uvm_km.c revision 1.120
1/*	$NetBSD: uvm_km.c,v 1.120 2012/02/10 17:35:47 para 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 1.120 2012/02/10 17:35:47 para Exp $");
124
125#include "opt_uvmhist.h"
126
127#include "opt_kmempages.h"
128
129#ifndef NKMEMPAGES
130#define NKMEMPAGES 0
131#endif
132
133/*
134 * Defaults for lower and upper-bounds for the kmem_arena page count.
135 * Can be overridden by kernel config options.
136 */
137#ifndef NKMEMPAGES_MIN
138#define NKMEMPAGES_MIN NKMEMPAGES_MIN_DEFAULT
139#endif
140
141#ifndef NKMEMPAGES_MAX
142#define NKMEMPAGES_MAX NKMEMPAGES_MAX_DEFAULT
143#endif
144
145
146#include <sys/param.h>
147#include <sys/systm.h>
148#include <sys/proc.h>
149#include <sys/pool.h>
150#include <sys/vmem.h>
151#include <sys/kmem.h>
152
153#include <uvm/uvm.h>
154
155/*
156 * global data structures
157 */
158
159struct vm_map *kernel_map = NULL;
160
161/*
162 * local data structues
163 */
164
165static struct vm_map		kernel_map_store;
166static struct vm_map_entry	kernel_image_mapent_store;
167static struct vm_map_entry	kernel_kmem_mapent_store;
168
169int nkmempages = 0;
170vaddr_t kmembase;
171vsize_t kmemsize;
172
173vmem_t *kmem_arena;
174vmem_t *kmem_va_arena;
175
176/*
177 * kmeminit_nkmempages: calculate the size of kmem_arena.
178 */
179void
180kmeminit_nkmempages(void)
181{
182	int npages;
183
184	if (nkmempages != 0) {
185		/*
186		 * It's already been set (by us being here before)
187		 * bail out now;
188		 */
189		return;
190	}
191
192#if defined(PMAP_MAP_POOLPAGE)
193	npages = (physmem / 4);
194#else
195	npages = (physmem / 3) * 2;
196#endif /* defined(PMAP_MAP_POOLPAGE) */
197
198#ifndef NKMEMPAGES_MAX_UNLIMITED
199	if (npages > NKMEMPAGES_MAX)
200		npages = NKMEMPAGES_MAX;
201#endif
202
203	if (npages < NKMEMPAGES_MIN)
204		npages = NKMEMPAGES_MIN;
205
206	nkmempages = npages;
207}
208
209/*
210 * uvm_km_bootstrap: init kernel maps and objects to reflect reality (i.e.
211 * KVM already allocated for text, data, bss, and static data structures).
212 *
213 * => KVM is defined by VM_MIN_KERNEL_ADDRESS/VM_MAX_KERNEL_ADDRESS.
214 *    we assume that [vmin -> start] has already been allocated and that
215 *    "end" is the end.
216 */
217
218void
219uvm_km_bootstrap(vaddr_t start, vaddr_t end)
220{
221	bool kmem_arena_small;
222	vaddr_t base = VM_MIN_KERNEL_ADDRESS;
223	struct uvm_map_args args;
224	int error;
225
226	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
227	UVMHIST_LOG(maphist, "start=%"PRIxVADDR" end=%#"PRIxVADDR,
228	    start, end, 0,0);
229
230	kmeminit_nkmempages();
231	kmemsize = (vsize_t)nkmempages * PAGE_SIZE;
232	kmem_arena_small = kmemsize < 64 * 1024 * 1024;
233
234	UVMHIST_LOG(maphist, "kmemsize=%#"PRIxVSIZE, kmemsize, 0,0,0);
235
236	/*
237	 * next, init kernel memory objects.
238	 */
239
240	/* kernel_object: for pageable anonymous kernel memory */
241	uvm_kernel_object = uao_create(VM_MAX_KERNEL_ADDRESS -
242				VM_MIN_KERNEL_ADDRESS, UAO_FLAG_KERNOBJ);
243
244	/*
245	 * init the map and reserve any space that might already
246	 * have been allocated kernel space before installing.
247	 */
248
249	uvm_map_setup(&kernel_map_store, base, end, VM_MAP_PAGEABLE);
250	kernel_map_store.pmap = pmap_kernel();
251	if (start != base) {
252		error = uvm_map_prepare(&kernel_map_store,
253		    base, start - base,
254		    NULL, UVM_UNKNOWN_OFFSET, 0,
255		    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
256		    		UVM_ADV_RANDOM, UVM_FLAG_FIXED), &args);
257		if (!error) {
258			kernel_image_mapent_store.flags =
259			    UVM_MAP_KERNEL | UVM_MAP_STATIC | UVM_MAP_NOMERGE;
260			error = uvm_map_enter(&kernel_map_store, &args,
261			    &kernel_image_mapent_store);
262		}
263
264		if (error)
265			panic(
266			    "uvm_km_bootstrap: could not reserve space for kernel");
267
268		kmembase = args.uma_start + args.uma_size;
269	} else {
270		kmembase = base;
271	}
272
273	error = uvm_map_prepare(&kernel_map_store,
274	    kmembase, kmemsize,
275	    NULL, UVM_UNKNOWN_OFFSET, 0,
276	    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
277	    		UVM_ADV_RANDOM, UVM_FLAG_FIXED), &args);
278	if (!error) {
279		kernel_kmem_mapent_store.flags =
280		    UVM_MAP_KERNEL | UVM_MAP_STATIC | UVM_MAP_NOMERGE;
281		error = uvm_map_enter(&kernel_map_store, &args,
282		    &kernel_kmem_mapent_store);
283	}
284
285	if (error)
286		panic("uvm_km_bootstrap: could not reserve kernel kmem");
287
288	/*
289	 * install!
290	 */
291
292	kernel_map = &kernel_map_store;
293
294	pool_subsystem_init();
295	vmem_bootstrap();
296
297	kmem_arena = vmem_create("kmem", kmembase, kmemsize, PAGE_SIZE,
298	    NULL, NULL, NULL,
299	    0, VM_NOSLEEP | VM_BOOTSTRAP, IPL_VM);
300
301	vmem_init(kmem_arena);
302
303	UVMHIST_LOG(maphist, "kmem vmem created (base=%#"PRIxVADDR
304	    ", size=%#"PRIxVSIZE, kmembase, kmemsize, 0,0);
305
306	kmem_va_arena = vmem_create("kva", 0, 0, PAGE_SIZE,
307	    vmem_alloc, vmem_free, kmem_arena,
308	    (kmem_arena_small ? 4 : 8) * PAGE_SIZE,
309	    VM_NOSLEEP | VM_BOOTSTRAP, IPL_VM);
310
311	UVMHIST_LOG(maphist, "<- done", 0,0,0,0);
312}
313
314/*
315 * uvm_km_init: init the kernel maps virtual memory caches
316 * and start the pool/kmem allocator.
317 */
318void
319uvm_km_init(void)
320{
321
322	kmem_init();
323
324	kmeminit(); // killme
325}
326
327/*
328 * uvm_km_suballoc: allocate a submap in the kernel map.   once a submap
329 * is allocated all references to that area of VM must go through it.  this
330 * allows the locking of VAs in kernel_map to be broken up into regions.
331 *
332 * => if `fixed' is true, *vmin specifies where the region described
333 *   pager_map => used to map "buf" structures into kernel space
334 *      by the submap must start
335 * => if submap is non NULL we use that as the submap, otherwise we
336 *	alloc a new map
337 */
338
339struct vm_map *
340uvm_km_suballoc(struct vm_map *map, vaddr_t *vmin /* IN/OUT */,
341    vaddr_t *vmax /* OUT */, vsize_t size, int flags, bool fixed,
342    struct vm_map *submap)
343{
344	int mapflags = UVM_FLAG_NOMERGE | (fixed ? UVM_FLAG_FIXED : 0);
345	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
346
347	KASSERT(vm_map_pmap(map) == pmap_kernel());
348
349	size = round_page(size);	/* round up to pagesize */
350
351	/*
352	 * first allocate a blank spot in the parent map
353	 */
354
355	if (uvm_map(map, vmin, size, NULL, UVM_UNKNOWN_OFFSET, 0,
356	    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
357	    UVM_ADV_RANDOM, mapflags)) != 0) {
358		panic("%s: unable to allocate space in parent map", __func__);
359	}
360
361	/*
362	 * set VM bounds (vmin is filled in by uvm_map)
363	 */
364
365	*vmax = *vmin + size;
366
367	/*
368	 * add references to pmap and create or init the submap
369	 */
370
371	pmap_reference(vm_map_pmap(map));
372	if (submap == NULL) {
373		submap = kmem_alloc(sizeof(*submap), KM_SLEEP);
374		if (submap == NULL)
375			panic("uvm_km_suballoc: unable to create submap");
376	}
377	uvm_map_setup(submap, *vmin, *vmax, flags);
378	submap->pmap = vm_map_pmap(map);
379
380	/*
381	 * now let uvm_map_submap plug in it...
382	 */
383
384	if (uvm_map_submap(map, *vmin, *vmax, submap) != 0)
385		panic("uvm_km_suballoc: submap allocation failed");
386
387	return(submap);
388}
389
390/*
391 * uvm_km_pgremove: remove pages from a kernel uvm_object and KVA.
392 */
393
394void
395uvm_km_pgremove(vaddr_t startva, vaddr_t endva)
396{
397	struct uvm_object * const uobj = uvm_kernel_object;
398	const voff_t start = startva - vm_map_min(kernel_map);
399	const voff_t end = endva - vm_map_min(kernel_map);
400	struct vm_page *pg;
401	voff_t curoff, nextoff;
402	int swpgonlydelta = 0;
403	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
404
405	KASSERT(VM_MIN_KERNEL_ADDRESS <= startva);
406	KASSERT(startva < endva);
407	KASSERT(endva <= VM_MAX_KERNEL_ADDRESS);
408
409	mutex_enter(uobj->vmobjlock);
410	pmap_remove(pmap_kernel(), startva, endva);
411	for (curoff = start; curoff < end; curoff = nextoff) {
412		nextoff = curoff + PAGE_SIZE;
413		pg = uvm_pagelookup(uobj, curoff);
414		if (pg != NULL && pg->flags & PG_BUSY) {
415			pg->flags |= PG_WANTED;
416			UVM_UNLOCK_AND_WAIT(pg, uobj->vmobjlock, 0,
417				    "km_pgrm", 0);
418			mutex_enter(uobj->vmobjlock);
419			nextoff = curoff;
420			continue;
421		}
422
423		/*
424		 * free the swap slot, then the page.
425		 */
426
427		if (pg == NULL &&
428		    uao_find_swslot(uobj, curoff >> PAGE_SHIFT) > 0) {
429			swpgonlydelta++;
430		}
431		uao_dropswap(uobj, curoff >> PAGE_SHIFT);
432		if (pg != NULL) {
433			mutex_enter(&uvm_pageqlock);
434			uvm_pagefree(pg);
435			mutex_exit(&uvm_pageqlock);
436		}
437	}
438	mutex_exit(uobj->vmobjlock);
439
440	if (swpgonlydelta > 0) {
441		mutex_enter(&uvm_swap_data_lock);
442		KASSERT(uvmexp.swpgonly >= swpgonlydelta);
443		uvmexp.swpgonly -= swpgonlydelta;
444		mutex_exit(&uvm_swap_data_lock);
445	}
446}
447
448
449/*
450 * uvm_km_pgremove_intrsafe: like uvm_km_pgremove(), but for non object backed
451 *    regions.
452 *
453 * => when you unmap a part of anonymous kernel memory you want to toss
454 *    the pages right away.    (this is called from uvm_unmap_...).
455 * => none of the pages will ever be busy, and none of them will ever
456 *    be on the active or inactive queues (because they have no object).
457 */
458
459void
460uvm_km_pgremove_intrsafe(struct vm_map *map, vaddr_t start, vaddr_t end)
461{
462	struct vm_page *pg;
463	paddr_t pa;
464	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
465
466	KASSERT(VM_MAP_IS_KERNEL(map));
467	KASSERT(vm_map_min(map) <= start);
468	KASSERT(start < end);
469	KASSERT(end <= vm_map_max(map));
470
471	for (; start < end; start += PAGE_SIZE) {
472		if (!pmap_extract(pmap_kernel(), start, &pa)) {
473			continue;
474		}
475		pg = PHYS_TO_VM_PAGE(pa);
476		KASSERT(pg);
477		KASSERT(pg->uobject == NULL && pg->uanon == NULL);
478		KASSERT((pg->flags & PG_BUSY) == 0);
479		uvm_pagefree(pg);
480	}
481}
482
483#if defined(DEBUG)
484void
485uvm_km_check_empty(struct vm_map *map, vaddr_t start, vaddr_t end)
486{
487	struct vm_page *pg;
488	vaddr_t va;
489	paddr_t pa;
490	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
491
492	KDASSERT(VM_MAP_IS_KERNEL(map));
493	KDASSERT(vm_map_min(map) <= start);
494	KDASSERT(start < end);
495	KDASSERT(end <= vm_map_max(map));
496
497	for (va = start; va < end; va += PAGE_SIZE) {
498		if (pmap_extract(pmap_kernel(), va, &pa)) {
499			panic("uvm_km_check_empty: va %p has pa 0x%llx",
500			    (void *)va, (long long)pa);
501		}
502		if ((map->flags & VM_MAP_INTRSAFE) == 0) {
503			mutex_enter(uvm_kernel_object->vmobjlock);
504			pg = uvm_pagelookup(uvm_kernel_object,
505			    va - vm_map_min(kernel_map));
506			mutex_exit(uvm_kernel_object->vmobjlock);
507			if (pg) {
508				panic("uvm_km_check_empty: "
509				    "has page hashed at %p", (const void *)va);
510			}
511		}
512	}
513}
514#endif /* defined(DEBUG) */
515
516/*
517 * uvm_km_alloc: allocate an area of kernel memory.
518 *
519 * => NOTE: we can return 0 even if we can wait if there is not enough
520 *	free VM space in the map... caller should be prepared to handle
521 *	this case.
522 * => we return KVA of memory allocated
523 */
524
525vaddr_t
526uvm_km_alloc(struct vm_map *map, vsize_t size, vsize_t align, uvm_flag_t flags)
527{
528	vaddr_t kva, loopva;
529	vaddr_t offset;
530	vsize_t loopsize;
531	struct vm_page *pg;
532	struct uvm_object *obj;
533	int pgaflags;
534	vm_prot_t prot;
535	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
536
537	KASSERT(vm_map_pmap(map) == pmap_kernel());
538	KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED ||
539		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE ||
540		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY);
541	KASSERT((flags & UVM_KMF_VAONLY) != 0 || (flags & UVM_KMF_COLORMATCH) == 0);
542	KASSERT((flags & UVM_KMF_COLORMATCH) == 0 || (flags & UVM_KMF_VAONLY) != 0);
543
544	/*
545	 * setup for call
546	 */
547
548	kva = vm_map_min(map);	/* hint */
549	size = round_page(size);
550	obj = (flags & UVM_KMF_PAGEABLE) ? uvm_kernel_object : NULL;
551	UVMHIST_LOG(maphist,"  (map=0x%x, obj=0x%x, size=0x%x, flags=%d)",
552		    map, obj, size, flags);
553
554	/*
555	 * allocate some virtual space
556	 */
557
558	if (__predict_false(uvm_map(map, &kva, size, obj, UVM_UNKNOWN_OFFSET,
559	    align, UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
560	    UVM_ADV_RANDOM,
561	    (flags & (UVM_KMF_TRYLOCK | UVM_KMF_NOWAIT | UVM_KMF_WAITVA
562	     | UVM_KMF_COLORMATCH)))) != 0)) {
563		UVMHIST_LOG(maphist, "<- done (no VM)",0,0,0,0);
564		return(0);
565	}
566
567	/*
568	 * if all we wanted was VA, return now
569	 */
570
571	if (flags & (UVM_KMF_VAONLY | UVM_KMF_PAGEABLE)) {
572		UVMHIST_LOG(maphist,"<- done valloc (kva=0x%x)", kva,0,0,0);
573		return(kva);
574	}
575
576	/*
577	 * recover object offset from virtual address
578	 */
579
580	offset = kva - vm_map_min(kernel_map);
581	UVMHIST_LOG(maphist, "  kva=0x%x, offset=0x%x", kva, offset,0,0);
582
583	/*
584	 * now allocate and map in the memory... note that we are the only ones
585	 * whom should ever get a handle on this area of VM.
586	 */
587
588	loopva = kva;
589	loopsize = size;
590
591	pgaflags = UVM_FLAG_COLORMATCH;
592	if (flags & UVM_KMF_NOWAIT)
593		pgaflags |= UVM_PGA_USERESERVE;
594	if (flags & UVM_KMF_ZERO)
595		pgaflags |= UVM_PGA_ZERO;
596	prot = VM_PROT_READ | VM_PROT_WRITE;
597	if (flags & UVM_KMF_EXEC)
598		prot |= VM_PROT_EXECUTE;
599	while (loopsize) {
600		KASSERTMSG(!pmap_extract(pmap_kernel(), loopva, NULL),
601		    "loopva=%#"PRIxVADDR, loopva);
602
603		pg = uvm_pagealloc_strat(NULL, offset, NULL, pgaflags,
604#ifdef UVM_KM_VMFREELIST
605		   UVM_PGA_STRAT_ONLY, UVM_KM_VMFREELIST
606#else
607		   UVM_PGA_STRAT_NORMAL, 0
608#endif
609		   );
610
611		/*
612		 * out of memory?
613		 */
614
615		if (__predict_false(pg == NULL)) {
616			if ((flags & UVM_KMF_NOWAIT) ||
617			    ((flags & UVM_KMF_CANFAIL) && !uvm_reclaimable())) {
618				/* free everything! */
619				uvm_km_free(map, kva, size,
620				    flags & UVM_KMF_TYPEMASK);
621				return (0);
622			} else {
623				uvm_wait("km_getwait2");	/* sleep here */
624				continue;
625			}
626		}
627
628		pg->flags &= ~PG_BUSY;	/* new page */
629		UVM_PAGE_OWN(pg, NULL);
630
631		/*
632		 * map it in
633		 */
634
635		pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg),
636		    prot, PMAP_KMPAGE);
637		loopva += PAGE_SIZE;
638		offset += PAGE_SIZE;
639		loopsize -= PAGE_SIZE;
640	}
641
642	pmap_update(pmap_kernel());
643
644	UVMHIST_LOG(maphist,"<- done (kva=0x%x)", kva,0,0,0);
645	return(kva);
646}
647
648/*
649 * uvm_km_free: free an area of kernel memory
650 */
651
652void
653uvm_km_free(struct vm_map *map, vaddr_t addr, vsize_t size, uvm_flag_t flags)
654{
655	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
656
657	KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED ||
658		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE ||
659		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY);
660	KASSERT((addr & PAGE_MASK) == 0);
661	KASSERT(vm_map_pmap(map) == pmap_kernel());
662
663	size = round_page(size);
664
665	if (flags & UVM_KMF_PAGEABLE) {
666		uvm_km_pgremove(addr, addr + size);
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_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
690int
691uvm_km_kmem_alloc(vmem_t *vm, vmem_size_t size, vm_flag_t flags,
692    vmem_addr_t *addr)
693{
694	struct vm_page *pg;
695	vmem_addr_t va;
696	int rc;
697	vaddr_t loopva;
698	vsize_t loopsize;
699
700	size = round_page(size);
701
702#if defined(PMAP_MAP_POOLPAGE)
703	if (size == PAGE_SIZE) {
704again:
705#ifdef PMAP_ALLOC_POOLPAGE
706		pg = PMAP_ALLOC_POOLPAGE((flags & VM_SLEEP) ?
707		   0 : UVM_PGA_USERESERVE);
708#else
709		pg = uvm_pagealloc(NULL, 0, NULL,
710		   (flags & VM_SLEEP) ? 0 : UVM_PGA_USERESERVE);
711#endif /* PMAP_ALLOC_POOLPAGE */
712		if (__predict_false(pg == NULL)) {
713			if (flags & VM_SLEEP) {
714				uvm_wait("plpg");
715				goto again;
716			}
717		}
718		va = PMAP_MAP_POOLPAGE(VM_PAGE_TO_PHYS(pg));
719		if (__predict_false(va == 0)) {
720			uvm_pagefree(pg);
721			return ENOMEM;
722		}
723		*addr = va;
724		return 0;
725	}
726#endif /* PMAP_MAP_POOLPAGE */
727
728	rc = vmem_alloc(vm, size, flags, &va);
729	if (rc != 0)
730		return rc;
731
732	loopva = va;
733	loopsize = size;
734
735	while (loopsize) {
736		KASSERTMSG(!pmap_extract(pmap_kernel(), loopva, NULL),
737		    "loopva=%#"PRIxVADDR" loopsize=%#"PRIxVSIZE" vmem=%p",
738		    loopva, loopsize, vm);
739
740		pg = uvm_pagealloc(NULL, loopva, NULL,
741		    UVM_FLAG_COLORMATCH
742		    | ((flags & VM_SLEEP) ? 0 : UVM_PGA_USERESERVE));
743		if (__predict_false(pg == NULL)) {
744			if (flags & VM_SLEEP) {
745				uvm_wait("plpg");
746				continue;
747			} else {
748				uvm_km_pgremove_intrsafe(kernel_map, va,
749				    va + size);
750				pmap_kremove(va, size);
751				vmem_free(kmem_va_arena, va, size);
752				return ENOMEM;
753			}
754		}
755
756		pg->flags &= ~PG_BUSY;	/* new page */
757		UVM_PAGE_OWN(pg, NULL);
758		pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg),
759		    VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE);
760
761		loopva += PAGE_SIZE;
762		loopsize -= PAGE_SIZE;
763	}
764	pmap_update(pmap_kernel());
765
766	*addr = va;
767
768	return 0;
769}
770
771void
772uvm_km_kmem_free(vmem_t *vm, vmem_addr_t addr, size_t size)
773{
774
775	size = round_page(size);
776#if defined(PMAP_UNMAP_POOLPAGE)
777	if (size == PAGE_SIZE) {
778		paddr_t pa;
779
780		pa = PMAP_UNMAP_POOLPAGE(addr);
781		uvm_pagefree(PHYS_TO_VM_PAGE(pa));
782		return;
783	}
784#endif /* PMAP_UNMAP_POOLPAGE */
785	uvm_km_pgremove_intrsafe(kernel_map, addr, addr + size);
786	pmap_kremove(addr, size);
787	pmap_update(pmap_kernel());
788
789	vmem_free(vm, addr, size);
790}
791
792bool
793uvm_km_va_starved_p(void)
794{
795	vmem_size_t total;
796	vmem_size_t free;
797
798	total = vmem_size(kmem_arena, VMEM_ALLOC|VMEM_FREE);
799	free = vmem_size(kmem_arena, VMEM_FREE);
800
801	return (free < (total / 10));
802}
803
804