4 * Copyright (c) 1997, 1998 John S. Dyson
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice immediately at the beginning of the file, without modification,
12 * this list of conditions, and the following disclaimer.
13 * 2. Absolutely no warranty of function or purpose is made by the author
16 * $FreeBSD: src/sys/vm/vm_zone.c,v 1.30.2.6 2002/10/10 19:50:16 dillon Exp $
19 #include <sys/param.h>
20 #include <sys/queue.h>
21 #include <sys/systm.h>
22 #include <sys/kernel.h>
24 #include <sys/malloc.h>
25 #include <sys/sysctl.h>
26 #include <sys/vmmeter.h>
29 #include <vm/vm_object.h>
30 #include <vm/vm_page.h>
31 #include <vm/vm_map.h>
32 #include <vm/vm_kern.h>
33 #include <vm/vm_extern.h>
34 #include <vm/vm_zone.h>
36 #include <sys/spinlock2.h>
37 #include <vm/vm_page2.h>
39 static MALLOC_DEFINE(M_ZONE, "ZONE", "Zone header");
41 #define ZONE_ERROR_INVALID 0
42 #define ZONE_ERROR_NOTFREE 1
43 #define ZONE_ERROR_ALREADYFREE 2
45 #define ZONE_ROUNDING 32
47 #define ZENTRY_FREE 0x12342378
51 static void *zget(vm_zone_t z);
54 * Return an item from the specified zone. This function is non-blocking for
55 * ZONE_INTERRUPT zones.
62 globaldata_t gd = mycpu;
68 zerror(ZONE_ERROR_INVALID);
72 * Avoid spinlock contention by allocating from a per-cpu queue
74 if (z->zfreecnt_pcpu[gd->gd_cpuid] > 0) {
76 if (z->zfreecnt_pcpu[gd->gd_cpuid] > 0) {
77 item = z->zitems_pcpu[gd->gd_cpuid];
80 ("zitems_pcpu unexpectedly NULL"));
81 if (((void **)item)[1] != (void *)ZENTRY_FREE)
82 zerror(ZONE_ERROR_NOTFREE);
83 ((void **)item)[1] = NULL;
85 z->zitems_pcpu[gd->gd_cpuid] = ((void **) item)[0];
86 --z->zfreecnt_pcpu[gd->gd_cpuid];
95 * Per-zone spinlock for the remainder. Always load at least one
99 if (z->zfreecnt > z->zfreemin) {
104 KASSERT(item != NULL, ("zitems unexpectedly NULL"));
105 if (((void **)item)[1] != (void *)ZENTRY_FREE)
106 zerror(ZONE_ERROR_NOTFREE);
108 z->zitems = ((void **)item)[0];
110 ((void **)item)[0] = z->zitems_pcpu[gd->gd_cpuid];
111 z->zitems_pcpu[gd->gd_cpuid] = item;
112 ++z->zfreecnt_pcpu[gd->gd_cpuid];
113 } while (--n > 0 && z->zfreecnt > z->zfreemin);
114 spin_unlock(&z->zlock);
117 spin_unlock(&z->zlock);
120 * PANICFAIL allows the caller to assume that the zalloc()
121 * will always succeed. If it doesn't, we panic here.
123 if (item == NULL && (z->zflags & ZONE_PANICFAIL))
124 panic("zalloc(%s) failed", z->zname);
130 * Free an item to the specified zone.
135 zfree(vm_zone_t z, void *item)
137 globaldata_t gd = mycpu;
141 * Avoid spinlock contention by freeing into a per-cpu queue
143 if ((zmax = z->zmax) != 0)
144 zmax = zmax / ncpus / 16;
148 if (z->zfreecnt_pcpu[gd->gd_cpuid] < zmax) {
150 ((void **)item)[0] = z->zitems_pcpu[gd->gd_cpuid];
152 if (((void **)item)[1] == (void *)ZENTRY_FREE)
153 zerror(ZONE_ERROR_ALREADYFREE);
154 ((void **)item)[1] = (void *)ZENTRY_FREE;
156 z->zitems_pcpu[gd->gd_cpuid] = item;
157 ++z->zfreecnt_pcpu[gd->gd_cpuid];
163 * Per-zone spinlock for the remainder.
165 spin_lock(&z->zlock);
166 ((void **)item)[0] = z->zitems;
168 if (((void **)item)[1] == (void *)ZENTRY_FREE)
169 zerror(ZONE_ERROR_ALREADYFREE);
170 ((void **)item)[1] = (void *)ZENTRY_FREE;
174 spin_unlock(&z->zlock);
178 * This file comprises a very simple zone allocator. This is used
179 * in lieu of the malloc allocator, where needed or more optimal.
181 * Note that the initial implementation of this had coloring, and
182 * absolutely no improvement (actually perf degradation) occurred.
184 * Note also that the zones are type stable. The only restriction is
185 * that the first two longwords of a data structure can be changed
186 * between allocations. Any data that must be stable between allocations
187 * must reside in areas after the first two longwords.
189 * zinitna, zinit, zbootinit are the initialization routines.
190 * zalloc, zfree, are the allocation/free routines.
193 LIST_HEAD(zlist, vm_zone) zlist = LIST_HEAD_INITIALIZER(zlist);
194 static int sysctl_vm_zone(SYSCTL_HANDLER_ARGS);
195 static int zone_kmem_pages, zone_kern_pages;
196 static long zone_kmem_kvaspace;
199 * Create a zone, but don't allocate the zone structure. If the
200 * zone had been previously created by the zone boot code, initialize
201 * various parts of the zone code.
203 * If waits are not allowed during allocation (e.g. during interrupt
204 * code), a-priori allocate the kernel virtual space, and allocate
205 * only pages when needed.
208 * z pointer to zone structure.
209 * obj pointer to VM object (opt).
211 * size size of zone entries.
212 * nentries number of zone entries allocated (only ZONE_INTERRUPT.)
213 * flags ZONE_INTERRUPT -- items can be allocated at interrupt time.
214 * zalloc number of pages allocated when memory is needed.
216 * Note that when using ZONE_INTERRUPT, the size of the zone is limited
217 * by the nentries argument. The size of the memory allocatable is
218 * unlimited if ZONE_INTERRUPT is not set.
223 zinitna(vm_zone_t z, vm_object_t obj, char *name, int size,
224 int nentries, int flags, int zalloc)
229 * Only zones created with zinit() are destroyable.
231 if (z->zflags & ZONE_DESTROYABLE)
232 panic("zinitna: can't create destroyable zone");
235 * NOTE: We can only adjust zsize if we previously did not
238 if ((z->zflags & ZONE_BOOT) == 0) {
239 z->zsize = (size + ZONE_ROUNDING - 1) & ~(ZONE_ROUNDING - 1);
240 spin_init(&z->zlock);
248 lwkt_gettoken(&vm_token);
249 LIST_INSERT_HEAD(&zlist, z, zlink);
250 lwkt_reltoken(&vm_token);
252 bzero(z->zitems_pcpu, sizeof(z->zitems_pcpu));
253 bzero(z->zfreecnt_pcpu, sizeof(z->zfreecnt_pcpu));
257 z->zkmcur = z->zkmmax = 0;
261 * If we cannot wait, allocate KVA space up front, and we will fill
262 * in pages as needed. This is particularly required when creating
263 * an allocation space for map entries in kernel_map, because we
264 * do not want to go into a recursion deadlock with
265 * vm_map_entry_reserve().
267 if (z->zflags & ZONE_INTERRUPT) {
268 totsize = round_page((size_t)z->zsize * nentries);
269 atomic_add_long(&zone_kmem_kvaspace, totsize);
271 z->zkva = kmem_alloc_pageable(&kernel_map, totsize);
273 LIST_REMOVE(z, zlink);
277 z->zpagemax = totsize / PAGE_SIZE;
279 z->zobj = vm_object_allocate(OBJT_DEFAULT, z->zpagemax);
282 _vm_object_allocate(OBJT_DEFAULT, z->zpagemax, obj);
285 z->zallocflag = VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT |
286 VM_ALLOC_NORMAL | VM_ALLOC_RETRY;
289 z->zallocflag = VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM;
294 if (z->zsize > PAGE_SIZE)
297 z->zfreemin = PAGE_SIZE / z->zsize;
306 * Populate the interrrupt zone at creation time rather than
307 * on first allocation, as this is a potentially long operation.
309 if (z->zflags & ZONE_INTERRUPT) {
320 * Subroutine same as zinitna, except zone data structure is allocated
321 * automatically by malloc. This routine should normally be used, except
322 * in certain tricky startup conditions in the VM system -- then
323 * zbootinit and zinitna can be used. Zinit is the standard zone
324 * initialization call.
329 zinit(char *name, int size, int nentries, int flags, int zalloc)
333 z = (vm_zone_t) kmalloc(sizeof (struct vm_zone), M_ZONE, M_NOWAIT);
338 if (zinitna(z, NULL, name, size, nentries,
339 flags & ~ZONE_DESTROYABLE, zalloc) == 0) {
344 if (flags & ZONE_DESTROYABLE)
345 z->zflags |= ZONE_DESTROYABLE;
351 * Initialize a zone before the system is fully up. This routine should
352 * only be called before full VM startup.
354 * Called from the low level boot code only.
357 zbootinit(vm_zone_t z, char *name, int size, void *item, int nitems)
361 bzero(z->zitems_pcpu, sizeof(z->zitems_pcpu));
362 bzero(z->zfreecnt_pcpu, sizeof(z->zfreecnt_pcpu));
368 z->zflags = ZONE_BOOT;
374 spin_init(&z->zlock);
376 bzero(item, (size_t)nitems * z->zsize);
378 for (i = 0; i < nitems; i++) {
379 ((void **)item)[0] = z->zitems;
381 ((void **)item)[1] = (void *)ZENTRY_FREE;
384 item = (uint8_t *)item + z->zsize;
386 z->zfreecnt = nitems;
390 lwkt_gettoken(&vm_token);
391 LIST_INSERT_HEAD(&zlist, z, zlink);
392 lwkt_reltoken(&vm_token);
396 * Release all resources owned by zone created with zinit().
401 zdestroy(vm_zone_t z)
407 panic("zdestroy: null zone");
408 if ((z->zflags & ZONE_DESTROYABLE) == 0)
409 panic("zdestroy: undestroyable zone");
411 lwkt_gettoken(&vm_token);
412 LIST_REMOVE(z, zlink);
413 lwkt_reltoken(&vm_token);
416 * Release virtual mappings, physical memory and update sysctl stats.
418 if (z->zflags & ZONE_INTERRUPT) {
420 * Pages mapped via pmap_kenter() must be removed from the
421 * kernel_pmap() before calling kmem_free() to avoid issues
422 * with kernel_pmap.pm_stats.resident_count.
424 pmap_qremove(z->zkva, z->zpagemax);
425 vm_object_hold(z->zobj);
426 for (i = 0; i < z->zpagecount; ++i) {
427 m = vm_page_lookup_busy_wait(z->zobj, i, TRUE, "vmzd");
428 vm_page_unwire(m, 0);
435 kmem_free(&kernel_map, z->zkva,
436 (size_t)z->zpagemax * PAGE_SIZE);
437 atomic_subtract_long(&zone_kmem_kvaspace,
438 (size_t)z->zpagemax * PAGE_SIZE);
441 * Free the backing object and physical pages.
443 vm_object_deallocate(z->zobj);
444 vm_object_drop(z->zobj);
445 atomic_subtract_int(&zone_kmem_pages, z->zpagecount);
447 for (i=0; i < z->zkmcur; i++) {
448 kmem_free(&kernel_map, z->zkmvec[i],
449 (size_t)z->zalloc * PAGE_SIZE);
450 atomic_subtract_int(&zone_kern_pages, z->zalloc);
452 if (z->zkmvec != NULL)
453 kfree(z->zkmvec, M_ZONE);
456 spin_uninit(&z->zlock);
462 * void *zalloc(vm_zone_t zone) --
463 * Returns an item from a specified zone. May not be called from a
464 * FAST interrupt or IPI function.
466 * void zfree(vm_zone_t zone, void *item) --
467 * Frees an item back to a specified zone. May not be called from a
468 * FAST interrupt or IPI function.
472 * Internal zone routine. Not to be called from external (non vm_zone) code.
489 panic("zget: null zone");
491 if (z->zflags & ZONE_INTERRUPT) {
493 * Interrupt zones do not mess with the kernel_map, they
494 * simply populate an existing mapping.
496 * First reserve the required space.
498 vm_object_hold(z->zobj);
499 noffset = (size_t)z->zpagecount * PAGE_SIZE;
500 noffset -= noffset % z->zsize;
501 savezpc = z->zpagecount;
502 if (z->zpagecount + z->zalloc > z->zpagemax)
503 z->zpagecount = z->zpagemax;
505 z->zpagecount += z->zalloc;
506 item = (char *)z->zkva + noffset;
507 npages = z->zpagecount - savezpc;
508 nitems = ((size_t)(savezpc + npages) * PAGE_SIZE - noffset) /
510 atomic_add_int(&zone_kmem_pages, npages);
513 * Now allocate the pages. Note that we can block in the
514 * loop, so we've already done all the necessary calculations
515 * and reservations above.
517 for (i = 0; i < npages; ++i) {
520 m = vm_page_alloc(z->zobj, savezpc + i, z->zallocflag);
522 /* note: z might be modified due to blocking */
524 KKASSERT(m->queue == PQ_NONE);
525 m->valid = VM_PAGE_BITS_ALL;
529 zkva = z->zkva + (size_t)(savezpc + i) * PAGE_SIZE;
530 pmap_kenter(zkva, VM_PAGE_TO_PHYS(m));
531 bzero((void *)zkva, PAGE_SIZE);
533 vm_object_drop(z->zobj);
534 } else if (z->zflags & ZONE_SPECIAL) {
536 * The special zone is the one used for vm_map_entry_t's.
537 * We have to avoid an infinite recursion in
538 * vm_map_entry_reserve() by using vm_map_entry_kreserve()
539 * instead. The map entries are pre-reserved by the kernel
540 * by vm_map_entry_reserve_cpu_init().
542 nbytes = (size_t)z->zalloc * PAGE_SIZE;
544 item = (void *)kmem_alloc3(&kernel_map, nbytes, KM_KRESERVE);
546 /* note: z might be modified due to blocking */
548 zone_kern_pages += z->zalloc; /* not MP-safe XXX */
553 nitems = nbytes / z->zsize;
556 * Otherwise allocate KVA from the kernel_map.
558 nbytes = (size_t)z->zalloc * PAGE_SIZE;
560 item = (void *)kmem_alloc3(&kernel_map, nbytes, 0);
562 /* note: z might be modified due to blocking */
564 zone_kern_pages += z->zalloc; /* not MP-safe XXX */
567 if (z->zflags & ZONE_DESTROYABLE) {
568 if (z->zkmcur == z->zkmmax) {
570 z->zkmmax==0 ? 1 : z->zkmmax*2;
571 z->zkmvec = krealloc(z->zkmvec,
572 z->zkmmax * sizeof(z->zkmvec[0]),
575 z->zkmvec[z->zkmcur++] = (vm_offset_t)item;
580 nitems = nbytes / z->zsize;
583 spin_lock(&z->zlock);
586 * Save one for immediate allocation
590 for (i = 0; i < nitems; i++) {
591 ((void **)item)[0] = z->zitems;
593 ((void **)item)[1] = (void *)ZENTRY_FREE;
596 item = (uint8_t *)item + z->zsize;
598 z->zfreecnt += nitems;
600 } else if (z->zfreecnt > 0) {
602 z->zitems = ((void **)item)[0];
604 if (((void **)item)[1] != (void *)ZENTRY_FREE)
605 zerror(ZONE_ERROR_NOTFREE);
606 ((void **) item)[1] = NULL;
613 spin_unlock(&z->zlock);
616 * A special zone may have used a kernel-reserved vm_map_entry. If
617 * so we have to be sure to recover our reserve so we don't run out.
618 * We will panic if we run out.
620 if (z->zflags & ZONE_SPECIAL)
621 vm_map_entry_reserve(0);
630 sysctl_vm_zone(SYSCTL_HANDLER_ARGS)
637 ksnprintf(tmpbuf, sizeof(tmpbuf),
638 "\nITEM SIZE LIMIT USED FREE REQUESTS\n");
639 error = SYSCTL_OUT(req, tmpbuf, strlen(tmpbuf));
643 lwkt_gettoken(&vm_token);
644 LIST_FOREACH(curzone, &zlist, zlink) {
651 len = strlen(curzone->zname);
652 if (len >= (sizeof(tmpname) - 1))
653 len = (sizeof(tmpname) - 1);
654 for(i = 0; i < sizeof(tmpname) - 1; i++)
657 memcpy(tmpname, curzone->zname, len);
660 if (curzone == LIST_FIRST(&zlist)) {
664 freecnt = curzone->zfreecnt;
665 for (n = 0; n < ncpus; ++n)
666 freecnt += curzone->zfreecnt_pcpu[n];
668 ksnprintf(tmpbuf + offset, sizeof(tmpbuf) - offset,
669 "%s %6.6u, %8.8u, %6.6u, %6.6u, %8.8u\n",
670 tmpname, curzone->zsize, curzone->zmax,
671 (curzone->ztotal - freecnt),
672 freecnt, curzone->znalloc);
674 len = strlen((char *)tmpbuf);
675 if (LIST_NEXT(curzone, zlink) == NULL)
678 error = SYSCTL_OUT(req, tmpbuf, len);
683 lwkt_reltoken(&vm_token);
687 #if defined(INVARIANTS)
698 case ZONE_ERROR_INVALID:
699 msg = "zone: invalid zone";
701 case ZONE_ERROR_NOTFREE:
702 msg = "zone: entry not free";
704 case ZONE_ERROR_ALREADYFREE:
705 msg = "zone: freeing free entry";
708 msg = "zone: invalid error";
715 SYSCTL_OID(_vm, OID_AUTO, zone, CTLTYPE_STRING|CTLFLAG_RD, \
716 NULL, 0, sysctl_vm_zone, "A", "Zone Info");
718 SYSCTL_INT(_vm, OID_AUTO, zone_kmem_pages,
719 CTLFLAG_RD, &zone_kmem_pages, 0, "Number of interrupt safe pages allocated by zone");
720 SYSCTL_INT(_vm, OID_AUTO, zone_burst,
721 CTLFLAG_RW, &zone_burst, 0, "Burst from depot to pcpu cache");
722 SYSCTL_LONG(_vm, OID_AUTO, zone_kmem_kvaspace,
723 CTLFLAG_RD, &zone_kmem_kvaspace, 0, "KVA space allocated by zone");
724 SYSCTL_INT(_vm, OID_AUTO, zone_kern_pages,
725 CTLFLAG_RD, &zone_kern_pages, 0, "Number of non-interrupt safe pages allocated by zone");