2 * KERN_SLABALLOC.C - Kernel SLAB memory allocator
4 * Copyright (c) 2003,2004,2010-2019 The DragonFly Project.
7 * This code is derived from software contributed to The DragonFly Project
8 * by Matthew Dillon <dillon@backplane.com>
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
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
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.
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
37 * This module implements a slab allocator drop-in replacement for the
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.
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*
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.
60 * XXX Balancing is needed between cpus. Balance will be handled through
61 * asynchronous IPIs primarily by reassigning the z_Cpu ownership of chunks.
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.
66 * Alloc Size Chunking Number of zones
76 * (if PAGE_SIZE is 4K the maximum zone allocation is 16383)
78 * Allocations >= ZoneLimit go directly to kmem.
79 * (n * PAGE_SIZE, n > 2) allocations go directly to kmem.
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).
88 * API REQUIREMENTS AND SIDE EFFECTS
90 * To operate as a drop-in replacement to the FreeBSD-4.x malloc() we
91 * have remained compatible with the following API requirements:
93 * + malloc(0) is allowed and returns non-NULL (ahc driver)
94 * + ability to allocate arbitrarily large chunks of memory
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>
110 #include <sys/kthread.h>
111 #include <sys/malloc.h>
114 #include <vm/vm_param.h>
115 #include <vm/vm_kern.h>
116 #include <vm/vm_extern.h>
117 #include <vm/vm_object.h>
119 #include <vm/vm_map.h>
120 #include <vm/vm_page.h>
121 #include <vm/vm_pageout.h>
123 #include <machine/cpu.h>
125 #include <sys/thread2.h>
126 #include <vm/vm_page2.h>
128 #if (__VM_CACHELINE_SIZE == 32)
129 #define CAN_CACHEALIGN(sz) ((sz) >= 256)
130 #elif (__VM_CACHELINE_SIZE == 64)
131 #define CAN_CACHEALIGN(sz) ((sz) >= 512)
132 #elif (__VM_CACHELINE_SIZE == 128)
133 #define CAN_CACHEALIGN(sz) ((sz) >= 1024)
135 #error "unsupported cacheline size"
138 #define btokup(z) (&pmap_kvtom((vm_offset_t)(z))->ku_pagecnt)
140 #define MEMORY_STRING "ptr=%p type=%p size=%lu flags=%04x"
141 #define MEMORY_ARGS void *ptr, void *type, unsigned long size, int flags
143 #if !defined(KTR_MEMORY)
144 #define KTR_MEMORY KTR_ALL
146 KTR_INFO_MASTER(memory);
147 KTR_INFO(KTR_MEMORY, memory, malloc_beg, 0, "malloc begin");
148 KTR_INFO(KTR_MEMORY, memory, malloc_end, 1, MEMORY_STRING, MEMORY_ARGS);
149 KTR_INFO(KTR_MEMORY, memory, free_zero, 2, MEMORY_STRING, MEMORY_ARGS);
150 KTR_INFO(KTR_MEMORY, memory, free_ovsz, 3, MEMORY_STRING, MEMORY_ARGS);
151 KTR_INFO(KTR_MEMORY, memory, free_ovsz_delayed, 4, MEMORY_STRING, MEMORY_ARGS);
152 KTR_INFO(KTR_MEMORY, memory, free_chunk, 5, MEMORY_STRING, MEMORY_ARGS);
153 KTR_INFO(KTR_MEMORY, memory, free_request, 6, MEMORY_STRING, MEMORY_ARGS);
154 KTR_INFO(KTR_MEMORY, memory, free_rem_beg, 7, MEMORY_STRING, MEMORY_ARGS);
155 KTR_INFO(KTR_MEMORY, memory, free_rem_end, 8, MEMORY_STRING, MEMORY_ARGS);
156 KTR_INFO(KTR_MEMORY, memory, free_beg, 9, "free begin");
157 KTR_INFO(KTR_MEMORY, memory, free_end, 10, "free end");
159 #define logmemory(name, ptr, type, size, flags) \
160 KTR_LOG(memory_ ## name, ptr, type, size, flags)
161 #define logmemory_quick(name) \
162 KTR_LOG(memory_ ## name)
165 * Fixed globals (not per-cpu)
167 __read_frequently static int ZoneSize;
168 __read_frequently static int ZoneLimit;
169 __read_frequently static int ZonePageCount;
170 __read_frequently static uintptr_t ZoneMask;
171 __read_frequently struct malloc_type *kmemstatistics; /* exported to vmstat */
173 #if defined(INVARIANTS)
174 static void chunk_mark_allocated(SLZone *z, void *chunk);
175 static void chunk_mark_free(SLZone *z, void *chunk);
177 #define chunk_mark_allocated(z, chunk)
178 #define chunk_mark_free(z, chunk)
182 * Misc constants. Note that allocations that are exact multiples of
183 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module.
185 #define ZONE_RELS_THRESH 32 /* threshold number of zones */
189 * The WEIRD_ADDR is used as known text to copy into free objects to
190 * try to create deterministic failure cases if the data is accessed after
193 #define WEIRD_ADDR 0xdeadc0de
195 #define ZERO_LENGTH_PTR ((void *)-8)
198 * Misc global malloc buckets
201 MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches");
202 MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory");
203 MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers");
204 MALLOC_DEFINE(M_DRM, "m_drm", "DRM memory allocations");
206 MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options");
207 MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery");
210 * Initialize the slab memory allocator. We have to choose a zone size based
211 * on available physical memory. We choose a zone side which is approximately
212 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of
213 * 128K. The zone size is limited to the bounds set in slaballoc.h
214 * (typically 32K min, 128K max).
216 static void kmeminit(void *dummy);
217 static void kmemfinishinit(void *dummy);
221 SYSINIT(kmem1, SI_BOOT1_ALLOCATOR, SI_ORDER_FIRST, kmeminit, NULL);
222 SYSINIT(kmem2, SI_BOOT2_POST_SMP, SI_ORDER_FIRST, kmemfinishinit, NULL);
226 * If enabled any memory allocated without M_ZERO is initialized to -1.
228 __read_frequently static int use_malloc_pattern;
229 SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW,
230 &use_malloc_pattern, 0,
231 "Initialize memory to -1 if M_ZERO not specified");
233 __read_frequently static int32_t weirdary[16];
234 __read_frequently static int use_weird_array;
235 SYSCTL_INT(_debug, OID_AUTO, use_weird_array, CTLFLAG_RW,
237 "Initialize memory to weird values on kfree()");
240 __read_frequently static int ZoneRelsThresh = ZONE_RELS_THRESH;
241 SYSCTL_INT(_kern, OID_AUTO, zone_cache, CTLFLAG_RW, &ZoneRelsThresh, 0, "");
242 __read_frequently static int kzone_pollfreq = 1;
243 SYSCTL_INT(_kern, OID_AUTO, kzone_pollfreq, CTLFLAG_RW, &kzone_pollfreq, 0, "");
245 static struct spinlock kmemstat_spin =
246 SPINLOCK_INITIALIZER(&kmemstat_spin, "malinit");
247 static struct malloc_type *kmemstat_poll;
250 * Returns the kernel memory size limit for the purposes of initializing
251 * various subsystem caches. The smaller of available memory and the KVM
252 * memory space is returned.
254 * The size in megabytes is returned.
261 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE;
262 if (limsize > KvaSize)
264 return (limsize / (1024 * 1024));
268 kmeminit(void *dummy)
276 limsize = kmem_lim_size();
277 usesize = (int)(limsize * 1024); /* convert to KB */
280 * If the machine has a large KVM space and more than 8G of ram,
281 * double the zone release threshold to reduce SMP invalidations.
282 * If more than 16G of ram, do it again.
284 * The BIOS eats a little ram so add some slop. We want 8G worth of
285 * memory sticks to trigger the first adjustment.
287 if (ZoneRelsThresh == ZONE_RELS_THRESH) {
288 if (limsize >= 7 * 1024)
290 if (limsize >= 15 * 1024)
292 if (limsize >= 31 * 1024)
294 if (limsize >= 63 * 1024)
296 if (limsize >= 127 * 1024)
301 * Calculate the zone size. This typically calculates to
302 * ZALLOC_MAX_ZONE_SIZE
304 ZoneSize = ZALLOC_MIN_ZONE_SIZE;
305 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize)
307 ZoneLimit = ZoneSize / 4;
308 if (ZoneLimit > ZALLOC_ZONE_LIMIT)
309 ZoneLimit = ZALLOC_ZONE_LIMIT;
310 ZoneMask = ~(uintptr_t)(ZoneSize - 1);
311 ZonePageCount = ZoneSize / PAGE_SIZE;
314 for (i = 0; i < NELEM(weirdary); ++i)
315 weirdary[i] = WEIRD_ADDR;
318 ZeroPage = kmem_slab_alloc(PAGE_SIZE, PAGE_SIZE, M_WAITOK|M_ZERO);
321 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024);
325 * Once we know how many cpus are configured reduce ZoneRelsThresh
326 * based on multiples of 32 cpu threads.
329 kmemfinishinit(void *dummy)
332 ZoneRelsThresh = ZoneRelsThresh * 32 / ncpus;
336 * (low level) Initialize slab-related elements in the globaldata structure.
338 * Occurs after kmeminit().
341 slab_gdinit(globaldata_t gd)
347 for (i = 0; i < NZONES; ++i)
348 TAILQ_INIT(&slgd->ZoneAry[i]);
349 TAILQ_INIT(&slgd->FreeZones);
350 TAILQ_INIT(&slgd->FreeOvZones);
354 * Initialize a malloc type tracking structure.
357 malloc_init(void *data)
359 struct malloc_type *type = data;
360 struct kmalloc_use *use;
364 if (type->ks_magic != M_MAGIC)
365 panic("malloc type lacks magic");
367 if (type->ks_limit != 0)
370 if (vmstats.v_page_count == 0)
371 panic("malloc_init not allowed before vm init");
373 limsize = kmem_lim_size() * (1024 * 1024);
374 type->ks_limit = limsize / 10;
375 if (type->ks_flags & KSF_OBJSIZE)
376 malloc_mgt_init(type, &type->ks_mgt, type->ks_objsize);
379 use = &type->ks_use0;
381 use = kmalloc(ncpus * sizeof(*use), M_TEMP, M_WAITOK | M_ZERO);
382 if (type->ks_flags & KSF_OBJSIZE) {
383 for (n = 0; n < ncpus; ++n)
384 malloc_mgt_init(type, &use[n].mgt, type->ks_objsize);
387 spin_lock(&kmemstat_spin);
388 type->ks_next = kmemstatistics;
390 kmemstatistics = type;
391 spin_unlock(&kmemstat_spin);
395 malloc_uninit(void *data)
397 struct malloc_type *type = data;
398 struct malloc_type *t;
404 if (type->ks_magic != M_MAGIC)
405 panic("malloc type lacks magic");
407 if (vmstats.v_page_count == 0)
408 panic("malloc_uninit not allowed before vm init");
410 if (type->ks_limit == 0)
411 panic("malloc_uninit on uninitialized type");
413 /* Make sure that all pending kfree()s are finished. */
414 lwkt_synchronize_ipiqs("muninit");
417 * Remove from the kmemstatistics list, blocking if the removal races
418 * the kmalloc poller.
420 * Advance kmemstat_poll if necessary.
422 spin_lock(&kmemstat_spin);
423 while (type->ks_flags & KSF_POLLING)
424 ssleep(type, &kmemstat_spin, 0, "kmuninit", 0);
426 if (kmemstat_poll == type)
427 kmemstat_poll = type->ks_next;
429 if (kmemstatistics == type) {
430 kmemstatistics = type->ks_next;
432 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) {
433 if (t->ks_next == type) {
434 t->ks_next = type->ks_next;
439 type->ks_next = NULL;
441 spin_unlock(&kmemstat_spin);
444 * memuse is only correct in aggregation. Due to memory being allocated
445 * on one cpu and freed on another individual array entries may be
446 * negative or positive (canceling each other out).
451 for (i = 0; i < ncpus; ++i) {
453 ttl += type->ks_use[i].memuse;
455 if (type->ks_flags & KSF_OBJSIZE)
456 malloc_mgt_uninit(type, &type->ks_use[i].mgt);
458 if (type->ks_flags & KSF_OBJSIZE)
459 malloc_mgt_uninit(type, &type->ks_mgt);
462 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n",
463 ttl, type->ks_shortdesc, i);
467 if (type->ks_use != &type->ks_use0) {
468 kfree(type->ks_use, M_TEMP);
474 * Slowly polls all kmalloc zones for cleanup
477 kmalloc_poller_thread(void)
479 struct malloc_type *type;
483 * Very slow poll by default, adjustable with sysctl
487 sticks = kzone_pollfreq;
490 sticks = hz / sticks + 1; /* approximate */
492 sticks = hz; /* safety */
493 tsleep((caddr_t)&sticks, 0, "kmslp", sticks);
496 * [re]poll one zone each period.
498 spin_lock(&kmemstat_spin);
499 type = kmemstat_poll;
502 type = kmemstatistics;
504 atomic_set_int(&type->ks_flags, KSF_POLLING);
505 spin_unlock(&kmemstat_spin);
506 if (malloc_mgt_poll(type)) {
507 spin_lock(&kmemstat_spin);
508 kmemstat_poll = type->ks_next;
510 spin_lock(&kmemstat_spin);
512 atomic_clear_int(&type->ks_flags, KSF_POLLING);
515 kmemstat_poll = NULL;
517 spin_unlock(&kmemstat_spin);
521 static struct thread *kmalloc_poller_td;
522 static struct kproc_desc kmalloc_poller_kp = {
524 kmalloc_poller_thread,
527 SYSINIT(kmalloc_polller, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST,
528 kproc_start, &kmalloc_poller_kp);
531 * Reinitialize all installed malloc regions after ncpus has been
532 * determined. type->ks_use0 is initially set to &type->ks_use0,
533 * this function will dynamically allocate it as appropriate for ncpus.
536 malloc_reinit_ncpus(void)
538 struct malloc_type *t;
539 struct kmalloc_use *use;
543 * If only one cpu we can leave ks_use set to ks_use0
549 * Expand ks_use for all kmalloc blocks
551 for (t = kmemstatistics; t; t = t->ks_next) {
552 KKASSERT(t->ks_use == &t->ks_use0);
553 t->ks_use = kmalloc(sizeof(*use) * ncpus, M_TEMP, M_WAITOK|M_ZERO);
554 t->ks_use[0] = t->ks_use0;
555 if (t->ks_flags & KSF_OBJSIZE) {
556 malloc_mgt_relocate(&t->ks_use0.mgt, &t->ks_use[0].mgt);
557 for (n = 1; n < ncpus; ++n)
558 malloc_mgt_init(t, &t->ks_use[n].mgt, t->ks_objsize);
564 * Increase the kmalloc pool limit for the specified pool. No changes
565 * are the made if the pool would shrink.
568 kmalloc_raise_limit(struct malloc_type *type, size_t bytes)
570 KKASSERT(type->ks_limit != 0);
573 if (type->ks_limit < bytes)
574 type->ks_limit = bytes;
578 kmalloc_set_unlimited(struct malloc_type *type)
580 type->ks_limit = kmem_lim_size() * (1024 * 1024);
584 * Dynamically create a malloc pool. This function is a NOP if *typep is
588 kmalloc_create(struct malloc_type **typep, const char *descr)
590 struct malloc_type *type;
592 if (*typep == NULL) {
593 type = kmalloc(sizeof(*type), M_TEMP, M_WAITOK | M_ZERO);
594 type->ks_magic = M_MAGIC;
595 type->ks_shortdesc = descr;
602 _kmalloc_create_obj(struct malloc_type **typep, const char *descr,
605 struct malloc_type *type;
607 if (*typep == NULL) {
608 type = kmalloc(sizeof(*type), M_TEMP, M_WAITOK | M_ZERO);
609 type->ks_magic = M_MAGIC;
610 type->ks_shortdesc = descr;
611 type->ks_flags = KSF_OBJSIZE;
612 type->ks_objsize = __VM_CACHELINE_ALIGN(objsize);
619 * Destroy a dynamically created malloc pool. This function is a NOP if
620 * the pool has already been destroyed.
622 * WARNING! For kmalloc_obj's, the exis state for related slabs is ignored,
623 * only call once all references are 100% known to be gone.
626 kmalloc_destroy(struct malloc_type **typep)
628 if (*typep != NULL) {
629 malloc_uninit(*typep);
630 kfree(*typep, M_TEMP);
636 * Calculate the zone index for the allocation request size and set the
637 * allocation request size to that particular zone's chunk size.
640 zoneindex(unsigned long *bytes, unsigned long *align)
642 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */
645 *bytes = n = (n + 7) & ~7;
647 return(n / 8 - 1); /* 8 byte chunks, 16 zones */
650 *bytes = n = (n + 15) & ~15;
656 *bytes = n = (n + 31) & ~31;
661 *bytes = n = (n + 63) & ~63;
666 *bytes = n = (n + 127) & ~127;
668 return(n / 128 + 31);
671 *bytes = n = (n + 255) & ~255;
673 return(n / 256 + 39);
675 *bytes = n = (n + 511) & ~511;
677 return(n / 512 + 47);
679 #if ZALLOC_ZONE_LIMIT > 8192
681 *bytes = n = (n + 1023) & ~1023;
683 return(n / 1024 + 55);
686 #if ZALLOC_ZONE_LIMIT > 16384
688 *bytes = n = (n + 2047) & ~2047;
690 return(n / 2048 + 63);
693 panic("Unexpected byte count %d", n);
698 clean_zone_rchunks(SLZone *z)
702 while ((bchunk = z->z_RChunks) != NULL) {
704 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) {
705 *z->z_LChunksp = bchunk;
707 chunk_mark_free(z, bchunk);
708 z->z_LChunksp = &bchunk->c_Next;
709 bchunk = bchunk->c_Next;
719 * If the zone becomes totally free and is not the only zone listed for a
720 * chunk size we move it to the FreeZones list. We always leave at least
721 * one zone per chunk size listed, even if it is freeable.
723 * Do not move the zone if there is an IPI in_flight (z_RCount != 0),
724 * otherwise MP races can result in our free_remote code accessing a
725 * destroyed zone. The remote end interlocks z_RCount with z_RChunks
726 * so one has to test both z_NFree and z_RCount.
728 * Since this code can be called from an IPI callback, do *NOT* try to mess
729 * with kernel_map here. Hysteresis will be performed at kmalloc() time.
731 static __inline SLZone *
732 check_zone_free(SLGlobalData *slgd, SLZone *z)
736 znext = TAILQ_NEXT(z, z_Entry);
737 if (z->z_NFree == z->z_NMax && z->z_RCount == 0 &&
738 (TAILQ_FIRST(&slgd->ZoneAry[z->z_ZoneIndex]) != z || znext)) {
741 TAILQ_REMOVE(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
744 TAILQ_INSERT_HEAD(&slgd->FreeZones, z, z_Entry);
754 * Used to debug memory corruption issues. Record up to (typically 32)
755 * allocation sources for this zone (for a particular chunk size).
759 slab_record_source(SLZone *z, const char *file, int line)
762 int b = line & (SLAB_DEBUG_ENTRIES - 1);
766 if (z->z_Sources[i].file == file && z->z_Sources[i].line == line)
768 if (z->z_Sources[i].file == NULL)
770 i = (i + 1) & (SLAB_DEBUG_ENTRIES - 1);
772 z->z_Sources[i].file = file;
773 z->z_Sources[i].line = line;
778 static __inline unsigned long
779 powerof2_size(unsigned long size)
783 if (size == 0 || powerof2(size))
791 * kmalloc() (SLAB ALLOCATOR)
793 * Allocate memory via the slab allocator. If the request is too large,
794 * or if it page-aligned beyond a certain size, we fall back to the
795 * KMEM subsystem. A SLAB tracking descriptor must be specified, use
796 * &SlabMisc if you don't care.
798 * M_RNOWAIT - don't block.
799 * M_NULLOK - return NULL instead of blocking.
800 * M_ZERO - zero the returned memory.
801 * M_USE_RESERVE - allow greater drawdown of the free list
802 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted
803 * M_POWEROF2 - roundup size to the nearest power of 2
808 /* don't let kmalloc macro mess up function declaration */
813 _kmalloc_debug(unsigned long size, struct malloc_type *type, int flags,
814 const char *file, int line)
817 _kmalloc(unsigned long size, struct malloc_type *type, int flags)
823 struct globaldata *gd;
830 logmemory_quick(malloc_beg);
835 * XXX silly to have this in the critical path.
837 KKASSERT(type->ks_limit != 0);
838 ++type->ks_use[gd->gd_cpuid].calls;
841 * Flagged for cache-alignment
843 if (flags & M_CACHEALIGN) {
844 if (size < __VM_CACHELINE_SIZE)
845 size = __VM_CACHELINE_SIZE;
846 else if (!CAN_CACHEALIGN(size))
851 * Flagged to force nearest power-of-2 (higher or same)
853 if (flags & M_POWEROF2)
854 size = powerof2_size(size);
857 * Handle the case where the limit is reached. Panic if we can't return
858 * NULL. The original malloc code looped, but this tended to
859 * simply deadlock the computer.
861 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used
862 * to determine if a more complete limit check should be done. The
863 * actual memory use is tracked via ks_use[cpu].memuse.
865 while (type->ks_loosememuse >= type->ks_limit) {
869 for (i = ttl = 0; i < ncpus; ++i)
870 ttl += type->ks_use[i].memuse;
871 type->ks_loosememuse = ttl; /* not MP synchronized */
872 if ((ssize_t)ttl < 0) /* deal with occassional race */
874 if (ttl >= type->ks_limit) {
875 if (flags & M_NULLOK) {
876 logmemory(malloc_end, NULL, type, size, flags);
879 panic("%s: malloc limit exceeded", type->ks_shortdesc);
884 * Handle the degenerate size == 0 case. Yes, this does happen.
885 * Return a special pointer. This is to maintain compatibility with
886 * the original malloc implementation. Certain devices, such as the
887 * adaptec driver, not only allocate 0 bytes, they check for NULL and
888 * also realloc() later on. Joy.
891 logmemory(malloc_end, ZERO_LENGTH_PTR, type, size, flags);
892 return(ZERO_LENGTH_PTR);
896 * Handle hysteresis from prior frees here in malloc(). We cannot
897 * safely manipulate the kernel_map in free() due to free() possibly
898 * being called via an IPI message or from sensitive interrupt code.
900 * NOTE: ku_pagecnt must be cleared before we free the slab or we
901 * might race another cpu allocating the kva and setting
904 while (slgd->NFreeZones > ZoneRelsThresh && (flags & M_RNOWAIT) == 0) {
906 if (slgd->NFreeZones > ZoneRelsThresh) { /* crit sect race */
909 z = TAILQ_LAST(&slgd->FreeZones, SLZoneList);
911 TAILQ_REMOVE(&slgd->FreeZones, z, z_Entry);
915 kmem_slab_free(z, ZoneSize); /* may block */
921 * XXX handle oversized frees that were queued from kfree().
923 while (TAILQ_FIRST(&slgd->FreeOvZones) && (flags & M_RNOWAIT) == 0) {
925 if ((z = TAILQ_LAST(&slgd->FreeOvZones, SLZoneList)) != NULL) {
928 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC);
929 TAILQ_REMOVE(&slgd->FreeOvZones, z, z_Entry);
930 tsize = z->z_ChunkSize;
931 kmem_slab_free(z, tsize); /* may block */
937 * Handle large allocations directly. There should not be very many of
938 * these so performance is not a big issue.
940 * The backend allocator is pretty nasty on a SMP system. Use the
941 * slab allocator for one and two page-sized chunks even though we lose
942 * some efficiency. XXX maybe fix mmio and the elf loader instead.
944 if (size >= ZoneLimit || ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) {
947 size = round_page(size);
948 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags);
950 logmemory(malloc_end, NULL, type, size, flags);
953 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */
954 flags |= M_PASSIVE_ZERO;
956 *kup = size / PAGE_SIZE;
962 * Attempt to allocate out of an existing zone. First try the free list,
963 * then allocate out of unallocated space. If we find a good zone move
964 * it to the head of the list so later allocations find it quickly
965 * (we might have thousands of zones in the list).
967 * Note: zoneindex() will panic of size is too large.
969 zi = zoneindex(&size, &align);
970 KKASSERT(zi < NZONES);
973 if ((z = TAILQ_LAST(&slgd->ZoneAry[zi], SLZoneList)) != NULL) {
975 * Locate a chunk - we have to have at least one. If this is the
976 * last chunk go ahead and do the work to retrieve chunks freed
977 * from remote cpus, and if the zone is still empty move it off
980 if (--z->z_NFree <= 0) {
981 KKASSERT(z->z_NFree == 0);
984 * WARNING! This code competes with other cpus. It is ok
985 * for us to not drain RChunks here but we might as well, and
986 * it is ok if more accumulate after we're done.
988 * Set RSignal before pulling rchunks off, indicating that we
989 * will be moving ourselves off of the ZoneAry. Remote ends will
990 * read RSignal before putting rchunks on thus interlocking
991 * their IPI signaling.
993 if (z->z_RChunks == NULL)
994 atomic_swap_int(&z->z_RSignal, 1);
996 clean_zone_rchunks(z);
999 * Remove from the zone list if no free chunks remain.
1002 if (z->z_NFree == 0) {
1003 TAILQ_REMOVE(&slgd->ZoneAry[zi], z, z_Entry);
1010 * Fast path, we have chunks available in z_LChunks.
1012 chunk = z->z_LChunks;
1014 chunk_mark_allocated(z, chunk);
1015 z->z_LChunks = chunk->c_Next;
1016 if (z->z_LChunks == NULL)
1017 z->z_LChunksp = &z->z_LChunks;
1019 slab_record_source(z, file, line);
1025 * No chunks are available in LChunks, the free chunk MUST be
1026 * in the never-before-used memory area, controlled by UIndex.
1028 * The consequences are very serious if our zone got corrupted so
1029 * we use an explicit panic rather than a KASSERT.
1031 if (z->z_UIndex + 1 != z->z_NMax)
1036 if (z->z_UIndex == z->z_UEndIndex)
1037 panic("slaballoc: corrupted zone");
1039 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
1040 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
1042 flags |= M_PASSIVE_ZERO;
1044 chunk_mark_allocated(z, chunk);
1046 slab_record_source(z, file, line);
1052 * If all zones are exhausted we need to allocate a new zone for this
1053 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see
1054 * UAlloc use above in regards to M_ZERO. Note that when we are reusing
1055 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and
1056 * we do not pre-zero it because we do not want to mess up the L1 cache.
1058 * At least one subsystem, the tty code (see CROUND) expects power-of-2
1059 * allocations to be power-of-2 aligned. We maintain compatibility by
1060 * adjusting the base offset below.
1066 if ((z = TAILQ_FIRST(&slgd->FreeZones)) != NULL) {
1067 TAILQ_REMOVE(&slgd->FreeZones, z, z_Entry);
1069 bzero(z, sizeof(SLZone));
1070 z->z_Flags |= SLZF_UNOTZEROD;
1072 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO);
1078 * How big is the base structure?
1080 #if defined(INVARIANTS)
1082 * Make room for z_Bitmap. An exact calculation is somewhat more
1083 * complicated so don't make an exact calculation.
1085 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]);
1086 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8);
1088 off = sizeof(SLZone);
1092 * Guarentee power-of-2 alignment for power-of-2-sized chunks.
1093 * Otherwise properly align the data according to the chunk size.
1097 off = roundup2(off, align);
1099 z->z_Magic = ZALLOC_SLAB_MAGIC;
1100 z->z_ZoneIndex = zi;
1101 z->z_NMax = (ZoneSize - off) / size;
1102 z->z_NFree = z->z_NMax - 1;
1103 z->z_BasePtr = (char *)z + off;
1104 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax;
1105 z->z_ChunkSize = size;
1107 z->z_Cpu = gd->gd_cpuid;
1108 z->z_LChunksp = &z->z_LChunks;
1110 bcopy(z->z_Sources, z->z_AltSources, sizeof(z->z_Sources));
1111 bzero(z->z_Sources, sizeof(z->z_Sources));
1113 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size);
1114 TAILQ_INSERT_HEAD(&slgd->ZoneAry[zi], z, z_Entry);
1115 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) {
1116 flags &= ~M_ZERO; /* already zero'd */
1117 flags |= M_PASSIVE_ZERO;
1120 *kup = -(z->z_Cpu + 1); /* -1 to -(N+1) */
1121 chunk_mark_allocated(z, chunk);
1123 slab_record_source(z, file, line);
1127 * Slide the base index for initial allocations out of the next
1128 * zone we create so we do not over-weight the lower part of the
1129 * cpu memory caches.
1131 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE)
1132 & (ZALLOC_MAX_ZONE_SIZE - 1);
1136 ++type->ks_use[gd->gd_cpuid].inuse;
1137 type->ks_use[gd->gd_cpuid].memuse += size;
1138 type->ks_use[gd->gd_cpuid].loosememuse += size;
1139 if (type->ks_use[gd->gd_cpuid].loosememuse >= ZoneSize) {
1140 /* not MP synchronized */
1141 type->ks_loosememuse += type->ks_use[gd->gd_cpuid].loosememuse;
1142 type->ks_use[gd->gd_cpuid].loosememuse = 0;
1149 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) {
1150 if (use_malloc_pattern) {
1151 for (i = 0; i < size; i += sizeof(int)) {
1152 *(int *)((char *)chunk + i) = -1;
1155 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */
1158 logmemory(malloc_end, chunk, type, size, flags);
1162 logmemory(malloc_end, NULL, type, size, flags);
1167 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE)
1169 * Generally speaking this routine is not called very often and we do
1170 * not attempt to optimize it beyond reusing the same pointer if the
1171 * new size fits within the chunking of the old pointer's zone.
1175 krealloc_debug(void *ptr, unsigned long size,
1176 struct malloc_type *type, int flags,
1177 const char *file, int line)
1180 krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags)
1183 unsigned long osize;
1184 unsigned long align;
1189 KKASSERT((flags & M_ZERO) == 0); /* not supported */
1191 if (ptr == NULL || ptr == ZERO_LENGTH_PTR)
1192 return(_kmalloc_debug(size, type, flags, file, line));
1199 * Handle oversized allocations. XXX we really should require that a
1200 * size be passed to free() instead of this nonsense.
1204 osize = *kup << PAGE_SHIFT;
1205 if (osize == round_page(size))
1207 if ((nptr = _kmalloc_debug(size, type, flags, file, line)) == NULL)
1209 bcopy(ptr, nptr, min(size, osize));
1215 * Get the original allocation's zone. If the new request winds up
1216 * using the same chunk size we do not have to do anything.
1218 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1221 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1224 * Allocate memory for the new request size. Note that zoneindex has
1225 * already adjusted the request size to the appropriate chunk size, which
1226 * should optimize our bcopy(). Then copy and return the new pointer.
1228 * Resizing a non-power-of-2 allocation to a power-of-2 size does not
1229 * necessary align the result.
1231 * We can only zoneindex (to align size to the chunk size) if the new
1232 * size is not too large.
1234 if (size < ZoneLimit) {
1235 zoneindex(&size, &align);
1236 if (z->z_ChunkSize == size)
1239 if ((nptr = _kmalloc_debug(size, type, flags, file, line)) == NULL)
1241 bcopy(ptr, nptr, min(size, z->z_ChunkSize));
1247 * Return the kmalloc limit for this type, in bytes.
1250 kmalloc_limit(struct malloc_type *type)
1252 KKASSERT(type->ks_limit != 0);
1253 return(type->ks_limit);
1257 * Allocate a copy of the specified string.
1259 * (MP SAFE) (MAY BLOCK)
1263 kstrdup_debug(const char *str, struct malloc_type *type,
1264 const char *file, int line)
1267 kstrdup(const char *str, struct malloc_type *type)
1270 int zlen; /* length inclusive of terminating NUL */
1275 zlen = strlen(str) + 1;
1276 nstr = _kmalloc_debug(zlen, type, M_WAITOK, file, line);
1277 bcopy(str, nstr, zlen);
1283 kstrndup_debug(const char *str, size_t maxlen, struct malloc_type *type,
1284 const char *file, int line)
1287 kstrndup(const char *str, size_t maxlen, struct malloc_type *type)
1290 int zlen; /* length inclusive of terminating NUL */
1295 zlen = strnlen(str, maxlen) + 1;
1296 nstr = _kmalloc_debug(zlen, type, M_WAITOK, file, line);
1297 bcopy(str, nstr, zlen);
1298 nstr[zlen - 1] = '\0';
1303 * Notify our cpu that a remote cpu has freed some chunks in a zone that
1304 * we own. RCount will be bumped so the memory should be good, but validate
1305 * that it really is.
1308 kfree_remote(void *ptr)
1315 slgd = &mycpu->gd_slab;
1318 KKASSERT(*kup == -((int)mycpuid + 1));
1319 KKASSERT(z->z_RCount > 0);
1320 atomic_subtract_int(&z->z_RCount, 1);
1322 logmemory(free_rem_beg, z, NULL, 0L, 0);
1323 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1324 KKASSERT(z->z_Cpu == mycpu->gd_cpuid);
1328 * Indicate that we will no longer be off of the ZoneAry by
1335 * Atomically extract the bchunks list and then process it back
1336 * into the lchunks list. We want to append our bchunks to the
1337 * lchunks list and not prepend since we likely do not have
1338 * cache mastership of the related data (not that it helps since
1339 * we are using c_Next).
1341 clean_zone_rchunks(z);
1342 if (z->z_NFree && nfree == 0) {
1343 TAILQ_INSERT_HEAD(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
1346 check_zone_free(slgd, z);
1347 logmemory(free_rem_end, z, NULL, 0L, 0);
1351 * free (SLAB ALLOCATOR)
1353 * Free a memory block previously allocated by malloc.
1355 * Note: We do not attempt to update ks_loosememuse as MP races could
1356 * prevent us from checking memory limits in malloc. YYY we may
1357 * consider updating ks_cpu.loosememuse.
1362 _kfree(void *ptr, struct malloc_type *type)
1367 struct globaldata *gd;
1373 logmemory_quick(free_beg);
1375 slgd = &gd->gd_slab;
1378 panic("trying to free NULL pointer");
1381 * Handle special 0-byte allocations
1383 if (ptr == ZERO_LENGTH_PTR) {
1384 logmemory(free_zero, ptr, type, -1UL, 0);
1385 logmemory_quick(free_end);
1390 * Panic on bad malloc type
1392 if (type->ks_magic != M_MAGIC)
1393 panic("free: malloc type lacks magic");
1396 * Handle oversized allocations. XXX we really should require that a
1397 * size be passed to free() instead of this nonsense.
1399 * This code is never called via an ipi.
1403 size = *kup << PAGE_SHIFT;
1406 if (use_weird_array) {
1407 KKASSERT(sizeof(weirdary) <= size);
1408 bcopy(weirdary, ptr, sizeof(weirdary));
1412 * NOTE: For oversized allocations we do not record the
1413 * originating cpu. It gets freed on the cpu calling
1414 * kfree(). The statistics are in aggregate.
1416 * note: XXX we have still inherited the interrupts-can't-block
1417 * assumption. An interrupt thread does not bump
1418 * gd_intr_nesting_level so check TDF_INTTHREAD. This is
1419 * primarily until we can fix softupdate's assumptions about free().
1422 --type->ks_use[gd->gd_cpuid].inuse;
1423 type->ks_use[gd->gd_cpuid].memuse -= size;
1424 if (mycpu->gd_intr_nesting_level ||
1425 (gd->gd_curthread->td_flags & TDF_INTTHREAD)) {
1426 logmemory(free_ovsz_delayed, ptr, type, size, 0);
1428 z->z_Magic = ZALLOC_OVSZ_MAGIC;
1429 z->z_ChunkSize = size;
1431 TAILQ_INSERT_HEAD(&slgd->FreeOvZones, z, z_Entry);
1435 logmemory(free_ovsz, ptr, type, size, 0);
1436 kmem_slab_free(ptr, size); /* may block */
1438 logmemory_quick(free_end);
1443 * Zone case. Figure out the zone based on the fact that it is
1446 z = (SLZone *)((uintptr_t)ptr & ZoneMask);
1449 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC);
1452 * If we do not own the zone then use atomic ops to free to the
1453 * remote cpu linked list and notify the target zone using a
1456 * The target zone cannot be deallocated while we own a chunk of it,
1457 * so the zone header's storage is stable until the very moment
1458 * we adjust z_RChunks. After that we cannot safely dereference (z).
1460 * (no critical section needed)
1462 if (z->z_CpuGd != gd) {
1464 * Making these adjustments now allow us to avoid passing (type)
1465 * to the remote cpu. Note that inuse/memuse is being
1466 * adjusted on OUR cpu, not the zone cpu, but it should all still
1467 * sum up properly and cancel out.
1470 --type->ks_use[gd->gd_cpuid].inuse;
1471 type->ks_use[gd->gd_cpuid].memuse -= z->z_ChunkSize;
1475 * WARNING! This code competes with other cpus. Once we
1476 * successfully link the chunk to RChunks the remote
1477 * cpu can rip z's storage out from under us.
1479 * Bumping RCount prevents z's storage from getting
1482 rsignal = z->z_RSignal;
1485 atomic_add_int(&z->z_RCount, 1);
1489 bchunk = z->z_RChunks;
1491 chunk->c_Next = bchunk;
1494 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, chunk))
1499 * We have to signal the remote cpu if our actions will cause
1500 * the remote zone to be placed back on ZoneAry so it can
1501 * move the zone back on.
1503 * We only need to deal with NULL->non-NULL RChunk transitions
1504 * and only if z_RSignal is set. We interlock by reading rsignal
1505 * before adding our chunk to RChunks. This should result in
1506 * virtually no IPI traffic.
1508 * We can use a passive IPI to reduce overhead even further.
1510 if (bchunk == NULL && rsignal) {
1511 logmemory(free_request, ptr, type,
1512 (unsigned long)z->z_ChunkSize, 0);
1513 lwkt_send_ipiq_passive(z->z_CpuGd, kfree_remote, z);
1514 /* z can get ripped out from under us from this point on */
1515 } else if (rsignal) {
1516 atomic_subtract_int(&z->z_RCount, 1);
1517 /* z can get ripped out from under us from this point on */
1519 logmemory_quick(free_end);
1526 logmemory(free_chunk, ptr, type, (unsigned long)z->z_ChunkSize, 0);
1530 chunk_mark_free(z, chunk);
1533 * Put weird data into the memory to detect modifications after freeing,
1534 * illegal pointer use after freeing (we should fault on the odd address),
1535 * and so forth. XXX needs more work, see the old malloc code.
1538 if (use_weird_array) {
1539 if (z->z_ChunkSize < sizeof(weirdary))
1540 bcopy(weirdary, chunk, z->z_ChunkSize);
1542 bcopy(weirdary, chunk, sizeof(weirdary));
1547 * Add this free non-zero'd chunk to a linked list for reuse. Add
1548 * to the front of the linked list so it is more likely to be
1549 * reallocated, since it is already in our L1 cache.
1552 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd)
1553 panic("BADFREE %p", chunk);
1555 chunk->c_Next = z->z_LChunks;
1556 z->z_LChunks = chunk;
1557 if (chunk->c_Next == NULL)
1558 z->z_LChunksp = &chunk->c_Next;
1561 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart)
1566 * Bump the number of free chunks. If it becomes non-zero the zone
1567 * must be added back onto the appropriate list. A fully allocated
1568 * zone that sees its first free is considered 'mature' and is placed
1569 * at the head, giving the system time to potentially free the remaining
1570 * entries even while other allocations are going on and making the zone
1573 if (z->z_NFree++ == 0)
1574 TAILQ_INSERT_HEAD(&slgd->ZoneAry[z->z_ZoneIndex], z, z_Entry);
1576 --type->ks_use[gd->gd_cpuid].inuse;
1577 type->ks_use[gd->gd_cpuid].memuse -= z->z_ChunkSize;
1579 check_zone_free(slgd, z);
1580 logmemory_quick(free_end);
1585 * Cleanup slabs which are hanging around due to RChunks or which are wholely
1586 * free and can be moved to the free list if not moved by other means.
1588 * Called once every 10 seconds on all cpus.
1593 SLGlobalData *slgd = &mycpu->gd_slab;
1598 for (i = 0; i < NZONES; ++i) {
1599 if ((z = TAILQ_FIRST(&slgd->ZoneAry[i])) == NULL)
1607 * Shift all RChunks to the end of the LChunks list. This is
1608 * an O(1) operation.
1610 * Then free the zone if possible.
1612 clean_zone_rchunks(z);
1613 z = check_zone_free(slgd, z);
1619 #if defined(INVARIANTS)
1622 * Helper routines for sanity checks
1625 chunk_mark_allocated(SLZone *z, void *chunk)
1627 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1630 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1631 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1632 ("memory chunk %p bit index %d is illegal", chunk, bitdex));
1633 bitptr = &z->z_Bitmap[bitdex >> 5];
1635 KASSERT((*bitptr & (1 << bitdex)) == 0,
1636 ("memory chunk %p is already allocated!", chunk));
1637 *bitptr |= 1 << bitdex;
1641 chunk_mark_free(SLZone *z, void *chunk)
1643 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize;
1646 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0);
1647 KASSERT(bitdex >= 0 && bitdex < z->z_NMax,
1648 ("memory chunk %p bit index %d is illegal!", chunk, bitdex));
1649 bitptr = &z->z_Bitmap[bitdex >> 5];
1651 KASSERT((*bitptr & (1 << bitdex)) != 0,
1652 ("memory chunk %p is already free!", chunk));
1653 *bitptr &= ~(1 << bitdex);
1661 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the
1662 * specified alignment. M_* flags are expected in the flags field.
1664 * Alignment must be a multiple of PAGE_SIZE.
1666 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(),
1667 * but when we move zalloc() over to use this function as its backend
1668 * we will have to switch to kreserve/krelease and call reserve(0)
1669 * after the new space is made available.
1671 * Interrupt code which has preempted other code is not allowed to
1672 * use PQ_CACHE pages. However, if an interrupt thread is run
1673 * non-preemptively or blocks and then runs non-preemptively, then
1674 * it is free to use PQ_CACHE pages. <--- may not apply any longer XXX
1677 kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags)
1681 int count, vmflags, base_vmflags;
1682 vm_page_t mbase = NULL;
1686 size = round_page(size);
1687 addr = vm_map_min(kernel_map);
1689 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1691 vm_map_lock(kernel_map);
1692 if (vm_map_findspace(kernel_map, addr, size, align, 0, &addr)) {
1693 vm_map_unlock(kernel_map);
1694 if ((flags & M_NULLOK) == 0)
1695 panic("kmem_slab_alloc(): kernel_map ran out of space!");
1696 vm_map_entry_release(count);
1702 * kernel_object maps 1:1 to kernel_map.
1704 vm_object_hold(kernel_object);
1705 vm_object_reference_locked(kernel_object);
1706 vm_map_insert(kernel_map, &count,
1707 kernel_object, NULL,
1712 VM_PROT_ALL, VM_PROT_ALL, 0);
1713 vm_object_drop(kernel_object);
1714 vm_map_set_wired_quick(kernel_map, addr, size, &count);
1715 vm_map_unlock(kernel_map);
1721 base_vmflags |= VM_ALLOC_ZERO;
1722 if (flags & M_USE_RESERVE)
1723 base_vmflags |= VM_ALLOC_SYSTEM;
1724 if (flags & M_USE_INTERRUPT_RESERVE)
1725 base_vmflags |= VM_ALLOC_INTERRUPT;
1726 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) {
1727 panic("kmem_slab_alloc: bad flags %08x (%p)",
1728 flags, ((int **)&size)[-1]);
1732 * Allocate the pages. Do not map them yet. VM_ALLOC_NORMAL can only
1733 * be set if we are not preempting.
1735 * VM_ALLOC_SYSTEM is automatically set if we are preempting and
1736 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is
1737 * implied in this case), though I'm not sure if we really need to
1740 vmflags = base_vmflags;
1741 if (flags & M_WAITOK) {
1742 if (td->td_preempted)
1743 vmflags |= VM_ALLOC_SYSTEM;
1745 vmflags |= VM_ALLOC_NORMAL;
1748 vm_object_hold(kernel_object);
1749 for (i = 0; i < size; i += PAGE_SIZE) {
1750 m = vm_page_alloc(kernel_object, OFF_TO_IDX(addr + i), vmflags);
1755 * If the allocation failed we either return NULL or we retry.
1757 * If M_WAITOK is specified we wait for more memory and retry.
1758 * If M_WAITOK is specified from a preemption we yield instead of
1759 * wait. Livelock will not occur because the interrupt thread
1760 * will not be preempting anyone the second time around after the
1764 if (flags & M_WAITOK) {
1765 if (td->td_preempted) {
1770 i -= PAGE_SIZE; /* retry */
1778 * Check and deal with an allocation failure
1783 m = vm_page_lookup(kernel_object, OFF_TO_IDX(addr + i));
1784 /* page should already be busy */
1787 vm_map_lock(kernel_map);
1788 vm_map_delete(kernel_map, addr, addr + size, &count);
1789 vm_map_unlock(kernel_map);
1790 vm_object_drop(kernel_object);
1792 vm_map_entry_release(count);
1800 * NOTE: The VM pages are still busied. mbase points to the first one
1801 * but we have to iterate via vm_page_next()
1803 vm_object_drop(kernel_object);
1807 * Enter the pages into the pmap and deal with M_ZERO.
1814 * page should already be busy
1816 m->valid = VM_PAGE_BITS_ALL;
1818 pmap_enter(kernel_pmap, addr + i, m,
1819 VM_PROT_ALL | VM_PROT_NOSYNC, 1, NULL);
1821 pagezero((char *)addr + i);
1822 KKASSERT(m->flags & (PG_WRITEABLE | PG_MAPPED));
1823 vm_page_flag_set(m, PG_REFERENCED);
1827 vm_object_hold(kernel_object);
1828 m = vm_page_next(m);
1829 vm_object_drop(kernel_object);
1832 vm_map_entry_release(count);
1833 return((void *)addr);
1840 kmem_slab_free(void *ptr, vm_size_t size)
1843 vm_map_remove(kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size);