uvm_km.c revision 1.114
1/*	$NetBSD: uvm_km.c,v 1.114 2012/01/31 00:30:52 matt 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 *   pager_map => used to map "buf" structures into kernel space
88 *   exec_map => used during exec to handle exec args
89 *   etc...
90 *
91 * the kernel allocates its private memory out of special uvm_objects whose
92 * reference count is set to UVM_OBJ_KERN (thus indicating that the objects
93 * are "special" and never die).   all kernel objects should be thought of
94 * as large, fixed-sized, sparsely populated uvm_objects.   each kernel
95 * object is equal to the size of kernel virtual address space (i.e. the
96 * value "VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS").
97 *
98 * note that just because a kernel object spans the entire kernel virtual
99 * address space doesn't mean that it has to be mapped into the entire space.
100 * large chunks of a kernel object's space go unused either because
101 * that area of kernel VM is unmapped, or there is some other type of
102 * object mapped into that range (e.g. a vnode).    for submap's kernel
103 * objects, the only part of the object that can ever be populated is the
104 * offsets that are managed by the submap.
105 *
106 * note that the "offset" in a kernel object is always the kernel virtual
107 * address minus the VM_MIN_KERNEL_ADDRESS (aka vm_map_min(kernel_map)).
108 * example:
109 *   suppose VM_MIN_KERNEL_ADDRESS is 0xf8000000 and the kernel does a
110 *   uvm_km_alloc(kernel_map, PAGE_SIZE) [allocate 1 wired down page in the
111 *   kernel map].    if uvm_km_alloc returns virtual address 0xf8235000,
112 *   then that means that the page at offset 0x235000 in kernel_object is
113 *   mapped at 0xf8235000.
114 *
115 * kernel object have one other special property: when the kernel virtual
116 * memory mapping them is unmapped, the backing memory in the object is
117 * freed right away.   this is done with the uvm_km_pgremove() function.
118 * this has to be done because there is no backing store for kernel pages
119 * and no need to save them after they are no longer referenced.
120 */
121
122#include <sys/cdefs.h>
123__KERNEL_RCSID(0, "$NetBSD: uvm_km.c,v 1.114 2012/01/31 00:30:52 matt Exp $");
124
125#include "opt_uvmhist.h"
126
127#include <sys/param.h>
128#include <sys/systm.h>
129#include <sys/proc.h>
130#include <sys/pool.h>
131#include <sys/vmem.h>
132#include <sys/kmem.h>
133
134#include <uvm/uvm.h>
135
136/*
137 * global data structures
138 */
139
140struct vm_map *kernel_map = NULL;
141
142/*
143 * local data structues
144 */
145
146static struct vm_map		kernel_map_store;
147static struct vm_map_entry	kernel_image_mapent_store;
148static struct vm_map_entry	kernel_kmem_mapent_store;
149
150vaddr_t kmembase;
151vsize_t kmemsize;
152
153vmem_t *kmem_arena;
154vmem_t *kmem_va_arena;
155
156/*
157 * uvm_km_bootstrap: init kernel maps and objects to reflect reality (i.e.
158 * KVM already allocated for text, data, bss, and static data structures).
159 *
160 * => KVM is defined by VM_MIN_KERNEL_ADDRESS/VM_MAX_KERNEL_ADDRESS.
161 *    we assume that [vmin -> start] has already been allocated and that
162 *    "end" is the end.
163 */
164
165void
166uvm_km_bootstrap(vaddr_t start, vaddr_t end)
167{
168	vaddr_t base = VM_MIN_KERNEL_ADDRESS;
169
170	kmemsize = MIN(((((vsize_t)(end - start)) / 3) * 2),
171	    ((((vsize_t)uvmexp.npages) * PAGE_SIZE)));
172	kmemsize = round_page(kmemsize);
173
174	/*
175	 * next, init kernel memory objects.
176	 */
177
178	/* kernel_object: for pageable anonymous kernel memory */
179	uvm_kernel_object = uao_create(VM_MAX_KERNEL_ADDRESS -
180				VM_MIN_KERNEL_ADDRESS, UAO_FLAG_KERNOBJ);
181
182	/*
183	 * init the map and reserve any space that might already
184	 * have been allocated kernel space before installing.
185	 */
186
187	uvm_map_setup(&kernel_map_store, base, end, VM_MAP_PAGEABLE);
188	kernel_map_store.pmap = pmap_kernel();
189	if (start != base) {
190		int error;
191		struct uvm_map_args args;
192
193		error = uvm_map_prepare(&kernel_map_store,
194		    base, start - base,
195		    NULL, UVM_UNKNOWN_OFFSET, 0,
196		    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
197		    		UVM_ADV_RANDOM, UVM_FLAG_FIXED), &args);
198		if (!error) {
199			kernel_image_mapent_store.flags =
200			    UVM_MAP_KERNEL | UVM_MAP_STATIC | UVM_MAP_NOMERGE;
201			error = uvm_map_enter(&kernel_map_store, &args,
202			    &kernel_image_mapent_store);
203		}
204
205		if (error)
206			panic(
207			    "uvm_km_bootstrap: could not reserve space for kernel");
208
209		kmembase = args.uma_start + args.uma_size;
210		error = uvm_map_prepare(&kernel_map_store,
211		    kmembase, kmemsize,
212		    NULL, UVM_UNKNOWN_OFFSET, 0,
213		    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
214		    		UVM_ADV_RANDOM, UVM_FLAG_FIXED), &args);
215		if (!error) {
216			kernel_kmem_mapent_store.flags =
217			    UVM_MAP_KERNEL | UVM_MAP_STATIC | UVM_MAP_NOMERGE;
218			error = uvm_map_enter(&kernel_map_store, &args,
219			    &kernel_kmem_mapent_store);
220		}
221
222		if (error)
223			panic(
224			    "uvm_km_bootstrap: could not reserve kernel kmem");
225	} else {
226		kmembase = base;
227	}
228
229	/*
230	 * install!
231	 */
232
233	kernel_map = &kernel_map_store;
234
235	pool_subsystem_init();
236	vmem_bootstrap();
237
238	kmem_arena = vmem_create("kmem", kmembase, kmemsize, PAGE_SIZE,
239	    NULL, NULL, NULL,
240	    0, VM_NOSLEEP | VM_BOOTSTRAP, IPL_VM);
241
242	vmem_init(kmem_arena);
243
244	kmem_va_arena = vmem_create("kva", 0, 0, PAGE_SIZE,
245	    vmem_alloc, vmem_free, kmem_arena,
246	    16 * PAGE_SIZE, VM_NOSLEEP | VM_BOOTSTRAP, IPL_VM);
247}
248
249/*
250 * uvm_km_init: init the kernel maps virtual memory caches
251 * and start the pool/kmem allocator.
252 */
253void
254uvm_km_init(void)
255{
256
257	kmem_init();
258
259	kmeminit(); // killme
260}
261
262/*
263 * uvm_km_suballoc: allocate a submap in the kernel map.   once a submap
264 * is allocated all references to that area of VM must go through it.  this
265 * allows the locking of VAs in kernel_map to be broken up into regions.
266 *
267 * => if `fixed' is true, *vmin specifies where the region described
268 *   pager_map => used to map "buf" structures into kernel space
269 *      by the submap must start
270 * => if submap is non NULL we use that as the submap, otherwise we
271 *	alloc a new map
272 */
273
274struct vm_map *
275uvm_km_suballoc(struct vm_map *map, vaddr_t *vmin /* IN/OUT */,
276    vaddr_t *vmax /* OUT */, vsize_t size, int flags, bool fixed,
277    struct vm_map *submap)
278{
279	int mapflags = UVM_FLAG_NOMERGE | (fixed ? UVM_FLAG_FIXED : 0);
280
281	KASSERT(vm_map_pmap(map) == pmap_kernel());
282
283	size = round_page(size);	/* round up to pagesize */
284
285	/*
286	 * first allocate a blank spot in the parent map
287	 */
288
289	if (uvm_map(map, vmin, size, NULL, UVM_UNKNOWN_OFFSET, 0,
290	    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
291	    UVM_ADV_RANDOM, mapflags)) != 0) {
292	       panic("uvm_km_suballoc: unable to allocate space in parent map");
293	}
294
295	/*
296	 * set VM bounds (vmin is filled in by uvm_map)
297	 */
298
299	*vmax = *vmin + size;
300
301	/*
302	 * add references to pmap and create or init the submap
303	 */
304
305	pmap_reference(vm_map_pmap(map));
306	if (submap == NULL) {
307		submap = kmem_alloc(sizeof(*submap), KM_SLEEP);
308		if (submap == NULL)
309			panic("uvm_km_suballoc: unable to create submap");
310	}
311	uvm_map_setup(submap, *vmin, *vmax, flags);
312	submap->pmap = vm_map_pmap(map);
313
314	/*
315	 * now let uvm_map_submap plug in it...
316	 */
317
318	if (uvm_map_submap(map, *vmin, *vmax, submap) != 0)
319		panic("uvm_km_suballoc: submap allocation failed");
320
321	return(submap);
322}
323
324/*
325 * uvm_km_pgremove: remove pages from a kernel uvm_object and KVA.
326 */
327
328void
329uvm_km_pgremove(vaddr_t startva, vaddr_t endva)
330{
331	struct uvm_object * const uobj = uvm_kernel_object;
332	const voff_t start = startva - vm_map_min(kernel_map);
333	const voff_t end = endva - vm_map_min(kernel_map);
334	struct vm_page *pg;
335	voff_t curoff, nextoff;
336	int swpgonlydelta = 0;
337	UVMHIST_FUNC("uvm_km_pgremove"); UVMHIST_CALLED(maphist);
338
339	KASSERT(VM_MIN_KERNEL_ADDRESS <= startva);
340	KASSERT(startva < endva);
341	KASSERT(endva <= VM_MAX_KERNEL_ADDRESS);
342
343	mutex_enter(uobj->vmobjlock);
344	pmap_remove(pmap_kernel(), startva, endva);
345	for (curoff = start; curoff < end; curoff = nextoff) {
346		nextoff = curoff + PAGE_SIZE;
347		pg = uvm_pagelookup(uobj, curoff);
348		if (pg != NULL && pg->flags & PG_BUSY) {
349			pg->flags |= PG_WANTED;
350			UVM_UNLOCK_AND_WAIT(pg, uobj->vmobjlock, 0,
351				    "km_pgrm", 0);
352			mutex_enter(uobj->vmobjlock);
353			nextoff = curoff;
354			continue;
355		}
356
357		/*
358		 * free the swap slot, then the page.
359		 */
360
361		if (pg == NULL &&
362		    uao_find_swslot(uobj, curoff >> PAGE_SHIFT) > 0) {
363			swpgonlydelta++;
364		}
365		uao_dropswap(uobj, curoff >> PAGE_SHIFT);
366		if (pg != NULL) {
367			mutex_enter(&uvm_pageqlock);
368			uvm_pagefree(pg);
369			mutex_exit(&uvm_pageqlock);
370		}
371	}
372	mutex_exit(uobj->vmobjlock);
373
374	if (swpgonlydelta > 0) {
375		mutex_enter(&uvm_swap_data_lock);
376		KASSERT(uvmexp.swpgonly >= swpgonlydelta);
377		uvmexp.swpgonly -= swpgonlydelta;
378		mutex_exit(&uvm_swap_data_lock);
379	}
380}
381
382
383/*
384 * uvm_km_pgremove_intrsafe: like uvm_km_pgremove(), but for non object backed
385 *    regions.
386 *
387 * => when you unmap a part of anonymous kernel memory you want to toss
388 *    the pages right away.    (this is called from uvm_unmap_...).
389 * => none of the pages will ever be busy, and none of them will ever
390 *    be on the active or inactive queues (because they have no object).
391 */
392
393void
394uvm_km_pgremove_intrsafe(struct vm_map *map, vaddr_t start, vaddr_t end)
395{
396	struct vm_page *pg;
397	paddr_t pa;
398	UVMHIST_FUNC("uvm_km_pgremove_intrsafe"); UVMHIST_CALLED(maphist);
399
400	KASSERT(VM_MAP_IS_KERNEL(map));
401	KASSERT(vm_map_min(map) <= start);
402	KASSERT(start < end);
403	KASSERT(end <= vm_map_max(map));
404
405	for (; start < end; start += PAGE_SIZE) {
406		if (!pmap_extract(pmap_kernel(), start, &pa)) {
407			continue;
408		}
409		pg = PHYS_TO_VM_PAGE(pa);
410		KASSERT(pg);
411		KASSERT(pg->uobject == NULL && pg->uanon == NULL);
412		KASSERT((pg->flags & PG_BUSY) == 0);
413		uvm_pagefree(pg);
414	}
415}
416
417#if defined(DEBUG)
418void
419uvm_km_check_empty(struct vm_map *map, vaddr_t start, vaddr_t end)
420{
421	struct vm_page *pg;
422	vaddr_t va;
423	paddr_t pa;
424
425	KDASSERT(VM_MAP_IS_KERNEL(map));
426	KDASSERT(vm_map_min(map) <= start);
427	KDASSERT(start < end);
428	KDASSERT(end <= vm_map_max(map));
429
430	for (va = start; va < end; va += PAGE_SIZE) {
431		if (pmap_extract(pmap_kernel(), va, &pa)) {
432			panic("uvm_km_check_empty: va %p has pa 0x%llx",
433			    (void *)va, (long long)pa);
434		}
435		if ((map->flags & VM_MAP_INTRSAFE) == 0) {
436			mutex_enter(uvm_kernel_object->vmobjlock);
437			pg = uvm_pagelookup(uvm_kernel_object,
438			    va - vm_map_min(kernel_map));
439			mutex_exit(uvm_kernel_object->vmobjlock);
440			if (pg) {
441				panic("uvm_km_check_empty: "
442				    "has page hashed at %p", (const void *)va);
443			}
444		}
445	}
446}
447#endif /* defined(DEBUG) */
448
449/*
450 * uvm_km_alloc: allocate an area of kernel memory.
451 *
452 * => NOTE: we can return 0 even if we can wait if there is not enough
453 *	free VM space in the map... caller should be prepared to handle
454 *	this case.
455 * => we return KVA of memory allocated
456 */
457
458vaddr_t
459uvm_km_alloc(struct vm_map *map, vsize_t size, vsize_t align, uvm_flag_t flags)
460{
461	vaddr_t kva, loopva;
462	vaddr_t offset;
463	vsize_t loopsize;
464	struct vm_page *pg;
465	struct uvm_object *obj;
466	int pgaflags;
467	vm_prot_t prot;
468	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
469
470	KASSERT(vm_map_pmap(map) == pmap_kernel());
471	KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED ||
472		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE ||
473		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY);
474	KASSERT((flags & UVM_KMF_VAONLY) != 0 || (flags & UVM_KMF_COLORMATCH) == 0);
475	KASSERT((flags & UVM_KMF_COLORMATCH) == 0 || (flags & UVM_KMF_VAONLY) != 0);
476
477	/*
478	 * setup for call
479	 */
480
481	kva = vm_map_min(map);	/* hint */
482	size = round_page(size);
483	obj = (flags & UVM_KMF_PAGEABLE) ? uvm_kernel_object : NULL;
484	UVMHIST_LOG(maphist,"  (map=0x%x, obj=0x%x, size=0x%x, flags=%d)",
485		    map, obj, size, flags);
486
487	/*
488	 * allocate some virtual space
489	 */
490
491	if (__predict_false(uvm_map(map, &kva, size, obj, UVM_UNKNOWN_OFFSET,
492	    align, UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
493	    UVM_ADV_RANDOM,
494	    (flags & (UVM_KMF_TRYLOCK | UVM_KMF_NOWAIT | UVM_KMF_WAITVA
495	     | UVM_KMF_COLORMATCH)))) != 0)) {
496		UVMHIST_LOG(maphist, "<- done (no VM)",0,0,0,0);
497		return(0);
498	}
499
500	/*
501	 * if all we wanted was VA, return now
502	 */
503
504	if (flags & (UVM_KMF_VAONLY | UVM_KMF_PAGEABLE)) {
505		UVMHIST_LOG(maphist,"<- done valloc (kva=0x%x)", kva,0,0,0);
506		return(kva);
507	}
508
509	/*
510	 * recover object offset from virtual address
511	 */
512
513	offset = kva - vm_map_min(kernel_map);
514	UVMHIST_LOG(maphist, "  kva=0x%x, offset=0x%x", kva, offset,0,0);
515
516	/*
517	 * now allocate and map in the memory... note that we are the only ones
518	 * whom should ever get a handle on this area of VM.
519	 */
520
521	loopva = kva;
522	loopsize = size;
523
524	pgaflags = UVM_FLAG_COLORMATCH;
525	if (flags & UVM_KMF_NOWAIT)
526		pgaflags |= UVM_PGA_USERESERVE;
527	if (flags & UVM_KMF_ZERO)
528		pgaflags |= UVM_PGA_ZERO;
529	prot = VM_PROT_READ | VM_PROT_WRITE;
530	if (flags & UVM_KMF_EXEC)
531		prot |= VM_PROT_EXECUTE;
532	while (loopsize) {
533		KASSERTMSG(!pmap_extract(pmap_kernel(), loopva, NULL),
534		    "loopva=%#"PRIxVADDR, loopva);
535
536		pg = uvm_pagealloc_strat(NULL, offset, NULL, pgaflags,
537#ifdef UVM_KM_VMFREELIST
538		   UVM_PGA_STRAT_ONLY, UVM_KM_VMFREELIST
539#else
540		   UVM_PGA_STRAT_NORMAL, 0
541#endif
542		   );
543
544		/*
545		 * out of memory?
546		 */
547
548		if (__predict_false(pg == NULL)) {
549			if ((flags & UVM_KMF_NOWAIT) ||
550			    ((flags & UVM_KMF_CANFAIL) && !uvm_reclaimable())) {
551				/* free everything! */
552				uvm_km_free(map, kva, size,
553				    flags & UVM_KMF_TYPEMASK);
554				return (0);
555			} else {
556				uvm_wait("km_getwait2");	/* sleep here */
557				continue;
558			}
559		}
560
561		pg->flags &= ~PG_BUSY;	/* new page */
562		UVM_PAGE_OWN(pg, NULL);
563
564		/*
565		 * map it in
566		 */
567
568		pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg),
569		    prot, PMAP_KMPAGE);
570		loopva += PAGE_SIZE;
571		offset += PAGE_SIZE;
572		loopsize -= PAGE_SIZE;
573	}
574
575	pmap_update(pmap_kernel());
576
577	UVMHIST_LOG(maphist,"<- done (kva=0x%x)", kva,0,0,0);
578	return(kva);
579}
580
581/*
582 * uvm_km_free: free an area of kernel memory
583 */
584
585void
586uvm_km_free(struct vm_map *map, vaddr_t addr, vsize_t size, uvm_flag_t flags)
587{
588
589	KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED ||
590		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE ||
591		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY);
592	KASSERT((addr & PAGE_MASK) == 0);
593	KASSERT(vm_map_pmap(map) == pmap_kernel());
594
595	size = round_page(size);
596
597	if (flags & UVM_KMF_PAGEABLE) {
598		uvm_km_pgremove(addr, addr + size);
599	} else if (flags & UVM_KMF_WIRED) {
600		/*
601		 * Note: uvm_km_pgremove_intrsafe() extracts mapping, thus
602		 * remove it after.  See comment below about KVA visibility.
603		 */
604		uvm_km_pgremove_intrsafe(map, addr, addr + size);
605		pmap_kremove(addr, size);
606	}
607
608	/*
609	 * Note: uvm_unmap_remove() calls pmap_update() for us, before
610	 * KVA becomes globally available.
611	 */
612
613	uvm_unmap1(map, addr, addr + size, UVM_FLAG_VAONLY);
614}
615
616/* Sanity; must specify both or none. */
617#if (defined(PMAP_MAP_POOLPAGE) || defined(PMAP_UNMAP_POOLPAGE)) && \
618    (!defined(PMAP_MAP_POOLPAGE) || !defined(PMAP_UNMAP_POOLPAGE))
619#error Must specify MAP and UNMAP together.
620#endif
621
622int
623uvm_km_kmem_alloc(vmem_t *vm, vmem_size_t size, vm_flag_t flags,
624    vmem_addr_t *addr)
625{
626	struct vm_page *pg;
627	vmem_addr_t va;
628	int rc;
629	vaddr_t loopva;
630	vsize_t loopsize;
631
632	size = round_page(size);
633
634#if defined(PMAP_MAP_POOLPAGE)
635	if (size == PAGE_SIZE) {
636again:
637#ifdef PMAP_ALLOC_POOLPAGE
638		pg = PMAP_ALLOC_POOLPAGE((flags & VM_SLEEP) ?
639		   0 : UVM_PGA_USERESERVE);
640#else
641		pg = uvm_pagealloc(NULL, 0, NULL,
642		   (flags & VM_SLEEP) ? 0 : UVM_PGA_USERESERVE);
643#endif /* PMAP_ALLOC_POOLPAGE */
644		if (__predict_false(pg == NULL)) {
645			if (flags & VM_SLEEP) {
646				uvm_wait("plpg");
647				goto again;
648			}
649		}
650		va = PMAP_MAP_POOLPAGE(VM_PAGE_TO_PHYS(pg));
651		if (__predict_false(va == 0)) {
652			uvm_pagefree(pg);
653			return ENOMEM;
654		}
655		*addr = va;
656		return 0;
657	}
658#endif /* PMAP_MAP_POOLPAGE */
659
660	rc = vmem_alloc(vm, size, flags, &va);
661	if (rc != 0)
662		return rc;
663
664	loopva = va;
665	loopsize = size;
666
667	while (loopsize) {
668		KASSERTMSG(!pmap_extract(pmap_kernel(), loopva, NULL),
669		    "loopva=%#"PRIxVADDR" loopsize=%#"PRIxVSIZE" vmem=%p",
670		    loopva, loopsize, vm);
671
672		pg = uvm_pagealloc(NULL, loopva, NULL,
673		    UVM_KMF_COLORMATCH
674		    | ((flags & VM_SLEEP) ? 0 : UVM_PGA_USERESERVE));
675		if (__predict_false(pg == NULL)) {
676			if (flags & VM_SLEEP) {
677				uvm_wait("plpg");
678				continue;
679			} else {
680				uvm_km_pgremove_intrsafe(kernel_map, va,
681				    va + size);
682				pmap_kremove(va, size);
683				vmem_free(kmem_va_arena, va, size);
684				return ENOMEM;
685			}
686		}
687
688		pg->flags &= ~PG_BUSY;	/* new page */
689		UVM_PAGE_OWN(pg, NULL);
690		pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg),
691		    VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE);
692
693		loopva += PAGE_SIZE;
694		loopsize -= PAGE_SIZE;
695	}
696	pmap_update(pmap_kernel());
697
698	*addr = va;
699
700	return 0;
701}
702
703void
704uvm_km_kmem_free(vmem_t *vm, vmem_addr_t addr, size_t size)
705{
706
707	size = round_page(size);
708#if defined(PMAP_UNMAP_POOLPAGE)
709	if (size == PAGE_SIZE) {
710		paddr_t pa;
711
712		pa = PMAP_UNMAP_POOLPAGE(addr);
713		uvm_pagefree(PHYS_TO_VM_PAGE(pa));
714		return;
715	}
716#endif /* PMAP_UNMAP_POOLPAGE */
717	uvm_km_pgremove_intrsafe(kernel_map, addr, addr + size);
718	pmap_kremove(addr, size);
719	pmap_update(pmap_kernel());
720
721	vmem_free(vm, addr, size);
722}
723
724bool
725uvm_km_va_starved_p(void)
726{
727	vmem_size_t total;
728	vmem_size_t free;
729
730	total = vmem_size(kmem_arena, VMEM_ALLOC|VMEM_FREE);
731	free = vmem_size(kmem_arena, VMEM_FREE);
732
733	return (free < (total / 10));
734}
735
736