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 #include "opt_vm.h" 69 70 #include <sys/param.h> 71 #include <sys/systm.h> 72 #include <sys/asan.h> 73 #include <sys/domainset.h> 74 #include <sys/eventhandler.h> 75 #include <sys/kernel.h> 76 #include <sys/lock.h> 77 #include <sys/malloc.h> 78 #include <sys/msan.h> 79 #include <sys/proc.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_pagequeue.h> 95 #include <vm/vm_phys.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 (1ul << 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 TSENTER(); 150 size = round_page(size); 151 if (vmem_alloc(kernel_arena, size, M_BESTFIT | M_NOWAIT, &addr)) 152 return (0); 153 TSEXIT(); 154 155 return (addr); 156 } 157 158 /* 159 * kva_free: 160 * 161 * Release a region of kernel virtual memory allocated 162 * with kva_alloc, and return the physical pages 163 * associated with that region. 164 * 165 * This routine may not block on kernel maps. 166 */ 167 void 168 kva_free(vm_offset_t addr, vm_size_t size) 169 { 170 171 size = round_page(size); 172 vmem_free(kernel_arena, addr, size); 173 } 174 175 /* 176 * Update sanitizer shadow state to reflect a new allocation. Force inlining to 177 * help make KMSAN origin tracking more precise. 178 */ 179 static __always_inline void 180 kmem_alloc_san(vm_offset_t addr, vm_size_t size, vm_size_t asize, int flags) 181 { 182 if ((flags & M_ZERO) == 0) { 183 kmsan_mark((void *)addr, asize, KMSAN_STATE_UNINIT); 184 kmsan_orig((void *)addr, asize, KMSAN_TYPE_KMEM, 185 KMSAN_RET_ADDR); 186 } else { 187 kmsan_mark((void *)addr, asize, KMSAN_STATE_INITED); 188 } 189 kasan_mark((void *)addr, size, asize, KASAN_KMEM_REDZONE); 190 } 191 192 static vm_page_t 193 kmem_alloc_contig_pages(vm_object_t object, vm_pindex_t pindex, int domain, 194 int pflags, u_long npages, vm_paddr_t low, vm_paddr_t high, 195 u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) 196 { 197 vm_page_t m; 198 int tries; 199 bool wait, reclaim; 200 201 VM_OBJECT_ASSERT_WLOCKED(object); 202 203 wait = (pflags & VM_ALLOC_WAITOK) != 0; 204 reclaim = (pflags & VM_ALLOC_NORECLAIM) == 0; 205 pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL); 206 pflags |= VM_ALLOC_NOWAIT; 207 for (tries = wait ? 3 : 1;; tries--) { 208 m = vm_page_alloc_contig_domain(object, pindex, domain, pflags, 209 npages, low, high, alignment, boundary, memattr); 210 if (m != NULL || tries == 0 || !reclaim) 211 break; 212 213 VM_OBJECT_WUNLOCK(object); 214 if (!vm_page_reclaim_contig_domain(domain, pflags, npages, 215 low, high, alignment, boundary) && wait) 216 vm_wait_domain(domain); 217 VM_OBJECT_WLOCK(object); 218 } 219 return (m); 220 } 221 222 /* 223 * Allocates a region from the kernel address map and physical pages 224 * within the specified address range to the kernel object. Creates a 225 * wired mapping from this region to these pages, and returns the 226 * region's starting virtual address. The allocated pages are not 227 * necessarily physically contiguous. If M_ZERO is specified through the 228 * given flags, then the pages are zeroed before they are mapped. 229 */ 230 static void * 231 kmem_alloc_attr_domain(int domain, vm_size_t size, int flags, vm_paddr_t low, 232 vm_paddr_t high, vm_memattr_t memattr) 233 { 234 vmem_t *vmem; 235 vm_object_t object; 236 vm_offset_t addr, i, offset; 237 vm_page_t m; 238 vm_size_t asize; 239 int pflags; 240 vm_prot_t prot; 241 242 object = kernel_object; 243 asize = round_page(size); 244 vmem = vm_dom[domain].vmd_kernel_arena; 245 if (vmem_alloc(vmem, asize, M_BESTFIT | flags, &addr)) 246 return (0); 247 offset = addr - VM_MIN_KERNEL_ADDRESS; 248 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED; 249 prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW; 250 VM_OBJECT_WLOCK(object); 251 for (i = 0; i < asize; i += PAGE_SIZE) { 252 m = kmem_alloc_contig_pages(object, atop(offset + i), 253 domain, pflags, 1, low, high, PAGE_SIZE, 0, memattr); 254 if (m == NULL) { 255 VM_OBJECT_WUNLOCK(object); 256 kmem_unback(object, addr, i); 257 vmem_free(vmem, addr, asize); 258 return (0); 259 } 260 KASSERT(vm_page_domain(m) == domain, 261 ("kmem_alloc_attr_domain: Domain mismatch %d != %d", 262 vm_page_domain(m), domain)); 263 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0) 264 pmap_zero_page(m); 265 vm_page_valid(m); 266 pmap_enter(kernel_pmap, addr + i, m, prot, 267 prot | PMAP_ENTER_WIRED, 0); 268 } 269 VM_OBJECT_WUNLOCK(object); 270 kmem_alloc_san(addr, size, asize, flags); 271 return ((void *)addr); 272 } 273 274 void * 275 kmem_alloc_attr(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high, 276 vm_memattr_t memattr) 277 { 278 279 return (kmem_alloc_attr_domainset(DOMAINSET_RR(), size, flags, low, 280 high, memattr)); 281 } 282 283 void * 284 kmem_alloc_attr_domainset(struct domainset *ds, vm_size_t size, int flags, 285 vm_paddr_t low, vm_paddr_t high, vm_memattr_t memattr) 286 { 287 struct vm_domainset_iter di; 288 void *addr; 289 int domain; 290 291 vm_domainset_iter_policy_init(&di, ds, &domain, &flags); 292 do { 293 addr = kmem_alloc_attr_domain(domain, size, flags, low, high, 294 memattr); 295 if (addr != NULL) 296 break; 297 } while (vm_domainset_iter_policy(&di, &domain) == 0); 298 299 return (addr); 300 } 301 302 /* 303 * Allocates a region from the kernel address map and physically 304 * contiguous pages within the specified address range to the kernel 305 * object. Creates a wired mapping from this region to these pages, and 306 * returns the region's starting virtual address. If M_ZERO is specified 307 * through the given flags, then the pages are zeroed before they are 308 * mapped. 309 */ 310 static void * 311 kmem_alloc_contig_domain(int domain, vm_size_t size, int flags, vm_paddr_t low, 312 vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 313 vm_memattr_t memattr) 314 { 315 vmem_t *vmem; 316 vm_object_t object; 317 vm_offset_t addr, offset, tmp; 318 vm_page_t end_m, m; 319 vm_size_t asize; 320 u_long npages; 321 int pflags; 322 323 object = kernel_object; 324 asize = round_page(size); 325 vmem = vm_dom[domain].vmd_kernel_arena; 326 if (vmem_alloc(vmem, asize, flags | M_BESTFIT, &addr)) 327 return (NULL); 328 offset = addr - VM_MIN_KERNEL_ADDRESS; 329 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED; 330 npages = atop(asize); 331 VM_OBJECT_WLOCK(object); 332 m = kmem_alloc_contig_pages(object, atop(offset), domain, 333 pflags, npages, low, high, alignment, boundary, memattr); 334 if (m == NULL) { 335 VM_OBJECT_WUNLOCK(object); 336 vmem_free(vmem, addr, asize); 337 return (NULL); 338 } 339 KASSERT(vm_page_domain(m) == domain, 340 ("kmem_alloc_contig_domain: Domain mismatch %d != %d", 341 vm_page_domain(m), domain)); 342 end_m = m + npages; 343 tmp = addr; 344 for (; m < end_m; m++) { 345 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0) 346 pmap_zero_page(m); 347 vm_page_valid(m); 348 pmap_enter(kernel_pmap, tmp, m, VM_PROT_RW, 349 VM_PROT_RW | PMAP_ENTER_WIRED, 0); 350 tmp += PAGE_SIZE; 351 } 352 VM_OBJECT_WUNLOCK(object); 353 kmem_alloc_san(addr, size, asize, flags); 354 return ((void *)addr); 355 } 356 357 void * 358 kmem_alloc_contig(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high, 359 u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) 360 { 361 362 return (kmem_alloc_contig_domainset(DOMAINSET_RR(), size, flags, low, 363 high, alignment, boundary, memattr)); 364 } 365 366 void * 367 kmem_alloc_contig_domainset(struct domainset *ds, vm_size_t size, int flags, 368 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 369 vm_memattr_t memattr) 370 { 371 struct vm_domainset_iter di; 372 void *addr; 373 int domain; 374 375 vm_domainset_iter_policy_init(&di, ds, &domain, &flags); 376 do { 377 addr = kmem_alloc_contig_domain(domain, size, flags, low, high, 378 alignment, boundary, memattr); 379 if (addr != NULL) 380 break; 381 } while (vm_domainset_iter_policy(&di, &domain) == 0); 382 383 return (addr); 384 } 385 386 /* 387 * kmem_subinit: 388 * 389 * Initializes a map to manage a subrange 390 * of the kernel virtual address space. 391 * 392 * Arguments are as follows: 393 * 394 * parent Map to take range from 395 * min, max Returned endpoints of map 396 * size Size of range to find 397 * superpage_align Request that min is superpage aligned 398 */ 399 void 400 kmem_subinit(vm_map_t map, vm_map_t parent, vm_offset_t *min, vm_offset_t *max, 401 vm_size_t size, bool superpage_align) 402 { 403 int ret; 404 405 size = round_page(size); 406 407 *min = vm_map_min(parent); 408 ret = vm_map_find(parent, NULL, 0, min, size, 0, superpage_align ? 409 VMFS_SUPER_SPACE : VMFS_ANY_SPACE, VM_PROT_ALL, VM_PROT_ALL, 410 MAP_ACC_NO_CHARGE); 411 if (ret != KERN_SUCCESS) 412 panic("kmem_subinit: bad status return of %d", ret); 413 *max = *min + size; 414 vm_map_init(map, vm_map_pmap(parent), *min, *max); 415 if (vm_map_submap(parent, *min, *max, map) != KERN_SUCCESS) 416 panic("kmem_subinit: unable to change range to submap"); 417 } 418 419 /* 420 * kmem_malloc_domain: 421 * 422 * Allocate wired-down pages in the kernel's address space. 423 */ 424 static void * 425 kmem_malloc_domain(int domain, vm_size_t size, int flags) 426 { 427 vmem_t *arena; 428 vm_offset_t addr; 429 vm_size_t asize; 430 int rv; 431 432 if (__predict_true((flags & M_EXEC) == 0)) 433 arena = vm_dom[domain].vmd_kernel_arena; 434 else 435 arena = vm_dom[domain].vmd_kernel_rwx_arena; 436 asize = round_page(size); 437 if (vmem_alloc(arena, asize, flags | M_BESTFIT, &addr)) 438 return (0); 439 440 rv = kmem_back_domain(domain, kernel_object, addr, asize, flags); 441 if (rv != KERN_SUCCESS) { 442 vmem_free(arena, addr, asize); 443 return (0); 444 } 445 kasan_mark((void *)addr, size, asize, KASAN_KMEM_REDZONE); 446 return ((void *)addr); 447 } 448 449 void * 450 kmem_malloc(vm_size_t size, int flags) 451 { 452 void * p; 453 454 TSENTER(); 455 p = kmem_malloc_domainset(DOMAINSET_RR(), size, flags); 456 TSEXIT(); 457 return (p); 458 } 459 460 void * 461 kmem_malloc_domainset(struct domainset *ds, vm_size_t size, int flags) 462 { 463 struct vm_domainset_iter di; 464 void *addr; 465 int domain; 466 467 vm_domainset_iter_policy_init(&di, ds, &domain, &flags); 468 do { 469 addr = kmem_malloc_domain(domain, size, flags); 470 if (addr != NULL) 471 break; 472 } while (vm_domainset_iter_policy(&di, &domain) == 0); 473 474 return (addr); 475 } 476 477 /* 478 * kmem_back_domain: 479 * 480 * Allocate physical pages from the specified domain for the specified 481 * virtual address range. 482 */ 483 int 484 kmem_back_domain(int domain, vm_object_t object, vm_offset_t addr, 485 vm_size_t size, int flags) 486 { 487 vm_offset_t offset, i; 488 vm_page_t m, mpred; 489 vm_prot_t prot; 490 int pflags; 491 492 KASSERT(object == kernel_object, 493 ("kmem_back_domain: only supports kernel object.")); 494 495 offset = addr - VM_MIN_KERNEL_ADDRESS; 496 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED; 497 pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL); 498 if (flags & M_WAITOK) 499 pflags |= VM_ALLOC_WAITFAIL; 500 prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW; 501 502 i = 0; 503 VM_OBJECT_WLOCK(object); 504 retry: 505 mpred = vm_radix_lookup_le(&object->rtree, atop(offset + i)); 506 for (; i < size; i += PAGE_SIZE, mpred = m) { 507 m = vm_page_alloc_domain_after(object, atop(offset + i), 508 domain, pflags, mpred); 509 510 /* 511 * Ran out of space, free everything up and return. Don't need 512 * to lock page queues here as we know that the pages we got 513 * aren't on any queues. 514 */ 515 if (m == NULL) { 516 if ((flags & M_NOWAIT) == 0) 517 goto retry; 518 VM_OBJECT_WUNLOCK(object); 519 kmem_unback(object, addr, i); 520 return (KERN_NO_SPACE); 521 } 522 KASSERT(vm_page_domain(m) == domain, 523 ("kmem_back_domain: Domain mismatch %d != %d", 524 vm_page_domain(m), domain)); 525 if (flags & M_ZERO && (m->flags & PG_ZERO) == 0) 526 pmap_zero_page(m); 527 KASSERT((m->oflags & VPO_UNMANAGED) != 0, 528 ("kmem_malloc: page %p is managed", m)); 529 vm_page_valid(m); 530 pmap_enter(kernel_pmap, addr + i, m, prot, 531 prot | PMAP_ENTER_WIRED, 0); 532 if (__predict_false((prot & VM_PROT_EXECUTE) != 0)) 533 m->oflags |= VPO_KMEM_EXEC; 534 } 535 VM_OBJECT_WUNLOCK(object); 536 kmem_alloc_san(addr, size, size, flags); 537 return (KERN_SUCCESS); 538 } 539 540 /* 541 * kmem_back: 542 * 543 * Allocate physical pages for the specified virtual address range. 544 */ 545 int 546 kmem_back(vm_object_t object, vm_offset_t addr, vm_size_t size, int flags) 547 { 548 vm_offset_t end, next, start; 549 int domain, rv; 550 551 KASSERT(object == kernel_object, 552 ("kmem_back: only supports kernel object.")); 553 554 for (start = addr, end = addr + size; addr < end; addr = next) { 555 /* 556 * We must ensure that pages backing a given large virtual page 557 * all come from the same physical domain. 558 */ 559 if (vm_ndomains > 1) { 560 domain = (addr >> KVA_QUANTUM_SHIFT) % vm_ndomains; 561 while (VM_DOMAIN_EMPTY(domain)) 562 domain++; 563 next = roundup2(addr + 1, KVA_QUANTUM); 564 if (next > end || next < start) 565 next = end; 566 } else { 567 domain = 0; 568 next = end; 569 } 570 rv = kmem_back_domain(domain, object, addr, next - addr, flags); 571 if (rv != KERN_SUCCESS) { 572 kmem_unback(object, start, addr - start); 573 break; 574 } 575 } 576 return (rv); 577 } 578 579 /* 580 * kmem_unback: 581 * 582 * Unmap and free the physical pages underlying the specified virtual 583 * address range. 584 * 585 * A physical page must exist within the specified object at each index 586 * that is being unmapped. 587 */ 588 static struct vmem * 589 _kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size) 590 { 591 struct vmem *arena; 592 vm_page_t m, next; 593 vm_offset_t end, offset; 594 int domain; 595 596 KASSERT(object == kernel_object, 597 ("kmem_unback: only supports kernel object.")); 598 599 if (size == 0) 600 return (NULL); 601 pmap_remove(kernel_pmap, addr, addr + size); 602 offset = addr - VM_MIN_KERNEL_ADDRESS; 603 end = offset + size; 604 VM_OBJECT_WLOCK(object); 605 m = vm_page_lookup(object, atop(offset)); 606 domain = vm_page_domain(m); 607 if (__predict_true((m->oflags & VPO_KMEM_EXEC) == 0)) 608 arena = vm_dom[domain].vmd_kernel_arena; 609 else 610 arena = vm_dom[domain].vmd_kernel_rwx_arena; 611 for (; offset < end; offset += PAGE_SIZE, m = next) { 612 next = vm_page_next(m); 613 vm_page_xbusy_claim(m); 614 vm_page_unwire_noq(m); 615 vm_page_free(m); 616 } 617 VM_OBJECT_WUNLOCK(object); 618 619 return (arena); 620 } 621 622 void 623 kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size) 624 { 625 626 (void)_kmem_unback(object, addr, size); 627 } 628 629 /* 630 * kmem_free: 631 * 632 * Free memory allocated with kmem_malloc. The size must match the 633 * original allocation. 634 */ 635 void 636 kmem_free(void *addr, vm_size_t size) 637 { 638 struct vmem *arena; 639 640 size = round_page(size); 641 kasan_mark(addr, size, size, 0); 642 arena = _kmem_unback(kernel_object, (uintptr_t)addr, size); 643 if (arena != NULL) 644 vmem_free(arena, (uintptr_t)addr, size); 645 } 646 647 /* 648 * kmap_alloc_wait: 649 * 650 * Allocates pageable memory from a sub-map of the kernel. If the submap 651 * has no room, the caller sleeps waiting for more memory in the submap. 652 * 653 * This routine may block. 654 */ 655 vm_offset_t 656 kmap_alloc_wait(vm_map_t map, vm_size_t size) 657 { 658 vm_offset_t addr; 659 660 size = round_page(size); 661 if (!swap_reserve(size)) 662 return (0); 663 664 for (;;) { 665 /* 666 * To make this work for more than one map, use the map's lock 667 * to lock out sleepers/wakers. 668 */ 669 vm_map_lock(map); 670 addr = vm_map_findspace(map, vm_map_min(map), size); 671 if (addr + size <= vm_map_max(map)) 672 break; 673 /* no space now; see if we can ever get space */ 674 if (vm_map_max(map) - vm_map_min(map) < size) { 675 vm_map_unlock(map); 676 swap_release(size); 677 return (0); 678 } 679 map->needs_wakeup = TRUE; 680 vm_map_unlock_and_wait(map, 0); 681 } 682 vm_map_insert(map, NULL, 0, addr, addr + size, VM_PROT_RW, VM_PROT_RW, 683 MAP_ACC_CHARGED); 684 vm_map_unlock(map); 685 return (addr); 686 } 687 688 /* 689 * kmap_free_wakeup: 690 * 691 * Returns memory to a submap of the kernel, and wakes up any processes 692 * waiting for memory in that map. 693 */ 694 void 695 kmap_free_wakeup(vm_map_t map, vm_offset_t addr, vm_size_t size) 696 { 697 698 vm_map_lock(map); 699 (void) vm_map_delete(map, trunc_page(addr), round_page(addr + size)); 700 if (map->needs_wakeup) { 701 map->needs_wakeup = FALSE; 702 vm_map_wakeup(map); 703 } 704 vm_map_unlock(map); 705 } 706 707 void 708 kmem_init_zero_region(void) 709 { 710 vm_offset_t addr, i; 711 vm_page_t m; 712 713 /* 714 * Map a single physical page of zeros to a larger virtual range. 715 * This requires less looping in places that want large amounts of 716 * zeros, while not using much more physical resources. 717 */ 718 addr = kva_alloc(ZERO_REGION_SIZE); 719 m = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_ZERO); 720 for (i = 0; i < ZERO_REGION_SIZE; i += PAGE_SIZE) 721 pmap_qenter(addr + i, &m, 1); 722 pmap_protect(kernel_pmap, addr, addr + ZERO_REGION_SIZE, VM_PROT_READ); 723 724 zero_region = (const void *)addr; 725 } 726 727 /* 728 * Import KVA from the kernel map into the kernel arena. 729 */ 730 static int 731 kva_import(void *unused, vmem_size_t size, int flags, vmem_addr_t *addrp) 732 { 733 vm_offset_t addr; 734 int result; 735 736 TSENTER(); 737 KASSERT((size % KVA_QUANTUM) == 0, 738 ("kva_import: Size %jd is not a multiple of %d", 739 (intmax_t)size, (int)KVA_QUANTUM)); 740 addr = vm_map_min(kernel_map); 741 result = vm_map_find(kernel_map, NULL, 0, &addr, size, 0, 742 VMFS_SUPER_SPACE, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 743 if (result != KERN_SUCCESS) { 744 TSEXIT(); 745 return (ENOMEM); 746 } 747 748 *addrp = addr; 749 750 TSEXIT(); 751 return (0); 752 } 753 754 /* 755 * Import KVA from a parent arena into a per-domain arena. Imports must be 756 * KVA_QUANTUM-aligned and a multiple of KVA_QUANTUM in size. 757 */ 758 static int 759 kva_import_domain(void *arena, vmem_size_t size, int flags, vmem_addr_t *addrp) 760 { 761 762 KASSERT((size % KVA_QUANTUM) == 0, 763 ("kva_import_domain: Size %jd is not a multiple of %d", 764 (intmax_t)size, (int)KVA_QUANTUM)); 765 return (vmem_xalloc(arena, size, KVA_QUANTUM, 0, 0, VMEM_ADDR_MIN, 766 VMEM_ADDR_MAX, flags, addrp)); 767 } 768 769 /* 770 * kmem_init: 771 * 772 * Create the kernel map; insert a mapping covering kernel text, 773 * data, bss, and all space allocated thus far (`boostrap' data). The 774 * new map will thus map the range between VM_MIN_KERNEL_ADDRESS and 775 * `start' as allocated, and the range between `start' and `end' as free. 776 * Create the kernel vmem arena and its per-domain children. 777 */ 778 void 779 kmem_init(vm_offset_t start, vm_offset_t end) 780 { 781 vm_size_t quantum; 782 int domain; 783 784 vm_map_init(kernel_map, kernel_pmap, VM_MIN_KERNEL_ADDRESS, end); 785 kernel_map->system_map = 1; 786 vm_map_lock(kernel_map); 787 /* N.B.: cannot use kgdb to debug, starting with this assignment ... */ 788 (void)vm_map_insert(kernel_map, NULL, 0, 789 #ifdef __amd64__ 790 KERNBASE, 791 #else 792 VM_MIN_KERNEL_ADDRESS, 793 #endif 794 start, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 795 /* ... and ending with the completion of the above `insert' */ 796 797 #ifdef __amd64__ 798 /* 799 * Mark KVA used for the page array as allocated. Other platforms 800 * that handle vm_page_array allocation can simply adjust virtual_avail 801 * instead. 802 */ 803 (void)vm_map_insert(kernel_map, NULL, 0, (vm_offset_t)vm_page_array, 804 (vm_offset_t)vm_page_array + round_2mpage(vm_page_array_size * 805 sizeof(struct vm_page)), 806 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT); 807 #endif 808 vm_map_unlock(kernel_map); 809 810 /* 811 * Use a large import quantum on NUMA systems. This helps minimize 812 * interleaving of superpages, reducing internal fragmentation within 813 * the per-domain arenas. 814 */ 815 if (vm_ndomains > 1 && PMAP_HAS_DMAP) 816 quantum = KVA_NUMA_IMPORT_QUANTUM; 817 else 818 quantum = KVA_QUANTUM; 819 820 /* 821 * Initialize the kernel_arena. This can grow on demand. 822 */ 823 vmem_init(kernel_arena, "kernel arena", 0, 0, PAGE_SIZE, 0, 0); 824 vmem_set_import(kernel_arena, kva_import, NULL, NULL, quantum); 825 826 for (domain = 0; domain < vm_ndomains; domain++) { 827 /* 828 * Initialize the per-domain arenas. These are used to color 829 * the KVA space in a way that ensures that virtual large pages 830 * are backed by memory from the same physical domain, 831 * maximizing the potential for superpage promotion. 832 */ 833 vm_dom[domain].vmd_kernel_arena = vmem_create( 834 "kernel arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK); 835 vmem_set_import(vm_dom[domain].vmd_kernel_arena, 836 kva_import_domain, NULL, kernel_arena, quantum); 837 838 /* 839 * In architectures with superpages, maintain separate arenas 840 * for allocations with permissions that differ from the 841 * "standard" read/write permissions used for kernel memory, 842 * so as not to inhibit superpage promotion. 843 * 844 * Use the base import quantum since this arena is rarely used. 845 */ 846 #if VM_NRESERVLEVEL > 0 847 vm_dom[domain].vmd_kernel_rwx_arena = vmem_create( 848 "kernel rwx arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK); 849 vmem_set_import(vm_dom[domain].vmd_kernel_rwx_arena, 850 kva_import_domain, (vmem_release_t *)vmem_xfree, 851 kernel_arena, KVA_QUANTUM); 852 #else 853 vm_dom[domain].vmd_kernel_rwx_arena = 854 vm_dom[domain].vmd_kernel_arena; 855 #endif 856 } 857 858 /* 859 * This must be the very first call so that the virtual address 860 * space used for early allocations is properly marked used in 861 * the map. 862 */ 863 uma_startup2(); 864 } 865 866 /* 867 * kmem_bootstrap_free: 868 * 869 * Free pages backing preloaded data (e.g., kernel modules) to the 870 * system. Currently only supported on platforms that create a 871 * vm_phys segment for preloaded data. 872 */ 873 void 874 kmem_bootstrap_free(vm_offset_t start, vm_size_t size) 875 { 876 #if defined(__i386__) || defined(__amd64__) 877 struct vm_domain *vmd; 878 vm_offset_t end, va; 879 vm_paddr_t pa; 880 vm_page_t m; 881 882 end = trunc_page(start + size); 883 start = round_page(start); 884 885 #ifdef __amd64__ 886 /* 887 * Preloaded files do not have execute permissions by default on amd64. 888 * Restore the default permissions to ensure that the direct map alias 889 * is updated. 890 */ 891 pmap_change_prot(start, end - start, VM_PROT_RW); 892 #endif 893 for (va = start; va < end; va += PAGE_SIZE) { 894 pa = pmap_kextract(va); 895 m = PHYS_TO_VM_PAGE(pa); 896 897 vmd = vm_pagequeue_domain(m); 898 vm_domain_free_lock(vmd); 899 vm_phys_free_pages(m, 0); 900 vm_domain_free_unlock(vmd); 901 902 vm_domain_freecnt_inc(vmd, 1); 903 vm_cnt.v_page_count++; 904 } 905 pmap_remove(kernel_pmap, start, end); 906 (void)vmem_add(kernel_arena, start, end - start, M_WAITOK); 907 #endif 908 } 909 910 /* 911 * Allow userspace to directly trigger the VM drain routine for testing 912 * purposes. 913 */ 914 static int 915 debug_vm_lowmem(SYSCTL_HANDLER_ARGS) 916 { 917 int error, i; 918 919 i = 0; 920 error = sysctl_handle_int(oidp, &i, 0, req); 921 if (error != 0) 922 return (error); 923 if ((i & ~(VM_LOW_KMEM | VM_LOW_PAGES)) != 0) 924 return (EINVAL); 925 if (i != 0) 926 EVENTHANDLER_INVOKE(vm_lowmem, i); 927 return (0); 928 } 929 SYSCTL_PROC(_debug, OID_AUTO, vm_lowmem, 930 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, debug_vm_lowmem, "I", 931 "set to trigger vm_lowmem event with given flags"); 932 933 static int 934 debug_uma_reclaim(SYSCTL_HANDLER_ARGS) 935 { 936 int error, i; 937 938 i = 0; 939 error = sysctl_handle_int(oidp, &i, 0, req); 940 if (error != 0 || req->newptr == NULL) 941 return (error); 942 if (i != UMA_RECLAIM_TRIM && i != UMA_RECLAIM_DRAIN && 943 i != UMA_RECLAIM_DRAIN_CPU) 944 return (EINVAL); 945 uma_reclaim(i); 946 return (0); 947 } 948 SYSCTL_PROC(_debug, OID_AUTO, uma_reclaim, 949 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, debug_uma_reclaim, "I", 950 "set to generate request to reclaim uma caches"); 951 952 static int 953 debug_uma_reclaim_domain(SYSCTL_HANDLER_ARGS) 954 { 955 int domain, error, request; 956 957 request = 0; 958 error = sysctl_handle_int(oidp, &request, 0, req); 959 if (error != 0 || req->newptr == NULL) 960 return (error); 961 962 domain = request >> 4; 963 request &= 0xf; 964 if (request != UMA_RECLAIM_TRIM && request != UMA_RECLAIM_DRAIN && 965 request != UMA_RECLAIM_DRAIN_CPU) 966 return (EINVAL); 967 if (domain < 0 || domain >= vm_ndomains) 968 return (EINVAL); 969 uma_reclaim_domain(request, domain); 970 return (0); 971 } 972 SYSCTL_PROC(_debug, OID_AUTO, uma_reclaim_domain, 973 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, 974 debug_uma_reclaim_domain, "I", 975 ""); 976