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 * 35 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 36 * All rights reserved. 37 * 38 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 39 * 40 * Permission to use, copy, modify and distribute this software and 41 * its documentation is hereby granted, provided that both the copyright 42 * notice and this permission notice appear in all copies of the 43 * software, derivative works or modified versions, and any portions 44 * thereof, and that both notices appear in supporting documentation. 45 * 46 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 47 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 48 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 49 * 50 * Carnegie Mellon requests users of this software to return to 51 * 52 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 53 * School of Computer Science 54 * Carnegie Mellon University 55 * Pittsburgh PA 15213-3890 56 * 57 * any improvements or extensions that they make and grant Carnegie the 58 * rights to redistribute these changes. 59 */ 60 61 /* 62 * Kernel memory management. 63 */ 64 65 #include <sys/cdefs.h> 66 #include "opt_vm.h" 67 68 #include <sys/param.h> 69 #include <sys/systm.h> 70 #include <sys/asan.h> 71 #include <sys/domainset.h> 72 #include <sys/eventhandler.h> 73 #include <sys/kernel.h> 74 #include <sys/lock.h> 75 #include <sys/malloc.h> 76 #include <sys/msan.h> 77 #include <sys/proc.h> 78 #include <sys/rwlock.h> 79 #include <sys/smp.h> 80 #include <sys/sysctl.h> 81 #include <sys/vmem.h> 82 #include <sys/vmmeter.h> 83 84 #include <vm/vm.h> 85 #include <vm/vm_param.h> 86 #include <vm/vm_domainset.h> 87 #include <vm/vm_kern.h> 88 #include <vm/pmap.h> 89 #include <vm/vm_map.h> 90 #include <vm/vm_object.h> 91 #include <vm/vm_page.h> 92 #include <vm/vm_pageout.h> 93 #include <vm/vm_pagequeue.h> 94 #include <vm/vm_phys.h> 95 #include <vm/vm_radix.h> 96 #include <vm/vm_extern.h> 97 #include <vm/uma.h> 98 99 struct vm_map kernel_map_store; 100 struct vm_map exec_map_store; 101 struct vm_map pipe_map_store; 102 103 const void *zero_region; 104 CTASSERT((ZERO_REGION_SIZE & PAGE_MASK) == 0); 105 106 /* NB: Used by kernel debuggers. */ 107 const u_long vm_maxuser_address = VM_MAXUSER_ADDRESS; 108 109 u_int exec_map_entry_size; 110 u_int exec_map_entries; 111 112 SYSCTL_ULONG(_vm, OID_AUTO, min_kernel_address, CTLFLAG_RD, 113 SYSCTL_NULL_ULONG_PTR, VM_MIN_KERNEL_ADDRESS, "Min kernel address"); 114 115 SYSCTL_ULONG(_vm, OID_AUTO, max_kernel_address, CTLFLAG_RD, 116 #if defined(__arm__) 117 &vm_max_kernel_address, 0, 118 #else 119 SYSCTL_NULL_ULONG_PTR, VM_MAX_KERNEL_ADDRESS, 120 #endif 121 "Max kernel address"); 122 123 #if VM_NRESERVLEVEL > 0 124 #define KVA_QUANTUM_SHIFT (VM_LEVEL_0_ORDER + PAGE_SHIFT) 125 #else 126 /* On non-superpage architectures we want large import sizes. */ 127 #define KVA_QUANTUM_SHIFT (8 + PAGE_SHIFT) 128 #endif 129 #define KVA_QUANTUM (1ul << KVA_QUANTUM_SHIFT) 130 #define KVA_NUMA_IMPORT_QUANTUM (KVA_QUANTUM * 128) 131 132 extern void uma_startup2(void); 133 134 /* 135 * kva_alloc: 136 * 137 * Allocate a virtual address range with no underlying object and 138 * no initial mapping to physical memory. Any mapping from this 139 * range to physical memory must be explicitly created prior to 140 * its use, typically with pmap_qenter(). Any attempt to create 141 * a mapping on demand through vm_fault() will result in a panic. 142 */ 143 vm_offset_t 144 kva_alloc(vm_size_t size) 145 { 146 vm_offset_t addr; 147 148 TSENTER(); 149 size = round_page(size); 150 if (vmem_alloc(kernel_arena, size, M_BESTFIT | M_NOWAIT, &addr)) 151 return (0); 152 TSEXIT(); 153 154 return (addr); 155 } 156 157 /* 158 * kva_free: 159 * 160 * Release a region of kernel virtual memory allocated 161 * with kva_alloc, and return the physical pages 162 * associated with that region. 163 * 164 * This routine may not block on kernel maps. 165 */ 166 void 167 kva_free(vm_offset_t addr, vm_size_t size) 168 { 169 170 size = round_page(size); 171 vmem_free(kernel_arena, addr, size); 172 } 173 174 /* 175 * Update sanitizer shadow state to reflect a new allocation. Force inlining to 176 * help make KMSAN origin tracking more precise. 177 */ 178 static __always_inline void 179 kmem_alloc_san(vm_offset_t addr, vm_size_t size, vm_size_t asize, int flags) 180 { 181 if ((flags & M_ZERO) == 0) { 182 kmsan_mark((void *)addr, asize, KMSAN_STATE_UNINIT); 183 kmsan_orig((void *)addr, asize, KMSAN_TYPE_KMEM, 184 KMSAN_RET_ADDR); 185 } else { 186 kmsan_mark((void *)addr, asize, KMSAN_STATE_INITED); 187 } 188 kasan_mark((void *)addr, size, asize, KASAN_KMEM_REDZONE); 189 } 190 191 static vm_page_t 192 kmem_alloc_contig_pages(vm_object_t object, vm_pindex_t pindex, int domain, 193 int pflags, u_long npages, vm_paddr_t low, vm_paddr_t high, 194 u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) 195 { 196 vm_page_t m; 197 int tries; 198 bool wait, reclaim; 199 200 VM_OBJECT_ASSERT_WLOCKED(object); 201 202 wait = (pflags & VM_ALLOC_WAITOK) != 0; 203 reclaim = (pflags & VM_ALLOC_NORECLAIM) == 0; 204 pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL); 205 pflags |= VM_ALLOC_NOWAIT; 206 for (tries = wait ? 3 : 1;; tries--) { 207 m = vm_page_alloc_contig_domain(object, pindex, domain, pflags, 208 npages, low, high, alignment, boundary, memattr); 209 if (m != NULL || tries == 0 || !reclaim) 210 break; 211 212 VM_OBJECT_WUNLOCK(object); 213 if (!vm_page_reclaim_contig_domain(domain, pflags, npages, 214 low, high, alignment, boundary) && wait) 215 vm_wait_domain(domain); 216 VM_OBJECT_WLOCK(object); 217 } 218 return (m); 219 } 220 221 /* 222 * Allocates a region from the kernel address map and physical pages 223 * within the specified address range to the kernel object. Creates a 224 * wired mapping from this region to these pages, and returns the 225 * region's starting virtual address. The allocated pages are not 226 * necessarily physically contiguous. If M_ZERO is specified through the 227 * given flags, then the pages are zeroed before they are mapped. 228 */ 229 static void * 230 kmem_alloc_attr_domain(int domain, vm_size_t size, int flags, vm_paddr_t low, 231 vm_paddr_t high, vm_memattr_t memattr) 232 { 233 vmem_t *vmem; 234 vm_object_t object; 235 vm_offset_t addr, i, offset; 236 vm_page_t m; 237 vm_size_t asize; 238 int pflags; 239 vm_prot_t prot; 240 241 object = kernel_object; 242 asize = round_page(size); 243 vmem = vm_dom[domain].vmd_kernel_arena; 244 if (vmem_alloc(vmem, asize, M_BESTFIT | flags, &addr)) 245 return (0); 246 offset = addr - VM_MIN_KERNEL_ADDRESS; 247 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED; 248 prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW; 249 VM_OBJECT_WLOCK(object); 250 for (i = 0; i < asize; i += PAGE_SIZE) { 251 m = kmem_alloc_contig_pages(object, atop(offset + i), 252 domain, pflags, 1, low, high, PAGE_SIZE, 0, memattr); 253 if (m == NULL) { 254 VM_OBJECT_WUNLOCK(object); 255 kmem_unback(object, addr, i); 256 vmem_free(vmem, addr, asize); 257 return (0); 258 } 259 KASSERT(vm_page_domain(m) == domain, 260 ("kmem_alloc_attr_domain: Domain mismatch %d != %d", 261 vm_page_domain(m), domain)); 262 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0) 263 pmap_zero_page(m); 264 vm_page_valid(m); 265 pmap_enter(kernel_pmap, addr + i, m, prot, 266 prot | PMAP_ENTER_WIRED, 0); 267 } 268 VM_OBJECT_WUNLOCK(object); 269 kmem_alloc_san(addr, size, asize, flags); 270 return ((void *)addr); 271 } 272 273 void * 274 kmem_alloc_attr(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high, 275 vm_memattr_t memattr) 276 { 277 278 return (kmem_alloc_attr_domainset(DOMAINSET_RR(), size, flags, low, 279 high, memattr)); 280 } 281 282 void * 283 kmem_alloc_attr_domainset(struct domainset *ds, vm_size_t size, int flags, 284 vm_paddr_t low, vm_paddr_t high, vm_memattr_t memattr) 285 { 286 struct vm_domainset_iter di; 287 void *addr; 288 int domain; 289 290 vm_domainset_iter_policy_init(&di, ds, &domain, &flags); 291 do { 292 addr = kmem_alloc_attr_domain(domain, size, flags, low, high, 293 memattr); 294 if (addr != NULL) 295 break; 296 } while (vm_domainset_iter_policy(&di, &domain) == 0); 297 298 return (addr); 299 } 300 301 /* 302 * Allocates a region from the kernel address map and physically 303 * contiguous pages within the specified address range to the kernel 304 * object. Creates a wired mapping from this region to these pages, and 305 * returns the region's starting virtual address. If M_ZERO is specified 306 * through the given flags, then the pages are zeroed before they are 307 * mapped. 308 */ 309 static void * 310 kmem_alloc_contig_domain(int domain, vm_size_t size, int flags, vm_paddr_t low, 311 vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 312 vm_memattr_t memattr) 313 { 314 vmem_t *vmem; 315 vm_object_t object; 316 vm_offset_t addr, offset, tmp; 317 vm_page_t end_m, m; 318 vm_size_t asize; 319 u_long npages; 320 int pflags; 321 322 object = kernel_object; 323 asize = round_page(size); 324 vmem = vm_dom[domain].vmd_kernel_arena; 325 if (vmem_alloc(vmem, asize, flags | M_BESTFIT, &addr)) 326 return (NULL); 327 offset = addr - VM_MIN_KERNEL_ADDRESS; 328 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED; 329 npages = atop(asize); 330 VM_OBJECT_WLOCK(object); 331 m = kmem_alloc_contig_pages(object, atop(offset), domain, 332 pflags, npages, low, high, alignment, boundary, memattr); 333 if (m == NULL) { 334 VM_OBJECT_WUNLOCK(object); 335 vmem_free(vmem, addr, asize); 336 return (NULL); 337 } 338 KASSERT(vm_page_domain(m) == domain, 339 ("kmem_alloc_contig_domain: Domain mismatch %d != %d", 340 vm_page_domain(m), domain)); 341 end_m = m + npages; 342 tmp = addr; 343 for (; m < end_m; m++) { 344 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0) 345 pmap_zero_page(m); 346 vm_page_valid(m); 347 pmap_enter(kernel_pmap, tmp, m, VM_PROT_RW, 348 VM_PROT_RW | PMAP_ENTER_WIRED, 0); 349 tmp += PAGE_SIZE; 350 } 351 VM_OBJECT_WUNLOCK(object); 352 kmem_alloc_san(addr, size, asize, flags); 353 return ((void *)addr); 354 } 355 356 void * 357 kmem_alloc_contig(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high, 358 u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) 359 { 360 361 return (kmem_alloc_contig_domainset(DOMAINSET_RR(), size, flags, low, 362 high, alignment, boundary, memattr)); 363 } 364 365 void * 366 kmem_alloc_contig_domainset(struct domainset *ds, vm_size_t size, int flags, 367 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, 368 vm_memattr_t memattr) 369 { 370 struct vm_domainset_iter di; 371 void *addr; 372 int domain; 373 374 vm_domainset_iter_policy_init(&di, ds, &domain, &flags); 375 do { 376 addr = kmem_alloc_contig_domain(domain, size, flags, low, high, 377 alignment, boundary, memattr); 378 if (addr != NULL) 379 break; 380 } while (vm_domainset_iter_policy(&di, &domain) == 0); 381 382 return (addr); 383 } 384 385 /* 386 * kmem_subinit: 387 * 388 * Initializes a map to manage a subrange 389 * of the kernel virtual address space. 390 * 391 * Arguments are as follows: 392 * 393 * parent Map to take range from 394 * min, max Returned endpoints of map 395 * size Size of range to find 396 * superpage_align Request that min is superpage aligned 397 */ 398 void 399 kmem_subinit(vm_map_t map, vm_map_t parent, vm_offset_t *min, vm_offset_t *max, 400 vm_size_t size, bool superpage_align) 401 { 402 int ret; 403 404 size = round_page(size); 405 406 *min = vm_map_min(parent); 407 ret = vm_map_find(parent, NULL, 0, min, size, 0, superpage_align ? 408 VMFS_SUPER_SPACE : VMFS_ANY_SPACE, VM_PROT_ALL, VM_PROT_ALL, 409 MAP_ACC_NO_CHARGE); 410 if (ret != KERN_SUCCESS) 411 panic("kmem_subinit: bad status return of %d", ret); 412 *max = *min + size; 413 vm_map_init(map, vm_map_pmap(parent), *min, *max); 414 if (vm_map_submap(parent, *min, *max, map) != KERN_SUCCESS) 415 panic("kmem_subinit: unable to change range to submap"); 416 } 417 418 /* 419 * kmem_malloc_domain: 420 * 421 * Allocate wired-down pages in the kernel's address space. 422 */ 423 static void * 424 kmem_malloc_domain(int domain, vm_size_t size, int flags) 425 { 426 vmem_t *arena; 427 vm_offset_t addr; 428 vm_size_t asize; 429 int rv; 430 431 if (__predict_true((flags & M_EXEC) == 0)) 432 arena = vm_dom[domain].vmd_kernel_arena; 433 else 434 arena = vm_dom[domain].vmd_kernel_rwx_arena; 435 asize = round_page(size); 436 if (vmem_alloc(arena, asize, flags | M_BESTFIT, &addr)) 437 return (0); 438 439 rv = kmem_back_domain(domain, kernel_object, addr, asize, flags); 440 if (rv != KERN_SUCCESS) { 441 vmem_free(arena, addr, asize); 442 return (0); 443 } 444 kasan_mark((void *)addr, size, asize, KASAN_KMEM_REDZONE); 445 return ((void *)addr); 446 } 447 448 void * 449 kmem_malloc(vm_size_t size, int flags) 450 { 451 void * p; 452 453 TSENTER(); 454 p = kmem_malloc_domainset(DOMAINSET_RR(), size, flags); 455 TSEXIT(); 456 return (p); 457 } 458 459 void * 460 kmem_malloc_domainset(struct domainset *ds, vm_size_t size, int flags) 461 { 462 struct vm_domainset_iter di; 463 void *addr; 464 int domain; 465 466 vm_domainset_iter_policy_init(&di, ds, &domain, &flags); 467 do { 468 addr = kmem_malloc_domain(domain, size, flags); 469 if (addr != NULL) 470 break; 471 } while (vm_domainset_iter_policy(&di, &domain) == 0); 472 473 return (addr); 474 } 475 476 /* 477 * kmem_back_domain: 478 * 479 * Allocate physical pages from the specified domain for the specified 480 * virtual address range. 481 */ 482 int 483 kmem_back_domain(int domain, vm_object_t object, vm_offset_t addr, 484 vm_size_t size, int flags) 485 { 486 vm_offset_t offset, i; 487 vm_page_t m, mpred; 488 vm_prot_t prot; 489 int pflags; 490 491 KASSERT(object == kernel_object, 492 ("kmem_back_domain: only supports kernel object.")); 493 494 offset = addr - VM_MIN_KERNEL_ADDRESS; 495 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED; 496 pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL); 497 if (flags & M_WAITOK) 498 pflags |= VM_ALLOC_WAITFAIL; 499 prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW; 500 501 i = 0; 502 VM_OBJECT_WLOCK(object); 503 retry: 504 mpred = vm_radix_lookup_le(&object->rtree, atop(offset + i)); 505 for (; i < size; i += PAGE_SIZE, mpred = m) { 506 m = vm_page_alloc_domain_after(object, atop(offset + i), 507 domain, pflags, mpred); 508 509 /* 510 * Ran out of space, free everything up and return. Don't need 511 * to lock page queues here as we know that the pages we got 512 * aren't on any queues. 513 */ 514 if (m == NULL) { 515 if ((flags & M_NOWAIT) == 0) 516 goto retry; 517 VM_OBJECT_WUNLOCK(object); 518 kmem_unback(object, addr, i); 519 return (KERN_NO_SPACE); 520 } 521 KASSERT(vm_page_domain(m) == domain, 522 ("kmem_back_domain: Domain mismatch %d != %d", 523 vm_page_domain(m), domain)); 524 if (flags & M_ZERO && (m->flags & PG_ZERO) == 0) 525 pmap_zero_page(m); 526 KASSERT((m->oflags & VPO_UNMANAGED) != 0, 527 ("kmem_malloc: page %p is managed", m)); 528 vm_page_valid(m); 529 pmap_enter(kernel_pmap, addr + i, m, prot, 530 prot | PMAP_ENTER_WIRED, 0); 531 if (__predict_false((prot & VM_PROT_EXECUTE) != 0)) 532 m->oflags |= VPO_KMEM_EXEC; 533 } 534 VM_OBJECT_WUNLOCK(object); 535 kmem_alloc_san(addr, size, size, flags); 536 return (KERN_SUCCESS); 537 } 538 539 /* 540 * kmem_back: 541 * 542 * Allocate physical pages for the specified virtual address range. 543 */ 544 int 545 kmem_back(vm_object_t object, vm_offset_t addr, vm_size_t size, int flags) 546 { 547 vm_offset_t end, next, start; 548 int domain, rv; 549 550 KASSERT(object == kernel_object, 551 ("kmem_back: only supports kernel object.")); 552 553 for (start = addr, end = addr + size; addr < end; addr = next) { 554 /* 555 * We must ensure that pages backing a given large virtual page 556 * all come from the same physical domain. 557 */ 558 if (vm_ndomains > 1) { 559 domain = (addr >> KVA_QUANTUM_SHIFT) % vm_ndomains; 560 while (VM_DOMAIN_EMPTY(domain)) 561 domain++; 562 next = roundup2(addr + 1, KVA_QUANTUM); 563 if (next > end || next < start) 564 next = end; 565 } else { 566 domain = 0; 567 next = end; 568 } 569 rv = kmem_back_domain(domain, object, addr, next - addr, flags); 570 if (rv != KERN_SUCCESS) { 571 kmem_unback(object, start, addr - start); 572 break; 573 } 574 } 575 return (rv); 576 } 577 578 /* 579 * kmem_unback: 580 * 581 * Unmap and free the physical pages underlying the specified virtual 582 * address range. 583 * 584 * A physical page must exist within the specified object at each index 585 * that is being unmapped. 586 */ 587 static struct vmem * 588 _kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size) 589 { 590 struct vmem *arena; 591 vm_page_t m, next; 592 vm_offset_t end, offset; 593 int domain; 594 595 KASSERT(object == kernel_object, 596 ("kmem_unback: only supports kernel object.")); 597 598 if (size == 0) 599 return (NULL); 600 pmap_remove(kernel_pmap, addr, addr + size); 601 offset = addr - VM_MIN_KERNEL_ADDRESS; 602 end = offset + size; 603 VM_OBJECT_WLOCK(object); 604 m = vm_page_lookup(object, atop(offset)); 605 domain = vm_page_domain(m); 606 if (__predict_true((m->oflags & VPO_KMEM_EXEC) == 0)) 607 arena = vm_dom[domain].vmd_kernel_arena; 608 else 609 arena = vm_dom[domain].vmd_kernel_rwx_arena; 610 for (; offset < end; offset += PAGE_SIZE, m = next) { 611 next = vm_page_next(m); 612 vm_page_xbusy_claim(m); 613 vm_page_unwire_noq(m); 614 vm_page_free(m); 615 } 616 VM_OBJECT_WUNLOCK(object); 617 618 return (arena); 619 } 620 621 void 622 kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size) 623 { 624 625 (void)_kmem_unback(object, addr, size); 626 } 627 628 /* 629 * kmem_free: 630 * 631 * Free memory allocated with kmem_malloc. The size must match the 632 * original allocation. 633 */ 634 void 635 kmem_free(void *addr, vm_size_t size) 636 { 637 struct vmem *arena; 638 639 size = round_page(size); 640 kasan_mark(addr, size, size, 0); 641 arena = _kmem_unback(kernel_object, (uintptr_t)addr, size); 642 if (arena != NULL) 643 vmem_free(arena, (uintptr_t)addr, size); 644 } 645 646 /* 647 * kmap_alloc_wait: 648 * 649 * Allocates pageable memory from a sub-map of the kernel. If the submap 650 * has no room, the caller sleeps waiting for more memory in the submap. 651 * 652 * This routine may block. 653 */ 654 vm_offset_t 655 kmap_alloc_wait(vm_map_t map, vm_size_t size) 656 { 657 vm_offset_t addr; 658 659 size = round_page(size); 660 if (!swap_reserve(size)) 661 return (0); 662 663 for (;;) { 664 /* 665 * To make this work for more than one map, use the map's lock 666 * to lock out sleepers/wakers. 667 */ 668 vm_map_lock(map); 669 addr = vm_map_findspace(map, vm_map_min(map), size); 670 if (addr + size <= vm_map_max(map)) 671 break; 672 /* no space now; see if we can ever get space */ 673 if (vm_map_max(map) - vm_map_min(map) < size) { 674 vm_map_unlock(map); 675 swap_release(size); 676 return (0); 677 } 678 map->needs_wakeup = TRUE; 679 vm_map_unlock_and_wait(map, 0); 680 } 681 vm_map_insert(map, NULL, 0, addr, addr + size, VM_PROT_RW, VM_PROT_RW, 682 MAP_ACC_CHARGED); 683 vm_map_unlock(map); 684 return (addr); 685 } 686 687 /* 688 * kmap_free_wakeup: 689 * 690 * Returns memory to a submap of the kernel, and wakes up any processes 691 * waiting for memory in that map. 692 */ 693 void 694 kmap_free_wakeup(vm_map_t map, vm_offset_t addr, vm_size_t size) 695 { 696 697 vm_map_lock(map); 698 (void) vm_map_delete(map, trunc_page(addr), round_page(addr + size)); 699 if (map->needs_wakeup) { 700 map->needs_wakeup = FALSE; 701 vm_map_wakeup(map); 702 } 703 vm_map_unlock(map); 704 } 705 706 void 707 kmem_init_zero_region(void) 708 { 709 vm_offset_t addr, i; 710 vm_page_t m; 711 712 /* 713 * Map a single physical page of zeros to a larger virtual range. 714 * This requires less looping in places that want large amounts of 715 * zeros, while not using much more physical resources. 716 */ 717 addr = kva_alloc(ZERO_REGION_SIZE); 718 m = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_ZERO); 719 for (i = 0; i < ZERO_REGION_SIZE; i += PAGE_SIZE) 720 pmap_qenter(addr + i, &m, 1); 721 pmap_protect(kernel_pmap, addr, addr + ZERO_REGION_SIZE, VM_PROT_READ); 722 723 zero_region = (const void *)addr; 724 } 725 726 /* 727 * Import KVA from the kernel map into the kernel arena. 728 */ 729 static int 730 kva_import(void *unused, vmem_size_t size, int flags, vmem_addr_t *addrp) 731 { 732 vm_offset_t addr; 733 int result; 734 735 TSENTER(); 736 KASSERT((size % KVA_QUANTUM) == 0, 737 ("kva_import: Size %jd is not a multiple of %d", 738 (intmax_t)size, (int)KVA_QUANTUM)); 739 addr = vm_map_min(kernel_map); 740 result = vm_map_find(kernel_map, NULL, 0, &addr, size, 0, 741 VMFS_SUPER_SPACE, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 742 if (result != KERN_SUCCESS) { 743 TSEXIT(); 744 return (ENOMEM); 745 } 746 747 *addrp = addr; 748 749 TSEXIT(); 750 return (0); 751 } 752 753 /* 754 * Import KVA from a parent arena into a per-domain arena. Imports must be 755 * KVA_QUANTUM-aligned and a multiple of KVA_QUANTUM in size. 756 */ 757 static int 758 kva_import_domain(void *arena, vmem_size_t size, int flags, vmem_addr_t *addrp) 759 { 760 761 KASSERT((size % KVA_QUANTUM) == 0, 762 ("kva_import_domain: Size %jd is not a multiple of %d", 763 (intmax_t)size, (int)KVA_QUANTUM)); 764 return (vmem_xalloc(arena, size, KVA_QUANTUM, 0, 0, VMEM_ADDR_MIN, 765 VMEM_ADDR_MAX, flags, addrp)); 766 } 767 768 /* 769 * kmem_init: 770 * 771 * Create the kernel map; insert a mapping covering kernel text, 772 * data, bss, and all space allocated thus far (`boostrap' data). The 773 * new map will thus map the range between VM_MIN_KERNEL_ADDRESS and 774 * `start' as allocated, and the range between `start' and `end' as free. 775 * Create the kernel vmem arena and its per-domain children. 776 */ 777 void 778 kmem_init(vm_offset_t start, vm_offset_t end) 779 { 780 vm_size_t quantum; 781 int domain; 782 783 vm_map_init(kernel_map, kernel_pmap, VM_MIN_KERNEL_ADDRESS, end); 784 kernel_map->system_map = 1; 785 vm_map_lock(kernel_map); 786 /* N.B.: cannot use kgdb to debug, starting with this assignment ... */ 787 (void)vm_map_insert(kernel_map, NULL, 0, 788 #ifdef __amd64__ 789 KERNBASE, 790 #else 791 VM_MIN_KERNEL_ADDRESS, 792 #endif 793 start, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 794 /* ... and ending with the completion of the above `insert' */ 795 796 #ifdef __amd64__ 797 /* 798 * Mark KVA used for the page array as allocated. Other platforms 799 * that handle vm_page_array allocation can simply adjust virtual_avail 800 * instead. 801 */ 802 (void)vm_map_insert(kernel_map, NULL, 0, (vm_offset_t)vm_page_array, 803 (vm_offset_t)vm_page_array + round_2mpage(vm_page_array_size * 804 sizeof(struct vm_page)), 805 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT); 806 #endif 807 vm_map_unlock(kernel_map); 808 809 /* 810 * Use a large import quantum on NUMA systems. This helps minimize 811 * interleaving of superpages, reducing internal fragmentation within 812 * the per-domain arenas. 813 */ 814 if (vm_ndomains > 1 && PMAP_HAS_DMAP) 815 quantum = KVA_NUMA_IMPORT_QUANTUM; 816 else 817 quantum = KVA_QUANTUM; 818 819 /* 820 * Initialize the kernel_arena. This can grow on demand. 821 */ 822 vmem_init(kernel_arena, "kernel arena", 0, 0, PAGE_SIZE, 0, 0); 823 vmem_set_import(kernel_arena, kva_import, NULL, NULL, quantum); 824 825 for (domain = 0; domain < vm_ndomains; domain++) { 826 /* 827 * Initialize the per-domain arenas. These are used to color 828 * the KVA space in a way that ensures that virtual large pages 829 * are backed by memory from the same physical domain, 830 * maximizing the potential for superpage promotion. 831 */ 832 vm_dom[domain].vmd_kernel_arena = vmem_create( 833 "kernel arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK); 834 vmem_set_import(vm_dom[domain].vmd_kernel_arena, 835 kva_import_domain, NULL, kernel_arena, quantum); 836 837 /* 838 * In architectures with superpages, maintain separate arenas 839 * for allocations with permissions that differ from the 840 * "standard" read/write permissions used for kernel memory, 841 * so as not to inhibit superpage promotion. 842 * 843 * Use the base import quantum since this arena is rarely used. 844 */ 845 #if VM_NRESERVLEVEL > 0 846 vm_dom[domain].vmd_kernel_rwx_arena = vmem_create( 847 "kernel rwx arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK); 848 vmem_set_import(vm_dom[domain].vmd_kernel_rwx_arena, 849 kva_import_domain, (vmem_release_t *)vmem_xfree, 850 kernel_arena, KVA_QUANTUM); 851 #else 852 vm_dom[domain].vmd_kernel_rwx_arena = 853 vm_dom[domain].vmd_kernel_arena; 854 #endif 855 } 856 857 /* 858 * This must be the very first call so that the virtual address 859 * space used for early allocations is properly marked used in 860 * the map. 861 */ 862 uma_startup2(); 863 } 864 865 /* 866 * kmem_bootstrap_free: 867 * 868 * Free pages backing preloaded data (e.g., kernel modules) to the 869 * system. Currently only supported on platforms that create a 870 * vm_phys segment for preloaded data. 871 */ 872 void 873 kmem_bootstrap_free(vm_offset_t start, vm_size_t size) 874 { 875 #if defined(__i386__) || defined(__amd64__) 876 struct vm_domain *vmd; 877 vm_offset_t end, va; 878 vm_paddr_t pa; 879 vm_page_t m; 880 881 end = trunc_page(start + size); 882 start = round_page(start); 883 884 #ifdef __amd64__ 885 /* 886 * Preloaded files do not have execute permissions by default on amd64. 887 * Restore the default permissions to ensure that the direct map alias 888 * is updated. 889 */ 890 pmap_change_prot(start, end - start, VM_PROT_RW); 891 #endif 892 for (va = start; va < end; va += PAGE_SIZE) { 893 pa = pmap_kextract(va); 894 m = PHYS_TO_VM_PAGE(pa); 895 896 vmd = vm_pagequeue_domain(m); 897 vm_domain_free_lock(vmd); 898 vm_phys_free_pages(m, 0); 899 vm_domain_free_unlock(vmd); 900 901 vm_domain_freecnt_inc(vmd, 1); 902 vm_cnt.v_page_count++; 903 } 904 pmap_remove(kernel_pmap, start, end); 905 (void)vmem_add(kernel_arena, start, end - start, M_WAITOK); 906 #endif 907 } 908 909 #ifdef PMAP_WANT_ACTIVE_CPUS_NAIVE 910 void 911 pmap_active_cpus(pmap_t pmap, cpuset_t *res) 912 { 913 struct thread *td; 914 struct proc *p; 915 struct vmspace *vm; 916 int c; 917 918 CPU_ZERO(res); 919 CPU_FOREACH(c) { 920 td = cpuid_to_pcpu[c]->pc_curthread; 921 p = td->td_proc; 922 if (p == NULL) 923 continue; 924 vm = vmspace_acquire_ref(p); 925 if (vm == NULL) 926 continue; 927 if (pmap == vmspace_pmap(vm)) 928 CPU_SET(c, res); 929 vmspace_free(vm); 930 } 931 } 932 #endif 933 934 /* 935 * Allow userspace to directly trigger the VM drain routine for testing 936 * purposes. 937 */ 938 static int 939 debug_vm_lowmem(SYSCTL_HANDLER_ARGS) 940 { 941 int error, i; 942 943 i = 0; 944 error = sysctl_handle_int(oidp, &i, 0, req); 945 if (error != 0) 946 return (error); 947 if ((i & ~(VM_LOW_KMEM | VM_LOW_PAGES)) != 0) 948 return (EINVAL); 949 if (i != 0) 950 EVENTHANDLER_INVOKE(vm_lowmem, i); 951 return (0); 952 } 953 SYSCTL_PROC(_debug, OID_AUTO, vm_lowmem, 954 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, debug_vm_lowmem, "I", 955 "set to trigger vm_lowmem event with given flags"); 956 957 static int 958 debug_uma_reclaim(SYSCTL_HANDLER_ARGS) 959 { 960 int error, i; 961 962 i = 0; 963 error = sysctl_handle_int(oidp, &i, 0, req); 964 if (error != 0 || req->newptr == NULL) 965 return (error); 966 if (i != UMA_RECLAIM_TRIM && i != UMA_RECLAIM_DRAIN && 967 i != UMA_RECLAIM_DRAIN_CPU) 968 return (EINVAL); 969 uma_reclaim(i); 970 return (0); 971 } 972 SYSCTL_PROC(_debug, OID_AUTO, uma_reclaim, 973 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, debug_uma_reclaim, "I", 974 "set to generate request to reclaim uma caches"); 975 976 static int 977 debug_uma_reclaim_domain(SYSCTL_HANDLER_ARGS) 978 { 979 int domain, error, request; 980 981 request = 0; 982 error = sysctl_handle_int(oidp, &request, 0, req); 983 if (error != 0 || req->newptr == NULL) 984 return (error); 985 986 domain = request >> 4; 987 request &= 0xf; 988 if (request != UMA_RECLAIM_TRIM && request != UMA_RECLAIM_DRAIN && 989 request != UMA_RECLAIM_DRAIN_CPU) 990 return (EINVAL); 991 if (domain < 0 || domain >= vm_ndomains) 992 return (EINVAL); 993 uma_reclaim_domain(request, domain); 994 return (0); 995 } 996 SYSCTL_PROC(_debug, OID_AUTO, uma_reclaim_domain, 997 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, 998 debug_uma_reclaim_domain, "I", 999 ""); 1000