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_guard_pages;
111 u_int exec_map_entries;
112
113 SYSCTL_ULONG(_vm, OID_AUTO, min_kernel_address, CTLFLAG_RD,
114 #if defined(__amd64__)
115 &kva_layout.km_low, 0,
116 #else
117 SYSCTL_NULL_ULONG_PTR, VM_MIN_KERNEL_ADDRESS,
118 #endif
119 "Min kernel address");
120
121 SYSCTL_ULONG(_vm, OID_AUTO, max_kernel_address, CTLFLAG_RD,
122 #if defined(__arm__)
123 &vm_max_kernel_address, 0,
124 #elif defined(__amd64__)
125 &kva_layout.km_high, 0,
126 #else
127 SYSCTL_NULL_ULONG_PTR, VM_MAX_KERNEL_ADDRESS,
128 #endif
129 "Max kernel address");
130
131 #if VM_NRESERVLEVEL > 1
132 #define KVA_QUANTUM_SHIFT (VM_LEVEL_1_ORDER + VM_LEVEL_0_ORDER + \
133 PAGE_SHIFT)
134 #elif VM_NRESERVLEVEL > 0
135 #define KVA_QUANTUM_SHIFT (VM_LEVEL_0_ORDER + PAGE_SHIFT)
136 #else
137 /* On non-superpage architectures we want large import sizes. */
138 #define KVA_QUANTUM_SHIFT (8 + PAGE_SHIFT)
139 #endif
140 #define KVA_QUANTUM (1ul << KVA_QUANTUM_SHIFT)
141 #define KVA_NUMA_IMPORT_QUANTUM (KVA_QUANTUM * 128)
142
143 extern void uma_startup2(void);
144
145 /*
146 * kva_alloc:
147 *
148 * Allocate a virtual address range with no underlying object and
149 * no initial mapping to physical memory. Any mapping from this
150 * range to physical memory must be explicitly created prior to
151 * its use, typically with pmap_qenter(). Any attempt to create
152 * a mapping on demand through vm_fault() will result in a panic.
153 */
154 void *
kva_alloc(vm_size_t size)155 kva_alloc(vm_size_t size)
156 {
157 vmem_addr_t addr;
158
159 TSENTER();
160 size = round_page(size);
161 if (vmem_xalloc(kernel_arena, size, 0, 0, 0, VMEM_ADDR_MIN,
162 VMEM_ADDR_MAX, M_BESTFIT | M_NOWAIT, &addr))
163 return (0);
164 TSEXIT();
165
166 return ((void *)addr);
167 }
168
169 /*
170 * kva_alloc_aligned:
171 *
172 * Allocate a virtual address range as in kva_alloc where the base
173 * address is aligned to align.
174 */
175 void *
kva_alloc_aligned(vm_size_t size,vm_size_t align)176 kva_alloc_aligned(vm_size_t size, vm_size_t align)
177 {
178 vmem_addr_t addr;
179
180 TSENTER();
181 size = round_page(size);
182 if (vmem_xalloc(kernel_arena, size, align, 0, 0, VMEM_ADDR_MIN,
183 VMEM_ADDR_MAX, M_BESTFIT | M_NOWAIT, &addr))
184 return (0);
185 TSEXIT();
186
187 return ((void *)addr);
188 }
189
190 /*
191 * kva_free:
192 *
193 * Release a region of kernel virtual memory allocated
194 * with kva_alloc, and return the physical pages
195 * associated with that region.
196 *
197 * This routine may not block on kernel maps.
198 */
199 void
kva_free(void * addr,vm_size_t size)200 kva_free(void *addr, vm_size_t size)
201 {
202
203 size = round_page(size);
204 vmem_xfree(kernel_arena, (uintptr_t)addr, size);
205 }
206
207 /*
208 * Update sanitizer shadow state to reflect a new allocation. Force inlining to
209 * help make KMSAN origin tracking more precise.
210 */
211 static __always_inline void
kmem_alloc_san(vm_offset_t addr,vm_size_t size,vm_size_t asize,int flags)212 kmem_alloc_san(vm_offset_t addr, vm_size_t size, vm_size_t asize, int flags)
213 {
214 if ((flags & M_ZERO) == 0) {
215 kmsan_mark((void *)addr, asize, KMSAN_STATE_UNINIT);
216 kmsan_orig((void *)addr, asize, KMSAN_TYPE_KMEM,
217 KMSAN_RET_ADDR);
218 } else {
219 kmsan_mark((void *)addr, asize, KMSAN_STATE_INITED);
220 }
221 kasan_mark((void *)addr, size, asize, KASAN_KMEM_REDZONE);
222 }
223
224 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)225 kmem_alloc_contig_pages(vm_object_t object, vm_pindex_t pindex, int domain,
226 int pflags, u_long npages, vm_paddr_t low, vm_paddr_t high,
227 u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr)
228 {
229 vm_page_t m;
230 int tries;
231 bool wait, reclaim;
232
233 VM_OBJECT_ASSERT_WLOCKED(object);
234
235 wait = (pflags & VM_ALLOC_WAITOK) != 0;
236 reclaim = (pflags & VM_ALLOC_NORECLAIM) == 0;
237 pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
238 pflags |= VM_ALLOC_NOWAIT;
239 for (tries = wait ? 3 : 1;; tries--) {
240 m = vm_page_alloc_contig_domain(object, pindex, domain, pflags,
241 npages, low, high, alignment, boundary, memattr);
242 if (m != NULL || tries == 0 || !reclaim)
243 break;
244
245 VM_OBJECT_WUNLOCK(object);
246 if (vm_page_reclaim_contig_domain(domain, pflags, npages,
247 low, high, alignment, boundary) == ENOMEM && wait)
248 vm_wait_domain(domain);
249 VM_OBJECT_WLOCK(object);
250 }
251 return (m);
252 }
253
254 /*
255 * Allocates a region from the kernel address map and physical pages
256 * within the specified address range to the kernel object. Creates a
257 * wired mapping from this region to these pages, and returns the
258 * region's starting virtual address. The allocated pages are not
259 * necessarily physically contiguous. If M_ZERO is specified through the
260 * given flags, then the pages are zeroed before they are mapped.
261 */
262 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)263 kmem_alloc_attr_domain(int domain, vm_size_t size, int flags, vm_paddr_t low,
264 vm_paddr_t high, vm_memattr_t memattr)
265 {
266 vmem_t *vmem;
267 vm_object_t object;
268 vm_offset_t addr, i, offset;
269 vm_page_t m;
270 vm_size_t asize;
271 int pflags;
272 vm_prot_t prot;
273 u_int pmap_enter_flags;
274
275 object = kernel_object;
276 asize = round_page(size);
277 vmem = vm_dom[domain].vmd_kernel_arena;
278 if (vmem_alloc(vmem, asize, M_BESTFIT | flags, &addr))
279 return (0);
280 offset = addr - VM_MIN_KERNEL_ADDRESS;
281 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
282 prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW;
283
284 pmap_enter_flags = prot | PMAP_ENTER_WIRED;
285 if ((flags & M_UNPROTECTED) != 0)
286 pmap_enter_flags |= PMAP_ENTER_UNPROTECTED;
287
288 VM_OBJECT_WLOCK(object);
289 for (i = 0; i < asize; i += PAGE_SIZE) {
290 m = kmem_alloc_contig_pages(object, atop(offset + i),
291 domain, pflags, 1, low, high, PAGE_SIZE, 0, memattr);
292 if (m == NULL) {
293 VM_OBJECT_WUNLOCK(object);
294 kmem_unback(object, addr, i);
295 vmem_free(vmem, addr, asize);
296 return (0);
297 }
298 KASSERT(vm_page_domain(m) == domain,
299 ("kmem_alloc_attr_domain: Domain mismatch %d != %d",
300 vm_page_domain(m), domain));
301 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
302 pmap_zero_page(m);
303 vm_page_valid(m);
304 pmap_enter(kernel_pmap, addr + i, m, prot,
305 pmap_enter_flags, 0);
306 }
307 VM_OBJECT_WUNLOCK(object);
308 kmem_alloc_san(addr, size, asize, flags);
309 return ((void *)addr);
310 }
311
312 void *
kmem_alloc_attr(vm_size_t size,int flags,vm_paddr_t low,vm_paddr_t high,vm_memattr_t memattr)313 kmem_alloc_attr(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high,
314 vm_memattr_t memattr)
315 {
316
317 return (kmem_alloc_attr_domainset(DOMAINSET_RR(), size, flags, low,
318 high, memattr));
319 }
320
321 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)322 kmem_alloc_attr_domainset(struct domainset *ds, vm_size_t size, int flags,
323 vm_paddr_t low, vm_paddr_t high, vm_memattr_t memattr)
324 {
325 struct vm_domainset_iter di;
326 vm_page_t bounds[2];
327 void *addr;
328 int domain;
329 int start_segind;
330
331 start_segind = -1;
332
333 if (vm_domainset_iter_policy_init(&di, ds, &domain, &flags) != 0)
334 return (NULL);
335
336 do {
337 addr = kmem_alloc_attr_domain(domain, size, flags, low, high,
338 memattr);
339 if (addr != NULL)
340 break;
341 if (start_segind == -1)
342 start_segind = vm_phys_lookup_segind(low);
343 if (vm_phys_find_range(bounds, start_segind, domain,
344 atop(round_page(size)), low, high) == -1) {
345 vm_domainset_iter_ignore(&di, domain);
346 }
347 } while (vm_domainset_iter_policy(&di, &domain) == 0);
348
349 return (addr);
350 }
351
352 /*
353 * Allocates a region from the kernel address map and physically
354 * contiguous pages within the specified address range to the kernel
355 * object. Creates a wired mapping from this region to these pages, and
356 * returns the region's starting virtual address. If M_ZERO is specified
357 * through the given flags, then the pages are zeroed before they are
358 * mapped.
359 */
360 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)361 kmem_alloc_contig_domain(int domain, vm_size_t size, int flags, vm_paddr_t low,
362 vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
363 vm_memattr_t memattr)
364 {
365 vmem_t *vmem;
366 vm_object_t object;
367 vm_offset_t addr, offset, tmp;
368 vm_page_t end_m, m;
369 vm_size_t asize;
370 u_long npages;
371 int pflags;
372 u_int pmap_enter_flags;
373
374 object = kernel_object;
375 asize = round_page(size);
376 vmem = vm_dom[domain].vmd_kernel_arena;
377 if (vmem_alloc(vmem, asize, flags | M_BESTFIT, &addr))
378 return (NULL);
379 offset = addr - VM_MIN_KERNEL_ADDRESS;
380 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
381 npages = atop(asize);
382 VM_OBJECT_WLOCK(object);
383 m = kmem_alloc_contig_pages(object, atop(offset), domain,
384 pflags, npages, low, high, alignment, boundary, memattr);
385 if (m == NULL) {
386 VM_OBJECT_WUNLOCK(object);
387 vmem_free(vmem, addr, asize);
388 return (NULL);
389 }
390 KASSERT(vm_page_domain(m) == domain,
391 ("kmem_alloc_contig_domain: Domain mismatch %d != %d",
392 vm_page_domain(m), domain));
393 end_m = m + npages;
394 tmp = addr;
395
396 pmap_enter_flags = VM_PROT_RW | PMAP_ENTER_WIRED;
397 if ((flags & M_UNPROTECTED) != 0)
398 pmap_enter_flags |= PMAP_ENTER_UNPROTECTED;
399
400 for (; m < end_m; m++) {
401 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
402 pmap_zero_page(m);
403 vm_page_valid(m);
404 pmap_enter(kernel_pmap, tmp, m, VM_PROT_RW,
405 pmap_enter_flags, 0);
406 tmp += PAGE_SIZE;
407 }
408 VM_OBJECT_WUNLOCK(object);
409 kmem_alloc_san(addr, size, asize, flags);
410 return ((void *)addr);
411 }
412
413 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)414 kmem_alloc_contig(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high,
415 u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr)
416 {
417
418 return (kmem_alloc_contig_domainset(DOMAINSET_RR(), size, flags, low,
419 high, alignment, boundary, memattr));
420 }
421
422 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)423 kmem_alloc_contig_domainset(struct domainset *ds, vm_size_t size, int flags,
424 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
425 vm_memattr_t memattr)
426 {
427 struct vm_domainset_iter di;
428 vm_page_t bounds[2];
429 void *addr;
430 int domain;
431 int start_segind;
432
433 start_segind = -1;
434
435 if (vm_domainset_iter_policy_init(&di, ds, &domain, &flags))
436 return (NULL);
437
438 do {
439 addr = kmem_alloc_contig_domain(domain, size, flags, low, high,
440 alignment, boundary, memattr);
441 if (addr != NULL)
442 break;
443 if (start_segind == -1)
444 start_segind = vm_phys_lookup_segind(low);
445 if (vm_phys_find_range(bounds, start_segind, domain,
446 atop(round_page(size)), low, high) == -1) {
447 vm_domainset_iter_ignore(&di, domain);
448 }
449 } while (vm_domainset_iter_policy(&di, &domain) == 0);
450
451 return (addr);
452 }
453
454 /*
455 * kmem_subinit:
456 *
457 * Initializes a map to manage a subrange
458 * of the kernel virtual address space.
459 *
460 * Arguments are as follows:
461 *
462 * parent Map to take range from
463 * min, max Returned endpoints of map
464 * size Size of range to find
465 * superpage_align Request that min is superpage aligned
466 */
467 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)468 kmem_subinit(vm_map_t map, vm_map_t parent, vm_offset_t *min, vm_offset_t *max,
469 vm_size_t size, bool superpage_align)
470 {
471 int ret;
472
473 size = round_page(size);
474
475 *min = vm_map_min(parent);
476 ret = vm_map_find(parent, NULL, 0, min, size, 0, superpage_align ?
477 VMFS_SUPER_SPACE : VMFS_ANY_SPACE, VM_PROT_ALL, VM_PROT_ALL,
478 MAP_ACC_NO_CHARGE);
479 if (ret != KERN_SUCCESS)
480 panic("kmem_subinit: bad status return of %d", ret);
481 *max = *min + size;
482 vm_map_init(map, vm_map_pmap(parent), *min, *max);
483 if (vm_map_submap(parent, *min, *max, map) != KERN_SUCCESS)
484 panic("kmem_subinit: unable to change range to submap");
485 }
486
487 /*
488 * kmem_malloc_domain:
489 *
490 * Allocate wired-down pages in the kernel's address space.
491 */
492 static void *
kmem_malloc_domain(int domain,vm_size_t size,int flags)493 kmem_malloc_domain(int domain, vm_size_t size, int flags)
494 {
495 vmem_t *arena;
496 vm_offset_t addr;
497 vm_size_t asize;
498 int rv;
499
500 if (__predict_true((flags & (M_EXEC | M_NEVERFREED)) == 0))
501 arena = vm_dom[domain].vmd_kernel_arena;
502 else if ((flags & M_EXEC) != 0)
503 arena = vm_dom[domain].vmd_kernel_rwx_arena;
504 else
505 arena = vm_dom[domain].vmd_kernel_nofree_arena;
506 asize = round_page(size);
507 if (vmem_alloc(arena, asize, flags | M_BESTFIT, &addr))
508 return (0);
509
510 rv = kmem_back_domain(domain, kernel_object, addr, asize, flags);
511 if (rv != KERN_SUCCESS) {
512 vmem_free(arena, addr, asize);
513 return (0);
514 }
515 kasan_mark((void *)addr, size, asize, KASAN_KMEM_REDZONE);
516 return ((void *)addr);
517 }
518
519 void *
kmem_malloc(vm_size_t size,int flags)520 kmem_malloc(vm_size_t size, int flags)
521 {
522 void * p;
523
524 TSENTER();
525 p = kmem_malloc_domainset(DOMAINSET_RR(), size, flags);
526 TSEXIT();
527 return (p);
528 }
529
530 void *
kmem_malloc_domainset(struct domainset * ds,vm_size_t size,int flags)531 kmem_malloc_domainset(struct domainset *ds, vm_size_t size, int flags)
532 {
533 struct vm_domainset_iter di;
534 void *addr;
535 int domain;
536
537 if (vm_domainset_iter_policy_init(&di, ds, &domain, &flags) != 0)
538 return (NULL);
539
540 do {
541 addr = kmem_malloc_domain(domain, size, flags);
542 if (addr != NULL)
543 break;
544 } while (vm_domainset_iter_policy(&di, &domain) == 0);
545
546 return (addr);
547 }
548
549 /*
550 * kmem_back_domain:
551 *
552 * Allocate physical pages from the specified domain for the specified
553 * virtual address range.
554 */
555 int
kmem_back_domain(int domain,vm_object_t object,vm_offset_t addr,vm_size_t size,int flags)556 kmem_back_domain(int domain, vm_object_t object, vm_offset_t addr,
557 vm_size_t size, int flags)
558 {
559 struct pctrie_iter pages;
560 vm_offset_t offset, i;
561 vm_page_t m;
562 vm_prot_t prot;
563 int pflags;
564 u_int pmap_enter_flags;
565
566 KASSERT(object == kernel_object,
567 ("kmem_back_domain: only supports kernel object."));
568
569 offset = addr - VM_MIN_KERNEL_ADDRESS;
570 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
571 pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
572 if (flags & M_WAITOK)
573 pflags |= VM_ALLOC_WAITFAIL;
574 prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW;
575
576 pmap_enter_flags = prot | PMAP_ENTER_WIRED;
577 if ((flags & M_UNPROTECTED) != 0)
578 pmap_enter_flags |= PMAP_ENTER_UNPROTECTED;
579
580 i = 0;
581 vm_page_iter_init(&pages, object);
582 VM_OBJECT_WLOCK(object);
583 retry:
584 for (; i < size; i += PAGE_SIZE) {
585 m = vm_page_alloc_domain_iter(object, atop(offset + i),
586 domain, pflags, &pages);
587
588 /*
589 * Ran out of space, free everything up and return. Don't need
590 * to lock page queues here as we know that the pages we got
591 * aren't on any queues.
592 */
593 if (m == NULL) {
594 if ((flags & M_NOWAIT) == 0)
595 goto retry;
596 VM_OBJECT_WUNLOCK(object);
597 kmem_unback(object, addr, i);
598 return (KERN_NO_SPACE);
599 }
600 KASSERT(vm_page_domain(m) == domain,
601 ("kmem_back_domain: Domain mismatch %d != %d",
602 vm_page_domain(m), domain));
603 if (flags & M_ZERO && (m->flags & PG_ZERO) == 0)
604 pmap_zero_page(m);
605 KASSERT((m->oflags & VPO_UNMANAGED) != 0,
606 ("kmem_malloc: page %p is managed", m));
607 vm_page_valid(m);
608 pmap_enter(kernel_pmap, addr + i, m, prot,
609 pmap_enter_flags, 0);
610 if (__predict_false((prot & VM_PROT_EXECUTE) != 0))
611 m->oflags |= VPO_KMEM_EXEC;
612 }
613 VM_OBJECT_WUNLOCK(object);
614 kmem_alloc_san(addr, size, size, flags);
615 return (KERN_SUCCESS);
616 }
617
618 /*
619 * kmem_back:
620 *
621 * Allocate physical pages for the specified virtual address range.
622 */
623 int
kmem_back(vm_object_t object,vm_offset_t addr,vm_size_t size,int flags)624 kmem_back(vm_object_t object, vm_offset_t addr, vm_size_t size, int flags)
625 {
626 vm_offset_t end, next, start;
627 int domain, rv;
628
629 KASSERT(object == kernel_object,
630 ("kmem_back: only supports kernel object."));
631
632 for (start = addr, end = addr + size; addr < end; addr = next) {
633 /*
634 * We must ensure that pages backing a given large virtual page
635 * all come from the same physical domain.
636 */
637 if (vm_ndomains > 1) {
638 domain = (addr >> KVA_QUANTUM_SHIFT) % vm_ndomains;
639 while (VM_DOMAIN_EMPTY(domain))
640 domain++;
641 next = roundup2(addr + 1, KVA_QUANTUM);
642 if (next > end || next < start)
643 next = end;
644 } else {
645 domain = 0;
646 next = end;
647 }
648 rv = kmem_back_domain(domain, object, addr, next - addr, flags);
649 if (rv != KERN_SUCCESS) {
650 kmem_unback(object, start, addr - start);
651 break;
652 }
653 }
654 return (rv);
655 }
656
657 /*
658 * kmem_unback:
659 *
660 * Unmap and free the physical pages underlying the specified virtual
661 * address range.
662 *
663 * A physical page must exist within the specified object at each index
664 * that is being unmapped.
665 */
666 static struct vmem *
_kmem_unback(vm_object_t object,vm_offset_t addr,vm_size_t size)667 _kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size)
668 {
669 struct pctrie_iter pages;
670 struct vmem *arena;
671 vm_page_t m;
672 vm_offset_t end, offset;
673 int domain;
674
675 KASSERT(object == kernel_object,
676 ("kmem_unback: only supports kernel object."));
677
678 if (size == 0)
679 return (NULL);
680 pmap_remove(kernel_pmap, addr, addr + size);
681 offset = addr - VM_MIN_KERNEL_ADDRESS;
682 end = offset + size;
683 vm_page_iter_init(&pages, object);
684 VM_OBJECT_WLOCK(object);
685 m = vm_radix_iter_lookup(&pages, atop(offset));
686 domain = vm_page_domain(m);
687 if (__predict_true((m->oflags & VPO_KMEM_EXEC) == 0))
688 arena = vm_dom[domain].vmd_kernel_arena;
689 else
690 arena = vm_dom[domain].vmd_kernel_rwx_arena;
691 for (; offset < end; offset += PAGE_SIZE,
692 m = vm_radix_iter_lookup(&pages, atop(offset))) {
693 vm_page_xbusy_claim(m);
694 vm_page_unwire_noq(m);
695 vm_page_iter_free(&pages, m);
696 }
697 VM_OBJECT_WUNLOCK(object);
698
699 return (arena);
700 }
701
702 void
kmem_unback(vm_object_t object,vm_offset_t addr,vm_size_t size)703 kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size)
704 {
705
706 (void)_kmem_unback(object, addr, size);
707 }
708
709 /*
710 * kmem_free:
711 *
712 * Free memory allocated with kmem_malloc. The size must match the
713 * original allocation.
714 */
715 void
kmem_free(void * addr,vm_size_t size)716 kmem_free(void *addr, vm_size_t size)
717 {
718 struct vmem *arena;
719
720 size = round_page(size);
721 kasan_mark(addr, size, size, 0);
722 arena = _kmem_unback(kernel_object, (uintptr_t)addr, size);
723 if (arena != NULL)
724 vmem_free(arena, (uintptr_t)addr, size);
725 }
726
727 static void
kmap_alloc_map(vm_map_t map,vm_offset_t addr,vm_size_t size,vm_prot_t prot,int flags)728 kmap_alloc_map(vm_map_t map, vm_offset_t addr, vm_size_t size,
729 vm_prot_t prot, int flags)
730 {
731 int error __diagused;
732
733 error = vm_map_insert(map, NULL, 0,
734 addr, addr + size, prot, prot, flags);
735 KASSERT(error == KERN_SUCCESS,
736 ("%s: unexpected error %d", __func__, error));
737 }
738
739 /*
740 * kmap_alloc_wait:
741 *
742 * Allocates pageable memory from a sub-map of the kernel. If the submap
743 * has no room, the caller sleeps waiting for more memory in the submap.
744 * If "guard_size" is non-zero, then unmapped KVA is left at the beginning
745 * and end of the allocated range.
746 *
747 * This routine may block.
748 */
749 void *
kmap_alloc_wait(vm_map_t map,vm_size_t size,vm_size_t guard_size)750 kmap_alloc_wait(vm_map_t map, vm_size_t size, vm_size_t guard_size)
751 {
752 vm_offset_t addr;
753 vm_size_t total_size;
754
755 KASSERT(size % PAGE_SIZE == 0 && guard_size % PAGE_SIZE == 0,
756 ("%s: size %zu guard_size %zu", __func__, size, guard_size));
757
758 if (!swap_reserve(size))
759 return (NULL);
760
761 total_size = size + 2 * guard_size;
762 for (;;) {
763 /*
764 * To make this work for more than one map, use the map's lock
765 * to lock out sleepers/wakers.
766 */
767 vm_map_lock(map);
768 addr = vm_map_findspace(map, vm_map_min(map), total_size);
769 if (addr + total_size <= vm_map_max(map))
770 break;
771 /* no space now; see if we can ever get space */
772 if (vm_map_max(map) - vm_map_min(map) < total_size) {
773 vm_map_unlock(map);
774 swap_release(size);
775 return (0);
776 }
777 vm_map_modflags(map, MAP_NEEDS_WAKEUP, 0);
778 vm_map_unlock_and_wait(map, 0);
779 }
780 if (guard_size != 0) {
781 kmap_alloc_map(map, addr, guard_size,
782 VM_PROT_NONE, MAP_CREATE_GUARD);
783 kmap_alloc_map(map, addr + guard_size + size, guard_size,
784 VM_PROT_NONE, MAP_CREATE_GUARD);
785 }
786 kmap_alloc_map(map, addr + guard_size, size, VM_PROT_RW,
787 MAP_ACC_CHARGED);
788 vm_map_unlock(map);
789 return ((void *)(addr + guard_size));
790 }
791
792 /*
793 * kmap_free_wakeup:
794 *
795 * Returns memory to a submap of the kernel, and wakes up any processes
796 * waiting for memory in that map.
797 */
798 void
kmap_free_wakeup(vm_map_t map,void * va,vm_size_t size)799 kmap_free_wakeup(vm_map_t map, void *va, vm_size_t size)
800 {
801 vm_offset_t addr;
802
803 addr = (vm_offset_t)va;
804 vm_map_lock(map);
805 (void) vm_map_delete(map, trunc_page(addr), round_page(addr + size));
806 if ((map->flags & MAP_NEEDS_WAKEUP) != 0) {
807 vm_map_modflags(map, 0, MAP_NEEDS_WAKEUP);
808 vm_map_wakeup(map);
809 }
810 vm_map_unlock(map);
811 }
812
813 void
kmem_init_zero_region(void)814 kmem_init_zero_region(void)
815 {
816 char *addr;
817 vm_offset_t i;
818 vm_page_t m;
819
820 /*
821 * Map a single physical page of zeros to a larger virtual range.
822 * This requires less looping in places that want large amounts of
823 * zeros, while not using much more physical resources.
824 */
825 addr = kva_alloc(ZERO_REGION_SIZE);
826 m = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_ZERO |
827 VM_ALLOC_NOFREE);
828 for (i = 0; i < ZERO_REGION_SIZE; i += PAGE_SIZE)
829 pmap_qenter(addr + i, &m, 1);
830 pmap_protect(kernel_pmap, (vm_offset_t)addr,
831 (vm_offset_t)addr + ZERO_REGION_SIZE, VM_PROT_READ);
832
833 zero_region = (const void *)addr;
834 }
835
836 /*
837 * Import KVA from the kernel map into the kernel arena.
838 */
839 static int
kva_import(void * unused,vmem_size_t size,int flags,vmem_addr_t * addrp)840 kva_import(void *unused, vmem_size_t size, int flags, vmem_addr_t *addrp)
841 {
842 vm_offset_t addr;
843 int result;
844
845 TSENTER();
846 KASSERT((size % KVA_QUANTUM) == 0,
847 ("kva_import: Size %jd is not a multiple of %d",
848 (intmax_t)size, (int)KVA_QUANTUM));
849 addr = vm_map_min(kernel_map);
850 result = vm_map_find(kernel_map, NULL, 0, &addr, size, 0,
851 VMFS_SUPER_SPACE, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
852 if (result != KERN_SUCCESS) {
853 TSEXIT();
854 return (ENOMEM);
855 }
856
857 *addrp = addr;
858
859 TSEXIT();
860 return (0);
861 }
862
863 /*
864 * Import KVA from a parent arena into a per-domain arena. Imports must be
865 * KVA_QUANTUM-aligned and a multiple of KVA_QUANTUM in size.
866 */
867 static int
kva_import_domain(void * arena,vmem_size_t size,int flags,vmem_addr_t * addrp)868 kva_import_domain(void *arena, vmem_size_t size, int flags, vmem_addr_t *addrp)
869 {
870
871 KASSERT((size % KVA_QUANTUM) == 0,
872 ("kva_import_domain: Size %jd is not a multiple of %d",
873 (intmax_t)size, (int)KVA_QUANTUM));
874 return (vmem_xalloc(arena, size, KVA_QUANTUM, 0, 0, VMEM_ADDR_MIN,
875 VMEM_ADDR_MAX, flags, addrp));
876 }
877
878 /*
879 * kmem_init:
880 *
881 * Create the kernel map; insert a mapping covering kernel text,
882 * data, bss, and all space allocated thus far (`boostrap' data). The
883 * new map will thus map the range between VM_MIN_KERNEL_ADDRESS and
884 * `start' as allocated, and the range between `start' and `end' as free.
885 * Create the kernel vmem arena and its per-domain children.
886 */
887 void
kmem_init(vm_offset_t start,vm_offset_t end)888 kmem_init(vm_offset_t start, vm_offset_t end)
889 {
890 vm_size_t quantum;
891 int domain;
892
893 vm_map_init_system(kernel_map, kernel_pmap, VM_MIN_KERNEL_ADDRESS, end);
894 vm_map_lock(kernel_map);
895 /* N.B.: cannot use kgdb to debug, starting with this assignment ... */
896 (void)vm_map_insert(kernel_map, NULL, 0,
897 #ifdef __amd64__
898 KERNBASE,
899 #else
900 VM_MIN_KERNEL_ADDRESS,
901 #endif
902 start, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
903 /* ... and ending with the completion of the above `insert' */
904
905 #ifdef __amd64__
906 /*
907 * Mark KVA used for the page array as allocated. Other platforms
908 * that handle vm_page_array allocation can simply adjust virtual_avail
909 * instead.
910 */
911 (void)vm_map_insert(kernel_map, NULL, 0, (vm_offset_t)vm_page_array,
912 (vm_offset_t)vm_page_array + round_2mpage(vm_page_array_size *
913 sizeof(struct vm_page)),
914 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
915 #endif
916 vm_map_unlock(kernel_map);
917
918 /*
919 * Use a large import quantum on NUMA systems. This helps minimize
920 * interleaving of superpages, reducing internal fragmentation within
921 * the per-domain arenas.
922 */
923 if (vm_ndomains > 1 && PMAP_HAS_DMAP)
924 quantum = KVA_NUMA_IMPORT_QUANTUM;
925 else
926 quantum = KVA_QUANTUM;
927
928 /*
929 * Initialize the kernel_arena. This can grow on demand.
930 */
931 vmem_init(kernel_arena, "kernel arena", 0, 0, PAGE_SIZE, 0, 0);
932 vmem_set_import(kernel_arena, kva_import, NULL, NULL, quantum);
933
934 for (domain = 0; domain < vm_ndomains; domain++) {
935 /*
936 * Initialize the per-domain arenas. These are used to color
937 * the KVA space in a way that ensures that virtual large pages
938 * are backed by memory from the same physical domain,
939 * maximizing the potential for superpage promotion.
940 */
941 vm_dom[domain].vmd_kernel_arena = vmem_create(
942 "kernel arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
943 vmem_set_import(vm_dom[domain].vmd_kernel_arena,
944 kva_import_domain, NULL, kernel_arena, quantum);
945
946 /*
947 * In architectures with superpages, maintain separate arenas
948 * for allocations with permissions that differ from the
949 * "standard" read/write permissions used for kernel memory
950 * and pages that are never released, so as not to inhibit
951 * superpage promotion.
952 *
953 * Use the base import quantum since these arenas are rarely
954 * used.
955 */
956 #if VM_NRESERVLEVEL > 0
957 vm_dom[domain].vmd_kernel_rwx_arena = vmem_create(
958 "kernel rwx arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
959 vm_dom[domain].vmd_kernel_nofree_arena = vmem_create(
960 "kernel NOFREE arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
961 vmem_set_import(vm_dom[domain].vmd_kernel_rwx_arena,
962 kva_import_domain, (vmem_release_t *)vmem_xfree,
963 kernel_arena, KVA_QUANTUM);
964 vmem_set_import(vm_dom[domain].vmd_kernel_nofree_arena,
965 kva_import_domain, (vmem_release_t *)vmem_xfree,
966 kernel_arena, KVA_QUANTUM);
967 #else
968 vm_dom[domain].vmd_kernel_rwx_arena =
969 vm_dom[domain].vmd_kernel_arena;
970 vm_dom[domain].vmd_kernel_nofree_arena =
971 vm_dom[domain].vmd_kernel_arena;
972 #endif
973 }
974
975 /*
976 * This must be the very first call so that the virtual address
977 * space used for early allocations is properly marked used in
978 * the map.
979 */
980 uma_startup2();
981 }
982
983 /*
984 * kmem_bootstrap_free:
985 *
986 * Free pages backing preloaded data (e.g., kernel modules) to the
987 * system. Currently only supported on platforms that create a
988 * vm_phys segment for preloaded data.
989 */
990 void
kmem_bootstrap_free(vm_offset_t start,vm_size_t size)991 kmem_bootstrap_free(vm_offset_t start, vm_size_t size)
992 {
993 #if defined(__i386__) || defined(__amd64__)
994 struct vm_domain *vmd;
995 vm_offset_t end, va;
996 vm_paddr_t pa;
997 vm_page_t m;
998
999 end = trunc_page(start + size);
1000 start = round_page(start);
1001
1002 for (va = start; va < end; va += PAGE_SIZE) {
1003 pa = pmap_kextract(va);
1004 m = PHYS_TO_VM_PAGE(pa);
1005
1006 vmd = vm_pagequeue_domain(m);
1007 vm_domain_free_lock(vmd);
1008 vm_phys_free_pages(m, m->pool, 0);
1009 vm_domain_free_unlock(vmd);
1010
1011 vm_domain_freecnt_inc(vmd, 1);
1012 vm_cnt.v_page_count++;
1013 }
1014 pmap_remove(kernel_pmap, start, end);
1015 (void)vmem_add(kernel_arena, start, end - start, M_WAITOK);
1016 #endif
1017 }
1018
1019 #ifdef PMAP_WANT_ACTIVE_CPUS_NAIVE
1020 void
pmap_active_cpus(pmap_t pmap,cpuset_t * res)1021 pmap_active_cpus(pmap_t pmap, cpuset_t *res)
1022 {
1023 struct thread *td;
1024 struct proc *p;
1025 struct vmspace *vm;
1026 int c;
1027
1028 CPU_ZERO(res);
1029 CPU_FOREACH(c) {
1030 td = cpuid_to_pcpu[c]->pc_curthread;
1031 p = td->td_proc;
1032 if (p == NULL)
1033 continue;
1034 vm = vmspace_acquire_ref(p);
1035 if (vm == NULL)
1036 continue;
1037 if (pmap == vmspace_pmap(vm))
1038 CPU_SET(c, res);
1039 vmspace_free(vm);
1040 }
1041 }
1042 #endif
1043
1044 /*
1045 * Allow userspace to directly trigger the VM drain routine for testing
1046 * purposes.
1047 */
1048 static int
debug_vm_lowmem(SYSCTL_HANDLER_ARGS)1049 debug_vm_lowmem(SYSCTL_HANDLER_ARGS)
1050 {
1051 int error, i;
1052
1053 i = 0;
1054 error = sysctl_handle_int(oidp, &i, 0, req);
1055 if (error != 0)
1056 return (error);
1057 if ((i & ~(VM_LOW_KMEM | VM_LOW_PAGES)) != 0)
1058 return (EINVAL);
1059 if (i != 0)
1060 EVENTHANDLER_INVOKE(vm_lowmem, i);
1061 return (0);
1062 }
1063 SYSCTL_PROC(_debug, OID_AUTO, vm_lowmem,
1064 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, debug_vm_lowmem, "I",
1065 "set to trigger vm_lowmem event with given flags");
1066
1067 static int
debug_uma_reclaim(SYSCTL_HANDLER_ARGS)1068 debug_uma_reclaim(SYSCTL_HANDLER_ARGS)
1069 {
1070 int error, i;
1071
1072 i = 0;
1073 error = sysctl_handle_int(oidp, &i, 0, req);
1074 if (error != 0 || req->newptr == NULL)
1075 return (error);
1076 if (i != UMA_RECLAIM_TRIM && i != UMA_RECLAIM_DRAIN &&
1077 i != UMA_RECLAIM_DRAIN_CPU)
1078 return (EINVAL);
1079 uma_reclaim(i);
1080 return (0);
1081 }
1082 SYSCTL_PROC(_debug, OID_AUTO, uma_reclaim,
1083 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, debug_uma_reclaim, "I",
1084 "set to generate request to reclaim uma caches");
1085
1086 static int
debug_uma_reclaim_domain(SYSCTL_HANDLER_ARGS)1087 debug_uma_reclaim_domain(SYSCTL_HANDLER_ARGS)
1088 {
1089 int domain, error, request;
1090
1091 request = 0;
1092 error = sysctl_handle_int(oidp, &request, 0, req);
1093 if (error != 0 || req->newptr == NULL)
1094 return (error);
1095
1096 domain = request >> 4;
1097 request &= 0xf;
1098 if (request != UMA_RECLAIM_TRIM && request != UMA_RECLAIM_DRAIN &&
1099 request != UMA_RECLAIM_DRAIN_CPU)
1100 return (EINVAL);
1101 if (domain < 0 || domain >= vm_ndomains)
1102 return (EINVAL);
1103 uma_reclaim_domain(request, domain);
1104 return (0);
1105 }
1106 SYSCTL_PROC(_debug, OID_AUTO, uma_reclaim_domain,
1107 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0,
1108 debug_uma_reclaim_domain, "I",
1109 "");
1110