, struct ifnet *ifp)
1569 refcount_init(&mst->refcount, 1);
1570 counter_u64_add(snd_tag_count, 1);
1574 m_snd_tag_destroy(struct m_snd_tag *mst)
1579 ifp->if_snd_tag_free(mst);
1581 counter_u64_add(snd_tag_count, -1);