2 * Copyright (c) 2004, 2005,
3 * Bosko Milekic <bmilekic@FreeBSD.org>. All rights reserved.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice unmodified, this list of conditions and the following
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
31 #include "opt_param.h"
33 #include <sys/param.h>
34 #include <sys/malloc.h>
35 #include <sys/systm.h>
37 #include <sys/domain.h>
38 #include <sys/eventhandler.h>
39 #include <sys/kernel.h>
40 #include <sys/protosw.h>
42 #include <sys/sysctl.h>
44 #include <security/mac/mac_framework.h>
47 #include <vm/vm_extern.h>
48 #include <vm/vm_kern.h>
49 #include <vm/vm_page.h>
50 #include <vm/vm_map.h>
52 #include <vm/uma_int.h>
53 #include <vm/uma_dbg.h>
56 * In FreeBSD, Mbufs and Mbuf Clusters are allocated from UMA
59 * Mbuf Clusters (2K, contiguous) are allocated from the Cluster
60 * Zone. The Zone can be capped at kern.ipc.nmbclusters, if the
61 * administrator so desires.
63 * Mbufs are allocated from a UMA Master Zone called the Mbuf
66 * Additionally, FreeBSD provides a Packet Zone, which it
67 * configures as a Secondary Zone to the Mbuf Master Zone,
68 * thus sharing backend Slab kegs with the Mbuf Master Zone.
70 * Thus common-case allocations and locking are simplified:
74 * | .------------>[(Packet Cache)] m_get(), m_gethdr()
76 * [(Cluster Cache)] [ Secondary ] [ (Mbuf Cache) ]
77 * [ Cluster Zone ] [ Zone ] [ Mbuf Master Zone ]
83 * \____________(VM)_________________/
86 * Whenever an object is allocated with uma_zalloc() out of
87 * one of the Zones its _ctor_ function is executed. The same
88 * for any deallocation through uma_zfree() the _dtor_ function
91 * Caches are per-CPU and are filled from the Master Zone.
93 * Whenever an object is allocated from the underlying global
94 * memory pool it gets pre-initialized with the _zinit_ functions.
95 * When the Keg's are overfull objects get decomissioned with
96 * _zfini_ functions and free'd back to the global memory pool.
100 int nmbufs; /* limits number of mbufs */
101 int nmbclusters; /* limits number of mbuf clusters */
102 int nmbjumbop; /* limits number of page size jumbo clusters */
103 int nmbjumbo9; /* limits number of 9k jumbo clusters */
104 int nmbjumbo16; /* limits number of 16k jumbo clusters */
106 static quad_t maxmbufmem; /* overall real memory limit for all mbufs */
108 SYSCTL_QUAD(_kern_ipc, OID_AUTO, maxmbufmem, CTLFLAG_RDTUN, &maxmbufmem, 0,
109 "Maximum real memory allocateable to various mbuf types");
112 * tunable_mbinit() has to be run before any mbuf allocations are done.
115 tunable_mbinit(void *dummy)
120 * The default limit for all mbuf related memory is 1/2 of all
121 * available kernel memory (physical or kmem).
122 * At most it can be 3/4 of available kernel memory.
124 realmem = qmin((quad_t)physmem * PAGE_SIZE,
125 vm_map_max(kmem_map) - vm_map_min(kmem_map));
126 maxmbufmem = realmem / 2;
127 TUNABLE_QUAD_FETCH("kern.ipc.maxmbufmem", &maxmbufmem);
128 if (maxmbufmem > realmem / 4 * 3)
129 maxmbufmem = realmem / 4 * 3;
131 TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
132 if (nmbclusters == 0)
133 nmbclusters = maxmbufmem / MCLBYTES / 4;
135 TUNABLE_INT_FETCH("kern.ipc.nmbjumbop", &nmbjumbop);
137 nmbjumbop = maxmbufmem / MJUMPAGESIZE / 4;
139 TUNABLE_INT_FETCH("kern.ipc.nmbjumbo9", &nmbjumbo9);
141 nmbjumbo9 = maxmbufmem / MJUM9BYTES / 6;
143 TUNABLE_INT_FETCH("kern.ipc.nmbjumbo16", &nmbjumbo16);
145 nmbjumbo16 = maxmbufmem / MJUM16BYTES / 6;
148 * We need at least as many mbufs as we have clusters of
149 * the various types added together.
151 TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
152 if (nmbufs < nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16)
153 nmbufs = lmax(maxmbufmem / MSIZE / 5,
154 nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16);
156 SYSINIT(tunable_mbinit, SI_SUB_KMEM, SI_ORDER_MIDDLE, tunable_mbinit, NULL);
159 sysctl_nmbclusters(SYSCTL_HANDLER_ARGS)
161 int error, newnmbclusters;
163 newnmbclusters = nmbclusters;
164 error = sysctl_handle_int(oidp, &newnmbclusters, 0, req);
165 if (error == 0 && req->newptr) {
166 if (newnmbclusters > nmbclusters &&
167 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
168 nmbclusters = newnmbclusters;
169 nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
170 EVENTHANDLER_INVOKE(nmbclusters_change);
176 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbclusters, CTLTYPE_INT|CTLFLAG_RW,
177 &nmbclusters, 0, sysctl_nmbclusters, "IU",
178 "Maximum number of mbuf clusters allowed");
181 sysctl_nmbjumbop(SYSCTL_HANDLER_ARGS)
183 int error, newnmbjumbop;
185 newnmbjumbop = nmbjumbop;
186 error = sysctl_handle_int(oidp, &newnmbjumbop, 0, req);
187 if (error == 0 && req->newptr) {
188 if (newnmbjumbop > nmbjumbop &&
189 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
190 nmbjumbop = newnmbjumbop;
191 nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
197 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbop, CTLTYPE_INT|CTLFLAG_RW,
198 &nmbjumbop, 0, sysctl_nmbjumbop, "IU",
199 "Maximum number of mbuf page size jumbo clusters allowed");
202 sysctl_nmbjumbo9(SYSCTL_HANDLER_ARGS)
204 int error, newnmbjumbo9;
206 newnmbjumbo9 = nmbjumbo9;
207 error = sysctl_handle_int(oidp, &newnmbjumbo9, 0, req);
208 if (error == 0 && req->newptr) {
209 if (newnmbjumbo9 > nmbjumbo9 &&
210 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
211 nmbjumbo9 = newnmbjumbo9;
212 nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
218 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo9, CTLTYPE_INT|CTLFLAG_RW,
219 &nmbjumbo9, 0, sysctl_nmbjumbo9, "IU",
220 "Maximum number of mbuf 9k jumbo clusters allowed");
223 sysctl_nmbjumbo16(SYSCTL_HANDLER_ARGS)
225 int error, newnmbjumbo16;
227 newnmbjumbo16 = nmbjumbo16;
228 error = sysctl_handle_int(oidp, &newnmbjumbo16, 0, req);
229 if (error == 0 && req->newptr) {
230 if (newnmbjumbo16 > nmbjumbo16 &&
231 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
232 nmbjumbo16 = newnmbjumbo16;
233 nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
239 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo16, CTLTYPE_INT|CTLFLAG_RW,
240 &nmbjumbo16, 0, sysctl_nmbjumbo16, "IU",
241 "Maximum number of mbuf 16k jumbo clusters allowed");
244 sysctl_nmbufs(SYSCTL_HANDLER_ARGS)
246 int error, newnmbufs;
249 error = sysctl_handle_int(oidp, &newnmbufs, 0, req);
250 if (error == 0 && req->newptr) {
251 if (newnmbufs > nmbufs) {
253 nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
254 EVENTHANDLER_INVOKE(nmbufs_change);
260 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbufs, CTLTYPE_INT|CTLFLAG_RW,
261 &nmbufs, 0, sysctl_nmbufs, "IU",
262 "Maximum number of mbufs allowed");
265 * Zones from which we allocate.
267 uma_zone_t zone_mbuf;
268 uma_zone_t zone_clust;
269 uma_zone_t zone_pack;
270 uma_zone_t zone_jumbop;
271 uma_zone_t zone_jumbo9;
272 uma_zone_t zone_jumbo16;
273 uma_zone_t zone_ext_refcnt;
278 static int mb_ctor_mbuf(void *, int, void *, int);
279 static int mb_ctor_clust(void *, int, void *, int);
280 static int mb_ctor_pack(void *, int, void *, int);
281 static void mb_dtor_mbuf(void *, int, void *);
282 static void mb_dtor_clust(void *, int, void *);
283 static void mb_dtor_pack(void *, int, void *);
284 static int mb_zinit_pack(void *, int, int);
285 static void mb_zfini_pack(void *, int);
287 static void mb_reclaim(void *);
288 static void *mbuf_jumbo_alloc(uma_zone_t, int, uint8_t *, int);
290 /* Ensure that MSIZE is a power of 2. */
291 CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE);
294 * Initialize FreeBSD Network buffer allocation.
297 mbuf_init(void *dummy)
301 * Configure UMA zones for Mbufs, Clusters, and Packets.
303 zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE,
304 mb_ctor_mbuf, mb_dtor_mbuf,
306 trash_init, trash_fini,
310 MSIZE - 1, UMA_ZONE_MAXBUCKET);
312 nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
313 uma_zone_set_warning(zone_mbuf, "kern.ipc.nmbufs limit reached");
315 zone_clust = uma_zcreate(MBUF_CLUSTER_MEM_NAME, MCLBYTES,
316 mb_ctor_clust, mb_dtor_clust,
318 trash_init, trash_fini,
322 UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
324 nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
325 uma_zone_set_warning(zone_clust, "kern.ipc.nmbclusters limit reached");
327 zone_pack = uma_zsecond_create(MBUF_PACKET_MEM_NAME, mb_ctor_pack,
328 mb_dtor_pack, mb_zinit_pack, mb_zfini_pack, zone_mbuf);
330 /* Make jumbo frame zone too. Page size, 9k and 16k. */
331 zone_jumbop = uma_zcreate(MBUF_JUMBOP_MEM_NAME, MJUMPAGESIZE,
332 mb_ctor_clust, mb_dtor_clust,
334 trash_init, trash_fini,
338 UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
340 nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
341 uma_zone_set_warning(zone_jumbop, "kern.ipc.nmbjumbop limit reached");
343 zone_jumbo9 = uma_zcreate(MBUF_JUMBO9_MEM_NAME, MJUM9BYTES,
344 mb_ctor_clust, mb_dtor_clust,
346 trash_init, trash_fini,
350 UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
351 uma_zone_set_allocf(zone_jumbo9, mbuf_jumbo_alloc);
353 nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
354 uma_zone_set_warning(zone_jumbo9, "kern.ipc.nmbjumbo9 limit reached");
356 zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES,
357 mb_ctor_clust, mb_dtor_clust,
359 trash_init, trash_fini,
363 UMA_ALIGN_PTR, UMA_ZONE_REFCNT);
364 uma_zone_set_allocf(zone_jumbo16, mbuf_jumbo_alloc);
366 nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
367 uma_zone_set_warning(zone_jumbo16, "kern.ipc.nmbjumbo16 limit reached");
369 zone_ext_refcnt = uma_zcreate(MBUF_EXTREFCNT_MEM_NAME, sizeof(u_int),
372 UMA_ALIGN_PTR, UMA_ZONE_ZINIT);
374 /* uma_prealloc() goes here... */
377 * Hook event handler for low-memory situation, used to
378 * drain protocols and push data back to the caches (UMA
379 * later pushes it back to VM).
381 EVENTHANDLER_REGISTER(vm_lowmem, mb_reclaim, NULL,
382 EVENTHANDLER_PRI_FIRST);
384 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL);
387 * UMA backend page allocator for the jumbo frame zones.
389 * Allocates kernel virtual memory that is backed by contiguous physical
393 mbuf_jumbo_alloc(uma_zone_t zone, int bytes, uint8_t *flags, int wait)
396 /* Inform UMA that this allocator uses kernel_map/object. */
397 *flags = UMA_SLAB_KERNEL;
398 return ((void *)kmem_alloc_contig(kernel_map, bytes, wait,
399 (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
403 * Constructor for Mbuf master zone.
405 * The 'arg' pointer points to a mb_args structure which
406 * contains call-specific information required to support the
407 * mbuf allocation API. See mbuf.h.
410 mb_ctor_mbuf(void *mem, int size, void *arg, int how)
413 struct mb_args *args;
421 trash_ctor(mem, size, arg, how);
423 m = (struct mbuf *)mem;
424 args = (struct mb_args *)arg;
429 * The mbuf is initialized later. The caller has the
430 * responsibility to set up any MAC labels too.
432 if (type == MT_NOINIT)
440 if (flags & M_PKTHDR) {
441 m->m_data = m->m_pktdat;
442 m->m_pkthdr.rcvif = NULL;
443 m->m_pkthdr.header = NULL;
445 m->m_pkthdr.csum_flags = 0;
446 m->m_pkthdr.csum_data = 0;
447 m->m_pkthdr.tso_segsz = 0;
448 m->m_pkthdr.ether_vtag = 0;
449 m->m_pkthdr.flowid = 0;
450 m->m_pkthdr.fibnum = 0;
451 SLIST_INIT(&m->m_pkthdr.tags);
453 /* If the label init fails, fail the alloc */
454 error = mac_mbuf_init(m, how);
459 m->m_data = m->m_dat;
464 * The Mbuf master zone destructor.
467 mb_dtor_mbuf(void *mem, int size, void *arg)
472 m = (struct mbuf *)mem;
473 flags = (unsigned long)arg;
475 if ((flags & MB_NOTAGS) == 0 && (m->m_flags & M_PKTHDR) != 0)
476 m_tag_delete_chain(m, NULL);
477 KASSERT((m->m_flags & M_EXT) == 0, ("%s: M_EXT set", __func__));
478 KASSERT((m->m_flags & M_NOFREE) == 0, ("%s: M_NOFREE set", __func__));
480 trash_dtor(mem, size, arg);
485 * The Mbuf Packet zone destructor.
488 mb_dtor_pack(void *mem, int size, void *arg)
492 m = (struct mbuf *)mem;
493 if ((m->m_flags & M_PKTHDR) != 0)
494 m_tag_delete_chain(m, NULL);
496 /* Make sure we've got a clean cluster back. */
497 KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__));
498 KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__));
499 KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__));
500 KASSERT(m->m_ext.ext_arg1 == NULL, ("%s: ext_arg1 != NULL", __func__));
501 KASSERT(m->m_ext.ext_arg2 == NULL, ("%s: ext_arg2 != NULL", __func__));
502 KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__));
503 KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__));
504 KASSERT(*m->m_ext.ref_cnt == 1, ("%s: ref_cnt != 1", __func__));
506 trash_dtor(m->m_ext.ext_buf, MCLBYTES, arg);
509 * If there are processes blocked on zone_clust, waiting for pages
510 * to be freed up, * cause them to be woken up by draining the
511 * packet zone. We are exposed to a race here * (in the check for
512 * the UMA_ZFLAG_FULL) where we might miss the flag set, but that
513 * is deliberate. We don't want to acquire the zone lock for every
516 if (uma_zone_exhausted_nolock(zone_clust))
517 zone_drain(zone_pack);
521 * The Cluster and Jumbo[PAGESIZE|9|16] zone constructor.
523 * Here the 'arg' pointer points to the Mbuf which we
524 * are configuring cluster storage for. If 'arg' is
525 * empty we allocate just the cluster without setting
526 * the mbuf to it. See mbuf.h.
529 mb_ctor_clust(void *mem, int size, void *arg, int how)
537 trash_ctor(mem, size, arg, how);
544 #if MJUMPAGESIZE != MCLBYTES
559 panic("unknown cluster size");
563 m = (struct mbuf *)arg;
564 refcnt = uma_find_refcnt(zone, mem);
567 m->m_ext.ext_buf = (caddr_t)mem;
568 m->m_data = m->m_ext.ext_buf;
570 m->m_ext.ext_free = NULL;
571 m->m_ext.ext_arg1 = NULL;
572 m->m_ext.ext_arg2 = NULL;
573 m->m_ext.ext_size = size;
574 m->m_ext.ext_type = type;
575 m->m_ext.ref_cnt = refcnt;
582 * The Mbuf Cluster zone destructor.
585 mb_dtor_clust(void *mem, int size, void *arg)
590 zone = m_getzone(size);
591 KASSERT(*(uma_find_refcnt(zone, mem)) <= 1,
592 ("%s: refcnt incorrect %u", __func__,
593 *(uma_find_refcnt(zone, mem))) );
595 trash_dtor(mem, size, arg);
600 * The Packet secondary zone's init routine, executed on the
601 * object's transition from mbuf keg slab to zone cache.
604 mb_zinit_pack(void *mem, int size, int how)
608 m = (struct mbuf *)mem; /* m is virgin. */
609 if (uma_zalloc_arg(zone_clust, m, how) == NULL ||
610 m->m_ext.ext_buf == NULL)
612 m->m_ext.ext_type = EXT_PACKET; /* Override. */
614 trash_init(m->m_ext.ext_buf, MCLBYTES, how);
620 * The Packet secondary zone's fini routine, executed on the
621 * object's transition from zone cache to keg slab.
624 mb_zfini_pack(void *mem, int size)
628 m = (struct mbuf *)mem;
630 trash_fini(m->m_ext.ext_buf, MCLBYTES);
632 uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL);
634 trash_dtor(mem, size, NULL);
639 * The "packet" keg constructor.
642 mb_ctor_pack(void *mem, int size, void *arg, int how)
645 struct mb_args *args;
652 m = (struct mbuf *)mem;
653 args = (struct mb_args *)arg;
658 trash_ctor(m->m_ext.ext_buf, MCLBYTES, arg, how);
662 m->m_data = m->m_ext.ext_buf;
664 m->m_flags = (flags | M_EXT);
667 if (flags & M_PKTHDR) {
668 m->m_pkthdr.rcvif = NULL;
670 m->m_pkthdr.header = NULL;
671 m->m_pkthdr.csum_flags = 0;
672 m->m_pkthdr.csum_data = 0;
673 m->m_pkthdr.tso_segsz = 0;
674 m->m_pkthdr.ether_vtag = 0;
675 m->m_pkthdr.flowid = 0;
676 m->m_pkthdr.fibnum = 0;
677 SLIST_INIT(&m->m_pkthdr.tags);
679 /* If the label init fails, fail the alloc */
680 error = mac_mbuf_init(m, how);
685 /* m_ext is already initialized. */
691 m_pkthdr_init(struct mbuf *m, int how)
696 m->m_data = m->m_pktdat;
697 SLIST_INIT(&m->m_pkthdr.tags);
698 m->m_pkthdr.rcvif = NULL;
699 m->m_pkthdr.header = NULL;
701 m->m_pkthdr.flowid = 0;
702 m->m_pkthdr.fibnum = 0;
703 m->m_pkthdr.csum_flags = 0;
704 m->m_pkthdr.csum_data = 0;
705 m->m_pkthdr.tso_segsz = 0;
706 m->m_pkthdr.ether_vtag = 0;
708 /* If the label init fails, fail the alloc */
709 error = mac_mbuf_init(m, how);
718 * This is the protocol drain routine.
720 * No locks should be held when this is called. The drain routines have to
721 * presently acquire some locks which raises the possibility of lock order
725 mb_reclaim(void *junk)
730 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK | WARN_PANIC, NULL,
733 for (dp = domains; dp != NULL; dp = dp->dom_next)
734 for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++)
735 if (pr->pr_drain != NULL)