Home | History | Annotate | Line # | Download | only in uvm
uvm_km.c revision 1.113
      1 /*	$NetBSD: uvm_km.c,v 1.113 2012/01/29 12:37:01 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 (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 1.113 2012/01/29 12:37:01 para Exp $");
    124 
    125 #include "opt_uvmhist.h"
    126 
    127 #include <sys/param.h>
    128 #include <sys/systm.h>
    129 #include <sys/proc.h>
    130 #include <sys/pool.h>
    131 #include <sys/vmem.h>
    132 #include <sys/kmem.h>
    133 
    134 #include <uvm/uvm.h>
    135 
    136 /*
    137  * global data structures
    138  */
    139 
    140 struct vm_map *kernel_map = NULL;
    141 
    142 /*
    143  * local data structues
    144  */
    145 
    146 static struct vm_map		kernel_map_store;
    147 static struct vm_map_entry	kernel_image_mapent_store;
    148 static struct vm_map_entry	kernel_kmem_mapent_store;
    149 
    150 vaddr_t kmembase;
    151 vsize_t kmemsize;
    152 
    153 vmem_t *kmem_arena;
    154 vmem_t *kmem_va_arena;
    155 
    156 /*
    157  * uvm_km_bootstrap: init kernel maps and objects to reflect reality (i.e.
    158  * KVM already allocated for text, data, bss, and static data structures).
    159  *
    160  * => KVM is defined by VM_MIN_KERNEL_ADDRESS/VM_MAX_KERNEL_ADDRESS.
    161  *    we assume that [vmin -> start] has already been allocated and that
    162  *    "end" is the end.
    163  */
    164 
    165 void
    166 uvm_km_bootstrap(vaddr_t start, vaddr_t end)
    167 {
    168 	vaddr_t base = VM_MIN_KERNEL_ADDRESS;
    169 
    170 	kmemsize = MIN(((((vsize_t)(end - start)) / 3) * 2),
    171 	    ((((vsize_t)uvmexp.npages) * PAGE_SIZE)));
    172 	kmemsize = round_page(kmemsize);
    173 
    174 	/*
    175 	 * next, init kernel memory objects.
    176 	 */
    177 
    178 	/* kernel_object: for pageable anonymous kernel memory */
    179 	uvm_kernel_object = uao_create(VM_MAX_KERNEL_ADDRESS -
    180 				VM_MIN_KERNEL_ADDRESS, UAO_FLAG_KERNOBJ);
    181 
    182 	/*
    183 	 * init the map and reserve any space that might already
    184 	 * have been allocated kernel space before installing.
    185 	 */
    186 
    187 	uvm_map_setup(&kernel_map_store, base, end, VM_MAP_PAGEABLE);
    188 	kernel_map_store.pmap = pmap_kernel();
    189 	if (start != base) {
    190 		int error;
    191 		struct uvm_map_args args;
    192 
    193 		error = uvm_map_prepare(&kernel_map_store,
    194 		    base, start - base,
    195 		    NULL, UVM_UNKNOWN_OFFSET, 0,
    196 		    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
    197 		    		UVM_ADV_RANDOM, UVM_FLAG_FIXED), &args);
    198 		if (!error) {
    199 			kernel_image_mapent_store.flags =
    200 			    UVM_MAP_KERNEL | UVM_MAP_STATIC | UVM_MAP_NOMERGE;
    201 			error = uvm_map_enter(&kernel_map_store, &args,
    202 			    &kernel_image_mapent_store);
    203 		}
    204 
    205 		if (error)
    206 			panic(
    207 			    "uvm_km_bootstrap: could not reserve space for kernel");
    208 
    209 		kmembase = args.uma_start + args.uma_size;
    210 		error = uvm_map_prepare(&kernel_map_store,
    211 		    kmembase, kmemsize,
    212 		    NULL, UVM_UNKNOWN_OFFSET, 0,
    213 		    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
    214 		    		UVM_ADV_RANDOM, UVM_FLAG_FIXED), &args);
    215 		if (!error) {
    216 			kernel_kmem_mapent_store.flags =
    217 			    UVM_MAP_KERNEL | UVM_MAP_STATIC | UVM_MAP_NOMERGE;
    218 			error = uvm_map_enter(&kernel_map_store, &args,
    219 			    &kernel_kmem_mapent_store);
    220 		}
    221 
    222 		if (error)
    223 			panic(
    224 			    "uvm_km_bootstrap: could not reserve kernel kmem");
    225 	}
    226 
    227 	/*
    228 	 * install!
    229 	 */
    230 
    231 	kernel_map = &kernel_map_store;
    232 
    233 	pool_subsystem_init();
    234 	vmem_bootstrap();
    235 
    236 	kmem_arena = vmem_create("kmem", kmembase, kmemsize, PAGE_SIZE,
    237 	    NULL, NULL, NULL,
    238 	    0, VM_NOSLEEP | VM_BOOTSTRAP, IPL_VM);
    239 
    240 	vmem_init(kmem_arena);
    241 
    242 	kmem_va_arena = vmem_create("kva", 0, 0, PAGE_SIZE,
    243 	    vmem_alloc, vmem_free, kmem_arena,
    244 	    16 * PAGE_SIZE, VM_NOSLEEP | VM_BOOTSTRAP, IPL_VM);
    245 }
    246 
    247 /*
    248  * uvm_km_init: init the kernel maps virtual memory caches
    249  * and start the pool/kmem allocator.
    250  */
    251 void
    252 uvm_km_init(void)
    253 {
    254 
    255 	kmem_init();
    256 
    257 	kmeminit(); // killme
    258 }
    259 
    260 /*
    261  * uvm_km_suballoc: allocate a submap in the kernel map.   once a submap
    262  * is allocated all references to that area of VM must go through it.  this
    263  * allows the locking of VAs in kernel_map to be broken up into regions.
    264  *
    265  * => if `fixed' is true, *vmin specifies where the region described
    266  *   pager_map => used to map "buf" structures into kernel space
    267  *      by the submap must start
    268  * => if submap is non NULL we use that as the submap, otherwise we
    269  *	alloc a new map
    270  */
    271 
    272 struct vm_map *
    273 uvm_km_suballoc(struct vm_map *map, vaddr_t *vmin /* IN/OUT */,
    274     vaddr_t *vmax /* OUT */, vsize_t size, int flags, bool fixed,
    275     struct vm_map *submap)
    276 {
    277 	int mapflags = UVM_FLAG_NOMERGE | (fixed ? UVM_FLAG_FIXED : 0);
    278 
    279 	KASSERT(vm_map_pmap(map) == pmap_kernel());
    280 
    281 	size = round_page(size);	/* round up to pagesize */
    282 
    283 	/*
    284 	 * first allocate a blank spot in the parent map
    285 	 */
    286 
    287 	if (uvm_map(map, vmin, size, NULL, UVM_UNKNOWN_OFFSET, 0,
    288 	    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
    289 	    UVM_ADV_RANDOM, mapflags)) != 0) {
    290 	       panic("uvm_km_suballoc: unable to allocate space in parent map");
    291 	}
    292 
    293 	/*
    294 	 * set VM bounds (vmin is filled in by uvm_map)
    295 	 */
    296 
    297 	*vmax = *vmin + size;
    298 
    299 	/*
    300 	 * add references to pmap and create or init the submap
    301 	 */
    302 
    303 	pmap_reference(vm_map_pmap(map));
    304 	if (submap == NULL) {
    305 		submap = kmem_alloc(sizeof(*submap), KM_SLEEP);
    306 		if (submap == NULL)
    307 			panic("uvm_km_suballoc: unable to create submap");
    308 	}
    309 	uvm_map_setup(submap, *vmin, *vmax, flags);
    310 	submap->pmap = vm_map_pmap(map);
    311 
    312 	/*
    313 	 * now let uvm_map_submap plug in it...
    314 	 */
    315 
    316 	if (uvm_map_submap(map, *vmin, *vmax, submap) != 0)
    317 		panic("uvm_km_suballoc: submap allocation failed");
    318 
    319 	return(submap);
    320 }
    321 
    322 /*
    323  * uvm_km_pgremove: remove pages from a kernel uvm_object and KVA.
    324  */
    325 
    326 void
    327 uvm_km_pgremove(vaddr_t startva, vaddr_t endva)
    328 {
    329 	struct uvm_object * const uobj = uvm_kernel_object;
    330 	const voff_t start = startva - vm_map_min(kernel_map);
    331 	const voff_t end = endva - vm_map_min(kernel_map);
    332 	struct vm_page *pg;
    333 	voff_t curoff, nextoff;
    334 	int swpgonlydelta = 0;
    335 	UVMHIST_FUNC("uvm_km_pgremove"); UVMHIST_CALLED(maphist);
    336 
    337 	KASSERT(VM_MIN_KERNEL_ADDRESS <= startva);
    338 	KASSERT(startva < endva);
    339 	KASSERT(endva <= VM_MAX_KERNEL_ADDRESS);
    340 
    341 	mutex_enter(uobj->vmobjlock);
    342 	pmap_remove(pmap_kernel(), startva, endva);
    343 	for (curoff = start; curoff < end; curoff = nextoff) {
    344 		nextoff = curoff + PAGE_SIZE;
    345 		pg = uvm_pagelookup(uobj, curoff);
    346 		if (pg != NULL && pg->flags & PG_BUSY) {
    347 			pg->flags |= PG_WANTED;
    348 			UVM_UNLOCK_AND_WAIT(pg, uobj->vmobjlock, 0,
    349 				    "km_pgrm", 0);
    350 			mutex_enter(uobj->vmobjlock);
    351 			nextoff = curoff;
    352 			continue;
    353 		}
    354 
    355 		/*
    356 		 * free the swap slot, then the page.
    357 		 */
    358 
    359 		if (pg == NULL &&
    360 		    uao_find_swslot(uobj, curoff >> PAGE_SHIFT) > 0) {
    361 			swpgonlydelta++;
    362 		}
    363 		uao_dropswap(uobj, curoff >> PAGE_SHIFT);
    364 		if (pg != NULL) {
    365 			mutex_enter(&uvm_pageqlock);
    366 			uvm_pagefree(pg);
    367 			mutex_exit(&uvm_pageqlock);
    368 		}
    369 	}
    370 	mutex_exit(uobj->vmobjlock);
    371 
    372 	if (swpgonlydelta > 0) {
    373 		mutex_enter(&uvm_swap_data_lock);
    374 		KASSERT(uvmexp.swpgonly >= swpgonlydelta);
    375 		uvmexp.swpgonly -= swpgonlydelta;
    376 		mutex_exit(&uvm_swap_data_lock);
    377 	}
    378 }
    379 
    380 
    381 /*
    382  * uvm_km_pgremove_intrsafe: like uvm_km_pgremove(), but for non object backed
    383  *    regions.
    384  *
    385  * => when you unmap a part of anonymous kernel memory you want to toss
    386  *    the pages right away.    (this is called from uvm_unmap_...).
    387  * => none of the pages will ever be busy, and none of them will ever
    388  *    be on the active or inactive queues (because they have no object).
    389  */
    390 
    391 void
    392 uvm_km_pgremove_intrsafe(struct vm_map *map, vaddr_t start, vaddr_t end)
    393 {
    394 	struct vm_page *pg;
    395 	paddr_t pa;
    396 	UVMHIST_FUNC("uvm_km_pgremove_intrsafe"); UVMHIST_CALLED(maphist);
    397 
    398 	KASSERT(VM_MAP_IS_KERNEL(map));
    399 	KASSERT(vm_map_min(map) <= start);
    400 	KASSERT(start < end);
    401 	KASSERT(end <= vm_map_max(map));
    402 
    403 	for (; start < end; start += PAGE_SIZE) {
    404 		if (!pmap_extract(pmap_kernel(), start, &pa)) {
    405 			continue;
    406 		}
    407 		pg = PHYS_TO_VM_PAGE(pa);
    408 		KASSERT(pg);
    409 		KASSERT(pg->uobject == NULL && pg->uanon == NULL);
    410 		KASSERT((pg->flags & PG_BUSY) == 0);
    411 		uvm_pagefree(pg);
    412 	}
    413 }
    414 
    415 #if defined(DEBUG)
    416 void
    417 uvm_km_check_empty(struct vm_map *map, vaddr_t start, vaddr_t end)
    418 {
    419 	struct vm_page *pg;
    420 	vaddr_t va;
    421 	paddr_t pa;
    422 
    423 	KDASSERT(VM_MAP_IS_KERNEL(map));
    424 	KDASSERT(vm_map_min(map) <= start);
    425 	KDASSERT(start < end);
    426 	KDASSERT(end <= vm_map_max(map));
    427 
    428 	for (va = start; va < end; va += PAGE_SIZE) {
    429 		if (pmap_extract(pmap_kernel(), va, &pa)) {
    430 			panic("uvm_km_check_empty: va %p has pa 0x%llx",
    431 			    (void *)va, (long long)pa);
    432 		}
    433 		if ((map->flags & VM_MAP_INTRSAFE) == 0) {
    434 			mutex_enter(uvm_kernel_object->vmobjlock);
    435 			pg = uvm_pagelookup(uvm_kernel_object,
    436 			    va - vm_map_min(kernel_map));
    437 			mutex_exit(uvm_kernel_object->vmobjlock);
    438 			if (pg) {
    439 				panic("uvm_km_check_empty: "
    440 				    "has page hashed at %p", (const void *)va);
    441 			}
    442 		}
    443 	}
    444 }
    445 #endif /* defined(DEBUG) */
    446 
    447 /*
    448  * uvm_km_alloc: allocate an area of kernel memory.
    449  *
    450  * => NOTE: we can return 0 even if we can wait if there is not enough
    451  *	free VM space in the map... caller should be prepared to handle
    452  *	this case.
    453  * => we return KVA of memory allocated
    454  */
    455 
    456 vaddr_t
    457 uvm_km_alloc(struct vm_map *map, vsize_t size, vsize_t align, uvm_flag_t flags)
    458 {
    459 	vaddr_t kva, loopva;
    460 	vaddr_t offset;
    461 	vsize_t loopsize;
    462 	struct vm_page *pg;
    463 	struct uvm_object *obj;
    464 	int pgaflags;
    465 	vm_prot_t prot;
    466 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
    467 
    468 	KASSERT(vm_map_pmap(map) == pmap_kernel());
    469 	KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED ||
    470 		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE ||
    471 		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY);
    472 	KASSERT((flags & UVM_KMF_VAONLY) != 0 || (flags & UVM_KMF_COLORMATCH) == 0);
    473 	KASSERT((flags & UVM_KMF_COLORMATCH) == 0 || (flags & UVM_KMF_VAONLY) != 0);
    474 
    475 	/*
    476 	 * setup for call
    477 	 */
    478 
    479 	kva = vm_map_min(map);	/* hint */
    480 	size = round_page(size);
    481 	obj = (flags & UVM_KMF_PAGEABLE) ? uvm_kernel_object : NULL;
    482 	UVMHIST_LOG(maphist,"  (map=0x%x, obj=0x%x, size=0x%x, flags=%d)",
    483 		    map, obj, size, flags);
    484 
    485 	/*
    486 	 * allocate some virtual space
    487 	 */
    488 
    489 	if (__predict_false(uvm_map(map, &kva, size, obj, UVM_UNKNOWN_OFFSET,
    490 	    align, UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
    491 	    UVM_ADV_RANDOM,
    492 	    (flags & (UVM_KMF_TRYLOCK | UVM_KMF_NOWAIT | UVM_KMF_WAITVA
    493 	     | UVM_KMF_COLORMATCH)))) != 0)) {
    494 		UVMHIST_LOG(maphist, "<- done (no VM)",0,0,0,0);
    495 		return(0);
    496 	}
    497 
    498 	/*
    499 	 * if all we wanted was VA, return now
    500 	 */
    501 
    502 	if (flags & (UVM_KMF_VAONLY | UVM_KMF_PAGEABLE)) {
    503 		UVMHIST_LOG(maphist,"<- done valloc (kva=0x%x)", kva,0,0,0);
    504 		return(kva);
    505 	}
    506 
    507 	/*
    508 	 * recover object offset from virtual address
    509 	 */
    510 
    511 	offset = kva - vm_map_min(kernel_map);
    512 	UVMHIST_LOG(maphist, "  kva=0x%x, offset=0x%x", kva, offset,0,0);
    513 
    514 	/*
    515 	 * now allocate and map in the memory... note that we are the only ones
    516 	 * whom should ever get a handle on this area of VM.
    517 	 */
    518 
    519 	loopva = kva;
    520 	loopsize = size;
    521 
    522 	pgaflags = UVM_FLAG_COLORMATCH;
    523 	if (flags & UVM_KMF_NOWAIT)
    524 		pgaflags |= UVM_PGA_USERESERVE;
    525 	if (flags & UVM_KMF_ZERO)
    526 		pgaflags |= UVM_PGA_ZERO;
    527 	prot = VM_PROT_READ | VM_PROT_WRITE;
    528 	if (flags & UVM_KMF_EXEC)
    529 		prot |= VM_PROT_EXECUTE;
    530 	while (loopsize) {
    531 		KASSERT(!pmap_extract(pmap_kernel(), loopva, NULL));
    532 
    533 		pg = uvm_pagealloc_strat(NULL, offset, NULL, pgaflags,
    534 #ifdef UVM_KM_VMFREELIST
    535 		   UVM_PGA_STRAT_ONLY, UVM_KM_VMFREELIST
    536 #else
    537 		   UVM_PGA_STRAT_NORMAL, 0
    538 #endif
    539 		   );
    540 
    541 		/*
    542 		 * out of memory?
    543 		 */
    544 
    545 		if (__predict_false(pg == NULL)) {
    546 			if ((flags & UVM_KMF_NOWAIT) ||
    547 			    ((flags & UVM_KMF_CANFAIL) && !uvm_reclaimable())) {
    548 				/* free everything! */
    549 				uvm_km_free(map, kva, size,
    550 				    flags & UVM_KMF_TYPEMASK);
    551 				return (0);
    552 			} else {
    553 				uvm_wait("km_getwait2");	/* sleep here */
    554 				continue;
    555 			}
    556 		}
    557 
    558 		pg->flags &= ~PG_BUSY;	/* new page */
    559 		UVM_PAGE_OWN(pg, NULL);
    560 
    561 		/*
    562 		 * map it in
    563 		 */
    564 
    565 		pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg),
    566 		    prot, PMAP_KMPAGE);
    567 		loopva += PAGE_SIZE;
    568 		offset += PAGE_SIZE;
    569 		loopsize -= PAGE_SIZE;
    570 	}
    571 
    572 	pmap_update(pmap_kernel());
    573 
    574 	UVMHIST_LOG(maphist,"<- done (kva=0x%x)", kva,0,0,0);
    575 	return(kva);
    576 }
    577 
    578 /*
    579  * uvm_km_free: free an area of kernel memory
    580  */
    581 
    582 void
    583 uvm_km_free(struct vm_map *map, vaddr_t addr, vsize_t size, uvm_flag_t flags)
    584 {
    585 
    586 	KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED ||
    587 		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE ||
    588 		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY);
    589 	KASSERT((addr & PAGE_MASK) == 0);
    590 	KASSERT(vm_map_pmap(map) == pmap_kernel());
    591 
    592 	size = round_page(size);
    593 
    594 	if (flags & UVM_KMF_PAGEABLE) {
    595 		uvm_km_pgremove(addr, addr + size);
    596 	} else if (flags & UVM_KMF_WIRED) {
    597 		/*
    598 		 * Note: uvm_km_pgremove_intrsafe() extracts mapping, thus
    599 		 * remove it after.  See comment below about KVA visibility.
    600 		 */
    601 		uvm_km_pgremove_intrsafe(map, addr, addr + size);
    602 		pmap_kremove(addr, size);
    603 	}
    604 
    605 	/*
    606 	 * Note: uvm_unmap_remove() calls pmap_update() for us, before
    607 	 * KVA becomes globally available.
    608 	 */
    609 
    610 	uvm_unmap1(map, addr, addr + size, UVM_FLAG_VAONLY);
    611 }
    612 
    613 /* Sanity; must specify both or none. */
    614 #if (defined(PMAP_MAP_POOLPAGE) || defined(PMAP_UNMAP_POOLPAGE)) && \
    615     (!defined(PMAP_MAP_POOLPAGE) || !defined(PMAP_UNMAP_POOLPAGE))
    616 #error Must specify MAP and UNMAP together.
    617 #endif
    618 
    619 int
    620 uvm_km_kmem_alloc(vmem_t *vm, vmem_size_t size, vm_flag_t flags,
    621     vmem_addr_t *addr)
    622 {
    623 	struct vm_page *pg;
    624 	vmem_addr_t va;
    625 	int rc;
    626 	vaddr_t loopva;
    627 	vsize_t loopsize;
    628 
    629 	size = round_page(size);
    630 
    631 #if defined(PMAP_MAP_POOLPAGE)
    632 	if (size == PAGE_SIZE) {
    633 again:
    634 #ifdef PMAP_ALLOC_POOLPAGE
    635 		pg = PMAP_ALLOC_POOLPAGE((flags & VM_SLEEP) ?
    636 		   0 : UVM_PGA_USERESERVE);
    637 #else
    638 		pg = uvm_pagealloc(NULL, 0, NULL,
    639 		   (flags & VM_SLEEP) ? 0 : UVM_PGA_USERESERVE);
    640 #endif /* PMAP_ALLOC_POOLPAGE */
    641 		if (__predict_false(pg == NULL)) {
    642 			if (flags & VM_SLEEP) {
    643 				uvm_wait("plpg");
    644 				goto again;
    645 			}
    646 		}
    647 		va = PMAP_MAP_POOLPAGE(VM_PAGE_TO_PHYS(pg));
    648 		if (__predict_false(va == 0)) {
    649 			uvm_pagefree(pg);
    650 			return ENOMEM;
    651 		}
    652 		*addr = va;
    653 		return 0;
    654 	}
    655 #endif /* PMAP_MAP_POOLPAGE */
    656 
    657 	rc = vmem_alloc(vm, size, flags, &va);
    658 	if (rc != 0)
    659 		return rc;
    660 
    661 	loopva = va;
    662 	loopsize = size;
    663 
    664 	while (loopsize) {
    665 		KASSERT(!pmap_extract(pmap_kernel(), loopva, NULL));
    666 
    667 		pg = uvm_pagealloc(NULL, 0, NULL,
    668 		    (flags & VM_SLEEP) ? 0 : UVM_PGA_USERESERVE);
    669 		if (__predict_false(pg == NULL)) {
    670 			if (flags & VM_SLEEP) {
    671 				uvm_wait("plpg");
    672 				continue;
    673 			} else {
    674 				uvm_km_pgremove_intrsafe(kernel_map, va,
    675 				    va + size);
    676 				pmap_kremove(va, size);
    677 				vmem_free(kmem_va_arena, va, size);
    678 				return ENOMEM;
    679 			}
    680 		}
    681 
    682 		pg->flags &= ~PG_BUSY;	/* new page */
    683 		UVM_PAGE_OWN(pg, NULL);
    684 		pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg),
    685 		    VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE);
    686 
    687 		loopva += PAGE_SIZE;
    688 		loopsize -= PAGE_SIZE;
    689 	}
    690 	pmap_update(pmap_kernel());
    691 
    692 	*addr = va;
    693 
    694 	return 0;
    695 }
    696 
    697 void
    698 uvm_km_kmem_free(vmem_t *vm, vmem_addr_t addr, size_t size)
    699 {
    700 
    701 	size = round_page(size);
    702 #if defined(PMAP_UNMAP_POOLPAGE)
    703 	if (size == PAGE_SIZE) {
    704 		paddr_t pa;
    705 
    706 		pa = PMAP_UNMAP_POOLPAGE(addr);
    707 		uvm_pagefree(PHYS_TO_VM_PAGE(pa));
    708 		return;
    709 	}
    710 #endif /* PMAP_UNMAP_POOLPAGE */
    711 	uvm_km_pgremove_intrsafe(kernel_map, addr, addr + size);
    712 	pmap_kremove(addr, size);
    713 	pmap_update(pmap_kernel());
    714 
    715 	vmem_free(vm, addr, size);
    716 }
    717 
    718 bool
    719 uvm_km_va_starved_p(void)
    720 {
    721 	vmem_size_t total;
    722 	vmem_size_t free;
    723 
    724 	total = vmem_size(kmem_arena, VMEM_ALLOC|VMEM_FREE);
    725 	free = vmem_size(kmem_arena, VMEM_FREE);
    726 
    727 	return (free < (total / 10));
    728 }
    729 
    730