1 /* $NetBSD: uvm_km.c,v 1.138 2013/01/29 21:29:40 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 1.138 2013/01/29 21:29:40 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/vmem_impl.h> 184 #include <sys/kmem.h> 185 186 #include <uvm/uvm.h> 187 188 /* 189 * global data structures 190 */ 191 192 struct vm_map *kernel_map = NULL; 193 194 /* 195 * local data structues 196 */ 197 198 static struct vm_map kernel_map_store; 199 static struct vm_map_entry kernel_image_mapent_store; 200 static struct vm_map_entry kernel_kmem_mapent_store; 201 202 int nkmempages = 0; 203 vaddr_t kmembase; 204 vsize_t kmemsize; 205 206 static struct vmem kmem_arena_store; 207 vmem_t *kmem_arena = NULL; 208 static struct vmem kmem_va_arena_store; 209 vmem_t *kmem_va_arena; 210 211 /* 212 * kmeminit_nkmempages: calculate the size of kmem_arena. 213 */ 214 void 215 kmeminit_nkmempages(void) 216 { 217 int npages; 218 219 if (nkmempages != 0) { 220 /* 221 * It's already been set (by us being here before) 222 * bail out now; 223 */ 224 return; 225 } 226 227 #if defined(PMAP_MAP_POOLPAGE) 228 npages = (physmem / 4); 229 #else 230 npages = (physmem / 3) * 2; 231 #endif /* defined(PMAP_MAP_POOLPAGE) */ 232 233 #ifndef NKMEMPAGES_MAX_UNLIMITED 234 if (npages > NKMEMPAGES_MAX) 235 npages = NKMEMPAGES_MAX; 236 #endif 237 238 if (npages < NKMEMPAGES_MIN) 239 npages = NKMEMPAGES_MIN; 240 241 nkmempages = npages; 242 } 243 244 /* 245 * uvm_km_bootstrap: init kernel maps and objects to reflect reality (i.e. 246 * KVM already allocated for text, data, bss, and static data structures). 247 * 248 * => KVM is defined by VM_MIN_KERNEL_ADDRESS/VM_MAX_KERNEL_ADDRESS. 249 * we assume that [vmin -> start] has already been allocated and that 250 * "end" is the end. 251 */ 252 253 void 254 uvm_km_bootstrap(vaddr_t start, vaddr_t end) 255 { 256 bool kmem_arena_small; 257 vaddr_t base = VM_MIN_KERNEL_ADDRESS; 258 struct uvm_map_args args; 259 int error; 260 261 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 262 UVMHIST_LOG(maphist, "start=%"PRIxVADDR" end=%#"PRIxVADDR, 263 start, end, 0,0); 264 265 kmeminit_nkmempages(); 266 kmemsize = (vsize_t)nkmempages * PAGE_SIZE; 267 kmem_arena_small = kmemsize < 64 * 1024 * 1024; 268 269 UVMHIST_LOG(maphist, "kmemsize=%#"PRIxVSIZE, kmemsize, 0,0,0); 270 271 /* 272 * next, init kernel memory objects. 273 */ 274 275 /* kernel_object: for pageable anonymous kernel memory */ 276 uvm_kernel_object = uao_create(VM_MAX_KERNEL_ADDRESS - 277 VM_MIN_KERNEL_ADDRESS, UAO_FLAG_KERNOBJ); 278 279 /* 280 * init the map and reserve any space that might already 281 * have been allocated kernel space before installing. 282 */ 283 284 uvm_map_setup(&kernel_map_store, base, end, VM_MAP_PAGEABLE); 285 kernel_map_store.pmap = pmap_kernel(); 286 if (start != base) { 287 error = uvm_map_prepare(&kernel_map_store, 288 base, start - base, 289 NULL, UVM_UNKNOWN_OFFSET, 0, 290 UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE, 291 UVM_ADV_RANDOM, UVM_FLAG_FIXED), &args); 292 if (!error) { 293 kernel_image_mapent_store.flags = 294 UVM_MAP_KERNEL | UVM_MAP_STATIC | UVM_MAP_NOMERGE; 295 error = uvm_map_enter(&kernel_map_store, &args, 296 &kernel_image_mapent_store); 297 } 298 299 if (error) 300 panic( 301 "uvm_km_bootstrap: could not reserve space for kernel"); 302 303 kmembase = args.uma_start + args.uma_size; 304 } else { 305 kmembase = base; 306 } 307 308 error = uvm_map_prepare(&kernel_map_store, 309 kmembase, kmemsize, 310 NULL, UVM_UNKNOWN_OFFSET, 0, 311 UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE, 312 UVM_ADV_RANDOM, UVM_FLAG_FIXED), &args); 313 if (!error) { 314 kernel_kmem_mapent_store.flags = 315 UVM_MAP_KERNEL | UVM_MAP_STATIC | UVM_MAP_NOMERGE; 316 error = uvm_map_enter(&kernel_map_store, &args, 317 &kernel_kmem_mapent_store); 318 } 319 320 if (error) 321 panic("uvm_km_bootstrap: could not reserve kernel kmem"); 322 323 /* 324 * install! 325 */ 326 327 kernel_map = &kernel_map_store; 328 329 pool_subsystem_init(); 330 331 kmem_arena = vmem_init(&kmem_arena_store, "kmem", 332 kmembase, kmemsize, PAGE_SIZE, NULL, NULL, NULL, 333 0, VM_NOSLEEP | VM_BOOTSTRAP, IPL_VM); 334 #ifdef PMAP_GROWKERNEL 335 /* 336 * kmem_arena VA allocations happen independently of uvm_map. 337 * grow kernel to accommodate the kmem_arena. 338 */ 339 if (uvm_maxkaddr < kmembase + kmemsize) { 340 uvm_maxkaddr = pmap_growkernel(kmembase + kmemsize); 341 KASSERTMSG(uvm_maxkaddr >= kmembase + kmemsize, 342 "%#"PRIxVADDR" %#"PRIxVADDR" %#"PRIxVSIZE, 343 uvm_maxkaddr, kmembase, kmemsize); 344 } 345 #endif 346 347 vmem_subsystem_init(kmem_arena); 348 349 UVMHIST_LOG(maphist, "kmem vmem created (base=%#"PRIxVADDR 350 ", size=%#"PRIxVSIZE, kmembase, kmemsize, 0,0); 351 352 kmem_va_arena = vmem_init(&kmem_va_arena_store, "kva", 353 0, 0, PAGE_SIZE, vmem_alloc, vmem_free, kmem_arena, 354 (kmem_arena_small ? 4 : VMEM_QCACHE_IDX_MAX) * PAGE_SIZE, 355 VM_NOSLEEP, IPL_VM); 356 357 UVMHIST_LOG(maphist, "<- done", 0,0,0,0); 358 } 359 360 /* 361 * uvm_km_init: init the kernel maps virtual memory caches 362 * and start the pool/kmem allocator. 363 */ 364 void 365 uvm_km_init(void) 366 { 367 368 kmem_init(); 369 370 kmeminit(); // killme 371 } 372 373 /* 374 * uvm_km_suballoc: allocate a submap in the kernel map. once a submap 375 * is allocated all references to that area of VM must go through it. this 376 * allows the locking of VAs in kernel_map to be broken up into regions. 377 * 378 * => if `fixed' is true, *vmin specifies where the region described 379 * pager_map => used to map "buf" structures into kernel space 380 * by the submap must start 381 * => if submap is non NULL we use that as the submap, otherwise we 382 * alloc a new map 383 */ 384 385 struct vm_map * 386 uvm_km_suballoc(struct vm_map *map, vaddr_t *vmin /* IN/OUT */, 387 vaddr_t *vmax /* OUT */, vsize_t size, int flags, bool fixed, 388 struct vm_map *submap) 389 { 390 int mapflags = UVM_FLAG_NOMERGE | (fixed ? UVM_FLAG_FIXED : 0); 391 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 392 393 KASSERT(vm_map_pmap(map) == pmap_kernel()); 394 395 size = round_page(size); /* round up to pagesize */ 396 397 /* 398 * first allocate a blank spot in the parent map 399 */ 400 401 if (uvm_map(map, vmin, size, NULL, UVM_UNKNOWN_OFFSET, 0, 402 UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE, 403 UVM_ADV_RANDOM, mapflags)) != 0) { 404 panic("%s: unable to allocate space in parent map", __func__); 405 } 406 407 /* 408 * set VM bounds (vmin is filled in by uvm_map) 409 */ 410 411 *vmax = *vmin + size; 412 413 /* 414 * add references to pmap and create or init the submap 415 */ 416 417 pmap_reference(vm_map_pmap(map)); 418 if (submap == NULL) { 419 submap = kmem_alloc(sizeof(*submap), KM_SLEEP); 420 if (submap == NULL) 421 panic("uvm_km_suballoc: unable to create submap"); 422 } 423 uvm_map_setup(submap, *vmin, *vmax, flags); 424 submap->pmap = vm_map_pmap(map); 425 426 /* 427 * now let uvm_map_submap plug in it... 428 */ 429 430 if (uvm_map_submap(map, *vmin, *vmax, submap) != 0) 431 panic("uvm_km_suballoc: submap allocation failed"); 432 433 return(submap); 434 } 435 436 /* 437 * uvm_km_pgremove: remove pages from a kernel uvm_object and KVA. 438 */ 439 440 void 441 uvm_km_pgremove(vaddr_t startva, vaddr_t endva) 442 { 443 struct uvm_object * const uobj = uvm_kernel_object; 444 const voff_t start = startva - vm_map_min(kernel_map); 445 const voff_t end = endva - vm_map_min(kernel_map); 446 struct vm_page *pg; 447 voff_t curoff, nextoff; 448 int swpgonlydelta = 0; 449 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 450 451 KASSERT(VM_MIN_KERNEL_ADDRESS <= startva); 452 KASSERT(startva < endva); 453 KASSERT(endva <= VM_MAX_KERNEL_ADDRESS); 454 455 mutex_enter(uobj->vmobjlock); 456 pmap_remove(pmap_kernel(), startva, endva); 457 for (curoff = start; curoff < end; curoff = nextoff) { 458 nextoff = curoff + PAGE_SIZE; 459 pg = uvm_pagelookup(uobj, curoff); 460 if (pg != NULL && pg->flags & PG_BUSY) { 461 pg->flags |= PG_WANTED; 462 UVM_UNLOCK_AND_WAIT(pg, uobj->vmobjlock, 0, 463 "km_pgrm", 0); 464 mutex_enter(uobj->vmobjlock); 465 nextoff = curoff; 466 continue; 467 } 468 469 /* 470 * free the swap slot, then the page. 471 */ 472 473 if (pg == NULL && 474 uao_find_swslot(uobj, curoff >> PAGE_SHIFT) > 0) { 475 swpgonlydelta++; 476 } 477 uao_dropswap(uobj, curoff >> PAGE_SHIFT); 478 if (pg != NULL) { 479 mutex_enter(&uvm_pageqlock); 480 uvm_pagefree(pg); 481 mutex_exit(&uvm_pageqlock); 482 } 483 } 484 mutex_exit(uobj->vmobjlock); 485 486 if (swpgonlydelta > 0) { 487 mutex_enter(&uvm_swap_data_lock); 488 KASSERT(uvmexp.swpgonly >= swpgonlydelta); 489 uvmexp.swpgonly -= swpgonlydelta; 490 mutex_exit(&uvm_swap_data_lock); 491 } 492 } 493 494 495 /* 496 * uvm_km_pgremove_intrsafe: like uvm_km_pgremove(), but for non object backed 497 * regions. 498 * 499 * => when you unmap a part of anonymous kernel memory you want to toss 500 * the pages right away. (this is called from uvm_unmap_...). 501 * => none of the pages will ever be busy, and none of them will ever 502 * be on the active or inactive queues (because they have no object). 503 */ 504 505 void 506 uvm_km_pgremove_intrsafe(struct vm_map *map, vaddr_t start, vaddr_t end) 507 { 508 #define __PGRM_BATCH 16 509 struct vm_page *pg; 510 paddr_t pa[__PGRM_BATCH]; 511 int npgrm, i; 512 vaddr_t va, batch_vastart; 513 514 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 515 516 KASSERT(VM_MAP_IS_KERNEL(map)); 517 KASSERTMSG(vm_map_min(map) <= start, 518 "vm_map_min(map) [%#"PRIxVADDR"] <= start [%#"PRIxVADDR"]" 519 " (size=%#"PRIxVSIZE")", 520 vm_map_min(map), start, end - start); 521 KASSERT(start < end); 522 KASSERT(end <= vm_map_max(map)); 523 524 for (va = start; va < end;) { 525 batch_vastart = va; 526 /* create a batch of at most __PGRM_BATCH pages to free */ 527 for (i = 0; 528 i < __PGRM_BATCH && va < end; 529 va += PAGE_SIZE) { 530 if (!pmap_extract(pmap_kernel(), va, &pa[i])) { 531 continue; 532 } 533 i++; 534 } 535 npgrm = i; 536 /* now remove the mappings */ 537 pmap_kremove(batch_vastart, va - batch_vastart); 538 /* and free the pages */ 539 for (i = 0; i < npgrm; i++) { 540 pg = PHYS_TO_VM_PAGE(pa[i]); 541 KASSERT(pg); 542 KASSERT(pg->uobject == NULL && pg->uanon == NULL); 543 KASSERT((pg->flags & PG_BUSY) == 0); 544 uvm_pagefree(pg); 545 } 546 } 547 #undef __PGRM_BATCH 548 } 549 550 #if defined(DEBUG) 551 void 552 uvm_km_check_empty(struct vm_map *map, vaddr_t start, vaddr_t end) 553 { 554 struct vm_page *pg; 555 vaddr_t va; 556 paddr_t pa; 557 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 558 559 KDASSERT(VM_MAP_IS_KERNEL(map)); 560 KDASSERT(vm_map_min(map) <= start); 561 KDASSERT(start < end); 562 KDASSERT(end <= vm_map_max(map)); 563 564 for (va = start; va < end; va += PAGE_SIZE) { 565 if (pmap_extract(pmap_kernel(), va, &pa)) { 566 panic("uvm_km_check_empty: va %p has pa 0x%llx", 567 (void *)va, (long long)pa); 568 } 569 mutex_enter(uvm_kernel_object->vmobjlock); 570 pg = uvm_pagelookup(uvm_kernel_object, 571 va - vm_map_min(kernel_map)); 572 mutex_exit(uvm_kernel_object->vmobjlock); 573 if (pg) { 574 panic("uvm_km_check_empty: " 575 "has page hashed at %p", (const void *)va); 576 } 577 } 578 } 579 #endif /* defined(DEBUG) */ 580 581 /* 582 * uvm_km_alloc: allocate an area of kernel memory. 583 * 584 * => NOTE: we can return 0 even if we can wait if there is not enough 585 * free VM space in the map... caller should be prepared to handle 586 * this case. 587 * => we return KVA of memory allocated 588 */ 589 590 vaddr_t 591 uvm_km_alloc(struct vm_map *map, vsize_t size, vsize_t align, uvm_flag_t flags) 592 { 593 vaddr_t kva, loopva; 594 vaddr_t offset; 595 vsize_t loopsize; 596 struct vm_page *pg; 597 struct uvm_object *obj; 598 int pgaflags; 599 vm_prot_t prot; 600 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 601 602 KASSERT(vm_map_pmap(map) == pmap_kernel()); 603 KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED || 604 (flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE || 605 (flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY); 606 KASSERT((flags & UVM_KMF_VAONLY) != 0 || (flags & UVM_KMF_COLORMATCH) == 0); 607 KASSERT((flags & UVM_KMF_COLORMATCH) == 0 || (flags & UVM_KMF_VAONLY) != 0); 608 609 /* 610 * setup for call 611 */ 612 613 kva = vm_map_min(map); /* hint */ 614 size = round_page(size); 615 obj = (flags & UVM_KMF_PAGEABLE) ? uvm_kernel_object : NULL; 616 UVMHIST_LOG(maphist," (map=0x%x, obj=0x%x, size=0x%x, flags=%d)", 617 map, obj, size, flags); 618 619 /* 620 * allocate some virtual space 621 */ 622 623 if (__predict_false(uvm_map(map, &kva, size, obj, UVM_UNKNOWN_OFFSET, 624 align, UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE, 625 UVM_ADV_RANDOM, 626 (flags & (UVM_KMF_TRYLOCK | UVM_KMF_NOWAIT | UVM_KMF_WAITVA 627 | UVM_KMF_COLORMATCH)))) != 0)) { 628 UVMHIST_LOG(maphist, "<- done (no VM)",0,0,0,0); 629 return(0); 630 } 631 632 /* 633 * if all we wanted was VA, return now 634 */ 635 636 if (flags & (UVM_KMF_VAONLY | UVM_KMF_PAGEABLE)) { 637 UVMHIST_LOG(maphist,"<- done valloc (kva=0x%x)", kva,0,0,0); 638 return(kva); 639 } 640 641 /* 642 * recover object offset from virtual address 643 */ 644 645 offset = kva - vm_map_min(kernel_map); 646 UVMHIST_LOG(maphist, " kva=0x%x, offset=0x%x", kva, offset,0,0); 647 648 /* 649 * now allocate and map in the memory... note that we are the only ones 650 * whom should ever get a handle on this area of VM. 651 */ 652 653 loopva = kva; 654 loopsize = size; 655 656 pgaflags = UVM_FLAG_COLORMATCH; 657 if (flags & UVM_KMF_NOWAIT) 658 pgaflags |= UVM_PGA_USERESERVE; 659 if (flags & UVM_KMF_ZERO) 660 pgaflags |= UVM_PGA_ZERO; 661 prot = VM_PROT_READ | VM_PROT_WRITE; 662 if (flags & UVM_KMF_EXEC) 663 prot |= VM_PROT_EXECUTE; 664 while (loopsize) { 665 KASSERTMSG(!pmap_extract(pmap_kernel(), loopva, NULL), 666 "loopva=%#"PRIxVADDR, loopva); 667 668 pg = uvm_pagealloc_strat(NULL, offset, NULL, pgaflags, 669 #ifdef UVM_KM_VMFREELIST 670 UVM_PGA_STRAT_ONLY, UVM_KM_VMFREELIST 671 #else 672 UVM_PGA_STRAT_NORMAL, 0 673 #endif 674 ); 675 676 /* 677 * out of memory? 678 */ 679 680 if (__predict_false(pg == NULL)) { 681 if ((flags & UVM_KMF_NOWAIT) || 682 ((flags & UVM_KMF_CANFAIL) && !uvm_reclaimable())) { 683 /* free everything! */ 684 uvm_km_free(map, kva, size, 685 flags & UVM_KMF_TYPEMASK); 686 return (0); 687 } else { 688 uvm_wait("km_getwait2"); /* sleep here */ 689 continue; 690 } 691 } 692 693 pg->flags &= ~PG_BUSY; /* new page */ 694 UVM_PAGE_OWN(pg, NULL); 695 696 /* 697 * map it in 698 */ 699 700 pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg), 701 prot, PMAP_KMPAGE); 702 loopva += PAGE_SIZE; 703 offset += PAGE_SIZE; 704 loopsize -= PAGE_SIZE; 705 } 706 707 pmap_update(pmap_kernel()); 708 709 UVMHIST_LOG(maphist,"<- done (kva=0x%x)", kva,0,0,0); 710 return(kva); 711 } 712 713 /* 714 * uvm_km_free: free an area of kernel memory 715 */ 716 717 void 718 uvm_km_free(struct vm_map *map, vaddr_t addr, vsize_t size, uvm_flag_t flags) 719 { 720 UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist); 721 722 KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED || 723 (flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE || 724 (flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY); 725 KASSERT((addr & PAGE_MASK) == 0); 726 KASSERT(vm_map_pmap(map) == pmap_kernel()); 727 728 size = round_page(size); 729 730 if (flags & UVM_KMF_PAGEABLE) { 731 uvm_km_pgremove(addr, addr + size); 732 } else if (flags & UVM_KMF_WIRED) { 733 /* 734 * Note: uvm_km_pgremove_intrsafe() extracts mapping, thus 735 * remove it after. See comment below about KVA visibility. 736 */ 737 uvm_km_pgremove_intrsafe(map, addr, addr + size); 738 } 739 740 /* 741 * Note: uvm_unmap_remove() calls pmap_update() for us, before 742 * KVA becomes globally available. 743 */ 744 745 uvm_unmap1(map, addr, addr + size, UVM_FLAG_VAONLY); 746 } 747 748 /* Sanity; must specify both or none. */ 749 #if (defined(PMAP_MAP_POOLPAGE) || defined(PMAP_UNMAP_POOLPAGE)) && \ 750 (!defined(PMAP_MAP_POOLPAGE) || !defined(PMAP_UNMAP_POOLPAGE)) 751 #error Must specify MAP and UNMAP together. 752 #endif 753 754 int 755 uvm_km_kmem_alloc(vmem_t *vm, vmem_size_t size, vm_flag_t flags, 756 vmem_addr_t *addr) 757 { 758 struct vm_page *pg; 759 vmem_addr_t va; 760 int rc; 761 vaddr_t loopva; 762 vsize_t loopsize; 763 764 size = round_page(size); 765 766 #if defined(PMAP_MAP_POOLPAGE) 767 if (size == PAGE_SIZE) { 768 again: 769 #ifdef PMAP_ALLOC_POOLPAGE 770 pg = PMAP_ALLOC_POOLPAGE((flags & VM_SLEEP) ? 771 0 : UVM_PGA_USERESERVE); 772 #else 773 pg = uvm_pagealloc(NULL, 0, NULL, 774 (flags & VM_SLEEP) ? 0 : UVM_PGA_USERESERVE); 775 #endif /* PMAP_ALLOC_POOLPAGE */ 776 if (__predict_false(pg == NULL)) { 777 if (flags & VM_SLEEP) { 778 uvm_wait("plpg"); 779 goto again; 780 } 781 return ENOMEM; 782 } 783 va = PMAP_MAP_POOLPAGE(VM_PAGE_TO_PHYS(pg)); 784 if (__predict_false(va == 0)) { 785 uvm_pagefree(pg); 786 return ENOMEM; 787 } 788 *addr = va; 789 return 0; 790 } 791 #endif /* PMAP_MAP_POOLPAGE */ 792 793 rc = vmem_alloc(vm, size, flags, &va); 794 if (rc != 0) 795 return rc; 796 797 #ifdef PMAP_GROWKERNEL 798 /* 799 * These VA allocations happen independently of uvm_map 800 * so this allocation must not extend beyond the current limit. 801 */ 802 KASSERTMSG(uvm_maxkaddr >= va + size, 803 "%#"PRIxVADDR" %#"PRIxPTR" %#zx", 804 uvm_maxkaddr, va, size); 805 #endif 806 807 loopva = va; 808 loopsize = size; 809 810 while (loopsize) { 811 #ifdef DIAGNOSTIC 812 paddr_t pa; 813 #endif 814 KASSERTMSG(!pmap_extract(pmap_kernel(), loopva, &pa), 815 "loopva=%#"PRIxVADDR" loopsize=%#"PRIxVSIZE 816 " pa=%#"PRIxPADDR" vmem=%p", 817 loopva, loopsize, pa, vm); 818 819 pg = uvm_pagealloc(NULL, loopva, NULL, 820 UVM_FLAG_COLORMATCH 821 | ((flags & VM_SLEEP) ? 0 : UVM_PGA_USERESERVE)); 822 if (__predict_false(pg == NULL)) { 823 if (flags & VM_SLEEP) { 824 uvm_wait("plpg"); 825 continue; 826 } else { 827 uvm_km_pgremove_intrsafe(kernel_map, va, 828 va + size); 829 vmem_free(vm, va, size); 830 return ENOMEM; 831 } 832 } 833 834 pg->flags &= ~PG_BUSY; /* new page */ 835 UVM_PAGE_OWN(pg, NULL); 836 pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg), 837 VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE); 838 839 loopva += PAGE_SIZE; 840 loopsize -= PAGE_SIZE; 841 } 842 pmap_update(pmap_kernel()); 843 844 *addr = va; 845 846 return 0; 847 } 848 849 void 850 uvm_km_kmem_free(vmem_t *vm, vmem_addr_t addr, size_t size) 851 { 852 853 size = round_page(size); 854 #if defined(PMAP_UNMAP_POOLPAGE) 855 if (size == PAGE_SIZE) { 856 paddr_t pa; 857 858 pa = PMAP_UNMAP_POOLPAGE(addr); 859 uvm_pagefree(PHYS_TO_VM_PAGE(pa)); 860 return; 861 } 862 #endif /* PMAP_UNMAP_POOLPAGE */ 863 uvm_km_pgremove_intrsafe(kernel_map, addr, addr + size); 864 pmap_update(pmap_kernel()); 865 866 vmem_free(vm, addr, size); 867 } 868 869 bool 870 uvm_km_va_starved_p(void) 871 { 872 vmem_size_t total; 873 vmem_size_t free; 874 875 if (kmem_arena == NULL) 876 return false; 877 878 total = vmem_size(kmem_arena, VMEM_ALLOC|VMEM_FREE); 879 free = vmem_size(kmem_arena, VMEM_FREE); 880 881 return (free < (total / 10)); 882 } 883