Home | History | Annotate | Line # | Download | only in uvm
uvm_km.c revision 1.87.2.1
      1 /*	$NetBSD: uvm_km.c,v 1.87.2.1 2006/06/19 04:11:44 chap 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 (at) 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 /*
     70  * uvm_km.c: handle kernel memory allocation and management
     71  */
     72 
     73 /*
     74  * overview of kernel memory management:
     75  *
     76  * the kernel virtual address space is mapped by "kernel_map."   kernel_map
     77  * starts at VM_MIN_KERNEL_ADDRESS and goes to VM_MAX_KERNEL_ADDRESS.
     78  * note that VM_MIN_KERNEL_ADDRESS is equal to vm_map_min(kernel_map).
     79  *
     80  * the kernel_map has several "submaps."   submaps can only appear in
     81  * the kernel_map (user processes can't use them).   submaps "take over"
     82  * the management of a sub-range of the kernel's address space.  submaps
     83  * are typically allocated at boot time and are never released.   kernel
     84  * virtual address space that is mapped by a submap is locked by the
     85  * submap's lock -- not the kernel_map's lock.
     86  *
     87  * thus, the useful feature of submaps is that they allow us to break
     88  * up the locking and protection of the kernel address space into smaller
     89  * chunks.
     90  *
     91  * the vm system has several standard kernel submaps, including:
     92  *   kmem_map => contains only wired kernel memory for the kernel
     93  *		malloc.   *** access to kmem_map must be protected
     94  *		by splvm() because we are allowed to call malloc()
     95  *		at interrupt time ***
     96  *   mb_map => memory for large mbufs,  *** protected by splvm ***
     97  *   pager_map => used to map "buf" structures into kernel space
     98  *   exec_map => used during exec to handle exec args
     99  *   etc...
    100  *
    101  * the kernel allocates its private memory out of special uvm_objects whose
    102  * reference count is set to UVM_OBJ_KERN (thus indicating that the objects
    103  * are "special" and never die).   all kernel objects should be thought of
    104  * as large, fixed-sized, sparsely populated uvm_objects.   each kernel
    105  * object is equal to the size of kernel virtual address space (i.e. the
    106  * value "VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS").
    107  *
    108  * note that just because a kernel object spans the entire kernel virutal
    109  * address space doesn't mean that it has to be mapped into the entire space.
    110  * large chunks of a kernel object's space go unused either because
    111  * that area of kernel VM is unmapped, or there is some other type of
    112  * object mapped into that range (e.g. a vnode).    for submap's kernel
    113  * objects, the only part of the object that can ever be populated is the
    114  * offsets that are managed by the submap.
    115  *
    116  * note that the "offset" in a kernel object is always the kernel virtual
    117  * address minus the VM_MIN_KERNEL_ADDRESS (aka vm_map_min(kernel_map)).
    118  * example:
    119  *   suppose VM_MIN_KERNEL_ADDRESS is 0xf8000000 and the kernel does a
    120  *   uvm_km_alloc(kernel_map, PAGE_SIZE) [allocate 1 wired down page in the
    121  *   kernel map].    if uvm_km_alloc returns virtual address 0xf8235000,
    122  *   then that means that the page at offset 0x235000 in kernel_object is
    123  *   mapped at 0xf8235000.
    124  *
    125  * kernel object have one other special property: when the kernel virtual
    126  * memory mapping them is unmapped, the backing memory in the object is
    127  * freed right away.   this is done with the uvm_km_pgremove() function.
    128  * this has to be done because there is no backing store for kernel pages
    129  * and no need to save them after they are no longer referenced.
    130  */
    131 
    132 #include <sys/cdefs.h>
    133 __KERNEL_RCSID(0, "$NetBSD: uvm_km.c,v 1.87.2.1 2006/06/19 04:11:44 chap Exp $");
    134 
    135 #include "opt_uvmhist.h"
    136 
    137 #include <sys/param.h>
    138 #include <sys/malloc.h>
    139 #include <sys/systm.h>
    140 #include <sys/proc.h>
    141 #include <sys/pool.h>
    142 
    143 #include <uvm/uvm.h>
    144 
    145 /*
    146  * global data structures
    147  */
    148 
    149 struct vm_map *kernel_map = NULL;
    150 
    151 /*
    152  * local data structues
    153  */
    154 
    155 static struct vm_map_kernel	kernel_map_store;
    156 static struct vm_map_entry	kernel_first_mapent_store;
    157 
    158 #if !defined(PMAP_MAP_POOLPAGE)
    159 
    160 /*
    161  * kva cache
    162  *
    163  * XXX maybe it's better to do this at the uvm_map layer.
    164  */
    165 
    166 #define	KM_VACACHE_SIZE	(32 * PAGE_SIZE) /* XXX tune */
    167 
    168 static void *km_vacache_alloc(struct pool *, int);
    169 static void km_vacache_free(struct pool *, void *);
    170 static void km_vacache_init(struct vm_map *, const char *, size_t);
    171 
    172 /* XXX */
    173 #define	KM_VACACHE_POOL_TO_MAP(pp) \
    174 	((struct vm_map *)((char *)(pp) - \
    175 	    offsetof(struct vm_map_kernel, vmk_vacache)))
    176 
    177 static void *
    178 km_vacache_alloc(struct pool *pp, int flags)
    179 {
    180 	vaddr_t va;
    181 	size_t size;
    182 	struct vm_map *map;
    183 	size = pp->pr_alloc->pa_pagesz;
    184 
    185 	map = KM_VACACHE_POOL_TO_MAP(pp);
    186 
    187 	va = vm_map_min(map); /* hint */
    188 	if (uvm_map(map, &va, size, NULL, UVM_UNKNOWN_OFFSET, size,
    189 	    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
    190 	    UVM_ADV_RANDOM, UVM_FLAG_QUANTUM |
    191 	    ((flags & PR_WAITOK) ? UVM_FLAG_WAITVA :
    192 	    UVM_FLAG_TRYLOCK | UVM_FLAG_NOWAIT))))
    193 		return NULL;
    194 
    195 	return (void *)va;
    196 }
    197 
    198 static void
    199 km_vacache_free(struct pool *pp, void *v)
    200 {
    201 	vaddr_t va = (vaddr_t)v;
    202 	size_t size = pp->pr_alloc->pa_pagesz;
    203 	struct vm_map *map;
    204 
    205 	map = KM_VACACHE_POOL_TO_MAP(pp);
    206 	uvm_unmap1(map, va, va + size, UVM_FLAG_QUANTUM|UVM_FLAG_VAONLY);
    207 }
    208 
    209 /*
    210  * km_vacache_init: initialize kva cache.
    211  */
    212 
    213 static void
    214 km_vacache_init(struct vm_map *map, const char *name, size_t size)
    215 {
    216 	struct vm_map_kernel *vmk;
    217 	struct pool *pp;
    218 	struct pool_allocator *pa;
    219 
    220 	KASSERT(VM_MAP_IS_KERNEL(map));
    221 	KASSERT(size < (vm_map_max(map) - vm_map_min(map)) / 2); /* sanity */
    222 
    223 	vmk = vm_map_to_kernel(map);
    224 	pp = &vmk->vmk_vacache;
    225 	pa = &vmk->vmk_vacache_allocator;
    226 	memset(pa, 0, sizeof(*pa));
    227 	pa->pa_alloc = km_vacache_alloc;
    228 	pa->pa_free = km_vacache_free;
    229 	pa->pa_pagesz = (unsigned int)size;
    230 	pa->pa_backingmap = map;
    231 	pa->pa_backingmapptr = NULL;
    232 	pool_init(pp, PAGE_SIZE, 0, 0, PR_NOTOUCH | PR_RECURSIVE, name, pa);
    233 }
    234 
    235 void
    236 uvm_km_vacache_init(struct vm_map *map, const char *name, size_t size)
    237 {
    238 
    239 	map->flags |= VM_MAP_VACACHE;
    240 	if (size == 0)
    241 		size = KM_VACACHE_SIZE;
    242 	km_vacache_init(map, name, size);
    243 }
    244 
    245 #else /* !defined(PMAP_MAP_POOLPAGE) */
    246 
    247 void
    248 uvm_km_vacache_init(struct vm_map *map, const char *name, size_t size)
    249 {
    250 
    251 	/* nothing */
    252 }
    253 
    254 #endif /* !defined(PMAP_MAP_POOLPAGE) */
    255 
    256 void
    257 uvm_km_va_drain(struct vm_map *map, uvm_flag_t flags)
    258 {
    259 	struct vm_map_kernel *vmk = vm_map_to_kernel(map);
    260 	const boolean_t intrsafe = (map->flags & VM_MAP_INTRSAFE) != 0;
    261 	int s = 0xdeadbeaf; /* XXX: gcc */
    262 
    263 	if (intrsafe) {
    264 		s = splvm();
    265 	}
    266 	callback_run_roundrobin(&vmk->vmk_reclaim_callback, NULL);
    267 	if (intrsafe) {
    268 		splx(s);
    269 	}
    270 }
    271 
    272 /*
    273  * uvm_km_init: init kernel maps and objects to reflect reality (i.e.
    274  * KVM already allocated for text, data, bss, and static data structures).
    275  *
    276  * => KVM is defined by VM_MIN_KERNEL_ADDRESS/VM_MAX_KERNEL_ADDRESS.
    277  *    we assume that [vmin -> start] has already been allocated and that
    278  *    "end" is the end.
    279  */
    280 
    281 void
    282 uvm_km_init(vaddr_t start, vaddr_t end)
    283 {
    284 	vaddr_t base = VM_MIN_KERNEL_ADDRESS;
    285 
    286 	/*
    287 	 * next, init kernel memory objects.
    288 	 */
    289 
    290 	/* kernel_object: for pageable anonymous kernel memory */
    291 	uao_init();
    292 	uvm.kernel_object = uao_create(VM_MAX_KERNEL_ADDRESS -
    293 				 VM_MIN_KERNEL_ADDRESS, UAO_FLAG_KERNOBJ);
    294 
    295 	/*
    296 	 * init the map and reserve any space that might already
    297 	 * have been allocated kernel space before installing.
    298 	 */
    299 
    300 	uvm_map_setup_kernel(&kernel_map_store, base, end, VM_MAP_PAGEABLE);
    301 	kernel_map_store.vmk_map.pmap = pmap_kernel();
    302 	if (start != base) {
    303 		int error;
    304 		struct uvm_map_args args;
    305 
    306 		error = uvm_map_prepare(&kernel_map_store.vmk_map,
    307 		    base, start - base,
    308 		    NULL, UVM_UNKNOWN_OFFSET, 0,
    309 		    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
    310 		    		UVM_ADV_RANDOM, UVM_FLAG_FIXED), &args);
    311 		if (!error) {
    312 			kernel_first_mapent_store.flags =
    313 			    UVM_MAP_KERNEL | UVM_MAP_FIRST;
    314 			error = uvm_map_enter(&kernel_map_store.vmk_map, &args,
    315 			    &kernel_first_mapent_store);
    316 		}
    317 
    318 		if (error)
    319 			panic(
    320 			    "uvm_km_init: could not reserve space for kernel");
    321 	}
    322 
    323 	/*
    324 	 * install!
    325 	 */
    326 
    327 	kernel_map = &kernel_map_store.vmk_map;
    328 	uvm_km_vacache_init(kernel_map, "kvakernel", 0);
    329 }
    330 
    331 /*
    332  * uvm_km_suballoc: allocate a submap in the kernel map.   once a submap
    333  * is allocated all references to that area of VM must go through it.  this
    334  * allows the locking of VAs in kernel_map to be broken up into regions.
    335  *
    336  * => if `fixed' is true, *vmin specifies where the region described
    337  *      by the submap must start
    338  * => if submap is non NULL we use that as the submap, otherwise we
    339  *	alloc a new map
    340  */
    341 
    342 struct vm_map *
    343 uvm_km_suballoc(struct vm_map *map, vaddr_t *vmin /* IN/OUT */,
    344     vaddr_t *vmax /* OUT */, vsize_t size, int flags, boolean_t fixed,
    345     struct vm_map_kernel *submap)
    346 {
    347 	int mapflags = UVM_FLAG_NOMERGE | (fixed ? UVM_FLAG_FIXED : 0);
    348 
    349 	KASSERT(vm_map_pmap(map) == pmap_kernel());
    350 
    351 	size = round_page(size);	/* round up to pagesize */
    352 	size += uvm_mapent_overhead(size, flags);
    353 
    354 	/*
    355 	 * first allocate a blank spot in the parent map
    356 	 */
    357 
    358 	if (uvm_map(map, vmin, size, NULL, UVM_UNKNOWN_OFFSET, 0,
    359 	    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
    360 	    UVM_ADV_RANDOM, mapflags)) != 0) {
    361 	       panic("uvm_km_suballoc: unable to allocate space in parent map");
    362 	}
    363 
    364 	/*
    365 	 * set VM bounds (vmin is filled in by uvm_map)
    366 	 */
    367 
    368 	*vmax = *vmin + size;
    369 
    370 	/*
    371 	 * add references to pmap and create or init the submap
    372 	 */
    373 
    374 	pmap_reference(vm_map_pmap(map));
    375 	if (submap == NULL) {
    376 		submap = malloc(sizeof(*submap), M_VMMAP, M_WAITOK);
    377 		if (submap == NULL)
    378 			panic("uvm_km_suballoc: unable to create submap");
    379 	}
    380 	uvm_map_setup_kernel(submap, *vmin, *vmax, flags);
    381 	submap->vmk_map.pmap = vm_map_pmap(map);
    382 
    383 	/*
    384 	 * now let uvm_map_submap plug in it...
    385 	 */
    386 
    387 	if (uvm_map_submap(map, *vmin, *vmax, &submap->vmk_map) != 0)
    388 		panic("uvm_km_suballoc: submap allocation failed");
    389 
    390 	return(&submap->vmk_map);
    391 }
    392 
    393 /*
    394  * uvm_km_pgremove: remove pages from a kernel uvm_object.
    395  *
    396  * => when you unmap a part of anonymous kernel memory you want to toss
    397  *    the pages right away.    (this gets called from uvm_unmap_...).
    398  */
    399 
    400 void
    401 uvm_km_pgremove(vaddr_t startva, vaddr_t endva)
    402 {
    403 	struct uvm_object * const uobj = uvm.kernel_object;
    404 	const voff_t start = startva - vm_map_min(kernel_map);
    405 	const voff_t end = endva - vm_map_min(kernel_map);
    406 	struct vm_page *pg;
    407 	voff_t curoff, nextoff;
    408 	int swpgonlydelta = 0;
    409 	UVMHIST_FUNC("uvm_km_pgremove"); UVMHIST_CALLED(maphist);
    410 
    411 	KASSERT(VM_MIN_KERNEL_ADDRESS <= startva);
    412 	KASSERT(startva < endva);
    413 	KASSERT(endva <= VM_MAX_KERNEL_ADDRESS);
    414 
    415 	simple_lock(&uobj->vmobjlock);
    416 
    417 	for (curoff = start; curoff < end; curoff = nextoff) {
    418 		nextoff = curoff + PAGE_SIZE;
    419 		pg = uvm_pagelookup(uobj, curoff);
    420 		if (pg != NULL && pg->flags & PG_BUSY) {
    421 			pg->flags |= PG_WANTED;
    422 			UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, 0,
    423 				    "km_pgrm", 0);
    424 			simple_lock(&uobj->vmobjlock);
    425 			nextoff = curoff;
    426 			continue;
    427 		}
    428 
    429 		/*
    430 		 * free the swap slot, then the page.
    431 		 */
    432 
    433 		if (pg == NULL &&
    434 		    uao_find_swslot(uobj, curoff >> PAGE_SHIFT) > 0) {
    435 			swpgonlydelta++;
    436 		}
    437 		uao_dropswap(uobj, curoff >> PAGE_SHIFT);
    438 		if (pg != NULL) {
    439 			uvm_lock_pageq();
    440 			uvm_pagefree(pg);
    441 			uvm_unlock_pageq();
    442 		}
    443 	}
    444 	simple_unlock(&uobj->vmobjlock);
    445 
    446 	if (swpgonlydelta > 0) {
    447 		simple_lock(&uvm.swap_data_lock);
    448 		KASSERT(uvmexp.swpgonly >= swpgonlydelta);
    449 		uvmexp.swpgonly -= swpgonlydelta;
    450 		simple_unlock(&uvm.swap_data_lock);
    451 	}
    452 }
    453 
    454 
    455 /*
    456  * uvm_km_pgremove_intrsafe: like uvm_km_pgremove(), but for non object backed
    457  *    regions.
    458  *
    459  * => when you unmap a part of anonymous kernel memory you want to toss
    460  *    the pages right away.    (this is called from uvm_unmap_...).
    461  * => none of the pages will ever be busy, and none of them will ever
    462  *    be on the active or inactive queues (because they have no object).
    463  */
    464 
    465 void
    466 uvm_km_pgremove_intrsafe(vaddr_t start, vaddr_t end)
    467 {
    468 	struct vm_page *pg;
    469 	paddr_t pa;
    470 	UVMHIST_FUNC("uvm_km_pgremove_intrsafe"); UVMHIST_CALLED(maphist);
    471 
    472 	KASSERT(VM_MIN_KERNEL_ADDRESS <= start);
    473 	KASSERT(start < end);
    474 	KASSERT(end <= VM_MAX_KERNEL_ADDRESS);
    475 
    476 	for (; start < end; start += PAGE_SIZE) {
    477 		if (!pmap_extract(pmap_kernel(), start, &pa)) {
    478 			continue;
    479 		}
    480 		pg = PHYS_TO_VM_PAGE(pa);
    481 		KASSERT(pg);
    482 		KASSERT(pg->uobject == NULL && pg->uanon == NULL);
    483 		uvm_pagefree(pg);
    484 	}
    485 }
    486 
    487 #if defined(DEBUG)
    488 void
    489 uvm_km_check_empty(vaddr_t start, vaddr_t end, boolean_t intrsafe)
    490 {
    491 	vaddr_t va;
    492 	paddr_t pa;
    493 
    494 	KDASSERT(VM_MIN_KERNEL_ADDRESS <= start);
    495 	KDASSERT(start < end);
    496 	KDASSERT(end <= VM_MAX_KERNEL_ADDRESS);
    497 
    498 	for (va = start; va < end; va += PAGE_SIZE) {
    499 		if (pmap_extract(pmap_kernel(), va, &pa)) {
    500 			panic("uvm_km_check_empty: va %p has pa 0x%llx",
    501 			    (void *)va, (long long)pa);
    502 		}
    503 		if (!intrsafe) {
    504 			const struct vm_page *pg;
    505 
    506 			simple_lock(&uvm.kernel_object->vmobjlock);
    507 			pg = uvm_pagelookup(uvm.kernel_object,
    508 			    va - vm_map_min(kernel_map));
    509 			simple_unlock(&uvm.kernel_object->vmobjlock);
    510 			if (pg) {
    511 				panic("uvm_km_check_empty: "
    512 				    "has page hashed at %p", (const void *)va);
    513 			}
    514 		}
    515 	}
    516 }
    517 #endif /* defined(DEBUG) */
    518 
    519 /*
    520  * uvm_km_alloc: allocate an area of kernel memory.
    521  *
    522  * => NOTE: we can return 0 even if we can wait if there is not enough
    523  *	free VM space in the map... caller should be prepared to handle
    524  *	this case.
    525  * => we return KVA of memory allocated
    526  */
    527 
    528 vaddr_t
    529 uvm_km_alloc(struct vm_map *map, vsize_t size, vsize_t align, uvm_flag_t flags)
    530 {
    531 	vaddr_t kva, loopva;
    532 	vaddr_t offset;
    533 	vsize_t loopsize;
    534 	struct vm_page *pg;
    535 	struct uvm_object *obj;
    536 	int pgaflags;
    537 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
    538 
    539 	KASSERT(vm_map_pmap(map) == pmap_kernel());
    540 	KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED ||
    541 		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE ||
    542 		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY);
    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_FLAG_QUANTUM)) != 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_PGA_USERESERVE;
    592 	if (flags & UVM_KMF_ZERO)
    593 		pgaflags |= UVM_PGA_ZERO;
    594 	while (loopsize) {
    595 		KASSERT(!pmap_extract(pmap_kernel(), loopva, NULL));
    596 
    597 		pg = uvm_pagealloc(NULL, offset, NULL, pgaflags);
    598 
    599 		/*
    600 		 * out of memory?
    601 		 */
    602 
    603 		if (__predict_false(pg == NULL)) {
    604 			if ((flags & UVM_KMF_NOWAIT) ||
    605 			    ((flags & UVM_KMF_CANFAIL) && !uvm_reclaimable())) {
    606 				/* free everything! */
    607 				uvm_km_free(map, kva, size,
    608 				    flags & UVM_KMF_TYPEMASK);
    609 				return (0);
    610 			} else {
    611 				uvm_wait("km_getwait2");	/* sleep here */
    612 				continue;
    613 			}
    614 		}
    615 
    616 		pg->flags &= ~PG_BUSY;	/* new page */
    617 		UVM_PAGE_OWN(pg, NULL);
    618 
    619 		/*
    620 		 * map it in
    621 		 */
    622 
    623 		pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg),
    624 		    VM_PROT_READ | VM_PROT_WRITE);
    625 		loopva += PAGE_SIZE;
    626 		offset += PAGE_SIZE;
    627 		loopsize -= PAGE_SIZE;
    628 	}
    629 
    630        	pmap_update(pmap_kernel());
    631 
    632 	UVMHIST_LOG(maphist,"<- done (kva=0x%x)", kva,0,0,0);
    633 	return(kva);
    634 }
    635 
    636 /*
    637  * uvm_km_free: free an area of kernel memory
    638  */
    639 
    640 void
    641 uvm_km_free(struct vm_map *map, vaddr_t addr, vsize_t size, uvm_flag_t flags)
    642 {
    643 
    644 	KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED ||
    645 		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE ||
    646 		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY);
    647 	KASSERT((addr & PAGE_MASK) == 0);
    648 	KASSERT(vm_map_pmap(map) == pmap_kernel());
    649 
    650 	size = round_page(size);
    651 
    652 	if (flags & UVM_KMF_PAGEABLE) {
    653 		uvm_km_pgremove(addr, addr + size);
    654 		pmap_remove(pmap_kernel(), addr, addr + size);
    655 	} else if (flags & UVM_KMF_WIRED) {
    656 		uvm_km_pgremove_intrsafe(addr, addr + size);
    657 		pmap_kremove(addr, size);
    658 	}
    659 
    660 	uvm_unmap1(map, addr, addr + size, UVM_FLAG_QUANTUM|UVM_FLAG_VAONLY);
    661 }
    662 
    663 /* Sanity; must specify both or none. */
    664 #if (defined(PMAP_MAP_POOLPAGE) || defined(PMAP_UNMAP_POOLPAGE)) && \
    665     (!defined(PMAP_MAP_POOLPAGE) || !defined(PMAP_UNMAP_POOLPAGE))
    666 #error Must specify MAP and UNMAP together.
    667 #endif
    668 
    669 /*
    670  * uvm_km_alloc_poolpage: allocate a page for the pool allocator
    671  *
    672  * => if the pmap specifies an alternate mapping method, we use it.
    673  */
    674 
    675 /* ARGSUSED */
    676 vaddr_t
    677 uvm_km_alloc_poolpage_cache(struct vm_map *map, boolean_t waitok)
    678 {
    679 #if defined(PMAP_MAP_POOLPAGE)
    680 	return uvm_km_alloc_poolpage(map, waitok);
    681 #else
    682 	struct vm_page *pg;
    683 	struct pool *pp = &vm_map_to_kernel(map)->vmk_vacache;
    684 	vaddr_t va;
    685 	int s = 0xdeadbeaf; /* XXX: gcc */
    686 	const boolean_t intrsafe = (map->flags & VM_MAP_INTRSAFE) != 0;
    687 
    688 	if ((map->flags & VM_MAP_VACACHE) == 0)
    689 		return uvm_km_alloc_poolpage(map, waitok);
    690 
    691 	if (intrsafe)
    692 		s = splvm();
    693 	va = (vaddr_t)pool_get(pp, waitok ? PR_WAITOK : PR_NOWAIT);
    694 	if (intrsafe)
    695 		splx(s);
    696 	if (va == 0)
    697 		return 0;
    698 	KASSERT(!pmap_extract(pmap_kernel(), va, NULL));
    699 again:
    700 	pg = uvm_pagealloc(NULL, 0, NULL, UVM_PGA_USERESERVE);
    701 	if (__predict_false(pg == NULL)) {
    702 		if (waitok) {
    703 			uvm_wait("plpg");
    704 			goto again;
    705 		} else {
    706 			if (intrsafe)
    707 				s = splvm();
    708 			pool_put(pp, (void *)va);
    709 			if (intrsafe)
    710 				splx(s);
    711 			return 0;
    712 		}
    713 	}
    714 	pmap_kenter_pa(va, VM_PAGE_TO_PHYS(pg), VM_PROT_READ|VM_PROT_WRITE);
    715 	pmap_update(pmap_kernel());
    716 
    717 	return va;
    718 #endif /* PMAP_MAP_POOLPAGE */
    719 }
    720 
    721 vaddr_t
    722 uvm_km_alloc_poolpage(struct vm_map *map, boolean_t waitok)
    723 {
    724 #if defined(PMAP_MAP_POOLPAGE)
    725 	struct vm_page *pg;
    726 	vaddr_t va;
    727 
    728  again:
    729 	pg = uvm_pagealloc(NULL, 0, NULL, UVM_PGA_USERESERVE);
    730 	if (__predict_false(pg == NULL)) {
    731 		if (waitok) {
    732 			uvm_wait("plpg");
    733 			goto again;
    734 		} else
    735 			return (0);
    736 	}
    737 	va = PMAP_MAP_POOLPAGE(VM_PAGE_TO_PHYS(pg));
    738 	if (__predict_false(va == 0))
    739 		uvm_pagefree(pg);
    740 	return (va);
    741 #else
    742 	vaddr_t va;
    743 	int s = 0xdeadbeaf; /* XXX: gcc */
    744 	const boolean_t intrsafe = (map->flags & VM_MAP_INTRSAFE) != 0;
    745 
    746 	if (intrsafe)
    747 		s = splvm();
    748 	va = uvm_km_alloc(map, PAGE_SIZE, 0,
    749 	    (waitok ? 0 : UVM_KMF_NOWAIT | UVM_KMF_TRYLOCK) | UVM_KMF_WIRED);
    750 	if (intrsafe)
    751 		splx(s);
    752 	return (va);
    753 #endif /* PMAP_MAP_POOLPAGE */
    754 }
    755 
    756 /*
    757  * uvm_km_free_poolpage: free a previously allocated pool page
    758  *
    759  * => if the pmap specifies an alternate unmapping method, we use it.
    760  */
    761 
    762 /* ARGSUSED */
    763 void
    764 uvm_km_free_poolpage_cache(struct vm_map *map, vaddr_t addr)
    765 {
    766 #if defined(PMAP_UNMAP_POOLPAGE)
    767 	uvm_km_free_poolpage(map, addr);
    768 #else
    769 	struct pool *pp;
    770 	int s = 0xdeadbeaf; /* XXX: gcc */
    771 	const boolean_t intrsafe = (map->flags & VM_MAP_INTRSAFE) != 0;
    772 
    773 	if ((map->flags & VM_MAP_VACACHE) == 0) {
    774 		uvm_km_free_poolpage(map, addr);
    775 		return;
    776 	}
    777 
    778 	KASSERT(pmap_extract(pmap_kernel(), addr, NULL));
    779 	uvm_km_pgremove_intrsafe(addr, addr + PAGE_SIZE);
    780 	pmap_kremove(addr, PAGE_SIZE);
    781 #if defined(DEBUG)
    782 	pmap_update(pmap_kernel());
    783 #endif
    784 	KASSERT(!pmap_extract(pmap_kernel(), addr, NULL));
    785 	pp = &vm_map_to_kernel(map)->vmk_vacache;
    786 	if (intrsafe)
    787 		s = splvm();
    788 	pool_put(pp, (void *)addr);
    789 	if (intrsafe)
    790 		splx(s);
    791 #endif
    792 }
    793 
    794 /* ARGSUSED */
    795 void
    796 uvm_km_free_poolpage(struct vm_map *map, vaddr_t addr)
    797 {
    798 #if defined(PMAP_UNMAP_POOLPAGE)
    799 	paddr_t pa;
    800 
    801 	pa = PMAP_UNMAP_POOLPAGE(addr);
    802 	uvm_pagefree(PHYS_TO_VM_PAGE(pa));
    803 #else
    804 	int s = 0xdeadbeaf; /* XXX: gcc */
    805 	const boolean_t intrsafe = (map->flags & VM_MAP_INTRSAFE) != 0;
    806 
    807 	if (intrsafe)
    808 		s = splvm();
    809 	uvm_km_free(map, addr, PAGE_SIZE, UVM_KMF_WIRED);
    810 	if (intrsafe)
    811 		splx(s);
    812 #endif /* PMAP_UNMAP_POOLPAGE */
    813 }
    814