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 [email protected]
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/smp.h>
82 #include <sys/sysctl.h>
83 #include <sys/vmem.h>
84 #include <sys/vmmeter.h>
85
86 #include <vm/vm.h>
87 #include <vm/vm_param.h>
88 #include <vm/vm_domainset.h>
89 #include <vm/vm_kern.h>
90 #include <vm/pmap.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_page.h>
94 #include <vm/vm_pageout.h>
95 #include <vm/vm_pagequeue.h>
96 #include <vm/vm_phys.h>
97 #include <vm/vm_radix.h>
98 #include <vm/vm_extern.h>
99 #include <vm/uma.h>
100
101 struct vm_map kernel_map_store;
102 struct vm_map exec_map_store;
103 struct vm_map pipe_map_store;
104
105 const void *zero_region;
106 CTASSERT((ZERO_REGION_SIZE & PAGE_MASK) == 0);
107
108 /* NB: Used by kernel debuggers. */
109 const u_long vm_maxuser_address = VM_MAXUSER_ADDRESS;
110
111 u_int exec_map_entry_size;
112 u_int exec_map_entries;
113
114 SYSCTL_ULONG(_vm, OID_AUTO, min_kernel_address, CTLFLAG_RD,
115 SYSCTL_NULL_ULONG_PTR, VM_MIN_KERNEL_ADDRESS, "Min kernel address");
116
117 SYSCTL_ULONG(_vm, OID_AUTO, max_kernel_address, CTLFLAG_RD,
118 #if defined(__arm__)
119 &vm_max_kernel_address, 0,
120 #else
121 SYSCTL_NULL_ULONG_PTR, VM_MAX_KERNEL_ADDRESS,
122 #endif
123 "Max kernel address");
124
125 #if VM_NRESERVLEVEL > 0
126 #define KVA_QUANTUM_SHIFT (VM_LEVEL_0_ORDER + PAGE_SHIFT)
127 #else
128 /* On non-superpage architectures we want large import sizes. */
129 #define KVA_QUANTUM_SHIFT (8 + PAGE_SHIFT)
130 #endif
131 #define KVA_QUANTUM (1ul << KVA_QUANTUM_SHIFT)
132 #define KVA_NUMA_IMPORT_QUANTUM (KVA_QUANTUM * 128)
133
134 extern void uma_startup2(void);
135
136 /*
137 * kva_alloc:
138 *
139 * Allocate a virtual address range with no underlying object and
140 * no initial mapping to physical memory. Any mapping from this
141 * range to physical memory must be explicitly created prior to
142 * its use, typically with pmap_qenter(). Any attempt to create
143 * a mapping on demand through vm_fault() will result in a panic.
144 */
145 vm_offset_t
kva_alloc(vm_size_t size)146 kva_alloc(vm_size_t size)
147 {
148 vm_offset_t addr;
149
150 TSENTER();
151 size = round_page(size);
152 if (vmem_xalloc(kernel_arena, size, 0, 0, 0, VMEM_ADDR_MIN,
153 VMEM_ADDR_MAX, M_BESTFIT | M_NOWAIT, &addr))
154 return (0);
155 TSEXIT();
156
157 return (addr);
158 }
159
160 /*
161 * kva_alloc_aligned:
162 *
163 * Allocate a virtual address range as in kva_alloc where the base
164 * address is aligned to align.
165 */
166 vm_offset_t
kva_alloc_aligned(vm_size_t size,vm_size_t align)167 kva_alloc_aligned(vm_size_t size, vm_size_t align)
168 {
169 vm_offset_t addr;
170
171 TSENTER();
172 size = round_page(size);
173 if (vmem_xalloc(kernel_arena, size, align, 0, 0, VMEM_ADDR_MIN,
174 VMEM_ADDR_MAX, M_BESTFIT | M_NOWAIT, &addr))
175 return (0);
176 TSEXIT();
177
178 return (addr);
179 }
180
181 /*
182 * kva_free:
183 *
184 * Release a region of kernel virtual memory allocated
185 * with kva_alloc, and return the physical pages
186 * associated with that region.
187 *
188 * This routine may not block on kernel maps.
189 */
190 void
kva_free(vm_offset_t addr,vm_size_t size)191 kva_free(vm_offset_t addr, vm_size_t size)
192 {
193
194 size = round_page(size);
195 vmem_xfree(kernel_arena, addr, size);
196 }
197
198 /*
199 * Update sanitizer shadow state to reflect a new allocation. Force inlining to
200 * help make KMSAN origin tracking more precise.
201 */
202 static __always_inline void
kmem_alloc_san(vm_offset_t addr,vm_size_t size,vm_size_t asize,int flags)203 kmem_alloc_san(vm_offset_t addr, vm_size_t size, vm_size_t asize, int flags)
204 {
205 if ((flags & M_ZERO) == 0) {
206 kmsan_mark((void *)addr, asize, KMSAN_STATE_UNINIT);
207 kmsan_orig((void *)addr, asize, KMSAN_TYPE_KMEM,
208 KMSAN_RET_ADDR);
209 } else {
210 kmsan_mark((void *)addr, asize, KMSAN_STATE_INITED);
211 }
212 kasan_mark((void *)addr, size, asize, KASAN_KMEM_REDZONE);
213 }
214
215 static vm_page_t
kmem_alloc_contig_pages(vm_object_t object,vm_pindex_t pindex,int domain,int pflags,u_long npages,vm_paddr_t low,vm_paddr_t high,u_long alignment,vm_paddr_t boundary,vm_memattr_t memattr)216 kmem_alloc_contig_pages(vm_object_t object, vm_pindex_t pindex, int domain,
217 int pflags, u_long npages, vm_paddr_t low, vm_paddr_t high,
218 u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr)
219 {
220 vm_page_t m;
221 int tries;
222 bool wait, reclaim;
223
224 VM_OBJECT_ASSERT_WLOCKED(object);
225
226 wait = (pflags & VM_ALLOC_WAITOK) != 0;
227 reclaim = (pflags & VM_ALLOC_NORECLAIM) == 0;
228 pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
229 pflags |= VM_ALLOC_NOWAIT;
230 for (tries = wait ? 3 : 1;; tries--) {
231 m = vm_page_alloc_contig_domain(object, pindex, domain, pflags,
232 npages, low, high, alignment, boundary, memattr);
233 if (m != NULL || tries == 0 || !reclaim)
234 break;
235
236 VM_OBJECT_WUNLOCK(object);
237 if (!vm_page_reclaim_contig_domain(domain, pflags, npages,
238 low, high, alignment, boundary) && wait)
239 vm_wait_domain(domain);
240 VM_OBJECT_WLOCK(object);
241 }
242 return (m);
243 }
244
245 /*
246 * Allocates a region from the kernel address map and physical pages
247 * within the specified address range to the kernel object. Creates a
248 * wired mapping from this region to these pages, and returns the
249 * region's starting virtual address. The allocated pages are not
250 * necessarily physically contiguous. If M_ZERO is specified through the
251 * given flags, then the pages are zeroed before they are mapped.
252 */
253 static void *
kmem_alloc_attr_domain(int domain,vm_size_t size,int flags,vm_paddr_t low,vm_paddr_t high,vm_memattr_t memattr)254 kmem_alloc_attr_domain(int domain, vm_size_t size, int flags, vm_paddr_t low,
255 vm_paddr_t high, vm_memattr_t memattr)
256 {
257 vmem_t *vmem;
258 vm_object_t object;
259 vm_offset_t addr, i, offset;
260 vm_page_t m;
261 vm_size_t asize;
262 int pflags;
263 vm_prot_t prot;
264
265 object = kernel_object;
266 asize = round_page(size);
267 vmem = vm_dom[domain].vmd_kernel_arena;
268 if (vmem_alloc(vmem, asize, M_BESTFIT | flags, &addr))
269 return (0);
270 offset = addr - VM_MIN_KERNEL_ADDRESS;
271 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
272 prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW;
273 VM_OBJECT_WLOCK(object);
274 for (i = 0; i < asize; i += PAGE_SIZE) {
275 m = kmem_alloc_contig_pages(object, atop(offset + i),
276 domain, pflags, 1, low, high, PAGE_SIZE, 0, memattr);
277 if (m == NULL) {
278 VM_OBJECT_WUNLOCK(object);
279 kmem_unback(object, addr, i);
280 vmem_free(vmem, addr, asize);
281 return (0);
282 }
283 KASSERT(vm_page_domain(m) == domain,
284 ("kmem_alloc_attr_domain: Domain mismatch %d != %d",
285 vm_page_domain(m), domain));
286 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
287 pmap_zero_page(m);
288 vm_page_valid(m);
289 pmap_enter(kernel_pmap, addr + i, m, prot,
290 prot | PMAP_ENTER_WIRED, 0);
291 }
292 VM_OBJECT_WUNLOCK(object);
293 kmem_alloc_san(addr, size, asize, flags);
294 return ((void *)addr);
295 }
296
297 void *
kmem_alloc_attr(vm_size_t size,int flags,vm_paddr_t low,vm_paddr_t high,vm_memattr_t memattr)298 kmem_alloc_attr(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high,
299 vm_memattr_t memattr)
300 {
301
302 return (kmem_alloc_attr_domainset(DOMAINSET_RR(), size, flags, low,
303 high, memattr));
304 }
305
306 void *
kmem_alloc_attr_domainset(struct domainset * ds,vm_size_t size,int flags,vm_paddr_t low,vm_paddr_t high,vm_memattr_t memattr)307 kmem_alloc_attr_domainset(struct domainset *ds, vm_size_t size, int flags,
308 vm_paddr_t low, vm_paddr_t high, vm_memattr_t memattr)
309 {
310 struct vm_domainset_iter di;
311 void *addr;
312 int domain;
313
314 vm_domainset_iter_policy_init(&di, ds, &domain, &flags);
315 do {
316 addr = kmem_alloc_attr_domain(domain, size, flags, low, high,
317 memattr);
318 if (addr != NULL)
319 break;
320 } while (vm_domainset_iter_policy(&di, &domain) == 0);
321
322 return (addr);
323 }
324
325 /*
326 * Allocates a region from the kernel address map and physically
327 * contiguous pages within the specified address range to the kernel
328 * object. Creates a wired mapping from this region to these pages, and
329 * returns the region's starting virtual address. If M_ZERO is specified
330 * through the given flags, then the pages are zeroed before they are
331 * mapped.
332 */
333 static void *
kmem_alloc_contig_domain(int domain,vm_size_t size,int flags,vm_paddr_t low,vm_paddr_t high,u_long alignment,vm_paddr_t boundary,vm_memattr_t memattr)334 kmem_alloc_contig_domain(int domain, vm_size_t size, int flags, vm_paddr_t low,
335 vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
336 vm_memattr_t memattr)
337 {
338 vmem_t *vmem;
339 vm_object_t object;
340 vm_offset_t addr, offset, tmp;
341 vm_page_t end_m, m;
342 vm_size_t asize;
343 u_long npages;
344 int pflags;
345
346 object = kernel_object;
347 asize = round_page(size);
348 vmem = vm_dom[domain].vmd_kernel_arena;
349 if (vmem_alloc(vmem, asize, flags | M_BESTFIT, &addr))
350 return (NULL);
351 offset = addr - VM_MIN_KERNEL_ADDRESS;
352 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
353 npages = atop(asize);
354 VM_OBJECT_WLOCK(object);
355 m = kmem_alloc_contig_pages(object, atop(offset), domain,
356 pflags, npages, low, high, alignment, boundary, memattr);
357 if (m == NULL) {
358 VM_OBJECT_WUNLOCK(object);
359 vmem_free(vmem, addr, asize);
360 return (NULL);
361 }
362 KASSERT(vm_page_domain(m) == domain,
363 ("kmem_alloc_contig_domain: Domain mismatch %d != %d",
364 vm_page_domain(m), domain));
365 end_m = m + npages;
366 tmp = addr;
367 for (; m < end_m; m++) {
368 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
369 pmap_zero_page(m);
370 vm_page_valid(m);
371 pmap_enter(kernel_pmap, tmp, m, VM_PROT_RW,
372 VM_PROT_RW | PMAP_ENTER_WIRED, 0);
373 tmp += PAGE_SIZE;
374 }
375 VM_OBJECT_WUNLOCK(object);
376 kmem_alloc_san(addr, size, asize, flags);
377 return ((void *)addr);
378 }
379
380 void *
kmem_alloc_contig(vm_size_t size,int flags,vm_paddr_t low,vm_paddr_t high,u_long alignment,vm_paddr_t boundary,vm_memattr_t memattr)381 kmem_alloc_contig(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high,
382 u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr)
383 {
384
385 return (kmem_alloc_contig_domainset(DOMAINSET_RR(), size, flags, low,
386 high, alignment, boundary, memattr));
387 }
388
389 void *
kmem_alloc_contig_domainset(struct domainset * ds,vm_size_t size,int flags,vm_paddr_t low,vm_paddr_t high,u_long alignment,vm_paddr_t boundary,vm_memattr_t memattr)390 kmem_alloc_contig_domainset(struct domainset *ds, vm_size_t size, int flags,
391 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
392 vm_memattr_t memattr)
393 {
394 struct vm_domainset_iter di;
395 void *addr;
396 int domain;
397
398 vm_domainset_iter_policy_init(&di, ds, &domain, &flags);
399 do {
400 addr = kmem_alloc_contig_domain(domain, size, flags, low, high,
401 alignment, boundary, memattr);
402 if (addr != NULL)
403 break;
404 } while (vm_domainset_iter_policy(&di, &domain) == 0);
405
406 return (addr);
407 }
408
409 /*
410 * kmem_subinit:
411 *
412 * Initializes a map to manage a subrange
413 * of the kernel virtual address space.
414 *
415 * Arguments are as follows:
416 *
417 * parent Map to take range from
418 * min, max Returned endpoints of map
419 * size Size of range to find
420 * superpage_align Request that min is superpage aligned
421 */
422 void
kmem_subinit(vm_map_t map,vm_map_t parent,vm_offset_t * min,vm_offset_t * max,vm_size_t size,bool superpage_align)423 kmem_subinit(vm_map_t map, vm_map_t parent, vm_offset_t *min, vm_offset_t *max,
424 vm_size_t size, bool superpage_align)
425 {
426 int ret;
427
428 size = round_page(size);
429
430 *min = vm_map_min(parent);
431 ret = vm_map_find(parent, NULL, 0, min, size, 0, superpage_align ?
432 VMFS_SUPER_SPACE : VMFS_ANY_SPACE, VM_PROT_ALL, VM_PROT_ALL,
433 MAP_ACC_NO_CHARGE);
434 if (ret != KERN_SUCCESS)
435 panic("kmem_subinit: bad status return of %d", ret);
436 *max = *min + size;
437 vm_map_init(map, vm_map_pmap(parent), *min, *max);
438 if (vm_map_submap(parent, *min, *max, map) != KERN_SUCCESS)
439 panic("kmem_subinit: unable to change range to submap");
440 }
441
442 /*
443 * kmem_malloc_domain:
444 *
445 * Allocate wired-down pages in the kernel's address space.
446 */
447 static void *
kmem_malloc_domain(int domain,vm_size_t size,int flags)448 kmem_malloc_domain(int domain, vm_size_t size, int flags)
449 {
450 vmem_t *arena;
451 vm_offset_t addr;
452 vm_size_t asize;
453 int rv;
454
455 if (__predict_true((flags & M_EXEC) == 0))
456 arena = vm_dom[domain].vmd_kernel_arena;
457 else
458 arena = vm_dom[domain].vmd_kernel_rwx_arena;
459 asize = round_page(size);
460 if (vmem_alloc(arena, asize, flags | M_BESTFIT, &addr))
461 return (0);
462
463 rv = kmem_back_domain(domain, kernel_object, addr, asize, flags);
464 if (rv != KERN_SUCCESS) {
465 vmem_free(arena, addr, asize);
466 return (0);
467 }
468 kasan_mark((void *)addr, size, asize, KASAN_KMEM_REDZONE);
469 return ((void *)addr);
470 }
471
472 void *
kmem_malloc(vm_size_t size,int flags)473 kmem_malloc(vm_size_t size, int flags)
474 {
475 void * p;
476
477 TSENTER();
478 p = kmem_malloc_domainset(DOMAINSET_RR(), size, flags);
479 TSEXIT();
480 return (p);
481 }
482
483 void *
kmem_malloc_domainset(struct domainset * ds,vm_size_t size,int flags)484 kmem_malloc_domainset(struct domainset *ds, vm_size_t size, int flags)
485 {
486 struct vm_domainset_iter di;
487 void *addr;
488 int domain;
489
490 vm_domainset_iter_policy_init(&di, ds, &domain, &flags);
491 do {
492 addr = kmem_malloc_domain(domain, size, flags);
493 if (addr != NULL)
494 break;
495 } while (vm_domainset_iter_policy(&di, &domain) == 0);
496
497 return (addr);
498 }
499
500 /*
501 * kmem_back_domain:
502 *
503 * Allocate physical pages from the specified domain for the specified
504 * virtual address range.
505 */
506 int
kmem_back_domain(int domain,vm_object_t object,vm_offset_t addr,vm_size_t size,int flags)507 kmem_back_domain(int domain, vm_object_t object, vm_offset_t addr,
508 vm_size_t size, int flags)
509 {
510 vm_offset_t offset, i;
511 vm_page_t m, mpred;
512 vm_prot_t prot;
513 int pflags;
514
515 KASSERT(object == kernel_object,
516 ("kmem_back_domain: only supports kernel object."));
517
518 offset = addr - VM_MIN_KERNEL_ADDRESS;
519 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
520 pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
521 if (flags & M_WAITOK)
522 pflags |= VM_ALLOC_WAITFAIL;
523 prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW;
524
525 i = 0;
526 VM_OBJECT_WLOCK(object);
527 retry:
528 mpred = vm_radix_lookup_le(&object->rtree, atop(offset + i));
529 for (; i < size; i += PAGE_SIZE, mpred = m) {
530 m = vm_page_alloc_domain_after(object, atop(offset + i),
531 domain, pflags, mpred);
532
533 /*
534 * Ran out of space, free everything up and return. Don't need
535 * to lock page queues here as we know that the pages we got
536 * aren't on any queues.
537 */
538 if (m == NULL) {
539 if ((flags & M_NOWAIT) == 0)
540 goto retry;
541 VM_OBJECT_WUNLOCK(object);
542 kmem_unback(object, addr, i);
543 return (KERN_NO_SPACE);
544 }
545 KASSERT(vm_page_domain(m) == domain,
546 ("kmem_back_domain: Domain mismatch %d != %d",
547 vm_page_domain(m), domain));
548 if (flags & M_ZERO && (m->flags & PG_ZERO) == 0)
549 pmap_zero_page(m);
550 KASSERT((m->oflags & VPO_UNMANAGED) != 0,
551 ("kmem_malloc: page %p is managed", m));
552 vm_page_valid(m);
553 pmap_enter(kernel_pmap, addr + i, m, prot,
554 prot | PMAP_ENTER_WIRED, 0);
555 if (__predict_false((prot & VM_PROT_EXECUTE) != 0))
556 m->oflags |= VPO_KMEM_EXEC;
557 }
558 VM_OBJECT_WUNLOCK(object);
559 kmem_alloc_san(addr, size, size, flags);
560 return (KERN_SUCCESS);
561 }
562
563 /*
564 * kmem_back:
565 *
566 * Allocate physical pages for the specified virtual address range.
567 */
568 int
kmem_back(vm_object_t object,vm_offset_t addr,vm_size_t size,int flags)569 kmem_back(vm_object_t object, vm_offset_t addr, vm_size_t size, int flags)
570 {
571 vm_offset_t end, next, start;
572 int domain, rv;
573
574 KASSERT(object == kernel_object,
575 ("kmem_back: only supports kernel object."));
576
577 for (start = addr, end = addr + size; addr < end; addr = next) {
578 /*
579 * We must ensure that pages backing a given large virtual page
580 * all come from the same physical domain.
581 */
582 if (vm_ndomains > 1) {
583 domain = (addr >> KVA_QUANTUM_SHIFT) % vm_ndomains;
584 while (VM_DOMAIN_EMPTY(domain))
585 domain++;
586 next = roundup2(addr + 1, KVA_QUANTUM);
587 if (next > end || next < start)
588 next = end;
589 } else {
590 domain = 0;
591 next = end;
592 }
593 rv = kmem_back_domain(domain, object, addr, next - addr, flags);
594 if (rv != KERN_SUCCESS) {
595 kmem_unback(object, start, addr - start);
596 break;
597 }
598 }
599 return (rv);
600 }
601
602 /*
603 * kmem_unback:
604 *
605 * Unmap and free the physical pages underlying the specified virtual
606 * address range.
607 *
608 * A physical page must exist within the specified object at each index
609 * that is being unmapped.
610 */
611 static struct vmem *
_kmem_unback(vm_object_t object,vm_offset_t addr,vm_size_t size)612 _kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size)
613 {
614 struct vmem *arena;
615 vm_page_t m, next;
616 vm_offset_t end, offset;
617 int domain;
618
619 KASSERT(object == kernel_object,
620 ("kmem_unback: only supports kernel object."));
621
622 if (size == 0)
623 return (NULL);
624 pmap_remove(kernel_pmap, addr, addr + size);
625 offset = addr - VM_MIN_KERNEL_ADDRESS;
626 end = offset + size;
627 VM_OBJECT_WLOCK(object);
628 m = vm_page_lookup(object, atop(offset));
629 domain = vm_page_domain(m);
630 if (__predict_true((m->oflags & VPO_KMEM_EXEC) == 0))
631 arena = vm_dom[domain].vmd_kernel_arena;
632 else
633 arena = vm_dom[domain].vmd_kernel_rwx_arena;
634 for (; offset < end; offset += PAGE_SIZE, m = next) {
635 next = vm_page_next(m);
636 vm_page_xbusy_claim(m);
637 vm_page_unwire_noq(m);
638 vm_page_free(m);
639 }
640 VM_OBJECT_WUNLOCK(object);
641
642 return (arena);
643 }
644
645 void
kmem_unback(vm_object_t object,vm_offset_t addr,vm_size_t size)646 kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size)
647 {
648
649 (void)_kmem_unback(object, addr, size);
650 }
651
652 /*
653 * kmem_free:
654 *
655 * Free memory allocated with kmem_malloc. The size must match the
656 * original allocation.
657 */
658 void
kmem_free(void * addr,vm_size_t size)659 kmem_free(void *addr, vm_size_t size)
660 {
661 struct vmem *arena;
662
663 size = round_page(size);
664 kasan_mark(addr, size, size, 0);
665 arena = _kmem_unback(kernel_object, (uintptr_t)addr, size);
666 if (arena != NULL)
667 vmem_free(arena, (uintptr_t)addr, size);
668 }
669
670 /*
671 * kmap_alloc_wait:
672 *
673 * Allocates pageable memory from a sub-map of the kernel. If the submap
674 * has no room, the caller sleeps waiting for more memory in the submap.
675 *
676 * This routine may block.
677 */
678 vm_offset_t
kmap_alloc_wait(vm_map_t map,vm_size_t size)679 kmap_alloc_wait(vm_map_t map, vm_size_t size)
680 {
681 vm_offset_t addr;
682
683 size = round_page(size);
684 if (!swap_reserve(size))
685 return (0);
686
687 for (;;) {
688 /*
689 * To make this work for more than one map, use the map's lock
690 * to lock out sleepers/wakers.
691 */
692 vm_map_lock(map);
693 addr = vm_map_findspace(map, vm_map_min(map), size);
694 if (addr + size <= vm_map_max(map))
695 break;
696 /* no space now; see if we can ever get space */
697 if (vm_map_max(map) - vm_map_min(map) < size) {
698 vm_map_unlock(map);
699 swap_release(size);
700 return (0);
701 }
702 map->needs_wakeup = TRUE;
703 vm_map_unlock_and_wait(map, 0);
704 }
705 vm_map_insert(map, NULL, 0, addr, addr + size, VM_PROT_RW, VM_PROT_RW,
706 MAP_ACC_CHARGED);
707 vm_map_unlock(map);
708 return (addr);
709 }
710
711 /*
712 * kmap_free_wakeup:
713 *
714 * Returns memory to a submap of the kernel, and wakes up any processes
715 * waiting for memory in that map.
716 */
717 void
kmap_free_wakeup(vm_map_t map,vm_offset_t addr,vm_size_t size)718 kmap_free_wakeup(vm_map_t map, vm_offset_t addr, vm_size_t size)
719 {
720
721 vm_map_lock(map);
722 (void) vm_map_delete(map, trunc_page(addr), round_page(addr + size));
723 if (map->needs_wakeup) {
724 map->needs_wakeup = FALSE;
725 vm_map_wakeup(map);
726 }
727 vm_map_unlock(map);
728 }
729
730 void
kmem_init_zero_region(void)731 kmem_init_zero_region(void)
732 {
733 vm_offset_t addr, i;
734 vm_page_t m;
735
736 /*
737 * Map a single physical page of zeros to a larger virtual range.
738 * This requires less looping in places that want large amounts of
739 * zeros, while not using much more physical resources.
740 */
741 addr = kva_alloc(ZERO_REGION_SIZE);
742 m = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_ZERO);
743 for (i = 0; i < ZERO_REGION_SIZE; i += PAGE_SIZE)
744 pmap_qenter(addr + i, &m, 1);
745 pmap_protect(kernel_pmap, addr, addr + ZERO_REGION_SIZE, VM_PROT_READ);
746
747 zero_region = (const void *)addr;
748 }
749
750 /*
751 * Import KVA from the kernel map into the kernel arena.
752 */
753 static int
kva_import(void * unused,vmem_size_t size,int flags,vmem_addr_t * addrp)754 kva_import(void *unused, vmem_size_t size, int flags, vmem_addr_t *addrp)
755 {
756 vm_offset_t addr;
757 int result;
758
759 TSENTER();
760 KASSERT((size % KVA_QUANTUM) == 0,
761 ("kva_import: Size %jd is not a multiple of %d",
762 (intmax_t)size, (int)KVA_QUANTUM));
763 addr = vm_map_min(kernel_map);
764 result = vm_map_find(kernel_map, NULL, 0, &addr, size, 0,
765 VMFS_SUPER_SPACE, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
766 if (result != KERN_SUCCESS) {
767 TSEXIT();
768 return (ENOMEM);
769 }
770
771 *addrp = addr;
772
773 TSEXIT();
774 return (0);
775 }
776
777 /*
778 * Import KVA from a parent arena into a per-domain arena. Imports must be
779 * KVA_QUANTUM-aligned and a multiple of KVA_QUANTUM in size.
780 */
781 static int
kva_import_domain(void * arena,vmem_size_t size,int flags,vmem_addr_t * addrp)782 kva_import_domain(void *arena, vmem_size_t size, int flags, vmem_addr_t *addrp)
783 {
784
785 KASSERT((size % KVA_QUANTUM) == 0,
786 ("kva_import_domain: Size %jd is not a multiple of %d",
787 (intmax_t)size, (int)KVA_QUANTUM));
788 return (vmem_xalloc(arena, size, KVA_QUANTUM, 0, 0, VMEM_ADDR_MIN,
789 VMEM_ADDR_MAX, flags, addrp));
790 }
791
792 /*
793 * kmem_init:
794 *
795 * Create the kernel map; insert a mapping covering kernel text,
796 * data, bss, and all space allocated thus far (`boostrap' data). The
797 * new map will thus map the range between VM_MIN_KERNEL_ADDRESS and
798 * `start' as allocated, and the range between `start' and `end' as free.
799 * Create the kernel vmem arena and its per-domain children.
800 */
801 void
kmem_init(vm_offset_t start,vm_offset_t end)802 kmem_init(vm_offset_t start, vm_offset_t end)
803 {
804 vm_size_t quantum;
805 int domain;
806
807 vm_map_init(kernel_map, kernel_pmap, VM_MIN_KERNEL_ADDRESS, end);
808 kernel_map->system_map = 1;
809 vm_map_lock(kernel_map);
810 /* N.B.: cannot use kgdb to debug, starting with this assignment ... */
811 (void)vm_map_insert(kernel_map, NULL, 0,
812 #ifdef __amd64__
813 KERNBASE,
814 #else
815 VM_MIN_KERNEL_ADDRESS,
816 #endif
817 start, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
818 /* ... and ending with the completion of the above `insert' */
819
820 #ifdef __amd64__
821 /*
822 * Mark KVA used for the page array as allocated. Other platforms
823 * that handle vm_page_array allocation can simply adjust virtual_avail
824 * instead.
825 */
826 (void)vm_map_insert(kernel_map, NULL, 0, (vm_offset_t)vm_page_array,
827 (vm_offset_t)vm_page_array + round_2mpage(vm_page_array_size *
828 sizeof(struct vm_page)),
829 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
830 #endif
831 vm_map_unlock(kernel_map);
832
833 /*
834 * Use a large import quantum on NUMA systems. This helps minimize
835 * interleaving of superpages, reducing internal fragmentation within
836 * the per-domain arenas.
837 */
838 if (vm_ndomains > 1 && PMAP_HAS_DMAP)
839 quantum = KVA_NUMA_IMPORT_QUANTUM;
840 else
841 quantum = KVA_QUANTUM;
842
843 /*
844 * Initialize the kernel_arena. This can grow on demand.
845 */
846 vmem_init(kernel_arena, "kernel arena", 0, 0, PAGE_SIZE, 0, 0);
847 vmem_set_import(kernel_arena, kva_import, NULL, NULL, quantum);
848
849 for (domain = 0; domain < vm_ndomains; domain++) {
850 /*
851 * Initialize the per-domain arenas. These are used to color
852 * the KVA space in a way that ensures that virtual large pages
853 * are backed by memory from the same physical domain,
854 * maximizing the potential for superpage promotion.
855 */
856 vm_dom[domain].vmd_kernel_arena = vmem_create(
857 "kernel arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
858 vmem_set_import(vm_dom[domain].vmd_kernel_arena,
859 kva_import_domain, NULL, kernel_arena, quantum);
860
861 /*
862 * In architectures with superpages, maintain separate arenas
863 * for allocations with permissions that differ from the
864 * "standard" read/write permissions used for kernel memory,
865 * so as not to inhibit superpage promotion.
866 *
867 * Use the base import quantum since this arena is rarely used.
868 */
869 #if VM_NRESERVLEVEL > 0
870 vm_dom[domain].vmd_kernel_rwx_arena = vmem_create(
871 "kernel rwx arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
872 vmem_set_import(vm_dom[domain].vmd_kernel_rwx_arena,
873 kva_import_domain, (vmem_release_t *)vmem_xfree,
874 kernel_arena, KVA_QUANTUM);
875 #else
876 vm_dom[domain].vmd_kernel_rwx_arena =
877 vm_dom[domain].vmd_kernel_arena;
878 #endif
879 }
880
881 /*
882 * This must be the very first call so that the virtual address
883 * space used for early allocations is properly marked used in
884 * the map.
885 */
886 uma_startup2();
887 }
888
889 /*
890 * kmem_bootstrap_free:
891 *
892 * Free pages backing preloaded data (e.g., kernel modules) to the
893 * system. Currently only supported on platforms that create a
894 * vm_phys segment for preloaded data.
895 */
896 void
kmem_bootstrap_free(vm_offset_t start,vm_size_t size)897 kmem_bootstrap_free(vm_offset_t start, vm_size_t size)
898 {
899 #if defined(__i386__) || defined(__amd64__)
900 struct vm_domain *vmd;
901 vm_offset_t end, va;
902 vm_paddr_t pa;
903 vm_page_t m;
904
905 end = trunc_page(start + size);
906 start = round_page(start);
907
908 #ifdef __amd64__
909 /*
910 * Preloaded files do not have execute permissions by default on amd64.
911 * Restore the default permissions to ensure that the direct map alias
912 * is updated.
913 */
914 pmap_change_prot(start, end - start, VM_PROT_RW);
915 #endif
916 for (va = start; va < end; va += PAGE_SIZE) {
917 pa = pmap_kextract(va);
918 m = PHYS_TO_VM_PAGE(pa);
919
920 vmd = vm_pagequeue_domain(m);
921 vm_domain_free_lock(vmd);
922 vm_phys_free_pages(m, 0);
923 vm_domain_free_unlock(vmd);
924
925 vm_domain_freecnt_inc(vmd, 1);
926 vm_cnt.v_page_count++;
927 }
928 pmap_remove(kernel_pmap, start, end);
929 (void)vmem_add(kernel_arena, start, end - start, M_WAITOK);
930 #endif
931 }
932
933 #ifdef PMAP_WANT_ACTIVE_CPUS_NAIVE
934 void
pmap_active_cpus(pmap_t pmap,cpuset_t * res)935 pmap_active_cpus(pmap_t pmap, cpuset_t *res)
936 {
937 struct thread *td;
938 struct proc *p;
939 struct vmspace *vm;
940 int c;
941
942 CPU_ZERO(res);
943 CPU_FOREACH(c) {
944 td = cpuid_to_pcpu[c]->pc_curthread;
945 p = td->td_proc;
946 if (p == NULL)
947 continue;
948 vm = vmspace_acquire_ref(p);
949 if (vm == NULL)
950 continue;
951 if (pmap == vmspace_pmap(vm))
952 CPU_SET(c, res);
953 vmspace_free(vm);
954 }
955 }
956 #endif
957
958 /*
959 * Allow userspace to directly trigger the VM drain routine for testing
960 * purposes.
961 */
962 static int
debug_vm_lowmem(SYSCTL_HANDLER_ARGS)963 debug_vm_lowmem(SYSCTL_HANDLER_ARGS)
964 {
965 int error, i;
966
967 i = 0;
968 error = sysctl_handle_int(oidp, &i, 0, req);
969 if (error != 0)
970 return (error);
971 if ((i & ~(VM_LOW_KMEM | VM_LOW_PAGES)) != 0)
972 return (EINVAL);
973 if (i != 0)
974 EVENTHANDLER_INVOKE(vm_lowmem, i);
975 return (0);
976 }
977 SYSCTL_PROC(_debug, OID_AUTO, vm_lowmem,
978 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, debug_vm_lowmem, "I",
979 "set to trigger vm_lowmem event with given flags");
980
981 static int
debug_uma_reclaim(SYSCTL_HANDLER_ARGS)982 debug_uma_reclaim(SYSCTL_HANDLER_ARGS)
983 {
984 int error, i;
985
986 i = 0;
987 error = sysctl_handle_int(oidp, &i, 0, req);
988 if (error != 0 || req->newptr == NULL)
989 return (error);
990 if (i != UMA_RECLAIM_TRIM && i != UMA_RECLAIM_DRAIN &&
991 i != UMA_RECLAIM_DRAIN_CPU)
992 return (EINVAL);
993 uma_reclaim(i);
994 return (0);
995 }
996 SYSCTL_PROC(_debug, OID_AUTO, uma_reclaim,
997 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, debug_uma_reclaim, "I",
998 "set to generate request to reclaim uma caches");
999
1000 static int
debug_uma_reclaim_domain(SYSCTL_HANDLER_ARGS)1001 debug_uma_reclaim_domain(SYSCTL_HANDLER_ARGS)
1002 {
1003 int domain, error, request;
1004
1005 request = 0;
1006 error = sysctl_handle_int(oidp, &request, 0, req);
1007 if (error != 0 || req->newptr == NULL)
1008 return (error);
1009
1010 domain = request >> 4;
1011 request &= 0xf;
1012 if (request != UMA_RECLAIM_TRIM && request != UMA_RECLAIM_DRAIN &&
1013 request != UMA_RECLAIM_DRAIN_CPU)
1014 return (EINVAL);
1015 if (domain < 0 || domain >= vm_ndomains)
1016 return (EINVAL);
1017 uma_reclaim_domain(request, domain);
1018 return (0);
1019 }
1020 SYSCTL_PROC(_debug, OID_AUTO, uma_reclaim_domain,
1021 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0,
1022 debug_uma_reclaim_domain, "I",
1023 "");
1024