xref: /netbsd-src/sys/uvm/uvm_km.c (revision 3b01aba77a7a698587faaae455bbfe740923c1f5)
1 /*	$NetBSD: uvm_km.c,v 1.50 2001/06/26 17:55:15 thorpej 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. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by Charles D. Cranor,
23  *      Washington University, the University of California, Berkeley and
24  *      its contributors.
25  * 4. Neither the name of the University nor the names of its contributors
26  *    may be used to endorse or promote products derived from this software
27  *    without specific prior written permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39  * SUCH DAMAGE.
40  *
41  *	@(#)vm_kern.c   8.3 (Berkeley) 1/12/94
42  * from: Id: uvm_km.c,v 1.1.2.14 1998/02/06 05:19:27 chs Exp
43  *
44  *
45  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46  * All rights reserved.
47  *
48  * Permission to use, copy, modify and distribute this software and
49  * its documentation is hereby granted, provided that both the copyright
50  * notice and this permission notice appear in all copies of the
51  * software, derivative works or modified versions, and any portions
52  * thereof, and that both notices appear in supporting documentation.
53  *
54  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
55  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
56  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
57  *
58  * Carnegie Mellon requests users of this software to return to
59  *
60  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
61  *  School of Computer Science
62  *  Carnegie Mellon University
63  *  Pittsburgh PA 15213-3890
64  *
65  * any improvements or extensions that they make and grant Carnegie the
66  * rights to redistribute these changes.
67  */
68 
69 #include "opt_uvmhist.h"
70 
71 /*
72  * uvm_km.c: handle kernel memory allocation and management
73  */
74 
75 /*
76  * overview of kernel memory management:
77  *
78  * the kernel virtual address space is mapped by "kernel_map."   kernel_map
79  * starts at VM_MIN_KERNEL_ADDRESS and goes to VM_MAX_KERNEL_ADDRESS.
80  * note that VM_MIN_KERNEL_ADDRESS is equal to vm_map_min(kernel_map).
81  *
82  * the kernel_map has several "submaps."   submaps can only appear in
83  * the kernel_map (user processes can't use them).   submaps "take over"
84  * the management of a sub-range of the kernel's address space.  submaps
85  * are typically allocated at boot time and are never released.   kernel
86  * virtual address space that is mapped by a submap is locked by the
87  * submap's lock -- not the kernel_map's lock.
88  *
89  * thus, the useful feature of submaps is that they allow us to break
90  * up the locking and protection of the kernel address space into smaller
91  * chunks.
92  *
93  * the vm system has several standard kernel submaps, including:
94  *   kmem_map => contains only wired kernel memory for the kernel
95  *		malloc.   *** access to kmem_map must be protected
96  *		by splvm() because we are allowed to call malloc()
97  *		at interrupt time ***
98  *   mb_map => memory for large mbufs,  *** protected by splvm ***
99  *   pager_map => used to map "buf" structures into kernel space
100  *   exec_map => used during exec to handle exec args
101  *   etc...
102  *
103  * the kernel allocates its private memory out of special uvm_objects whose
104  * reference count is set to UVM_OBJ_KERN (thus indicating that the objects
105  * are "special" and never die).   all kernel objects should be thought of
106  * as large, fixed-sized, sparsely populated uvm_objects.   each kernel
107  * object is equal to the size of kernel virtual address space (i.e. the
108  * value "VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS").
109  *
110  * most kernel private memory lives in kernel_object.   the only exception
111  * to this is for memory that belongs to submaps that must be protected
112  * by splvm().    each of these submaps has their own private kernel
113  * object (e.g. kmem_object, mb_object).
114  *
115  * note that just because a kernel object spans the entire kernel virutal
116  * address space doesn't mean that it has to be mapped into the entire space.
117  * large chunks of a kernel object's space go unused either because
118  * that area of kernel VM is unmapped, or there is some other type of
119  * object mapped into that range (e.g. a vnode).    for submap's kernel
120  * objects, the only part of the object that can ever be populated is the
121  * offsets that are managed by the submap.
122  *
123  * note that the "offset" in a kernel object is always the kernel virtual
124  * address minus the VM_MIN_KERNEL_ADDRESS (aka vm_map_min(kernel_map)).
125  * example:
126  *   suppose VM_MIN_KERNEL_ADDRESS is 0xf8000000 and the kernel does a
127  *   uvm_km_alloc(kernel_map, PAGE_SIZE) [allocate 1 wired down page in the
128  *   kernel map].    if uvm_km_alloc returns virtual address 0xf8235000,
129  *   then that means that the page at offset 0x235000 in kernel_object is
130  *   mapped at 0xf8235000.
131  *
132  * note that the offsets in kmem_object and mb_object also follow this
133  * rule.   this means that the offsets for kmem_object must fall in the
134  * range of [vm_map_min(kmem_object) - vm_map_min(kernel_map)] to
135  * [vm_map_max(kmem_object) - vm_map_min(kernel_map)], so the offsets
136  * in those objects will typically not start at zero.
137  *
138  * kernel object have one other special property: when the kernel virtual
139  * memory mapping them is unmapped, the backing memory in the object is
140  * freed right away.   this is done with the uvm_km_pgremove() function.
141  * this has to be done because there is no backing store for kernel pages
142  * and no need to save them after they are no longer referenced.
143  */
144 
145 #include <sys/param.h>
146 #include <sys/systm.h>
147 #include <sys/proc.h>
148 
149 #include <uvm/uvm.h>
150 
151 /*
152  * global data structures
153  */
154 
155 struct vm_map *kernel_map = NULL;
156 
157 /*
158  * local data structues
159  */
160 
161 static struct vm_map		kernel_map_store;
162 static struct uvm_object	kmem_object_store;
163 static struct uvm_object	mb_object_store;
164 
165 /*
166  * All pager operations here are NULL, but the object must have
167  * a pager ops vector associated with it; various places assume
168  * it to be so.
169  */
170 static struct uvm_pagerops	km_pager;
171 
172 /*
173  * uvm_km_init: init kernel maps and objects to reflect reality (i.e.
174  * KVM already allocated for text, data, bss, and static data structures).
175  *
176  * => KVM is defined by VM_MIN_KERNEL_ADDRESS/VM_MAX_KERNEL_ADDRESS.
177  *    we assume that [min -> start] has already been allocated and that
178  *    "end" is the end.
179  */
180 
181 void
182 uvm_km_init(start, end)
183 	vaddr_t start, end;
184 {
185 	vaddr_t base = VM_MIN_KERNEL_ADDRESS;
186 
187 	/*
188 	 * next, init kernel memory objects.
189 	 */
190 
191 	/* kernel_object: for pageable anonymous kernel memory */
192 	uao_init();
193 	uvm.kernel_object = uao_create(VM_MAX_KERNEL_ADDRESS -
194 				 VM_MIN_KERNEL_ADDRESS, UAO_FLAG_KERNOBJ);
195 
196 	/*
197 	 * kmem_object: for use by the kernel malloc().  Memory is always
198 	 * wired, and this object (and the kmem_map) can be accessed at
199 	 * interrupt time.
200 	 */
201 	simple_lock_init(&kmem_object_store.vmobjlock);
202 	kmem_object_store.pgops = &km_pager;
203 	TAILQ_INIT(&kmem_object_store.memq);
204 	kmem_object_store.uo_npages = 0;
205 	/* we are special.  we never die */
206 	kmem_object_store.uo_refs = UVM_OBJ_KERN_INTRSAFE;
207 	uvmexp.kmem_object = &kmem_object_store;
208 
209 	/*
210 	 * mb_object: for mbuf cluster pages on platforms which use the
211 	 * mb_map.  Memory is always wired, and this object (and the mb_map)
212 	 * can be accessed at interrupt time.
213 	 */
214 	simple_lock_init(&mb_object_store.vmobjlock);
215 	mb_object_store.pgops = &km_pager;
216 	TAILQ_INIT(&mb_object_store.memq);
217 	mb_object_store.uo_npages = 0;
218 	/* we are special.  we never die */
219 	mb_object_store.uo_refs = UVM_OBJ_KERN_INTRSAFE;
220 	uvmexp.mb_object = &mb_object_store;
221 
222 	/*
223 	 * init the map and reserve allready allocated kernel space
224 	 * before installing.
225 	 */
226 
227 	uvm_map_setup(&kernel_map_store, base, end, VM_MAP_PAGEABLE);
228 	kernel_map_store.pmap = pmap_kernel();
229 	if (uvm_map(&kernel_map_store, &base, start - base, NULL,
230 	    UVM_UNKNOWN_OFFSET, 0, UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL,
231 	    UVM_INH_NONE, UVM_ADV_RANDOM,UVM_FLAG_FIXED)) != 0)
232 		panic("uvm_km_init: could not reserve space for kernel");
233 
234 	/*
235 	 * install!
236 	 */
237 
238 	kernel_map = &kernel_map_store;
239 }
240 
241 /*
242  * uvm_km_suballoc: allocate a submap in the kernel map.   once a submap
243  * is allocated all references to that area of VM must go through it.  this
244  * allows the locking of VAs in kernel_map to be broken up into regions.
245  *
246  * => if `fixed' is true, *min specifies where the region described
247  *      by the submap must start
248  * => if submap is non NULL we use that as the submap, otherwise we
249  *	alloc a new map
250  */
251 struct vm_map *
252 uvm_km_suballoc(map, min, max, size, flags, fixed, submap)
253 	struct vm_map *map;
254 	vaddr_t *min, *max;		/* OUT, OUT */
255 	vsize_t size;
256 	int flags;
257 	boolean_t fixed;
258 	struct vm_map *submap;
259 {
260 	int mapflags = UVM_FLAG_NOMERGE | (fixed ? UVM_FLAG_FIXED : 0);
261 
262 	size = round_page(size);	/* round up to pagesize */
263 
264 	/*
265 	 * first allocate a blank spot in the parent map
266 	 */
267 
268 	if (uvm_map(map, min, size, NULL, UVM_UNKNOWN_OFFSET, 0,
269 	    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
270 	    UVM_ADV_RANDOM, mapflags)) != 0) {
271 	       panic("uvm_km_suballoc: unable to allocate space in parent map");
272 	}
273 
274 	/*
275 	 * set VM bounds (min is filled in by uvm_map)
276 	 */
277 
278 	*max = *min + size;
279 
280 	/*
281 	 * add references to pmap and create or init the submap
282 	 */
283 
284 	pmap_reference(vm_map_pmap(map));
285 	if (submap == NULL) {
286 		submap = uvm_map_create(vm_map_pmap(map), *min, *max, flags);
287 		if (submap == NULL)
288 			panic("uvm_km_suballoc: unable to create submap");
289 	} else {
290 		uvm_map_setup(submap, *min, *max, flags);
291 		submap->pmap = vm_map_pmap(map);
292 	}
293 
294 	/*
295 	 * now let uvm_map_submap plug in it...
296 	 */
297 
298 	if (uvm_map_submap(map, *min, *max, submap) != 0)
299 		panic("uvm_km_suballoc: submap allocation failed");
300 
301 	return(submap);
302 }
303 
304 /*
305  * uvm_km_pgremove: remove pages from a kernel uvm_object.
306  *
307  * => when you unmap a part of anonymous kernel memory you want to toss
308  *    the pages right away.    (this gets called from uvm_unmap_...).
309  */
310 
311 #define UKM_HASH_PENALTY 4      /* a guess */
312 
313 void
314 uvm_km_pgremove(uobj, start, end)
315 	struct uvm_object *uobj;
316 	vaddr_t start, end;
317 {
318 	boolean_t by_list;
319 	struct vm_page *pp, *ppnext;
320 	vaddr_t curoff;
321 	UVMHIST_FUNC("uvm_km_pgremove"); UVMHIST_CALLED(maphist);
322 
323 	KASSERT(uobj->pgops == &aobj_pager);
324 	simple_lock(&uobj->vmobjlock);
325 
326 	/* choose cheapest traversal */
327 	by_list = (uobj->uo_npages <=
328 	     ((end - start) >> PAGE_SHIFT) * UKM_HASH_PENALTY);
329 
330 	if (by_list)
331 		goto loop_by_list;
332 
333 	/* by hash */
334 
335 	for (curoff = start ; curoff < end ; curoff += PAGE_SIZE) {
336 		pp = uvm_pagelookup(uobj, curoff);
337 		if (pp == NULL)
338 			continue;
339 
340 		UVMHIST_LOG(maphist,"  page 0x%x, busy=%d", pp,
341 		    pp->flags & PG_BUSY, 0, 0);
342 
343 		/* now do the actual work */
344 		if (pp->flags & PG_BUSY) {
345 			/* owner must check for this when done */
346 			pp->flags |= PG_RELEASED;
347 		} else {
348 			/* free the swap slot... */
349 			uao_dropswap(uobj, curoff >> PAGE_SHIFT);
350 
351 			/*
352 			 * ...and free the page; note it may be on the
353 			 * active or inactive queues.
354 			 */
355 			uvm_lock_pageq();
356 			uvm_pagefree(pp);
357 			uvm_unlock_pageq();
358 		}
359 	}
360 	simple_unlock(&uobj->vmobjlock);
361 	return;
362 
363 loop_by_list:
364 
365 	for (pp = TAILQ_FIRST(&uobj->memq); pp != NULL; pp = ppnext) {
366 		ppnext = TAILQ_NEXT(pp, listq);
367 		if (pp->offset < start || pp->offset >= end) {
368 			continue;
369 		}
370 
371 		UVMHIST_LOG(maphist,"  page 0x%x, busy=%d", pp,
372 		    pp->flags & PG_BUSY, 0, 0);
373 
374 		if (pp->flags & PG_BUSY) {
375 			/* owner must check for this when done */
376 			pp->flags |= PG_RELEASED;
377 		} else {
378 			/* free the swap slot... */
379 			uao_dropswap(uobj, pp->offset >> PAGE_SHIFT);
380 
381 			/*
382 			 * ...and free the page; note it may be on the
383 			 * active or inactive queues.
384 			 */
385 			uvm_lock_pageq();
386 			uvm_pagefree(pp);
387 			uvm_unlock_pageq();
388 		}
389 	}
390 	simple_unlock(&uobj->vmobjlock);
391 }
392 
393 
394 /*
395  * uvm_km_pgremove_intrsafe: like uvm_km_pgremove(), but for "intrsafe"
396  *    objects
397  *
398  * => when you unmap a part of anonymous kernel memory you want to toss
399  *    the pages right away.    (this gets called from uvm_unmap_...).
400  * => none of the pages will ever be busy, and none of them will ever
401  *    be on the active or inactive queues (because these objects are
402  *    never allowed to "page").
403  */
404 
405 void
406 uvm_km_pgremove_intrsafe(uobj, start, end)
407 	struct uvm_object *uobj;
408 	vaddr_t start, end;
409 {
410 	boolean_t by_list;
411 	struct vm_page *pp, *ppnext;
412 	vaddr_t curoff;
413 	UVMHIST_FUNC("uvm_km_pgremove_intrsafe"); UVMHIST_CALLED(maphist);
414 
415 	KASSERT(UVM_OBJ_IS_INTRSAFE_OBJECT(uobj));
416 	simple_lock(&uobj->vmobjlock);		/* lock object */
417 
418 	/* choose cheapest traversal */
419 	by_list = (uobj->uo_npages <=
420 	     ((end - start) >> PAGE_SHIFT) * UKM_HASH_PENALTY);
421 
422 	if (by_list)
423 		goto loop_by_list;
424 
425 	/* by hash */
426 
427 	for (curoff = start ; curoff < end ; curoff += PAGE_SIZE) {
428 		pp = uvm_pagelookup(uobj, curoff);
429 		if (pp == NULL) {
430 			continue;
431 		}
432 
433 		UVMHIST_LOG(maphist,"  page 0x%x, busy=%d", pp,
434 		    pp->flags & PG_BUSY, 0, 0);
435 		KASSERT((pp->flags & PG_BUSY) == 0);
436 		KASSERT((pp->pqflags & PQ_ACTIVE) == 0);
437 		KASSERT((pp->pqflags & PQ_INACTIVE) == 0);
438 		uvm_pagefree(pp);
439 	}
440 	simple_unlock(&uobj->vmobjlock);
441 	return;
442 
443 loop_by_list:
444 
445 	for (pp = TAILQ_FIRST(&uobj->memq); pp != NULL; pp = ppnext) {
446 		ppnext = TAILQ_NEXT(pp, listq);
447 		if (pp->offset < start || pp->offset >= end) {
448 			continue;
449 		}
450 
451 		UVMHIST_LOG(maphist,"  page 0x%x, busy=%d", pp,
452 		    pp->flags & PG_BUSY, 0, 0);
453 		KASSERT((pp->flags & PG_BUSY) == 0);
454 		KASSERT((pp->pqflags & PQ_ACTIVE) == 0);
455 		KASSERT((pp->pqflags & PQ_INACTIVE) == 0);
456 		uvm_pagefree(pp);
457 	}
458 	simple_unlock(&uobj->vmobjlock);
459 }
460 
461 
462 /*
463  * uvm_km_kmemalloc: lower level kernel memory allocator for malloc()
464  *
465  * => we map wired memory into the specified map using the obj passed in
466  * => NOTE: we can return NULL even if we can wait if there is not enough
467  *	free VM space in the map... caller should be prepared to handle
468  *	this case.
469  * => we return KVA of memory allocated
470  * => flags: NOWAIT, VALLOC - just allocate VA, TRYLOCK - fail if we can't
471  *	lock the map
472  */
473 
474 vaddr_t
475 uvm_km_kmemalloc(map, obj, size, flags)
476 	struct vm_map *map;
477 	struct uvm_object *obj;
478 	vsize_t size;
479 	int flags;
480 {
481 	vaddr_t kva, loopva;
482 	vaddr_t offset;
483 	vsize_t loopsize;
484 	struct vm_page *pg;
485 	UVMHIST_FUNC("uvm_km_kmemalloc"); UVMHIST_CALLED(maphist);
486 
487 	UVMHIST_LOG(maphist,"  (map=0x%x, obj=0x%x, size=0x%x, flags=%d)",
488 		    map, obj, size, flags);
489 	KASSERT(vm_map_pmap(map) == pmap_kernel());
490 
491 	/*
492 	 * setup for call
493 	 */
494 
495 	size = round_page(size);
496 	kva = vm_map_min(map);	/* hint */
497 
498 	/*
499 	 * allocate some virtual space
500 	 */
501 
502 	if (__predict_false(uvm_map(map, &kva, size, obj, UVM_UNKNOWN_OFFSET,
503 	      0, UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
504 			  UVM_ADV_RANDOM, (flags & UVM_KMF_TRYLOCK)))
505 			!= 0)) {
506 		UVMHIST_LOG(maphist, "<- done (no VM)",0,0,0,0);
507 		return(0);
508 	}
509 
510 	/*
511 	 * if all we wanted was VA, return now
512 	 */
513 
514 	if (flags & UVM_KMF_VALLOC) {
515 		UVMHIST_LOG(maphist,"<- done valloc (kva=0x%x)", kva,0,0,0);
516 		return(kva);
517 	}
518 
519 	/*
520 	 * recover object offset from virtual address
521 	 */
522 
523 	offset = kva - vm_map_min(kernel_map);
524 	UVMHIST_LOG(maphist, "  kva=0x%x, offset=0x%x", kva, offset,0,0);
525 
526 	/*
527 	 * now allocate and map in the memory... note that we are the only ones
528 	 * whom should ever get a handle on this area of VM.
529 	 */
530 
531 	loopva = kva;
532 	loopsize = size;
533 	while (loopsize) {
534 		simple_lock(&obj->vmobjlock);
535 		pg = uvm_pagealloc(obj, offset, NULL, 0);
536 		if (__predict_true(pg != NULL)) {
537 			pg->flags &= ~PG_BUSY;	/* new page */
538 			UVM_PAGE_OWN(pg, NULL);
539 		}
540 		simple_unlock(&obj->vmobjlock);
541 
542 		/*
543 		 * out of memory?
544 		 */
545 
546 		if (__predict_false(pg == NULL)) {
547 			if (flags & UVM_KMF_NOWAIT) {
548 				/* free everything! */
549 				uvm_unmap(map, kva, kva + size);
550 				return(0);
551 			} else {
552 				uvm_wait("km_getwait2");	/* sleep here */
553 				continue;
554 			}
555 		}
556 
557 		/*
558 		 * map it in: note that we call pmap_enter with the map and
559 		 * object unlocked in case we are kmem_map/kmem_object
560 		 * (because if pmap_enter wants to allocate out of kmem_object
561 		 * it will need to lock it itself!)
562 		 */
563 
564 		if (UVM_OBJ_IS_INTRSAFE_OBJECT(obj)) {
565 			pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg),
566 			    VM_PROT_ALL);
567 		} else {
568 			pmap_enter(map->pmap, loopva, VM_PAGE_TO_PHYS(pg),
569 			    UVM_PROT_ALL,
570 			    PMAP_WIRED | VM_PROT_READ | VM_PROT_WRITE);
571 		}
572 		loopva += PAGE_SIZE;
573 		offset += PAGE_SIZE;
574 		loopsize -= PAGE_SIZE;
575 	}
576 	pmap_update();
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 
585 void
586 uvm_km_free(map, addr, size)
587 	struct vm_map *map;
588 	vaddr_t addr;
589 	vsize_t size;
590 {
591 	uvm_unmap(map, trunc_page(addr), round_page(addr+size));
592 }
593 
594 /*
595  * uvm_km_free_wakeup: free an area of kernel memory and wake up
596  * anyone waiting for vm space.
597  *
598  * => XXX: "wanted" bit + unlock&wait on other end?
599  */
600 
601 void
602 uvm_km_free_wakeup(map, addr, size)
603 	struct vm_map *map;
604 	vaddr_t addr;
605 	vsize_t size;
606 {
607 	struct vm_map_entry *dead_entries;
608 
609 	vm_map_lock(map);
610 	uvm_unmap_remove(map, trunc_page(addr), round_page(addr + size),
611 	    &dead_entries);
612 	wakeup(map);
613 	vm_map_unlock(map);
614 	if (dead_entries != NULL)
615 		uvm_unmap_detach(dead_entries, 0);
616 }
617 
618 /*
619  * uvm_km_alloc1: allocate wired down memory in the kernel map.
620  *
621  * => we can sleep if needed
622  */
623 
624 vaddr_t
625 uvm_km_alloc1(map, size, zeroit)
626 	struct vm_map *map;
627 	vsize_t size;
628 	boolean_t zeroit;
629 {
630 	vaddr_t kva, loopva, offset;
631 	struct vm_page *pg;
632 	UVMHIST_FUNC("uvm_km_alloc1"); UVMHIST_CALLED(maphist);
633 
634 	UVMHIST_LOG(maphist,"(map=0x%x, size=0x%x)", map, size,0,0);
635 	KASSERT(vm_map_pmap(map) == pmap_kernel());
636 
637 	size = round_page(size);
638 	kva = vm_map_min(map);		/* hint */
639 
640 	/*
641 	 * allocate some virtual space
642 	 */
643 
644 	if (__predict_false(uvm_map(map, &kva, size, uvm.kernel_object,
645 	      UVM_UNKNOWN_OFFSET, 0, UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL,
646 					      UVM_INH_NONE, UVM_ADV_RANDOM,
647 					      0)) != 0)) {
648 		UVMHIST_LOG(maphist,"<- done (no VM)",0,0,0,0);
649 		return(0);
650 	}
651 
652 	/*
653 	 * recover object offset from virtual address
654 	 */
655 
656 	offset = kva - vm_map_min(kernel_map);
657 	UVMHIST_LOG(maphist,"  kva=0x%x, offset=0x%x", kva, offset,0,0);
658 
659 	/*
660 	 * now allocate the memory.  we must be careful about released pages.
661 	 */
662 
663 	loopva = kva;
664 	while (size) {
665 		simple_lock(&uvm.kernel_object->vmobjlock);
666 		pg = uvm_pagelookup(uvm.kernel_object, offset);
667 
668 		/*
669 		 * if we found a page in an unallocated region, it must be
670 		 * released
671 		 */
672 		if (pg) {
673 			if ((pg->flags & PG_RELEASED) == 0)
674 				panic("uvm_km_alloc1: non-released page");
675 			pg->flags |= PG_WANTED;
676 			UVM_UNLOCK_AND_WAIT(pg, &uvm.kernel_object->vmobjlock,
677 			    FALSE, "km_alloc", 0);
678 			continue;   /* retry */
679 		}
680 
681 		/* allocate ram */
682 		pg = uvm_pagealloc(uvm.kernel_object, offset, NULL, 0);
683 		if (pg) {
684 			pg->flags &= ~PG_BUSY;	/* new page */
685 			UVM_PAGE_OWN(pg, NULL);
686 		}
687 		simple_unlock(&uvm.kernel_object->vmobjlock);
688 		if (__predict_false(pg == NULL)) {
689 			uvm_wait("km_alloc1w");	/* wait for memory */
690 			continue;
691 		}
692 
693 		/*
694 		 * map it in; note we're never called with an intrsafe
695 		 * object, so we always use regular old pmap_enter().
696 		 */
697 		pmap_enter(map->pmap, loopva, VM_PAGE_TO_PHYS(pg),
698 		    UVM_PROT_ALL, PMAP_WIRED | VM_PROT_READ | VM_PROT_WRITE);
699 
700 		loopva += PAGE_SIZE;
701 		offset += PAGE_SIZE;
702 		size -= PAGE_SIZE;
703 	}
704 
705 	pmap_update();
706 
707 	/*
708 	 * zero on request (note that "size" is now zero due to the above loop
709 	 * so we need to subtract kva from loopva to reconstruct the size).
710 	 */
711 
712 	if (zeroit)
713 		memset((caddr_t)kva, 0, loopva - kva);
714 
715 	UVMHIST_LOG(maphist,"<- done (kva=0x%x)", kva,0,0,0);
716 	return(kva);
717 }
718 
719 /*
720  * uvm_km_valloc: allocate zero-fill memory in the kernel's address space
721  *
722  * => memory is not allocated until fault time
723  */
724 
725 vaddr_t
726 uvm_km_valloc(map, size)
727 	struct vm_map *map;
728 	vsize_t size;
729 {
730 	return(uvm_km_valloc_align(map, size, 0));
731 }
732 
733 vaddr_t
734 uvm_km_valloc_align(map, size, align)
735 	struct vm_map *map;
736 	vsize_t size;
737 	vsize_t align;
738 {
739 	vaddr_t kva;
740 	UVMHIST_FUNC("uvm_km_valloc"); UVMHIST_CALLED(maphist);
741 
742 	UVMHIST_LOG(maphist, "(map=0x%x, size=0x%x)", map, size, 0,0);
743 	KASSERT(vm_map_pmap(map) == pmap_kernel());
744 
745 	size = round_page(size);
746 	kva = vm_map_min(map);		/* hint */
747 
748 	/*
749 	 * allocate some virtual space.  will be demand filled by kernel_object.
750 	 */
751 
752 	if (__predict_false(uvm_map(map, &kva, size, uvm.kernel_object,
753 	    UVM_UNKNOWN_OFFSET, align, UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL,
754 					    UVM_INH_NONE, UVM_ADV_RANDOM,
755 					    0)) != 0)) {
756 		UVMHIST_LOG(maphist, "<- done (no VM)", 0,0,0,0);
757 		return(0);
758 	}
759 
760 	UVMHIST_LOG(maphist, "<- done (kva=0x%x)", kva,0,0,0);
761 	return(kva);
762 }
763 
764 /*
765  * uvm_km_valloc_wait: allocate zero-fill memory in the kernel's address space
766  *
767  * => memory is not allocated until fault time
768  * => if no room in map, wait for space to free, unless requested size
769  *    is larger than map (in which case we return 0)
770  */
771 
772 vaddr_t
773 uvm_km_valloc_prefer_wait(map, size, prefer)
774 	struct vm_map *map;
775 	vsize_t size;
776 	voff_t prefer;
777 {
778 	vaddr_t kva;
779 	UVMHIST_FUNC("uvm_km_valloc_prefer_wait"); UVMHIST_CALLED(maphist);
780 
781 	UVMHIST_LOG(maphist, "(map=0x%x, size=0x%x)", map, size, 0,0);
782 	KASSERT(vm_map_pmap(map) == pmap_kernel());
783 
784 	size = round_page(size);
785 	if (size > vm_map_max(map) - vm_map_min(map))
786 		return(0);
787 
788 	while (1) {
789 		kva = vm_map_min(map);		/* hint */
790 
791 		/*
792 		 * allocate some virtual space.   will be demand filled
793 		 * by kernel_object.
794 		 */
795 
796 		if (__predict_true(uvm_map(map, &kva, size, uvm.kernel_object,
797 		    prefer, 0, UVM_MAPFLAG(UVM_PROT_ALL,
798 		    UVM_PROT_ALL, UVM_INH_NONE, UVM_ADV_RANDOM, 0))
799 		    == 0)) {
800 			UVMHIST_LOG(maphist,"<- done (kva=0x%x)", kva,0,0,0);
801 			return(kva);
802 		}
803 
804 		/*
805 		 * failed.  sleep for a while (on map)
806 		 */
807 
808 		UVMHIST_LOG(maphist,"<<<sleeping>>>",0,0,0,0);
809 		tsleep((caddr_t)map, PVM, "vallocwait", 0);
810 	}
811 	/*NOTREACHED*/
812 }
813 
814 vaddr_t
815 uvm_km_valloc_wait(map, size)
816 	struct vm_map *map;
817 	vsize_t size;
818 {
819 	return uvm_km_valloc_prefer_wait(map, size, UVM_UNKNOWN_OFFSET);
820 }
821 
822 /* Sanity; must specify both or none. */
823 #if (defined(PMAP_MAP_POOLPAGE) || defined(PMAP_UNMAP_POOLPAGE)) && \
824     (!defined(PMAP_MAP_POOLPAGE) || !defined(PMAP_UNMAP_POOLPAGE))
825 #error Must specify MAP and UNMAP together.
826 #endif
827 
828 /*
829  * uvm_km_alloc_poolpage: allocate a page for the pool allocator
830  *
831  * => if the pmap specifies an alternate mapping method, we use it.
832  */
833 
834 /* ARGSUSED */
835 vaddr_t
836 uvm_km_alloc_poolpage1(map, obj, waitok)
837 	struct vm_map *map;
838 	struct uvm_object *obj;
839 	boolean_t waitok;
840 {
841 #if defined(PMAP_MAP_POOLPAGE)
842 	struct vm_page *pg;
843 	vaddr_t va;
844 
845  again:
846 	pg = uvm_pagealloc(NULL, 0, NULL, UVM_PGA_USERESERVE);
847 	if (__predict_false(pg == NULL)) {
848 		if (waitok) {
849 			uvm_wait("plpg");
850 			goto again;
851 		} else
852 			return (0);
853 	}
854 	va = PMAP_MAP_POOLPAGE(VM_PAGE_TO_PHYS(pg));
855 	if (__predict_false(va == 0))
856 		uvm_pagefree(pg);
857 	return (va);
858 #else
859 	vaddr_t va;
860 	int s;
861 
862 	/*
863 	 * NOTE: We may be called with a map that doens't require splvm
864 	 * protection (e.g. kernel_map).  However, it does not hurt to
865 	 * go to splvm in this case (since unprocted maps will never be
866 	 * accessed in interrupt context).
867 	 *
868 	 * XXX We may want to consider changing the interface to this
869 	 * XXX function.
870 	 */
871 
872 	s = splvm();
873 	va = uvm_km_kmemalloc(map, obj, PAGE_SIZE, waitok ? 0 : UVM_KMF_NOWAIT);
874 	splx(s);
875 	return (va);
876 #endif /* PMAP_MAP_POOLPAGE */
877 }
878 
879 /*
880  * uvm_km_free_poolpage: free a previously allocated pool page
881  *
882  * => if the pmap specifies an alternate unmapping method, we use it.
883  */
884 
885 /* ARGSUSED */
886 void
887 uvm_km_free_poolpage1(map, addr)
888 	struct vm_map *map;
889 	vaddr_t addr;
890 {
891 #if defined(PMAP_UNMAP_POOLPAGE)
892 	paddr_t pa;
893 
894 	pa = PMAP_UNMAP_POOLPAGE(addr);
895 	uvm_pagefree(PHYS_TO_VM_PAGE(pa));
896 #else
897 	int s;
898 
899 	/*
900 	 * NOTE: We may be called with a map that doens't require splvm
901 	 * protection (e.g. kernel_map).  However, it does not hurt to
902 	 * go to splvm in this case (since unprocted maps will never be
903 	 * accessed in interrupt context).
904 	 *
905 	 * XXX We may want to consider changing the interface to this
906 	 * XXX function.
907 	 */
908 
909 	s = splvm();
910 	uvm_km_free(map, addr, PAGE_SIZE);
911 	splx(s);
912 #endif /* PMAP_UNMAP_POOLPAGE */
913 }
914