uvm_km.c revision 1.137
1/* $NetBSD: uvm_km.c,v 1.137 2013/01/26 15:18:01 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@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/arenas, including: 87 * kmem_arena => used for kmem/pool (memoryallocators(9)) 88 * pager_map => used to map "buf" structures into kernel space 89 * exec_map => used during exec to handle exec args 90 * etc... 91 * 92 * The kmem_arena is a "special submap", as it lives in a fixed map entry 93 * within the kernel_map and is controlled by vmem(9). 94 * 95 * the kernel allocates its private memory out of special uvm_objects whose 96 * reference count is set to UVM_OBJ_KERN (thus indicating that the objects 97 * are "special" and never die). all kernel objects should be thought of 98 * as large, fixed-sized, sparsely populated uvm_objects. each kernel 99 * object is equal to the size of kernel virtual address space (i.e. the 100 * value "VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS"). 101 * 102 * note that just because a kernel object spans the entire kernel virtual 103 * address space doesn't mean that it has to be mapped into the entire space. 104 * large chunks of a kernel object's space go unused either because 105 * that area of kernel VM is unmapped, or there is some other type of 106 * object mapped into that range (e.g. a vnode). for submap's kernel 107 * objects, the only part of the object that can ever be populated is the 108 * offsets that are managed by the submap. 109 * 110 * note that the "offset" in a kernel object is always the kernel virtual 111 * address minus the VM_MIN_KERNEL_ADDRESS (aka vm_map_min(kernel_map)). 112 * example: 113 * suppose VM_MIN_KERNEL_ADDRESS is 0xf8000000 and the kernel does a 114 * uvm_km_alloc(kernel_map, PAGE_SIZE) [allocate 1 wired down page in the 115 * kernel map]. if uvm_km_alloc returns virtual address 0xf8235000, 116 * then that means that the page at offset 0x235000 in kernel_object is 117 * mapped at 0xf8235000. 118 * 119 * kernel object have one other special property: when the kernel virtual 120 * memory mapping them is unmapped, the backing memory in the object is 121 * freed right away. this is done with the uvm_km_pgremove() function. 122 * this has to be done because there is no backing store for kernel pages 123 * and no need to save them after they are no longer referenced. 124 * 125 * Generic arenas: 126 * 127 * kmem_arena: 128 * Main arena controlling the kernel KVA used by other arenas. 129 * 130 * kmem_va_arena: 131 * Implements quantum caching in order to speedup allocations and 132 * reduce fragmentation. The pool(9), unless created with a custom 133 * meta-data allocator, and kmem(9) subsystems use this arena. 134 * 135 * Arenas for meta-data allocations are used by vmem(9) and pool(9). 136 * These arenas cannot use quantum cache. However, kmem_va_meta_arena 137 * compensates this by importing larger chunks from kmem_arena. 138 * 139 * kmem_va_meta_arena: 140 * Space for meta-data. 141 * 142 * kmem_meta_arena: 143 * Imports from kmem_va_meta_arena. Allocations from this arena are 144 * backed with the pages. 145 * 146 * Arena stacking: 147 * 148 * kmem_arena 149 * kmem_va_arena 150 * kmem_va_meta_arena 151 * kmem_meta_arena 152 */ 153 154#include <sys/cdefs.h> 155__KERNEL_RCSID(0, "$NetBSD: uvm_km.c,v 1.137 2013/01/26 15:18:01 para Exp $"); 156 157#include "opt_uvmhist.h" 158 159#include "opt_kmempages.h" 160 161#ifndef NKMEMPAGES 162#define NKMEMPAGES 0 163#endif 164 165/* 166 * Defaults for lower and upper-bounds for the kmem_arena page count. 167 * Can be overridden by kernel config options. 168 */ 169#ifndef NKMEMPAGES_MIN 170#define NKMEMPAGES_MIN NKMEMPAGES_MIN_DEFAULT 171#endif 172 173#ifndef NKMEMPAGES_MAX 174#define NKMEMPAGES_MAX NKMEMPAGES_MAX_DEFAULT 175#endif 176 177 178#include <sys/param.h> 179#include <sys/systm.h> 180#include <sys/proc.h> 181#include <sys/pool.h> 182#include <sys/vmem.h> 183#include <sys/kmem.h> 184 185#include <uvm/uvm.h> 186 187/* 188 * global data structures 189 */ 190 191struct vm_map *kernel_map = NULL; 192 193/* 194 * local data structues 195 */ 196 197static struct vm_map kernel_map_store; 198static struct vm_map_entry kernel_image_mapent_store; 199static struct vm_map_entry kernel_kmem_mapent_store; 200 201int nkmempages = 0; 202vaddr_t kmembase; 203vsize_t kmemsize; 204 205vmem_t *kmem_arena = NULL; 206vmem_t *kmem_va_arena; 207 208/* 209 * kmeminit_nkmempages: calculate the size of kmem_arena. 210 */ 211void 212kmeminit_nkmempages(void) 213{ 214 int npages; 215 216 if (nkmempages != 0) { 217 /* 218 * It's already been set (by us being here before) 219 * bail out now; 220 */ 221 return; 222 } 223 224#if defined(PMAP_MAP_POOLPAGE) 225 npages = (physmem / 4); 226#else 227 npages = (physmem / 3) * 2; 228#endif /* defined(PMAP_MAP_POOLPAGE) */ 229 230#ifndef NKMEMPAGES_MAX_UNLIMITED 231 if (npages > NKMEMPAGES_MAX) 232 npages = NKMEMPAGES_MAX; 233#endif 234 235 if (npages < NKMEMPAGES_MIN) 236 npages = NKMEMPAGES_MIN; 237 238 nkmempages = npages; 239} 240 241/* 242 * uvm_km_bootstrap: init kernel maps and objects to reflect reality (i.e. 243 * KVM already allocated for text, data, bss, and static data structures). 244 * 245 * => KVM is defined by VM_MIN_KERNEL_ADDRESS/VM_MAX_KERNEL_ADDRESS. 246 * we assume that [vmin -> start] has already been allocated and that 247 * "end" is the end. 248 */ 249 250void 251uvm_km_bootstrap(vaddr_t start, vaddr_t end) 252{ 253 bool kmem_arena_small; 254 vaddr_t base = VM_MIN_KERNEL_ADDRESS; 255 struct uvm_map_args args; 256 int error; 257 258 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 259 UVMHIST_LOG(maphist, "start=%"PRIxVADDR" end=%#"PRIxVADDR, 260 start, end, 0,0); 261 262 kmeminit_nkmempages(); 263 kmemsize = (vsize_t)nkmempages * PAGE_SIZE; 264 kmem_arena_small = kmemsize < 64 * 1024 * 1024; 265 266 UVMHIST_LOG(maphist, "kmemsize=%#"PRIxVSIZE, kmemsize, 0,0,0); 267 268 /* 269 * next, init kernel memory objects. 270 */ 271 272 /* kernel_object: for pageable anonymous kernel memory */ 273 uvm_kernel_object = uao_create(VM_MAX_KERNEL_ADDRESS - 274 VM_MIN_KERNEL_ADDRESS, UAO_FLAG_KERNOBJ); 275 276 /* 277 * init the map and reserve any space that might already 278 * have been allocated kernel space before installing. 279 */ 280 281 uvm_map_setup(&kernel_map_store, base, end, VM_MAP_PAGEABLE); 282 kernel_map_store.pmap = pmap_kernel(); 283 if (start != base) { 284 error = uvm_map_prepare(&kernel_map_store, 285 base, start - base, 286 NULL, UVM_UNKNOWN_OFFSET, 0, 287 UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE, 288 UVM_ADV_RANDOM, UVM_FLAG_FIXED), &args); 289 if (!error) { 290 kernel_image_mapent_store.flags = 291 UVM_MAP_KERNEL | UVM_MAP_STATIC | UVM_MAP_NOMERGE; 292 error = uvm_map_enter(&kernel_map_store, &args, 293 &kernel_image_mapent_store); 294 } 295 296 if (error) 297 panic( 298 "uvm_km_bootstrap: could not reserve space for kernel"); 299 300 kmembase = args.uma_start + args.uma_size; 301 } else { 302 kmembase = base; 303 } 304 305 error = uvm_map_prepare(&kernel_map_store, 306 kmembase, kmemsize, 307 NULL, UVM_UNKNOWN_OFFSET, 0, 308 UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE, 309 UVM_ADV_RANDOM, UVM_FLAG_FIXED), &args); 310 if (!error) { 311 kernel_kmem_mapent_store.flags = 312 UVM_MAP_KERNEL | UVM_MAP_STATIC | UVM_MAP_NOMERGE; 313 error = uvm_map_enter(&kernel_map_store, &args, 314 &kernel_kmem_mapent_store); 315 } 316 317 if (error) 318 panic("uvm_km_bootstrap: could not reserve kernel kmem"); 319 320 /* 321 * install! 322 */ 323 324 kernel_map = &kernel_map_store; 325 326 pool_subsystem_init(); 327 vmem_bootstrap(); 328 329 kmem_arena = vmem_create("kmem", kmembase, kmemsize, PAGE_SIZE, 330 NULL, NULL, NULL, 331 0, VM_NOSLEEP | VM_BOOTSTRAP, IPL_VM); 332#ifdef PMAP_GROWKERNEL 333 /* 334 * kmem_arena VA allocations happen independently of uvm_map. 335 * grow kernel to accommodate the kmem_arena. 336 */ 337 if (uvm_maxkaddr < kmembase + kmemsize) { 338 uvm_maxkaddr = pmap_growkernel(kmembase + kmemsize); 339 KASSERTMSG(uvm_maxkaddr >= kmembase + kmemsize, 340 "%#"PRIxVADDR" %#"PRIxVADDR" %#"PRIxVSIZE, 341 uvm_maxkaddr, kmembase, kmemsize); 342 } 343#endif 344 345 vmem_init(kmem_arena); 346 347 UVMHIST_LOG(maphist, "kmem vmem created (base=%#"PRIxVADDR 348 ", size=%#"PRIxVSIZE, kmembase, kmemsize, 0,0); 349 350 kmem_va_arena = vmem_create("kva", 0, 0, PAGE_SIZE, 351 vmem_alloc, vmem_free, kmem_arena, 352 (kmem_arena_small ? 4 : 8) * PAGE_SIZE, 353 VM_NOSLEEP | VM_BOOTSTRAP, IPL_VM); 354 355 UVMHIST_LOG(maphist, "<- done", 0,0,0,0); 356} 357 358/* 359 * uvm_km_init: init the kernel maps virtual memory caches 360 * and start the pool/kmem allocator. 361 */ 362void 363uvm_km_init(void) 364{ 365 366 kmem_init(); 367 368 kmeminit(); // killme 369} 370 371/* 372 * uvm_km_suballoc: allocate a submap in the kernel map. once a submap 373 * is allocated all references to that area of VM must go through it. this 374 * allows the locking of VAs in kernel_map to be broken up into regions. 375 * 376 * => if `fixed' is true, *vmin specifies where the region described 377 * pager_map => used to map "buf" structures into kernel space 378 * by the submap must start 379 * => if submap is non NULL we use that as the submap, otherwise we 380 * alloc a new map 381 */ 382 383struct vm_map * 384uvm_km_suballoc(struct vm_map *map, vaddr_t *vmin /* IN/OUT */, 385 vaddr_t *vmax /* OUT */, vsize_t size, int flags, bool fixed, 386 struct vm_map *submap) 387{ 388 int mapflags = UVM_FLAG_NOMERGE | (fixed ? UVM_FLAG_FIXED : 0); 389 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 390 391 KASSERT(vm_map_pmap(map) == pmap_kernel()); 392 393 size = round_page(size); /* round up to pagesize */ 394 395 /* 396 * first allocate a blank spot in the parent map 397 */ 398 399 if (uvm_map(map, vmin, size, NULL, UVM_UNKNOWN_OFFSET, 0, 400 UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE, 401 UVM_ADV_RANDOM, mapflags)) != 0) { 402 panic("%s: unable to allocate space in parent map", __func__); 403 } 404 405 /* 406 * set VM bounds (vmin is filled in by uvm_map) 407 */ 408 409 *vmax = *vmin + size; 410 411 /* 412 * add references to pmap and create or init the submap 413 */ 414 415 pmap_reference(vm_map_pmap(map)); 416 if (submap == NULL) { 417 submap = kmem_alloc(sizeof(*submap), KM_SLEEP); 418 if (submap == NULL) 419 panic("uvm_km_suballoc: unable to create submap"); 420 } 421 uvm_map_setup(submap, *vmin, *vmax, flags); 422 submap->pmap = vm_map_pmap(map); 423 424 /* 425 * now let uvm_map_submap plug in it... 426 */ 427 428 if (uvm_map_submap(map, *vmin, *vmax, submap) != 0) 429 panic("uvm_km_suballoc: submap allocation failed"); 430 431 return(submap); 432} 433 434/* 435 * uvm_km_pgremove: remove pages from a kernel uvm_object and KVA. 436 */ 437 438void 439uvm_km_pgremove(vaddr_t startva, vaddr_t endva) 440{ 441 struct uvm_object * const uobj = uvm_kernel_object; 442 const voff_t start = startva - vm_map_min(kernel_map); 443 const voff_t end = endva - vm_map_min(kernel_map); 444 struct vm_page *pg; 445 voff_t curoff, nextoff; 446 int swpgonlydelta = 0; 447 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 448 449 KASSERT(VM_MIN_KERNEL_ADDRESS <= startva); 450 KASSERT(startva < endva); 451 KASSERT(endva <= VM_MAX_KERNEL_ADDRESS); 452 453 mutex_enter(uobj->vmobjlock); 454 pmap_remove(pmap_kernel(), startva, endva); 455 for (curoff = start; curoff < end; curoff = nextoff) { 456 nextoff = curoff + PAGE_SIZE; 457 pg = uvm_pagelookup(uobj, curoff); 458 if (pg != NULL && pg->flags & PG_BUSY) { 459 pg->flags |= PG_WANTED; 460 UVM_UNLOCK_AND_WAIT(pg, uobj->vmobjlock, 0, 461 "km_pgrm", 0); 462 mutex_enter(uobj->vmobjlock); 463 nextoff = curoff; 464 continue; 465 } 466 467 /* 468 * free the swap slot, then the page. 469 */ 470 471 if (pg == NULL && 472 uao_find_swslot(uobj, curoff >> PAGE_SHIFT) > 0) { 473 swpgonlydelta++; 474 } 475 uao_dropswap(uobj, curoff >> PAGE_SHIFT); 476 if (pg != NULL) { 477 mutex_enter(&uvm_pageqlock); 478 uvm_pagefree(pg); 479 mutex_exit(&uvm_pageqlock); 480 } 481 } 482 mutex_exit(uobj->vmobjlock); 483 484 if (swpgonlydelta > 0) { 485 mutex_enter(&uvm_swap_data_lock); 486 KASSERT(uvmexp.swpgonly >= swpgonlydelta); 487 uvmexp.swpgonly -= swpgonlydelta; 488 mutex_exit(&uvm_swap_data_lock); 489 } 490} 491 492 493/* 494 * uvm_km_pgremove_intrsafe: like uvm_km_pgremove(), but for non object backed 495 * regions. 496 * 497 * => when you unmap a part of anonymous kernel memory you want to toss 498 * the pages right away. (this is called from uvm_unmap_...). 499 * => none of the pages will ever be busy, and none of them will ever 500 * be on the active or inactive queues (because they have no object). 501 */ 502 503void 504uvm_km_pgremove_intrsafe(struct vm_map *map, vaddr_t start, vaddr_t end) 505{ 506#define __PGRM_BATCH 16 507 struct vm_page *pg; 508 paddr_t pa[__PGRM_BATCH]; 509 int npgrm, i; 510 vaddr_t va, batch_vastart; 511 512 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 513 514 KASSERT(VM_MAP_IS_KERNEL(map)); 515 KASSERTMSG(vm_map_min(map) <= start, 516 "vm_map_min(map) [%#"PRIxVADDR"] <= start [%#"PRIxVADDR"]" 517 " (size=%#"PRIxVSIZE")", 518 vm_map_min(map), start, end - start); 519 KASSERT(start < end); 520 KASSERT(end <= vm_map_max(map)); 521 522 for (va = start; va < end;) { 523 batch_vastart = va; 524 /* create a batch of at most __PGRM_BATCH pages to free */ 525 for (i = 0; 526 i < __PGRM_BATCH && va < end; 527 va += PAGE_SIZE) { 528 if (!pmap_extract(pmap_kernel(), va, &pa[i])) { 529 continue; 530 } 531 i++; 532 } 533 npgrm = i; 534 /* now remove the mappings */ 535 pmap_kremove(batch_vastart, va - batch_vastart); 536 /* and free the pages */ 537 for (i = 0; i < npgrm; i++) { 538 pg = PHYS_TO_VM_PAGE(pa[i]); 539 KASSERT(pg); 540 KASSERT(pg->uobject == NULL && pg->uanon == NULL); 541 KASSERT((pg->flags & PG_BUSY) == 0); 542 uvm_pagefree(pg); 543 } 544 } 545#undef __PGRM_BATCH 546} 547 548#if defined(DEBUG) 549void 550uvm_km_check_empty(struct vm_map *map, vaddr_t start, vaddr_t end) 551{ 552 struct vm_page *pg; 553 vaddr_t va; 554 paddr_t pa; 555 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 556 557 KDASSERT(VM_MAP_IS_KERNEL(map)); 558 KDASSERT(vm_map_min(map) <= start); 559 KDASSERT(start < end); 560 KDASSERT(end <= vm_map_max(map)); 561 562 for (va = start; va < end; va += PAGE_SIZE) { 563 if (pmap_extract(pmap_kernel(), va, &pa)) { 564 panic("uvm_km_check_empty: va %p has pa 0x%llx", 565 (void *)va, (long long)pa); 566 } 567 mutex_enter(uvm_kernel_object->vmobjlock); 568 pg = uvm_pagelookup(uvm_kernel_object, 569 va - vm_map_min(kernel_map)); 570 mutex_exit(uvm_kernel_object->vmobjlock); 571 if (pg) { 572 panic("uvm_km_check_empty: " 573 "has page hashed at %p", (const void *)va); 574 } 575 } 576} 577#endif /* defined(DEBUG) */ 578 579/* 580 * uvm_km_alloc: allocate an area of kernel memory. 581 * 582 * => NOTE: we can return 0 even if we can wait if there is not enough 583 * free VM space in the map... caller should be prepared to handle 584 * this case. 585 * => we return KVA of memory allocated 586 */ 587 588vaddr_t 589uvm_km_alloc(struct vm_map *map, vsize_t size, vsize_t align, uvm_flag_t flags) 590{ 591 vaddr_t kva, loopva; 592 vaddr_t offset; 593 vsize_t loopsize; 594 struct vm_page *pg; 595 struct uvm_object *obj; 596 int pgaflags; 597 vm_prot_t prot; 598 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 599 600 KASSERT(vm_map_pmap(map) == pmap_kernel()); 601 KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED || 602 (flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE || 603 (flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY); 604 KASSERT((flags & UVM_KMF_VAONLY) != 0 || (flags & UVM_KMF_COLORMATCH) == 0); 605 KASSERT((flags & UVM_KMF_COLORMATCH) == 0 || (flags & UVM_KMF_VAONLY) != 0); 606 607 /* 608 * setup for call 609 */ 610 611 kva = vm_map_min(map); /* hint */ 612 size = round_page(size); 613 obj = (flags & UVM_KMF_PAGEABLE) ? uvm_kernel_object : NULL; 614 UVMHIST_LOG(maphist," (map=0x%x, obj=0x%x, size=0x%x, flags=%d)", 615 map, obj, size, flags); 616 617 /* 618 * allocate some virtual space 619 */ 620 621 if (__predict_false(uvm_map(map, &kva, size, obj, UVM_UNKNOWN_OFFSET, 622 align, UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE, 623 UVM_ADV_RANDOM, 624 (flags & (UVM_KMF_TRYLOCK | UVM_KMF_NOWAIT | UVM_KMF_WAITVA 625 | UVM_KMF_COLORMATCH)))) != 0)) { 626 UVMHIST_LOG(maphist, "<- done (no VM)",0,0,0,0); 627 return(0); 628 } 629 630 /* 631 * if all we wanted was VA, return now 632 */ 633 634 if (flags & (UVM_KMF_VAONLY | UVM_KMF_PAGEABLE)) { 635 UVMHIST_LOG(maphist,"<- done valloc (kva=0x%x)", kva,0,0,0); 636 return(kva); 637 } 638 639 /* 640 * recover object offset from virtual address 641 */ 642 643 offset = kva - vm_map_min(kernel_map); 644 UVMHIST_LOG(maphist, " kva=0x%x, offset=0x%x", kva, offset,0,0); 645 646 /* 647 * now allocate and map in the memory... note that we are the only ones 648 * whom should ever get a handle on this area of VM. 649 */ 650 651 loopva = kva; 652 loopsize = size; 653 654 pgaflags = UVM_FLAG_COLORMATCH; 655 if (flags & UVM_KMF_NOWAIT) 656 pgaflags |= UVM_PGA_USERESERVE; 657 if (flags & UVM_KMF_ZERO) 658 pgaflags |= UVM_PGA_ZERO; 659 prot = VM_PROT_READ | VM_PROT_WRITE; 660 if (flags & UVM_KMF_EXEC) 661 prot |= VM_PROT_EXECUTE; 662 while (loopsize) { 663 KASSERTMSG(!pmap_extract(pmap_kernel(), loopva, NULL), 664 "loopva=%#"PRIxVADDR, loopva); 665 666 pg = uvm_pagealloc_strat(NULL, offset, NULL, pgaflags, 667#ifdef UVM_KM_VMFREELIST 668 UVM_PGA_STRAT_ONLY, UVM_KM_VMFREELIST 669#else 670 UVM_PGA_STRAT_NORMAL, 0 671#endif 672 ); 673 674 /* 675 * out of memory? 676 */ 677 678 if (__predict_false(pg == NULL)) { 679 if ((flags & UVM_KMF_NOWAIT) || 680 ((flags & UVM_KMF_CANFAIL) && !uvm_reclaimable())) { 681 /* free everything! */ 682 uvm_km_free(map, kva, size, 683 flags & UVM_KMF_TYPEMASK); 684 return (0); 685 } else { 686 uvm_wait("km_getwait2"); /* sleep here */ 687 continue; 688 } 689 } 690 691 pg->flags &= ~PG_BUSY; /* new page */ 692 UVM_PAGE_OWN(pg, NULL); 693 694 /* 695 * map it in 696 */ 697 698 pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg), 699 prot, PMAP_KMPAGE); 700 loopva += PAGE_SIZE; 701 offset += PAGE_SIZE; 702 loopsize -= PAGE_SIZE; 703 } 704 705 pmap_update(pmap_kernel()); 706 707 UVMHIST_LOG(maphist,"<- done (kva=0x%x)", kva,0,0,0); 708 return(kva); 709} 710 711/* 712 * uvm_km_free: free an area of kernel memory 713 */ 714 715void 716uvm_km_free(struct vm_map *map, vaddr_t addr, vsize_t size, uvm_flag_t flags) 717{ 718 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 719 720 KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED || 721 (flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE || 722 (flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY); 723 KASSERT((addr & PAGE_MASK) == 0); 724 KASSERT(vm_map_pmap(map) == pmap_kernel()); 725 726 size = round_page(size); 727 728 if (flags & UVM_KMF_PAGEABLE) { 729 uvm_km_pgremove(addr, addr + size); 730 } else if (flags & UVM_KMF_WIRED) { 731 /* 732 * Note: uvm_km_pgremove_intrsafe() extracts mapping, thus 733 * remove it after. See comment below about KVA visibility. 734 */ 735 uvm_km_pgremove_intrsafe(map, addr, addr + size); 736 } 737 738 /* 739 * Note: uvm_unmap_remove() calls pmap_update() for us, before 740 * KVA becomes globally available. 741 */ 742 743 uvm_unmap1(map, addr, addr + size, UVM_FLAG_VAONLY); 744} 745 746/* Sanity; must specify both or none. */ 747#if (defined(PMAP_MAP_POOLPAGE) || defined(PMAP_UNMAP_POOLPAGE)) && \ 748 (!defined(PMAP_MAP_POOLPAGE) || !defined(PMAP_UNMAP_POOLPAGE)) 749#error Must specify MAP and UNMAP together. 750#endif 751 752int 753uvm_km_kmem_alloc(vmem_t *vm, vmem_size_t size, vm_flag_t flags, 754 vmem_addr_t *addr) 755{ 756 struct vm_page *pg; 757 vmem_addr_t va; 758 int rc; 759 vaddr_t loopva; 760 vsize_t loopsize; 761 762 size = round_page(size); 763 764#if defined(PMAP_MAP_POOLPAGE) 765 if (size == PAGE_SIZE) { 766again: 767#ifdef PMAP_ALLOC_POOLPAGE 768 pg = PMAP_ALLOC_POOLPAGE((flags & VM_SLEEP) ? 769 0 : UVM_PGA_USERESERVE); 770#else 771 pg = uvm_pagealloc(NULL, 0, NULL, 772 (flags & VM_SLEEP) ? 0 : UVM_PGA_USERESERVE); 773#endif /* PMAP_ALLOC_POOLPAGE */ 774 if (__predict_false(pg == NULL)) { 775 if (flags & VM_SLEEP) { 776 uvm_wait("plpg"); 777 goto again; 778 } 779 return ENOMEM; 780 } 781 va = PMAP_MAP_POOLPAGE(VM_PAGE_TO_PHYS(pg)); 782 if (__predict_false(va == 0)) { 783 uvm_pagefree(pg); 784 return ENOMEM; 785 } 786 *addr = va; 787 return 0; 788 } 789#endif /* PMAP_MAP_POOLPAGE */ 790 791 rc = vmem_alloc(vm, size, flags, &va); 792 if (rc != 0) 793 return rc; 794 795#ifdef PMAP_GROWKERNEL 796 /* 797 * These VA allocations happen independently of uvm_map 798 * so this allocation must not extend beyond the current limit. 799 */ 800 KASSERTMSG(uvm_maxkaddr >= va + size, 801 "%#"PRIxVADDR" %#"PRIxPTR" %#zx", 802 uvm_maxkaddr, va, size); 803#endif 804 805 loopva = va; 806 loopsize = size; 807 808 while (loopsize) { 809#ifdef DIAGNOSTIC 810 paddr_t pa; 811#endif 812 KASSERTMSG(!pmap_extract(pmap_kernel(), loopva, &pa), 813 "loopva=%#"PRIxVADDR" loopsize=%#"PRIxVSIZE 814 " pa=%#"PRIxPADDR" vmem=%p", 815 loopva, loopsize, pa, vm); 816 817 pg = uvm_pagealloc(NULL, loopva, NULL, 818 UVM_FLAG_COLORMATCH 819 | ((flags & VM_SLEEP) ? 0 : UVM_PGA_USERESERVE)); 820 if (__predict_false(pg == NULL)) { 821 if (flags & VM_SLEEP) { 822 uvm_wait("plpg"); 823 continue; 824 } else { 825 uvm_km_pgremove_intrsafe(kernel_map, va, 826 va + size); 827 vmem_free(vm, va, size); 828 return ENOMEM; 829 } 830 } 831 832 pg->flags &= ~PG_BUSY; /* new page */ 833 UVM_PAGE_OWN(pg, NULL); 834 pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg), 835 VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE); 836 837 loopva += PAGE_SIZE; 838 loopsize -= PAGE_SIZE; 839 } 840 pmap_update(pmap_kernel()); 841 842 *addr = va; 843 844 return 0; 845} 846 847void 848uvm_km_kmem_free(vmem_t *vm, vmem_addr_t addr, size_t size) 849{ 850 851 size = round_page(size); 852#if defined(PMAP_UNMAP_POOLPAGE) 853 if (size == PAGE_SIZE) { 854 paddr_t pa; 855 856 pa = PMAP_UNMAP_POOLPAGE(addr); 857 uvm_pagefree(PHYS_TO_VM_PAGE(pa)); 858 return; 859 } 860#endif /* PMAP_UNMAP_POOLPAGE */ 861 uvm_km_pgremove_intrsafe(kernel_map, addr, addr + size); 862 pmap_update(pmap_kernel()); 863 864 vmem_free(vm, addr, size); 865} 866 867bool 868uvm_km_va_starved_p(void) 869{ 870 vmem_size_t total; 871 vmem_size_t free; 872 873 if (kmem_arena == NULL) 874 return false; 875 876 total = vmem_size(kmem_arena, VMEM_ALLOC|VMEM_FREE); 877 free = vmem_size(kmem_arena, VMEM_FREE); 878 879 return (free < (total / 10)); 880} 881