1 /*- 2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) 3 * 4 * Copyright (c) 1991, 1993 5 * The Regents of the University of California. All rights reserved. 6 * 7 * This code is derived from software contributed to Berkeley by 8 * The Mach Operating System project at Carnegie-Mellon University. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * from: @(#)vm_kern.c 8.3 (Berkeley) 1/12/94 35 * 36 * 37 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 38 * All rights reserved. 39 * 40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 */ 62 63 /* 64 * Kernel memory management. 65 */ 66 67 #include <sys/cdefs.h> 68 __FBSDID("$FreeBSD$"); 69 70 #include "opt_vm.h" 71 72 #include <sys/param.h> 73 #include <sys/systm.h> 74 #include <sys/kernel.h> /* for ticks and hz */ 75 #include <sys/domainset.h> 76 #include <sys/eventhandler.h> 77 #include <sys/lock.h> 78 #include <sys/proc.h> 79 #include <sys/malloc.h> 80 #include <sys/rwlock.h> 81 #include <sys/sysctl.h> 82 #include <sys/vmem.h> 83 #include <sys/vmmeter.h> 84 85 #include <vm/vm.h> 86 #include <vm/vm_param.h> 87 #include <vm/vm_domainset.h> 88 #include <vm/vm_kern.h> 89 #include <vm/pmap.h> 90 #include <vm/vm_map.h> 91 #include <vm/vm_object.h> 92 #include <vm/vm_page.h> 93 #include <vm/vm_pageout.h> 94 #include <vm/vm_phys.h> 95 #include <vm/vm_pagequeue.h> 96 #include <vm/vm_radix.h> 97 #include <vm/vm_extern.h> 98 #include <vm/uma.h> 99 100 struct vm_map kernel_map_store; 101 struct vm_map exec_map_store; 102 struct vm_map pipe_map_store; 103 104 const void *zero_region; 105 CTASSERT((ZERO_REGION_SIZE & PAGE_MASK) == 0); 106 107 /* NB: Used by kernel debuggers. */ 108 const u_long vm_maxuser_address = VM_MAXUSER_ADDRESS; 109 110 u_int exec_map_entry_size; 111 u_int exec_map_entries; 112 113 SYSCTL_ULONG(_vm, OID_AUTO, min_kernel_address, CTLFLAG_RD, 114 SYSCTL_NULL_ULONG_PTR, VM_MIN_KERNEL_ADDRESS, "Min kernel address"); 115 116 SYSCTL_ULONG(_vm, OID_AUTO, max_kernel_address, CTLFLAG_RD, 117 #if defined(__arm__) 118 &vm_max_kernel_address, 0, 119 #else 120 SYSCTL_NULL_ULONG_PTR, VM_MAX_KERNEL_ADDRESS, 121 #endif 122 "Max kernel address"); 123 124 #if VM_NRESERVLEVEL > 0 125 #define KVA_QUANTUM_SHIFT (VM_LEVEL_0_ORDER + PAGE_SHIFT) 126 #else 127 /* On non-superpage architectures we want large import sizes. */ 128 #define KVA_QUANTUM_SHIFT (8 + PAGE_SHIFT) 129 #endif 130 #define KVA_QUANTUM (1 << KVA_QUANTUM_SHIFT) 131 #define KVA_NUMA_IMPORT_QUANTUM (KVA_QUANTUM * 128) 132 133 extern void uma_startup2(void); 134 135 /* 136 * kva_alloc: 137 * 138 * Allocate a virtual address range with no underlying object and 139 * no initial mapping to physical memory. Any mapping from this 140 * range to physical memory must be explicitly created prior to 141 * its use, typically with pmap_qenter(). Any attempt to create 142 * a mapping on demand through vm_fault() will result in a panic. 143 */ 144 vm_offset_t 145 kva_alloc(vm_size_t size) 146 { 147 vm_offset_t addr; 148 149 size = round_page(size); 150 if (vmem_alloc(kernel_arena, size, M_BESTFIT | M_NOWAIT, &addr)) 151 return (0); 152 153 return (addr); 154 } 155 156 /* 157 * kva_free: 158 * 159 * Release a region of kernel virtual memory allocated 160 * with kva_alloc, and return the physical pages 161 * associated with that region. 162 * 163 * This routine may not block on kernel maps. 164 */ 165 void 166 kva_free(vm_offset_t addr, vm_size_t size) 167 { 168 169 size = round_page(size); 170 vmem_free(kernel_arena, addr, size); 171 } 172 173 static vm_page_t 174 kmem_alloc_contig_pages(vm_object_t object, vm_pindex_t pindex, int domain, 175 int pflags, u_long npages, vm_paddr_t low, vm_paddr_t high, 176 u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) 177 { 178 vm_page_t m; 179 int tries; 180 bool wait; 181 182 VM_OBJECT_ASSERT_WLOCKED(object); 183 184 wait = (pflags & VM_ALLOC_WAITOK) != 0; 185 pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL); 186 pflags |= VM_ALLOC_NOWAIT; 187 for (tries = wait ? 3 : 1;; tries--) { 188 m = vm_page_alloc_contig_domain(object, pindex, domain, pflags, 189 npages, low, high, alignment, boundary, memattr); 190 if (m != NULL || tries == 0) 191 break; 192 193 VM_OBJECT_WUNLOCK(object); 194 if (!vm_page_reclaim_contig_domain(domain, pflags, npages, 195 low, high, alignment, boundary) && wait) 196 vm_wait_domain(domain); 197 VM_OBJECT_WLOCK(object); 198 } 199 return (m); 200 } 201 202 /* 203 * Allocates a region from the kernel address map and physical pages 204 * within the specified address range to the kernel object. Creates a 205 * wired mapping from this region to these pages, and returns the 206 * region's starting virtual address. The allocated pages are not 207 * necessarily physically contiguous. If M_ZERO is specified through the 208 * given flags, then the pages are zeroed before they are mapped. 209 */ 210 static vm_offset_t 211 kmem_alloc_attr_domain(int domain, vm_size_t size, int flags, vm_paddr_t low, 212 vm_paddr_t high, vm_memattr_t memattr) 213 { 214 vmem_t *vmem; 215 vm_object_t object; 216 vm_offset_t addr, i, offset; 217 vm_page_t m; 218 int pflags; 219 vm_prot_t prot; 220 221 object = kernel_object; 222 size = round_page(size); 223 vmem = vm_dom[domain].vmd_kernel_arena; 224 if (vmem_alloc(vmem, size, M_BESTFIT | flags, &addr)) 225 return (0); 226 offset = addr - VM_MIN_KERNEL_ADDRESS; 227 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED; 228 prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW; 229 VM_OBJECT_WLOCK(object); 230 for (i = 0; i < size; i += PAGE_SIZE) { 231 m = kmem_alloc_contig_pages(object, atop(offset + i), 232 domain, pflags, 1, low, high, PAGE_SIZE, 0, memattr); 233 if (m == NULL) { 234 VM_OBJECT_WUNLOCK(object); 235 kmem_unback(object, addr, i); 236 vmem_free(vmem, addr, size); 237 return (0); 238 } 239 KASSERT(vm_phys_domain(m) == domain, 240 ("kmem_alloc_attr_domain: Domain mismatch %d != %d", 241 vm_phys_domain(m), domain)); 242 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0) 243 pmap_zero_page(m); 244 vm_page_valid(m); 245 pmap_enter(kernel_pmap, addr + i, m, prot, 246 prot | PMAP_ENTER_WIRED, 0); 247 } 248 VM_OBJECT_WUNLOCK(object); 249 return (addr); 250 } 251 252 vm_offset_t 253 kmem_alloc_attr(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high, 254 vm_memattr_t memattr) 255 { 256 257 return (kmem_alloc_attr_domainset(DOMAINSET_RR(), size, flags, low, 258 high, memattr)); 259 } 260 261 vm_offset_t 262 kmem_alloc_attr_domainset(struct domainset *ds, vm_size_t size, int flags, 263 vm_paddr_t low, vm_paddr_t high, vm_memattr_t memattr) 264 { 265 struct vm_domainset_iter di; 266 vm_offset_t addr; 267 int domain; 268 269 vm_domainset_iter_policy_init(&di, ds, &domain, &flags); 270 do { 271 addr = kmem_alloc_attr_domain(domain, size, flags, low, high, 272 memattr); 273 if (addr != 0) 274 break; 275 } while (vm_domainset_iter_policy(&di, &domain) == 0); 276 277 return (addr); 278 } 279 280 /* 281 * Allocates a region from the kernel address map and physically 282 * contiguous pages within the specified address range to the kernel 283 * object. Creates a wired mapping from this region to these pages, and 284 * returns the region's starting virtual address. If M_ZERO is specified 285 * through the given flags, then the pages are zeroed before they are 286 * mapped. 287 */ 288 static vm_offset_t 289 kmem_alloc_contig_domain(int domain, vm_size_t size, int flags, vm_paddr_t low, 290 vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 291 vm_memattr_t memattr) 292 { 293 vmem_t *vmem; 294 vm_object_t object; 295 vm_offset_t addr, offset, tmp; 296 vm_page_t end_m, m; 297 u_long npages; 298 int pflags; 299 300 object = kernel_object; 301 size = round_page(size); 302 vmem = vm_dom[domain].vmd_kernel_arena; 303 if (vmem_alloc(vmem, size, flags | M_BESTFIT, &addr)) 304 return (0); 305 offset = addr - VM_MIN_KERNEL_ADDRESS; 306 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED; 307 npages = atop(size); 308 VM_OBJECT_WLOCK(object); 309 m = kmem_alloc_contig_pages(object, atop(offset), domain, 310 pflags, npages, low, high, alignment, boundary, memattr); 311 if (m == NULL) { 312 VM_OBJECT_WUNLOCK(object); 313 vmem_free(vmem, addr, size); 314 return (0); 315 } 316 KASSERT(vm_phys_domain(m) == domain, 317 ("kmem_alloc_contig_domain: Domain mismatch %d != %d", 318 vm_phys_domain(m), domain)); 319 end_m = m + npages; 320 tmp = addr; 321 for (; m < end_m; m++) { 322 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0) 323 pmap_zero_page(m); 324 vm_page_valid(m); 325 pmap_enter(kernel_pmap, tmp, m, VM_PROT_RW, 326 VM_PROT_RW | PMAP_ENTER_WIRED, 0); 327 tmp += PAGE_SIZE; 328 } 329 VM_OBJECT_WUNLOCK(object); 330 return (addr); 331 } 332 333 vm_offset_t 334 kmem_alloc_contig(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high, 335 u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) 336 { 337 338 return (kmem_alloc_contig_domainset(DOMAINSET_RR(), size, flags, low, 339 high, alignment, boundary, memattr)); 340 } 341 342 vm_offset_t 343 kmem_alloc_contig_domainset(struct domainset *ds, vm_size_t size, int flags, 344 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 345 vm_memattr_t memattr) 346 { 347 struct vm_domainset_iter di; 348 vm_offset_t addr; 349 int domain; 350 351 vm_domainset_iter_policy_init(&di, ds, &domain, &flags); 352 do { 353 addr = kmem_alloc_contig_domain(domain, size, flags, low, high, 354 alignment, boundary, memattr); 355 if (addr != 0) 356 break; 357 } while (vm_domainset_iter_policy(&di, &domain) == 0); 358 359 return (addr); 360 } 361 362 /* 363 * kmem_subinit: 364 * 365 * Initializes a map to manage a subrange 366 * of the kernel virtual address space. 367 * 368 * Arguments are as follows: 369 * 370 * parent Map to take range from 371 * min, max Returned endpoints of map 372 * size Size of range to find 373 * superpage_align Request that min is superpage aligned 374 */ 375 void 376 kmem_subinit(vm_map_t map, vm_map_t parent, vm_offset_t *min, vm_offset_t *max, 377 vm_size_t size, bool superpage_align) 378 { 379 int ret; 380 381 size = round_page(size); 382 383 *min = vm_map_min(parent); 384 ret = vm_map_find(parent, NULL, 0, min, size, 0, superpage_align ? 385 VMFS_SUPER_SPACE : VMFS_ANY_SPACE, VM_PROT_ALL, VM_PROT_ALL, 386 MAP_ACC_NO_CHARGE); 387 if (ret != KERN_SUCCESS) 388 panic("kmem_subinit: bad status return of %d", ret); 389 *max = *min + size; 390 vm_map_init(map, vm_map_pmap(parent), *min, *max); 391 if (vm_map_submap(parent, *min, *max, map) != KERN_SUCCESS) 392 panic("kmem_subinit: unable to change range to submap"); 393 } 394 395 /* 396 * kmem_malloc_domain: 397 * 398 * Allocate wired-down pages in the kernel's address space. 399 */ 400 static vm_offset_t 401 kmem_malloc_domain(int domain, vm_size_t size, int flags) 402 { 403 vmem_t *arena; 404 vm_offset_t addr; 405 int rv; 406 407 if (__predict_true((flags & M_EXEC) == 0)) 408 arena = vm_dom[domain].vmd_kernel_arena; 409 else 410 arena = vm_dom[domain].vmd_kernel_rwx_arena; 411 size = round_page(size); 412 if (vmem_alloc(arena, size, flags | M_BESTFIT, &addr)) 413 return (0); 414 415 rv = kmem_back_domain(domain, kernel_object, addr, size, flags); 416 if (rv != KERN_SUCCESS) { 417 vmem_free(arena, addr, size); 418 return (0); 419 } 420 return (addr); 421 } 422 423 vm_offset_t 424 kmem_malloc(vm_size_t size, int flags) 425 { 426 427 return (kmem_malloc_domainset(DOMAINSET_RR(), size, flags)); 428 } 429 430 vm_offset_t 431 kmem_malloc_domainset(struct domainset *ds, vm_size_t size, int flags) 432 { 433 struct vm_domainset_iter di; 434 vm_offset_t addr; 435 int domain; 436 437 vm_domainset_iter_policy_init(&di, ds, &domain, &flags); 438 do { 439 addr = kmem_malloc_domain(domain, size, flags); 440 if (addr != 0) 441 break; 442 } while (vm_domainset_iter_policy(&di, &domain) == 0); 443 444 return (addr); 445 } 446 447 /* 448 * kmem_back_domain: 449 * 450 * Allocate physical pages from the specified domain for the specified 451 * virtual address range. 452 */ 453 int 454 kmem_back_domain(int domain, vm_object_t object, vm_offset_t addr, 455 vm_size_t size, int flags) 456 { 457 vm_offset_t offset, i; 458 vm_page_t m, mpred; 459 vm_prot_t prot; 460 int pflags; 461 462 KASSERT(object == kernel_object, 463 ("kmem_back_domain: only supports kernel object.")); 464 465 offset = addr - VM_MIN_KERNEL_ADDRESS; 466 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED; 467 pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL); 468 if (flags & M_WAITOK) 469 pflags |= VM_ALLOC_WAITFAIL; 470 prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW; 471 472 i = 0; 473 VM_OBJECT_WLOCK(object); 474 retry: 475 mpred = vm_radix_lookup_le(&object->rtree, atop(offset + i)); 476 for (; i < size; i += PAGE_SIZE, mpred = m) { 477 m = vm_page_alloc_domain_after(object, atop(offset + i), 478 domain, pflags, mpred); 479 480 /* 481 * Ran out of space, free everything up and return. Don't need 482 * to lock page queues here as we know that the pages we got 483 * aren't on any queues. 484 */ 485 if (m == NULL) { 486 if ((flags & M_NOWAIT) == 0) 487 goto retry; 488 VM_OBJECT_WUNLOCK(object); 489 kmem_unback(object, addr, i); 490 return (KERN_NO_SPACE); 491 } 492 KASSERT(vm_phys_domain(m) == domain, 493 ("kmem_back_domain: Domain mismatch %d != %d", 494 vm_phys_domain(m), domain)); 495 if (flags & M_ZERO && (m->flags & PG_ZERO) == 0) 496 pmap_zero_page(m); 497 KASSERT((m->oflags & VPO_UNMANAGED) != 0, 498 ("kmem_malloc: page %p is managed", m)); 499 vm_page_valid(m); 500 pmap_enter(kernel_pmap, addr + i, m, prot, 501 prot | PMAP_ENTER_WIRED, 0); 502 if (__predict_false((prot & VM_PROT_EXECUTE) != 0)) 503 m->oflags |= VPO_KMEM_EXEC; 504 } 505 VM_OBJECT_WUNLOCK(object); 506 507 return (KERN_SUCCESS); 508 } 509 510 /* 511 * kmem_back: 512 * 513 * Allocate physical pages for the specified virtual address range. 514 */ 515 int 516 kmem_back(vm_object_t object, vm_offset_t addr, vm_size_t size, int flags) 517 { 518 vm_offset_t end, next, start; 519 int domain, rv; 520 521 KASSERT(object == kernel_object, 522 ("kmem_back: only supports kernel object.")); 523 524 for (start = addr, end = addr + size; addr < end; addr = next) { 525 /* 526 * We must ensure that pages backing a given large virtual page 527 * all come from the same physical domain. 528 */ 529 if (vm_ndomains > 1) { 530 domain = (addr >> KVA_QUANTUM_SHIFT) % vm_ndomains; 531 while (VM_DOMAIN_EMPTY(domain)) 532 domain++; 533 next = roundup2(addr + 1, KVA_QUANTUM); 534 if (next > end || next < start) 535 next = end; 536 } else { 537 domain = 0; 538 next = end; 539 } 540 rv = kmem_back_domain(domain, object, addr, next - addr, flags); 541 if (rv != KERN_SUCCESS) { 542 kmem_unback(object, start, addr - start); 543 break; 544 } 545 } 546 return (rv); 547 } 548 549 /* 550 * kmem_unback: 551 * 552 * Unmap and free the physical pages underlying the specified virtual 553 * address range. 554 * 555 * A physical page must exist within the specified object at each index 556 * that is being unmapped. 557 */ 558 static struct vmem * 559 _kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size) 560 { 561 struct vmem *arena; 562 vm_page_t m, next; 563 vm_offset_t end, offset; 564 int domain; 565 566 KASSERT(object == kernel_object, 567 ("kmem_unback: only supports kernel object.")); 568 569 if (size == 0) 570 return (NULL); 571 pmap_remove(kernel_pmap, addr, addr + size); 572 offset = addr - VM_MIN_KERNEL_ADDRESS; 573 end = offset + size; 574 VM_OBJECT_WLOCK(object); 575 m = vm_page_lookup(object, atop(offset)); 576 domain = vm_phys_domain(m); 577 if (__predict_true((m->oflags & VPO_KMEM_EXEC) == 0)) 578 arena = vm_dom[domain].vmd_kernel_arena; 579 else 580 arena = vm_dom[domain].vmd_kernel_rwx_arena; 581 for (; offset < end; offset += PAGE_SIZE, m = next) { 582 next = vm_page_next(m); 583 vm_page_xbusy_claim(m); 584 vm_page_unwire_noq(m); 585 vm_page_free(m); 586 } 587 VM_OBJECT_WUNLOCK(object); 588 589 return (arena); 590 } 591 592 void 593 kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size) 594 { 595 596 (void)_kmem_unback(object, addr, size); 597 } 598 599 /* 600 * kmem_free: 601 * 602 * Free memory allocated with kmem_malloc. The size must match the 603 * original allocation. 604 */ 605 void 606 kmem_free(vm_offset_t addr, vm_size_t size) 607 { 608 struct vmem *arena; 609 610 size = round_page(size); 611 arena = _kmem_unback(kernel_object, addr, size); 612 if (arena != NULL) 613 vmem_free(arena, addr, size); 614 } 615 616 /* 617 * kmap_alloc_wait: 618 * 619 * Allocates pageable memory from a sub-map of the kernel. If the submap 620 * has no room, the caller sleeps waiting for more memory in the submap. 621 * 622 * This routine may block. 623 */ 624 vm_offset_t 625 kmap_alloc_wait(vm_map_t map, vm_size_t size) 626 { 627 vm_offset_t addr; 628 629 size = round_page(size); 630 if (!swap_reserve(size)) 631 return (0); 632 633 for (;;) { 634 /* 635 * To make this work for more than one map, use the map's lock 636 * to lock out sleepers/wakers. 637 */ 638 vm_map_lock(map); 639 addr = vm_map_findspace(map, vm_map_min(map), size); 640 if (addr + size <= vm_map_max(map)) 641 break; 642 /* no space now; see if we can ever get space */ 643 if (vm_map_max(map) - vm_map_min(map) < size) { 644 vm_map_unlock(map); 645 swap_release(size); 646 return (0); 647 } 648 map->needs_wakeup = TRUE; 649 vm_map_unlock_and_wait(map, 0); 650 } 651 vm_map_insert(map, NULL, 0, addr, addr + size, VM_PROT_RW, VM_PROT_RW, 652 MAP_ACC_CHARGED); 653 vm_map_unlock(map); 654 return (addr); 655 } 656 657 /* 658 * kmap_free_wakeup: 659 * 660 * Returns memory to a submap of the kernel, and wakes up any processes 661 * waiting for memory in that map. 662 */ 663 void 664 kmap_free_wakeup(vm_map_t map, vm_offset_t addr, vm_size_t size) 665 { 666 667 vm_map_lock(map); 668 (void) vm_map_delete(map, trunc_page(addr), round_page(addr + size)); 669 if (map->needs_wakeup) { 670 map->needs_wakeup = FALSE; 671 vm_map_wakeup(map); 672 } 673 vm_map_unlock(map); 674 } 675 676 void 677 kmem_init_zero_region(void) 678 { 679 vm_offset_t addr, i; 680 vm_page_t m; 681 682 /* 683 * Map a single physical page of zeros to a larger virtual range. 684 * This requires less looping in places that want large amounts of 685 * zeros, while not using much more physical resources. 686 */ 687 addr = kva_alloc(ZERO_REGION_SIZE); 688 m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | 689 VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); 690 if ((m->flags & PG_ZERO) == 0) 691 pmap_zero_page(m); 692 for (i = 0; i < ZERO_REGION_SIZE; i += PAGE_SIZE) 693 pmap_qenter(addr + i, &m, 1); 694 pmap_protect(kernel_pmap, addr, addr + ZERO_REGION_SIZE, VM_PROT_READ); 695 696 zero_region = (const void *)addr; 697 } 698 699 /* 700 * Import KVA from the kernel map into the kernel arena. 701 */ 702 static int 703 kva_import(void *unused, vmem_size_t size, int flags, vmem_addr_t *addrp) 704 { 705 vm_offset_t addr; 706 int result; 707 708 KASSERT((size % KVA_QUANTUM) == 0, 709 ("kva_import: Size %jd is not a multiple of %d", 710 (intmax_t)size, (int)KVA_QUANTUM)); 711 addr = vm_map_min(kernel_map); 712 result = vm_map_find(kernel_map, NULL, 0, &addr, size, 0, 713 VMFS_SUPER_SPACE, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 714 if (result != KERN_SUCCESS) 715 return (ENOMEM); 716 717 *addrp = addr; 718 719 return (0); 720 } 721 722 /* 723 * Import KVA from a parent arena into a per-domain arena. Imports must be 724 * KVA_QUANTUM-aligned and a multiple of KVA_QUANTUM in size. 725 */ 726 static int 727 kva_import_domain(void *arena, vmem_size_t size, int flags, vmem_addr_t *addrp) 728 { 729 730 KASSERT((size % KVA_QUANTUM) == 0, 731 ("kva_import_domain: Size %jd is not a multiple of %d", 732 (intmax_t)size, (int)KVA_QUANTUM)); 733 return (vmem_xalloc(arena, size, KVA_QUANTUM, 0, 0, VMEM_ADDR_MIN, 734 VMEM_ADDR_MAX, flags, addrp)); 735 } 736 737 /* 738 * kmem_init: 739 * 740 * Create the kernel map; insert a mapping covering kernel text, 741 * data, bss, and all space allocated thus far (`boostrap' data). The 742 * new map will thus map the range between VM_MIN_KERNEL_ADDRESS and 743 * `start' as allocated, and the range between `start' and `end' as free. 744 * Create the kernel vmem arena and its per-domain children. 745 */ 746 void 747 kmem_init(vm_offset_t start, vm_offset_t end) 748 { 749 vm_size_t quantum; 750 int domain; 751 752 vm_map_init(kernel_map, kernel_pmap, VM_MIN_KERNEL_ADDRESS, end); 753 kernel_map->system_map = 1; 754 vm_map_lock(kernel_map); 755 /* N.B.: cannot use kgdb to debug, starting with this assignment ... */ 756 (void)vm_map_insert(kernel_map, NULL, 0, 757 #ifdef __amd64__ 758 KERNBASE, 759 #else 760 VM_MIN_KERNEL_ADDRESS, 761 #endif 762 start, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 763 /* ... and ending with the completion of the above `insert' */ 764 765 #ifdef __amd64__ 766 /* 767 * Mark KVA used for the page array as allocated. Other platforms 768 * that handle vm_page_array allocation can simply adjust virtual_avail 769 * instead. 770 */ 771 (void)vm_map_insert(kernel_map, NULL, 0, (vm_offset_t)vm_page_array, 772 (vm_offset_t)vm_page_array + round_2mpage(vm_page_array_size * 773 sizeof(struct vm_page)), 774 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT); 775 #endif 776 vm_map_unlock(kernel_map); 777 778 /* 779 * Use a large import quantum on NUMA systems. This helps minimize 780 * interleaving of superpages, reducing internal fragmentation within 781 * the per-domain arenas. 782 */ 783 if (vm_ndomains > 1 && PMAP_HAS_DMAP) 784 quantum = KVA_NUMA_IMPORT_QUANTUM; 785 else 786 quantum = KVA_QUANTUM; 787 788 /* 789 * Initialize the kernel_arena. This can grow on demand. 790 */ 791 vmem_init(kernel_arena, "kernel arena", 0, 0, PAGE_SIZE, 0, 0); 792 vmem_set_import(kernel_arena, kva_import, NULL, NULL, quantum); 793 794 for (domain = 0; domain < vm_ndomains; domain++) { 795 /* 796 * Initialize the per-domain arenas. These are used to color 797 * the KVA space in a way that ensures that virtual large pages 798 * are backed by memory from the same physical domain, 799 * maximizing the potential for superpage promotion. 800 */ 801 vm_dom[domain].vmd_kernel_arena = vmem_create( 802 "kernel arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK); 803 vmem_set_import(vm_dom[domain].vmd_kernel_arena, 804 kva_import_domain, NULL, kernel_arena, quantum); 805 806 /* 807 * In architectures with superpages, maintain separate arenas 808 * for allocations with permissions that differ from the 809 * "standard" read/write permissions used for kernel memory, 810 * so as not to inhibit superpage promotion. 811 * 812 * Use the base import quantum since this arena is rarely used. 813 */ 814 #if VM_NRESERVLEVEL > 0 815 vm_dom[domain].vmd_kernel_rwx_arena = vmem_create( 816 "kernel rwx arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK); 817 vmem_set_import(vm_dom[domain].vmd_kernel_rwx_arena, 818 kva_import_domain, (vmem_release_t *)vmem_xfree, 819 kernel_arena, KVA_QUANTUM); 820 #else 821 vm_dom[domain].vmd_kernel_rwx_arena = 822 vm_dom[domain].vmd_kernel_arena; 823 #endif 824 } 825 826 /* 827 * This must be the very first call so that the virtual address 828 * space used for early allocations is properly marked used in 829 * the map. 830 */ 831 uma_startup2(); 832 } 833 834 /* 835 * kmem_bootstrap_free: 836 * 837 * Free pages backing preloaded data (e.g., kernel modules) to the 838 * system. Currently only supported on platforms that create a 839 * vm_phys segment for preloaded data. 840 */ 841 void 842 kmem_bootstrap_free(vm_offset_t start, vm_size_t size) 843 { 844 #if defined(__i386__) || defined(__amd64__) 845 struct vm_domain *vmd; 846 vm_offset_t end, va; 847 vm_paddr_t pa; 848 vm_page_t m; 849 850 end = trunc_page(start + size); 851 start = round_page(start); 852 853 #ifdef __amd64__ 854 /* 855 * Preloaded files do not have execute permissions by default on amd64. 856 * Restore the default permissions to ensure that the direct map alias 857 * is updated. 858 */ 859 pmap_change_prot(start, end - start, VM_PROT_RW); 860 #endif 861 for (va = start; va < end; va += PAGE_SIZE) { 862 pa = pmap_kextract(va); 863 m = PHYS_TO_VM_PAGE(pa); 864 865 vmd = vm_pagequeue_domain(m); 866 vm_domain_free_lock(vmd); 867 vm_phys_free_pages(m, 0); 868 vm_domain_free_unlock(vmd); 869 870 vm_domain_freecnt_inc(vmd, 1); 871 vm_cnt.v_page_count++; 872 } 873 pmap_remove(kernel_pmap, start, end); 874 (void)vmem_add(kernel_arena, start, end - start, M_WAITOK); 875 #endif 876 } 877 878 /* 879 * Allow userspace to directly trigger the VM drain routine for testing 880 * purposes. 881 */ 882 static int 883 debug_vm_lowmem(SYSCTL_HANDLER_ARGS) 884 { 885 int error, i; 886 887 i = 0; 888 error = sysctl_handle_int(oidp, &i, 0, req); 889 if (error) 890 return (error); 891 if ((i & ~(VM_LOW_KMEM | VM_LOW_PAGES)) != 0) 892 return (EINVAL); 893 if (i != 0) 894 EVENTHANDLER_INVOKE(vm_lowmem, i); 895 return (0); 896 } 897 898 SYSCTL_PROC(_debug, OID_AUTO, vm_lowmem, CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, 899 debug_vm_lowmem, "I", "set to trigger vm_lowmem event with given flags"); 900