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 static quad_t maxmbufmem; /* overall real memory limit for all mbufs */
117 SYSCTL_QUAD(_kern_ipc, OID_AUTO, maxmbufmem, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &maxmbufmem, 0,
118 "Maximum real memory allocatable to various mbuf types");
120 static counter_u64_t snd_tag_count;
121 SYSCTL_COUNTER_U64(_kern_ipc, OID_AUTO, num_snd_tags, CTLFLAG_RW,
122 &snd_tag_count, "# of active mbuf send tags");
125 * tunable_mbinit() has to be run before any mbuf allocations are done.
128 tunable_mbinit(void *dummy)
133 * The default limit for all mbuf related memory is 1/2 of all
134 * available kernel memory (physical or kmem).
135 * At most it can be 3/4 of available kernel memory.
137 realmem = qmin((quad_t)physmem * PAGE_SIZE, vm_kmem_size);
138 maxmbufmem = realmem / 2;
139 TUNABLE_QUAD_FETCH("kern.ipc.maxmbufmem", &maxmbufmem);
140 if (maxmbufmem > realmem / 4 * 3)
141 maxmbufmem = realmem / 4 * 3;
143 TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
144 if (nmbclusters == 0)
145 nmbclusters = maxmbufmem / MCLBYTES / 4;
147 TUNABLE_INT_FETCH("kern.ipc.nmbjumbop", &nmbjumbop);
149 nmbjumbop = maxmbufmem / MJUMPAGESIZE / 4;
151 TUNABLE_INT_FETCH("kern.ipc.nmbjumbo9", &nmbjumbo9);
153 nmbjumbo9 = maxmbufmem / MJUM9BYTES / 6;
155 TUNABLE_INT_FETCH("kern.ipc.nmbjumbo16", &nmbjumbo16);
157 nmbjumbo16 = maxmbufmem / MJUM16BYTES / 6;
160 * We need at least as many mbufs as we have clusters of
161 * the various types added together.
163 TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
164 if (nmbufs < nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16)
165 nmbufs = lmax(maxmbufmem / MSIZE / 5,
166 nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16);
168 SYSINIT(tunable_mbinit, SI_SUB_KMEM, SI_ORDER_MIDDLE, tunable_mbinit, NULL);
171 sysctl_nmbclusters(SYSCTL_HANDLER_ARGS)
173 int error, newnmbclusters;
175 newnmbclusters = nmbclusters;
176 error = sysctl_handle_int(oidp, &newnmbclusters, 0, req);
177 if (error == 0 && req->newptr && newnmbclusters != nmbclusters) {
178 if (newnmbclusters > nmbclusters &&
179 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
180 nmbclusters = newnmbclusters;
181 nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
182 EVENTHANDLER_INVOKE(nmbclusters_change);
188 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbclusters, CTLTYPE_INT|CTLFLAG_RW,
189 &nmbclusters, 0, sysctl_nmbclusters, "IU",
190 "Maximum number of mbuf clusters allowed");
193 sysctl_nmbjumbop(SYSCTL_HANDLER_ARGS)
195 int error, newnmbjumbop;
197 newnmbjumbop = nmbjumbop;
198 error = sysctl_handle_int(oidp, &newnmbjumbop, 0, req);
199 if (error == 0 && req->newptr && newnmbjumbop != nmbjumbop) {
200 if (newnmbjumbop > nmbjumbop &&
201 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
202 nmbjumbop = newnmbjumbop;
203 nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
209 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbop, CTLTYPE_INT|CTLFLAG_RW,
210 &nmbjumbop, 0, sysctl_nmbjumbop, "IU",
211 "Maximum number of mbuf page size jumbo clusters allowed");
214 sysctl_nmbjumbo9(SYSCTL_HANDLER_ARGS)
216 int error, newnmbjumbo9;
218 newnmbjumbo9 = nmbjumbo9;
219 error = sysctl_handle_int(oidp, &newnmbjumbo9, 0, req);
220 if (error == 0 && req->newptr && newnmbjumbo9 != nmbjumbo9) {
221 if (newnmbjumbo9 > nmbjumbo9 &&
222 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
223 nmbjumbo9 = newnmbjumbo9;
224 nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
230 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo9, CTLTYPE_INT|CTLFLAG_RW,
231 &nmbjumbo9, 0, sysctl_nmbjumbo9, "IU",
232 "Maximum number of mbuf 9k jumbo clusters allowed");
235 sysctl_nmbjumbo16(SYSCTL_HANDLER_ARGS)
237 int error, newnmbjumbo16;
239 newnmbjumbo16 = nmbjumbo16;
240 error = sysctl_handle_int(oidp, &newnmbjumbo16, 0, req);
241 if (error == 0 && req->newptr && newnmbjumbo16 != nmbjumbo16) {
242 if (newnmbjumbo16 > nmbjumbo16 &&
243 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
244 nmbjumbo16 = newnmbjumbo16;
245 nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
251 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo16, CTLTYPE_INT|CTLFLAG_RW,
252 &nmbjumbo16, 0, sysctl_nmbjumbo16, "IU",
253 "Maximum number of mbuf 16k jumbo clusters allowed");
256 sysctl_nmbufs(SYSCTL_HANDLER_ARGS)
258 int error, newnmbufs;
261 error = sysctl_handle_int(oidp, &newnmbufs, 0, req);
262 if (error == 0 && req->newptr && newnmbufs != nmbufs) {
263 if (newnmbufs > nmbufs) {
265 nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
266 EVENTHANDLER_INVOKE(nmbufs_change);
272 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbufs, CTLTYPE_INT|CTLFLAG_RW,
273 &nmbufs, 0, sysctl_nmbufs, "IU",
274 "Maximum number of mbufs allowed");
277 * Zones from which we allocate.
279 uma_zone_t zone_mbuf;
280 uma_zone_t zone_clust;
281 uma_zone_t zone_pack;
282 uma_zone_t zone_jumbop;
283 uma_zone_t zone_jumbo9;
284 uma_zone_t zone_jumbo16;
285 uma_zone_t zone_extpgs;
290 static int mb_ctor_mbuf(void *, int, void *, int);
291 static int mb_ctor_clust(void *, int, void *, int);
292 static int mb_ctor_pack(void *, int, void *, int);
293 static void mb_dtor_mbuf(void *, int, void *);
294 static void mb_dtor_pack(void *, int, void *);
295 static int mb_zinit_pack(void *, int, int);
296 static void mb_zfini_pack(void *, int);
297 static void mb_reclaim(uma_zone_t, int);
298 static void *mbuf_jumbo_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
300 /* Ensure that MSIZE is a power of 2. */
301 CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE);
303 _Static_assert(sizeof(struct mbuf_ext_pgs) == 256,
304 "mbuf_ext_pgs size mismatch");
307 * Initialize FreeBSD Network buffer allocation.
310 mbuf_init(void *dummy)
314 * Configure UMA zones for Mbufs, Clusters, and Packets.
316 zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE,
317 mb_ctor_mbuf, mb_dtor_mbuf,
319 trash_init, trash_fini,
323 MSIZE - 1, UMA_ZONE_MAXBUCKET);
325 nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
326 uma_zone_set_warning(zone_mbuf, "kern.ipc.nmbufs limit reached");
327 uma_zone_set_maxaction(zone_mbuf, mb_reclaim);
329 zone_clust = uma_zcreate(MBUF_CLUSTER_MEM_NAME, MCLBYTES,
332 trash_dtor, trash_init, trash_fini,
338 nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
339 uma_zone_set_warning(zone_clust, "kern.ipc.nmbclusters limit reached");
340 uma_zone_set_maxaction(zone_clust, mb_reclaim);
342 zone_pack = uma_zsecond_create(MBUF_PACKET_MEM_NAME, mb_ctor_pack,
343 mb_dtor_pack, mb_zinit_pack, mb_zfini_pack, zone_mbuf);
345 /* Make jumbo frame zone too. Page size, 9k and 16k. */
346 zone_jumbop = uma_zcreate(MBUF_JUMBOP_MEM_NAME, MJUMPAGESIZE,
349 trash_dtor, trash_init, trash_fini,
355 nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
356 uma_zone_set_warning(zone_jumbop, "kern.ipc.nmbjumbop limit reached");
357 uma_zone_set_maxaction(zone_jumbop, mb_reclaim);
359 zone_jumbo9 = uma_zcreate(MBUF_JUMBO9_MEM_NAME, MJUM9BYTES,
362 trash_dtor, trash_init, trash_fini,
367 uma_zone_set_allocf(zone_jumbo9, mbuf_jumbo_alloc);
369 nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
370 uma_zone_set_warning(zone_jumbo9, "kern.ipc.nmbjumbo9 limit reached");
371 uma_zone_set_maxaction(zone_jumbo9, mb_reclaim);
373 zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES,
376 trash_dtor, trash_init, trash_fini,
381 uma_zone_set_allocf(zone_jumbo16, mbuf_jumbo_alloc);
383 nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
384 uma_zone_set_warning(zone_jumbo16, "kern.ipc.nmbjumbo16 limit reached");
385 uma_zone_set_maxaction(zone_jumbo16, mb_reclaim);
387 zone_extpgs = uma_zcreate(MBUF_EXTPGS_MEM_NAME,
388 sizeof(struct mbuf_ext_pgs),
390 trash_ctor, trash_dtor, trash_init, trash_fini,
392 NULL, NULL, NULL, NULL,
397 * Hook event handler for low-memory situation, used to
398 * drain protocols and push data back to the caches (UMA
399 * later pushes it back to VM).
401 EVENTHANDLER_REGISTER(vm_lowmem, mb_reclaim, NULL,
402 EVENTHANDLER_PRI_FIRST);
404 snd_tag_count = counter_u64_alloc(M_WAITOK);
406 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL);
410 * netdump makes use of a pre-allocated pool of mbufs and clusters. When
411 * netdump is configured, we initialize a set of UMA cache zones which return
412 * items from this pool. At panic-time, the regular UMA zone pointers are
413 * overwritten with those of the cache zones so that drivers may allocate and
414 * free mbufs and clusters without attempting to allocate physical memory.
416 * We keep mbufs and clusters in a pair of mbuf queues. In particular, for
417 * the purpose of caching clusters, we treat them as mbufs.
419 static struct mbufq nd_mbufq =
420 { STAILQ_HEAD_INITIALIZER(nd_mbufq.mq_head), 0, INT_MAX };
421 static struct mbufq nd_clustq =
422 { STAILQ_HEAD_INITIALIZER(nd_clustq.mq_head), 0, INT_MAX };
424 static int nd_clsize;
425 static uma_zone_t nd_zone_mbuf;
426 static uma_zone_t nd_zone_clust;
427 static uma_zone_t nd_zone_pack;
430 nd_buf_import(void *arg, void **store, int count, int domain __unused,
439 for (i = 0; i < count; i++) {
440 m = mbufq_dequeue(q);
443 trash_init(m, q == &nd_mbufq ? MSIZE : nd_clsize, flags);
446 KASSERT((flags & M_WAITOK) == 0 || i == count,
447 ("%s: ran out of pre-allocated mbufs", __func__));
452 nd_buf_release(void *arg, void **store, int count)
460 for (i = 0; i < count; i++) {
462 (void)mbufq_enqueue(q, m);
467 nd_pack_import(void *arg __unused, void **store, int count, int domain __unused,
474 for (i = 0; i < count; i++) {
475 m = m_get(MT_DATA, M_NOWAIT);
478 clust = uma_zalloc(nd_zone_clust, M_NOWAIT);
483 mb_ctor_clust(clust, nd_clsize, m, 0);
486 KASSERT((flags & M_WAITOK) == 0 || i == count,
487 ("%s: ran out of pre-allocated mbufs", __func__));
492 nd_pack_release(void *arg __unused, void **store, int count)
498 for (i = 0; i < count; i++) {
500 clust = m->m_ext.ext_buf;
501 uma_zfree(nd_zone_clust, clust);
502 uma_zfree(nd_zone_mbuf, m);
507 * Free the pre-allocated mbufs and clusters reserved for netdump, and destroy
508 * the corresponding UMA cache zones.
511 netdump_mbuf_drain(void)
516 if (nd_zone_mbuf != NULL) {
517 uma_zdestroy(nd_zone_mbuf);
520 if (nd_zone_clust != NULL) {
521 uma_zdestroy(nd_zone_clust);
522 nd_zone_clust = NULL;
524 if (nd_zone_pack != NULL) {
525 uma_zdestroy(nd_zone_pack);
529 while ((m = mbufq_dequeue(&nd_mbufq)) != NULL)
531 while ((item = mbufq_dequeue(&nd_clustq)) != NULL)
532 uma_zfree(m_getzone(nd_clsize), item);
536 * Callback invoked immediately prior to starting a netdump.
539 netdump_mbuf_dump(void)
543 * All cluster zones return buffers of the size requested by the
544 * drivers. It's up to the driver to reinitialize the zones if the
545 * MTU of a netdump-enabled interface changes.
547 printf("netdump: overwriting mbuf zone pointers\n");
548 zone_mbuf = nd_zone_mbuf;
549 zone_clust = nd_zone_clust;
550 zone_pack = nd_zone_pack;
551 zone_jumbop = nd_zone_clust;
552 zone_jumbo9 = nd_zone_clust;
553 zone_jumbo16 = nd_zone_clust;
557 * Reinitialize the netdump mbuf+cluster pool and cache zones.
560 netdump_mbuf_reinit(int nmbuf, int nclust, int clsize)
565 netdump_mbuf_drain();
569 nd_zone_mbuf = uma_zcache_create("netdump_" MBUF_MEM_NAME,
570 MSIZE, mb_ctor_mbuf, mb_dtor_mbuf,
572 trash_init, trash_fini,
576 nd_buf_import, nd_buf_release,
577 &nd_mbufq, UMA_ZONE_NOBUCKET);
579 nd_zone_clust = uma_zcache_create("netdump_" MBUF_CLUSTER_MEM_NAME,
580 clsize, mb_ctor_clust,
582 trash_dtor, trash_init, trash_fini,
586 nd_buf_import, nd_buf_release,
587 &nd_clustq, UMA_ZONE_NOBUCKET);
589 nd_zone_pack = uma_zcache_create("netdump_" MBUF_PACKET_MEM_NAME,
590 MCLBYTES, mb_ctor_pack, mb_dtor_pack, NULL, NULL,
591 nd_pack_import, nd_pack_release,
592 NULL, UMA_ZONE_NOBUCKET);
594 while (nmbuf-- > 0) {
595 m = m_get(MT_DATA, M_WAITOK);
596 uma_zfree(nd_zone_mbuf, m);
598 while (nclust-- > 0) {
599 item = uma_zalloc(m_getzone(nd_clsize), M_WAITOK);
600 uma_zfree(nd_zone_clust, item);
606 * UMA backend page allocator for the jumbo frame zones.
608 * Allocates kernel virtual memory that is backed by contiguous physical
612 mbuf_jumbo_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
616 /* Inform UMA that this allocator uses kernel_map/object. */
617 *flags = UMA_SLAB_KERNEL;
618 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
619 bytes, wait, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
620 VM_MEMATTR_DEFAULT));
624 * Constructor for Mbuf master zone.
626 * The 'arg' pointer points to a mb_args structure which
627 * contains call-specific information required to support the
628 * mbuf allocation API. See mbuf.h.
631 mb_ctor_mbuf(void *mem, int size, void *arg, int how)
634 struct mb_args *args;
640 trash_ctor(mem, size, arg, how);
642 args = (struct mb_args *)arg;
646 * The mbuf is initialized later. The caller has the
647 * responsibility to set up any MAC labels too.
649 if (type == MT_NOINIT)
652 m = (struct mbuf *)mem;
654 MPASS((flags & M_NOFREE) == 0);
656 error = m_init(m, how, type, flags);
662 * The Mbuf master zone destructor.
665 mb_dtor_mbuf(void *mem, int size, void *arg)
670 m = (struct mbuf *)mem;
671 flags = (unsigned long)arg;
673 KASSERT((m->m_flags & M_NOFREE) == 0, ("%s: M_NOFREE set", __func__));
674 if (!(flags & MB_DTOR_SKIP) && (m->m_flags & M_PKTHDR) && !SLIST_EMPTY(&m->m_pkthdr.tags))
675 m_tag_delete_chain(m, NULL);
677 trash_dtor(mem, size, arg);
682 * The Mbuf Packet zone destructor.
685 mb_dtor_pack(void *mem, int size, void *arg)
689 m = (struct mbuf *)mem;
690 if ((m->m_flags & M_PKTHDR) != 0)
691 m_tag_delete_chain(m, NULL);
693 /* Make sure we've got a clean cluster back. */
694 KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__));
695 KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__));
696 KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__));
697 KASSERT(m->m_ext.ext_arg1 == NULL, ("%s: ext_arg1 != NULL", __func__));
698 KASSERT(m->m_ext.ext_arg2 == NULL, ("%s: ext_arg2 != NULL", __func__));
699 KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__));
700 KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__));
702 trash_dtor(m->m_ext.ext_buf, MCLBYTES, arg);
705 * If there are processes blocked on zone_clust, waiting for pages
706 * to be freed up, * cause them to be woken up by draining the
707 * packet zone. We are exposed to a race here * (in the check for
708 * the UMA_ZFLAG_FULL) where we might miss the flag set, but that
709 * is deliberate. We don't want to acquire the zone lock for every
712 if (uma_zone_exhausted_nolock(zone_clust))
713 zone_drain(zone_pack);
717 * The Cluster and Jumbo[PAGESIZE|9|16] zone constructor.
719 * Here the 'arg' pointer points to the Mbuf which we
720 * are configuring cluster storage for. If 'arg' is
721 * empty we allocate just the cluster without setting
722 * the mbuf to it. See mbuf.h.
725 mb_ctor_clust(void *mem, int size, void *arg, int how)
730 trash_ctor(mem, size, arg, how);
732 m = (struct mbuf *)arg;
734 m->m_ext.ext_buf = (char *)mem;
735 m->m_data = m->m_ext.ext_buf;
737 m->m_ext.ext_free = NULL;
738 m->m_ext.ext_arg1 = NULL;
739 m->m_ext.ext_arg2 = NULL;
740 m->m_ext.ext_size = size;
741 m->m_ext.ext_type = m_gettype(size);
742 m->m_ext.ext_flags = EXT_FLAG_EMBREF;
743 m->m_ext.ext_count = 1;
750 * The Packet secondary zone's init routine, executed on the
751 * object's transition from mbuf keg slab to zone cache.
754 mb_zinit_pack(void *mem, int size, int how)
758 m = (struct mbuf *)mem; /* m is virgin. */
759 if (uma_zalloc_arg(zone_clust, m, how) == NULL ||
760 m->m_ext.ext_buf == NULL)
762 m->m_ext.ext_type = EXT_PACKET; /* Override. */
764 trash_init(m->m_ext.ext_buf, MCLBYTES, how);
770 * The Packet secondary zone's fini routine, executed on the
771 * object's transition from zone cache to keg slab.
774 mb_zfini_pack(void *mem, int size)
778 m = (struct mbuf *)mem;
780 trash_fini(m->m_ext.ext_buf, MCLBYTES);
782 uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL);
784 trash_dtor(mem, size, NULL);
789 * The "packet" keg constructor.
792 mb_ctor_pack(void *mem, int size, void *arg, int how)
795 struct mb_args *args;
799 m = (struct mbuf *)mem;
800 args = (struct mb_args *)arg;
803 MPASS((flags & M_NOFREE) == 0);
806 trash_ctor(m->m_ext.ext_buf, MCLBYTES, arg, how);
809 error = m_init(m, how, type, flags);
811 /* m_ext is already initialized. */
812 m->m_data = m->m_ext.ext_buf;
813 m->m_flags = (flags | M_EXT);
819 * This is the protocol drain routine. Called by UMA whenever any of the
820 * mbuf zones is closed to its limit.
822 * No locks should be held when this is called. The drain routines have to
823 * presently acquire some locks which raises the possibility of lock order
827 mb_reclaim(uma_zone_t zone __unused, int pending __unused)
832 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK | WARN_PANIC, NULL, __func__);
834 for (dp = domains; dp != NULL; dp = dp->dom_next)
835 for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++)
836 if (pr->pr_drain != NULL)
841 * Free "count" units of I/O from an mbuf chain. They could be held
842 * in EXT_PGS or just as a normal mbuf. This code is intended to be
843 * called in an error path (I/O error, closed connection, etc).
846 mb_free_notready(struct mbuf *m, int count)
850 for (i = 0; i < count && m != NULL; i++) {
851 if ((m->m_flags & M_EXT) != 0 &&
852 m->m_ext.ext_type == EXT_PGS) {
853 m->m_ext.ext_pgs->nrdy--;
854 if (m->m_ext.ext_pgs->nrdy != 0)
859 KASSERT(i == count, ("Removed only %d items from %p", i, m));
863 * Compress an unmapped mbuf into a simple mbuf when it holds a small
864 * amount of data. This is used as a DOS defense to avoid having
865 * small packets tie up wired pages, an ext_pgs structure, and an
866 * mbuf. Since this converts the existing mbuf in place, it can only
867 * be used if there are no other references to 'm'.
870 mb_unmapped_compress(struct mbuf *m)
872 volatile u_int *refcnt;
876 * Assert that 'm' does not have a packet header. If 'm' had
877 * a packet header, it would only be able to hold MHLEN bytes
878 * and m_data would have to be initialized differently.
880 KASSERT((m->m_flags & M_PKTHDR) == 0 && (m->m_flags & M_EXT) &&
881 m->m_ext.ext_type == EXT_PGS,
882 ("%s: m %p !M_EXT or !EXT_PGS or M_PKTHDR", __func__, m));
883 KASSERT(m->m_len <= MLEN, ("m_len too large %p", m));
885 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
886 refcnt = &m->m_ext.ext_count;
888 KASSERT(m->m_ext.ext_cnt != NULL,
889 ("%s: no refcounting pointer on %p", __func__, m));
890 refcnt = m->m_ext.ext_cnt;
897 * Copy mbuf header and m_ext portion of 'm' to 'm_temp' to
898 * create a "fake" EXT_PGS mbuf that can be used with
899 * m_copydata() as well as the ext_free callback.
901 memcpy(&m_temp, m, offsetof(struct mbuf, m_ext) + sizeof (m->m_ext));
902 m_temp.m_next = NULL;
903 m_temp.m_nextpkt = NULL;
905 /* Turn 'm' into a "normal" mbuf. */
906 m->m_flags &= ~(M_EXT | M_RDONLY | M_NOMAP);
907 m->m_data = m->m_dat;
909 /* Copy data from template's ext_pgs. */
910 m_copydata(&m_temp, 0, m_temp.m_len, mtod(m, caddr_t));
912 /* Free the backing pages. */
913 m_temp.m_ext.ext_free(&m_temp);
915 /* Finally, free the ext_pgs struct. */
916 uma_zfree(zone_extpgs, m_temp.m_ext.ext_pgs);
921 * These next few routines are used to permit downgrading an unmapped
922 * mbuf to a chain of mapped mbufs. This is used when an interface
923 * doesn't supported unmapped mbufs or if checksums need to be
924 * computed in software.
926 * Each unmapped mbuf is converted to a chain of mbufs. First, any
927 * TLS header data is stored in a regular mbuf. Second, each page of
928 * unmapped data is stored in an mbuf with an EXT_SFBUF external
929 * cluster. These mbufs use an sf_buf to provide a valid KVA for the
930 * associated physical page. They also hold a reference on the
931 * original EXT_PGS mbuf to ensure the physical page doesn't go away.
932 * Finally, any TLS trailer data is stored in a regular mbuf.
934 * mb_unmapped_free_mext() is the ext_free handler for the EXT_SFBUF
935 * mbufs. It frees the associated sf_buf and releases its reference
936 * on the original EXT_PGS mbuf.
938 * _mb_unmapped_to_ext() is a helper function that converts a single
939 * unmapped mbuf into a chain of mbufs.
941 * mb_unmapped_to_ext() is the public function that walks an mbuf
942 * chain converting any unmapped mbufs to mapped mbufs. It returns
943 * the new chain of unmapped mbufs on success. On failure it frees
944 * the original mbuf chain and returns NULL.
947 mb_unmapped_free_mext(struct mbuf *m)
952 sf = m->m_ext.ext_arg1;
955 /* Drop the reference on the backing EXT_PGS mbuf. */
956 old_m = m->m_ext.ext_arg2;
961 _mb_unmapped_to_ext(struct mbuf *m)
963 struct mbuf_ext_pgs *ext_pgs;
964 struct mbuf *m_new, *top, *prev, *mref;
967 int i, len, off, pglen, pgoff, seglen, segoff;
968 volatile u_int *refcnt;
971 MBUF_EXT_PGS_ASSERT(m);
972 ext_pgs = m->m_ext.ext_pgs;
974 KASSERT(ext_pgs->tls == NULL, ("%s: can't convert TLS mbuf %p",
977 /* See if this is the mbuf that holds the embedded refcount. */
978 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
979 refcnt = &m->m_ext.ext_count;
982 KASSERT(m->m_ext.ext_cnt != NULL,
983 ("%s: no refcounting pointer on %p", __func__, m));
984 refcnt = m->m_ext.ext_cnt;
985 mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
988 /* Skip over any data removed from the front. */
989 off = mtod(m, vm_offset_t);
992 if (ext_pgs->hdr_len != 0) {
993 if (off >= ext_pgs->hdr_len) {
994 off -= ext_pgs->hdr_len;
996 seglen = ext_pgs->hdr_len - off;
998 seglen = min(seglen, len);
1001 m_new = m_get(M_NOWAIT, MT_DATA);
1004 m_new->m_len = seglen;
1006 memcpy(mtod(m_new, void *), &ext_pgs->hdr[segoff],
1010 pgoff = ext_pgs->first_pg_off;
1011 for (i = 0; i < ext_pgs->npgs && len > 0; i++) {
1012 pglen = mbuf_ext_pg_len(ext_pgs, i, pgoff);
1018 seglen = pglen - off;
1019 segoff = pgoff + off;
1021 seglen = min(seglen, len);
1024 pg = PHYS_TO_VM_PAGE(ext_pgs->pa[i]);
1025 m_new = m_get(M_NOWAIT, MT_DATA);
1031 prev->m_next = m_new;
1034 sf = sf_buf_alloc(pg, SFB_NOWAIT);
1039 m_extadd(m_new, (char *)sf_buf_kva(sf), PAGE_SIZE,
1040 mb_unmapped_free_mext, sf, mref, M_RDONLY, EXT_SFBUF);
1041 m_new->m_data += segoff;
1042 m_new->m_len = seglen;
1047 KASSERT((off + len) <= ext_pgs->trail_len,
1048 ("off + len > trail (%d + %d > %d)", off, len,
1049 ext_pgs->trail_len));
1050 m_new = m_get(M_NOWAIT, MT_DATA);
1056 prev->m_next = m_new;
1058 memcpy(mtod(m_new, void *), &ext_pgs->trail[off], len);
1063 * Obtain an additional reference on the old mbuf for
1064 * each created EXT_SFBUF mbuf. They will be dropped
1065 * in mb_unmapped_free_mext().
1070 atomic_add_int(refcnt, ref_inc);
1078 * Obtain an additional reference on the old mbuf for
1079 * each created EXT_SFBUF mbuf. They will be
1080 * immediately dropped when these mbufs are freed
1086 atomic_add_int(refcnt, ref_inc);
1094 mb_unmapped_to_ext(struct mbuf *top)
1096 struct mbuf *m, *next, *prev = NULL;
1099 for (m = top; m != NULL; m = next) {
1100 /* m might be freed, so cache the next pointer. */
1102 if (m->m_flags & M_NOMAP) {
1105 * Remove 'm' from the new chain so
1106 * that the 'top' chain terminates
1107 * before 'm' in case 'top' is freed
1110 prev->m_next = NULL;
1112 m = _mb_unmapped_to_ext(m);
1125 * Replaced one mbuf with a chain, so we must
1126 * find the end of chain.
1140 * Allocate an empty EXT_PGS mbuf. The ext_free routine is
1141 * responsible for freeing any pages backing this mbuf when it is
1145 mb_alloc_ext_pgs(int how, bool pkthdr, m_ext_free_t ext_free)
1148 struct mbuf_ext_pgs *ext_pgs;
1151 m = m_gethdr(how, MT_DATA);
1153 m = m_get(how, MT_DATA);
1157 ext_pgs = uma_zalloc(zone_extpgs, how);
1158 if (ext_pgs == NULL) {
1164 ext_pgs->first_pg_off = 0;
1165 ext_pgs->last_pg_len = 0;
1166 ext_pgs->hdr_len = 0;
1167 ext_pgs->trail_len = 0;
1168 ext_pgs->tls = NULL;
1171 m->m_flags |= (M_EXT | M_RDONLY | M_NOMAP);
1172 m->m_ext.ext_type = EXT_PGS;
1173 m->m_ext.ext_flags = EXT_FLAG_EMBREF;
1174 m->m_ext.ext_count = 1;
1175 m->m_ext.ext_pgs = ext_pgs;
1176 m->m_ext.ext_size = 0;
1177 m->m_ext.ext_free = ext_free;
1181 #ifdef INVARIANT_SUPPORT
1183 mb_ext_pgs_check(struct mbuf_ext_pgs *ext_pgs)
1187 * NB: This expects a non-empty buffer (npgs > 0 and
1190 KASSERT(ext_pgs->npgs > 0,
1191 ("ext_pgs with no valid pages: %p", ext_pgs));
1192 KASSERT(ext_pgs->npgs <= nitems(ext_pgs->pa),
1193 ("ext_pgs with too many pages: %p", ext_pgs));
1194 KASSERT(ext_pgs->nrdy <= ext_pgs->npgs,
1195 ("ext_pgs with too many ready pages: %p", ext_pgs));
1196 KASSERT(ext_pgs->first_pg_off < PAGE_SIZE,
1197 ("ext_pgs with too large page offset: %p", ext_pgs));
1198 KASSERT(ext_pgs->last_pg_len > 0,
1199 ("ext_pgs with zero last page length: %p", ext_pgs));
1200 KASSERT(ext_pgs->last_pg_len <= PAGE_SIZE,
1201 ("ext_pgs with too large last page length: %p", ext_pgs));
1202 if (ext_pgs->npgs == 1) {
1203 KASSERT(ext_pgs->first_pg_off + ext_pgs->last_pg_len <=
1204 PAGE_SIZE, ("ext_pgs with single page too large: %p",
1207 KASSERT(ext_pgs->hdr_len <= sizeof(ext_pgs->hdr),
1208 ("ext_pgs with too large header length: %p", ext_pgs));
1209 KASSERT(ext_pgs->trail_len <= sizeof(ext_pgs->trail),
1210 ("ext_pgs with too large header length: %p", ext_pgs));
1215 * Clean up after mbufs with M_EXT storage attached to them if the
1216 * reference count hits 1.
1219 mb_free_ext(struct mbuf *m)
1221 volatile u_int *refcnt;
1225 KASSERT(m->m_flags & M_EXT, ("%s: M_EXT not set on %p", __func__, m));
1227 /* See if this is the mbuf that holds the embedded refcount. */
1228 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
1229 refcnt = &m->m_ext.ext_count;
1232 KASSERT(m->m_ext.ext_cnt != NULL,
1233 ("%s: no refcounting pointer on %p", __func__, m));
1234 refcnt = m->m_ext.ext_cnt;
1235 mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
1239 * Check if the header is embedded in the cluster. It is
1240 * important that we can't touch any of the mbuf fields
1241 * after we have freed the external storage, since mbuf
1242 * could have been embedded in it. For now, the mbufs
1243 * embedded into the cluster are always of type EXT_EXTREF,
1244 * and for this type we won't free the mref.
1246 if (m->m_flags & M_NOFREE) {
1248 KASSERT(m->m_ext.ext_type == EXT_EXTREF ||
1249 m->m_ext.ext_type == EXT_RXRING,
1250 ("%s: no-free mbuf %p has wrong type", __func__, m));
1254 /* Free attached storage if this mbuf is the only reference to it. */
1255 if (*refcnt == 1 || atomic_fetchadd_int(refcnt, -1) == 1) {
1256 switch (m->m_ext.ext_type) {
1258 /* The packet zone is special. */
1261 uma_zfree(zone_pack, mref);
1264 uma_zfree(zone_clust, m->m_ext.ext_buf);
1265 uma_zfree(zone_mbuf, mref);
1268 uma_zfree(zone_jumbop, m->m_ext.ext_buf);
1269 uma_zfree(zone_mbuf, mref);
1272 uma_zfree(zone_jumbo9, m->m_ext.ext_buf);
1273 uma_zfree(zone_mbuf, mref);
1276 uma_zfree(zone_jumbo16, m->m_ext.ext_buf);
1277 uma_zfree(zone_mbuf, mref);
1280 uma_zfree(zone_extpgs, mref->m_ext.ext_pgs);
1281 uma_zfree(zone_mbuf, mref);
1286 case EXT_DISPOSABLE:
1287 KASSERT(mref->m_ext.ext_free != NULL,
1288 ("%s: ext_free not set", __func__));
1289 mref->m_ext.ext_free(mref);
1290 uma_zfree(zone_mbuf, mref);
1293 KASSERT(m->m_ext.ext_free != NULL,
1294 ("%s: ext_free not set", __func__));
1295 m->m_ext.ext_free(m);
1298 KASSERT(m->m_ext.ext_free == NULL,
1299 ("%s: ext_free is set", __func__));
1302 KASSERT(m->m_ext.ext_type == 0,
1303 ("%s: unknown ext_type", __func__));
1307 if (freembuf && m != mref)
1308 uma_zfree(zone_mbuf, m);
1312 * Official mbuf(9) allocation KPI for stack and drivers:
1314 * m_get() - a single mbuf without any attachments, sys/mbuf.h.
1315 * m_gethdr() - a single mbuf initialized as M_PKTHDR, sys/mbuf.h.
1316 * m_getcl() - an mbuf + 2k cluster, sys/mbuf.h.
1317 * m_clget() - attach cluster to already allocated mbuf.
1318 * m_cljget() - attach jumbo cluster to already allocated mbuf.
1319 * m_get2() - allocate minimum mbuf that would fit size argument.
1320 * m_getm2() - allocate a chain of mbufs/clusters.
1321 * m_extadd() - attach external cluster to mbuf.
1323 * m_free() - free single mbuf with its tags and ext, sys/mbuf.h.
1324 * m_freem() - free chain of mbufs.
1328 m_clget(struct mbuf *m, int how)
1331 KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
1333 m->m_ext.ext_buf = (char *)NULL;
1334 uma_zalloc_arg(zone_clust, m, how);
1336 * On a cluster allocation failure, drain the packet zone and retry,
1337 * we might be able to loosen a few clusters up on the drain.
1339 if ((how & M_NOWAIT) && (m->m_ext.ext_buf == NULL)) {
1340 zone_drain(zone_pack);
1341 uma_zalloc_arg(zone_clust, m, how);
1343 MBUF_PROBE2(m__clget, m, how);
1344 return (m->m_flags & M_EXT);
1348 * m_cljget() is different from m_clget() as it can allocate clusters without
1349 * attaching them to an mbuf. In that case the return value is the pointer
1350 * to the cluster of the requested size. If an mbuf was specified, it gets
1351 * the cluster attached to it and the return value can be safely ignored.
1352 * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
1355 m_cljget(struct mbuf *m, int how, int size)
1361 KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
1363 m->m_ext.ext_buf = NULL;
1366 zone = m_getzone(size);
1367 retval = uma_zalloc_arg(zone, m, how);
1369 MBUF_PROBE4(m__cljget, m, how, size, retval);
1375 * m_get2() allocates minimum mbuf that would fit "size" argument.
1378 m_get2(int size, int how, short type, int flags)
1380 struct mb_args args;
1386 if (size <= MHLEN || (size <= MLEN && (flags & M_PKTHDR) == 0))
1387 return (uma_zalloc_arg(zone_mbuf, &args, how));
1388 if (size <= MCLBYTES)
1389 return (uma_zalloc_arg(zone_pack, &args, how));
1391 if (size > MJUMPAGESIZE)
1394 m = uma_zalloc_arg(zone_mbuf, &args, how);
1398 n = uma_zalloc_arg(zone_jumbop, m, how);
1400 uma_zfree(zone_mbuf, m);
1408 * m_getjcl() returns an mbuf with a cluster of the specified size attached.
1409 * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
1412 m_getjcl(int how, short type, int flags, int size)
1414 struct mb_args args;
1418 if (size == MCLBYTES)
1419 return m_getcl(how, type, flags);
1424 m = uma_zalloc_arg(zone_mbuf, &args, how);
1428 zone = m_getzone(size);
1429 n = uma_zalloc_arg(zone, m, how);
1431 uma_zfree(zone_mbuf, m);
1438 * Allocate a given length worth of mbufs and/or clusters (whatever fits
1439 * best) and return a pointer to the top of the allocated chain. If an
1440 * existing mbuf chain is provided, then we will append the new chain
1441 * to the existing one and return a pointer to the provided mbuf.
1444 m_getm2(struct mbuf *m, int len, int how, short type, int flags)
1446 struct mbuf *mb, *nm = NULL, *mtail = NULL;
1448 KASSERT(len >= 0, ("%s: len is < 0", __func__));
1450 /* Validate flags. */
1451 flags &= (M_PKTHDR | M_EOR);
1453 /* Packet header mbuf must be first in chain. */
1454 if ((flags & M_PKTHDR) && m != NULL)
1457 /* Loop and append maximum sized mbufs to the chain tail. */
1460 mb = m_getjcl(how, type, (flags & M_PKTHDR),
1462 else if (len >= MINCLSIZE)
1463 mb = m_getcl(how, type, (flags & M_PKTHDR));
1464 else if (flags & M_PKTHDR)
1465 mb = m_gethdr(how, type);
1467 mb = m_get(how, type);
1469 /* Fail the whole operation if one mbuf can't be allocated. */
1483 flags &= ~M_PKTHDR; /* Only valid on the first mbuf. */
1486 mtail->m_flags |= M_EOR; /* Only valid on the last mbuf. */
1488 /* If mbuf was supplied, append new chain to the end of it. */
1490 for (mtail = m; mtail->m_next != NULL; mtail = mtail->m_next)
1493 mtail->m_flags &= ~M_EOR;
1501 * Configure a provided mbuf to refer to the provided external storage
1502 * buffer and setup a reference count for said buffer.
1505 * mb The existing mbuf to which to attach the provided buffer.
1506 * buf The address of the provided external storage buffer.
1507 * size The size of the provided buffer.
1508 * freef A pointer to a routine that is responsible for freeing the
1509 * provided external storage buffer.
1510 * args A pointer to an argument structure (of any type) to be passed
1511 * to the provided freef routine (may be NULL).
1512 * flags Any other flags to be passed to the provided mbuf.
1513 * type The type that the external storage buffer should be
1520 m_extadd(struct mbuf *mb, char *buf, u_int size, m_ext_free_t freef,
1521 void *arg1, void *arg2, int flags, int type)
1524 KASSERT(type != EXT_CLUSTER, ("%s: EXT_CLUSTER not allowed", __func__));
1526 mb->m_flags |= (M_EXT | flags);
1527 mb->m_ext.ext_buf = buf;
1528 mb->m_data = mb->m_ext.ext_buf;
1529 mb->m_ext.ext_size = size;
1530 mb->m_ext.ext_free = freef;
1531 mb->m_ext.ext_arg1 = arg1;
1532 mb->m_ext.ext_arg2 = arg2;
1533 mb->m_ext.ext_type = type;
1535 if (type != EXT_EXTREF) {
1536 mb->m_ext.ext_count = 1;
1537 mb->m_ext.ext_flags = EXT_FLAG_EMBREF;
1539 mb->m_ext.ext_flags = 0;
1543 * Free an entire chain of mbufs and associated external buffers, if
1547 m_freem(struct mbuf *mb)
1550 MBUF_PROBE1(m__freem, mb);
1556 m_snd_tag_init(struct m_snd_tag *mst, struct ifnet *ifp)
1561 refcount_init(&mst->refcount, 1);
1562 counter_u64_add(snd_tag_count, 1);
1566 m_snd_tag_destroy(struct m_snd_tag *mst)
1571 ifp->if_snd_tag_free(mst);
1573 counter_u64_add(snd_tag_count, -1);