xref: /netbsd-src/sys/uvm/uvm_km.c (revision 1b9578b8c2c1f848eeb16dabbfd7d1f0d9fdefbd)
1 /*	$NetBSD: uvm_km.c,v 1.110 2011/07/05 14:03:06 yamt 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 1.110 2011/07/05 14:03:06 yamt 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 
141 struct vm_map *kernel_map = NULL;
142 
143 /*
144  * local data structues
145  */
146 
147 static struct vm_map_kernel	kernel_map_store;
148 static 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 
160 static void *km_vacache_alloc(struct pool *, int);
161 static void km_vacache_free(struct pool *, void *);
162 static 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 
169 static void *
170 km_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 
190 static void
191 km_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 
205 static void
206 km_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 
236 void
237 uvm_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 
248 void
249 uvm_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 
257 void
258 uvm_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 
274 void
275 uvm_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 
335 struct vm_map *
336 uvm_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 and KVA.
388  */
389 
390 void
391 uvm_km_pgremove(vaddr_t startva, vaddr_t endva)
392 {
393 	struct uvm_object * const uobj = uvm_kernel_object;
394 	const voff_t start = startva - vm_map_min(kernel_map);
395 	const voff_t end = endva - vm_map_min(kernel_map);
396 	struct vm_page *pg;
397 	voff_t curoff, nextoff;
398 	int swpgonlydelta = 0;
399 	UVMHIST_FUNC("uvm_km_pgremove"); UVMHIST_CALLED(maphist);
400 
401 	KASSERT(VM_MIN_KERNEL_ADDRESS <= startva);
402 	KASSERT(startva < endva);
403 	KASSERT(endva <= VM_MAX_KERNEL_ADDRESS);
404 
405 	mutex_enter(uobj->vmobjlock);
406 	pmap_remove(pmap_kernel(), startva, endva);
407 	for (curoff = start; curoff < end; curoff = nextoff) {
408 		nextoff = curoff + PAGE_SIZE;
409 		pg = uvm_pagelookup(uobj, curoff);
410 		if (pg != NULL && pg->flags & PG_BUSY) {
411 			pg->flags |= PG_WANTED;
412 			UVM_UNLOCK_AND_WAIT(pg, uobj->vmobjlock, 0,
413 				    "km_pgrm", 0);
414 			mutex_enter(uobj->vmobjlock);
415 			nextoff = curoff;
416 			continue;
417 		}
418 
419 		/*
420 		 * free the swap slot, then the page.
421 		 */
422 
423 		if (pg == NULL &&
424 		    uao_find_swslot(uobj, curoff >> PAGE_SHIFT) > 0) {
425 			swpgonlydelta++;
426 		}
427 		uao_dropswap(uobj, curoff >> PAGE_SHIFT);
428 		if (pg != NULL) {
429 			mutex_enter(&uvm_pageqlock);
430 			uvm_pagefree(pg);
431 			mutex_exit(&uvm_pageqlock);
432 		}
433 	}
434 	mutex_exit(uobj->vmobjlock);
435 
436 	if (swpgonlydelta > 0) {
437 		mutex_enter(&uvm_swap_data_lock);
438 		KASSERT(uvmexp.swpgonly >= swpgonlydelta);
439 		uvmexp.swpgonly -= swpgonlydelta;
440 		mutex_exit(&uvm_swap_data_lock);
441 	}
442 }
443 
444 
445 /*
446  * uvm_km_pgremove_intrsafe: like uvm_km_pgremove(), but for non object backed
447  *    regions.
448  *
449  * => when you unmap a part of anonymous kernel memory you want to toss
450  *    the pages right away.    (this is called from uvm_unmap_...).
451  * => none of the pages will ever be busy, and none of them will ever
452  *    be on the active or inactive queues (because they have no object).
453  */
454 
455 void
456 uvm_km_pgremove_intrsafe(struct vm_map *map, vaddr_t start, vaddr_t end)
457 {
458 	struct vm_page *pg;
459 	paddr_t pa;
460 	UVMHIST_FUNC("uvm_km_pgremove_intrsafe"); UVMHIST_CALLED(maphist);
461 
462 	KASSERT(VM_MAP_IS_KERNEL(map));
463 	KASSERT(vm_map_min(map) <= start);
464 	KASSERT(start < end);
465 	KASSERT(end <= vm_map_max(map));
466 
467 	for (; start < end; start += PAGE_SIZE) {
468 		if (!pmap_extract(pmap_kernel(), start, &pa)) {
469 			continue;
470 		}
471 		pg = PHYS_TO_VM_PAGE(pa);
472 		KASSERT(pg);
473 		KASSERT(pg->uobject == NULL && pg->uanon == NULL);
474 		KASSERT((pg->flags & PG_BUSY) == 0);
475 		uvm_pagefree(pg);
476 	}
477 }
478 
479 #if defined(DEBUG)
480 void
481 uvm_km_check_empty(struct vm_map *map, vaddr_t start, vaddr_t end)
482 {
483 	struct vm_page *pg;
484 	vaddr_t va;
485 	paddr_t pa;
486 
487 	KDASSERT(VM_MAP_IS_KERNEL(map));
488 	KDASSERT(vm_map_min(map) <= start);
489 	KDASSERT(start < end);
490 	KDASSERT(end <= vm_map_max(map));
491 
492 	for (va = start; va < end; va += PAGE_SIZE) {
493 		if (pmap_extract(pmap_kernel(), va, &pa)) {
494 			panic("uvm_km_check_empty: va %p has pa 0x%llx",
495 			    (void *)va, (long long)pa);
496 		}
497 		if ((map->flags & VM_MAP_INTRSAFE) == 0) {
498 			mutex_enter(uvm_kernel_object->vmobjlock);
499 			pg = uvm_pagelookup(uvm_kernel_object,
500 			    va - vm_map_min(kernel_map));
501 			mutex_exit(uvm_kernel_object->vmobjlock);
502 			if (pg) {
503 				panic("uvm_km_check_empty: "
504 				    "has page hashed at %p", (const void *)va);
505 			}
506 		}
507 	}
508 }
509 #endif /* defined(DEBUG) */
510 
511 /*
512  * uvm_km_alloc: allocate an area of kernel memory.
513  *
514  * => NOTE: we can return 0 even if we can wait if there is not enough
515  *	free VM space in the map... caller should be prepared to handle
516  *	this case.
517  * => we return KVA of memory allocated
518  */
519 
520 vaddr_t
521 uvm_km_alloc(struct vm_map *map, vsize_t size, vsize_t align, uvm_flag_t flags)
522 {
523 	vaddr_t kva, loopva;
524 	vaddr_t offset;
525 	vsize_t loopsize;
526 	struct vm_page *pg;
527 	struct uvm_object *obj;
528 	int pgaflags;
529 	vm_prot_t prot;
530 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
531 
532 	KASSERT(vm_map_pmap(map) == pmap_kernel());
533 	KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED ||
534 		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE ||
535 		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY);
536 
537 	/*
538 	 * setup for call
539 	 */
540 
541 	kva = vm_map_min(map);	/* hint */
542 	size = round_page(size);
543 	obj = (flags & UVM_KMF_PAGEABLE) ? uvm_kernel_object : NULL;
544 	UVMHIST_LOG(maphist,"  (map=0x%x, obj=0x%x, size=0x%x, flags=%d)",
545 		    map, obj, size, flags);
546 
547 	/*
548 	 * allocate some virtual space
549 	 */
550 
551 	if (__predict_false(uvm_map(map, &kva, size, obj, UVM_UNKNOWN_OFFSET,
552 	    align, UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
553 	    UVM_ADV_RANDOM,
554 	    (flags & (UVM_KMF_TRYLOCK | UVM_KMF_NOWAIT | UVM_KMF_WAITVA))
555 	    | UVM_FLAG_QUANTUM)) != 0)) {
556 		UVMHIST_LOG(maphist, "<- done (no VM)",0,0,0,0);
557 		return(0);
558 	}
559 
560 	/*
561 	 * if all we wanted was VA, return now
562 	 */
563 
564 	if (flags & (UVM_KMF_VAONLY | UVM_KMF_PAGEABLE)) {
565 		UVMHIST_LOG(maphist,"<- done valloc (kva=0x%x)", kva,0,0,0);
566 		return(kva);
567 	}
568 
569 	/*
570 	 * recover object offset from virtual address
571 	 */
572 
573 	offset = kva - vm_map_min(kernel_map);
574 	UVMHIST_LOG(maphist, "  kva=0x%x, offset=0x%x", kva, offset,0,0);
575 
576 	/*
577 	 * now allocate and map in the memory... note that we are the only ones
578 	 * whom should ever get a handle on this area of VM.
579 	 */
580 
581 	loopva = kva;
582 	loopsize = size;
583 
584 	pgaflags = UVM_FLAG_COLORMATCH;
585 	if (flags & UVM_KMF_NOWAIT)
586 		pgaflags |= UVM_PGA_USERESERVE;
587 	if (flags & UVM_KMF_ZERO)
588 		pgaflags |= UVM_PGA_ZERO;
589 	prot = VM_PROT_READ | VM_PROT_WRITE;
590 	if (flags & UVM_KMF_EXEC)
591 		prot |= VM_PROT_EXECUTE;
592 	while (loopsize) {
593 		KASSERT(!pmap_extract(pmap_kernel(), loopva, NULL));
594 
595 		pg = uvm_pagealloc_strat(NULL, offset, NULL, pgaflags,
596 #ifdef UVM_KM_VMFREELIST
597 		   UVM_PGA_STRAT_ONLY, UVM_KM_VMFREELIST
598 #else
599 		   UVM_PGA_STRAT_NORMAL, 0
600 #endif
601 		   );
602 
603 		/*
604 		 * out of memory?
605 		 */
606 
607 		if (__predict_false(pg == NULL)) {
608 			if ((flags & UVM_KMF_NOWAIT) ||
609 			    ((flags & UVM_KMF_CANFAIL) && !uvm_reclaimable())) {
610 				/* free everything! */
611 				uvm_km_free(map, kva, size,
612 				    flags & UVM_KMF_TYPEMASK);
613 				return (0);
614 			} else {
615 				uvm_wait("km_getwait2");	/* sleep here */
616 				continue;
617 			}
618 		}
619 
620 		pg->flags &= ~PG_BUSY;	/* new page */
621 		UVM_PAGE_OWN(pg, NULL);
622 
623 		/*
624 		 * map it in
625 		 */
626 
627 		pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg),
628 		    prot, PMAP_KMPAGE);
629 		loopva += PAGE_SIZE;
630 		offset += PAGE_SIZE;
631 		loopsize -= PAGE_SIZE;
632 	}
633 
634        	pmap_update(pmap_kernel());
635 
636 	UVMHIST_LOG(maphist,"<- done (kva=0x%x)", kva,0,0,0);
637 	return(kva);
638 }
639 
640 /*
641  * uvm_km_free: free an area of kernel memory
642  */
643 
644 void
645 uvm_km_free(struct vm_map *map, vaddr_t addr, vsize_t size, uvm_flag_t flags)
646 {
647 
648 	KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED ||
649 		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE ||
650 		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY);
651 	KASSERT((addr & PAGE_MASK) == 0);
652 	KASSERT(vm_map_pmap(map) == pmap_kernel());
653 
654 	size = round_page(size);
655 
656 	if (flags & UVM_KMF_PAGEABLE) {
657 		uvm_km_pgremove(addr, addr + size);
658 	} else if (flags & UVM_KMF_WIRED) {
659 		/*
660 		 * Note: uvm_km_pgremove_intrsafe() extracts mapping, thus
661 		 * remove it after.  See comment below about KVA visibility.
662 		 */
663 		uvm_km_pgremove_intrsafe(map, addr, addr + size);
664 		pmap_kremove(addr, size);
665 	}
666 
667 	/*
668 	 * Note: uvm_unmap_remove() calls pmap_update() for us, before
669 	 * KVA becomes globally available.
670 	 */
671 
672 	uvm_unmap1(map, addr, addr + size, UVM_FLAG_QUANTUM|UVM_FLAG_VAONLY);
673 }
674 
675 /* Sanity; must specify both or none. */
676 #if (defined(PMAP_MAP_POOLPAGE) || defined(PMAP_UNMAP_POOLPAGE)) && \
677     (!defined(PMAP_MAP_POOLPAGE) || !defined(PMAP_UNMAP_POOLPAGE))
678 #error Must specify MAP and UNMAP together.
679 #endif
680 
681 /*
682  * uvm_km_alloc_poolpage: allocate a page for the pool allocator
683  *
684  * => if the pmap specifies an alternate mapping method, we use it.
685  */
686 
687 /* ARGSUSED */
688 vaddr_t
689 uvm_km_alloc_poolpage_cache(struct vm_map *map, bool waitok)
690 {
691 #if defined(PMAP_MAP_POOLPAGE)
692 	return uvm_km_alloc_poolpage(map, waitok);
693 #else
694 	struct vm_page *pg;
695 	struct pool *pp = &vm_map_to_kernel(map)->vmk_vacache;
696 	vaddr_t va;
697 
698 	if ((map->flags & VM_MAP_VACACHE) == 0)
699 		return uvm_km_alloc_poolpage(map, waitok);
700 
701 	va = (vaddr_t)pool_get(pp, waitok ? PR_WAITOK : PR_NOWAIT);
702 	if (va == 0)
703 		return 0;
704 	KASSERT(!pmap_extract(pmap_kernel(), va, NULL));
705 again:
706 	pg = uvm_pagealloc(NULL, 0, NULL, waitok ? 0 : UVM_PGA_USERESERVE);
707 	if (__predict_false(pg == NULL)) {
708 		if (waitok) {
709 			uvm_wait("plpg");
710 			goto again;
711 		} else {
712 			pool_put(pp, (void *)va);
713 			return 0;
714 		}
715 	}
716 	pg->flags &= ~PG_BUSY;	/* new page */
717 	UVM_PAGE_OWN(pg, NULL);
718 	pmap_kenter_pa(va, VM_PAGE_TO_PHYS(pg),
719 	    VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE);
720 	pmap_update(pmap_kernel());
721 
722 	return va;
723 #endif /* PMAP_MAP_POOLPAGE */
724 }
725 
726 vaddr_t
727 uvm_km_alloc_poolpage(struct vm_map *map, bool waitok)
728 {
729 #if defined(PMAP_MAP_POOLPAGE)
730 	struct vm_page *pg;
731 	vaddr_t va;
732 
733 
734  again:
735 #ifdef PMAP_ALLOC_POOLPAGE
736 	pg = PMAP_ALLOC_POOLPAGE(waitok ? 0 : UVM_PGA_USERESERVE);
737 #else
738 	pg = uvm_pagealloc(NULL, 0, NULL, waitok ? 0 : UVM_PGA_USERESERVE);
739 #endif
740 	if (__predict_false(pg == NULL)) {
741 		if (waitok) {
742 			uvm_wait("plpg");
743 			goto again;
744 		} else
745 			return (0);
746 	}
747 	va = PMAP_MAP_POOLPAGE(VM_PAGE_TO_PHYS(pg));
748 	if (__predict_false(va == 0))
749 		uvm_pagefree(pg);
750 	return (va);
751 #else
752 	vaddr_t va;
753 
754 	va = uvm_km_alloc(map, PAGE_SIZE, 0,
755 	    (waitok ? 0 : UVM_KMF_NOWAIT | UVM_KMF_TRYLOCK) | UVM_KMF_WIRED);
756 	return (va);
757 #endif /* PMAP_MAP_POOLPAGE */
758 }
759 
760 /*
761  * uvm_km_free_poolpage: free a previously allocated pool page
762  *
763  * => if the pmap specifies an alternate unmapping method, we use it.
764  */
765 
766 /* ARGSUSED */
767 void
768 uvm_km_free_poolpage_cache(struct vm_map *map, vaddr_t addr)
769 {
770 #if defined(PMAP_UNMAP_POOLPAGE)
771 	uvm_km_free_poolpage(map, addr);
772 #else
773 	struct pool *pp;
774 
775 	if ((map->flags & VM_MAP_VACACHE) == 0) {
776 		uvm_km_free_poolpage(map, addr);
777 		return;
778 	}
779 
780 	KASSERT(pmap_extract(pmap_kernel(), addr, NULL));
781 	uvm_km_pgremove_intrsafe(map, addr, addr + PAGE_SIZE);
782 	pmap_kremove(addr, PAGE_SIZE);
783 #if defined(DEBUG)
784 	pmap_update(pmap_kernel());
785 #endif
786 	KASSERT(!pmap_extract(pmap_kernel(), addr, NULL));
787 	pp = &vm_map_to_kernel(map)->vmk_vacache;
788 	pool_put(pp, (void *)addr);
789 #endif
790 }
791 
792 /* ARGSUSED */
793 void
794 uvm_km_free_poolpage(struct vm_map *map, vaddr_t addr)
795 {
796 #if defined(PMAP_UNMAP_POOLPAGE)
797 	paddr_t pa;
798 
799 	pa = PMAP_UNMAP_POOLPAGE(addr);
800 	uvm_pagefree(PHYS_TO_VM_PAGE(pa));
801 #else
802 	uvm_km_free(map, addr, PAGE_SIZE, UVM_KMF_WIRED);
803 #endif /* PMAP_UNMAP_POOLPAGE */
804 }
805