Home | History | Annotate | Line # | Download | only in uvm
uvm_km.c revision 1.118
      1 /*	$NetBSD: uvm_km.c,v 1.118 2012/02/03 19:25:07 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 (at) 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 1.118 2012/02/03 19:25:07 matt 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 
    159 struct vm_map *kernel_map = NULL;
    160 
    161 /*
    162  * local data structues
    163  */
    164 
    165 static struct vm_map		kernel_map_store;
    166 static struct vm_map_entry	kernel_image_mapent_store;
    167 static struct vm_map_entry	kernel_kmem_mapent_store;
    168 
    169 int nkmempages = 0;
    170 vaddr_t kmembase;
    171 vsize_t kmemsize;
    172 
    173 vmem_t *kmem_arena;
    174 vmem_t *kmem_va_arena;
    175 
    176 /*
    177  * kmeminit_nkmempages: calculate the size of kmem_arena.
    178  */
    179 void
    180 kmeminit_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 	npages = physmem;
    193 
    194 	if (npages > NKMEMPAGES_MAX)
    195 		npages = NKMEMPAGES_MAX;
    196 
    197 	if (npages < NKMEMPAGES_MIN)
    198 		npages = NKMEMPAGES_MIN;
    199 
    200 	nkmempages = npages;
    201 }
    202 
    203 /*
    204  * uvm_km_bootstrap: init kernel maps and objects to reflect reality (i.e.
    205  * KVM already allocated for text, data, bss, and static data structures).
    206  *
    207  * => KVM is defined by VM_MIN_KERNEL_ADDRESS/VM_MAX_KERNEL_ADDRESS.
    208  *    we assume that [vmin -> start] has already been allocated and that
    209  *    "end" is the end.
    210  */
    211 
    212 void
    213 uvm_km_bootstrap(vaddr_t start, vaddr_t end)
    214 {
    215 	vaddr_t base = VM_MIN_KERNEL_ADDRESS;
    216 	struct uvm_map_args args;
    217 	int error;
    218 
    219 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
    220 	UVMHIST_LOG(maphist, "start=%"PRIxVADDR" end=%#"PRIxVADDR,
    221 	    start, end, 0,0);
    222 
    223 	kmeminit_nkmempages();
    224 	kmemsize = nkmempages * PAGE_SIZE;
    225 
    226 	/* kmemsize = MIN((((vsize_t)(end - start)) / 3),
    227 	    ((((vsize_t)uvmexp.npages) * PAGE_SIZE) / 2));
    228 	kmemsize = round_page(kmemsize); */
    229 
    230 	UVMHIST_LOG(maphist, "kmemsize=%#"PRIxVSIZE, kmemsize, 0,0,0);
    231 
    232 	/*
    233 	 * next, init kernel memory objects.
    234 	 */
    235 
    236 	/* kernel_object: for pageable anonymous kernel memory */
    237 	uvm_kernel_object = uao_create(VM_MAX_KERNEL_ADDRESS -
    238 				VM_MIN_KERNEL_ADDRESS, UAO_FLAG_KERNOBJ);
    239 
    240 	/*
    241 	 * init the map and reserve any space that might already
    242 	 * have been allocated kernel space before installing.
    243 	 */
    244 
    245 	uvm_map_setup(&kernel_map_store, base, end, VM_MAP_PAGEABLE);
    246 	kernel_map_store.pmap = pmap_kernel();
    247 	if (start != base) {
    248 		error = uvm_map_prepare(&kernel_map_store,
    249 		    base, start - base,
    250 		    NULL, UVM_UNKNOWN_OFFSET, 0,
    251 		    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
    252 		    		UVM_ADV_RANDOM, UVM_FLAG_FIXED), &args);
    253 		if (!error) {
    254 			kernel_image_mapent_store.flags =
    255 			    UVM_MAP_KERNEL | UVM_MAP_STATIC | UVM_MAP_NOMERGE;
    256 			error = uvm_map_enter(&kernel_map_store, &args,
    257 			    &kernel_image_mapent_store);
    258 		}
    259 
    260 		if (error)
    261 			panic(
    262 			    "uvm_km_bootstrap: could not reserve space for kernel");
    263 
    264 		kmembase = args.uma_start + args.uma_size;
    265 	} else {
    266 		kmembase = base;
    267 	}
    268 
    269 	error = uvm_map_prepare(&kernel_map_store,
    270 	    kmembase, kmemsize,
    271 	    NULL, UVM_UNKNOWN_OFFSET, 0,
    272 	    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
    273 	    		UVM_ADV_RANDOM, UVM_FLAG_FIXED), &args);
    274 	if (!error) {
    275 		kernel_kmem_mapent_store.flags =
    276 		    UVM_MAP_KERNEL | UVM_MAP_STATIC | UVM_MAP_NOMERGE;
    277 		error = uvm_map_enter(&kernel_map_store, &args,
    278 		    &kernel_kmem_mapent_store);
    279 	}
    280 
    281 	if (error)
    282 		panic("uvm_km_bootstrap: could not reserve kernel kmem");
    283 
    284 	/*
    285 	 * install!
    286 	 */
    287 
    288 	kernel_map = &kernel_map_store;
    289 
    290 	pool_subsystem_init();
    291 	vmem_bootstrap();
    292 
    293 	kmem_arena = vmem_create("kmem", kmembase, kmemsize, PAGE_SIZE,
    294 	    NULL, NULL, NULL,
    295 	    0, VM_NOSLEEP | VM_BOOTSTRAP, IPL_VM);
    296 
    297 	vmem_init(kmem_arena);
    298 
    299 	UVMHIST_LOG(maphist, "kmem vmem created (base=%#"PRIxVADDR
    300 	    ", size=%#"PRIxVSIZE, kmembase, kmemsize, 0,0);
    301 
    302 	kmem_va_arena = vmem_create("kva", 0, 0, PAGE_SIZE,
    303 	    vmem_alloc, vmem_free, kmem_arena,
    304 	    16 * PAGE_SIZE, VM_NOSLEEP | VM_BOOTSTRAP, IPL_VM);
    305 
    306 	UVMHIST_LOG(maphist, "<- done", 0,0,0,0);
    307 }
    308 
    309 /*
    310  * uvm_km_init: init the kernel maps virtual memory caches
    311  * and start the pool/kmem allocator.
    312  */
    313 void
    314 uvm_km_init(void)
    315 {
    316 
    317 	kmem_init();
    318 
    319 	kmeminit(); // killme
    320 }
    321 
    322 /*
    323  * uvm_km_suballoc: allocate a submap in the kernel map.   once a submap
    324  * is allocated all references to that area of VM must go through it.  this
    325  * allows the locking of VAs in kernel_map to be broken up into regions.
    326  *
    327  * => if `fixed' is true, *vmin specifies where the region described
    328  *   pager_map => used to map "buf" structures into kernel space
    329  *      by the submap must start
    330  * => if submap is non NULL we use that as the submap, otherwise we
    331  *	alloc a new map
    332  */
    333 
    334 struct vm_map *
    335 uvm_km_suballoc(struct vm_map *map, vaddr_t *vmin /* IN/OUT */,
    336     vaddr_t *vmax /* OUT */, vsize_t size, int flags, bool fixed,
    337     struct vm_map *submap)
    338 {
    339 	int mapflags = UVM_FLAG_NOMERGE | (fixed ? UVM_FLAG_FIXED : 0);
    340 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
    341 
    342 	KASSERT(vm_map_pmap(map) == pmap_kernel());
    343 
    344 	size = round_page(size);	/* round up to pagesize */
    345 
    346 	/*
    347 	 * first allocate a blank spot in the parent map
    348 	 */
    349 
    350 	if (uvm_map(map, vmin, size, NULL, UVM_UNKNOWN_OFFSET, 0,
    351 	    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
    352 	    UVM_ADV_RANDOM, mapflags)) != 0) {
    353 		panic("%s: unable to allocate space in parent map", __func__);
    354 	}
    355 
    356 	/*
    357 	 * set VM bounds (vmin is filled in by uvm_map)
    358 	 */
    359 
    360 	*vmax = *vmin + size;
    361 
    362 	/*
    363 	 * add references to pmap and create or init the submap
    364 	 */
    365 
    366 	pmap_reference(vm_map_pmap(map));
    367 	if (submap == NULL) {
    368 		submap = kmem_alloc(sizeof(*submap), KM_SLEEP);
    369 		if (submap == NULL)
    370 			panic("uvm_km_suballoc: unable to create submap");
    371 	}
    372 	uvm_map_setup(submap, *vmin, *vmax, flags);
    373 	submap->pmap = vm_map_pmap(map);
    374 
    375 	/*
    376 	 * now let uvm_map_submap plug in it...
    377 	 */
    378 
    379 	if (uvm_map_submap(map, *vmin, *vmax, submap) != 0)
    380 		panic("uvm_km_suballoc: submap allocation failed");
    381 
    382 	return(submap);
    383 }
    384 
    385 /*
    386  * uvm_km_pgremove: remove pages from a kernel uvm_object and KVA.
    387  */
    388 
    389 void
    390 uvm_km_pgremove(vaddr_t startva, vaddr_t endva)
    391 {
    392 	struct uvm_object * const uobj = uvm_kernel_object;
    393 	const voff_t start = startva - vm_map_min(kernel_map);
    394 	const voff_t end = endva - vm_map_min(kernel_map);
    395 	struct vm_page *pg;
    396 	voff_t curoff, nextoff;
    397 	int swpgonlydelta = 0;
    398 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
    399 
    400 	KASSERT(VM_MIN_KERNEL_ADDRESS <= startva);
    401 	KASSERT(startva < endva);
    402 	KASSERT(endva <= VM_MAX_KERNEL_ADDRESS);
    403 
    404 	mutex_enter(uobj->vmobjlock);
    405 	pmap_remove(pmap_kernel(), startva, endva);
    406 	for (curoff = start; curoff < end; curoff = nextoff) {
    407 		nextoff = curoff + PAGE_SIZE;
    408 		pg = uvm_pagelookup(uobj, curoff);
    409 		if (pg != NULL && pg->flags & PG_BUSY) {
    410 			pg->flags |= PG_WANTED;
    411 			UVM_UNLOCK_AND_WAIT(pg, uobj->vmobjlock, 0,
    412 				    "km_pgrm", 0);
    413 			mutex_enter(uobj->vmobjlock);
    414 			nextoff = curoff;
    415 			continue;
    416 		}
    417 
    418 		/*
    419 		 * free the swap slot, then the page.
    420 		 */
    421 
    422 		if (pg == NULL &&
    423 		    uao_find_swslot(uobj, curoff >> PAGE_SHIFT) > 0) {
    424 			swpgonlydelta++;
    425 		}
    426 		uao_dropswap(uobj, curoff >> PAGE_SHIFT);
    427 		if (pg != NULL) {
    428 			mutex_enter(&uvm_pageqlock);
    429 			uvm_pagefree(pg);
    430 			mutex_exit(&uvm_pageqlock);
    431 		}
    432 	}
    433 	mutex_exit(uobj->vmobjlock);
    434 
    435 	if (swpgonlydelta > 0) {
    436 		mutex_enter(&uvm_swap_data_lock);
    437 		KASSERT(uvmexp.swpgonly >= swpgonlydelta);
    438 		uvmexp.swpgonly -= swpgonlydelta;
    439 		mutex_exit(&uvm_swap_data_lock);
    440 	}
    441 }
    442 
    443 
    444 /*
    445  * uvm_km_pgremove_intrsafe: like uvm_km_pgremove(), but for non object backed
    446  *    regions.
    447  *
    448  * => when you unmap a part of anonymous kernel memory you want to toss
    449  *    the pages right away.    (this is called from uvm_unmap_...).
    450  * => none of the pages will ever be busy, and none of them will ever
    451  *    be on the active or inactive queues (because they have no object).
    452  */
    453 
    454 void
    455 uvm_km_pgremove_intrsafe(struct vm_map *map, vaddr_t start, vaddr_t end)
    456 {
    457 	struct vm_page *pg;
    458 	paddr_t pa;
    459 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
    460 
    461 	KASSERT(VM_MAP_IS_KERNEL(map));
    462 	KASSERT(vm_map_min(map) <= start);
    463 	KASSERT(start < end);
    464 	KASSERT(end <= vm_map_max(map));
    465 
    466 	for (; start < end; start += PAGE_SIZE) {
    467 		if (!pmap_extract(pmap_kernel(), start, &pa)) {
    468 			continue;
    469 		}
    470 		pg = PHYS_TO_VM_PAGE(pa);
    471 		KASSERT(pg);
    472 		KASSERT(pg->uobject == NULL && pg->uanon == NULL);
    473 		KASSERT((pg->flags & PG_BUSY) == 0);
    474 		uvm_pagefree(pg);
    475 	}
    476 }
    477 
    478 #if defined(DEBUG)
    479 void
    480 uvm_km_check_empty(struct vm_map *map, vaddr_t start, vaddr_t end)
    481 {
    482 	struct vm_page *pg;
    483 	vaddr_t va;
    484 	paddr_t pa;
    485 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
    486 
    487 	KDASSERT(VM_MAP_IS_KERNEL(map));
    488 	KDASSERT(vm_map_min(map) <= start);
    489 	KDASSERT(start < end);
    490 	KDASSERT(end <= vm_map_max(map));
    491 
    492 	for (va = start; va < end; va += PAGE_SIZE) {
    493 		if (pmap_extract(pmap_kernel(), va, &pa)) {
    494 			panic("uvm_km_check_empty: va %p has pa 0x%llx",
    495 			    (void *)va, (long long)pa);
    496 		}
    497 		if ((map->flags & VM_MAP_INTRSAFE) == 0) {
    498 			mutex_enter(uvm_kernel_object->vmobjlock);
    499 			pg = uvm_pagelookup(uvm_kernel_object,
    500 			    va - vm_map_min(kernel_map));
    501 			mutex_exit(uvm_kernel_object->vmobjlock);
    502 			if (pg) {
    503 				panic("uvm_km_check_empty: "
    504 				    "has page hashed at %p", (const void *)va);
    505 			}
    506 		}
    507 	}
    508 }
    509 #endif /* defined(DEBUG) */
    510 
    511 /*
    512  * uvm_km_alloc: allocate an area of kernel memory.
    513  *
    514  * => NOTE: we can return 0 even if we can wait if there is not enough
    515  *	free VM space in the map... caller should be prepared to handle
    516  *	this case.
    517  * => we return KVA of memory allocated
    518  */
    519 
    520 vaddr_t
    521 uvm_km_alloc(struct vm_map *map, vsize_t size, vsize_t align, uvm_flag_t flags)
    522 {
    523 	vaddr_t kva, loopva;
    524 	vaddr_t offset;
    525 	vsize_t loopsize;
    526 	struct vm_page *pg;
    527 	struct uvm_object *obj;
    528 	int pgaflags;
    529 	vm_prot_t prot;
    530 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
    531 
    532 	KASSERT(vm_map_pmap(map) == pmap_kernel());
    533 	KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED ||
    534 		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE ||
    535 		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY);
    536 	KASSERT((flags & UVM_KMF_VAONLY) != 0 || (flags & UVM_KMF_COLORMATCH) == 0);
    537 	KASSERT((flags & UVM_KMF_COLORMATCH) == 0 || (flags & UVM_KMF_VAONLY) != 0);
    538 
    539 	/*
    540 	 * setup for call
    541 	 */
    542 
    543 	kva = vm_map_min(map);	/* hint */
    544 	size = round_page(size);
    545 	obj = (flags & UVM_KMF_PAGEABLE) ? uvm_kernel_object : NULL;
    546 	UVMHIST_LOG(maphist,"  (map=0x%x, obj=0x%x, size=0x%x, flags=%d)",
    547 		    map, obj, size, flags);
    548 
    549 	/*
    550 	 * allocate some virtual space
    551 	 */
    552 
    553 	if (__predict_false(uvm_map(map, &kva, size, obj, UVM_UNKNOWN_OFFSET,
    554 	    align, UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
    555 	    UVM_ADV_RANDOM,
    556 	    (flags & (UVM_KMF_TRYLOCK | UVM_KMF_NOWAIT | UVM_KMF_WAITVA
    557 	     | UVM_KMF_COLORMATCH)))) != 0)) {
    558 		UVMHIST_LOG(maphist, "<- done (no VM)",0,0,0,0);
    559 		return(0);
    560 	}
    561 
    562 	/*
    563 	 * if all we wanted was VA, return now
    564 	 */
    565 
    566 	if (flags & (UVM_KMF_VAONLY | UVM_KMF_PAGEABLE)) {
    567 		UVMHIST_LOG(maphist,"<- done valloc (kva=0x%x)", kva,0,0,0);
    568 		return(kva);
    569 	}
    570 
    571 	/*
    572 	 * recover object offset from virtual address
    573 	 */
    574 
    575 	offset = kva - vm_map_min(kernel_map);
    576 	UVMHIST_LOG(maphist, "  kva=0x%x, offset=0x%x", kva, offset,0,0);
    577 
    578 	/*
    579 	 * now allocate and map in the memory... note that we are the only ones
    580 	 * whom should ever get a handle on this area of VM.
    581 	 */
    582 
    583 	loopva = kva;
    584 	loopsize = size;
    585 
    586 	pgaflags = UVM_FLAG_COLORMATCH;
    587 	if (flags & UVM_KMF_NOWAIT)
    588 		pgaflags |= UVM_PGA_USERESERVE;
    589 	if (flags & UVM_KMF_ZERO)
    590 		pgaflags |= UVM_PGA_ZERO;
    591 	prot = VM_PROT_READ | VM_PROT_WRITE;
    592 	if (flags & UVM_KMF_EXEC)
    593 		prot |= VM_PROT_EXECUTE;
    594 	while (loopsize) {
    595 		KASSERTMSG(!pmap_extract(pmap_kernel(), loopva, NULL),
    596 		    "loopva=%#"PRIxVADDR, loopva);
    597 
    598 		pg = uvm_pagealloc_strat(NULL, offset, NULL, pgaflags,
    599 #ifdef UVM_KM_VMFREELIST
    600 		   UVM_PGA_STRAT_ONLY, UVM_KM_VMFREELIST
    601 #else
    602 		   UVM_PGA_STRAT_NORMAL, 0
    603 #endif
    604 		   );
    605 
    606 		/*
    607 		 * out of memory?
    608 		 */
    609 
    610 		if (__predict_false(pg == NULL)) {
    611 			if ((flags & UVM_KMF_NOWAIT) ||
    612 			    ((flags & UVM_KMF_CANFAIL) && !uvm_reclaimable())) {
    613 				/* free everything! */
    614 				uvm_km_free(map, kva, size,
    615 				    flags & UVM_KMF_TYPEMASK);
    616 				return (0);
    617 			} else {
    618 				uvm_wait("km_getwait2");	/* sleep here */
    619 				continue;
    620 			}
    621 		}
    622 
    623 		pg->flags &= ~PG_BUSY;	/* new page */
    624 		UVM_PAGE_OWN(pg, NULL);
    625 
    626 		/*
    627 		 * map it in
    628 		 */
    629 
    630 		pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg),
    631 		    prot, PMAP_KMPAGE);
    632 		loopva += PAGE_SIZE;
    633 		offset += PAGE_SIZE;
    634 		loopsize -= PAGE_SIZE;
    635 	}
    636 
    637 	pmap_update(pmap_kernel());
    638 
    639 	UVMHIST_LOG(maphist,"<- done (kva=0x%x)", kva,0,0,0);
    640 	return(kva);
    641 }
    642 
    643 /*
    644  * uvm_km_free: free an area of kernel memory
    645  */
    646 
    647 void
    648 uvm_km_free(struct vm_map *map, vaddr_t addr, vsize_t size, uvm_flag_t flags)
    649 {
    650 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
    651 
    652 	KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED ||
    653 		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE ||
    654 		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY);
    655 	KASSERT((addr & PAGE_MASK) == 0);
    656 	KASSERT(vm_map_pmap(map) == pmap_kernel());
    657 
    658 	size = round_page(size);
    659 
    660 	if (flags & UVM_KMF_PAGEABLE) {
    661 		uvm_km_pgremove(addr, addr + size);
    662 	} else if (flags & UVM_KMF_WIRED) {
    663 		/*
    664 		 * Note: uvm_km_pgremove_intrsafe() extracts mapping, thus
    665 		 * remove it after.  See comment below about KVA visibility.
    666 		 */
    667 		uvm_km_pgremove_intrsafe(map, addr, addr + size);
    668 		pmap_kremove(addr, size);
    669 	}
    670 
    671 	/*
    672 	 * Note: uvm_unmap_remove() calls pmap_update() for us, before
    673 	 * KVA becomes globally available.
    674 	 */
    675 
    676 	uvm_unmap1(map, addr, addr + size, UVM_FLAG_VAONLY);
    677 }
    678 
    679 /* Sanity; must specify both or none. */
    680 #if (defined(PMAP_MAP_POOLPAGE) || defined(PMAP_UNMAP_POOLPAGE)) && \
    681     (!defined(PMAP_MAP_POOLPAGE) || !defined(PMAP_UNMAP_POOLPAGE))
    682 #error Must specify MAP and UNMAP together.
    683 #endif
    684 
    685 int
    686 uvm_km_kmem_alloc(vmem_t *vm, vmem_size_t size, vm_flag_t flags,
    687     vmem_addr_t *addr)
    688 {
    689 	struct vm_page *pg;
    690 	vmem_addr_t va;
    691 	int rc;
    692 	vaddr_t loopva;
    693 	vsize_t loopsize;
    694 
    695 	size = round_page(size);
    696 
    697 #if defined(PMAP_MAP_POOLPAGE)
    698 	if (size == PAGE_SIZE) {
    699 again:
    700 #ifdef PMAP_ALLOC_POOLPAGE
    701 		pg = PMAP_ALLOC_POOLPAGE((flags & VM_SLEEP) ?
    702 		   0 : UVM_PGA_USERESERVE);
    703 #else
    704 		pg = uvm_pagealloc(NULL, 0, NULL,
    705 		   (flags & VM_SLEEP) ? 0 : UVM_PGA_USERESERVE);
    706 #endif /* PMAP_ALLOC_POOLPAGE */
    707 		if (__predict_false(pg == NULL)) {
    708 			if (flags & VM_SLEEP) {
    709 				uvm_wait("plpg");
    710 				goto again;
    711 			}
    712 		}
    713 		va = PMAP_MAP_POOLPAGE(VM_PAGE_TO_PHYS(pg));
    714 		if (__predict_false(va == 0)) {
    715 			uvm_pagefree(pg);
    716 			return ENOMEM;
    717 		}
    718 		*addr = va;
    719 		return 0;
    720 	}
    721 #endif /* PMAP_MAP_POOLPAGE */
    722 
    723 	rc = vmem_alloc(vm, size, flags, &va);
    724 	if (rc != 0)
    725 		return rc;
    726 
    727 	loopva = va;
    728 	loopsize = size;
    729 
    730 	while (loopsize) {
    731 		KASSERTMSG(!pmap_extract(pmap_kernel(), loopva, NULL),
    732 		    "loopva=%#"PRIxVADDR" loopsize=%#"PRIxVSIZE" vmem=%p",
    733 		    loopva, loopsize, vm);
    734 
    735 		pg = uvm_pagealloc(NULL, loopva, NULL,
    736 		    UVM_FLAG_COLORMATCH
    737 		    | ((flags & VM_SLEEP) ? 0 : UVM_PGA_USERESERVE));
    738 		if (__predict_false(pg == NULL)) {
    739 			if (flags & VM_SLEEP) {
    740 				uvm_wait("plpg");
    741 				continue;
    742 			} else {
    743 				uvm_km_pgremove_intrsafe(kernel_map, va,
    744 				    va + size);
    745 				pmap_kremove(va, size);
    746 				vmem_free(kmem_va_arena, va, size);
    747 				return ENOMEM;
    748 			}
    749 		}
    750 
    751 		pg->flags &= ~PG_BUSY;	/* new page */
    752 		UVM_PAGE_OWN(pg, NULL);
    753 		pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg),
    754 		    VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE);
    755 
    756 		loopva += PAGE_SIZE;
    757 		loopsize -= PAGE_SIZE;
    758 	}
    759 	pmap_update(pmap_kernel());
    760 
    761 	*addr = va;
    762 
    763 	return 0;
    764 }
    765 
    766 void
    767 uvm_km_kmem_free(vmem_t *vm, vmem_addr_t addr, size_t size)
    768 {
    769 
    770 	size = round_page(size);
    771 #if defined(PMAP_UNMAP_POOLPAGE)
    772 	if (size == PAGE_SIZE) {
    773 		paddr_t pa;
    774 
    775 		pa = PMAP_UNMAP_POOLPAGE(addr);
    776 		uvm_pagefree(PHYS_TO_VM_PAGE(pa));
    777 		return;
    778 	}
    779 #endif /* PMAP_UNMAP_POOLPAGE */
    780 	uvm_km_pgremove_intrsafe(kernel_map, addr, addr + size);
    781 	pmap_kremove(addr, size);
    782 	pmap_update(pmap_kernel());
    783 
    784 	vmem_free(vm, addr, size);
    785 }
    786 
    787 bool
    788 uvm_km_va_starved_p(void)
    789 {
    790 	vmem_size_t total;
    791 	vmem_size_t free;
    792 
    793 	total = vmem_size(kmem_arena, VMEM_ALLOC|VMEM_FREE);
    794 	free = vmem_size(kmem_arena, VMEM_FREE);
    795 
    796 	return (free < (total / 10));
    797 }
    798 
    799