1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (c) 2004, 2005,
5 * Bosko Milekic <bmilekic@FreeBSD.org>. All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice unmodified, this list of conditions and the following
12 * disclaimer.
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.
16 *
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
27 * SUCH DAMAGE.
28 */
29
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD: stable/12/sys/kern/kern_mbuf.c 373247 2023-10-12 10:13:13Z zlei $");
32
33 #include "opt_param.h"
34
35 #include <sys/param.h>
36 #include <sys/domainset.h>
37 #include <sys/malloc.h>
38 #include <sys/types.h>
39 #include <sys/systm.h>
40 #include <sys/mbuf.h>
41 #include <sys/domain.h>
42 #include <sys/eventhandler.h>
43 #include <sys/kernel.h>
44 #include <sys/limits.h>
45 #include <sys/lock.h>
46 #include <sys/mutex.h>
47 #include <sys/protosw.h>
48 #include <sys/smp.h>
49 #include <sys/sysctl.h>
50
51 #include <vm/vm.h>
52 #include <vm/vm_extern.h>
53 #include <vm/vm_kern.h>
54 #include <vm/vm_page.h>
55 #include <vm/vm_map.h>
56 #include <vm/uma.h>
57 #include <vm/uma_dbg.h>
58
59 /*
60 * In FreeBSD, Mbufs and Mbuf Clusters are allocated from UMA
61 * Zones.
62 *
63 * Mbuf Clusters (2K, contiguous) are allocated from the Cluster
64 * Zone. The Zone can be capped at kern.ipc.nmbclusters, if the
65 * administrator so desires.
66 *
67 * Mbufs are allocated from a UMA Master Zone called the Mbuf
68 * Zone.
69 *
70 * Additionally, FreeBSD provides a Packet Zone, which it
71 * configures as a Secondary Zone to the Mbuf Master Zone,
72 * thus sharing backend Slab kegs with the Mbuf Master Zone.
73 *
74 * Thus common-case allocations and locking are simplified:
75 *
76 * m_clget() m_getcl()
77 * | |
78 * | .------------>[(Packet Cache)] m_get(), m_gethdr()
79 * | | [ Packet ] |
80 * [(Cluster Cache)] [ Secondary ] [ (Mbuf Cache) ]
81 * [ Cluster Zone ] [ Zone ] [ Mbuf Master Zone ]
82 * | \________ |
83 * [ Cluster Keg ] \ /
84 * | [ Mbuf Keg ]
85 * [ Cluster Slabs ] |
86 * | [ Mbuf Slabs ]
87 * \____________(VM)_________________/
88 *
89 *
90 * Whenever an object is allocated with uma_zalloc() out of
91 * one of the Zones its _ctor_ function is executed. The same
92 * for any deallocation through uma_zfree() the _dtor_ function
93 * is executed.
94 *
95 * Caches are per-CPU and are filled from the Master Zone.
96 *
97 * Whenever an object is allocated from the underlying global
98 * memory pool it gets pre-initialized with the _zinit_ functions.
99 * When the Keg's are overfull objects get decommissioned with
100 * _zfini_ functions and free'd back to the global memory pool.
101 *
102 */
103
104 int nmbufs; /* limits number of mbufs */
105 int nmbclusters; /* limits number of mbuf clusters */
106 int nmbjumbop; /* limits number of page size jumbo clusters */
107 int nmbjumbo9; /* limits number of 9k jumbo clusters */
108 int nmbjumbo16; /* limits number of 16k jumbo clusters */
109
110 static quad_t maxmbufmem; /* overall real memory limit for all mbufs */
111
112 SYSCTL_QUAD(_kern_ipc, OID_AUTO, maxmbufmem, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &maxmbufmem, 0,
113 "Maximum real memory allocatable to various mbuf types");
114
115 /*
116 * tunable_mbinit() has to be run before any mbuf allocations are done.
117 */
118 static void
tunable_mbinit(void * dummy)119 tunable_mbinit(void *dummy)
120 {
121 quad_t realmem;
122
123 /*
124 * The default limit for all mbuf related memory is 1/2 of all
125 * available kernel memory (physical or kmem).
126 * At most it can be 3/4 of available kernel memory.
127 */
128 realmem = qmin((quad_t)physmem * PAGE_SIZE, vm_kmem_size);
129 maxmbufmem = realmem / 2;
130 TUNABLE_QUAD_FETCH("kern.ipc.maxmbufmem", &maxmbufmem);
131 if (maxmbufmem > realmem / 4 * 3)
132 maxmbufmem = realmem / 4 * 3;
133
134 TUNABLE_INT_FETCH("kern.ipc.nmbclusters", &nmbclusters);
135 if (nmbclusters == 0)
136 nmbclusters = maxmbufmem / MCLBYTES / 4;
137
138 TUNABLE_INT_FETCH("kern.ipc.nmbjumbop", &nmbjumbop);
139 if (nmbjumbop == 0)
140 nmbjumbop = maxmbufmem / MJUMPAGESIZE / 4;
141
142 TUNABLE_INT_FETCH("kern.ipc.nmbjumbo9", &nmbjumbo9);
143 if (nmbjumbo9 == 0)
144 nmbjumbo9 = maxmbufmem / MJUM9BYTES / 6;
145
146 TUNABLE_INT_FETCH("kern.ipc.nmbjumbo16", &nmbjumbo16);
147 if (nmbjumbo16 == 0)
148 nmbjumbo16 = maxmbufmem / MJUM16BYTES / 6;
149
150 /*
151 * We need at least as many mbufs as we have clusters of
152 * the various types added together.
153 */
154 TUNABLE_INT_FETCH("kern.ipc.nmbufs", &nmbufs);
155 if (nmbufs < nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16)
156 nmbufs = lmax(maxmbufmem / MSIZE / 5,
157 nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16);
158 }
159 SYSINIT(tunable_mbinit, SI_SUB_KMEM, SI_ORDER_MIDDLE, tunable_mbinit, NULL);
160
161 static int
sysctl_nmbclusters(SYSCTL_HANDLER_ARGS)162 sysctl_nmbclusters(SYSCTL_HANDLER_ARGS)
163 {
164 int error, newnmbclusters;
165
166 newnmbclusters = nmbclusters;
167 error = sysctl_handle_int(oidp, &newnmbclusters, 0, req);
168 if (error == 0 && req->newptr && newnmbclusters != nmbclusters) {
169 if (newnmbclusters > nmbclusters &&
170 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
171 nmbclusters = newnmbclusters;
172 nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
173 EVENTHANDLER_INVOKE(nmbclusters_change);
174 } else
175 error = EINVAL;
176 }
177 return (error);
178 }
179 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbclusters,
180 CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
181 &nmbclusters, 0, sysctl_nmbclusters, "IU",
182 "Maximum number of mbuf clusters allowed");
183
184 static int
sysctl_nmbjumbop(SYSCTL_HANDLER_ARGS)185 sysctl_nmbjumbop(SYSCTL_HANDLER_ARGS)
186 {
187 int error, newnmbjumbop;
188
189 newnmbjumbop = nmbjumbop;
190 error = sysctl_handle_int(oidp, &newnmbjumbop, 0, req);
191 if (error == 0 && req->newptr && newnmbjumbop != nmbjumbop) {
192 if (newnmbjumbop > nmbjumbop &&
193 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
194 nmbjumbop = newnmbjumbop;
195 nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
196 } else
197 error = EINVAL;
198 }
199 return (error);
200 }
201 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbop,
202 CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
203 &nmbjumbop, 0, sysctl_nmbjumbop, "IU",
204 "Maximum number of mbuf page size jumbo clusters allowed");
205
206 static int
sysctl_nmbjumbo9(SYSCTL_HANDLER_ARGS)207 sysctl_nmbjumbo9(SYSCTL_HANDLER_ARGS)
208 {
209 int error, newnmbjumbo9;
210
211 newnmbjumbo9 = nmbjumbo9;
212 error = sysctl_handle_int(oidp, &newnmbjumbo9, 0, req);
213 if (error == 0 && req->newptr && newnmbjumbo9 != nmbjumbo9) {
214 if (newnmbjumbo9 > nmbjumbo9 &&
215 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
216 nmbjumbo9 = newnmbjumbo9;
217 nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
218 } else
219 error = EINVAL;
220 }
221 return (error);
222 }
223 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo9,
224 CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
225 &nmbjumbo9, 0, sysctl_nmbjumbo9, "IU",
226 "Maximum number of mbuf 9k jumbo clusters allowed");
227
228 static int
sysctl_nmbjumbo16(SYSCTL_HANDLER_ARGS)229 sysctl_nmbjumbo16(SYSCTL_HANDLER_ARGS)
230 {
231 int error, newnmbjumbo16;
232
233 newnmbjumbo16 = nmbjumbo16;
234 error = sysctl_handle_int(oidp, &newnmbjumbo16, 0, req);
235 if (error == 0 && req->newptr && newnmbjumbo16 != nmbjumbo16) {
236 if (newnmbjumbo16 > nmbjumbo16 &&
237 nmbufs >= nmbclusters + nmbjumbop + nmbjumbo9 + nmbjumbo16) {
238 nmbjumbo16 = newnmbjumbo16;
239 nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
240 } else
241 error = EINVAL;
242 }
243 return (error);
244 }
245 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbjumbo16,
246 CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
247 &nmbjumbo16, 0, sysctl_nmbjumbo16, "IU",
248 "Maximum number of mbuf 16k jumbo clusters allowed");
249
250 static int
sysctl_nmbufs(SYSCTL_HANDLER_ARGS)251 sysctl_nmbufs(SYSCTL_HANDLER_ARGS)
252 {
253 int error, newnmbufs;
254
255 newnmbufs = nmbufs;
256 error = sysctl_handle_int(oidp, &newnmbufs, 0, req);
257 if (error == 0 && req->newptr && newnmbufs != nmbufs) {
258 if (newnmbufs > nmbufs) {
259 nmbufs = newnmbufs;
260 nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
261 EVENTHANDLER_INVOKE(nmbufs_change);
262 } else
263 error = EINVAL;
264 }
265 return (error);
266 }
267 SYSCTL_PROC(_kern_ipc, OID_AUTO, nmbufs,
268 CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
269 &nmbufs, 0, sysctl_nmbufs, "IU",
270 "Maximum number of mbufs allowed");
271
272 /*
273 * Zones from which we allocate.
274 */
275 uma_zone_t zone_mbuf;
276 uma_zone_t zone_clust;
277 uma_zone_t zone_pack;
278 uma_zone_t zone_jumbop;
279 uma_zone_t zone_jumbo9;
280 uma_zone_t zone_jumbo16;
281
282 /*
283 * Local prototypes.
284 */
285 static int mb_ctor_mbuf(void *, int, void *, int);
286 static int mb_ctor_clust(void *, int, void *, int);
287 static int mb_ctor_pack(void *, int, void *, int);
288 static void mb_dtor_mbuf(void *, int, void *);
289 static void mb_dtor_pack(void *, int, void *);
290 static int mb_zinit_pack(void *, int, int);
291 static void mb_zfini_pack(void *, int);
292 static void mb_reclaim(uma_zone_t, int);
293 static void *mbuf_jumbo_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
294
295 /* Ensure that MSIZE is a power of 2. */
296 CTASSERT((((MSIZE - 1) ^ MSIZE) + 1) >> 1 == MSIZE);
297
298 /*
299 * Initialize FreeBSD Network buffer allocation.
300 */
301 static void
mbuf_init(void * dummy)302 mbuf_init(void *dummy)
303 {
304
305 /*
306 * Configure UMA zones for Mbufs, Clusters, and Packets.
307 */
308 zone_mbuf = uma_zcreate(MBUF_MEM_NAME, MSIZE,
309 mb_ctor_mbuf, mb_dtor_mbuf,
310 #ifdef INVARIANTS
311 trash_init, trash_fini,
312 #else
313 NULL, NULL,
314 #endif
315 MSIZE - 1, UMA_ZONE_MAXBUCKET);
316 if (nmbufs > 0)
317 nmbufs = uma_zone_set_max(zone_mbuf, nmbufs);
318 uma_zone_set_warning(zone_mbuf, "kern.ipc.nmbufs limit reached");
319 uma_zone_set_maxaction(zone_mbuf, mb_reclaim);
320
321 zone_clust = uma_zcreate(MBUF_CLUSTER_MEM_NAME, MCLBYTES,
322 mb_ctor_clust,
323 #ifdef INVARIANTS
324 trash_dtor, trash_init, trash_fini,
325 #else
326 NULL, NULL, NULL,
327 #endif
328 UMA_ALIGN_PTR, 0);
329 if (nmbclusters > 0)
330 nmbclusters = uma_zone_set_max(zone_clust, nmbclusters);
331 uma_zone_set_warning(zone_clust, "kern.ipc.nmbclusters limit reached");
332 uma_zone_set_maxaction(zone_clust, mb_reclaim);
333
334 zone_pack = uma_zsecond_create(MBUF_PACKET_MEM_NAME, mb_ctor_pack,
335 mb_dtor_pack, mb_zinit_pack, mb_zfini_pack, zone_mbuf);
336
337 /* Make jumbo frame zone too. Page size, 9k and 16k. */
338 zone_jumbop = uma_zcreate(MBUF_JUMBOP_MEM_NAME, MJUMPAGESIZE,
339 mb_ctor_clust,
340 #ifdef INVARIANTS
341 trash_dtor, trash_init, trash_fini,
342 #else
343 NULL, NULL, NULL,
344 #endif
345 UMA_ALIGN_PTR, 0);
346 if (nmbjumbop > 0)
347 nmbjumbop = uma_zone_set_max(zone_jumbop, nmbjumbop);
348 uma_zone_set_warning(zone_jumbop, "kern.ipc.nmbjumbop limit reached");
349 uma_zone_set_maxaction(zone_jumbop, mb_reclaim);
350
351 zone_jumbo9 = uma_zcreate(MBUF_JUMBO9_MEM_NAME, MJUM9BYTES,
352 mb_ctor_clust,
353 #ifdef INVARIANTS
354 trash_dtor, trash_init, trash_fini,
355 #else
356 NULL, NULL, NULL,
357 #endif
358 UMA_ALIGN_PTR, 0);
359 uma_zone_set_allocf(zone_jumbo9, mbuf_jumbo_alloc);
360 if (nmbjumbo9 > 0)
361 nmbjumbo9 = uma_zone_set_max(zone_jumbo9, nmbjumbo9);
362 uma_zone_set_warning(zone_jumbo9, "kern.ipc.nmbjumbo9 limit reached");
363 uma_zone_set_maxaction(zone_jumbo9, mb_reclaim);
364
365 zone_jumbo16 = uma_zcreate(MBUF_JUMBO16_MEM_NAME, MJUM16BYTES,
366 mb_ctor_clust,
367 #ifdef INVARIANTS
368 trash_dtor, trash_init, trash_fini,
369 #else
370 NULL, NULL, NULL,
371 #endif
372 UMA_ALIGN_PTR, 0);
373 uma_zone_set_allocf(zone_jumbo16, mbuf_jumbo_alloc);
374 if (nmbjumbo16 > 0)
375 nmbjumbo16 = uma_zone_set_max(zone_jumbo16, nmbjumbo16);
376 uma_zone_set_warning(zone_jumbo16, "kern.ipc.nmbjumbo16 limit reached");
377 uma_zone_set_maxaction(zone_jumbo16, mb_reclaim);
378
379 /*
380 * Hook event handler for low-memory situation, used to
381 * drain protocols and push data back to the caches (UMA
382 * later pushes it back to VM).
383 */
384 EVENTHANDLER_REGISTER(vm_lowmem, mb_reclaim, NULL,
385 EVENTHANDLER_PRI_FIRST);
386 }
387 SYSINIT(mbuf, SI_SUB_MBUF, SI_ORDER_FIRST, mbuf_init, NULL);
388
389 #ifdef NETDUMP
390 /*
391 * netdump makes use of a pre-allocated pool of mbufs and clusters. When
392 * netdump is configured, we initialize a set of UMA cache zones which return
393 * items from this pool. At panic-time, the regular UMA zone pointers are
394 * overwritten with those of the cache zones so that drivers may allocate and
395 * free mbufs and clusters without attempting to allocate physical memory.
396 *
397 * We keep mbufs and clusters in a pair of mbuf queues. In particular, for
398 * the purpose of caching clusters, we treat them as mbufs.
399 */
400 static struct mbufq nd_mbufq =
401 { STAILQ_HEAD_INITIALIZER(nd_mbufq.mq_head), 0, INT_MAX };
402 static struct mbufq nd_clustq =
403 { STAILQ_HEAD_INITIALIZER(nd_clustq.mq_head), 0, INT_MAX };
404
405 static int nd_clsize;
406 static uma_zone_t nd_zone_mbuf;
407 static uma_zone_t nd_zone_clust;
408 static uma_zone_t nd_zone_pack;
409
410 static int
nd_buf_import(void * arg,void ** store,int count,int domain __unused,int flags)411 nd_buf_import(void *arg, void **store, int count, int domain __unused,
412 int flags)
413 {
414 struct mbufq *q;
415 struct mbuf *m;
416 int i;
417
418 q = arg;
419
420 for (i = 0; i < count; i++) {
421 m = mbufq_dequeue(q);
422 if (m == NULL)
423 break;
424 trash_init(m, q == &nd_mbufq ? MSIZE : nd_clsize, flags);
425 store[i] = m;
426 }
427 return (i);
428 }
429
430 static void
nd_buf_release(void * arg,void ** store,int count)431 nd_buf_release(void *arg, void **store, int count)
432 {
433 struct mbufq *q;
434 struct mbuf *m;
435 int i;
436
437 q = arg;
438
439 for (i = 0; i < count; i++) {
440 m = store[i];
441 (void)mbufq_enqueue(q, m);
442 }
443 }
444
445 static int
nd_pack_import(void * arg __unused,void ** store,int count,int domain __unused,int flags __unused)446 nd_pack_import(void *arg __unused, void **store, int count, int domain __unused,
447 int flags __unused)
448 {
449 struct mbuf *m;
450 void *clust;
451 int i;
452
453 for (i = 0; i < count; i++) {
454 m = m_get(MT_DATA, M_NOWAIT);
455 if (m == NULL)
456 break;
457 clust = uma_zalloc(nd_zone_clust, M_NOWAIT);
458 if (clust == NULL) {
459 m_free(m);
460 break;
461 }
462 mb_ctor_clust(clust, nd_clsize, m, 0);
463 store[i] = m;
464 }
465 return (i);
466 }
467
468 static void
nd_pack_release(void * arg __unused,void ** store,int count)469 nd_pack_release(void *arg __unused, void **store, int count)
470 {
471 struct mbuf *m;
472 void *clust;
473 int i;
474
475 for (i = 0; i < count; i++) {
476 m = store[i];
477 clust = m->m_ext.ext_buf;
478 uma_zfree(nd_zone_clust, clust);
479 uma_zfree(nd_zone_mbuf, m);
480 }
481 }
482
483 /*
484 * Free the pre-allocated mbufs and clusters reserved for netdump, and destroy
485 * the corresponding UMA cache zones.
486 */
487 void
netdump_mbuf_drain(void)488 netdump_mbuf_drain(void)
489 {
490 struct mbuf *m;
491 void *item;
492
493 if (nd_zone_mbuf != NULL) {
494 uma_zdestroy(nd_zone_mbuf);
495 nd_zone_mbuf = NULL;
496 }
497 if (nd_zone_clust != NULL) {
498 uma_zdestroy(nd_zone_clust);
499 nd_zone_clust = NULL;
500 }
501 if (nd_zone_pack != NULL) {
502 uma_zdestroy(nd_zone_pack);
503 nd_zone_pack = NULL;
504 }
505
506 while ((m = mbufq_dequeue(&nd_mbufq)) != NULL)
507 m_free(m);
508 while ((item = mbufq_dequeue(&nd_clustq)) != NULL)
509 uma_zfree(m_getzone(nd_clsize), item);
510 }
511
512 /*
513 * Callback invoked immediately prior to starting a netdump.
514 */
515 void
netdump_mbuf_dump(void)516 netdump_mbuf_dump(void)
517 {
518
519 /*
520 * All cluster zones return buffers of the size requested by the
521 * drivers. It's up to the driver to reinitialize the zones if the
522 * MTU of a netdump-enabled interface changes.
523 */
524 printf("netdump: overwriting mbuf zone pointers\n");
525 zone_mbuf = nd_zone_mbuf;
526 zone_clust = nd_zone_clust;
527 zone_pack = nd_zone_pack;
528 zone_jumbop = nd_zone_clust;
529 zone_jumbo9 = nd_zone_clust;
530 zone_jumbo16 = nd_zone_clust;
531 }
532
533 /*
534 * Reinitialize the netdump mbuf+cluster pool and cache zones.
535 */
536 void
netdump_mbuf_reinit(int nmbuf,int nclust,int clsize)537 netdump_mbuf_reinit(int nmbuf, int nclust, int clsize)
538 {
539 struct mbuf *m;
540 void *item;
541
542 netdump_mbuf_drain();
543
544 nd_clsize = clsize;
545
546 nd_zone_mbuf = uma_zcache_create("netdump_" MBUF_MEM_NAME,
547 MSIZE, mb_ctor_mbuf, mb_dtor_mbuf,
548 #ifdef INVARIANTS
549 trash_init, trash_fini,
550 #else
551 NULL, NULL,
552 #endif
553 nd_buf_import, nd_buf_release,
554 &nd_mbufq, UMA_ZONE_NOBUCKET);
555
556 nd_zone_clust = uma_zcache_create("netdump_" MBUF_CLUSTER_MEM_NAME,
557 clsize, mb_ctor_clust,
558 #ifdef INVARIANTS
559 trash_dtor, trash_init, trash_fini,
560 #else
561 NULL, NULL, NULL,
562 #endif
563 nd_buf_import, nd_buf_release,
564 &nd_clustq, UMA_ZONE_NOBUCKET);
565
566 nd_zone_pack = uma_zcache_create("netdump_" MBUF_PACKET_MEM_NAME,
567 MCLBYTES, mb_ctor_pack, mb_dtor_pack, NULL, NULL,
568 nd_pack_import, nd_pack_release,
569 NULL, UMA_ZONE_NOBUCKET);
570
571 while (nmbuf-- > 0) {
572 m = m_get(MT_DATA, M_WAITOK);
573 uma_zfree(nd_zone_mbuf, m);
574 }
575 while (nclust-- > 0) {
576 item = uma_zalloc(m_getzone(nd_clsize), M_WAITOK);
577 uma_zfree(nd_zone_clust, item);
578 }
579 }
580 #endif /* NETDUMP */
581
582 /*
583 * UMA backend page allocator for the jumbo frame zones.
584 *
585 * Allocates kernel virtual memory that is backed by contiguous physical
586 * pages.
587 */
588 static void *
mbuf_jumbo_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * flags,int wait)589 mbuf_jumbo_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
590 int wait)
591 {
592
593 /* Inform UMA that this allocator uses kernel_map/object. */
594 *flags = UMA_SLAB_KERNEL;
595 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
596 bytes, wait, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
597 VM_MEMATTR_DEFAULT));
598 }
599
600 /*
601 * Constructor for Mbuf master zone.
602 *
603 * The 'arg' pointer points to a mb_args structure which
604 * contains call-specific information required to support the
605 * mbuf allocation API. See mbuf.h.
606 */
607 static int
mb_ctor_mbuf(void * mem,int size,void * arg,int how)608 mb_ctor_mbuf(void *mem, int size, void *arg, int how)
609 {
610 struct mbuf *m;
611 struct mb_args *args;
612 int error;
613 int flags;
614 short type;
615
616 #ifdef INVARIANTS
617 trash_ctor(mem, size, arg, how);
618 #endif
619 args = (struct mb_args *)arg;
620 type = args->type;
621
622 /*
623 * The mbuf is initialized later. The caller has the
624 * responsibility to set up any MAC labels too.
625 */
626 if (type == MT_NOINIT)
627 return (0);
628
629 m = (struct mbuf *)mem;
630 flags = args->flags;
631 MPASS((flags & M_NOFREE) == 0);
632
633 error = m_init(m, how, type, flags);
634
635 return (error);
636 }
637
638 /*
639 * The Mbuf master zone destructor.
640 */
641 static void
mb_dtor_mbuf(void * mem,int size,void * arg)642 mb_dtor_mbuf(void *mem, int size, void *arg)
643 {
644 struct mbuf *m;
645 unsigned long flags;
646
647 m = (struct mbuf *)mem;
648 flags = (unsigned long)arg;
649
650 KASSERT((m->m_flags & M_NOFREE) == 0, ("%s: M_NOFREE set", __func__));
651 if (!(flags & MB_DTOR_SKIP) && (m->m_flags & M_PKTHDR) && !SLIST_EMPTY(&m->m_pkthdr.tags))
652 m_tag_delete_chain(m, NULL);
653 #ifdef INVARIANTS
654 trash_dtor(mem, size, arg);
655 #endif
656 }
657
658 /*
659 * The Mbuf Packet zone destructor.
660 */
661 static void
mb_dtor_pack(void * mem,int size,void * arg)662 mb_dtor_pack(void *mem, int size, void *arg)
663 {
664 struct mbuf *m;
665
666 m = (struct mbuf *)mem;
667 if ((m->m_flags & M_PKTHDR) != 0)
668 m_tag_delete_chain(m, NULL);
669
670 /* Make sure we've got a clean cluster back. */
671 KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set", __func__));
672 KASSERT(m->m_ext.ext_buf != NULL, ("%s: ext_buf == NULL", __func__));
673 KASSERT(m->m_ext.ext_free == NULL, ("%s: ext_free != NULL", __func__));
674 KASSERT(m->m_ext.ext_arg1 == NULL, ("%s: ext_arg1 != NULL", __func__));
675 KASSERT(m->m_ext.ext_arg2 == NULL, ("%s: ext_arg2 != NULL", __func__));
676 KASSERT(m->m_ext.ext_size == MCLBYTES, ("%s: ext_size != MCLBYTES", __func__));
677 KASSERT(m->m_ext.ext_type == EXT_PACKET, ("%s: ext_type != EXT_PACKET", __func__));
678 #ifdef INVARIANTS
679 trash_dtor(m->m_ext.ext_buf, MCLBYTES, arg);
680 #endif
681 /*
682 * If there are processes blocked on zone_clust, waiting for pages
683 * to be freed up, * cause them to be woken up by draining the
684 * packet zone. We are exposed to a race here * (in the check for
685 * the UMA_ZFLAG_FULL) where we might miss the flag set, but that
686 * is deliberate. We don't want to acquire the zone lock for every
687 * mbuf free.
688 */
689 if (uma_zone_exhausted_nolock(zone_clust))
690 zone_drain(zone_pack);
691 }
692
693 /*
694 * The Cluster and Jumbo[PAGESIZE|9|16] zone constructor.
695 *
696 * Here the 'arg' pointer points to the Mbuf which we
697 * are configuring cluster storage for. If 'arg' is
698 * empty we allocate just the cluster without setting
699 * the mbuf to it. See mbuf.h.
700 */
701 static int
mb_ctor_clust(void * mem,int size,void * arg,int how)702 mb_ctor_clust(void *mem, int size, void *arg, int how)
703 {
704 struct mbuf *m;
705
706 #ifdef INVARIANTS
707 trash_ctor(mem, size, arg, how);
708 #endif
709 m = (struct mbuf *)arg;
710 if (m != NULL) {
711 m->m_ext.ext_buf = (char *)mem;
712 m->m_data = m->m_ext.ext_buf;
713 m->m_flags |= M_EXT;
714 m->m_ext.ext_free = NULL;
715 m->m_ext.ext_arg1 = NULL;
716 m->m_ext.ext_arg2 = NULL;
717 m->m_ext.ext_size = size;
718 m->m_ext.ext_type = m_gettype(size);
719 m->m_ext.ext_flags = EXT_FLAG_EMBREF;
720 m->m_ext.ext_count = 1;
721 }
722
723 return (0);
724 }
725
726 /*
727 * The Packet secondary zone's init routine, executed on the
728 * object's transition from mbuf keg slab to zone cache.
729 */
730 static int
mb_zinit_pack(void * mem,int size,int how)731 mb_zinit_pack(void *mem, int size, int how)
732 {
733 struct mbuf *m;
734
735 m = (struct mbuf *)mem; /* m is virgin. */
736 if (uma_zalloc_arg(zone_clust, m, how) == NULL ||
737 m->m_ext.ext_buf == NULL)
738 return (ENOMEM);
739 m->m_ext.ext_type = EXT_PACKET; /* Override. */
740 #ifdef INVARIANTS
741 trash_init(m->m_ext.ext_buf, MCLBYTES, how);
742 #endif
743 return (0);
744 }
745
746 /*
747 * The Packet secondary zone's fini routine, executed on the
748 * object's transition from zone cache to keg slab.
749 */
750 static void
mb_zfini_pack(void * mem,int size)751 mb_zfini_pack(void *mem, int size)
752 {
753 struct mbuf *m;
754
755 m = (struct mbuf *)mem;
756 #ifdef INVARIANTS
757 trash_fini(m->m_ext.ext_buf, MCLBYTES);
758 #endif
759 uma_zfree_arg(zone_clust, m->m_ext.ext_buf, NULL);
760 #ifdef INVARIANTS
761 trash_dtor(mem, size, NULL);
762 #endif
763 }
764
765 /*
766 * The "packet" keg constructor.
767 */
768 static int
mb_ctor_pack(void * mem,int size,void * arg,int how)769 mb_ctor_pack(void *mem, int size, void *arg, int how)
770 {
771 struct mbuf *m;
772 struct mb_args *args;
773 int error, flags;
774 short type;
775
776 m = (struct mbuf *)mem;
777 args = (struct mb_args *)arg;
778 flags = args->flags;
779 type = args->type;
780 MPASS((flags & M_NOFREE) == 0);
781
782 #ifdef INVARIANTS
783 trash_ctor(m->m_ext.ext_buf, MCLBYTES, arg, how);
784 #endif
785
786 error = m_init(m, how, type, flags);
787
788 /* m_ext is already initialized. */
789 m->m_data = m->m_ext.ext_buf;
790 m->m_flags = (flags | M_EXT);
791
792 return (error);
793 }
794
795 /*
796 * This is the protocol drain routine. Called by UMA whenever any of the
797 * mbuf zones is closed to its limit.
798 *
799 * No locks should be held when this is called. The drain routines have to
800 * presently acquire some locks which raises the possibility of lock order
801 * reversal.
802 */
803 static void
mb_reclaim(uma_zone_t zone __unused,int pending __unused)804 mb_reclaim(uma_zone_t zone __unused, int pending __unused)
805 {
806 struct domain *dp;
807 struct protosw *pr;
808
809 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK | WARN_PANIC, NULL, __func__);
810
811 for (dp = domains; dp != NULL; dp = dp->dom_next)
812 for (pr = dp->dom_protosw; pr < dp->dom_protoswNPROTOSW; pr++)
813 if (pr->pr_drain != NULL)
814 (*pr->pr_drain)();
815 }
816
817 /*
818 * Clean up after mbufs with M_EXT storage attached to them if the
819 * reference count hits 1.
820 */
821 void
mb_free_ext(struct mbuf * m)822 mb_free_ext(struct mbuf *m)
823 {
824 volatile u_int *refcnt;
825 struct mbuf *mref;
826 int freembuf;
827
828 KASSERT(m->m_flags & M_EXT, ("%s: M_EXT not set on %p", __func__, m));
829
830 /* See if this is the mbuf that holds the embedded refcount. */
831 if (m->m_ext.ext_flags & EXT_FLAG_EMBREF) {
832 refcnt = &m->m_ext.ext_count;
833 mref = m;
834 } else {
835 KASSERT(m->m_ext.ext_cnt != NULL,
836 ("%s: no refcounting pointer on %p", __func__, m));
837 refcnt = m->m_ext.ext_cnt;
838 mref = __containerof(refcnt, struct mbuf, m_ext.ext_count);
839 }
840
841 /*
842 * Check if the header is embedded in the cluster. It is
843 * important that we can't touch any of the mbuf fields
844 * after we have freed the external storage, since mbuf
845 * could have been embedded in it. For now, the mbufs
846 * embedded into the cluster are always of type EXT_EXTREF,
847 * and for this type we won't free the mref.
848 */
849 if (m->m_flags & M_NOFREE) {
850 freembuf = 0;
851 KASSERT(m->m_ext.ext_type == EXT_EXTREF,
852 ("%s: no-free mbuf %p has wrong type", __func__, m));
853 } else
854 freembuf = 1;
855
856 /* Free attached storage if this mbuf is the only reference to it. */
857 if (*refcnt == 1 || atomic_fetchadd_int(refcnt, -1) == 1) {
858 switch (m->m_ext.ext_type) {
859 case EXT_PACKET:
860 /* The packet zone is special. */
861 if (*refcnt == 0)
862 *refcnt = 1;
863 uma_zfree(zone_pack, mref);
864 break;
865 case EXT_CLUSTER:
866 uma_zfree(zone_clust, m->m_ext.ext_buf);
867 uma_zfree(zone_mbuf, mref);
868 break;
869 case EXT_JUMBOP:
870 uma_zfree(zone_jumbop, m->m_ext.ext_buf);
871 uma_zfree(zone_mbuf, mref);
872 break;
873 case EXT_JUMBO9:
874 uma_zfree(zone_jumbo9, m->m_ext.ext_buf);
875 uma_zfree(zone_mbuf, mref);
876 break;
877 case EXT_JUMBO16:
878 uma_zfree(zone_jumbo16, m->m_ext.ext_buf);
879 uma_zfree(zone_mbuf, mref);
880 break;
881 case EXT_SFBUF:
882 case EXT_NET_DRV:
883 case EXT_MOD_TYPE:
884 case EXT_DISPOSABLE:
885 KASSERT(mref->m_ext.ext_free != NULL,
886 ("%s: ext_free not set", __func__));
887 mref->m_ext.ext_free(mref);
888 uma_zfree(zone_mbuf, mref);
889 break;
890 case EXT_EXTREF:
891 KASSERT(m->m_ext.ext_free != NULL,
892 ("%s: ext_free not set", __func__));
893 m->m_ext.ext_free(m);
894 break;
895 default:
896 KASSERT(m->m_ext.ext_type == 0,
897 ("%s: unknown ext_type", __func__));
898 }
899 }
900
901 if (freembuf && m != mref)
902 uma_zfree(zone_mbuf, m);
903 }
904
905 /*
906 * Official mbuf(9) allocation KPI for stack and drivers:
907 *
908 * m_get() - a single mbuf without any attachments, sys/mbuf.h.
909 * m_gethdr() - a single mbuf initialized as M_PKTHDR, sys/mbuf.h.
910 * m_getcl() - an mbuf + 2k cluster, sys/mbuf.h.
911 * m_clget() - attach cluster to already allocated mbuf.
912 * m_cljget() - attach jumbo cluster to already allocated mbuf.
913 * m_get2() - allocate minimum mbuf that would fit size argument.
914 * m_getm2() - allocate a chain of mbufs/clusters.
915 * m_extadd() - attach external cluster to mbuf.
916 *
917 * m_free() - free single mbuf with its tags and ext, sys/mbuf.h.
918 * m_freem() - free chain of mbufs.
919 */
920
921 int
m_clget(struct mbuf * m,int how)922 m_clget(struct mbuf *m, int how)
923 {
924
925 KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
926 __func__, m));
927 m->m_ext.ext_buf = (char *)NULL;
928 uma_zalloc_arg(zone_clust, m, how);
929 /*
930 * On a cluster allocation failure, drain the packet zone and retry,
931 * we might be able to loosen a few clusters up on the drain.
932 */
933 if ((how & M_NOWAIT) && (m->m_ext.ext_buf == NULL)) {
934 zone_drain(zone_pack);
935 uma_zalloc_arg(zone_clust, m, how);
936 }
937 MBUF_PROBE2(m__clget, m, how);
938 return (m->m_flags & M_EXT);
939 }
940
941 /*
942 * m_cljget() is different from m_clget() as it can allocate clusters without
943 * attaching them to an mbuf. In that case the return value is the pointer
944 * to the cluster of the requested size. If an mbuf was specified, it gets
945 * the cluster attached to it and the return value can be safely ignored.
946 * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
947 */
948 void *
m_cljget(struct mbuf * m,int how,int size)949 m_cljget(struct mbuf *m, int how, int size)
950 {
951 uma_zone_t zone;
952 void *retval;
953
954 if (m != NULL) {
955 KASSERT((m->m_flags & M_EXT) == 0, ("%s: mbuf %p has M_EXT",
956 __func__, m));
957 m->m_ext.ext_buf = NULL;
958 }
959
960 zone = m_getzone(size);
961 retval = uma_zalloc_arg(zone, m, how);
962
963 MBUF_PROBE4(m__cljget, m, how, size, retval);
964
965 return (retval);
966 }
967
968 /*
969 * m_get2() allocates minimum mbuf that would fit "size" argument.
970 */
971 struct mbuf *
m_get2(int size,int how,short type,int flags)972 m_get2(int size, int how, short type, int flags)
973 {
974 struct mb_args args;
975 struct mbuf *m, *n;
976
977 args.flags = flags;
978 args.type = type;
979
980 if (size <= MHLEN || (size <= MLEN && (flags & M_PKTHDR) == 0))
981 return (uma_zalloc_arg(zone_mbuf, &args, how));
982 if (size <= MCLBYTES)
983 return (uma_zalloc_arg(zone_pack, &args, how));
984
985 if (size > MJUMPAGESIZE)
986 return (NULL);
987
988 m = uma_zalloc_arg(zone_mbuf, &args, how);
989 if (m == NULL)
990 return (NULL);
991
992 n = uma_zalloc_arg(zone_jumbop, m, how);
993 if (n == NULL) {
994 uma_zfree(zone_mbuf, m);
995 return (NULL);
996 }
997
998 return (m);
999 }
1000
1001 /*
1002 * m_getjcl() returns an mbuf with a cluster of the specified size attached.
1003 * For size it takes MCLBYTES, MJUMPAGESIZE, MJUM9BYTES, MJUM16BYTES.
1004 */
1005 struct mbuf *
m_getjcl(int how,short type,int flags,int size)1006 m_getjcl(int how, short type, int flags, int size)
1007 {
1008 struct mb_args args;
1009 struct mbuf *m, *n;
1010 uma_zone_t zone;
1011
1012 if (size == MCLBYTES)
1013 return m_getcl(how, type, flags);
1014
1015 args.flags = flags;
1016 args.type = type;
1017
1018 m = uma_zalloc_arg(zone_mbuf, &args, how);
1019 if (m == NULL)
1020 return (NULL);
1021
1022 zone = m_getzone(size);
1023 n = uma_zalloc_arg(zone, m, how);
1024 if (n == NULL) {
1025 uma_zfree(zone_mbuf, m);
1026 return (NULL);
1027 }
1028 MBUF_PROBE5(m__getjcl, how, type, flags, size, m);
1029 return (m);
1030 }
1031
1032 /*
1033 * Allocate a given length worth of mbufs and/or clusters (whatever fits
1034 * best) and return a pointer to the top of the allocated chain. If an
1035 * existing mbuf chain is provided, then we will append the new chain
1036 * to the existing one but still return the top of the newly allocated
1037 * chain.
1038 */
1039 struct mbuf *
m_getm2(struct mbuf * m,int len,int how,short type,int flags)1040 m_getm2(struct mbuf *m, int len, int how, short type, int flags)
1041 {
1042 struct mbuf *mb, *nm = NULL, *mtail = NULL;
1043
1044 KASSERT(len >= 0, ("%s: len is < 0", __func__));
1045
1046 /* Validate flags. */
1047 flags &= (M_PKTHDR | M_EOR);
1048
1049 /* Packet header mbuf must be first in chain. */
1050 if ((flags & M_PKTHDR) && m != NULL)
1051 flags &= ~M_PKTHDR;
1052
1053 /* Loop and append maximum sized mbufs to the chain tail. */
1054 while (len > 0) {
1055 if (len > MCLBYTES)
1056 mb = m_getjcl(how, type, (flags & M_PKTHDR),
1057 MJUMPAGESIZE);
1058 else if (len >= MINCLSIZE)
1059 mb = m_getcl(how, type, (flags & M_PKTHDR));
1060 else if (flags & M_PKTHDR)
1061 mb = m_gethdr(how, type);
1062 else
1063 mb = m_get(how, type);
1064
1065 /* Fail the whole operation if one mbuf can't be allocated. */
1066 if (mb == NULL) {
1067 if (nm != NULL)
1068 m_freem(nm);
1069 return (NULL);
1070 }
1071
1072 /* Book keeping. */
1073 len -= M_SIZE(mb);
1074 if (mtail != NULL)
1075 mtail->m_next = mb;
1076 else
1077 nm = mb;
1078 mtail = mb;
1079 flags &= ~M_PKTHDR; /* Only valid on the first mbuf. */
1080 }
1081 if (flags & M_EOR)
1082 mtail->m_flags |= M_EOR; /* Only valid on the last mbuf. */
1083
1084 /* If mbuf was supplied, append new chain to the end of it. */
1085 if (m != NULL) {
1086 for (mtail = m; mtail->m_next != NULL; mtail = mtail->m_next)
1087 ;
1088 mtail->m_next = nm;
1089 mtail->m_flags &= ~M_EOR;
1090 } else
1091 m = nm;
1092
1093 return (m);
1094 }
1095
1096 /*-
1097 * Configure a provided mbuf to refer to the provided external storage
1098 * buffer and setup a reference count for said buffer.
1099 *
1100 * Arguments:
1101 * mb The existing mbuf to which to attach the provided buffer.
1102 * buf The address of the provided external storage buffer.
1103 * size The size of the provided buffer.
1104 * freef A pointer to a routine that is responsible for freeing the
1105 * provided external storage buffer.
1106 * args A pointer to an argument structure (of any type) to be passed
1107 * to the provided freef routine (may be NULL).
1108 * flags Any other flags to be passed to the provided mbuf.
1109 * type The type that the external storage buffer should be
1110 * labeled with.
1111 *
1112 * Returns:
1113 * Nothing.
1114 */
1115 void
m_extadd(struct mbuf * mb,char * buf,u_int size,m_ext_free_t freef,void * arg1,void * arg2,int flags,int type)1116 m_extadd(struct mbuf *mb, char *buf, u_int size, m_ext_free_t freef,
1117 void *arg1, void *arg2, int flags, int type)
1118 {
1119
1120 KASSERT(type != EXT_CLUSTER, ("%s: EXT_CLUSTER not allowed", __func__));
1121
1122 mb->m_flags |= (M_EXT | flags);
1123 mb->m_ext.ext_buf = buf;
1124 mb->m_data = mb->m_ext.ext_buf;
1125 mb->m_ext.ext_size = size;
1126 mb->m_ext.ext_free = freef;
1127 mb->m_ext.ext_arg1 = arg1;
1128 mb->m_ext.ext_arg2 = arg2;
1129 mb->m_ext.ext_type = type;
1130
1131 if (type != EXT_EXTREF) {
1132 mb->m_ext.ext_count = 1;
1133 mb->m_ext.ext_flags = EXT_FLAG_EMBREF;
1134 } else
1135 mb->m_ext.ext_flags = 0;
1136 }
1137
1138 /*
1139 * Free an entire chain of mbufs and associated external buffers, if
1140 * applicable.
1141 */
1142 void
m_freem(struct mbuf * mb)1143 m_freem(struct mbuf *mb)
1144 {
1145
1146 MBUF_PROBE1(m__freem, mb);
1147 while (mb != NULL)
1148 mb = m_free(mb);
1149 }
1150