1 /*
2 * KERN_SLABALLOC.C - Kernel SLAB memory allocator
3 *
4 * Copyright (c) 2003,2004,2010-2019 The DragonFly Project.
5 * All rights reserved.
6 *
7 * This code is derived from software contributed to The DragonFly Project
8 * by Matthew Dillon <dillon@backplane.com>
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 *
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in
18 * the documentation and/or other materials provided with the
19 * distribution.
20 * 3. Neither the name of The DragonFly Project nor the names of its
21 * contributors may be used to endorse or promote products derived
22 * from this software without specific, prior written permission.
23 *
24 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
25 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
26 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
27 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
28 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
29 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
30 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
31 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
32 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
33 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
34 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * SUCH DAMAGE.
36 *
37 * This module implements a slab allocator drop-in replacement for the
38 * kernel malloc().
39 *
40 * A slab allocator reserves a ZONE for each chunk size, then lays the
41 * chunks out in an array within the zone. Allocation and deallocation
42 * is nearly instantanious, and fragmentation/overhead losses are limited
43 * to a fixed worst-case amount.
44 *
45 * The downside of this slab implementation is in the chunk size
46 * multiplied by the number of zones. ~80 zones * 128K = 10MB of VM per cpu.
47 * In a kernel implementation all this memory will be physical so
48 * the zone size is adjusted downward on machines with less physical
49 * memory. The upside is that overhead is bounded... this is the *worst*
50 * case overhead.
51 *
52 * Slab management is done on a per-cpu basis and no locking or mutexes
53 * are required, only a critical section. When one cpu frees memory
54 * belonging to another cpu's slab manager an asynchronous IPI message
55 * will be queued to execute the operation. In addition, both the
56 * high level slab allocator and the low level zone allocator optimize
57 * M_ZERO requests, and the slab allocator does not have to pre initialize
58 * the linked list of chunks.
59 *
60 * XXX Balancing is needed between cpus. Balance will be handled through
61 * asynchronous IPIs primarily by reassigning the z_Cpu ownership of chunks.
62 *
63 * XXX If we have to allocate a new zone and M_USE_RESERVE is set, use of
64 * the new zone should be restricted to M_USE_RESERVE requests only.
65 *
66 * Alloc Size Chunking Number of zones
67 * 0-127 8 16
68 * 128-255 16 8
69 * 256-511 32 8
70 * 512-1023 64 8
71 * 1024-2047 128 8
72 * 2048-4095 256 8
73 * 4096-8191 512 8
74 * 8192-16383 1024 8
75 * 16384-32767 2048 8
76 * (if PAGE_SIZE is 4K the maximum zone allocation is 16383)
77 *
78 * Allocations >= ZoneLimit go directly to kmem.
79 * (n * PAGE_SIZE, n > 2) allocations go directly to kmem.
80 *
81 * Alignment properties:
82 * - All power-of-2 sized allocations are power-of-2 aligned.
83 * - Allocations with M_POWEROF2 are power-of-2 aligned on the nearest
84 * power-of-2 round up of 'size'.
85 * - Non-power-of-2 sized allocations are zone chunk size aligned (see the
86 * above table 'Chunking' column).
87 *
88 * API REQUIREMENTS AND SIDE EFFECTS
89 *
90 * To operate as a drop-in replacement to the FreeBSD-4.x malloc() we
91 * have remained compatible with the following API requirements:
92 *
93 * + malloc(0) is allowed and returns non-NULL (ahc driver)
94 * + ability to allocate arbitrarily large chunks of memory
95 */
96
97 #include <sys/param.h>
98 #include <sys/systm.h>
99 #include <sys/kernel.h>
100 #include <sys/slaballoc.h>
101 #include <sys/mbuf.h>
102 #include <sys/vmmeter.h>
103 #include <sys/lock.h>
104 #include <sys/thread.h>
105 #include <sys/globaldata.h>
106 #include <sys/sysctl.h>
107 #include <sys/ktr.h>
108 #include <sys/kthread.h>
109 #include <sys/malloc.h>
110
111 #include <vm/vm.h>
112 #include <vm/vm_param.h>
113 #include <vm/vm_kern.h>
114 #include <vm/vm_extern.h>
115 #include <vm/vm_object.h>
116 #include <vm/pmap.h>
117 #include <vm/vm_map.h>
118 #include <vm/vm_page.h>
119 #include <vm/vm_pageout.h>
120
121 #include <machine/cpu.h>
122
123 #include <sys/thread2.h>
124 #include <vm/vm_page2.h>
125
126 #if (__VM_CACHELINE_SIZE == 32)
127 #define CAN_CACHEALIGN(sz) ((sz) >= 256)
128 #elif (__VM_CACHELINE_SIZE == 64)
129 #define CAN_CACHEALIGN(sz) ((sz) >= 512)
130 #elif (__VM_CACHELINE_SIZE == 128)
131 #define CAN_CACHEALIGN(sz) ((sz) >= 1024)
132 #else
133 #error "unsupported cacheline size"
134 #endif
135
136 #define btokup(z) (&pmap_kvtom((vm_offset_t)(z))->ku_pagecnt)
137
138 #define MEMORY_STRING "ptr=%p type=%p size=%lu flags=%04x"
139 #define MEMORY_ARGS void *ptr, void *type, unsigned long size, int flags
140
141 #if !defined(KTR_MEMORY)
142 #define KTR_MEMORY KTR_ALL
143 #endif
144 KTR_INFO_MASTER(memory);
145 KTR_INFO(KTR_MEMORY, memory, malloc_beg, 0, "malloc begin");
146 KTR_INFO(KTR_MEMORY, memory, malloc_end, 1, MEMORY_STRING, MEMORY_ARGS);
147 KTR_INFO(KTR_MEMORY, memory, free_zero, 2, MEMORY_STRING, MEMORY_ARGS);
148 KTR_INFO(KTR_MEMORY, memory, free_ovsz, 3, MEMORY_STRING, MEMORY_ARGS);
149 KTR_INFO(KTR_MEMORY, memory, free_ovsz_delayed, 4, MEMORY_STRING, MEMORY_ARGS);
150 KTR_INFO(KTR_MEMORY, memory, free_chunk, 5, MEMORY_STRING, MEMORY_ARGS);
151 KTR_INFO(KTR_MEMORY, memory, free_request, 6, MEMORY_STRING, MEMORY_ARGS);
152 KTR_INFO(KTR_MEMORY, memory, free_rem_beg, 7, MEMORY_STRING, MEMORY_ARGS);
153 KTR_INFO(KTR_MEMORY, memory, free_rem_end, 8, MEMORY_STRING, MEMORY_ARGS);
154 KTR_INFO(KTR_MEMORY, memory, free_beg, 9, "free begin");
155 KTR_INFO(KTR_MEMORY, memory, free_end, 10, "free end");
156
157 #define logmemory(name, ptr, type, size, flags) \
158 KTR_LOG(memory_ ## name, ptr, type, size, flags)
159 #define logmemory_quick(name) \
160 KTR_LOG(memory_ ## name)
161
162 /*
163 * Fixed globals (not per-cpu)
164 */
165 __read_frequently static int ZoneSize;
166 __read_frequently static int ZoneLimit;
167 __read_frequently static int ZonePageCount;
168 __read_frequently static uintptr_t ZoneMask;
169 __read_frequently struct malloc_type *kmemstatistics; /* exported to vmstat */
170
171 #if defined(INVARIANTS)
172 static void chunk_mark_allocated(SLZone *z, void *chunk);
173 static void chunk_mark_free(SLZone *z, void *chunk);
174 #else
175 #define chunk_mark_allocated(z, chunk)
176 #define chunk_mark_free(z, chunk)
177 #endif
178
179 /*
180 * Misc constants. Note that allocations that are exact multiples of
181 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
182 */
183 #define ZONE_RELS_THRESH 32 /* threshold number of zones */
184
185 #ifdef INVARIANTS
186 /*
187 * The WEIRD_ADDR is used as known text to copy into free objects to
188 * try to create deterministic failure cases if the data is accessed after
189 * free.
190 */
191 #define WEIRD_ADDR 0xdeadc0de
192 #endif
193 #define ZERO_LENGTH_PTR ((void *)-8)
194
195 /*
196 * Misc global malloc buckets
197 */
198
199 MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
200 MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
201 MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
202 MALLOC_DEFINE(M_DRM, "m_drm", "DRM memory allocations");
203
204 MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
205 MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
206
207 /*
208 * Initialize the slab memory allocator. We have to choose a zone size based
209 * on available physical memory. We choose a zone side which is approximately
210 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
211 * 128K. The zone size is limited to the bounds set in slaballoc.h
212 * (typically 32K min, 128K max).
213 */
214 static void kmeminit(void *dummy);
215 static void kmemfinishinit(void *dummy);
216
217 char *ZeroPage;
218
219 SYSINIT(kmem1, SI_BOOT1_ALLOCATOR, SI_ORDER_FIRST, kmeminit, NULL);
220 SYSINIT(kmem2, SI_BOOT2_POST_SMP, SI_ORDER_FIRST, kmemfinishinit, NULL);
221
222 #ifdef INVARIANTS
223 /*
224 * If enabled any memory allocated without M_ZERO is initialized to -1.
225 */
226 __read_frequently static int use_malloc_pattern;
227 SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW,
228 &use_malloc_pattern, 0,
229 "Initialize memory to -1 if M_ZERO not specified");
230
231 __read_frequently static int32_t weirdary[16];
232 __read_frequently static int use_weird_array;
233 SYSCTL_INT(_debug, OID_AUTO, use_weird_array, CTLFLAG_RW,
234 &use_weird_array, 0,
235 "Initialize memory to weird values on kfree()");
236 #endif
237
238 __read_frequently static int ZoneRelsThresh = ZONE_RELS_THRESH;
239 SYSCTL_INT(_kern, OID_AUTO, zone_cache, CTLFLAG_RW, &ZoneRelsThresh, 0, "");
240 __read_frequently static int kzone_pollfreq = 1;
241 SYSCTL_INT(_kern, OID_AUTO, kzone_pollfreq, CTLFLAG_RW, &kzone_pollfreq, 0, "");
242
243 static struct spinlock kmemstat_spin =
244 SPINLOCK_INITIALIZER(&kmemstat_spin, "malinit");
245 static struct malloc_type *kmemstat_poll;
246
247 /*
248 * Returns the kernel memory size limit for the purposes of initializing
249 * various subsystem caches. The smaller of available memory and the KVM
250 * memory space is returned.
251 *
252 * The size in megabytes is returned.
253 */
254 size_t
kmem_lim_size(void)255 kmem_lim_size(void)
256 {
257 size_t limsize;
258
259 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
260 if (limsize > KvaSize)
261 limsize = KvaSize;
262 return (limsize / (1024 * 1024));
263 }
264
265 static void
kmeminit(void * dummy)266 kmeminit(void *dummy)
267 {
268 size_t limsize;
269 int usesize;
270 #ifdef INVARIANTS
271 int i;
272 #endif
273
274 limsize = kmem_lim_size();
275 usesize = (int)(limsize * 1024); /* convert to KB */
276
277 /*
278 * If the machine has a large KVM space and more than 8G of ram,
279 * double the zone release threshold to reduce SMP invalidations.
280 * If more than 16G of ram, do it again.
281 *
282 * The BIOS eats a little ram so add some slop. We want 8G worth of
283 * memory sticks to trigger the first adjustment.
284 */
285 if (ZoneRelsThresh == ZONE_RELS_THRESH) {
286 if (limsize >= 7 * 1024)
287 ZoneRelsThresh *= 2;
288 if (limsize >= 15 * 1024)
289 ZoneRelsThresh *= 2;
290 if (limsize >= 31 * 1024)
291 ZoneRelsThresh *= 2;
292 if (limsize >= 63 * 1024)
293 ZoneRelsThresh *= 2;
294 if (limsize >= 127 * 1024)
295 ZoneRelsThresh *= 2;
296 }
297
298 /*
299 * Calculate the zone size. This typically calculates to
300 * ZALLOC_MAX_ZONE_SIZE
301 */
302 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
303 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
304 ZoneSize <<= 1;
305 ZoneLimit = ZoneSize / 4;
306 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
307 ZoneLimit = ZALLOC_ZONE_LIMIT;
308 ZoneMask = ~(uintptr_t)(ZoneSize - 1);
309 ZonePageCount = ZoneSize / PAGE_SIZE;
310
311 #ifdef INVARIANTS
312 for (i = 0; i < NELEM(weirdary); ++i)
313 weirdary[i] = WEIRD_ADDR;
314 #endif
315
316 ZeroPage = kmem_slab_alloc(PAGE_SIZE, PAGE_SIZE, M_WAITOK|M_ZERO);
317
318 if (bootverbose)
319 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
320 }
321
322 /*
323 * Once we know how many cpus are configured reduce ZoneRelsThresh
324 * based on multiples of 32 cpu threads.
325 */
326 static void
kmemfinishinit(void * dummy)327 kmemfinishinit(void *dummy)
328 {
329 if (ncpus > 32)
330 ZoneRelsThresh = ZoneRelsThresh * 32 / ncpus;
331 }
332
333 /*
334 * (low level) Initialize slab-related elements in the globaldata structure.
335 *
336 * Occurs after kmeminit().
337 */
338 void
slab_gdinit(globaldata_t gd)339 slab_gdinit(globaldata_t gd)
340 {
341 SLGlobalData *slgd;
342 int i;
343
344 slgd = &gd->gd_slab;
345 for (i = 0; i < NZONES; ++i)
346 TAILQ_INIT(&slgd->ZoneAry[i]);
347 TAILQ_INIT(&slgd->FreeZones);
348 TAILQ_INIT(&slgd->FreeOvZones);
349 }
350
351 /*
352 * Initialize a malloc type tracking structure.
353 */
354 void
malloc_init(void * data)355 malloc_init(void *data)
356 {
357 struct malloc_type *type = data;
358 struct kmalloc_use *use;
359 size_t limsize;
360 int n;
361
362 if (type->ks_magic != M_MAGIC)
363 panic("malloc type lacks magic");
364
365 if (type->ks_limit != 0)
366 return;
367
368 if (vmstats.v_page_count == 0)
369 panic("malloc_init not allowed before vm init");
370
371 limsize = kmem_lim_size() * (1024 * 1024);
372 type->ks_limit = limsize / 10;
373 if (type->ks_flags & KSF_OBJSIZE)
374 malloc_mgt_init(type, &type->ks_mgt, type->ks_objsize);
375
376 if (ncpus == 1)
377 use = &type->ks_use0;
378 else
379 use = kmalloc(ncpus * sizeof(*use), M_TEMP, M_WAITOK | M_ZERO);
380 if (type->ks_flags & KSF_OBJSIZE) {
381 for (n = 0; n < ncpus; ++n)
382 malloc_mgt_init(type, &use[n].mgt, type->ks_objsize);
383 }
384
385 spin_lock(&kmemstat_spin);
386 type->ks_next = kmemstatistics;
387 type->ks_use = use;
388 kmemstatistics = type;
389 spin_unlock(&kmemstat_spin);
390 }
391
392 void
malloc_uninit(void * data)393 malloc_uninit(void *data)
394 {
395 struct malloc_type *type = data;
396 struct malloc_type *t;
397 int i;
398 #ifdef INVARIANTS
399 long ttl;
400 #endif
401
402 if (type->ks_magic != M_MAGIC)
403 panic("malloc type lacks magic");
404
405 if (vmstats.v_page_count == 0)
406 panic("malloc_uninit not allowed before vm init");
407
408 if (type->ks_limit == 0)
409 panic("malloc_uninit on uninitialized type");
410
411 /* Make sure that all pending kfree()s are finished. */
412 lwkt_synchronize_ipiqs("muninit");
413
414 /*
415 * Remove from the kmemstatistics list, blocking if the removal races
416 * the kmalloc poller.
417 *
418 * Advance kmemstat_poll if necessary.
419 */
420 spin_lock(&kmemstat_spin);
421 while (type->ks_flags & KSF_POLLING)
422 ssleep(type, &kmemstat_spin, 0, "kmuninit", 0);
423
424 if (kmemstat_poll == type)
425 kmemstat_poll = type->ks_next;
426
427 if (kmemstatistics == type) {
428 kmemstatistics = type->ks_next;
429 } else {
430 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
431 if (t->ks_next == type) {
432 t->ks_next = type->ks_next;
433 break;
434 }
435 }
436 }
437 type->ks_next = NULL;
438 type->ks_limit = 0;
439 spin_unlock(&kmemstat_spin);
440
441 /*
442 * memuse is only correct in aggregation. Due to memory being allocated
443 * on one cpu and freed on another individual array entries may be
444 * negative or positive (canceling each other out).
445 */
446 #ifdef INVARIANTS
447 ttl = 0;
448 #endif
449 for (i = 0; i < ncpus; ++i) {
450 #ifdef INVARIANTS
451 ttl += type->ks_use[i].memuse;
452 #endif
453 if (type->ks_flags & KSF_OBJSIZE)
454 malloc_mgt_uninit(type, &type->ks_use[i].mgt);
455 }
456 if (type->ks_flags & KSF_OBJSIZE)
457 malloc_mgt_uninit(type, &type->ks_mgt);
458 #ifdef INVARIANTS
459 if (ttl) {
460 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
461 ttl, type->ks_shortdesc, i);
462 }
463 #endif
464
465 if (type->ks_use != &type->ks_use0) {
466 kfree(type->ks_use, M_TEMP);
467 type->ks_use = NULL;
468 }
469 }
470
471 /*
472 * Slowly polls all kmalloc zones for cleanup
473 */
474 static void
kmalloc_poller_thread(void)475 kmalloc_poller_thread(void)
476 {
477 struct malloc_type *type;
478
479 for (;;) {
480 /*
481 * Very slow poll by default, adjustable with sysctl
482 */
483 int sticks;
484
485 sticks = kzone_pollfreq;
486 cpu_ccfence();
487 if (sticks > 0)
488 sticks = hz / sticks + 1; /* approximate */
489 else
490 sticks = hz; /* safety */
491 tsleep((caddr_t)&sticks, 0, "kmslp", sticks);
492
493 /*
494 * [re]poll one zone each period.
495 */
496 spin_lock(&kmemstat_spin);
497 type = kmemstat_poll;
498
499 if (type == NULL)
500 type = kmemstatistics;
501 if (type) {
502 atomic_set_int(&type->ks_flags, KSF_POLLING);
503 spin_unlock(&kmemstat_spin);
504 if (malloc_mgt_poll(type)) {
505 spin_lock(&kmemstat_spin);
506 kmemstat_poll = type->ks_next;
507 } else {
508 spin_lock(&kmemstat_spin);
509 }
510 atomic_clear_int(&type->ks_flags, KSF_POLLING);
511 wakeup(type);
512 } else {
513 kmemstat_poll = NULL;
514 }
515 spin_unlock(&kmemstat_spin);
516 }
517 }
518
519 static struct thread *kmalloc_poller_td;
520 static struct kproc_desc kmalloc_poller_kp = {
521 "kmalloc_poller",
522 kmalloc_poller_thread,
523 &kmalloc_poller_td
524 };
525 SYSINIT(kmalloc_polller, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST,
526 kproc_start, &kmalloc_poller_kp);
527
528 /*
529 * Reinitialize all installed malloc regions after ncpus has been
530 * determined. type->ks_use0 is initially set to &type->ks_use0,
531 * this function will dynamically allocate it as appropriate for ncpus.
532 */
533 void
malloc_reinit_ncpus(void)534 malloc_reinit_ncpus(void)
535 {
536 struct malloc_type *t;
537 struct kmalloc_use *use;
538 int n;
539
540 /*
541 * If only one cpu we can leave ks_use set to ks_use0
542 */
543 if (ncpus <= 1)
544 return;
545
546 /*
547 * Expand ks_use for all kmalloc blocks
548 */
549 for (t = kmemstatistics; t; t = t->ks_next) {
550 KKASSERT(t->ks_use == &t->ks_use0);
551 t->ks_use = kmalloc(sizeof(*use) * ncpus, M_TEMP, M_WAITOK|M_ZERO);
552 t->ks_use[0] = t->ks_use0;
553 if (t->ks_flags & KSF_OBJSIZE) {
554 malloc_mgt_relocate(&t->ks_use0.mgt, &t->ks_use[0].mgt);
555 for (n = 1; n < ncpus; ++n)
556 malloc_mgt_init(t, &t->ks_use[n].mgt, t->ks_objsize);
557 }
558 }
559 }
560
561 /*
562 * Increase the kmalloc pool limit for the specified pool. No changes
563 * are the made if the pool would shrink.
564 */
565 void
kmalloc_raise_limit(struct malloc_type * type,size_t bytes)566 kmalloc_raise_limit(struct malloc_type *type, size_t bytes)
567 {
568 KKASSERT(type->ks_limit != 0);
569 if (bytes == 0)
570 bytes = KvaSize;
571 if (type->ks_limit < bytes)
572 type->ks_limit = bytes;
573 }
574
575 void
kmalloc_set_unlimited(struct malloc_type * type)576 kmalloc_set_unlimited(struct malloc_type *type)
577 {
578 type->ks_limit = kmem_lim_size() * (1024 * 1024);
579 }
580
581 /*
582 * Dynamically create a malloc pool. This function is a NOP if *typep is
583 * already non-NULL.
584 */
585 void
kmalloc_create(struct malloc_type ** typep,const char * descr)586 kmalloc_create(struct malloc_type **typep, const char *descr)
587 {
588 struct malloc_type *type;
589
590 if (*typep == NULL) {
591 type = kmalloc(sizeof(*type), M_TEMP, M_WAITOK | M_ZERO);
592 type->ks_magic = M_MAGIC;
593 type->ks_shortdesc = descr;
594 malloc_init(type);
595 *typep = type;
596 }
597 }
598
599 void
_kmalloc_create_obj(struct malloc_type ** typep,const char * descr,size_t objsize)600 _kmalloc_create_obj(struct malloc_type **typep, const char *descr,
601 size_t objsize)
602 {
603 struct malloc_type *type;
604
605 if (*typep == NULL) {
606 type = kmalloc(sizeof(*type), M_TEMP, M_WAITOK | M_ZERO);
607 type->ks_magic = M_MAGIC;
608 type->ks_shortdesc = descr;
609 type->ks_flags = KSF_OBJSIZE;
610 type->ks_objsize = __VM_CACHELINE_ALIGN(objsize);
611 malloc_init(type);
612 *typep = type;
613 }
614 }
615
616 /*
617 * Destroy a dynamically created malloc pool. This function is a NOP if
618 * the pool has already been destroyed.
619 *
620 * WARNING! For kmalloc_obj's, the exis state for related slabs is ignored,
621 * only call once all references are 100% known to be gone.
622 */
623 void
kmalloc_destroy(struct malloc_type ** typep)624 kmalloc_destroy(struct malloc_type **typep)
625 {
626 if (*typep != NULL) {
627 malloc_uninit(*typep);
628 kfree(*typep, M_TEMP);
629 *typep = NULL;
630 }
631 }
632
633 /*
634 * Calculate the zone index for the allocation request size and set the
635 * allocation request size to that particular zone's chunk size.
636 */
637 static __inline int
zoneindex(unsigned long * bytes,unsigned long * align)638 zoneindex(unsigned long *bytes, unsigned long *align)
639 {
640 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
641
642 if (n < 128) {
643 *bytes = n = (n + 7) & ~7;
644 *align = 8;
645 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
646 }
647 if (n < 256) {
648 *bytes = n = (n + 15) & ~15;
649 *align = 16;
650 return(n / 16 + 7);
651 }
652 if (n < 8192) {
653 if (n < 512) {
654 *bytes = n = (n + 31) & ~31;
655 *align = 32;
656 return(n / 32 + 15);
657 }
658 if (n < 1024) {
659 *bytes = n = (n + 63) & ~63;
660 *align = 64;
661 return(n / 64 + 23);
662 }
663 if (n < 2048) {
664 *bytes = n = (n + 127) & ~127;
665 *align = 128;
666 return(n / 128 + 31);
667 }
668 if (n < 4096) {
669 *bytes = n = (n + 255) & ~255;
670 *align = 256;
671 return(n / 256 + 39);
672 }
673 *bytes = n = (n + 511) & ~511;
674 *align = 512;
675 return(n / 512 + 47);
676 }
677 #if ZALLOC_ZONE_LIMIT > 8192
678 if (n < 16384) {
679 *bytes = n = (n + 1023) & ~1023;
680 *align = 1024;
681 return(n / 1024 + 55);
682 }
683 #endif
684 #if ZALLOC_ZONE_LIMIT > 16384
685 if (n < 32768) {
686 *bytes = n = (n + 2047) & ~2047;
687 *align = 2048;
688 return(n / 2048 + 63);
689 }
690 #endif
691 panic("Unexpected byte count %d", n);
692 return(0);
693 }
694
695 static __inline void
clean_zone_rchunks(SLZone * z)696 clean_zone_rchunks(SLZone *z)
697 {
698 SLChunk *bchunk;
699
700 while ((bchunk = z->z_RChunks) != NULL) {
701 cpu_ccfence();
702 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
703 *z->z_LChunksp = bchunk;
704 while (bchunk) {
705 chunk_mark_free(z, bchunk);
706 z->z_LChunksp = &bchunk->c_Next;
707 bchunk = bchunk->c_Next;
708 ++z->z_NFree;
709 }
710 break;
711 }
712 /* retry */
713 }
714 }
715
716 /*
717 * If the zone becomes totally free and is not the only zone listed for a
718 * chunk size we move it to the FreeZones list. We always leave at least
719 * one zone per chunk size listed, even if it is freeable.
720 *
721 * Do not move the zone if there is an IPI in_flight (z_RCount != 0),
722 * otherwise MP races can result in our free_remote code accessing a
723 * destroyed zone. The remote end interlocks z_RCount with z_RChunks
724 * so one has to test both z_NFree and z_RCount.
725 *
726 * Since this code can be called from an IPI callback, do *NOT* try to mess
727 * with kernel_map here. Hysteresis will be performed at kmalloc() time.
728 */
729 static __inline SLZone *
check_zone_free(SLGlobalData * slgd,SLZone * z)730 check_zone_free(SLGlobalData *slgd, SLZone *z)
731 {
732 SLZone *znext;
733
734 znext = TAILQ_NEXT(z, z_Entry);
735 if (z->z_NFree == z->z_NMax && z->z_RCount == 0 &&
736 (TAILQ_FIRST(&slgd->ZoneAry[z->z_ZoneIndex]) != z || znext)) {
737 int *kup;
738
739 TAILQ_REMOVE(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
740
741 z->z_Magic = -1;
742 TAILQ_INSERT_HEAD(&slgd->FreeZones, z, z_Entry);
743 ++slgd->NFreeZones;
744 kup = btokup(z);
745 *kup = 0;
746 }
747 return znext;
748 }
749
750 #ifdef SLAB_DEBUG
751 /*
752 * Used to debug memory corruption issues. Record up to (typically 32)
753 * allocation sources for this zone (for a particular chunk size).
754 */
755
756 static void
slab_record_source(SLZone * z,const char * file,int line)757 slab_record_source(SLZone *z, const char *file, int line)
758 {
759 int i;
760 int b = line & (SLAB_DEBUG_ENTRIES - 1);
761
762 i = b;
763 do {
764 if (z->z_Sources[i].file == file && z->z_Sources[i].line == line)
765 return;
766 if (z->z_Sources[i].file == NULL)
767 break;
768 i = (i + 1) & (SLAB_DEBUG_ENTRIES - 1);
769 } while (i != b);
770 z->z_Sources[i].file = file;
771 z->z_Sources[i].line = line;
772 }
773
774 #endif
775
776 static __inline unsigned long
powerof2_size(unsigned long size)777 powerof2_size(unsigned long size)
778 {
779 int i;
780
781 if (size == 0 || powerof2(size))
782 return size;
783
784 i = flsl(size);
785 return (1UL << i);
786 }
787
788 /*
789 * kmalloc() (SLAB ALLOCATOR)
790 *
791 * Allocate memory via the slab allocator. If the request is too large,
792 * or if it page-aligned beyond a certain size, we fall back to the
793 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
794 * &SlabMisc if you don't care.
795 *
796 * M_RNOWAIT - don't block.
797 * M_NULLOK - return NULL instead of blocking.
798 * M_ZERO - zero the returned memory.
799 * M_USE_RESERVE - allow greater drawdown of the free list
800 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
801 * M_POWEROF2 - roundup size to the nearest power of 2
802 *
803 * MPSAFE
804 */
805
806 /* don't let kmalloc macro mess up function declaration */
807 #undef kmalloc
808
809 #ifdef SLAB_DEBUG
810 void *
_kmalloc_debug(unsigned long size,struct malloc_type * type,int flags,const char * file,int line)811 _kmalloc_debug(unsigned long size, struct malloc_type *type, int flags,
812 const char *file, int line)
813 #else
814 void *
815 _kmalloc(unsigned long size, struct malloc_type *type, int flags)
816 #endif
817 {
818 SLZone *z;
819 SLChunk *chunk;
820 SLGlobalData *slgd;
821 struct globaldata *gd;
822 unsigned long align;
823 int zi;
824 #ifdef INVARIANTS
825 int i;
826 #endif
827
828 logmemory_quick(malloc_beg);
829 gd = mycpu;
830 slgd = &gd->gd_slab;
831
832 /*
833 * XXX silly to have this in the critical path.
834 */
835 KKASSERT(type->ks_limit != 0);
836 ++type->ks_use[gd->gd_cpuid].calls;
837
838 /*
839 * Flagged for cache-alignment
840 */
841 if (flags & M_CACHEALIGN) {
842 if (size < __VM_CACHELINE_SIZE)
843 size = __VM_CACHELINE_SIZE;
844 else if (!CAN_CACHEALIGN(size))
845 flags |= M_POWEROF2;
846 }
847
848 /*
849 * Flagged to force nearest power-of-2 (higher or same)
850 */
851 if (flags & M_POWEROF2)
852 size = powerof2_size(size);
853
854 /*
855 * Handle the case where the limit is reached. Panic if we can't return
856 * NULL. The original malloc code looped, but this tended to
857 * simply deadlock the computer.
858 *
859 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
860 * to determine if a more complete limit check should be done. The
861 * actual memory use is tracked via ks_use[cpu].memuse.
862 */
863 while (type->ks_loosememuse >= type->ks_limit) {
864 int i;
865 long ttl;
866
867 for (i = ttl = 0; i < ncpus; ++i)
868 ttl += type->ks_use[i].memuse;
869 type->ks_loosememuse = ttl; /* not MP synchronized */
870 if ((ssize_t)ttl < 0) /* deal with occassional race */
871 ttl = 0;
872 if (ttl >= type->ks_limit) {
873 if (flags & M_NULLOK) {
874 logmemory(malloc_end, NULL, type, size, flags);
875 return(NULL);
876 }
877 panic("%s: malloc limit exceeded", type->ks_shortdesc);
878 }
879 }
880
881 /*
882 * Handle the degenerate size == 0 case. Yes, this does happen.
883 * Return a special pointer. This is to maintain compatibility with
884 * the original malloc implementation. Certain devices, such as the
885 * adaptec driver, not only allocate 0 bytes, they check for NULL and
886 * also realloc() later on. Joy.
887 */
888 if (size == 0) {
889 logmemory(malloc_end, ZERO_LENGTH_PTR, type, size, flags);
890 return(ZERO_LENGTH_PTR);
891 }
892
893 /*
894 * Handle hysteresis from prior frees here in malloc(). We cannot
895 * safely manipulate the kernel_map in free() due to free() possibly
896 * being called via an IPI message or from sensitive interrupt code.
897 *
898 * NOTE: ku_pagecnt must be cleared before we free the slab or we
899 * might race another cpu allocating the kva and setting
900 * ku_pagecnt.
901 */
902 while (slgd->NFreeZones > ZoneRelsThresh && (flags & M_RNOWAIT) == 0) {
903 crit_enter();
904 if (slgd->NFreeZones > ZoneRelsThresh) { /* crit sect race */
905 int *kup;
906
907 z = TAILQ_LAST(&slgd->FreeZones, SLZoneList);
908 KKASSERT(z != NULL);
909 TAILQ_REMOVE(&slgd->FreeZones, z, z_Entry);
910 --slgd->NFreeZones;
911 kup = btokup(z);
912 *kup = 0;
913 kmem_slab_free(z, ZoneSize); /* may block */
914 }
915 crit_exit();
916 }
917
918 /*
919 * XXX handle oversized frees that were queued from kfree().
920 */
921 while (TAILQ_FIRST(&slgd->FreeOvZones) && (flags & M_RNOWAIT) == 0) {
922 crit_enter();
923 if ((z = TAILQ_LAST(&slgd->FreeOvZones, SLZoneList)) != NULL) {
924 vm_size_t tsize;
925
926 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
927 TAILQ_REMOVE(&slgd->FreeOvZones, z, z_Entry);
928 tsize = z->z_ChunkSize;
929 kmem_slab_free(z, tsize); /* may block */
930 }
931 crit_exit();
932 }
933
934 /*
935 * Handle large allocations directly. There should not be very many of
936 * these so performance is not a big issue.
937 *
938 * The backend allocator is pretty nasty on a SMP system. Use the
939 * slab allocator for one and two page-sized chunks even though we lose
940 * some efficiency. XXX maybe fix mmio and the elf loader instead.
941 */
942 if (size >= ZoneLimit || ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) {
943 int *kup;
944
945 size = round_page(size);
946 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
947 if (chunk == NULL) {
948 logmemory(malloc_end, NULL, type, size, flags);
949 return(NULL);
950 }
951 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
952 flags |= M_PASSIVE_ZERO;
953 kup = btokup(chunk);
954 *kup = size / PAGE_SIZE;
955 crit_enter();
956 goto done;
957 }
958
959 /*
960 * Attempt to allocate out of an existing zone. First try the free list,
961 * then allocate out of unallocated space. If we find a good zone move
962 * it to the head of the list so later allocations find it quickly
963 * (we might have thousands of zones in the list).
964 *
965 * Note: zoneindex() will panic of size is too large.
966 */
967 zi = zoneindex(&size, &align);
968 KKASSERT(zi < NZONES);
969 crit_enter();
970
971 if ((z = TAILQ_LAST(&slgd->ZoneAry[zi], SLZoneList)) != NULL) {
972 /*
973 * Locate a chunk - we have to have at least one. If this is the
974 * last chunk go ahead and do the work to retrieve chunks freed
975 * from remote cpus, and if the zone is still empty move it off
976 * the ZoneAry.
977 */
978 if (--z->z_NFree <= 0) {
979 KKASSERT(z->z_NFree == 0);
980
981 /*
982 * WARNING! This code competes with other cpus. It is ok
983 * for us to not drain RChunks here but we might as well, and
984 * it is ok if more accumulate after we're done.
985 *
986 * Set RSignal before pulling rchunks off, indicating that we
987 * will be moving ourselves off of the ZoneAry. Remote ends will
988 * read RSignal before putting rchunks on thus interlocking
989 * their IPI signaling.
990 */
991 if (z->z_RChunks == NULL)
992 atomic_swap_int(&z->z_RSignal, 1);
993
994 clean_zone_rchunks(z);
995
996 /*
997 * Remove from the zone list if no free chunks remain.
998 * Clear RSignal
999 */
1000 if (z->z_NFree == 0) {
1001 TAILQ_REMOVE(&slgd->ZoneAry[zi], z, z_Entry);
1002 } else {
1003 z->z_RSignal = 0;
1004 }
1005 }
1006
1007 /*
1008 * Fast path, we have chunks available in z_LChunks.
1009 */
1010 chunk = z->z_LChunks;
1011 if (chunk) {
1012 chunk_mark_allocated(z, chunk);
1013 z->z_LChunks = chunk->c_Next;
1014 if (z->z_LChunks == NULL)
1015 z->z_LChunksp = &z->z_LChunks;
1016 #ifdef SLAB_DEBUG
1017 slab_record_source(z, file, line);
1018 #endif
1019 goto done;
1020 }
1021
1022 /*
1023 * No chunks are available in LChunks, the free chunk MUST be
1024 * in the never-before-used memory area, controlled by UIndex.
1025 *
1026 * The consequences are very serious if our zone got corrupted so
1027 * we use an explicit panic rather than a KASSERT.
1028 */
1029 if (z->z_UIndex + 1 != z->z_NMax)
1030 ++z->z_UIndex;
1031 else
1032 z->z_UIndex = 0;
1033
1034 if (z->z_UIndex == z->z_UEndIndex)
1035 panic("slaballoc: corrupted zone");
1036
1037 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
1038 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
1039 flags &= ~M_ZERO;
1040 flags |= M_PASSIVE_ZERO;
1041 }
1042 chunk_mark_allocated(z, chunk);
1043 #ifdef SLAB_DEBUG
1044 slab_record_source(z, file, line);
1045 #endif
1046 goto done;
1047 }
1048
1049 /*
1050 * If all zones are exhausted we need to allocate a new zone for this
1051 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
1052 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
1053 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
1054 * we do not pre-zero it because we do not want to mess up the L1 cache.
1055 *
1056 * At least one subsystem, the tty code (see CROUND) expects power-of-2
1057 * allocations to be power-of-2 aligned. We maintain compatibility by
1058 * adjusting the base offset below.
1059 */
1060 {
1061 int off;
1062 int *kup;
1063
1064 if ((z = TAILQ_FIRST(&slgd->FreeZones)) != NULL) {
1065 TAILQ_REMOVE(&slgd->FreeZones, z, z_Entry);
1066 --slgd->NFreeZones;
1067 bzero(z, sizeof(SLZone));
1068 z->z_Flags |= SLZF_UNOTZEROD;
1069 } else {
1070 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
1071 if (z == NULL)
1072 goto fail;
1073 }
1074
1075 /*
1076 * How big is the base structure?
1077 */
1078 #if defined(INVARIANTS)
1079 /*
1080 * Make room for z_Bitmap. An exact calculation is somewhat more
1081 * complicated so don't make an exact calculation.
1082 */
1083 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]);
1084 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8);
1085 #else
1086 off = sizeof(SLZone);
1087 #endif
1088
1089 /*
1090 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
1091 * Otherwise properly align the data according to the chunk size.
1092 */
1093 if (powerof2(size))
1094 align = size;
1095 off = roundup2(off, align);
1096
1097 z->z_Magic = ZALLOC_SLAB_MAGIC;
1098 z->z_ZoneIndex = zi;
1099 z->z_NMax = (ZoneSize - off) / size;
1100 z->z_NFree = z->z_NMax - 1;
1101 z->z_BasePtr = (char *)z + off;
1102 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
1103 z->z_ChunkSize = size;
1104 z->z_CpuGd = gd;
1105 z->z_Cpu = gd->gd_cpuid;
1106 z->z_LChunksp = &z->z_LChunks;
1107 #ifdef SLAB_DEBUG
1108 bcopy(z->z_Sources, z->z_AltSources, sizeof(z->z_Sources));
1109 bzero(z->z_Sources, sizeof(z->z_Sources));
1110 #endif
1111 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
1112 TAILQ_INSERT_HEAD(&slgd->ZoneAry[zi], z, z_Entry);
1113 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
1114 flags &= ~M_ZERO; /* already zero'd */
1115 flags |= M_PASSIVE_ZERO;
1116 }
1117 kup = btokup(z);
1118 *kup = -(z->z_Cpu + 1); /* -1 to -(N+1) */
1119 chunk_mark_allocated(z, chunk);
1120 #ifdef SLAB_DEBUG
1121 slab_record_source(z, file, line);
1122 #endif
1123
1124 /*
1125 * Slide the base index for initial allocations out of the next
1126 * zone we create so we do not over-weight the lower part of the
1127 * cpu memory caches.
1128 */
1129 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
1130 & (ZALLOC_MAX_ZONE_SIZE - 1);
1131 }
1132
1133 done:
1134 ++type->ks_use[gd->gd_cpuid].inuse;
1135 type->ks_use[gd->gd_cpuid].memuse += size;
1136 type->ks_use[gd->gd_cpuid].loosememuse += size;
1137 if (type->ks_use[gd->gd_cpuid].loosememuse >= ZoneSize) {
1138 /* not MP synchronized */
1139 type->ks_loosememuse += type->ks_use[gd->gd_cpuid].loosememuse;
1140 type->ks_use[gd->gd_cpuid].loosememuse = 0;
1141 }
1142 crit_exit();
1143
1144 if (flags & M_ZERO)
1145 bzero(chunk, size);
1146 #ifdef INVARIANTS
1147 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
1148 if (use_malloc_pattern) {
1149 for (i = 0; i < size; i += sizeof(int)) {
1150 *(int *)((char *)chunk + i) = -1;
1151 }
1152 }
1153 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
1154 }
1155 #endif
1156 logmemory(malloc_end, chunk, type, size, flags);
1157 return(chunk);
1158 fail:
1159 crit_exit();
1160 logmemory(malloc_end, NULL, type, size, flags);
1161 return(NULL);
1162 }
1163
1164 /*
1165 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
1166 *
1167 * Generally speaking this routine is not called very often and we do
1168 * not attempt to optimize it beyond reusing the same pointer if the
1169 * new size fits within the chunking of the old pointer's zone.
1170 */
1171 #ifdef SLAB_DEBUG
1172 void *
krealloc_debug(void * ptr,unsigned long size,struct malloc_type * type,int flags,const char * file,int line)1173 krealloc_debug(void *ptr, unsigned long size,
1174 struct malloc_type *type, int flags,
1175 const char *file, int line)
1176 #else
1177 void *
1178 krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
1179 #endif
1180 {
1181 unsigned long osize;
1182 unsigned long align;
1183 SLZone *z;
1184 void *nptr;
1185 int *kup;
1186
1187 KKASSERT((flags & M_ZERO) == 0); /* not supported */
1188
1189 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
1190 return(_kmalloc_debug(size, type, flags, file, line));
1191 if (size == 0) {
1192 kfree(ptr, type);
1193 return(NULL);
1194 }
1195
1196 /*
1197 * Handle oversized allocations. XXX we really should require that a
1198 * size be passed to free() instead of this nonsense.
1199 */
1200 kup = btokup(ptr);
1201 if (*kup > 0) {
1202 osize = *kup << PAGE_SHIFT;
1203 if (osize == round_page(size))
1204 return(ptr);
1205 if ((nptr = _kmalloc_debug(size, type, flags, file, line)) == NULL)
1206 return(NULL);
1207 bcopy(ptr, nptr, min(size, osize));
1208 kfree(ptr, type);
1209 return(nptr);
1210 }
1211
1212 /*
1213 * Get the original allocation's zone. If the new request winds up
1214 * using the same chunk size we do not have to do anything.
1215 */
1216 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1217 kup = btokup(z);
1218 KKASSERT(*kup < 0);
1219 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1220
1221 /*
1222 * Allocate memory for the new request size. Note that zoneindex has
1223 * already adjusted the request size to the appropriate chunk size, which
1224 * should optimize our bcopy(). Then copy and return the new pointer.
1225 *
1226 * Resizing a non-power-of-2 allocation to a power-of-2 size does not
1227 * necessary align the result.
1228 *
1229 * We can only zoneindex (to align size to the chunk size) if the new
1230 * size is not too large.
1231 */
1232 if (size < ZoneLimit) {
1233 zoneindex(&size, &align);
1234 if (z->z_ChunkSize == size)
1235 return(ptr);
1236 }
1237 if ((nptr = _kmalloc_debug(size, type, flags, file, line)) == NULL)
1238 return(NULL);
1239 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
1240 kfree(ptr, type);
1241 return(nptr);
1242 }
1243
1244 size_t
kmalloc_usable_size(const void * ptr)1245 kmalloc_usable_size(const void *ptr)
1246 {
1247 unsigned long size;
1248 SLZone *z;
1249 int *kup;
1250
1251 if (ptr == NULL)
1252 return 0;
1253 if (ptr == ZERO_LENGTH_PTR)
1254 return 0;
1255
1256 /*
1257 * Check to see if the pointer blongs to an oversized segment
1258 */
1259 kup = btokup(ptr);
1260 if (*kup > 0) {
1261 size = *kup << PAGE_SHIFT;
1262 return size;
1263 }
1264
1265 /*
1266 * Zone case. Figure out the zone based on the fact that it is
1267 * ZoneSize aligned.
1268 */
1269 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1270 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1271
1272 return (z->z_ChunkSize);
1273 }
1274
1275 /*
1276 * Return the kmalloc limit for this type, in bytes.
1277 */
1278 long
kmalloc_limit(struct malloc_type * type)1279 kmalloc_limit(struct malloc_type *type)
1280 {
1281 KKASSERT(type->ks_limit != 0);
1282 return(type->ks_limit);
1283 }
1284
1285 /*
1286 * Allocate a copy of the specified string.
1287 *
1288 * (MP SAFE) (MAY BLOCK)
1289 */
1290 #ifdef SLAB_DEBUG
1291 char *
kstrdup_debug(const char * str,struct malloc_type * type,const char * file,int line)1292 kstrdup_debug(const char *str, struct malloc_type *type,
1293 const char *file, int line)
1294 #else
1295 char *
1296 kstrdup(const char *str, struct malloc_type *type)
1297 #endif
1298 {
1299 int zlen; /* length inclusive of terminating NUL */
1300 char *nstr;
1301
1302 if (str == NULL)
1303 return(NULL);
1304 zlen = strlen(str) + 1;
1305 nstr = _kmalloc_debug(zlen, type, M_WAITOK, file, line);
1306 bcopy(str, nstr, zlen);
1307 return(nstr);
1308 }
1309
1310 #ifdef SLAB_DEBUG
1311 char *
kstrndup_debug(const char * str,size_t maxlen,struct malloc_type * type,const char * file,int line)1312 kstrndup_debug(const char *str, size_t maxlen, struct malloc_type *type,
1313 const char *file, int line)
1314 #else
1315 char *
1316 kstrndup(const char *str, size_t maxlen, struct malloc_type *type)
1317 #endif
1318 {
1319 int zlen; /* length inclusive of terminating NUL */
1320 char *nstr;
1321
1322 if (str == NULL)
1323 return(NULL);
1324 zlen = strnlen(str, maxlen) + 1;
1325 nstr = _kmalloc_debug(zlen, type, M_WAITOK, file, line);
1326 bcopy(str, nstr, zlen);
1327 nstr[zlen - 1] = '\0';
1328 return(nstr);
1329 }
1330
1331 /*
1332 * Notify our cpu that a remote cpu has freed some chunks in a zone that
1333 * we own. RCount will be bumped so the memory should be good, but validate
1334 * that it really is.
1335 */
1336 static void
kfree_remote(void * ptr)1337 kfree_remote(void *ptr)
1338 {
1339 SLGlobalData *slgd;
1340 SLZone *z;
1341 int nfree;
1342 int *kup;
1343
1344 slgd = &mycpu->gd_slab;
1345 z = ptr;
1346 kup = btokup(z);
1347 KKASSERT(*kup == -((int)mycpuid + 1));
1348 KKASSERT(z->z_RCount > 0);
1349 atomic_subtract_int(&z->z_RCount, 1);
1350
1351 logmemory(free_rem_beg, z, NULL, 0L, 0);
1352 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1353 KKASSERT(z->z_Cpu == mycpu->gd_cpuid);
1354 nfree = z->z_NFree;
1355
1356 /*
1357 * Indicate that we will no longer be off of the ZoneAry by
1358 * clearing RSignal.
1359 */
1360 if (z->z_RChunks)
1361 z->z_RSignal = 0;
1362
1363 /*
1364 * Atomically extract the bchunks list and then process it back
1365 * into the lchunks list. We want to append our bchunks to the
1366 * lchunks list and not prepend since we likely do not have
1367 * cache mastership of the related data (not that it helps since
1368 * we are using c_Next).
1369 */
1370 clean_zone_rchunks(z);
1371 if (z->z_NFree && nfree == 0) {
1372 TAILQ_INSERT_HEAD(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
1373 }
1374
1375 check_zone_free(slgd, z);
1376 logmemory(free_rem_end, z, NULL, 0L, 0);
1377 }
1378
1379 /*
1380 * free (SLAB ALLOCATOR)
1381 *
1382 * Free a memory block previously allocated by malloc.
1383 *
1384 * Note: We do not attempt to update ks_loosememuse as MP races could
1385 * prevent us from checking memory limits in malloc. YYY we may
1386 * consider updating ks_cpu.loosememuse.
1387 *
1388 * MPSAFE
1389 */
1390 void
_kfree(void * ptr,struct malloc_type * type)1391 _kfree(void *ptr, struct malloc_type *type)
1392 {
1393 SLZone *z;
1394 SLChunk *chunk;
1395 SLGlobalData *slgd;
1396 struct globaldata *gd;
1397 int *kup;
1398 unsigned long size;
1399 SLChunk *bchunk;
1400 int rsignal;
1401
1402 logmemory_quick(free_beg);
1403 gd = mycpu;
1404 slgd = &gd->gd_slab;
1405
1406 if (ptr == NULL)
1407 panic("trying to free NULL pointer");
1408
1409 /*
1410 * Handle special 0-byte allocations
1411 */
1412 if (ptr == ZERO_LENGTH_PTR) {
1413 logmemory(free_zero, ptr, type, -1UL, 0);
1414 logmemory_quick(free_end);
1415 return;
1416 }
1417
1418 /*
1419 * Panic on bad malloc type
1420 */
1421 if (type->ks_magic != M_MAGIC)
1422 panic("free: malloc type lacks magic");
1423
1424 /*
1425 * Handle oversized allocations. XXX we really should require that a
1426 * size be passed to free() instead of this nonsense.
1427 *
1428 * This code is never called via an ipi.
1429 */
1430 kup = btokup(ptr);
1431 if (*kup > 0) {
1432 size = *kup << PAGE_SHIFT;
1433 *kup = 0;
1434 #ifdef INVARIANTS
1435 if (use_weird_array) {
1436 KKASSERT(sizeof(weirdary) <= size);
1437 bcopy(weirdary, ptr, sizeof(weirdary));
1438 }
1439 #endif
1440 /*
1441 * NOTE: For oversized allocations we do not record the
1442 * originating cpu. It gets freed on the cpu calling
1443 * kfree(). The statistics are in aggregate.
1444 *
1445 * note: XXX we have still inherited the interrupts-can't-block
1446 * assumption. An interrupt thread does not bump
1447 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
1448 * primarily until we can fix softupdate's assumptions about free().
1449 */
1450 crit_enter();
1451 --type->ks_use[gd->gd_cpuid].inuse;
1452 type->ks_use[gd->gd_cpuid].memuse -= size;
1453 if (mycpu->gd_intr_nesting_level ||
1454 (gd->gd_curthread->td_flags & TDF_INTTHREAD)) {
1455 logmemory(free_ovsz_delayed, ptr, type, size, 0);
1456 z = (SLZone *)ptr;
1457 z->z_Magic = ZALLOC_OVSZ_MAGIC;
1458 z->z_ChunkSize = size;
1459
1460 TAILQ_INSERT_HEAD(&slgd->FreeOvZones, z, z_Entry);
1461 crit_exit();
1462 } else {
1463 crit_exit();
1464 logmemory(free_ovsz, ptr, type, size, 0);
1465 kmem_slab_free(ptr, size); /* may block */
1466 }
1467 logmemory_quick(free_end);
1468 return;
1469 }
1470
1471 /*
1472 * Zone case. Figure out the zone based on the fact that it is
1473 * ZoneSize aligned.
1474 */
1475 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1476 kup = btokup(z);
1477 KKASSERT(*kup < 0);
1478 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1479
1480 /*
1481 * If we do not own the zone then use atomic ops to free to the
1482 * remote cpu linked list and notify the target zone using a
1483 * passive message.
1484 *
1485 * The target zone cannot be deallocated while we own a chunk of it,
1486 * so the zone header's storage is stable until the very moment
1487 * we adjust z_RChunks. After that we cannot safely dereference (z).
1488 *
1489 * (no critical section needed)
1490 */
1491 if (z->z_CpuGd != gd) {
1492 /*
1493 * Making these adjustments now allow us to avoid passing (type)
1494 * to the remote cpu. Note that inuse/memuse is being
1495 * adjusted on OUR cpu, not the zone cpu, but it should all still
1496 * sum up properly and cancel out.
1497 */
1498 crit_enter();
1499 --type->ks_use[gd->gd_cpuid].inuse;
1500 type->ks_use[gd->gd_cpuid].memuse -= z->z_ChunkSize;
1501 crit_exit();
1502
1503 /*
1504 * WARNING! This code competes with other cpus. Once we
1505 * successfully link the chunk to RChunks the remote
1506 * cpu can rip z's storage out from under us.
1507 *
1508 * Bumping RCount prevents z's storage from getting
1509 * ripped out.
1510 */
1511 rsignal = z->z_RSignal;
1512 cpu_lfence();
1513 if (rsignal)
1514 atomic_add_int(&z->z_RCount, 1);
1515
1516 chunk = ptr;
1517 for (;;) {
1518 bchunk = z->z_RChunks;
1519 cpu_ccfence();
1520 chunk->c_Next = bchunk;
1521 cpu_sfence();
1522
1523 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, chunk))
1524 break;
1525 }
1526
1527 /*
1528 * We have to signal the remote cpu if our actions will cause
1529 * the remote zone to be placed back on ZoneAry so it can
1530 * move the zone back on.
1531 *
1532 * We only need to deal with NULL->non-NULL RChunk transitions
1533 * and only if z_RSignal is set. We interlock by reading rsignal
1534 * before adding our chunk to RChunks. This should result in
1535 * virtually no IPI traffic.
1536 *
1537 * We can use a passive IPI to reduce overhead even further.
1538 */
1539 if (bchunk == NULL && rsignal) {
1540 logmemory(free_request, ptr, type,
1541 (unsigned long)z->z_ChunkSize, 0);
1542 lwkt_send_ipiq_passive(z->z_CpuGd, kfree_remote, z);
1543 /* z can get ripped out from under us from this point on */
1544 } else if (rsignal) {
1545 atomic_subtract_int(&z->z_RCount, 1);
1546 /* z can get ripped out from under us from this point on */
1547 }
1548 logmemory_quick(free_end);
1549 return;
1550 }
1551
1552 /*
1553 * kfree locally
1554 */
1555 logmemory(free_chunk, ptr, type, (unsigned long)z->z_ChunkSize, 0);
1556
1557 crit_enter();
1558 chunk = ptr;
1559 chunk_mark_free(z, chunk);
1560
1561 /*
1562 * Put weird data into the memory to detect modifications after freeing,
1563 * illegal pointer use after freeing (we should fault on the odd address),
1564 * and so forth. XXX needs more work, see the old malloc code.
1565 */
1566 #ifdef INVARIANTS
1567 if (use_weird_array) {
1568 if (z->z_ChunkSize < sizeof(weirdary))
1569 bcopy(weirdary, chunk, z->z_ChunkSize);
1570 else
1571 bcopy(weirdary, chunk, sizeof(weirdary));
1572 }
1573 #endif
1574
1575 /*
1576 * Add this free non-zero'd chunk to a linked list for reuse. Add
1577 * to the front of the linked list so it is more likely to be
1578 * reallocated, since it is already in our L1 cache.
1579 */
1580 #ifdef INVARIANTS
1581 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd)
1582 panic("BADFREE %p", chunk);
1583 #endif
1584 chunk->c_Next = z->z_LChunks;
1585 z->z_LChunks = chunk;
1586 if (chunk->c_Next == NULL)
1587 z->z_LChunksp = &chunk->c_Next;
1588
1589 #ifdef INVARIANTS
1590 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart)
1591 panic("BADFREE2");
1592 #endif
1593
1594 /*
1595 * Bump the number of free chunks. If it becomes non-zero the zone
1596 * must be added back onto the appropriate list. A fully allocated
1597 * zone that sees its first free is considered 'mature' and is placed
1598 * at the head, giving the system time to potentially free the remaining
1599 * entries even while other allocations are going on and making the zone
1600 * freeable.
1601 */
1602 if (z->z_NFree++ == 0)
1603 TAILQ_INSERT_HEAD(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
1604
1605 --type->ks_use[gd->gd_cpuid].inuse;
1606 type->ks_use[gd->gd_cpuid].memuse -= z->z_ChunkSize;
1607
1608 check_zone_free(slgd, z);
1609 logmemory_quick(free_end);
1610 crit_exit();
1611 }
1612
1613 /*
1614 * Cleanup slabs which are hanging around due to RChunks or which are wholely
1615 * free and can be moved to the free list if not moved by other means.
1616 *
1617 * Called once every 10 seconds on all cpus.
1618 */
1619 void
slab_cleanup(void)1620 slab_cleanup(void)
1621 {
1622 SLGlobalData *slgd = &mycpu->gd_slab;
1623 SLZone *z;
1624 int i;
1625
1626 crit_enter();
1627 for (i = 0; i < NZONES; ++i) {
1628 if ((z = TAILQ_FIRST(&slgd->ZoneAry[i])) == NULL)
1629 continue;
1630
1631 /*
1632 * Scan zones.
1633 */
1634 while (z) {
1635 /*
1636 * Shift all RChunks to the end of the LChunks list. This is
1637 * an O(1) operation.
1638 *
1639 * Then free the zone if possible.
1640 */
1641 clean_zone_rchunks(z);
1642 z = check_zone_free(slgd, z);
1643 }
1644 }
1645 crit_exit();
1646 }
1647
1648 #if defined(INVARIANTS)
1649
1650 /*
1651 * Helper routines for sanity checks
1652 */
1653 static void
chunk_mark_allocated(SLZone * z,void * chunk)1654 chunk_mark_allocated(SLZone *z, void *chunk)
1655 {
1656 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1657 uint32_t *bitptr;
1658
1659 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1660 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1661 ("memory chunk %p bit index %d is illegal", chunk, bitdex));
1662 bitptr = &z->z_Bitmap[bitdex >> 5];
1663 bitdex &= 31;
1664 KASSERT((*bitptr & (1 << bitdex)) == 0,
1665 ("memory chunk %p is already allocated!", chunk));
1666 *bitptr |= 1 << bitdex;
1667 }
1668
1669 static void
chunk_mark_free(SLZone * z,void * chunk)1670 chunk_mark_free(SLZone *z, void *chunk)
1671 {
1672 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1673 uint32_t *bitptr;
1674
1675 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1676 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1677 ("memory chunk %p bit index %d is illegal!", chunk, bitdex));
1678 bitptr = &z->z_Bitmap[bitdex >> 5];
1679 bitdex &= 31;
1680 KASSERT((*bitptr & (1 << bitdex)) != 0,
1681 ("memory chunk %p is already free!", chunk));
1682 *bitptr &= ~(1 << bitdex);
1683 }
1684
1685 #endif
1686
1687 /*
1688 * kmem_slab_alloc()
1689 *
1690 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1691 * specified alignment. M_* flags are expected in the flags field.
1692 *
1693 * Alignment must be a multiple of PAGE_SIZE.
1694 *
1695 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1696 * but when we move zalloc() over to use this function as its backend
1697 * we will have to switch to kreserve/krelease and call reserve(0)
1698 * after the new space is made available.
1699 *
1700 * Interrupt code which has preempted other code is not allowed to
1701 * use PQ_CACHE pages. However, if an interrupt thread is run
1702 * non-preemptively or blocks and then runs non-preemptively, then
1703 * it is free to use PQ_CACHE pages. <--- may not apply any longer XXX
1704 */
1705 void *
kmem_slab_alloc(vm_size_t size,vm_offset_t align,int flags)1706 kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
1707 {
1708 vm_size_t i;
1709 vm_offset_t addr;
1710 int count, vmflags, base_vmflags;
1711 vm_page_t mbase = NULL;
1712 vm_page_t m;
1713 thread_t td;
1714
1715 size = round_page(size);
1716 addr = vm_map_min(kernel_map);
1717
1718 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1719 crit_enter();
1720 vm_map_lock(kernel_map);
1721 if (vm_map_findspace(kernel_map, addr, size, align, 0, &addr)) {
1722 vm_map_unlock(kernel_map);
1723 if ((flags & M_NULLOK) == 0)
1724 panic("kmem_slab_alloc(): kernel_map ran out of space!");
1725 vm_map_entry_release(count);
1726 crit_exit();
1727 return(NULL);
1728 }
1729
1730 /*
1731 * kernel_object maps 1:1 to kernel_map.
1732 */
1733 vm_object_hold(kernel_object);
1734 vm_object_reference_locked(kernel_object);
1735 vm_map_insert(kernel_map, &count,
1736 kernel_object, NULL,
1737 addr, NULL,
1738 addr, addr + size,
1739 VM_MAPTYPE_NORMAL,
1740 VM_SUBSYS_KMALLOC,
1741 VM_PROT_ALL, VM_PROT_ALL, 0);
1742 vm_object_drop(kernel_object);
1743 vm_map_set_wired_quick(kernel_map, addr, size, &count);
1744 vm_map_unlock(kernel_map);
1745
1746 td = curthread;
1747
1748 base_vmflags = 0;
1749 if (flags & M_ZERO)
1750 base_vmflags |= VM_ALLOC_ZERO;
1751 if (flags & M_USE_RESERVE)
1752 base_vmflags |= VM_ALLOC_SYSTEM;
1753 if (flags & M_USE_INTERRUPT_RESERVE)
1754 base_vmflags |= VM_ALLOC_INTERRUPT;
1755 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) {
1756 panic("kmem_slab_alloc: bad flags %08x (%p)",
1757 flags, ((int **)&size)[-1]);
1758 }
1759
1760 /*
1761 * Allocate the pages. Do not map them yet. VM_ALLOC_NORMAL can only
1762 * be set if we are not preempting.
1763 *
1764 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1765 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1766 * implied in this case), though I'm not sure if we really need to
1767 * do that.
1768 */
1769 vmflags = base_vmflags;
1770 if (flags & M_WAITOK) {
1771 if (td->td_preempted)
1772 vmflags |= VM_ALLOC_SYSTEM;
1773 else
1774 vmflags |= VM_ALLOC_NORMAL;
1775 }
1776
1777 vm_object_hold(kernel_object);
1778 for (i = 0; i < size; i += PAGE_SIZE) {
1779 m = vm_page_alloc(kernel_object, OFF_TO_IDX(addr + i), vmflags);
1780 if (i == 0)
1781 mbase = m;
1782
1783 /*
1784 * If the allocation failed we either return NULL or we retry.
1785 *
1786 * If M_WAITOK is specified we wait for more memory and retry.
1787 * If M_WAITOK is specified from a preemption we yield instead of
1788 * wait. Livelock will not occur because the interrupt thread
1789 * will not be preempting anyone the second time around after the
1790 * yield.
1791 */
1792 if (m == NULL) {
1793 if (flags & M_WAITOK) {
1794 if (td->td_preempted) {
1795 lwkt_switch();
1796 } else {
1797 vm_wait(0);
1798 }
1799 i -= PAGE_SIZE; /* retry */
1800 continue;
1801 }
1802 break;
1803 }
1804 }
1805
1806 /*
1807 * Check and deal with an allocation failure
1808 */
1809 if (i != size) {
1810 while (i != 0) {
1811 i -= PAGE_SIZE;
1812 m = vm_page_lookup(kernel_object, OFF_TO_IDX(addr + i));
1813 /* page should already be busy */
1814 vm_page_free(m);
1815 }
1816 vm_map_lock(kernel_map);
1817 vm_map_delete(kernel_map, addr, addr + size, &count);
1818 vm_map_unlock(kernel_map);
1819 vm_object_drop(kernel_object);
1820
1821 vm_map_entry_release(count);
1822 crit_exit();
1823 return(NULL);
1824 }
1825
1826 /*
1827 * Success!
1828 *
1829 * NOTE: The VM pages are still busied. mbase points to the first one
1830 * but we have to iterate via vm_page_next()
1831 */
1832 vm_object_drop(kernel_object);
1833 crit_exit();
1834
1835 /*
1836 * Enter the pages into the pmap and deal with M_ZERO.
1837 */
1838 m = mbase;
1839 i = 0;
1840
1841 while (i < size) {
1842 /*
1843 * page should already be busy
1844 */
1845 m->valid = VM_PAGE_BITS_ALL;
1846 vm_page_wire(m);
1847 pmap_enter(kernel_pmap, addr + i, m,
1848 VM_PROT_ALL | VM_PROT_NOSYNC, 1, NULL);
1849 if (flags & M_ZERO)
1850 pagezero((char *)addr + i);
1851 KKASSERT(m->flags & (PG_WRITEABLE | PG_MAPPED));
1852 vm_page_flag_set(m, PG_REFERENCED);
1853 vm_page_wakeup(m);
1854
1855 i += PAGE_SIZE;
1856 vm_object_hold(kernel_object);
1857 m = vm_page_next(m);
1858 vm_object_drop(kernel_object);
1859 }
1860 smp_invltlb();
1861 vm_map_entry_release(count);
1862 return((void *)addr);
1863 }
1864
1865 /*
1866 * kmem_slab_free()
1867 */
1868 void
kmem_slab_free(void * ptr,vm_size_t size)1869 kmem_slab_free(void *ptr, vm_size_t size)
1870 {
1871 crit_enter();
1872 vm_map_remove(kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
1873 crit_exit();
1874 }
1875