xref: /dragonfly/sys/kern/kern_slaballoc.c (revision a78dfe84)
1 /*
2  * (MPSAFE)
3  *
4  * KERN_SLABALLOC.C	- Kernel SLAB memory allocator
5  *
6  * Copyright (c) 2003,2004,2010 The DragonFly Project.  All rights reserved.
7  *
8  * This code is derived from software contributed to The DragonFly Project
9  * by Matthew Dillon <dillon@backplane.com>
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  *
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in
19  *    the documentation and/or other materials provided with the
20  *    distribution.
21  * 3. Neither the name of The DragonFly Project nor the names of its
22  *    contributors may be used to endorse or promote products derived
23  *    from this software without specific, prior written permission.
24  *
25  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
26  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
27  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
28  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
29  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
30  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
31  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
32  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
33  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
34  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
35  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  * This module implements a slab allocator drop-in replacement for the
39  * kernel malloc().
40  *
41  * A slab allocator reserves a ZONE for each chunk size, then lays the
42  * chunks out in an array within the zone.  Allocation and deallocation
43  * is nearly instantanious, and fragmentation/overhead losses are limited
44  * to a fixed worst-case amount.
45  *
46  * The downside of this slab implementation is in the chunk size
47  * multiplied by the number of zones.  ~80 zones * 128K = 10MB of VM per cpu.
48  * In a kernel implementation all this memory will be physical so
49  * the zone size is adjusted downward on machines with less physical
50  * memory.  The upside is that overhead is bounded... this is the *worst*
51  * case overhead.
52  *
53  * Slab management is done on a per-cpu basis and no locking or mutexes
54  * are required, only a critical section.  When one cpu frees memory
55  * belonging to another cpu's slab manager an asynchronous IPI message
56  * will be queued to execute the operation.   In addition, both the
57  * high level slab allocator and the low level zone allocator optimize
58  * M_ZERO requests, and the slab allocator does not have to pre initialize
59  * the linked list of chunks.
60  *
61  * XXX Balancing is needed between cpus.  Balance will be handled through
62  * asynchronous IPIs primarily by reassigning the z_Cpu ownership of chunks.
63  *
64  * XXX If we have to allocate a new zone and M_USE_RESERVE is set, use of
65  * the new zone should be restricted to M_USE_RESERVE requests only.
66  *
67  *	Alloc Size	Chunking        Number of zones
68  *	0-127		8		16
69  *	128-255		16		8
70  *	256-511		32		8
71  *	512-1023	64		8
72  *	1024-2047	128		8
73  *	2048-4095	256		8
74  *	4096-8191	512		8
75  *	8192-16383	1024		8
76  *	16384-32767	2048		8
77  *	(if PAGE_SIZE is 4K the maximum zone allocation is 16383)
78  *
79  *	Allocations >= ZoneLimit 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 "opt_vm.h"
98 
99 #include <sys/param.h>
100 #include <sys/systm.h>
101 #include <sys/kernel.h>
102 #include <sys/slaballoc.h>
103 #include <sys/mbuf.h>
104 #include <sys/vmmeter.h>
105 #include <sys/lock.h>
106 #include <sys/thread.h>
107 #include <sys/globaldata.h>
108 #include <sys/sysctl.h>
109 #include <sys/ktr.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 #define btokup(z)	(&pmap_kvtom((vm_offset_t)(z))->ku_pagecnt)
127 
128 #define MEMORY_STRING	"ptr=%p type=%p size=%lu flags=%04x"
129 #define MEMORY_ARGS	void *ptr, void *type, unsigned long size, int flags
130 
131 #if !defined(KTR_MEMORY)
132 #define KTR_MEMORY	KTR_ALL
133 #endif
134 KTR_INFO_MASTER(memory);
135 KTR_INFO(KTR_MEMORY, memory, malloc_beg, 0, "malloc begin");
136 KTR_INFO(KTR_MEMORY, memory, malloc_end, 1, MEMORY_STRING, MEMORY_ARGS);
137 KTR_INFO(KTR_MEMORY, memory, free_zero, 2, MEMORY_STRING, MEMORY_ARGS);
138 KTR_INFO(KTR_MEMORY, memory, free_ovsz, 3, MEMORY_STRING, MEMORY_ARGS);
139 KTR_INFO(KTR_MEMORY, memory, free_ovsz_delayed, 4, MEMORY_STRING, MEMORY_ARGS);
140 KTR_INFO(KTR_MEMORY, memory, free_chunk, 5, MEMORY_STRING, MEMORY_ARGS);
141 KTR_INFO(KTR_MEMORY, memory, free_request, 6, MEMORY_STRING, MEMORY_ARGS);
142 KTR_INFO(KTR_MEMORY, memory, free_rem_beg, 7, MEMORY_STRING, MEMORY_ARGS);
143 KTR_INFO(KTR_MEMORY, memory, free_rem_end, 8, MEMORY_STRING, MEMORY_ARGS);
144 KTR_INFO(KTR_MEMORY, memory, free_beg, 9, "free begin");
145 KTR_INFO(KTR_MEMORY, memory, free_end, 10, "free end");
146 
147 #define logmemory(name, ptr, type, size, flags)				\
148 	KTR_LOG(memory_ ## name, ptr, type, size, flags)
149 #define logmemory_quick(name)						\
150 	KTR_LOG(memory_ ## name)
151 
152 /*
153  * Fixed globals (not per-cpu)
154  */
155 static int ZoneSize;
156 static int ZoneLimit;
157 static int ZonePageCount;
158 static uintptr_t ZoneMask;
159 static int ZoneBigAlloc;		/* in KB */
160 static int ZoneGenAlloc;		/* in KB */
161 struct malloc_type *kmemstatistics;	/* exported to vmstat */
162 static int32_t weirdary[16];
163 
164 static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags);
165 static void kmem_slab_free(void *ptr, vm_size_t bytes);
166 
167 #if defined(INVARIANTS)
168 static void chunk_mark_allocated(SLZone *z, void *chunk);
169 static void chunk_mark_free(SLZone *z, void *chunk);
170 #else
171 #define chunk_mark_allocated(z, chunk)
172 #define chunk_mark_free(z, chunk)
173 #endif
174 
175 /*
176  * Misc constants.  Note that allocations that are exact multiples of
177  * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
178  */
179 #define ZONE_RELS_THRESH	32		/* threshold number of zones */
180 
181 /*
182  * The WEIRD_ADDR is used as known text to copy into free objects to
183  * try to create deterministic failure cases if the data is accessed after
184  * free.
185  */
186 #define WEIRD_ADDR      0xdeadc0de
187 #define MAX_COPY        sizeof(weirdary)
188 #define ZERO_LENGTH_PTR	((void *)-8)
189 
190 /*
191  * Misc global malloc buckets
192  */
193 
194 MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
195 MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
196 MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
197 MALLOC_DEFINE(M_DRM, "m_drm", "DRM memory allocations");
198 
199 MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
200 MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
201 
202 /*
203  * Initialize the slab memory allocator.  We have to choose a zone size based
204  * on available physical memory.  We choose a zone side which is approximately
205  * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
206  * 128K.  The zone size is limited to the bounds set in slaballoc.h
207  * (typically 32K min, 128K max).
208  */
209 static void kmeminit(void *dummy);
210 
211 char *ZeroPage;
212 
213 SYSINIT(kmem, SI_BOOT1_ALLOCATOR, SI_ORDER_FIRST, kmeminit, NULL);
214 
215 #ifdef INVARIANTS
216 /*
217  * If enabled any memory allocated without M_ZERO is initialized to -1.
218  */
219 static int  use_malloc_pattern;
220 SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW,
221     &use_malloc_pattern, 0,
222     "Initialize memory to -1 if M_ZERO not specified");
223 #endif
224 
225 static int ZoneRelsThresh = ZONE_RELS_THRESH;
226 SYSCTL_INT(_kern, OID_AUTO, zone_big_alloc, CTLFLAG_RD, &ZoneBigAlloc, 0, "");
227 SYSCTL_INT(_kern, OID_AUTO, zone_gen_alloc, CTLFLAG_RD, &ZoneGenAlloc, 0, "");
228 SYSCTL_INT(_kern, OID_AUTO, zone_cache, CTLFLAG_RW, &ZoneRelsThresh, 0, "");
229 static long SlabsAllocated;
230 static long SlabsFreed;
231 SYSCTL_LONG(_kern, OID_AUTO, slabs_allocated, CTLFLAG_RD,
232 	    &SlabsAllocated, 0, "");
233 SYSCTL_LONG(_kern, OID_AUTO, slabs_freed, CTLFLAG_RD,
234 	    &SlabsFreed, 0, "");
235 static int SlabFreeToTail;
236 SYSCTL_INT(_kern, OID_AUTO, slab_freetotail, CTLFLAG_RW,
237 	    &SlabFreeToTail, 0, "");
238 
239 static struct spinlock kmemstat_spin =
240 			SPINLOCK_INITIALIZER(&kmemstat_spin, "malinit");
241 
242 /*
243  * Returns the kernel memory size limit for the purposes of initializing
244  * various subsystem caches.  The smaller of available memory and the KVM
245  * memory space is returned.
246  *
247  * The size in megabytes is returned.
248  */
249 size_t
250 kmem_lim_size(void)
251 {
252     size_t limsize;
253 
254     limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
255     if (limsize > KvaSize)
256 	limsize = KvaSize;
257     return (limsize / (1024 * 1024));
258 }
259 
260 static void
261 kmeminit(void *dummy)
262 {
263     size_t limsize;
264     int usesize;
265     int i;
266 
267     limsize = kmem_lim_size();
268     usesize = (int)(limsize * 1024);	/* convert to KB */
269 
270     /*
271      * If the machine has a large KVM space and more than 8G of ram,
272      * double the zone release threshold to reduce SMP invalidations.
273      * If more than 16G of ram, do it again.
274      *
275      * The BIOS eats a little ram so add some slop.  We want 8G worth of
276      * memory sticks to trigger the first adjustment.
277      */
278     if (ZoneRelsThresh == ZONE_RELS_THRESH) {
279 	    if (limsize >= 7 * 1024)
280 		    ZoneRelsThresh *= 2;
281 	    if (limsize >= 15 * 1024)
282 		    ZoneRelsThresh *= 2;
283     }
284 
285     /*
286      * Calculate the zone size.  This typically calculates to
287      * ZALLOC_MAX_ZONE_SIZE
288      */
289     ZoneSize = ZALLOC_MIN_ZONE_SIZE;
290     while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
291 	ZoneSize <<= 1;
292     ZoneLimit = ZoneSize / 4;
293     if (ZoneLimit > ZALLOC_ZONE_LIMIT)
294 	ZoneLimit = ZALLOC_ZONE_LIMIT;
295     ZoneMask = ~(uintptr_t)(ZoneSize - 1);
296     ZonePageCount = ZoneSize / PAGE_SIZE;
297 
298     for (i = 0; i < NELEM(weirdary); ++i)
299 	weirdary[i] = WEIRD_ADDR;
300 
301     ZeroPage = kmem_slab_alloc(PAGE_SIZE, PAGE_SIZE, M_WAITOK|M_ZERO);
302 
303     if (bootverbose)
304 	kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
305 }
306 
307 /*
308  * (low level) Initialize slab-related elements in the globaldata structure.
309  *
310  * Occurs after kmeminit().
311  */
312 void
313 slab_gdinit(globaldata_t gd)
314 {
315 	SLGlobalData *slgd;
316 	int i;
317 
318 	slgd = &gd->gd_slab;
319 	for (i = 0; i < NZONES; ++i)
320 		TAILQ_INIT(&slgd->ZoneAry[i]);
321 	TAILQ_INIT(&slgd->FreeZones);
322 	TAILQ_INIT(&slgd->FreeOvZones);
323 }
324 
325 /*
326  * Initialize a malloc type tracking structure.
327  */
328 void
329 malloc_init(void *data)
330 {
331     struct malloc_type *type = data;
332     size_t limsize;
333 
334     if (type->ks_magic != M_MAGIC)
335 	panic("malloc type lacks magic");
336 
337     if (type->ks_limit != 0)
338 	return;
339 
340     if (vmstats.v_page_count == 0)
341 	panic("malloc_init not allowed before vm init");
342 
343     limsize = kmem_lim_size() * (1024 * 1024);
344     type->ks_limit = limsize / 10;
345 
346     spin_lock(&kmemstat_spin);
347     type->ks_next = kmemstatistics;
348     kmemstatistics = type;
349     spin_unlock(&kmemstat_spin);
350 }
351 
352 void
353 malloc_uninit(void *data)
354 {
355     struct malloc_type *type = data;
356     struct malloc_type *t;
357 #ifdef INVARIANTS
358     int i;
359     long ttl;
360 #endif
361 
362     if (type->ks_magic != M_MAGIC)
363 	panic("malloc type lacks magic");
364 
365     if (vmstats.v_page_count == 0)
366 	panic("malloc_uninit not allowed before vm init");
367 
368     if (type->ks_limit == 0)
369 	panic("malloc_uninit on uninitialized type");
370 
371     /* Make sure that all pending kfree()s are finished. */
372     lwkt_synchronize_ipiqs("muninit");
373 
374 #ifdef INVARIANTS
375     /*
376      * memuse is only correct in aggregation.  Due to memory being allocated
377      * on one cpu and freed on another individual array entries may be
378      * negative or positive (canceling each other out).
379      */
380     for (i = ttl = 0; i < ncpus; ++i)
381 	ttl += type->ks_use[i].memuse;
382     if (ttl) {
383 	kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
384 	    ttl, type->ks_shortdesc, i);
385     }
386 #endif
387     spin_lock(&kmemstat_spin);
388     if (type == kmemstatistics) {
389 	kmemstatistics = type->ks_next;
390     } else {
391 	for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
392 	    if (t->ks_next == type) {
393 		t->ks_next = type->ks_next;
394 		break;
395 	    }
396 	}
397     }
398     type->ks_next = NULL;
399     type->ks_limit = 0;
400     spin_unlock(&kmemstat_spin);
401 }
402 
403 /*
404  * Increase the kmalloc pool limit for the specified pool.  No changes
405  * are the made if the pool would shrink.
406  */
407 void
408 kmalloc_raise_limit(struct malloc_type *type, size_t bytes)
409 {
410     if (type->ks_limit == 0)
411 	malloc_init(type);
412     if (bytes == 0)
413 	bytes = KvaSize;
414     if (type->ks_limit < bytes)
415 	type->ks_limit = bytes;
416 }
417 
418 void
419 kmalloc_set_unlimited(struct malloc_type *type)
420 {
421     type->ks_limit = kmem_lim_size() * (1024 * 1024);
422 }
423 
424 /*
425  * Dynamically create a malloc pool.  This function is a NOP if *typep is
426  * already non-NULL.
427  */
428 void
429 kmalloc_create(struct malloc_type **typep, const char *descr)
430 {
431 	struct malloc_type *type;
432 
433 	if (*typep == NULL) {
434 		type = kmalloc(sizeof(*type), M_TEMP, M_WAITOK | M_ZERO);
435 		type->ks_magic = M_MAGIC;
436 		type->ks_shortdesc = descr;
437 		malloc_init(type);
438 		*typep = type;
439 	}
440 }
441 
442 /*
443  * Destroy a dynamically created malloc pool.  This function is a NOP if
444  * the pool has already been destroyed.
445  */
446 void
447 kmalloc_destroy(struct malloc_type **typep)
448 {
449 	if (*typep != NULL) {
450 		malloc_uninit(*typep);
451 		kfree(*typep, M_TEMP);
452 		*typep = NULL;
453 	}
454 }
455 
456 /*
457  * Calculate the zone index for the allocation request size and set the
458  * allocation request size to that particular zone's chunk size.
459  */
460 static __inline int
461 zoneindex(unsigned long *bytes, unsigned long *align)
462 {
463     unsigned int n = (unsigned int)*bytes;	/* unsigned for shift opt */
464     if (n < 128) {
465 	*bytes = n = (n + 7) & ~7;
466 	*align = 8;
467 	return(n / 8 - 1);		/* 8 byte chunks, 16 zones */
468     }
469     if (n < 256) {
470 	*bytes = n = (n + 15) & ~15;
471 	*align = 16;
472 	return(n / 16 + 7);
473     }
474     if (n < 8192) {
475 	if (n < 512) {
476 	    *bytes = n = (n + 31) & ~31;
477 	    *align = 32;
478 	    return(n / 32 + 15);
479 	}
480 	if (n < 1024) {
481 	    *bytes = n = (n + 63) & ~63;
482 	    *align = 64;
483 	    return(n / 64 + 23);
484 	}
485 	if (n < 2048) {
486 	    *bytes = n = (n + 127) & ~127;
487 	    *align = 128;
488 	    return(n / 128 + 31);
489 	}
490 	if (n < 4096) {
491 	    *bytes = n = (n + 255) & ~255;
492 	    *align = 256;
493 	    return(n / 256 + 39);
494 	}
495 	*bytes = n = (n + 511) & ~511;
496 	*align = 512;
497 	return(n / 512 + 47);
498     }
499 #if ZALLOC_ZONE_LIMIT > 8192
500     if (n < 16384) {
501 	*bytes = n = (n + 1023) & ~1023;
502 	*align = 1024;
503 	return(n / 1024 + 55);
504     }
505 #endif
506 #if ZALLOC_ZONE_LIMIT > 16384
507     if (n < 32768) {
508 	*bytes = n = (n + 2047) & ~2047;
509 	*align = 2048;
510 	return(n / 2048 + 63);
511     }
512 #endif
513     panic("Unexpected byte count %d", n);
514     return(0);
515 }
516 
517 static __inline
518 void
519 clean_zone_rchunks(SLZone *z)
520 {
521     SLChunk *bchunk;
522 
523     while ((bchunk = z->z_RChunks) != NULL) {
524 	cpu_ccfence();
525 	if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
526 	    *z->z_LChunksp = bchunk;
527 	    while (bchunk) {
528 		chunk_mark_free(z, bchunk);
529 		z->z_LChunksp = &bchunk->c_Next;
530 		bchunk = bchunk->c_Next;
531 		++z->z_NFree;
532 	    }
533 	    break;
534 	}
535 	/* retry */
536     }
537 }
538 
539 /*
540  * If the zone becomes totally free and is not the only zone listed for a
541  * chunk size we move it to the FreeZones list.  We always leave at least
542  * one zone per chunk size listed, even if it is freeable.
543  *
544  * Do not move the zone if there is an IPI in_flight (z_RCount != 0),
545  * otherwise MP races can result in our free_remote code accessing a
546  * destroyed zone.  The remote end interlocks z_RCount with z_RChunks
547  * so one has to test both z_NFree and z_RCount.
548  *
549  * Since this code can be called from an IPI callback, do *NOT* try to mess
550  * with kernel_map here.  Hysteresis will be performed at kmalloc() time.
551  */
552 static __inline
553 SLZone *
554 check_zone_free(SLGlobalData *slgd, SLZone *z)
555 {
556     SLZone *znext;
557 
558     znext = TAILQ_NEXT(z, z_Entry);
559     if (z->z_NFree == z->z_NMax && z->z_RCount == 0 &&
560 	(TAILQ_FIRST(&slgd->ZoneAry[z->z_ZoneIndex]) != z || znext)
561     ) {
562 	int *kup;
563 
564 	TAILQ_REMOVE(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
565 
566 	z->z_Magic = -1;
567 	TAILQ_INSERT_HEAD(&slgd->FreeZones, z, z_Entry);
568 	++slgd->NFreeZones;
569 	kup = btokup(z);
570 	*kup = 0;
571     }
572     return znext;
573 }
574 
575 #ifdef SLAB_DEBUG
576 /*
577  * Used to debug memory corruption issues.  Record up to (typically 32)
578  * allocation sources for this zone (for a particular chunk size).
579  */
580 
581 static void
582 slab_record_source(SLZone *z, const char *file, int line)
583 {
584     int i;
585     int b = line & (SLAB_DEBUG_ENTRIES - 1);
586 
587     i = b;
588     do {
589 	if (z->z_Sources[i].file == file && z->z_Sources[i].line == line)
590 		return;
591 	if (z->z_Sources[i].file == NULL)
592 		break;
593 	i = (i + 1) & (SLAB_DEBUG_ENTRIES - 1);
594     } while (i != b);
595     z->z_Sources[i].file = file;
596     z->z_Sources[i].line = line;
597 }
598 
599 #endif
600 
601 static __inline unsigned long
602 powerof2_size(unsigned long size)
603 {
604 	int i;
605 
606 	if (size == 0 || powerof2(size))
607 		return size;
608 
609 	i = flsl(size);
610 	return (1UL << i);
611 }
612 
613 /*
614  * kmalloc()	(SLAB ALLOCATOR)
615  *
616  *	Allocate memory via the slab allocator.  If the request is too large,
617  *	or if it page-aligned beyond a certain size, we fall back to the
618  *	KMEM subsystem.  A SLAB tracking descriptor must be specified, use
619  *	&SlabMisc if you don't care.
620  *
621  *	M_RNOWAIT	- don't block.
622  *	M_NULLOK	- return NULL instead of blocking.
623  *	M_ZERO		- zero the returned memory.
624  *	M_USE_RESERVE	- allow greater drawdown of the free list
625  *	M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
626  *	M_POWEROF2	- roundup size to the nearest power of 2
627  *
628  * MPSAFE
629  */
630 
631 #ifdef SLAB_DEBUG
632 void *
633 kmalloc_debug(unsigned long size, struct malloc_type *type, int flags,
634 	      const char *file, int line)
635 #else
636 void *
637 kmalloc(unsigned long size, struct malloc_type *type, int flags)
638 #endif
639 {
640     SLZone *z;
641     SLChunk *chunk;
642     SLGlobalData *slgd;
643     struct globaldata *gd;
644     unsigned long align;
645     int zi;
646 #ifdef INVARIANTS
647     int i;
648 #endif
649 
650     logmemory_quick(malloc_beg);
651     gd = mycpu;
652     slgd = &gd->gd_slab;
653 
654     /*
655      * XXX silly to have this in the critical path.
656      */
657     if (type->ks_limit == 0) {
658 	crit_enter();
659 	malloc_init(type);
660 	crit_exit();
661     }
662     ++type->ks_calls;
663 
664     if (flags & M_POWEROF2)
665 	size = powerof2_size(size);
666 
667     /*
668      * Handle the case where the limit is reached.  Panic if we can't return
669      * NULL.  The original malloc code looped, but this tended to
670      * simply deadlock the computer.
671      *
672      * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
673      * to determine if a more complete limit check should be done.  The
674      * actual memory use is tracked via ks_use[cpu].memuse.
675      */
676     while (type->ks_loosememuse >= type->ks_limit) {
677 	int i;
678 	long ttl;
679 
680 	for (i = ttl = 0; i < ncpus; ++i)
681 	    ttl += type->ks_use[i].memuse;
682 	type->ks_loosememuse = ttl;	/* not MP synchronized */
683 	if ((ssize_t)ttl < 0)		/* deal with occassional race */
684 		ttl = 0;
685 	if (ttl >= type->ks_limit) {
686 	    if (flags & M_NULLOK) {
687 		logmemory(malloc_end, NULL, type, size, flags);
688 		return(NULL);
689 	    }
690 	    panic("%s: malloc limit exceeded", type->ks_shortdesc);
691 	}
692     }
693 
694     /*
695      * Handle the degenerate size == 0 case.  Yes, this does happen.
696      * Return a special pointer.  This is to maintain compatibility with
697      * the original malloc implementation.  Certain devices, such as the
698      * adaptec driver, not only allocate 0 bytes, they check for NULL and
699      * also realloc() later on.  Joy.
700      */
701     if (size == 0) {
702 	logmemory(malloc_end, ZERO_LENGTH_PTR, type, size, flags);
703 	return(ZERO_LENGTH_PTR);
704     }
705 
706     /*
707      * Handle hysteresis from prior frees here in malloc().  We cannot
708      * safely manipulate the kernel_map in free() due to free() possibly
709      * being called via an IPI message or from sensitive interrupt code.
710      *
711      * NOTE: ku_pagecnt must be cleared before we free the slab or we
712      *	     might race another cpu allocating the kva and setting
713      *	     ku_pagecnt.
714      */
715     while (slgd->NFreeZones > ZoneRelsThresh && (flags & M_RNOWAIT) == 0) {
716 	crit_enter();
717 	if (slgd->NFreeZones > ZoneRelsThresh) {	/* crit sect race */
718 	    int *kup;
719 
720 	    z = TAILQ_LAST(&slgd->FreeZones, SLZoneList);
721 	    KKASSERT(z != NULL);
722 	    TAILQ_REMOVE(&slgd->FreeZones, z, z_Entry);
723 	    --slgd->NFreeZones;
724 	    kup = btokup(z);
725 	    *kup = 0;
726 	    kmem_slab_free(z, ZoneSize);	/* may block */
727 	    atomic_add_int(&ZoneGenAlloc, -ZoneSize / 1024);
728 	}
729 	crit_exit();
730     }
731 
732     /*
733      * XXX handle oversized frees that were queued from kfree().
734      */
735     while (TAILQ_FIRST(&slgd->FreeOvZones) && (flags & M_RNOWAIT) == 0) {
736 	crit_enter();
737 	if ((z = TAILQ_LAST(&slgd->FreeOvZones, SLZoneList)) != NULL) {
738 	    vm_size_t tsize;
739 
740 	    KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
741 	    TAILQ_REMOVE(&slgd->FreeOvZones, z, z_Entry);
742 	    tsize = z->z_ChunkSize;
743 	    kmem_slab_free(z, tsize);	/* may block */
744 	    atomic_add_int(&ZoneBigAlloc, -(int)tsize / 1024);
745 	}
746 	crit_exit();
747     }
748 
749     /*
750      * Handle large allocations directly.  There should not be very many of
751      * these so performance is not a big issue.
752      *
753      * The backend allocator is pretty nasty on a SMP system.   Use the
754      * slab allocator for one and two page-sized chunks even though we lose
755      * some efficiency.  XXX maybe fix mmio and the elf loader instead.
756      */
757     if (size >= ZoneLimit || ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) {
758 	int *kup;
759 
760 	size = round_page(size);
761 	chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
762 	if (chunk == NULL) {
763 	    logmemory(malloc_end, NULL, type, size, flags);
764 	    return(NULL);
765 	}
766 	atomic_add_int(&ZoneBigAlloc, (int)size / 1024);
767 	flags &= ~M_ZERO;	/* result already zero'd if M_ZERO was set */
768 	flags |= M_PASSIVE_ZERO;
769 	kup = btokup(chunk);
770 	*kup = size / PAGE_SIZE;
771 	crit_enter();
772 	goto done;
773     }
774 
775     /*
776      * Attempt to allocate out of an existing zone.  First try the free list,
777      * then allocate out of unallocated space.  If we find a good zone move
778      * it to the head of the list so later allocations find it quickly
779      * (we might have thousands of zones in the list).
780      *
781      * Note: zoneindex() will panic of size is too large.
782      */
783     zi = zoneindex(&size, &align);
784     KKASSERT(zi < NZONES);
785     crit_enter();
786 
787     if ((z = TAILQ_LAST(&slgd->ZoneAry[zi], SLZoneList)) != NULL) {
788 	/*
789 	 * Locate a chunk - we have to have at least one.  If this is the
790 	 * last chunk go ahead and do the work to retrieve chunks freed
791 	 * from remote cpus, and if the zone is still empty move it off
792 	 * the ZoneAry.
793 	 */
794 	if (--z->z_NFree <= 0) {
795 	    KKASSERT(z->z_NFree == 0);
796 
797 	    /*
798 	     * WARNING! This code competes with other cpus.  It is ok
799 	     * for us to not drain RChunks here but we might as well, and
800 	     * it is ok if more accumulate after we're done.
801 	     *
802 	     * Set RSignal before pulling rchunks off, indicating that we
803 	     * will be moving ourselves off of the ZoneAry.  Remote ends will
804 	     * read RSignal before putting rchunks on thus interlocking
805 	     * their IPI signaling.
806 	     */
807 	    if (z->z_RChunks == NULL)
808 		atomic_swap_int(&z->z_RSignal, 1);
809 
810 	    clean_zone_rchunks(z);
811 
812 	    /*
813 	     * Remove from the zone list if no free chunks remain.
814 	     * Clear RSignal
815 	     */
816 	    if (z->z_NFree == 0) {
817 		TAILQ_REMOVE(&slgd->ZoneAry[zi], z, z_Entry);
818 	    } else {
819 		z->z_RSignal = 0;
820 	    }
821 	}
822 
823 	/*
824 	 * Fast path, we have chunks available in z_LChunks.
825 	 */
826 	chunk = z->z_LChunks;
827 	if (chunk) {
828 		chunk_mark_allocated(z, chunk);
829 		z->z_LChunks = chunk->c_Next;
830 		if (z->z_LChunks == NULL)
831 			z->z_LChunksp = &z->z_LChunks;
832 #ifdef SLAB_DEBUG
833 		slab_record_source(z, file, line);
834 #endif
835 		goto done;
836 	}
837 
838 	/*
839 	 * No chunks are available in LChunks, the free chunk MUST be
840 	 * in the never-before-used memory area, controlled by UIndex.
841 	 *
842 	 * The consequences are very serious if our zone got corrupted so
843 	 * we use an explicit panic rather than a KASSERT.
844 	 */
845 	if (z->z_UIndex + 1 != z->z_NMax)
846 	    ++z->z_UIndex;
847 	else
848 	    z->z_UIndex = 0;
849 
850 	if (z->z_UIndex == z->z_UEndIndex)
851 	    panic("slaballoc: corrupted zone");
852 
853 	chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
854 	if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
855 	    flags &= ~M_ZERO;
856 	    flags |= M_PASSIVE_ZERO;
857 	}
858 	chunk_mark_allocated(z, chunk);
859 #ifdef SLAB_DEBUG
860 	slab_record_source(z, file, line);
861 #endif
862 	goto done;
863     }
864 
865     /*
866      * If all zones are exhausted we need to allocate a new zone for this
867      * index.  Use M_ZERO to take advantage of pre-zerod pages.  Also see
868      * UAlloc use above in regards to M_ZERO.  Note that when we are reusing
869      * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
870      * we do not pre-zero it because we do not want to mess up the L1 cache.
871      *
872      * At least one subsystem, the tty code (see CROUND) expects power-of-2
873      * allocations to be power-of-2 aligned.  We maintain compatibility by
874      * adjusting the base offset below.
875      */
876     {
877 	int off;
878 	int *kup;
879 
880 	if ((z = TAILQ_FIRST(&slgd->FreeZones)) != NULL) {
881 	    TAILQ_REMOVE(&slgd->FreeZones, z, z_Entry);
882 	    --slgd->NFreeZones;
883 	    bzero(z, sizeof(SLZone));
884 	    z->z_Flags |= SLZF_UNOTZEROD;
885 	} else {
886 	    z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
887 	    if (z == NULL)
888 		goto fail;
889 	    atomic_add_int(&ZoneGenAlloc, ZoneSize / 1024);
890 	}
891 
892 	/*
893 	 * How big is the base structure?
894 	 */
895 #if defined(INVARIANTS)
896 	/*
897 	 * Make room for z_Bitmap.  An exact calculation is somewhat more
898 	 * complicated so don't make an exact calculation.
899 	 */
900 	off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]);
901 	bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8);
902 #else
903 	off = sizeof(SLZone);
904 #endif
905 
906 	/*
907 	 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
908 	 * Otherwise properly align the data according to the chunk size.
909 	 */
910 	if (powerof2(size))
911 	    align = size;
912 	off = roundup2(off, align);
913 
914 	z->z_Magic = ZALLOC_SLAB_MAGIC;
915 	z->z_ZoneIndex = zi;
916 	z->z_NMax = (ZoneSize - off) / size;
917 	z->z_NFree = z->z_NMax - 1;
918 	z->z_BasePtr = (char *)z + off;
919 	z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
920 	z->z_ChunkSize = size;
921 	z->z_CpuGd = gd;
922 	z->z_Cpu = gd->gd_cpuid;
923 	z->z_LChunksp = &z->z_LChunks;
924 #ifdef SLAB_DEBUG
925 	bcopy(z->z_Sources, z->z_AltSources, sizeof(z->z_Sources));
926 	bzero(z->z_Sources, sizeof(z->z_Sources));
927 #endif
928 	chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
929 	TAILQ_INSERT_HEAD(&slgd->ZoneAry[zi], z, z_Entry);
930 	if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
931 	    flags &= ~M_ZERO;	/* already zero'd */
932 	    flags |= M_PASSIVE_ZERO;
933 	}
934 	kup = btokup(z);
935 	*kup = -(z->z_Cpu + 1);	/* -1 to -(N+1) */
936 	chunk_mark_allocated(z, chunk);
937 #ifdef SLAB_DEBUG
938 	slab_record_source(z, file, line);
939 #endif
940 
941 	/*
942 	 * Slide the base index for initial allocations out of the next
943 	 * zone we create so we do not over-weight the lower part of the
944 	 * cpu memory caches.
945 	 */
946 	slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
947 				& (ZALLOC_MAX_ZONE_SIZE - 1);
948     }
949 
950 done:
951     ++type->ks_use[gd->gd_cpuid].inuse;
952     type->ks_use[gd->gd_cpuid].memuse += size;
953     type->ks_loosememuse += size;	/* not MP synchronized */
954     crit_exit();
955 
956     if (flags & M_ZERO)
957 	bzero(chunk, size);
958 #ifdef INVARIANTS
959     else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
960 	if (use_malloc_pattern) {
961 	    for (i = 0; i < size; i += sizeof(int)) {
962 		*(int *)((char *)chunk + i) = -1;
963 	    }
964 	}
965 	chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
966     }
967 #endif
968     logmemory(malloc_end, chunk, type, size, flags);
969     return(chunk);
970 fail:
971     crit_exit();
972     logmemory(malloc_end, NULL, type, size, flags);
973     return(NULL);
974 }
975 
976 /*
977  * kernel realloc.  (SLAB ALLOCATOR) (MP SAFE)
978  *
979  * Generally speaking this routine is not called very often and we do
980  * not attempt to optimize it beyond reusing the same pointer if the
981  * new size fits within the chunking of the old pointer's zone.
982  */
983 #ifdef SLAB_DEBUG
984 void *
985 krealloc_debug(void *ptr, unsigned long size,
986 	       struct malloc_type *type, int flags,
987 	       const char *file, int line)
988 #else
989 void *
990 krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
991 #endif
992 {
993     unsigned long osize;
994     unsigned long align;
995     SLZone *z;
996     void *nptr;
997     int *kup;
998 
999     KKASSERT((flags & M_ZERO) == 0);	/* not supported */
1000 
1001     if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
1002 	return(kmalloc_debug(size, type, flags, file, line));
1003     if (size == 0) {
1004 	kfree(ptr, type);
1005 	return(NULL);
1006     }
1007 
1008     /*
1009      * Handle oversized allocations.  XXX we really should require that a
1010      * size be passed to free() instead of this nonsense.
1011      */
1012     kup = btokup(ptr);
1013     if (*kup > 0) {
1014 	osize = *kup << PAGE_SHIFT;
1015 	if (osize == round_page(size))
1016 	    return(ptr);
1017 	if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL)
1018 	    return(NULL);
1019 	bcopy(ptr, nptr, min(size, osize));
1020 	kfree(ptr, type);
1021 	return(nptr);
1022     }
1023 
1024     /*
1025      * Get the original allocation's zone.  If the new request winds up
1026      * using the same chunk size we do not have to do anything.
1027      */
1028     z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1029     kup = btokup(z);
1030     KKASSERT(*kup < 0);
1031     KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1032 
1033     /*
1034      * Allocate memory for the new request size.  Note that zoneindex has
1035      * already adjusted the request size to the appropriate chunk size, which
1036      * should optimize our bcopy().  Then copy and return the new pointer.
1037      *
1038      * Resizing a non-power-of-2 allocation to a power-of-2 size does not
1039      * necessary align the result.
1040      *
1041      * We can only zoneindex (to align size to the chunk size) if the new
1042      * size is not too large.
1043      */
1044     if (size < ZoneLimit) {
1045 	zoneindex(&size, &align);
1046 	if (z->z_ChunkSize == size)
1047 	    return(ptr);
1048     }
1049     if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL)
1050 	return(NULL);
1051     bcopy(ptr, nptr, min(size, z->z_ChunkSize));
1052     kfree(ptr, type);
1053     return(nptr);
1054 }
1055 
1056 /*
1057  * Return the kmalloc limit for this type, in bytes.
1058  */
1059 long
1060 kmalloc_limit(struct malloc_type *type)
1061 {
1062     if (type->ks_limit == 0) {
1063 	crit_enter();
1064 	if (type->ks_limit == 0)
1065 	    malloc_init(type);
1066 	crit_exit();
1067     }
1068     return(type->ks_limit);
1069 }
1070 
1071 /*
1072  * Allocate a copy of the specified string.
1073  *
1074  * (MP SAFE) (MAY BLOCK)
1075  */
1076 #ifdef SLAB_DEBUG
1077 char *
1078 kstrdup_debug(const char *str, struct malloc_type *type,
1079 	      const char *file, int line)
1080 #else
1081 char *
1082 kstrdup(const char *str, struct malloc_type *type)
1083 #endif
1084 {
1085     int zlen;	/* length inclusive of terminating NUL */
1086     char *nstr;
1087 
1088     if (str == NULL)
1089 	return(NULL);
1090     zlen = strlen(str) + 1;
1091     nstr = kmalloc_debug(zlen, type, M_WAITOK, file, line);
1092     bcopy(str, nstr, zlen);
1093     return(nstr);
1094 }
1095 
1096 #ifdef SLAB_DEBUG
1097 char *
1098 kstrndup_debug(const char *str, size_t maxlen, struct malloc_type *type,
1099 	      const char *file, int line)
1100 #else
1101 char *
1102 kstrndup(const char *str, size_t maxlen, struct malloc_type *type)
1103 #endif
1104 {
1105     int zlen;	/* length inclusive of terminating NUL */
1106     char *nstr;
1107 
1108     if (str == NULL)
1109 	return(NULL);
1110     zlen = strnlen(str, maxlen) + 1;
1111     nstr = kmalloc_debug(zlen, type, M_WAITOK, file, line);
1112     bcopy(str, nstr, zlen);
1113     nstr[zlen - 1] = '\0';
1114     return(nstr);
1115 }
1116 
1117 /*
1118  * Notify our cpu that a remote cpu has freed some chunks in a zone that
1119  * we own.  RCount will be bumped so the memory should be good, but validate
1120  * that it really is.
1121  */
1122 static
1123 void
1124 kfree_remote(void *ptr)
1125 {
1126     SLGlobalData *slgd;
1127     SLZone *z;
1128     int nfree;
1129     int *kup;
1130 
1131     slgd = &mycpu->gd_slab;
1132     z = ptr;
1133     kup = btokup(z);
1134     KKASSERT(*kup == -((int)mycpuid + 1));
1135     KKASSERT(z->z_RCount > 0);
1136     atomic_subtract_int(&z->z_RCount, 1);
1137 
1138     logmemory(free_rem_beg, z, NULL, 0L, 0);
1139     KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1140     KKASSERT(z->z_Cpu  == mycpu->gd_cpuid);
1141     nfree = z->z_NFree;
1142 
1143     /*
1144      * Indicate that we will no longer be off of the ZoneAry by
1145      * clearing RSignal.
1146      */
1147     if (z->z_RChunks)
1148 	z->z_RSignal = 0;
1149 
1150     /*
1151      * Atomically extract the bchunks list and then process it back
1152      * into the lchunks list.  We want to append our bchunks to the
1153      * lchunks list and not prepend since we likely do not have
1154      * cache mastership of the related data (not that it helps since
1155      * we are using c_Next).
1156      */
1157     clean_zone_rchunks(z);
1158     if (z->z_NFree && nfree == 0) {
1159 	TAILQ_INSERT_HEAD(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
1160     }
1161 
1162     /*
1163      * If the zone becomes totally free and is not the only zone listed for a
1164      * chunk size we move it to the FreeZones list.  We always leave at least
1165      * one zone per chunk size listed, even if it is freeable.
1166      *
1167      * Since this code can be called from an IPI callback, do *NOT* try to
1168      * mess with kernel_map here.  Hysteresis will be performed at malloc()
1169      * time.
1170      *
1171      * Do not move the zone if there is an IPI in_flight (z_RCount != 0),
1172      * otherwise MP races can result in our free_remote code accessing a
1173      * destroyed zone.  The remote end interlocks z_RCount with z_RChunks
1174      * so one has to test both z_NFree and z_RCount.
1175      */
1176     if (z->z_NFree == z->z_NMax && z->z_RCount == 0 &&
1177 	(TAILQ_FIRST(&slgd->ZoneAry[z->z_ZoneIndex]) != z ||
1178 	 TAILQ_NEXT(z, z_Entry))
1179     ) {
1180 	int *kup;
1181 
1182 	TAILQ_REMOVE(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
1183 	z->z_Magic = -1;
1184 	TAILQ_INSERT_HEAD(&slgd->FreeZones, z, z_Entry);
1185 	++slgd->NFreeZones;
1186 	kup = btokup(z);
1187 	*kup = 0;
1188     }
1189     logmemory(free_rem_end, z, NULL, 0L, 0);
1190 }
1191 
1192 /*
1193  * free (SLAB ALLOCATOR)
1194  *
1195  * Free a memory block previously allocated by malloc.  Note that we do not
1196  * attempt to update ks_loosememuse as MP races could prevent us from
1197  * checking memory limits in malloc.
1198  *
1199  * MPSAFE
1200  */
1201 void
1202 kfree(void *ptr, struct malloc_type *type)
1203 {
1204     SLZone *z;
1205     SLChunk *chunk;
1206     SLGlobalData *slgd;
1207     struct globaldata *gd;
1208     int *kup;
1209     unsigned long size;
1210     SLChunk *bchunk;
1211     int rsignal;
1212 
1213     logmemory_quick(free_beg);
1214     gd = mycpu;
1215     slgd = &gd->gd_slab;
1216 
1217     if (ptr == NULL)
1218 	panic("trying to free NULL pointer");
1219 
1220     /*
1221      * Handle special 0-byte allocations
1222      */
1223     if (ptr == ZERO_LENGTH_PTR) {
1224 	logmemory(free_zero, ptr, type, -1UL, 0);
1225 	logmemory_quick(free_end);
1226 	return;
1227     }
1228 
1229     /*
1230      * Panic on bad malloc type
1231      */
1232     if (type->ks_magic != M_MAGIC)
1233 	panic("free: malloc type lacks magic");
1234 
1235     /*
1236      * Handle oversized allocations.  XXX we really should require that a
1237      * size be passed to free() instead of this nonsense.
1238      *
1239      * This code is never called via an ipi.
1240      */
1241     kup = btokup(ptr);
1242     if (*kup > 0) {
1243 	size = *kup << PAGE_SHIFT;
1244 	*kup = 0;
1245 #ifdef INVARIANTS
1246 	KKASSERT(sizeof(weirdary) <= size);
1247 	bcopy(weirdary, ptr, sizeof(weirdary));
1248 #endif
1249 	/*
1250 	 * NOTE: For oversized allocations we do not record the
1251 	 *	     originating cpu.  It gets freed on the cpu calling
1252 	 *	     kfree().  The statistics are in aggregate.
1253 	 *
1254 	 * note: XXX we have still inherited the interrupts-can't-block
1255 	 * assumption.  An interrupt thread does not bump
1256 	 * gd_intr_nesting_level so check TDF_INTTHREAD.  This is
1257 	 * primarily until we can fix softupdate's assumptions about free().
1258 	 */
1259 	crit_enter();
1260 	--type->ks_use[gd->gd_cpuid].inuse;
1261 	type->ks_use[gd->gd_cpuid].memuse -= size;
1262 	if (mycpu->gd_intr_nesting_level ||
1263 	    (gd->gd_curthread->td_flags & TDF_INTTHREAD))
1264 	{
1265 	    logmemory(free_ovsz_delayed, ptr, type, size, 0);
1266 	    z = (SLZone *)ptr;
1267 	    z->z_Magic = ZALLOC_OVSZ_MAGIC;
1268 	    z->z_ChunkSize = size;
1269 
1270 	    TAILQ_INSERT_HEAD(&slgd->FreeOvZones, z, z_Entry);
1271 	    crit_exit();
1272 	} else {
1273 	    crit_exit();
1274 	    logmemory(free_ovsz, ptr, type, size, 0);
1275 	    kmem_slab_free(ptr, size);	/* may block */
1276 	    atomic_add_int(&ZoneBigAlloc, -(int)size / 1024);
1277 	}
1278 	logmemory_quick(free_end);
1279 	return;
1280     }
1281 
1282     /*
1283      * Zone case.  Figure out the zone based on the fact that it is
1284      * ZoneSize aligned.
1285      */
1286     z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1287     kup = btokup(z);
1288     KKASSERT(*kup < 0);
1289     KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1290 
1291     /*
1292      * If we do not own the zone then use atomic ops to free to the
1293      * remote cpu linked list and notify the target zone using a
1294      * passive message.
1295      *
1296      * The target zone cannot be deallocated while we own a chunk of it,
1297      * so the zone header's storage is stable until the very moment
1298      * we adjust z_RChunks.  After that we cannot safely dereference (z).
1299      *
1300      * (no critical section needed)
1301      */
1302     if (z->z_CpuGd != gd) {
1303 	/*
1304 	 * Making these adjustments now allow us to avoid passing (type)
1305 	 * to the remote cpu.  Note that inuse/memuse is being
1306 	 * adjusted on OUR cpu, not the zone cpu, but it should all still
1307 	 * sum up properly and cancel out.
1308 	 */
1309 	crit_enter();
1310 	--type->ks_use[gd->gd_cpuid].inuse;
1311 	type->ks_use[gd->gd_cpuid].memuse -= z->z_ChunkSize;
1312 	crit_exit();
1313 
1314 	/*
1315 	 * WARNING! This code competes with other cpus.  Once we
1316 	 *	    successfully link the chunk to RChunks the remote
1317 	 *	    cpu can rip z's storage out from under us.
1318 	 *
1319 	 *	    Bumping RCount prevents z's storage from getting
1320 	 *	    ripped out.
1321 	 */
1322 	rsignal = z->z_RSignal;
1323 	cpu_lfence();
1324 	if (rsignal)
1325 		atomic_add_int(&z->z_RCount, 1);
1326 
1327 	chunk = ptr;
1328 	for (;;) {
1329 	    bchunk = z->z_RChunks;
1330 	    cpu_ccfence();
1331 	    chunk->c_Next = bchunk;
1332 	    cpu_sfence();
1333 
1334 	    if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, chunk))
1335 		break;
1336 	}
1337 
1338 	/*
1339 	 * We have to signal the remote cpu if our actions will cause
1340 	 * the remote zone to be placed back on ZoneAry so it can
1341 	 * move the zone back on.
1342 	 *
1343 	 * We only need to deal with NULL->non-NULL RChunk transitions
1344 	 * and only if z_RSignal is set.  We interlock by reading rsignal
1345 	 * before adding our chunk to RChunks.  This should result in
1346 	 * virtually no IPI traffic.
1347 	 *
1348 	 * We can use a passive IPI to reduce overhead even further.
1349 	 */
1350 	if (bchunk == NULL && rsignal) {
1351 		logmemory(free_request, ptr, type,
1352 			  (unsigned long)z->z_ChunkSize, 0);
1353 	    lwkt_send_ipiq_passive(z->z_CpuGd, kfree_remote, z);
1354 	    /* z can get ripped out from under us from this point on */
1355 	} else if (rsignal) {
1356 	    atomic_subtract_int(&z->z_RCount, 1);
1357 	    /* z can get ripped out from under us from this point on */
1358 	}
1359 	logmemory_quick(free_end);
1360 	return;
1361     }
1362 
1363     /*
1364      * kfree locally
1365      */
1366     logmemory(free_chunk, ptr, type, (unsigned long)z->z_ChunkSize, 0);
1367 
1368     crit_enter();
1369     chunk = ptr;
1370     chunk_mark_free(z, chunk);
1371 
1372     /*
1373      * Put weird data into the memory to detect modifications after freeing,
1374      * illegal pointer use after freeing (we should fault on the odd address),
1375      * and so forth.  XXX needs more work, see the old malloc code.
1376      */
1377 #ifdef INVARIANTS
1378     if (z->z_ChunkSize < sizeof(weirdary))
1379 	bcopy(weirdary, chunk, z->z_ChunkSize);
1380     else
1381 	bcopy(weirdary, chunk, sizeof(weirdary));
1382 #endif
1383 
1384     /*
1385      * Add this free non-zero'd chunk to a linked list for reuse.  Add
1386      * to the front of the linked list so it is more likely to be
1387      * reallocated, since it is already in our L1 cache.
1388      */
1389 #ifdef INVARIANTS
1390     if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd)
1391 	panic("BADFREE %p", chunk);
1392 #endif
1393     chunk->c_Next = z->z_LChunks;
1394     z->z_LChunks = chunk;
1395     if (chunk->c_Next == NULL)
1396 	    z->z_LChunksp = &chunk->c_Next;
1397 
1398 #ifdef INVARIANTS
1399     if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart)
1400 	panic("BADFREE2");
1401 #endif
1402 
1403     /*
1404      * Bump the number of free chunks.  If it becomes non-zero the zone
1405      * must be added back onto the appropriate list.  A fully allocated
1406      * zone that sees its first free is considered 'mature' and is placed
1407      * at the head, giving the system time to potentially free the remaining
1408      * entries even while other allocations are going on and making the zone
1409      * freeable.
1410      */
1411     if (z->z_NFree++ == 0) {
1412 	if (SlabFreeToTail)
1413 		TAILQ_INSERT_TAIL(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
1414 	else
1415 		TAILQ_INSERT_HEAD(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
1416     }
1417 
1418     --type->ks_use[z->z_Cpu].inuse;
1419     type->ks_use[z->z_Cpu].memuse -= z->z_ChunkSize;
1420 
1421     check_zone_free(slgd, z);
1422     logmemory_quick(free_end);
1423     crit_exit();
1424 }
1425 
1426 /*
1427  * Cleanup slabs which are hanging around due to RChunks or which are wholely
1428  * free and can be moved to the free list if not moved by other means.
1429  *
1430  * Called once every 10 seconds on all cpus.
1431  */
1432 void
1433 slab_cleanup(void)
1434 {
1435     SLGlobalData *slgd = &mycpu->gd_slab;
1436     SLZone *z;
1437     int i;
1438 
1439     crit_enter();
1440     for (i = 0; i < NZONES; ++i) {
1441 	if ((z = TAILQ_FIRST(&slgd->ZoneAry[i])) == NULL)
1442 		continue;
1443 
1444 	/*
1445 	 * Scan zones.
1446 	 */
1447 	while (z) {
1448 	    /*
1449 	     * Shift all RChunks to the end of the LChunks list.  This is
1450 	     * an O(1) operation.
1451 	     *
1452 	     * Then free the zone if possible.
1453 	     */
1454 	    clean_zone_rchunks(z);
1455 	    z = check_zone_free(slgd, z);
1456 	}
1457     }
1458     crit_exit();
1459 }
1460 
1461 #if defined(INVARIANTS)
1462 
1463 /*
1464  * Helper routines for sanity checks
1465  */
1466 static
1467 void
1468 chunk_mark_allocated(SLZone *z, void *chunk)
1469 {
1470     int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1471     uint32_t *bitptr;
1472 
1473     KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1474     KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1475 	    ("memory chunk %p bit index %d is illegal", chunk, bitdex));
1476     bitptr = &z->z_Bitmap[bitdex >> 5];
1477     bitdex &= 31;
1478     KASSERT((*bitptr & (1 << bitdex)) == 0,
1479 	    ("memory chunk %p is already allocated!", chunk));
1480     *bitptr |= 1 << bitdex;
1481 }
1482 
1483 static
1484 void
1485 chunk_mark_free(SLZone *z, void *chunk)
1486 {
1487     int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1488     uint32_t *bitptr;
1489 
1490     KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1491     KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1492 	    ("memory chunk %p bit index %d is illegal!", chunk, bitdex));
1493     bitptr = &z->z_Bitmap[bitdex >> 5];
1494     bitdex &= 31;
1495     KASSERT((*bitptr & (1 << bitdex)) != 0,
1496 	    ("memory chunk %p is already free!", chunk));
1497     *bitptr &= ~(1 << bitdex);
1498 }
1499 
1500 #endif
1501 
1502 /*
1503  * kmem_slab_alloc()
1504  *
1505  *	Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1506  *	specified alignment.  M_* flags are expected in the flags field.
1507  *
1508  *	Alignment must be a multiple of PAGE_SIZE.
1509  *
1510  *	NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1511  *	but when we move zalloc() over to use this function as its backend
1512  *	we will have to switch to kreserve/krelease and call reserve(0)
1513  *	after the new space is made available.
1514  *
1515  *	Interrupt code which has preempted other code is not allowed to
1516  *	use PQ_CACHE pages.  However, if an interrupt thread is run
1517  *	non-preemptively or blocks and then runs non-preemptively, then
1518  *	it is free to use PQ_CACHE pages.  <--- may not apply any longer XXX
1519  */
1520 static void *
1521 kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
1522 {
1523     vm_size_t i;
1524     vm_offset_t addr;
1525     int count, vmflags, base_vmflags;
1526     vm_page_t mbase = NULL;
1527     vm_page_t m;
1528     thread_t td;
1529 
1530     size = round_page(size);
1531     addr = vm_map_min(&kernel_map);
1532 
1533     count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1534     crit_enter();
1535     vm_map_lock(&kernel_map);
1536     if (vm_map_findspace(&kernel_map, addr, size, align, 0, &addr)) {
1537 	vm_map_unlock(&kernel_map);
1538 	if ((flags & M_NULLOK) == 0)
1539 	    panic("kmem_slab_alloc(): kernel_map ran out of space!");
1540 	vm_map_entry_release(count);
1541 	crit_exit();
1542 	return(NULL);
1543     }
1544 
1545     /*
1546      * kernel_object maps 1:1 to kernel_map.
1547      */
1548     vm_object_hold(&kernel_object);
1549     vm_object_reference_locked(&kernel_object);
1550     vm_map_insert(&kernel_map, &count,
1551 		  &kernel_object, NULL,
1552 		  addr, addr, addr + size,
1553 		  VM_MAPTYPE_NORMAL,
1554 		  VM_SUBSYS_KMALLOC,
1555 		  VM_PROT_ALL, VM_PROT_ALL, 0);
1556     vm_object_drop(&kernel_object);
1557     vm_map_set_wired_quick(&kernel_map, addr, size, &count);
1558     vm_map_unlock(&kernel_map);
1559 
1560     td = curthread;
1561 
1562     base_vmflags = 0;
1563     if (flags & M_ZERO)
1564         base_vmflags |= VM_ALLOC_ZERO;
1565     if (flags & M_USE_RESERVE)
1566 	base_vmflags |= VM_ALLOC_SYSTEM;
1567     if (flags & M_USE_INTERRUPT_RESERVE)
1568         base_vmflags |= VM_ALLOC_INTERRUPT;
1569     if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) {
1570 	panic("kmem_slab_alloc: bad flags %08x (%p)",
1571 	      flags, ((int **)&size)[-1]);
1572     }
1573 
1574     /*
1575      * Allocate the pages.  Do not map them yet.  VM_ALLOC_NORMAL can only
1576      * be set if we are not preempting.
1577      *
1578      * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1579      * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1580      * implied in this case), though I'm not sure if we really need to
1581      * do that.
1582      */
1583     vmflags = base_vmflags;
1584     if (flags & M_WAITOK) {
1585 	if (td->td_preempted)
1586 	    vmflags |= VM_ALLOC_SYSTEM;
1587 	else
1588 	    vmflags |= VM_ALLOC_NORMAL;
1589     }
1590 
1591     vm_object_hold(&kernel_object);
1592     for (i = 0; i < size; i += PAGE_SIZE) {
1593 	m = vm_page_alloc(&kernel_object, OFF_TO_IDX(addr + i), vmflags);
1594 	if (i == 0)
1595 		mbase = m;
1596 
1597 	/*
1598 	 * If the allocation failed we either return NULL or we retry.
1599 	 *
1600 	 * If M_WAITOK is specified we wait for more memory and retry.
1601 	 * If M_WAITOK is specified from a preemption we yield instead of
1602 	 * wait.  Livelock will not occur because the interrupt thread
1603 	 * will not be preempting anyone the second time around after the
1604 	 * yield.
1605 	 */
1606 	if (m == NULL) {
1607 	    if (flags & M_WAITOK) {
1608 		if (td->td_preempted) {
1609 		    lwkt_switch();
1610 		} else {
1611 		    vm_wait(0);
1612 		}
1613 		i -= PAGE_SIZE;	/* retry */
1614 		continue;
1615 	    }
1616 	    break;
1617 	}
1618     }
1619 
1620     /*
1621      * Check and deal with an allocation failure
1622      */
1623     if (i != size) {
1624 	while (i != 0) {
1625 	    i -= PAGE_SIZE;
1626 	    m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i));
1627 	    /* page should already be busy */
1628 	    vm_page_free(m);
1629 	}
1630 	vm_map_lock(&kernel_map);
1631 	vm_map_delete(&kernel_map, addr, addr + size, &count);
1632 	vm_map_unlock(&kernel_map);
1633 	vm_object_drop(&kernel_object);
1634 
1635 	vm_map_entry_release(count);
1636 	crit_exit();
1637 	return(NULL);
1638     }
1639 
1640     /*
1641      * Success!
1642      *
1643      * NOTE: The VM pages are still busied.  mbase points to the first one
1644      *	     but we have to iterate via vm_page_next()
1645      */
1646     vm_object_drop(&kernel_object);
1647     crit_exit();
1648 
1649     /*
1650      * Enter the pages into the pmap and deal with M_ZERO.
1651      */
1652     m = mbase;
1653     i = 0;
1654 
1655     while (i < size) {
1656 	/*
1657 	 * page should already be busy
1658 	 */
1659 	m->valid = VM_PAGE_BITS_ALL;
1660 	vm_page_wire(m);
1661 	pmap_enter(&kernel_pmap, addr + i, m,
1662 		   VM_PROT_ALL | VM_PROT_NOSYNC, 1, NULL);
1663 	if (flags & M_ZERO)
1664 		pagezero((char *)addr + i);
1665 	KKASSERT(m->flags & (PG_WRITEABLE | PG_MAPPED));
1666 	vm_page_flag_set(m, PG_REFERENCED);
1667 	vm_page_wakeup(m);
1668 
1669 	i += PAGE_SIZE;
1670 	vm_object_hold(&kernel_object);
1671 	m = vm_page_next(m);
1672 	vm_object_drop(&kernel_object);
1673     }
1674     smp_invltlb();
1675     vm_map_entry_release(count);
1676     atomic_add_long(&SlabsAllocated, 1);
1677     return((void *)addr);
1678 }
1679 
1680 /*
1681  * kmem_slab_free()
1682  */
1683 static void
1684 kmem_slab_free(void *ptr, vm_size_t size)
1685 {
1686     crit_enter();
1687     vm_map_remove(&kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);
1688     atomic_add_long(&SlabsFreed, 1);
1689     crit_exit();
1690 }
1691 
1692 void *
1693 kmalloc_cachealign(unsigned long size_alloc, struct malloc_type *type,
1694     int flags)
1695 {
1696 #if (__VM_CACHELINE_SIZE == 32)
1697 #define CAN_CACHEALIGN(sz)	((sz) >= 256)
1698 #elif (__VM_CACHELINE_SIZE == 64)
1699 #define CAN_CACHEALIGN(sz)	((sz) >= 512)
1700 #elif (__VM_CACHELINE_SIZE == 128)
1701 #define CAN_CACHEALIGN(sz)	((sz) >= 1024)
1702 #else
1703 #error "unsupported cacheline size"
1704 #endif
1705 
1706 	void *ret;
1707 
1708 	if (size_alloc < __VM_CACHELINE_SIZE)
1709 		size_alloc = __VM_CACHELINE_SIZE;
1710 	else if (!CAN_CACHEALIGN(size_alloc))
1711 		flags |= M_POWEROF2;
1712 
1713 	ret = kmalloc(size_alloc, type, flags);
1714 	KASSERT(((uintptr_t)ret & (__VM_CACHELINE_SIZE - 1)) == 0,
1715 	    ("%p(%lu) not cacheline %d aligned",
1716 	     ret, size_alloc, __VM_CACHELINE_SIZE));
1717 	return ret;
1718 
1719 #undef CAN_CACHEALIGN
1720 }
1721