uvm_km.c revision 1.109
1/*	$NetBSD: uvm_km.c,v 1.109 2011/06/12 03:36:03 rmind Exp $	*/
2
3/*
4 * Copyright (c) 1997 Charles D. Cranor and Washington University.
5 * Copyright (c) 1991, 1993, The Regents of the University of California.
6 *
7 * All rights reserved.
8 *
9 * This code is derived from software contributed to Berkeley by
10 * The Mach Operating System project at Carnegie-Mellon University.
11 *
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
14 * are met:
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 the
19 *    documentation and/or other materials provided with the distribution.
20 * 3. Neither the name of the University nor the names of its contributors
21 *    may be used to endorse or promote products derived from this software
22 *    without specific prior written permission.
23 *
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * SUCH DAMAGE.
35 *
36 *	@(#)vm_kern.c   8.3 (Berkeley) 1/12/94
37 * from: Id: uvm_km.c,v 1.1.2.14 1998/02/06 05:19:27 chs Exp
38 *
39 *
40 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
41 * All rights reserved.
42 *
43 * Permission to use, copy, modify and distribute this software and
44 * its documentation is hereby granted, provided that both the copyright
45 * notice and this permission notice appear in all copies of the
46 * software, derivative works or modified versions, and any portions
47 * thereof, and that both notices appear in supporting documentation.
48 *
49 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
50 * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
51 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
52 *
53 * Carnegie Mellon requests users of this software to return to
54 *
55 *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
56 *  School of Computer Science
57 *  Carnegie Mellon University
58 *  Pittsburgh PA 15213-3890
59 *
60 * any improvements or extensions that they make and grant Carnegie the
61 * rights to redistribute these changes.
62 */
63
64/*
65 * uvm_km.c: handle kernel memory allocation and management
66 */
67
68/*
69 * overview of kernel memory management:
70 *
71 * the kernel virtual address space is mapped by "kernel_map."   kernel_map
72 * starts at VM_MIN_KERNEL_ADDRESS and goes to VM_MAX_KERNEL_ADDRESS.
73 * note that VM_MIN_KERNEL_ADDRESS is equal to vm_map_min(kernel_map).
74 *
75 * the kernel_map has several "submaps."   submaps can only appear in
76 * the kernel_map (user processes can't use them).   submaps "take over"
77 * the management of a sub-range of the kernel's address space.  submaps
78 * are typically allocated at boot time and are never released.   kernel
79 * virtual address space that is mapped by a submap is locked by the
80 * submap's lock -- not the kernel_map's lock.
81 *
82 * thus, the useful feature of submaps is that they allow us to break
83 * up the locking and protection of the kernel address space into smaller
84 * chunks.
85 *
86 * the vm system has several standard kernel submaps, including:
87 *   kmem_map => contains only wired kernel memory for the kernel
88 *		malloc.
89 *   pager_map => used to map "buf" structures into kernel space
90 *   exec_map => used during exec to handle exec args
91 *   etc...
92 *
93 * the kernel allocates its private memory out of special uvm_objects whose
94 * reference count is set to UVM_OBJ_KERN (thus indicating that the objects
95 * are "special" and never die).   all kernel objects should be thought of
96 * as large, fixed-sized, sparsely populated uvm_objects.   each kernel
97 * object is equal to the size of kernel virtual address space (i.e. the
98 * value "VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS").
99 *
100 * note that just because a kernel object spans the entire kernel virtual
101 * address space doesn't mean that it has to be mapped into the entire space.
102 * large chunks of a kernel object's space go unused either because
103 * that area of kernel VM is unmapped, or there is some other type of
104 * object mapped into that range (e.g. a vnode).    for submap's kernel
105 * objects, the only part of the object that can ever be populated is the
106 * offsets that are managed by the submap.
107 *
108 * note that the "offset" in a kernel object is always the kernel virtual
109 * address minus the VM_MIN_KERNEL_ADDRESS (aka vm_map_min(kernel_map)).
110 * example:
111 *   suppose VM_MIN_KERNEL_ADDRESS is 0xf8000000 and the kernel does a
112 *   uvm_km_alloc(kernel_map, PAGE_SIZE) [allocate 1 wired down page in the
113 *   kernel map].    if uvm_km_alloc returns virtual address 0xf8235000,
114 *   then that means that the page at offset 0x235000 in kernel_object is
115 *   mapped at 0xf8235000.
116 *
117 * kernel object have one other special property: when the kernel virtual
118 * memory mapping them is unmapped, the backing memory in the object is
119 * freed right away.   this is done with the uvm_km_pgremove() function.
120 * this has to be done because there is no backing store for kernel pages
121 * and no need to save them after they are no longer referenced.
122 */
123
124#include <sys/cdefs.h>
125__KERNEL_RCSID(0, "$NetBSD: uvm_km.c,v 1.109 2011/06/12 03:36:03 rmind Exp $");
126
127#include "opt_uvmhist.h"
128
129#include <sys/param.h>
130#include <sys/malloc.h>
131#include <sys/systm.h>
132#include <sys/proc.h>
133#include <sys/pool.h>
134
135#include <uvm/uvm.h>
136
137/*
138 * global data structures
139 */
140
141struct vm_map *kernel_map = NULL;
142
143/*
144 * local data structues
145 */
146
147static struct vm_map_kernel	kernel_map_store;
148static struct vm_map_entry	kernel_first_mapent_store;
149
150#if !defined(PMAP_MAP_POOLPAGE)
151
152/*
153 * kva cache
154 *
155 * XXX maybe it's better to do this at the uvm_map layer.
156 */
157
158#define	KM_VACACHE_SIZE	(32 * PAGE_SIZE) /* XXX tune */
159
160static void *km_vacache_alloc(struct pool *, int);
161static void km_vacache_free(struct pool *, void *);
162static void km_vacache_init(struct vm_map *, const char *, size_t);
163
164/* XXX */
165#define	KM_VACACHE_POOL_TO_MAP(pp) \
166	((struct vm_map *)((char *)(pp) - \
167	    offsetof(struct vm_map_kernel, vmk_vacache)))
168
169static void *
170km_vacache_alloc(struct pool *pp, int flags)
171{
172	vaddr_t va;
173	size_t size;
174	struct vm_map *map;
175	size = pp->pr_alloc->pa_pagesz;
176
177	map = KM_VACACHE_POOL_TO_MAP(pp);
178
179	va = vm_map_min(map); /* hint */
180	if (uvm_map(map, &va, size, NULL, UVM_UNKNOWN_OFFSET, size,
181	    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
182	    UVM_ADV_RANDOM, UVM_FLAG_QUANTUM |
183	    ((flags & PR_WAITOK) ? UVM_FLAG_WAITVA :
184	    UVM_FLAG_TRYLOCK | UVM_FLAG_NOWAIT))))
185		return NULL;
186
187	return (void *)va;
188}
189
190static void
191km_vacache_free(struct pool *pp, void *v)
192{
193	vaddr_t va = (vaddr_t)v;
194	size_t size = pp->pr_alloc->pa_pagesz;
195	struct vm_map *map;
196
197	map = KM_VACACHE_POOL_TO_MAP(pp);
198	uvm_unmap1(map, va, va + size, UVM_FLAG_QUANTUM|UVM_FLAG_VAONLY);
199}
200
201/*
202 * km_vacache_init: initialize kva cache.
203 */
204
205static void
206km_vacache_init(struct vm_map *map, const char *name, size_t size)
207{
208	struct vm_map_kernel *vmk;
209	struct pool *pp;
210	struct pool_allocator *pa;
211	int ipl;
212
213	KASSERT(VM_MAP_IS_KERNEL(map));
214	KASSERT(size < (vm_map_max(map) - vm_map_min(map)) / 2); /* sanity */
215
216
217	vmk = vm_map_to_kernel(map);
218	pp = &vmk->vmk_vacache;
219	pa = &vmk->vmk_vacache_allocator;
220	memset(pa, 0, sizeof(*pa));
221	pa->pa_alloc = km_vacache_alloc;
222	pa->pa_free = km_vacache_free;
223	pa->pa_pagesz = (unsigned int)size;
224	pa->pa_backingmap = map;
225	pa->pa_backingmapptr = NULL;
226
227	if ((map->flags & VM_MAP_INTRSAFE) != 0)
228		ipl = IPL_VM;
229	else
230		ipl = IPL_NONE;
231
232	pool_init(pp, PAGE_SIZE, 0, 0, PR_NOTOUCH | PR_RECURSIVE, name, pa,
233	    ipl);
234}
235
236void
237uvm_km_vacache_init(struct vm_map *map, const char *name, size_t size)
238{
239
240	map->flags |= VM_MAP_VACACHE;
241	if (size == 0)
242		size = KM_VACACHE_SIZE;
243	km_vacache_init(map, name, size);
244}
245
246#else /* !defined(PMAP_MAP_POOLPAGE) */
247
248void
249uvm_km_vacache_init(struct vm_map *map, const char *name, size_t size)
250{
251
252	/* nothing */
253}
254
255#endif /* !defined(PMAP_MAP_POOLPAGE) */
256
257void
258uvm_km_va_drain(struct vm_map *map, uvm_flag_t flags)
259{
260	struct vm_map_kernel *vmk = vm_map_to_kernel(map);
261
262	callback_run_roundrobin(&vmk->vmk_reclaim_callback, NULL);
263}
264
265/*
266 * uvm_km_init: init kernel maps and objects to reflect reality (i.e.
267 * KVM already allocated for text, data, bss, and static data structures).
268 *
269 * => KVM is defined by VM_MIN_KERNEL_ADDRESS/VM_MAX_KERNEL_ADDRESS.
270 *    we assume that [vmin -> start] has already been allocated and that
271 *    "end" is the end.
272 */
273
274void
275uvm_km_init(vaddr_t start, vaddr_t end)
276{
277	vaddr_t base = VM_MIN_KERNEL_ADDRESS;
278
279	/*
280	 * next, init kernel memory objects.
281	 */
282
283	/* kernel_object: for pageable anonymous kernel memory */
284	uao_init();
285	uvm_kernel_object = uao_create(VM_MAX_KERNEL_ADDRESS -
286				 VM_MIN_KERNEL_ADDRESS, UAO_FLAG_KERNOBJ);
287
288	/*
289	 * init the map and reserve any space that might already
290	 * have been allocated kernel space before installing.
291	 */
292
293	uvm_map_setup_kernel(&kernel_map_store, base, end, VM_MAP_PAGEABLE);
294	kernel_map_store.vmk_map.pmap = pmap_kernel();
295	if (start != base) {
296		int error;
297		struct uvm_map_args args;
298
299		error = uvm_map_prepare(&kernel_map_store.vmk_map,
300		    base, start - base,
301		    NULL, UVM_UNKNOWN_OFFSET, 0,
302		    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
303		    		UVM_ADV_RANDOM, UVM_FLAG_FIXED), &args);
304		if (!error) {
305			kernel_first_mapent_store.flags =
306			    UVM_MAP_KERNEL | UVM_MAP_FIRST;
307			error = uvm_map_enter(&kernel_map_store.vmk_map, &args,
308			    &kernel_first_mapent_store);
309		}
310
311		if (error)
312			panic(
313			    "uvm_km_init: could not reserve space for kernel");
314	}
315
316	/*
317	 * install!
318	 */
319
320	kernel_map = &kernel_map_store.vmk_map;
321	uvm_km_vacache_init(kernel_map, "kvakernel", 0);
322}
323
324/*
325 * uvm_km_suballoc: allocate a submap in the kernel map.   once a submap
326 * is allocated all references to that area of VM must go through it.  this
327 * allows the locking of VAs in kernel_map to be broken up into regions.
328 *
329 * => if `fixed' is true, *vmin specifies where the region described
330 *      by the submap must start
331 * => if submap is non NULL we use that as the submap, otherwise we
332 *	alloc a new map
333 */
334
335struct vm_map *
336uvm_km_suballoc(struct vm_map *map, vaddr_t *vmin /* IN/OUT */,
337    vaddr_t *vmax /* OUT */, vsize_t size, int flags, bool fixed,
338    struct vm_map_kernel *submap)
339{
340	int mapflags = UVM_FLAG_NOMERGE | (fixed ? UVM_FLAG_FIXED : 0);
341
342	KASSERT(vm_map_pmap(map) == pmap_kernel());
343
344	size = round_page(size);	/* round up to pagesize */
345	size += uvm_mapent_overhead(size, flags);
346
347	/*
348	 * first allocate a blank spot in the parent map
349	 */
350
351	if (uvm_map(map, vmin, size, NULL, UVM_UNKNOWN_OFFSET, 0,
352	    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
353	    UVM_ADV_RANDOM, mapflags)) != 0) {
354	       panic("uvm_km_suballoc: unable to allocate space in parent map");
355	}
356
357	/*
358	 * set VM bounds (vmin is filled in by uvm_map)
359	 */
360
361	*vmax = *vmin + size;
362
363	/*
364	 * add references to pmap and create or init the submap
365	 */
366
367	pmap_reference(vm_map_pmap(map));
368	if (submap == NULL) {
369		submap = malloc(sizeof(*submap), M_VMMAP, M_WAITOK);
370		if (submap == NULL)
371			panic("uvm_km_suballoc: unable to create submap");
372	}
373	uvm_map_setup_kernel(submap, *vmin, *vmax, flags);
374	submap->vmk_map.pmap = vm_map_pmap(map);
375
376	/*
377	 * now let uvm_map_submap plug in it...
378	 */
379
380	if (uvm_map_submap(map, *vmin, *vmax, &submap->vmk_map) != 0)
381		panic("uvm_km_suballoc: submap allocation failed");
382
383	return(&submap->vmk_map);
384}
385
386/*
387 * uvm_km_pgremove: remove pages from a kernel uvm_object.
388 *
389 * => when you unmap a part of anonymous kernel memory you want to toss
390 *    the pages right away.    (this gets called from uvm_unmap_...).
391 */
392
393void
394uvm_km_pgremove(vaddr_t startva, vaddr_t endva)
395{
396	struct uvm_object * const uobj = uvm_kernel_object;
397	const voff_t start = startva - vm_map_min(kernel_map);
398	const voff_t end = endva - vm_map_min(kernel_map);
399	struct vm_page *pg;
400	voff_t curoff, nextoff;
401	int swpgonlydelta = 0;
402	UVMHIST_FUNC("uvm_km_pgremove"); UVMHIST_CALLED(maphist);
403
404	KASSERT(VM_MIN_KERNEL_ADDRESS <= startva);
405	KASSERT(startva < endva);
406	KASSERT(endva <= VM_MAX_KERNEL_ADDRESS);
407
408	mutex_enter(uobj->vmobjlock);
409
410	for (curoff = start; curoff < end; curoff = nextoff) {
411		nextoff = curoff + PAGE_SIZE;
412		pg = uvm_pagelookup(uobj, curoff);
413		if (pg != NULL && pg->flags & PG_BUSY) {
414			pg->flags |= PG_WANTED;
415			UVM_UNLOCK_AND_WAIT(pg, uobj->vmobjlock, 0,
416				    "km_pgrm", 0);
417			mutex_enter(uobj->vmobjlock);
418			nextoff = curoff;
419			continue;
420		}
421
422		/*
423		 * free the swap slot, then the page.
424		 */
425
426		if (pg == NULL &&
427		    uao_find_swslot(uobj, curoff >> PAGE_SHIFT) > 0) {
428			swpgonlydelta++;
429		}
430		uao_dropswap(uobj, curoff >> PAGE_SHIFT);
431		if (pg != NULL) {
432			mutex_enter(&uvm_pageqlock);
433			uvm_pagefree(pg);
434			mutex_exit(&uvm_pageqlock);
435		}
436	}
437	mutex_exit(uobj->vmobjlock);
438
439	if (swpgonlydelta > 0) {
440		mutex_enter(&uvm_swap_data_lock);
441		KASSERT(uvmexp.swpgonly >= swpgonlydelta);
442		uvmexp.swpgonly -= swpgonlydelta;
443		mutex_exit(&uvm_swap_data_lock);
444	}
445}
446
447
448/*
449 * uvm_km_pgremove_intrsafe: like uvm_km_pgremove(), but for non object backed
450 *    regions.
451 *
452 * => when you unmap a part of anonymous kernel memory you want to toss
453 *    the pages right away.    (this is called from uvm_unmap_...).
454 * => none of the pages will ever be busy, and none of them will ever
455 *    be on the active or inactive queues (because they have no object).
456 */
457
458void
459uvm_km_pgremove_intrsafe(struct vm_map *map, vaddr_t start, vaddr_t end)
460{
461	struct vm_page *pg;
462	paddr_t pa;
463	UVMHIST_FUNC("uvm_km_pgremove_intrsafe"); UVMHIST_CALLED(maphist);
464
465	KASSERT(VM_MAP_IS_KERNEL(map));
466	KASSERT(vm_map_min(map) <= start);
467	KASSERT(start < end);
468	KASSERT(end <= vm_map_max(map));
469
470	for (; start < end; start += PAGE_SIZE) {
471		if (!pmap_extract(pmap_kernel(), start, &pa)) {
472			continue;
473		}
474		pg = PHYS_TO_VM_PAGE(pa);
475		KASSERT(pg);
476		KASSERT(pg->uobject == NULL && pg->uanon == NULL);
477		uvm_pagefree(pg);
478	}
479}
480
481#if defined(DEBUG)
482void
483uvm_km_check_empty(struct vm_map *map, vaddr_t start, vaddr_t end)
484{
485	struct vm_page *pg;
486	vaddr_t va;
487	paddr_t pa;
488
489	KDASSERT(VM_MAP_IS_KERNEL(map));
490	KDASSERT(vm_map_min(map) <= start);
491	KDASSERT(start < end);
492	KDASSERT(end <= vm_map_max(map));
493
494	for (va = start; va < end; va += PAGE_SIZE) {
495		if (pmap_extract(pmap_kernel(), va, &pa)) {
496			panic("uvm_km_check_empty: va %p has pa 0x%llx",
497			    (void *)va, (long long)pa);
498		}
499		if ((map->flags & VM_MAP_INTRSAFE) == 0) {
500			mutex_enter(uvm_kernel_object->vmobjlock);
501			pg = uvm_pagelookup(uvm_kernel_object,
502			    va - vm_map_min(kernel_map));
503			mutex_exit(uvm_kernel_object->vmobjlock);
504			if (pg) {
505				panic("uvm_km_check_empty: "
506				    "has page hashed at %p", (const void *)va);
507			}
508		}
509	}
510}
511#endif /* defined(DEBUG) */
512
513/*
514 * uvm_km_alloc: allocate an area of kernel memory.
515 *
516 * => NOTE: we can return 0 even if we can wait if there is not enough
517 *	free VM space in the map... caller should be prepared to handle
518 *	this case.
519 * => we return KVA of memory allocated
520 */
521
522vaddr_t
523uvm_km_alloc(struct vm_map *map, vsize_t size, vsize_t align, uvm_flag_t flags)
524{
525	vaddr_t kva, loopva;
526	vaddr_t offset;
527	vsize_t loopsize;
528	struct vm_page *pg;
529	struct uvm_object *obj;
530	int pgaflags;
531	vm_prot_t prot;
532	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
533
534	KASSERT(vm_map_pmap(map) == pmap_kernel());
535	KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED ||
536		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE ||
537		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY);
538
539	/*
540	 * setup for call
541	 */
542
543	kva = vm_map_min(map);	/* hint */
544	size = round_page(size);
545	obj = (flags & UVM_KMF_PAGEABLE) ? uvm_kernel_object : NULL;
546	UVMHIST_LOG(maphist,"  (map=0x%x, obj=0x%x, size=0x%x, flags=%d)",
547		    map, obj, size, flags);
548
549	/*
550	 * allocate some virtual space
551	 */
552
553	if (__predict_false(uvm_map(map, &kva, size, obj, UVM_UNKNOWN_OFFSET,
554	    align, UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
555	    UVM_ADV_RANDOM,
556	    (flags & (UVM_KMF_TRYLOCK | UVM_KMF_NOWAIT | UVM_KMF_WAITVA))
557	    | UVM_FLAG_QUANTUM)) != 0)) {
558		UVMHIST_LOG(maphist, "<- done (no VM)",0,0,0,0);
559		return(0);
560	}
561
562	/*
563	 * if all we wanted was VA, return now
564	 */
565
566	if (flags & (UVM_KMF_VAONLY | UVM_KMF_PAGEABLE)) {
567		UVMHIST_LOG(maphist,"<- done valloc (kva=0x%x)", kva,0,0,0);
568		return(kva);
569	}
570
571	/*
572	 * recover object offset from virtual address
573	 */
574
575	offset = kva - vm_map_min(kernel_map);
576	UVMHIST_LOG(maphist, "  kva=0x%x, offset=0x%x", kva, offset,0,0);
577
578	/*
579	 * now allocate and map in the memory... note that we are the only ones
580	 * whom should ever get a handle on this area of VM.
581	 */
582
583	loopva = kva;
584	loopsize = size;
585
586	pgaflags = UVM_FLAG_COLORMATCH;
587	if (flags & UVM_KMF_NOWAIT)
588		pgaflags |= UVM_PGA_USERESERVE;
589	if (flags & UVM_KMF_ZERO)
590		pgaflags |= UVM_PGA_ZERO;
591	prot = VM_PROT_READ | VM_PROT_WRITE;
592	if (flags & UVM_KMF_EXEC)
593		prot |= VM_PROT_EXECUTE;
594	while (loopsize) {
595		KASSERT(!pmap_extract(pmap_kernel(), loopva, NULL));
596
597		pg = uvm_pagealloc_strat(NULL, offset, NULL, pgaflags,
598#ifdef UVM_KM_VMFREELIST
599		   UVM_PGA_STRAT_ONLY, UVM_KM_VMFREELIST
600#else
601		   UVM_PGA_STRAT_NORMAL, 0
602#endif
603		   );
604
605		/*
606		 * out of memory?
607		 */
608
609		if (__predict_false(pg == NULL)) {
610			if ((flags & UVM_KMF_NOWAIT) ||
611			    ((flags & UVM_KMF_CANFAIL) && !uvm_reclaimable())) {
612				/* free everything! */
613				uvm_km_free(map, kva, size,
614				    flags & UVM_KMF_TYPEMASK);
615				return (0);
616			} else {
617				uvm_wait("km_getwait2");	/* sleep here */
618				continue;
619			}
620		}
621
622		pg->flags &= ~PG_BUSY;	/* new page */
623		UVM_PAGE_OWN(pg, NULL);
624
625		/*
626		 * map it in
627		 */
628
629		pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg),
630		    prot, PMAP_KMPAGE);
631		loopva += PAGE_SIZE;
632		offset += PAGE_SIZE;
633		loopsize -= PAGE_SIZE;
634	}
635
636       	pmap_update(pmap_kernel());
637
638	UVMHIST_LOG(maphist,"<- done (kva=0x%x)", kva,0,0,0);
639	return(kva);
640}
641
642/*
643 * uvm_km_free: free an area of kernel memory
644 */
645
646void
647uvm_km_free(struct vm_map *map, vaddr_t addr, vsize_t size, uvm_flag_t flags)
648{
649
650	KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED ||
651		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE ||
652		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY);
653	KASSERT((addr & PAGE_MASK) == 0);
654	KASSERT(vm_map_pmap(map) == pmap_kernel());
655
656	size = round_page(size);
657
658
659	if (flags & UVM_KMF_PAGEABLE) {
660		/*
661		 * No need to lock for pmap, since the kernel is always
662		 * self-consistent.  The pages cannot be in use elsewhere.
663		 */
664		uvm_km_pgremove(addr, addr + size);
665		pmap_remove(pmap_kernel(), addr, addr + size);
666
667	} else if (flags & UVM_KMF_WIRED) {
668		/*
669		 * Note: uvm_km_pgremove_intrsafe() extracts mapping, thus
670		 * remove it after.  See comment below about KVA visibility.
671		 */
672		uvm_km_pgremove_intrsafe(map, addr, addr + size);
673		pmap_kremove(addr, size);
674	}
675
676	/*
677	 * Note: uvm_unmap_remove() calls pmap_update() for us, before
678	 * KVA becomes globally available.
679	 */
680
681	uvm_unmap1(map, addr, addr + size, UVM_FLAG_QUANTUM|UVM_FLAG_VAONLY);
682}
683
684/* Sanity; must specify both or none. */
685#if (defined(PMAP_MAP_POOLPAGE) || defined(PMAP_UNMAP_POOLPAGE)) && \
686    (!defined(PMAP_MAP_POOLPAGE) || !defined(PMAP_UNMAP_POOLPAGE))
687#error Must specify MAP and UNMAP together.
688#endif
689
690/*
691 * uvm_km_alloc_poolpage: allocate a page for the pool allocator
692 *
693 * => if the pmap specifies an alternate mapping method, we use it.
694 */
695
696/* ARGSUSED */
697vaddr_t
698uvm_km_alloc_poolpage_cache(struct vm_map *map, bool waitok)
699{
700#if defined(PMAP_MAP_POOLPAGE)
701	return uvm_km_alloc_poolpage(map, waitok);
702#else
703	struct vm_page *pg;
704	struct pool *pp = &vm_map_to_kernel(map)->vmk_vacache;
705	vaddr_t va;
706
707	if ((map->flags & VM_MAP_VACACHE) == 0)
708		return uvm_km_alloc_poolpage(map, waitok);
709
710	va = (vaddr_t)pool_get(pp, waitok ? PR_WAITOK : PR_NOWAIT);
711	if (va == 0)
712		return 0;
713	KASSERT(!pmap_extract(pmap_kernel(), va, NULL));
714again:
715	pg = uvm_pagealloc(NULL, 0, NULL, waitok ? 0 : UVM_PGA_USERESERVE);
716	if (__predict_false(pg == NULL)) {
717		if (waitok) {
718			uvm_wait("plpg");
719			goto again;
720		} else {
721			pool_put(pp, (void *)va);
722			return 0;
723		}
724	}
725	pmap_kenter_pa(va, VM_PAGE_TO_PHYS(pg),
726	    VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE);
727	pmap_update(pmap_kernel());
728
729	return va;
730#endif /* PMAP_MAP_POOLPAGE */
731}
732
733vaddr_t
734uvm_km_alloc_poolpage(struct vm_map *map, bool waitok)
735{
736#if defined(PMAP_MAP_POOLPAGE)
737	struct vm_page *pg;
738	vaddr_t va;
739
740
741 again:
742#ifdef PMAP_ALLOC_POOLPAGE
743	pg = PMAP_ALLOC_POOLPAGE(waitok ? 0 : UVM_PGA_USERESERVE);
744#else
745	pg = uvm_pagealloc(NULL, 0, NULL, waitok ? 0 : UVM_PGA_USERESERVE);
746#endif
747	if (__predict_false(pg == NULL)) {
748		if (waitok) {
749			uvm_wait("plpg");
750			goto again;
751		} else
752			return (0);
753	}
754	va = PMAP_MAP_POOLPAGE(VM_PAGE_TO_PHYS(pg));
755	if (__predict_false(va == 0))
756		uvm_pagefree(pg);
757	return (va);
758#else
759	vaddr_t va;
760
761	va = uvm_km_alloc(map, PAGE_SIZE, 0,
762	    (waitok ? 0 : UVM_KMF_NOWAIT | UVM_KMF_TRYLOCK) | UVM_KMF_WIRED);
763	return (va);
764#endif /* PMAP_MAP_POOLPAGE */
765}
766
767/*
768 * uvm_km_free_poolpage: free a previously allocated pool page
769 *
770 * => if the pmap specifies an alternate unmapping method, we use it.
771 */
772
773/* ARGSUSED */
774void
775uvm_km_free_poolpage_cache(struct vm_map *map, vaddr_t addr)
776{
777#if defined(PMAP_UNMAP_POOLPAGE)
778	uvm_km_free_poolpage(map, addr);
779#else
780	struct pool *pp;
781
782	if ((map->flags & VM_MAP_VACACHE) == 0) {
783		uvm_km_free_poolpage(map, addr);
784		return;
785	}
786
787	KASSERT(pmap_extract(pmap_kernel(), addr, NULL));
788	uvm_km_pgremove_intrsafe(map, addr, addr + PAGE_SIZE);
789	pmap_kremove(addr, PAGE_SIZE);
790#if defined(DEBUG)
791	pmap_update(pmap_kernel());
792#endif
793	KASSERT(!pmap_extract(pmap_kernel(), addr, NULL));
794	pp = &vm_map_to_kernel(map)->vmk_vacache;
795	pool_put(pp, (void *)addr);
796#endif
797}
798
799/* ARGSUSED */
800void
801uvm_km_free_poolpage(struct vm_map *map, vaddr_t addr)
802{
803#if defined(PMAP_UNMAP_POOLPAGE)
804	paddr_t pa;
805
806	pa = PMAP_UNMAP_POOLPAGE(addr);
807	uvm_pagefree(PHYS_TO_VM_PAGE(pa));
808#else
809	uvm_km_free(map, addr, PAGE_SIZE, UVM_KMF_WIRED);
810#endif /* PMAP_UNMAP_POOLPAGE */
811}
812