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
49 long zone_burst = 128;
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;
69 zerror(ZONE_ERROR_INVALID);
71 zpcpu = &z->zpcpu[gd->gd_cpuid];
74 * Avoid spinlock contention by allocating from a per-cpu queue
76 if (zpcpu->zfreecnt > 0) {
78 if (zpcpu->zfreecnt > 0) {
82 ("zitems_pcpu unexpectedly NULL"));
83 if (((void **)item)[1] != (void *)ZENTRY_FREE)
84 zerror(ZONE_ERROR_NOTFREE);
85 ((void **)item)[1] = NULL;
87 zpcpu->zitems = ((void **) item)[0];
98 * Per-zone spinlock for the remainder. Always load at least one
101 spin_lock(&z->zspin);
102 if (z->zfreecnt > z->zfreemin) {
107 KASSERT(item != NULL, ("zitems unexpectedly NULL"));
108 if (((void **)item)[1] != (void *)ZENTRY_FREE)
109 zerror(ZONE_ERROR_NOTFREE);
111 z->zitems = ((void **)item)[0];
113 ((void **)item)[0] = zpcpu->zitems;
114 zpcpu->zitems = item;
116 } while (--n > 0 && z->zfreecnt > z->zfreemin);
117 spin_unlock(&z->zspin);
120 spin_unlock(&z->zspin);
123 * PANICFAIL allows the caller to assume that the zalloc()
124 * will always succeed. If it doesn't, we panic here.
126 if (item == NULL && (z->zflags & ZONE_PANICFAIL))
127 panic("zalloc(%s) failed", z->zname);
133 * Free an item to the specified zone.
138 zfree(vm_zone_t z, void *item)
140 globaldata_t gd = mycpu;
146 zpcpu = &z->zpcpu[gd->gd_cpuid];
149 * Avoid spinlock contention by freeing into a per-cpu queue
159 ((void **)item)[0] = zpcpu->zitems;
161 if (((void **)item)[1] == (void *)ZENTRY_FREE)
162 zerror(ZONE_ERROR_ALREADYFREE);
163 ((void **)item)[1] = (void *)ZENTRY_FREE;
165 zpcpu->zitems = item;
168 if (zpcpu->zfreecnt < zmax) {
174 * Hystereis, move (zmax) (calculated below) items to the pool.
177 if (zmax > zone_burst)
182 while (count < zmax) {
183 tail_item = ((void **)tail_item)[0];
186 zpcpu->zitems = ((void **)tail_item)[0];
187 zpcpu->zfreecnt -= count;
190 * Per-zone spinlock for the remainder.
192 * Also implement hysteresis by freeing a number of pcpu
195 spin_lock(&z->zspin);
196 ((void **)tail_item)[0] = z->zitems;
198 z->zfreecnt += count;
199 spin_unlock(&z->zspin);
205 * This file comprises a very simple zone allocator. This is used
206 * in lieu of the malloc allocator, where needed or more optimal.
208 * Note that the initial implementation of this had coloring, and
209 * absolutely no improvement (actually perf degradation) occurred.
211 * Note also that the zones are type stable. The only restriction is
212 * that the first two longwords of a data structure can be changed
213 * between allocations. Any data that must be stable between allocations
214 * must reside in areas after the first two longwords.
216 * zinitna, zinit, zbootinit are the initialization routines.
217 * zalloc, zfree, are the allocation/free routines.
220 LIST_HEAD(zlist, vm_zone) zlist = LIST_HEAD_INITIALIZER(zlist);
221 static int sysctl_vm_zone(SYSCTL_HANDLER_ARGS);
222 static vm_pindex_t zone_kmem_pages, zone_kern_pages;
223 static long zone_kmem_kvaspace;
226 * Create a zone, but don't allocate the zone structure. If the
227 * zone had been previously created by the zone boot code, initialize
228 * various parts of the zone code.
230 * If waits are not allowed during allocation (e.g. during interrupt
231 * code), a-priori allocate the kernel virtual space, and allocate
232 * only pages when needed.
235 * z pointer to zone structure.
236 * obj pointer to VM object (opt).
238 * size size of zone entries.
239 * nentries number of zone entries allocated (only ZONE_INTERRUPT.)
240 * flags ZONE_INTERRUPT -- items can be allocated at interrupt time.
241 * zalloc number of pages allocated when memory is needed.
243 * Note that when using ZONE_INTERRUPT, the size of the zone is limited
244 * by the nentries argument. The size of the memory allocatable is
245 * unlimited if ZONE_INTERRUPT is not set.
250 zinitna(vm_zone_t z, char *name, size_t size, long nentries, uint32_t flags)
255 * Only zones created with zinit() are destroyable.
257 if (z->zflags & ZONE_DESTROYABLE)
258 panic("zinitna: can't create destroyable zone");
261 * NOTE: We can only adjust zsize if we previously did not
264 if ((z->zflags & ZONE_BOOT) == 0) {
265 z->zsize = roundup2(size, ZONE_ROUNDING);
266 spin_init(&z->zspin, "zinitna");
273 lwkt_gettoken(&vm_token);
274 LIST_INSERT_HEAD(&zlist, z, zlink);
275 lwkt_reltoken(&vm_token);
277 bzero(z->zpcpu, sizeof(z->zpcpu));
281 z->zkmcur = z->zkmmax = 0;
285 * If we cannot wait, allocate KVA space up front, and we will fill
286 * in pages as needed. This is particularly required when creating
287 * an allocation space for map entries in kernel_map, because we
288 * do not want to go into a recursion deadlock with
289 * vm_map_entry_reserve().
291 if (z->zflags & ZONE_INTERRUPT) {
292 totsize = round_page((size_t)z->zsize * nentries);
293 atomic_add_long(&zone_kmem_kvaspace, totsize);
295 z->zkva = kmem_alloc_pageable(&kernel_map, totsize,
298 LIST_REMOVE(z, zlink);
302 z->zpagemax = totsize / PAGE_SIZE;
303 z->zallocflag = VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT |
304 VM_ALLOC_NORMAL | VM_ALLOC_RETRY;
306 z->zmax_pcpu = z->zmax / ncpus / 16;
309 * Set reasonable pcpu cache bounds. Low-memory systems
310 * might try to cache too little, large-memory systems
311 * might try to cache more than necessarsy.
313 * In particular, pvzone can wind up being excessive and
314 * waste memory unnecessarily.
316 if (z->zmax_pcpu < 1024)
318 if (z->zmax_pcpu * z->zsize > 16*1024*1024)
319 z->zmax_pcpu = 16*1024*1024 / z->zsize;
321 z->zallocflag = VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM;
327 if (z->zsize > PAGE_SIZE)
330 z->zfreemin = PAGE_SIZE / z->zsize;
335 * Reduce kernel_map spam by allocating in chunks of 4 pages.
340 * Populate the interrrupt zone at creation time rather than
341 * on first allocation, as this is a potentially long operation.
343 if (z->zflags & ZONE_INTERRUPT) {
355 * Subroutine same as zinitna, except zone data structure is allocated
356 * automatically by malloc. This routine should normally be used, except
357 * in certain tricky startup conditions in the VM system -- then
358 * zbootinit and zinitna can be used. Zinit is the standard zone
359 * initialization call.
364 zinit(char *name, size_t size, long nentries, uint32_t flags)
368 z = (vm_zone_t) kmalloc(sizeof (struct vm_zone), M_ZONE, M_NOWAIT);
373 if (zinitna(z, name, size, nentries, flags & ~ZONE_DESTROYABLE) == 0) {
378 if (flags & ZONE_DESTROYABLE)
379 z->zflags |= ZONE_DESTROYABLE;
385 * Initialize a zone before the system is fully up. This routine should
386 * only be called before full VM startup.
388 * Called from the low level boot code only.
391 zbootinit(vm_zone_t z, char *name, size_t size, void *item, long nitems)
395 spin_init(&z->zspin, "zbootinit");
396 bzero(z->zpcpu, sizeof(z->zpcpu));
400 z->zflags = ZONE_BOOT;
406 bzero(item, (size_t)nitems * z->zsize);
408 for (i = 0; i < nitems; i++) {
409 ((void **)item)[0] = z->zitems;
411 ((void **)item)[1] = (void *)ZENTRY_FREE;
414 item = (uint8_t *)item + z->zsize;
416 z->zfreecnt = nitems;
420 lwkt_gettoken(&vm_token);
421 LIST_INSERT_HEAD(&zlist, z, zlink);
422 lwkt_reltoken(&vm_token);
426 * Release all resources owned by zone created with zinit().
431 zdestroy(vm_zone_t z)
436 panic("zdestroy: null zone");
437 if ((z->zflags & ZONE_DESTROYABLE) == 0)
438 panic("zdestroy: undestroyable zone");
440 lwkt_gettoken(&vm_token);
441 LIST_REMOVE(z, zlink);
442 lwkt_reltoken(&vm_token);
445 * Release virtual mappings, physical memory and update sysctl stats.
447 KKASSERT((z->zflags & ZONE_INTERRUPT) == 0);
448 for (i = 0; i < z->zkmcur; i++) {
449 kmem_free(&kernel_map, z->zkmvec[i],
450 (size_t)z->zalloc * PAGE_SIZE);
451 atomic_subtract_long(&zone_kern_pages, z->zalloc);
453 if (z->zkmvec != NULL)
454 kfree(z->zkmvec, M_ZONE);
456 spin_uninit(&z->zspin);
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.
474 * This function may return NULL.
481 vm_page_t pgs[ZONE_MAXPGLOAD];
493 panic("zget: null zone");
495 if (z->zflags & ZONE_INTERRUPT) {
497 * Interrupt zones do not mess with the kernel_map, they
498 * simply populate an existing mapping.
500 * First allocate as many pages as we can, stopping at
501 * our limit or if the page allocation fails.
503 for (i = 0; i < ZONE_MAXPGLOAD && i < z->zalloc; ++i) {
504 m = vm_page_alloc(NULL,
505 mycpu->gd_rand_incr++,
514 * Account for the pages.
516 * NOTE! Do not allow overlap with a prior page as it
517 * may still be undergoing allocation on another
520 spin_lock(&z->zspin);
521 noffset = (size_t)z->zpagecount * PAGE_SIZE;
522 /* noffset -= noffset % z->zsize; */
523 savezpc = z->zpagecount;
524 if (z->zpagecount + nalloc > z->zpagemax)
525 z->zpagecount = z->zpagemax;
527 z->zpagecount += nalloc;
528 item = (char *)z->zkva + noffset;
529 npages = z->zpagecount - savezpc;
530 nitems = ((size_t)(savezpc + npages) * PAGE_SIZE - noffset) /
532 atomic_add_long(&zone_kmem_pages, npages);
533 spin_unlock(&z->zspin);
536 * Enter the pages into the reserved KVA space.
538 for (i = 0; i < npages; ++i) {
542 KKASSERT(m->queue == PQ_NONE);
543 m->valid = VM_PAGE_BITS_ALL;
547 zkva = z->zkva + (size_t)(savezpc + i) * PAGE_SIZE;
548 pmap_kenter(zkva, VM_PAGE_TO_PHYS(m));
549 bzero((void *)zkva, PAGE_SIZE);
551 for (i = npages; i < nalloc; ++i) {
555 } else if (z->zflags & ZONE_SPECIAL) {
557 * The special zone is the one used for vm_map_entry_t's.
558 * We have to avoid an infinite recursion in
559 * vm_map_entry_reserve() by using vm_map_entry_kreserve()
560 * instead. The map entries are pre-reserved by the kernel
561 * by vm_map_entry_reserve_cpu_init().
563 nbytes = (size_t)z->zalloc * PAGE_SIZE;
565 item = (void *)kmem_alloc3(&kernel_map, nbytes,
566 VM_SUBSYS_ZALLOC, KM_KRESERVE);
568 /* note: z might be modified due to blocking */
570 atomic_add_long(&zone_kern_pages, z->zalloc);
575 nitems = nbytes / z->zsize;
578 * Otherwise allocate KVA from the kernel_map.
580 nbytes = (size_t)z->zalloc * PAGE_SIZE;
582 item = (void *)kmem_alloc3(&kernel_map, nbytes,
583 VM_SUBSYS_ZALLOC, 0);
585 /* note: z might be modified due to blocking */
587 atomic_add_long(&zone_kern_pages, z->zalloc);
590 if (z->zflags & ZONE_DESTROYABLE) {
591 if (z->zkmcur == z->zkmmax) {
593 z->zkmmax==0 ? 1 : z->zkmmax*2;
594 z->zkmvec = krealloc(z->zkmvec,
595 z->zkmmax * sizeof(z->zkmvec[0]),
598 z->zkmvec[z->zkmcur++] = (vm_offset_t)item;
603 nitems = nbytes / z->zsize;
607 * Enter any new pages into the pool, reserving one, or get the
608 * item from the existing pool.
610 spin_lock(&z->zspin);
615 * Enter pages into the pool saving one for immediate
619 for (i = 0; i < nitems; i++) {
620 ((void **)item)[0] = z->zitems;
622 ((void **)item)[1] = (void *)ZENTRY_FREE;
625 item = (uint8_t *)item + z->zsize;
627 z->zfreecnt += nitems;
629 } else if (z->zfreecnt > 0) {
631 * Get an item from the existing pool.
634 z->zitems = ((void **)item)[0];
636 if (((void **)item)[1] != (void *)ZENTRY_FREE)
637 zerror(ZONE_ERROR_NOTFREE);
638 ((void **) item)[1] = NULL;
644 * No items available.
648 spin_unlock(&z->zspin);
651 * A special zone may have used a kernel-reserved vm_map_entry. If
652 * so we have to be sure to recover our reserve so we don't run out.
653 * We will panic if we run out.
655 if (z->zflags & ZONE_SPECIAL)
656 vm_map_entry_reserve(0);
665 sysctl_vm_zone(SYSCTL_HANDLER_ARGS)
672 ksnprintf(tmpbuf, sizeof(tmpbuf),
673 "\nITEM SIZE LIMIT USED FREE REQUESTS\n");
674 error = SYSCTL_OUT(req, tmpbuf, strlen(tmpbuf));
678 lwkt_gettoken(&vm_token);
679 LIST_FOREACH(curzone, &zlist, zlink) {
687 len = strlen(curzone->zname);
688 if (len >= (sizeof(tmpname) - 1))
689 len = (sizeof(tmpname) - 1);
690 for(i = 0; i < sizeof(tmpname) - 1; i++)
693 memcpy(tmpname, curzone->zname, len);
696 if (curzone == LIST_FIRST(&zlist)) {
700 freecnt = curzone->zfreecnt;
701 znalloc = curzone->znalloc;
702 for (n = 0; n < ncpus; ++n) {
703 freecnt += curzone->zpcpu[n].zfreecnt;
704 znalloc += curzone->zpcpu[n].znalloc;
707 ksnprintf(tmpbuf + offset, sizeof(tmpbuf) - offset,
708 "%s %6.6lu, %8.8lu, %6.6lu, %6.6lu, %8.8lu\n",
709 tmpname, curzone->zsize, curzone->zmax,
710 (curzone->ztotal - freecnt),
713 len = strlen((char *)tmpbuf);
714 if (LIST_NEXT(curzone, zlink) == NULL)
717 error = SYSCTL_OUT(req, tmpbuf, len);
722 lwkt_reltoken(&vm_token);
726 #if defined(INVARIANTS)
737 case ZONE_ERROR_INVALID:
738 msg = "zone: invalid zone";
740 case ZONE_ERROR_NOTFREE:
741 msg = "zone: entry not free";
743 case ZONE_ERROR_ALREADYFREE:
744 msg = "zone: freeing free entry";
747 msg = "zone: invalid error";
754 SYSCTL_OID(_vm, OID_AUTO, zone, CTLTYPE_STRING|CTLFLAG_RD, \
755 NULL, 0, sysctl_vm_zone, "A", "Zone Info");
757 SYSCTL_LONG(_vm, OID_AUTO, zone_kmem_pages,
758 CTLFLAG_RD, &zone_kmem_pages, 0, "Number of interrupt safe pages allocated by zone");
759 SYSCTL_LONG(_vm, OID_AUTO, zone_burst,
760 CTLFLAG_RW, &zone_burst, 0, "Burst from depot to pcpu cache");
761 SYSCTL_LONG(_vm, OID_AUTO, zone_kmem_kvaspace,
762 CTLFLAG_RD, &zone_kmem_kvaspace, 0, "KVA space allocated by zone");
763 SYSCTL_LONG(_vm, OID_AUTO, zone_kern_pages,
764 CTLFLAG_RD, &zone_kern_pages, 0, "Number of non-interrupt safe pages allocated by zone");