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"
34 #include "opt_kern_tls.h"
36 #include <sys/param.h>
38 #include <sys/domainset.h>
39 #include <sys/malloc.h>
40 #include <sys/systm.h>
42 #include <sys/domain.h>
43 #include <sys/eventhandler.h>
44 #include <sys/kernel.h>
46 #include <sys/limits.h>
48 #include <sys/mutex.h>
49 #include <sys/protosw.h>
50 #include <sys/refcount.h>
51 #include <sys/sf_buf.h>
53 #include <sys/socket.h>
54 #include <sys/sysctl.h>
57 #include <net/if_var.h>
60 #include <vm/vm_extern.h>
61 #include <vm/vm_kern.h>
62 #include <vm/vm_page.h>
63 #include <vm/vm_map.h>
65 #include <vm/uma_dbg.h>
68 * In FreeBSD, Mbufs and Mbuf Clusters are allocated from UMA
71 * Mbuf Clusters (2K, contiguous) are allocated from the Cluster
72 * Zone. The Zone can be capped at kern.ipc.nmbclusters, if the
73 * administrator so desires.
75 * Mbufs are allocated from a UMA Master Zone called the Mbuf
78 * Additionally, FreeBSD provides a Packet Zone, which it
79 * configures as a Secondary Zone to the Mbuf Master Zone,
80 * thus sharing backend Slab kegs with the Mbuf Master Zone.
82 * Thus common-case allocations and locking are simplified:
86 * | .------------>[(Packet Cache)] m_get(), m_gethdr()
88 * [(Cluster Cache)] [ Secondary ] [ (Mbuf Cache) ]
89 * [ Cluster Zone ] [ Zone ] [ Mbuf Master Zone ]
95 * \____________(VM)_________________/
98 * Whenever an object is allocated with uma_zalloc() out of
99 * one of the Zones its _ctor_ function is executed. The same
100 * for any deallocation through uma_zfree() the _dtor_ function
103 * Caches are per-CPU and are filled from the Master Zone.
105 * Whenever an object is allocated from the underlying global
106 * memory pool it gets pre-initialized with the _zinit_ functions.
107 * When the Keg's are overfull objects get decommissioned with
108 * _zfini_ functions and free'd back to the global memory pool.
112 int nmbufs; /* limits number of mbufs */
113 int nmbclusters; /* limits number of mbuf clusters */
114 int nmbjumbop; /* limits number of page size jumbo clusters */
115 int nmbjumbo9; /* limits number of 9k jumbo clusters */
116 int nmbjumbo16; /* limits number of 16k jumbo clusters */
118 bool mb_use_ext_pgs; /* use EXT_PGS mbufs for sendfile & TLS */
119 SYSCTL_BOOL(_kern_ipc, OID_AUTO, mb_use_ext_pgs, CTLFLAG_RWTUN,
121 "Use unmapped mbufs for sendfile(2) and TLS offload");
123 static quad_t maxmbufmem; /* overall real memory limit for all mbufs */
125 SYSCTL_QUAD(_kern_ipc, OID_AUTO, maxmbufmem, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &maxmbufmem, 0,
126 "Maximum real memory allocatable to various mbuf types");
128 static counter_u64_t snd_tag_count;
129 SYSCTL_COUNTER_U64(_kern_ipc, OID_AUTO, num_snd_tags, CTLFLAG_RW,
130 &snd_tag_count, "# of active mbuf send tags");
133 * tunable_mbinit() has to be run before any mbuf allocations are done.
136 tunable_mbinit(void *dummy)
141 * The default limit for all mbuf related memory is 1/2 of all
142 * available kernel memory (physical or kmem).
143 * At most it can be 3/4 of available kernel memory.
145 realmem = qmin((quad_t)physmem * PAGE_SIZE, vm_kmem_size);
146 maxmbufmem = realmem / 2;
147 TUNABLE_QUAD_FETCH("kern.ipc.maxmbufmem", &maxmbufmem);
148 if (maxmbufmem > realmem / 4 * 3)
149 maxmbufmem = realmem / 4 * 3;
151 TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
152 if (nmbclusters == 0)
153 nmbclusters = maxmbufmem / MCLBYTES / 4;
155 TUNABLE_INT_FETCH("kern.ipc.nmbjumbop", &nmbjumbop);
157 nmbjumbop = maxmbufmem / MJUMPAGESIZE / 4;
159 TUNABLE_INT_FETCH("kern.ipc.nmbjumbo9", &nmbjumbo9);
161 nmbjumbo9 = maxmbufmem / MJUM9BYTES / 6;
163 TUNABLE_INT_FETCH("kern.ipc.nmbjumbo16", &nmbjumbo16);
165 nmbjumbo16 = maxmbufmem / MJUM16BYTES / 6;
168 * We need at least as many mbufs as we have clusters of
169 * the various types added together.
171 TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
172 if (nmbufs < nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16)
173 nmbufs = lmax(maxmbufmem / MSIZE / 5,
174 nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16);
176 SYSINIT(tunable_mbinit, SI_SUB_KMEM, SI_ORDER_MIDDLE, tunable_mbinit, NULL);
179 sysctl_nmbclusters(SYSCTL_HANDLER_ARGS)
181 int error, newnmbclusters;
183 newnmbclusters = nmbclusters;
184 error = sysctl_handle_int(oidp, &newnmbclusters, 0, req);
185 if (error == 0 && req->newptr && newnmbclusters != nmbclusters) {
186 if (newnmbclusters > nmbclusters &&
187 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
188 nmbclusters = newnmbclusters;
189 nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
190 EVENTHANDLER_INVOKE(nmbclusters_change);
196 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbclusters,
197 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &nmbclusters, 0,
198 sysctl_nmbclusters, "IU",
199 "Maximum number of mbuf clusters allowed");
202 sysctl_nmbjumbop(SYSCTL_HANDLER_ARGS)
204 int error, newnmbjumbop;
206 newnmbjumbop = nmbjumbop;
207 error = sysctl_handle_int(oidp, &newnmbjumbop, 0, req);
208 if (error == 0 && req->newptr && newnmbjumbop != nmbjumbop) {
209 if (newnmbjumbop > nmbjumbop &&
210 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
211 nmbjumbop = newnmbjumbop;
212 nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
218 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbop,
219 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &nmbjumbop, 0,
220 sysctl_nmbjumbop, "IU",
221 "Maximum number of mbuf page size jumbo clusters allowed");
224 sysctl_nmbjumbo9(SYSCTL_HANDLER_ARGS)
226 int error, newnmbjumbo9;
228 newnmbjumbo9 = nmbjumbo9;
229 error = sysctl_handle_int(oidp, &newnmbjumbo9, 0, req);
230 if (error == 0 && req->newptr && newnmbjumbo9 != nmbjumbo9) {
231 if (newnmbjumbo9 > nmbjumbo9 &&
232 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
233 nmbjumbo9 = newnmbjumbo9;
234 nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
240 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo9,
241 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &nmbjumbo9, 0,
242 sysctl_nmbjumbo9, "IU",
243 "Maximum number of mbuf 9k jumbo clusters allowed");
246 sysctl_nmbjumbo16(SYSCTL_HANDLER_ARGS)
248 int error, newnmbjumbo16;
250 newnmbjumbo16 = nmbjumbo16;
251 error = sysctl_handle_int(oidp, &newnmbjumbo16, 0, req);
252 if (error == 0 && req->newptr && newnmbjumbo16 != nmbjumbo16) {
253 if (newnmbjumbo16 > nmbjumbo16 &&
254 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
255 nmbjumbo16 = newnmbjumbo16;
256 nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
262 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo16,
263 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &nmbjumbo16, 0,
264 sysctl_nmbjumbo16, "IU",
265 "Maximum number of mbuf 16k jumbo clusters allowed");
268 sysctl_nmbufs(SYSCTL_HANDLER_ARGS)
270 int error, newnmbufs;
273 error = sysctl_handle_int(oidp, &newnmbufs, 0, req);
274 if (error == 0 && req->newptr && newnmbufs != nmbufs) {
275 if (newnmbufs > nmbufs) {
277 nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
278 EVENTHANDLER_INVOKE(nmbufs_change);
284 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbufs,
285 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
286 &nmbufs, 0, sysctl_nmbufs, "IU",
287 "Maximum number of mbufs allowed");
290 * Zones from which we allocate.
292 uma_zone_t zone_mbuf;
293 uma_zone_t zone_clust;
294 uma_zone_t zone_pack;
295 uma_zone_t zone_jumbop;
296 uma_zone_t zone_jumbo9;
297 uma_zone_t zone_jumbo16;
302 static int mb_ctor_mbuf(void *, int, void *, int);
303 static int mb_ctor_clust(void *, int, void *, int);
304 static int mb_ctor_pack(void *, int, void *, int);
305 static void mb_dtor_mbuf(void *, int, void *);
306 static void mb_dtor_pack(void *, int, void *);
307 static int mb_zinit_pack(void *, int, int);
308 static void mb_zfini_pack(void *, int);
309 static void mb_reclaim(uma_zone_t, int);
311 /* Ensure that MSIZE is a power of 2. */
312 CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE);
314 _Static_assert(sizeof(struct mbuf) <= MSIZE,
315 "size of mbuf exceeds MSIZE");
317 * Initialize FreeBSD Network buffer allocation.
320 mbuf_init(void *dummy)
324 * Configure UMA zones for Mbufs, Clusters, and Packets.
326 zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE,
327 mb_ctor_mbuf, mb_dtor_mbuf, NULL, NULL,
328 MSIZE - 1, UMA_ZONE_CONTIG | 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,
335 mb_ctor_clust, NULL, NULL, NULL,
336 UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
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,
347 mb_ctor_clust, NULL, NULL, NULL,
348 UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
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,
355 mb_ctor_clust, NULL, NULL, NULL,
356 UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
358 nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
359 uma_zone_set_warning(zone_jumbo9, "kern.ipc.nmbjumbo9 limit reached");
360 uma_zone_set_maxaction(zone_jumbo9, mb_reclaim);
362 zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES,
363 mb_ctor_clust, NULL, NULL, NULL,
364 UMA_ALIGN_PTR, UMA_ZONE_CONTIG);
366 nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
367 uma_zone_set_warning(zone_jumbo16, "kern.ipc.nmbjumbo16 limit reached");
368 uma_zone_set_maxaction(zone_jumbo16, mb_reclaim);
371 * Hook event handler for low-memory situation, used to
372 * drain protocols and push data back to the caches (UMA
373 * later pushes it back to VM).
375 EVENTHANDLER_REGISTER(vm_lowmem, mb_reclaim, NULL,
376 EVENTHANDLER_PRI_FIRST);
378 snd_tag_count = counter_u64_alloc(M_WAITOK);
380 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL);
384 * debugnet makes use of a pre-allocated pool of mbufs and clusters. When
385 * debugnet is configured, we initialize a set of UMA cache zones which return
386 * items from this pool. At panic-time, the regular UMA zone pointers are
387 * overwritten with those of the cache zones so that drivers may allocate and
388 * free mbufs and clusters without attempting to allocate physical memory.
390 * We keep mbufs and clusters in a pair of mbuf queues. In particular, for
391 * the purpose of caching clusters, we treat them as mbufs.
393 static struct mbufq dn_mbufq =
394 { STAILQ_HEAD_INITIALIZER(dn_mbufq.mq_head), 0, INT_MAX };
395 static struct mbufq dn_clustq =
396 { STAILQ_HEAD_INITIALIZER(dn_clustq.mq_head), 0, INT_MAX };
398 static int dn_clsize;
399 static uma_zone_t dn_zone_mbuf;
400 static uma_zone_t dn_zone_clust;
401 static uma_zone_t dn_zone_pack;
403 static struct debugnet_saved_zones {
405 uma_zone_t dsz_clust;
407 uma_zone_t dsz_jumbop;
408 uma_zone_t dsz_jumbo9;
409 uma_zone_t dsz_jumbo16;
410 bool dsz_debugnet_zones_enabled;
414 dn_buf_import(void *arg, void **store, int count, int domain __unused,
423 for (i = 0; i < count; i++) {
424 m = mbufq_dequeue(q);
427 trash_init(m, q == &dn_mbufq ? MSIZE : dn_clsize, flags);
430 KASSERT((flags & M_WAITOK) == 0 || i == count,
431 ("%s: ran out of pre-allocated mbufs", __func__));
436 dn_buf_release(void *arg, void **store, int count)
444 for (i = 0; i < count; i++) {
446 (void)mbufq_enqueue(q, m);
451 dn_pack_import(void *arg __unused, void **store, int count, int domain __unused,
458 for (i = 0; i < count; i++) {
459 m = m_get(MT_DATA, M_NOWAIT);
462 clust = uma_zalloc(dn_zone_clust, M_NOWAIT);
467 mb_ctor_clust(clust, dn_clsize, m, 0);
470 KASSERT((flags & M_WAITOK) == 0 || i == count,
471 ("%s: ran out of pre-allocated mbufs", __func__));
476 dn_pack_release(void *arg __unused, void **store, int count)
482 for (i = 0; i < count; i++) {
484 clust = m->m_ext.ext_buf;
485 uma_zfree(dn_zone_clust, clust);
486 uma_zfree(dn_zone_mbuf, m);
491 * Free the pre-allocated mbufs and clusters reserved for debugnet, and destroy
492 * the corresponding UMA cache zones.
495 debugnet_mbuf_drain(void)
500 if (dn_zone_mbuf != NULL) {
501 uma_zdestroy(dn_zone_mbuf);
504 if (dn_zone_clust != NULL) {
505 uma_zdestroy(dn_zone_clust);
506 dn_zone_clust = NULL;
508 if (dn_zone_pack != NULL) {
509 uma_zdestroy(dn_zone_pack);
513 while ((m = mbufq_dequeue(&dn_mbufq)) != NULL)
515 while ((item = mbufq_dequeue(&dn_clustq)) != NULL)
516 uma_zfree(m_getzone(dn_clsize), item);
520 * Callback invoked immediately prior to starting a debugnet connection.
523 debugnet_mbuf_start(void)
526 MPASS(!dn_saved_zones.dsz_debugnet_zones_enabled);
528 /* Save the old zone pointers to restore when debugnet is closed. */
529 dn_saved_zones = (struct debugnet_saved_zones) {
530 .dsz_debugnet_zones_enabled = true,
531 .dsz_mbuf = zone_mbuf,
532 .dsz_clust = zone_clust,
533 .dsz_pack = zone_pack,
534 .dsz_jumbop = zone_jumbop,
535 .dsz_jumbo9 = zone_jumbo9,
536 .dsz_jumbo16 = zone_jumbo16,
540 * All cluster zones return buffers of the size requested by the
541 * drivers. It's up to the driver to reinitialize the zones if the
542 * MTU of a debugnet-enabled interface changes.
544 printf("debugnet: overwriting mbuf zone pointers\n");
545 zone_mbuf = dn_zone_mbuf;
546 zone_clust = dn_zone_clust;
547 zone_pack = dn_zone_pack;
548 zone_jumbop = dn_zone_clust;
549 zone_jumbo9 = dn_zone_clust;
550 zone_jumbo16 = dn_zone_clust;
554 * Callback invoked when a debugnet connection is closed/finished.
557 debugnet_mbuf_finish(void)
560 MPASS(dn_saved_zones.dsz_debugnet_zones_enabled);
562 printf("debugnet: restoring mbuf zone pointers\n");
563 zone_mbuf = dn_saved_zones.dsz_mbuf;
564 zone_clust = dn_saved_zones.dsz_clust;
565 zone_pack = dn_saved_zones.dsz_pack;
566 zone_jumbop = dn_saved_zones.dsz_jumbop;
567 zone_jumbo9 = dn_saved_zones.dsz_jumbo9;
568 zone_jumbo16 = dn_saved_zones.dsz_jumbo16;
570 memset(&dn_saved_zones, 0, sizeof(dn_saved_zones));
574 * Reinitialize the debugnet mbuf+cluster pool and cache zones.
577 debugnet_mbuf_reinit(int nmbuf, int nclust, int clsize)
582 debugnet_mbuf_drain();
586 dn_zone_mbuf = uma_zcache_create("debugnet_" MBUF_MEM_NAME,
587 MSIZE, mb_ctor_mbuf, mb_dtor_mbuf, NULL, NULL,
588 dn_buf_import, dn_buf_release,
589 &dn_mbufq, UMA_ZONE_NOBUCKET);
591 dn_zone_clust = uma_zcache_create("debugnet_" MBUF_CLUSTER_MEM_NAME,
592 clsize, mb_ctor_clust, NULL, NULL, NULL,
593 dn_buf_import, dn_buf_release,
594 &dn_clustq, UMA_ZONE_NOBUCKET);
596 dn_zone_pack = uma_zcache_create("debugnet_" MBUF_PACKET_MEM_NAME,
597 MCLBYTES, mb_ctor_pack, mb_dtor_pack, NULL, NULL,
598 dn_pack_import, dn_pack_release,
599 NULL, UMA_ZONE_NOBUCKET);
601 while (nmbuf-- > 0) {
602 m = m_get(MT_DATA, M_WAITOK);
603 uma_zfree(dn_zone_mbuf, m);
605 while (nclust-- > 0) {
606 item = uma_zalloc(m_getzone(dn_clsize), M_WAITOK);
607 uma_zfree(dn_zone_clust, item);
610 #endif /* DEBUGNET */
613 * Constructor for Mbuf master zone.
615 * The 'arg' pointer points to a mb_args structure which
616 * contains call-specific information required to support the
617 * mbuf allocation API. See mbuf.h.
620 mb_ctor_mbuf(void *mem, int size, void *arg, int how)
623 struct mb_args *args;
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);
665 * The Mbuf Packet zone destructor.
668 mb_dtor_pack(void *mem, int size, void *arg)
672 m = (struct mbuf *)mem;
673 if ((m->m_flags & M_PKTHDR) != 0)
674 m_tag_delete_chain(m, NULL);
676 /* Make sure we've got a clean cluster back. */
677 KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__));
678 KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__));
679 KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__));
680 KASSERT(m->m_ext.ext_arg1 == NULL, ("%s: ext_arg1 != NULL", __func__));
681 KASSERT(m->m_ext.ext_arg2 == NULL, ("%s: ext_arg2 != NULL", __func__));
682 KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__));
683 KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__));
685 trash_dtor(m->m_ext.ext_buf, MCLBYTES, arg);
688 * If there are processes blocked on zone_clust, waiting for pages
689 * to be freed up, cause them to be woken up by draining the
690 * packet zone. We are exposed to a race here (in the check for
691 * the UMA_ZFLAG_FULL) where we might miss the flag set, but that
692 * is deliberate. We don't want to acquire the zone lock for every
695 if (uma_zone_exhausted(zone_clust))
696 uma_zone_reclaim(zone_pack, UMA_RECLAIM_DRAIN);
700 * The Cluster and Jumbo[PAGESIZE|9|16] zone constructor.
702 * Here the 'arg' pointer points to the Mbuf which we
703 * are configuring cluster storage for. If 'arg' is
704 * empty we allocate just the cluster without setting
705 * the mbuf to it. See mbuf.h.
708 mb_ctor_clust(void *mem, int size, void *arg, int how)
712 m = (struct mbuf *)arg;
714 m->m_ext.ext_buf = (char *)mem;
715 m->m_data = m->m_ext.ext_buf;
717 m->m_ext.ext_free = NULL;
718 m->m_ext.ext_arg1 = NULL;
719 m->m_ext.ext_arg2 = NULL;
720 m->m_ext.ext_size = size;
721 m->m_ext.ext_type = m_gettype(size);
722 m->m_ext.ext_flags = EXT_FLAG_EMBREF;
723 m->m_ext.ext_count = 1;
730 * The Packet secondary zone's init routine, executed on the
731 * object's transition from mbuf keg slab to zone cache.
734 mb_zinit_pack(void *mem, int size, int how)
738 m = (struct mbuf *)mem; /* m is virgin. */
739 if (uma_zalloc_arg(zone_clust, m, how) == NULL ||
740 m->m_ext.ext_buf == NULL)
742 m->m_ext.ext_type = EXT_PACKET; /* Override. */
744 trash_init(m->m_ext.ext_buf, MCLBYTES, how);
750 * The Packet secondary zone's fini routine, executed on the
751 * object's transition from zone cache to keg slab.
754 mb_zfini_pack(void *mem, int size)
758 m = (struct mbuf *)mem;
760 trash_fini(m->m_ext.ext_buf, MCLBYTES);
762 uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL);
764 trash_dtor(mem, size, NULL);
769 * The "packet" keg constructor.
772 mb_ctor_pack(void *mem, int size, void *arg, int how)
775 struct mb_args *args;
779 m = (struct mbuf *)mem;
780 args = (struct mb_args *)arg;
783 MPASS((flags & M_NOFREE) == 0);
786 trash_ctor(m->m_ext.ext_buf, MCLBYTES, arg, how);
789 error = m_init(m, how, type, flags);
791 /* m_ext is already initialized. */
792 m->m_data = m->m_ext.ext_buf;
793 m->m_flags = (flags | M_EXT);
799 * This is the protocol drain routine. Called by UMA whenever any of the
800 * mbuf zones is closed to its limit.
802 * No locks should be held when this is called. The drain routines have to
803 * presently acquire some locks which raises the possibility of lock order
807 mb_reclaim(uma_zone_t zone __unused, int pending __unused)
809 struct epoch_tracker et;
813 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK | WARN_PANIC, NULL, __func__);
816 for (dp = domains; dp != NULL; dp = dp->dom_next)
817 for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++)
818 if (pr->pr_drain != NULL)
824 * Free "count" units of I/O from an mbuf chain. They could be held
825 * in EXT_PGS or just as a normal mbuf. This code is intended to be
826 * called in an error path (I/O error, closed connection, etc).
829 mb_free_notready(struct mbuf *m, int count)
833 for (i = 0; i < count && m != NULL; i++) {
834 if ((m->m_flags & M_EXT) != 0 &&
835 m->m_ext.ext_type == EXT_PGS) {
837 if (m->m_ext_pgs.nrdy != 0)
842 KASSERT(i == count, ("Removed only %d items from %p", i, m));
846 * Compress an unmapped mbuf into a simple mbuf when it holds a small
847 * amount of data. This is used as a DOS defense to avoid having
848 * small packets tie up wired pages, an ext_pgs structure, and an
849 * mbuf. Since this converts the existing mbuf in place, it can only
850 * be used if there are no other references to 'm'.
853 mb_unmapped_compress(struct mbuf *m)
855 volatile u_int *refcnt;
859 * Assert that 'm' does not have a packet header. If 'm' had
860 * a packet header, it would only be able to hold MHLEN bytes
861 * and m_data would have to be initialized differently.
863 KASSERT((m->m_flags & M_PKTHDR) == 0 && (m->m_flags & M_EXT) &&
864 m->m_ext.ext_type == EXT_PGS,
865 ("%s: m %p !M_EXT or !EXT_PGS or M_PKTHDR", __func__, m));
866 KASSERT(m->m_len <= MLEN, ("m_len too large %p", m));
868 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
869 refcnt = &m->m_ext.ext_count;
871 KASSERT(m->m_ext.ext_cnt != NULL,
872 ("%s: no refcounting pointer on %p", __func__, m));
873 refcnt = m->m_ext.ext_cnt;
879 m_copydata(m, 0, m->m_len, buf);
881 /* Free the backing pages. */
882 m->m_ext.ext_free(m);
884 /* Turn 'm' into a "normal" mbuf. */
885 m->m_flags &= ~(M_EXT | M_RDONLY | M_NOMAP);
886 m->m_data = m->m_dat;
888 /* Copy data back into m. */
889 bcopy(buf, mtod(m, char *), m->m_len);
895 * These next few routines are used to permit downgrading an unmapped
896 * mbuf to a chain of mapped mbufs. This is used when an interface
897 * doesn't supported unmapped mbufs or if checksums need to be
898 * computed in software.
900 * Each unmapped mbuf is converted to a chain of mbufs. First, any
901 * TLS header data is stored in a regular mbuf. Second, each page of
902 * unmapped data is stored in an mbuf with an EXT_SFBUF external
903 * cluster. These mbufs use an sf_buf to provide a valid KVA for the
904 * associated physical page. They also hold a reference on the
905 * original EXT_PGS mbuf to ensure the physical page doesn't go away.
906 * Finally, any TLS trailer data is stored in a regular mbuf.
908 * mb_unmapped_free_mext() is the ext_free handler for the EXT_SFBUF
909 * mbufs. It frees the associated sf_buf and releases its reference
910 * on the original EXT_PGS mbuf.
912 * _mb_unmapped_to_ext() is a helper function that converts a single
913 * unmapped mbuf into a chain of mbufs.
915 * mb_unmapped_to_ext() is the public function that walks an mbuf
916 * chain converting any unmapped mbufs to mapped mbufs. It returns
917 * the new chain of unmapped mbufs on success. On failure it frees
918 * the original mbuf chain and returns NULL.
921 mb_unmapped_free_mext(struct mbuf *m)
926 sf = m->m_ext.ext_arg1;
929 /* Drop the reference on the backing EXT_PGS mbuf. */
930 old_m = m->m_ext.ext_arg2;
935 _mb_unmapped_to_ext(struct mbuf *m)
937 struct mbuf_ext_pgs *ext_pgs;
938 struct mbuf *m_new, *top, *prev, *mref;
941 int i, len, off, pglen, pgoff, seglen, segoff;
942 volatile u_int *refcnt;
945 MBUF_EXT_PGS_ASSERT(m);
946 ext_pgs = &m->m_ext_pgs;
948 KASSERT(ext_pgs->tls == NULL, ("%s: can't convert TLS mbuf %p",
951 /* See if this is the mbuf that holds the embedded refcount. */
952 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
953 refcnt = &m->m_ext.ext_count;
956 KASSERT(m->m_ext.ext_cnt != NULL,
957 ("%s: no refcounting pointer on %p", __func__, m));
958 refcnt = m->m_ext.ext_cnt;
959 mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
962 /* Skip over any data removed from the front. */
963 off = mtod(m, vm_offset_t);
966 if (ext_pgs->hdr_len != 0) {
967 if (off >= ext_pgs->hdr_len) {
968 off -= ext_pgs->hdr_len;
970 seglen = ext_pgs->hdr_len - off;
972 seglen = min(seglen, len);
975 m_new = m_get(M_NOWAIT, MT_DATA);
978 m_new->m_len = seglen;
980 memcpy(mtod(m_new, void *), &m->m_epg_hdr[segoff],
984 pgoff = ext_pgs->first_pg_off;
985 for (i = 0; i < ext_pgs->npgs && len > 0; i++) {
986 pglen = m_epg_pagelen(m, i, pgoff);
992 seglen = pglen - off;
993 segoff = pgoff + off;
995 seglen = min(seglen, len);
998 pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]);
999 m_new = m_get(M_NOWAIT, MT_DATA);
1005 prev->m_next = m_new;
1008 sf = sf_buf_alloc(pg, SFB_NOWAIT);
1013 m_extadd(m_new, (char *)sf_buf_kva(sf), PAGE_SIZE,
1014 mb_unmapped_free_mext, sf, mref, M_RDONLY, EXT_SFBUF);
1015 m_new->m_data += segoff;
1016 m_new->m_len = seglen;
1021 KASSERT((off + len) <= ext_pgs->trail_len,
1022 ("off + len > trail (%d + %d > %d)", off, len,
1023 ext_pgs->trail_len));
1024 m_new = m_get(M_NOWAIT, MT_DATA);
1030 prev->m_next = m_new;
1032 memcpy(mtod(m_new, void *), &m->m_epg_trail[off], len);
1037 * Obtain an additional reference on the old mbuf for
1038 * each created EXT_SFBUF mbuf. They will be dropped
1039 * in mb_unmapped_free_mext().
1044 atomic_add_int(refcnt, ref_inc);
1052 * Obtain an additional reference on the old mbuf for
1053 * each created EXT_SFBUF mbuf. They will be
1054 * immediately dropped when these mbufs are freed
1060 atomic_add_int(refcnt, ref_inc);
1068 mb_unmapped_to_ext(struct mbuf *top)
1070 struct mbuf *m, *next, *prev = NULL;
1073 for (m = top; m != NULL; m = next) {
1074 /* m might be freed, so cache the next pointer. */
1076 if (m->m_flags & M_NOMAP) {
1079 * Remove 'm' from the new chain so
1080 * that the 'top' chain terminates
1081 * before 'm' in case 'top' is freed
1084 prev->m_next = NULL;
1086 m = _mb_unmapped_to_ext(m);
1099 * Replaced one mbuf with a chain, so we must
1100 * find the end of chain.
1114 * Allocate an empty EXT_PGS mbuf. The ext_free routine is
1115 * responsible for freeing any pages backing this mbuf when it is
1119 mb_alloc_ext_pgs(int how, m_ext_free_t ext_free)
1122 struct mbuf_ext_pgs *ext_pgs;
1124 m = m_get(how, MT_DATA);
1128 ext_pgs = &m->m_ext_pgs;
1131 ext_pgs->first_pg_off = 0;
1132 ext_pgs->last_pg_len = 0;
1134 ext_pgs->hdr_len = 0;
1135 ext_pgs->trail_len = 0;
1136 ext_pgs->tls = NULL;
1139 m->m_flags |= (M_EXT | M_RDONLY | M_NOMAP);
1140 m->m_ext.ext_type = EXT_PGS;
1141 m->m_ext.ext_flags = EXT_FLAG_EMBREF;
1142 m->m_ext.ext_count = 1;
1143 m->m_ext.ext_size = 0;
1144 m->m_ext.ext_free = ext_free;
1148 #ifdef INVARIANT_SUPPORT
1150 mb_ext_pgs_check(struct mbuf *m)
1152 struct mbuf_ext_pgs *ext_pgs = &m->m_ext_pgs;
1155 * NB: This expects a non-empty buffer (npgs > 0 and
1158 KASSERT(ext_pgs->npgs > 0,
1159 ("ext_pgs with no valid pages: %p", ext_pgs));
1160 KASSERT(ext_pgs->npgs <= nitems(m->m_epg_pa),
1161 ("ext_pgs with too many pages: %p", ext_pgs));
1162 KASSERT(ext_pgs->nrdy <= ext_pgs->npgs,
1163 ("ext_pgs with too many ready pages: %p", ext_pgs));
1164 KASSERT(ext_pgs->first_pg_off < PAGE_SIZE,
1165 ("ext_pgs with too large page offset: %p", ext_pgs));
1166 KASSERT(ext_pgs->last_pg_len > 0,
1167 ("ext_pgs with zero last page length: %p", ext_pgs));
1168 KASSERT(ext_pgs->last_pg_len <= PAGE_SIZE,
1169 ("ext_pgs with too large last page length: %p", ext_pgs));
1170 if (ext_pgs->npgs == 1) {
1171 KASSERT(ext_pgs->first_pg_off + ext_pgs->last_pg_len <=
1172 PAGE_SIZE, ("ext_pgs with single page too large: %p",
1175 KASSERT(ext_pgs->hdr_len <= sizeof(m->m_epg_hdr),
1176 ("ext_pgs with too large header length: %p", ext_pgs));
1177 KASSERT(ext_pgs->trail_len <= sizeof(m->m_epg_trail),
1178 ("ext_pgs with too large header length: %p", ext_pgs));
1183 * Clean up after mbufs with M_EXT storage attached to them if the
1184 * reference count hits 1.
1187 mb_free_ext(struct mbuf *m)
1189 volatile u_int *refcnt;
1193 KASSERT(m->m_flags & M_EXT, ("%s: M_EXT not set on %p", __func__, m));
1195 /* See if this is the mbuf that holds the embedded refcount. */
1196 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
1197 refcnt = &m->m_ext.ext_count;
1200 KASSERT(m->m_ext.ext_cnt != NULL,
1201 ("%s: no refcounting pointer on %p", __func__, m));
1202 refcnt = m->m_ext.ext_cnt;
1203 mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
1207 * Check if the header is embedded in the cluster. It is
1208 * important that we can't touch any of the mbuf fields
1209 * after we have freed the external storage, since mbuf
1210 * could have been embedded in it. For now, the mbufs
1211 * embedded into the cluster are always of type EXT_EXTREF,
1212 * and for this type we won't free the mref.
1214 if (m->m_flags & M_NOFREE) {
1216 KASSERT(m->m_ext.ext_type == EXT_EXTREF ||
1217 m->m_ext.ext_type == EXT_RXRING,
1218 ("%s: no-free mbuf %p has wrong type", __func__, m));
1222 /* Free attached storage if this mbuf is the only reference to it. */
1223 if (*refcnt == 1 || atomic_fetchadd_int(refcnt, -1) == 1) {
1224 switch (m->m_ext.ext_type) {
1226 /* The packet zone is special. */
1229 uma_zfree(zone_pack, mref);
1232 uma_zfree(zone_clust, m->m_ext.ext_buf);
1233 uma_zfree(zone_mbuf, mref);
1236 uma_zfree(zone_jumbop, m->m_ext.ext_buf);
1237 uma_zfree(zone_mbuf, mref);
1240 uma_zfree(zone_jumbo9, m->m_ext.ext_buf);
1241 uma_zfree(zone_mbuf, mref);
1244 uma_zfree(zone_jumbo16, m->m_ext.ext_buf);
1245 uma_zfree(zone_mbuf, mref);
1249 struct ktls_session *tls;
1252 KASSERT(mref->m_ext.ext_free != NULL,
1253 ("%s: ext_free not set", __func__));
1254 mref->m_ext.ext_free(mref);
1256 tls = mref->m_ext_pgs.tls;
1258 !refcount_release_if_not_last(&tls->refcount))
1259 ktls_enqueue_to_free(mref);
1262 uma_zfree(zone_mbuf, mref);
1268 case EXT_DISPOSABLE:
1269 KASSERT(mref->m_ext.ext_free != NULL,
1270 ("%s: ext_free not set", __func__));
1271 mref->m_ext.ext_free(mref);
1272 uma_zfree(zone_mbuf, mref);
1275 KASSERT(m->m_ext.ext_free != NULL,
1276 ("%s: ext_free not set", __func__));
1277 m->m_ext.ext_free(m);
1280 KASSERT(m->m_ext.ext_free == NULL,
1281 ("%s: ext_free is set", __func__));
1284 KASSERT(m->m_ext.ext_type == 0,
1285 ("%s: unknown ext_type", __func__));
1289 if (freembuf && m != mref)
1290 uma_zfree(zone_mbuf, m);
1294 * Official mbuf(9) allocation KPI for stack and drivers:
1296 * m_get() - a single mbuf without any attachments, sys/mbuf.h.
1297 * m_gethdr() - a single mbuf initialized as M_PKTHDR, sys/mbuf.h.
1298 * m_getcl() - an mbuf + 2k cluster, sys/mbuf.h.
1299 * m_clget() - attach cluster to already allocated mbuf.
1300 * m_cljget() - attach jumbo cluster to already allocated mbuf.
1301 * m_get2() - allocate minimum mbuf that would fit size argument.
1302 * m_getm2() - allocate a chain of mbufs/clusters.
1303 * m_extadd() - attach external cluster to mbuf.
1305 * m_free() - free single mbuf with its tags and ext, sys/mbuf.h.
1306 * m_freem() - free chain of mbufs.
1310 m_clget(struct mbuf *m, int how)
1313 KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
1315 m->m_ext.ext_buf = (char *)NULL;
1316 uma_zalloc_arg(zone_clust, m, how);
1318 * On a cluster allocation failure, drain the packet zone and retry,
1319 * we might be able to loosen a few clusters up on the drain.
1321 if ((how & M_NOWAIT) && (m->m_ext.ext_buf == NULL)) {
1322 uma_zone_reclaim(zone_pack, UMA_RECLAIM_DRAIN);
1323 uma_zalloc_arg(zone_clust, m, how);
1325 MBUF_PROBE2(m__clget, m, how);
1326 return (m->m_flags & M_EXT);
1330 * m_cljget() is different from m_clget() as it can allocate clusters without
1331 * attaching them to an mbuf. In that case the return value is the pointer
1332 * to the cluster of the requested size. If an mbuf was specified, it gets
1333 * the cluster attached to it and the return value can be safely ignored.
1334 * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
1337 m_cljget(struct mbuf *m, int how, int size)
1343 KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
1345 m->m_ext.ext_buf = NULL;
1348 zone = m_getzone(size);
1349 retval = uma_zalloc_arg(zone, m, how);
1351 MBUF_PROBE4(m__cljget, m, how, size, retval);
1357 * m_get2() allocates minimum mbuf that would fit "size" argument.
1360 m_get2(int size, int how, short type, int flags)
1362 struct mb_args args;
1368 if (size <= MHLEN || (size <= MLEN && (flags & M_PKTHDR) == 0))
1369 return (uma_zalloc_arg(zone_mbuf, &args, how));
1370 if (size <= MCLBYTES)
1371 return (uma_zalloc_arg(zone_pack, &args, how));
1373 if (size > MJUMPAGESIZE)
1376 m = uma_zalloc_arg(zone_mbuf, &args, how);
1380 n = uma_zalloc_arg(zone_jumbop, m, how);
1382 uma_zfree(zone_mbuf, m);
1390 * m_getjcl() returns an mbuf with a cluster of the specified size attached.
1391 * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
1394 m_getjcl(int how, short type, int flags, int size)
1396 struct mb_args args;
1400 if (size == MCLBYTES)
1401 return m_getcl(how, type, flags);
1406 m = uma_zalloc_arg(zone_mbuf, &args, how);
1410 zone = m_getzone(size);
1411 n = uma_zalloc_arg(zone, m, how);
1413 uma_zfree(zone_mbuf, m);
1420 * Allocate a given length worth of mbufs and/or clusters (whatever fits
1421 * best) and return a pointer to the top of the allocated chain. If an
1422 * existing mbuf chain is provided, then we will append the new chain
1423 * to the existing one and return a pointer to the provided mbuf.
1426 m_getm2(struct mbuf *m, int len, int how, short type, int flags)
1428 struct mbuf *mb, *nm = NULL, *mtail = NULL;
1430 KASSERT(len >= 0, ("%s: len is < 0", __func__));
1432 /* Validate flags. */
1433 flags &= (M_PKTHDR | M_EOR);
1435 /* Packet header mbuf must be first in chain. */
1436 if ((flags & M_PKTHDR) && m != NULL)
1439 /* Loop and append maximum sized mbufs to the chain tail. */
1442 mb = m_getjcl(how, type, (flags & M_PKTHDR),
1444 else if (len >= MINCLSIZE)
1445 mb = m_getcl(how, type, (flags & M_PKTHDR));
1446 else if (flags & M_PKTHDR)
1447 mb = m_gethdr(how, type);
1449 mb = m_get(how, type);
1451 /* Fail the whole operation if one mbuf can't be allocated. */
1465 flags &= ~M_PKTHDR; /* Only valid on the first mbuf. */
1468 mtail->m_flags |= M_EOR; /* Only valid on the last mbuf. */
1470 /* If mbuf was supplied, append new chain to the end of it. */
1472 for (mtail = m; mtail->m_next != NULL; mtail = mtail->m_next)
1475 mtail->m_flags &= ~M_EOR;
1483 * Configure a provided mbuf to refer to the provided external storage
1484 * buffer and setup a reference count for said buffer.
1487 * mb The existing mbuf to which to attach the provided buffer.
1488 * buf The address of the provided external storage buffer.
1489 * size The size of the provided buffer.
1490 * freef A pointer to a routine that is responsible for freeing the
1491 * provided external storage buffer.
1492 * args A pointer to an argument structure (of any type) to be passed
1493 * to the provided freef routine (may be NULL).
1494 * flags Any other flags to be passed to the provided mbuf.
1495 * type The type that the external storage buffer should be
1502 m_extadd(struct mbuf *mb, char *buf, u_int size, m_ext_free_t freef,
1503 void *arg1, void *arg2, int flags, int type)
1506 KASSERT(type != EXT_CLUSTER, ("%s: EXT_CLUSTER not allowed", __func__));
1508 mb->m_flags |= (M_EXT | flags);
1509 mb->m_ext.ext_buf = buf;
1510 mb->m_data = mb->m_ext.ext_buf;
1511 mb->m_ext.ext_size = size;
1512 mb->m_ext.ext_free = freef;
1513 mb->m_ext.ext_arg1 = arg1;
1514 mb->m_ext.ext_arg2 = arg2;
1515 mb->m_ext.ext_type = type;
1517 if (type != EXT_EXTREF) {
1518 mb->m_ext.ext_count = 1;
1519 mb->m_ext.ext_flags = EXT_FLAG_EMBREF;
1521 mb->m_ext.ext_flags = 0;
1525 * Free an entire chain of mbufs and associated external buffers, if
1529 m_freem(struct mbuf *mb)
1532 MBUF_PROBE1(m__freem, mb);
1538 m_snd_tag_init(struct m_snd_tag *mst, struct ifnet *ifp)
1543 refcount_init(&mst->refcount, 1);
1544 counter_u64_add(snd_tag_count, 1);
1548 m_snd_tag_destroy(struct m_snd_tag *mst)
1553 ifp->if_snd_tag_free(mst);
1555 counter_u64_add(snd_tag_count, -1);