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 #if defined(__amd64__)
114 &kva_layout.km_low, 0,
115 #else
116 SYSCTL_NULL_ULONG_PTR, VM_MIN_KERNEL_ADDRESS,
117 #endif
118 "Min kernel address");
119
120 SYSCTL_ULONG(_vm, OID_AUTO, max_kernel_address, CTLFLAG_RD,
121 #if defined(__arm__)
122 &vm_max_kernel_address, 0,
123 #elif defined(__amd64__)
124 &kva_layout.km_high, 0,
125 #else
126 SYSCTL_NULL_ULONG_PTR, VM_MAX_KERNEL_ADDRESS,
127 #endif
128 "Max kernel address");
129
130 #if VM_NRESERVLEVEL > 1
131 #define KVA_QUANTUM_SHIFT (VM_LEVEL_1_ORDER + VM_LEVEL_0_ORDER + \
132 PAGE_SHIFT)
133 #elif VM_NRESERVLEVEL > 0
134 #define KVA_QUANTUM_SHIFT (VM_LEVEL_0_ORDER + PAGE_SHIFT)
135 #else
136 /* On non-superpage architectures we want large import sizes. */
137 #define KVA_QUANTUM_SHIFT (8 + PAGE_SHIFT)
138 #endif
139 #define KVA_QUANTUM (1ul << KVA_QUANTUM_SHIFT)
140 #define KVA_NUMA_IMPORT_QUANTUM (KVA_QUANTUM * 128)
141
142 extern void uma_startup2(void);
143
144 /*
145 * kva_alloc:
146 *
147 * Allocate a virtual address range with no underlying object and
148 * no initial mapping to physical memory. Any mapping from this
149 * range to physical memory must be explicitly created prior to
150 * its use, typically with pmap_qenter(). Any attempt to create
151 * a mapping on demand through vm_fault() will result in a panic.
152 */
153 vm_offset_t
kva_alloc(vm_size_t size)154 kva_alloc(vm_size_t size)
155 {
156 vm_offset_t addr;
157
158 TSENTER();
159 size = round_page(size);
160 if (vmem_xalloc(kernel_arena, size, 0, 0, 0, VMEM_ADDR_MIN,
161 VMEM_ADDR_MAX, M_BESTFIT | M_NOWAIT, &addr))
162 return (0);
163 TSEXIT();
164
165 return (addr);
166 }
167
168 /*
169 * kva_alloc_aligned:
170 *
171 * Allocate a virtual address range as in kva_alloc where the base
172 * address is aligned to align.
173 */
174 vm_offset_t
kva_alloc_aligned(vm_size_t size,vm_size_t align)175 kva_alloc_aligned(vm_size_t size, vm_size_t align)
176 {
177 vm_offset_t addr;
178
179 TSENTER();
180 size = round_page(size);
181 if (vmem_xalloc(kernel_arena, size, align, 0, 0, VMEM_ADDR_MIN,
182 VMEM_ADDR_MAX, M_BESTFIT | M_NOWAIT, &addr))
183 return (0);
184 TSEXIT();
185
186 return (addr);
187 }
188
189 /*
190 * kva_free:
191 *
192 * Release a region of kernel virtual memory allocated
193 * with kva_alloc, and return the physical pages
194 * associated with that region.
195 *
196 * This routine may not block on kernel maps.
197 */
198 void
kva_free(vm_offset_t addr,vm_size_t size)199 kva_free(vm_offset_t addr, vm_size_t size)
200 {
201
202 size = round_page(size);
203 vmem_xfree(kernel_arena, addr, size);
204 }
205
206 /*
207 * Update sanitizer shadow state to reflect a new allocation. Force inlining to
208 * help make KMSAN origin tracking more precise.
209 */
210 static __always_inline void
kmem_alloc_san(vm_offset_t addr,vm_size_t size,vm_size_t asize,int flags)211 kmem_alloc_san(vm_offset_t addr, vm_size_t size, vm_size_t asize, int flags)
212 {
213 if ((flags & M_ZERO) == 0) {
214 kmsan_mark((void *)addr, asize, KMSAN_STATE_UNINIT);
215 kmsan_orig((void *)addr, asize, KMSAN_TYPE_KMEM,
216 KMSAN_RET_ADDR);
217 } else {
218 kmsan_mark((void *)addr, asize, KMSAN_STATE_INITED);
219 }
220 kasan_mark((void *)addr, size, asize, KASAN_KMEM_REDZONE);
221 }
222
223 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)224 kmem_alloc_contig_pages(vm_object_t object, vm_pindex_t pindex, int domain,
225 int pflags, u_long npages, vm_paddr_t low, vm_paddr_t high,
226 u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr)
227 {
228 vm_page_t m;
229 int tries;
230 bool wait, reclaim;
231
232 VM_OBJECT_ASSERT_WLOCKED(object);
233
234 wait = (pflags & VM_ALLOC_WAITOK) != 0;
235 reclaim = (pflags & VM_ALLOC_NORECLAIM) == 0;
236 pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
237 pflags |= VM_ALLOC_NOWAIT;
238 for (tries = wait ? 3 : 1;; tries--) {
239 m = vm_page_alloc_contig_domain(object, pindex, domain, pflags,
240 npages, low, high, alignment, boundary, memattr);
241 if (m != NULL || tries == 0 || !reclaim)
242 break;
243
244 VM_OBJECT_WUNLOCK(object);
245 if (vm_page_reclaim_contig_domain(domain, pflags, npages,
246 low, high, alignment, boundary) == ENOMEM && wait)
247 vm_wait_domain(domain);
248 VM_OBJECT_WLOCK(object);
249 }
250 return (m);
251 }
252
253 /*
254 * Allocates a region from the kernel address map and physical pages
255 * within the specified address range to the kernel object. Creates a
256 * wired mapping from this region to these pages, and returns the
257 * region's starting virtual address. The allocated pages are not
258 * necessarily physically contiguous. If M_ZERO is specified through the
259 * given flags, then the pages are zeroed before they are mapped.
260 */
261 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)262 kmem_alloc_attr_domain(int domain, vm_size_t size, int flags, vm_paddr_t low,
263 vm_paddr_t high, vm_memattr_t memattr)
264 {
265 vmem_t *vmem;
266 vm_object_t object;
267 vm_offset_t addr, i, offset;
268 vm_page_t m;
269 vm_size_t asize;
270 int pflags;
271 vm_prot_t prot;
272
273 object = kernel_object;
274 asize = round_page(size);
275 vmem = vm_dom[domain].vmd_kernel_arena;
276 if (vmem_alloc(vmem, asize, M_BESTFIT | flags, &addr))
277 return (0);
278 offset = addr - VM_MIN_KERNEL_ADDRESS;
279 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
280 prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW;
281 VM_OBJECT_WLOCK(object);
282 for (i = 0; i < asize; i += PAGE_SIZE) {
283 m = kmem_alloc_contig_pages(object, atop(offset + i),
284 domain, pflags, 1, low, high, PAGE_SIZE, 0, memattr);
285 if (m == NULL) {
286 VM_OBJECT_WUNLOCK(object);
287 kmem_unback(object, addr, i);
288 vmem_free(vmem, addr, asize);
289 return (0);
290 }
291 KASSERT(vm_page_domain(m) == domain,
292 ("kmem_alloc_attr_domain: Domain mismatch %d != %d",
293 vm_page_domain(m), domain));
294 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
295 pmap_zero_page(m);
296 vm_page_valid(m);
297 pmap_enter(kernel_pmap, addr + i, m, prot,
298 prot | PMAP_ENTER_WIRED, 0);
299 }
300 VM_OBJECT_WUNLOCK(object);
301 kmem_alloc_san(addr, size, asize, flags);
302 return ((void *)addr);
303 }
304
305 void *
kmem_alloc_attr(vm_size_t size,int flags,vm_paddr_t low,vm_paddr_t high,vm_memattr_t memattr)306 kmem_alloc_attr(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high,
307 vm_memattr_t memattr)
308 {
309
310 return (kmem_alloc_attr_domainset(DOMAINSET_RR(), size, flags, low,
311 high, memattr));
312 }
313
314 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)315 kmem_alloc_attr_domainset(struct domainset *ds, vm_size_t size, int flags,
316 vm_paddr_t low, vm_paddr_t high, vm_memattr_t memattr)
317 {
318 struct vm_domainset_iter di;
319 vm_page_t bounds[2];
320 void *addr;
321 int domain;
322 int start_segind;
323
324 start_segind = -1;
325
326 vm_domainset_iter_policy_init(&di, ds, &domain, &flags);
327 do {
328 addr = kmem_alloc_attr_domain(domain, size, flags, low, high,
329 memattr);
330 if (addr != NULL)
331 break;
332 if (start_segind == -1)
333 start_segind = vm_phys_lookup_segind(low);
334 if (vm_phys_find_range(bounds, start_segind, domain,
335 atop(round_page(size)), low, high) == -1) {
336 vm_domainset_iter_ignore(&di, domain);
337 }
338 } while (vm_domainset_iter_policy(&di, &domain) == 0);
339
340 return (addr);
341 }
342
343 /*
344 * Allocates a region from the kernel address map and physically
345 * contiguous pages within the specified address range to the kernel
346 * object. Creates a wired mapping from this region to these pages, and
347 * returns the region's starting virtual address. If M_ZERO is specified
348 * through the given flags, then the pages are zeroed before they are
349 * mapped.
350 */
351 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)352 kmem_alloc_contig_domain(int domain, vm_size_t size, int flags, vm_paddr_t low,
353 vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
354 vm_memattr_t memattr)
355 {
356 vmem_t *vmem;
357 vm_object_t object;
358 vm_offset_t addr, offset, tmp;
359 vm_page_t end_m, m;
360 vm_size_t asize;
361 u_long npages;
362 int pflags;
363
364 object = kernel_object;
365 asize = round_page(size);
366 vmem = vm_dom[domain].vmd_kernel_arena;
367 if (vmem_alloc(vmem, asize, flags | M_BESTFIT, &addr))
368 return (NULL);
369 offset = addr - VM_MIN_KERNEL_ADDRESS;
370 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
371 npages = atop(asize);
372 VM_OBJECT_WLOCK(object);
373 m = kmem_alloc_contig_pages(object, atop(offset), domain,
374 pflags, npages, low, high, alignment, boundary, memattr);
375 if (m == NULL) {
376 VM_OBJECT_WUNLOCK(object);
377 vmem_free(vmem, addr, asize);
378 return (NULL);
379 }
380 KASSERT(vm_page_domain(m) == domain,
381 ("kmem_alloc_contig_domain: Domain mismatch %d != %d",
382 vm_page_domain(m), domain));
383 end_m = m + npages;
384 tmp = addr;
385 for (; m < end_m; m++) {
386 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
387 pmap_zero_page(m);
388 vm_page_valid(m);
389 pmap_enter(kernel_pmap, tmp, m, VM_PROT_RW,
390 VM_PROT_RW | PMAP_ENTER_WIRED, 0);
391 tmp += PAGE_SIZE;
392 }
393 VM_OBJECT_WUNLOCK(object);
394 kmem_alloc_san(addr, size, asize, flags);
395 return ((void *)addr);
396 }
397
398 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)399 kmem_alloc_contig(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high,
400 u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr)
401 {
402
403 return (kmem_alloc_contig_domainset(DOMAINSET_RR(), size, flags, low,
404 high, alignment, boundary, memattr));
405 }
406
407 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)408 kmem_alloc_contig_domainset(struct domainset *ds, vm_size_t size, int flags,
409 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
410 vm_memattr_t memattr)
411 {
412 struct vm_domainset_iter di;
413 vm_page_t bounds[2];
414 void *addr;
415 int domain;
416 int start_segind;
417
418 start_segind = -1;
419
420 vm_domainset_iter_policy_init(&di, ds, &domain, &flags);
421 do {
422 addr = kmem_alloc_contig_domain(domain, size, flags, low, high,
423 alignment, boundary, memattr);
424 if (addr != NULL)
425 break;
426 if (start_segind == -1)
427 start_segind = vm_phys_lookup_segind(low);
428 if (vm_phys_find_range(bounds, start_segind, domain,
429 atop(round_page(size)), low, high) == -1) {
430 vm_domainset_iter_ignore(&di, domain);
431 }
432 } while (vm_domainset_iter_policy(&di, &domain) == 0);
433
434 return (addr);
435 }
436
437 /*
438 * kmem_subinit:
439 *
440 * Initializes a map to manage a subrange
441 * of the kernel virtual address space.
442 *
443 * Arguments are as follows:
444 *
445 * parent Map to take range from
446 * min, max Returned endpoints of map
447 * size Size of range to find
448 * superpage_align Request that min is superpage aligned
449 */
450 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)451 kmem_subinit(vm_map_t map, vm_map_t parent, vm_offset_t *min, vm_offset_t *max,
452 vm_size_t size, bool superpage_align)
453 {
454 int ret;
455
456 size = round_page(size);
457
458 *min = vm_map_min(parent);
459 ret = vm_map_find(parent, NULL, 0, min, size, 0, superpage_align ?
460 VMFS_SUPER_SPACE : VMFS_ANY_SPACE, VM_PROT_ALL, VM_PROT_ALL,
461 MAP_ACC_NO_CHARGE);
462 if (ret != KERN_SUCCESS)
463 panic("kmem_subinit: bad status return of %d", ret);
464 *max = *min + size;
465 vm_map_init(map, vm_map_pmap(parent), *min, *max);
466 if (vm_map_submap(parent, *min, *max, map) != KERN_SUCCESS)
467 panic("kmem_subinit: unable to change range to submap");
468 }
469
470 /*
471 * kmem_malloc_domain:
472 *
473 * Allocate wired-down pages in the kernel's address space.
474 */
475 static void *
kmem_malloc_domain(int domain,vm_size_t size,int flags)476 kmem_malloc_domain(int domain, vm_size_t size, int flags)
477 {
478 vmem_t *arena;
479 vm_offset_t addr;
480 vm_size_t asize;
481 int rv;
482
483 if (__predict_true((flags & (M_EXEC | M_NEVERFREED)) == 0))
484 arena = vm_dom[domain].vmd_kernel_arena;
485 else if ((flags & M_EXEC) != 0)
486 arena = vm_dom[domain].vmd_kernel_rwx_arena;
487 else
488 arena = vm_dom[domain].vmd_kernel_nofree_arena;
489 asize = round_page(size);
490 if (vmem_alloc(arena, asize, flags | M_BESTFIT, &addr))
491 return (0);
492
493 rv = kmem_back_domain(domain, kernel_object, addr, asize, flags);
494 if (rv != KERN_SUCCESS) {
495 vmem_free(arena, addr, asize);
496 return (0);
497 }
498 kasan_mark((void *)addr, size, asize, KASAN_KMEM_REDZONE);
499 return ((void *)addr);
500 }
501
502 void *
kmem_malloc(vm_size_t size,int flags)503 kmem_malloc(vm_size_t size, int flags)
504 {
505 void * p;
506
507 TSENTER();
508 p = kmem_malloc_domainset(DOMAINSET_RR(), size, flags);
509 TSEXIT();
510 return (p);
511 }
512
513 void *
kmem_malloc_domainset(struct domainset * ds,vm_size_t size,int flags)514 kmem_malloc_domainset(struct domainset *ds, vm_size_t size, int flags)
515 {
516 struct vm_domainset_iter di;
517 void *addr;
518 int domain;
519
520 vm_domainset_iter_policy_init(&di, ds, &domain, &flags);
521 do {
522 addr = kmem_malloc_domain(domain, size, flags);
523 if (addr != NULL)
524 break;
525 } while (vm_domainset_iter_policy(&di, &domain) == 0);
526
527 return (addr);
528 }
529
530 /*
531 * kmem_back_domain:
532 *
533 * Allocate physical pages from the specified domain for the specified
534 * virtual address range.
535 */
536 int
kmem_back_domain(int domain,vm_object_t object,vm_offset_t addr,vm_size_t size,int flags)537 kmem_back_domain(int domain, vm_object_t object, vm_offset_t addr,
538 vm_size_t size, int flags)
539 {
540 struct pctrie_iter pages;
541 vm_offset_t offset, i;
542 vm_page_t m;
543 vm_prot_t prot;
544 int pflags;
545
546 KASSERT(object == kernel_object,
547 ("kmem_back_domain: only supports kernel object."));
548
549 offset = addr - VM_MIN_KERNEL_ADDRESS;
550 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
551 pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
552 if (flags & M_WAITOK)
553 pflags |= VM_ALLOC_WAITFAIL;
554 prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW;
555
556 i = 0;
557 vm_page_iter_init(&pages, object);
558 VM_OBJECT_WLOCK(object);
559 retry:
560 for (; i < size; i += PAGE_SIZE) {
561 m = vm_page_alloc_domain_iter(object, atop(offset + i),
562 domain, pflags, &pages);
563
564 /*
565 * Ran out of space, free everything up and return. Don't need
566 * to lock page queues here as we know that the pages we got
567 * aren't on any queues.
568 */
569 if (m == NULL) {
570 if ((flags & M_NOWAIT) == 0)
571 goto retry;
572 VM_OBJECT_WUNLOCK(object);
573 kmem_unback(object, addr, i);
574 return (KERN_NO_SPACE);
575 }
576 KASSERT(vm_page_domain(m) == domain,
577 ("kmem_back_domain: Domain mismatch %d != %d",
578 vm_page_domain(m), domain));
579 if (flags & M_ZERO && (m->flags & PG_ZERO) == 0)
580 pmap_zero_page(m);
581 KASSERT((m->oflags & VPO_UNMANAGED) != 0,
582 ("kmem_malloc: page %p is managed", m));
583 vm_page_valid(m);
584 pmap_enter(kernel_pmap, addr + i, m, prot,
585 prot | PMAP_ENTER_WIRED, 0);
586 if (__predict_false((prot & VM_PROT_EXECUTE) != 0))
587 m->oflags |= VPO_KMEM_EXEC;
588 }
589 VM_OBJECT_WUNLOCK(object);
590 kmem_alloc_san(addr, size, size, flags);
591 return (KERN_SUCCESS);
592 }
593
594 /*
595 * kmem_back:
596 *
597 * Allocate physical pages for the specified virtual address range.
598 */
599 int
kmem_back(vm_object_t object,vm_offset_t addr,vm_size_t size,int flags)600 kmem_back(vm_object_t object, vm_offset_t addr, vm_size_t size, int flags)
601 {
602 vm_offset_t end, next, start;
603 int domain, rv;
604
605 KASSERT(object == kernel_object,
606 ("kmem_back: only supports kernel object."));
607
608 for (start = addr, end = addr + size; addr < end; addr = next) {
609 /*
610 * We must ensure that pages backing a given large virtual page
611 * all come from the same physical domain.
612 */
613 if (vm_ndomains > 1) {
614 domain = (addr >> KVA_QUANTUM_SHIFT) % vm_ndomains;
615 while (VM_DOMAIN_EMPTY(domain))
616 domain++;
617 next = roundup2(addr + 1, KVA_QUANTUM);
618 if (next > end || next < start)
619 next = end;
620 } else {
621 domain = 0;
622 next = end;
623 }
624 rv = kmem_back_domain(domain, object, addr, next - addr, flags);
625 if (rv != KERN_SUCCESS) {
626 kmem_unback(object, start, addr - start);
627 break;
628 }
629 }
630 return (rv);
631 }
632
633 /*
634 * kmem_unback:
635 *
636 * Unmap and free the physical pages underlying the specified virtual
637 * address range.
638 *
639 * A physical page must exist within the specified object at each index
640 * that is being unmapped.
641 */
642 static struct vmem *
_kmem_unback(vm_object_t object,vm_offset_t addr,vm_size_t size)643 _kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size)
644 {
645 struct pctrie_iter pages;
646 struct vmem *arena;
647 vm_page_t m;
648 vm_offset_t end, offset;
649 int domain;
650
651 KASSERT(object == kernel_object,
652 ("kmem_unback: only supports kernel object."));
653
654 if (size == 0)
655 return (NULL);
656 pmap_remove(kernel_pmap, addr, addr + size);
657 offset = addr - VM_MIN_KERNEL_ADDRESS;
658 end = offset + size;
659 vm_page_iter_init(&pages, object);
660 VM_OBJECT_WLOCK(object);
661 m = vm_radix_iter_lookup(&pages, atop(offset));
662 domain = vm_page_domain(m);
663 if (__predict_true((m->oflags & VPO_KMEM_EXEC) == 0))
664 arena = vm_dom[domain].vmd_kernel_arena;
665 else
666 arena = vm_dom[domain].vmd_kernel_rwx_arena;
667 for (; offset < end; offset += PAGE_SIZE,
668 m = vm_radix_iter_lookup(&pages, atop(offset))) {
669 vm_page_xbusy_claim(m);
670 vm_page_unwire_noq(m);
671 vm_page_iter_free(&pages, m);
672 }
673 VM_OBJECT_WUNLOCK(object);
674
675 return (arena);
676 }
677
678 void
kmem_unback(vm_object_t object,vm_offset_t addr,vm_size_t size)679 kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size)
680 {
681
682 (void)_kmem_unback(object, addr, size);
683 }
684
685 /*
686 * kmem_free:
687 *
688 * Free memory allocated with kmem_malloc. The size must match the
689 * original allocation.
690 */
691 void
kmem_free(void * addr,vm_size_t size)692 kmem_free(void *addr, vm_size_t size)
693 {
694 struct vmem *arena;
695
696 size = round_page(size);
697 kasan_mark(addr, size, size, 0);
698 arena = _kmem_unback(kernel_object, (uintptr_t)addr, size);
699 if (arena != NULL)
700 vmem_free(arena, (uintptr_t)addr, size);
701 }
702
703 /*
704 * kmap_alloc_wait:
705 *
706 * Allocates pageable memory from a sub-map of the kernel. If the submap
707 * has no room, the caller sleeps waiting for more memory in the submap.
708 *
709 * This routine may block.
710 */
711 vm_offset_t
kmap_alloc_wait(vm_map_t map,vm_size_t size)712 kmap_alloc_wait(vm_map_t map, vm_size_t size)
713 {
714 vm_offset_t addr;
715
716 size = round_page(size);
717 if (!swap_reserve(size))
718 return (0);
719
720 for (;;) {
721 /*
722 * To make this work for more than one map, use the map's lock
723 * to lock out sleepers/wakers.
724 */
725 vm_map_lock(map);
726 addr = vm_map_findspace(map, vm_map_min(map), size);
727 if (addr + size <= vm_map_max(map))
728 break;
729 /* no space now; see if we can ever get space */
730 if (vm_map_max(map) - vm_map_min(map) < size) {
731 vm_map_unlock(map);
732 swap_release(size);
733 return (0);
734 }
735 vm_map_modflags(map, MAP_NEEDS_WAKEUP, 0);
736 vm_map_unlock_and_wait(map, 0);
737 }
738 vm_map_insert(map, NULL, 0, addr, addr + size, VM_PROT_RW, VM_PROT_RW,
739 MAP_ACC_CHARGED);
740 vm_map_unlock(map);
741 return (addr);
742 }
743
744 /*
745 * kmap_free_wakeup:
746 *
747 * Returns memory to a submap of the kernel, and wakes up any processes
748 * waiting for memory in that map.
749 */
750 void
kmap_free_wakeup(vm_map_t map,vm_offset_t addr,vm_size_t size)751 kmap_free_wakeup(vm_map_t map, vm_offset_t addr, vm_size_t size)
752 {
753
754 vm_map_lock(map);
755 (void) vm_map_delete(map, trunc_page(addr), round_page(addr + size));
756 if ((map->flags & MAP_NEEDS_WAKEUP) != 0) {
757 vm_map_modflags(map, 0, MAP_NEEDS_WAKEUP);
758 vm_map_wakeup(map);
759 }
760 vm_map_unlock(map);
761 }
762
763 void
kmem_init_zero_region(void)764 kmem_init_zero_region(void)
765 {
766 vm_offset_t addr, i;
767 vm_page_t m;
768
769 /*
770 * Map a single physical page of zeros to a larger virtual range.
771 * This requires less looping in places that want large amounts of
772 * zeros, while not using much more physical resources.
773 */
774 addr = kva_alloc(ZERO_REGION_SIZE);
775 m = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_ZERO |
776 VM_ALLOC_NOFREE);
777 for (i = 0; i < ZERO_REGION_SIZE; i += PAGE_SIZE)
778 pmap_qenter(addr + i, &m, 1);
779 pmap_protect(kernel_pmap, addr, addr + ZERO_REGION_SIZE, VM_PROT_READ);
780
781 zero_region = (const void *)addr;
782 }
783
784 /*
785 * Import KVA from the kernel map into the kernel arena.
786 */
787 static int
kva_import(void * unused,vmem_size_t size,int flags,vmem_addr_t * addrp)788 kva_import(void *unused, vmem_size_t size, int flags, vmem_addr_t *addrp)
789 {
790 vm_offset_t addr;
791 int result;
792
793 TSENTER();
794 KASSERT((size % KVA_QUANTUM) == 0,
795 ("kva_import: Size %jd is not a multiple of %d",
796 (intmax_t)size, (int)KVA_QUANTUM));
797 addr = vm_map_min(kernel_map);
798 result = vm_map_find(kernel_map, NULL, 0, &addr, size, 0,
799 VMFS_SUPER_SPACE, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
800 if (result != KERN_SUCCESS) {
801 TSEXIT();
802 return (ENOMEM);
803 }
804
805 *addrp = addr;
806
807 TSEXIT();
808 return (0);
809 }
810
811 /*
812 * Import KVA from a parent arena into a per-domain arena. Imports must be
813 * KVA_QUANTUM-aligned and a multiple of KVA_QUANTUM in size.
814 */
815 static int
kva_import_domain(void * arena,vmem_size_t size,int flags,vmem_addr_t * addrp)816 kva_import_domain(void *arena, vmem_size_t size, int flags, vmem_addr_t *addrp)
817 {
818
819 KASSERT((size % KVA_QUANTUM) == 0,
820 ("kva_import_domain: Size %jd is not a multiple of %d",
821 (intmax_t)size, (int)KVA_QUANTUM));
822 return (vmem_xalloc(arena, size, KVA_QUANTUM, 0, 0, VMEM_ADDR_MIN,
823 VMEM_ADDR_MAX, flags, addrp));
824 }
825
826 /*
827 * kmem_init:
828 *
829 * Create the kernel map; insert a mapping covering kernel text,
830 * data, bss, and all space allocated thus far (`boostrap' data). The
831 * new map will thus map the range between VM_MIN_KERNEL_ADDRESS and
832 * `start' as allocated, and the range between `start' and `end' as free.
833 * Create the kernel vmem arena and its per-domain children.
834 */
835 void
kmem_init(vm_offset_t start,vm_offset_t end)836 kmem_init(vm_offset_t start, vm_offset_t end)
837 {
838 vm_size_t quantum;
839 int domain;
840
841 vm_map_init_system(kernel_map, kernel_pmap, VM_MIN_KERNEL_ADDRESS, end);
842 vm_map_lock(kernel_map);
843 /* N.B.: cannot use kgdb to debug, starting with this assignment ... */
844 (void)vm_map_insert(kernel_map, NULL, 0,
845 #ifdef __amd64__
846 KERNBASE,
847 #else
848 VM_MIN_KERNEL_ADDRESS,
849 #endif
850 start, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
851 /* ... and ending with the completion of the above `insert' */
852
853 #ifdef __amd64__
854 /*
855 * Mark KVA used for the page array as allocated. Other platforms
856 * that handle vm_page_array allocation can simply adjust virtual_avail
857 * instead.
858 */
859 (void)vm_map_insert(kernel_map, NULL, 0, (vm_offset_t)vm_page_array,
860 (vm_offset_t)vm_page_array + round_2mpage(vm_page_array_size *
861 sizeof(struct vm_page)),
862 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
863 #endif
864 vm_map_unlock(kernel_map);
865
866 /*
867 * Use a large import quantum on NUMA systems. This helps minimize
868 * interleaving of superpages, reducing internal fragmentation within
869 * the per-domain arenas.
870 */
871 if (vm_ndomains > 1 && PMAP_HAS_DMAP)
872 quantum = KVA_NUMA_IMPORT_QUANTUM;
873 else
874 quantum = KVA_QUANTUM;
875
876 /*
877 * Initialize the kernel_arena. This can grow on demand.
878 */
879 vmem_init(kernel_arena, "kernel arena", 0, 0, PAGE_SIZE, 0, 0);
880 vmem_set_import(kernel_arena, kva_import, NULL, NULL, quantum);
881
882 for (domain = 0; domain < vm_ndomains; domain++) {
883 /*
884 * Initialize the per-domain arenas. These are used to color
885 * the KVA space in a way that ensures that virtual large pages
886 * are backed by memory from the same physical domain,
887 * maximizing the potential for superpage promotion.
888 */
889 vm_dom[domain].vmd_kernel_arena = vmem_create(
890 "kernel arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
891 vmem_set_import(vm_dom[domain].vmd_kernel_arena,
892 kva_import_domain, NULL, kernel_arena, quantum);
893
894 /*
895 * In architectures with superpages, maintain separate arenas
896 * for allocations with permissions that differ from the
897 * "standard" read/write permissions used for kernel memory
898 * and pages that are never released, so as not to inhibit
899 * superpage promotion.
900 *
901 * Use the base import quantum since these arenas are rarely
902 * used.
903 */
904 #if VM_NRESERVLEVEL > 0
905 vm_dom[domain].vmd_kernel_rwx_arena = vmem_create(
906 "kernel rwx arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
907 vm_dom[domain].vmd_kernel_nofree_arena = vmem_create(
908 "kernel NOFREE arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
909 vmem_set_import(vm_dom[domain].vmd_kernel_rwx_arena,
910 kva_import_domain, (vmem_release_t *)vmem_xfree,
911 kernel_arena, KVA_QUANTUM);
912 vmem_set_import(vm_dom[domain].vmd_kernel_nofree_arena,
913 kva_import_domain, (vmem_release_t *)vmem_xfree,
914 kernel_arena, KVA_QUANTUM);
915 #else
916 vm_dom[domain].vmd_kernel_rwx_arena =
917 vm_dom[domain].vmd_kernel_arena;
918 vm_dom[domain].vmd_kernel_nofree_arena =
919 vm_dom[domain].vmd_kernel_arena;
920 #endif
921 }
922
923 /*
924 * This must be the very first call so that the virtual address
925 * space used for early allocations is properly marked used in
926 * the map.
927 */
928 uma_startup2();
929 }
930
931 /*
932 * kmem_bootstrap_free:
933 *
934 * Free pages backing preloaded data (e.g., kernel modules) to the
935 * system. Currently only supported on platforms that create a
936 * vm_phys segment for preloaded data.
937 */
938 void
kmem_bootstrap_free(vm_offset_t start,vm_size_t size)939 kmem_bootstrap_free(vm_offset_t start, vm_size_t size)
940 {
941 #if defined(__i386__) || defined(__amd64__)
942 struct vm_domain *vmd;
943 vm_offset_t end, va;
944 vm_paddr_t pa;
945 vm_page_t m;
946
947 end = trunc_page(start + size);
948 start = round_page(start);
949
950 #ifdef __amd64__
951 /*
952 * Preloaded files do not have execute permissions by default on amd64.
953 * Restore the default permissions to ensure that the direct map alias
954 * is updated.
955 */
956 pmap_change_prot(start, end - start, VM_PROT_RW);
957 #endif
958 for (va = start; va < end; va += PAGE_SIZE) {
959 pa = pmap_kextract(va);
960 m = PHYS_TO_VM_PAGE(pa);
961
962 vmd = vm_pagequeue_domain(m);
963 vm_domain_free_lock(vmd);
964 vm_phys_free_pages(m, m->pool, 0);
965 vm_domain_free_unlock(vmd);
966
967 vm_domain_freecnt_inc(vmd, 1);
968 vm_cnt.v_page_count++;
969 }
970 pmap_remove(kernel_pmap, start, end);
971 (void)vmem_add(kernel_arena, start, end - start, M_WAITOK);
972 #endif
973 }
974
975 #ifdef PMAP_WANT_ACTIVE_CPUS_NAIVE
976 void
pmap_active_cpus(pmap_t pmap,cpuset_t * res)977 pmap_active_cpus(pmap_t pmap, cpuset_t *res)
978 {
979 struct thread *td;
980 struct proc *p;
981 struct vmspace *vm;
982 int c;
983
984 CPU_ZERO(res);
985 CPU_FOREACH(c) {
986 td = cpuid_to_pcpu[c]->pc_curthread;
987 p = td->td_proc;
988 if (p == NULL)
989 continue;
990 vm = vmspace_acquire_ref(p);
991 if (vm == NULL)
992 continue;
993 if (pmap == vmspace_pmap(vm))
994 CPU_SET(c, res);
995 vmspace_free(vm);
996 }
997 }
998 #endif
999
1000 /*
1001 * Allow userspace to directly trigger the VM drain routine for testing
1002 * purposes.
1003 */
1004 static int
debug_vm_lowmem(SYSCTL_HANDLER_ARGS)1005 debug_vm_lowmem(SYSCTL_HANDLER_ARGS)
1006 {
1007 int error, i;
1008
1009 i = 0;
1010 error = sysctl_handle_int(oidp, &i, 0, req);
1011 if (error != 0)
1012 return (error);
1013 if ((i & ~(VM_LOW_KMEM | VM_LOW_PAGES)) != 0)
1014 return (EINVAL);
1015 if (i != 0)
1016 EVENTHANDLER_INVOKE(vm_lowmem, i);
1017 return (0);
1018 }
1019 SYSCTL_PROC(_debug, OID_AUTO, vm_lowmem,
1020 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, debug_vm_lowmem, "I",
1021 "set to trigger vm_lowmem event with given flags");
1022
1023 static int
debug_uma_reclaim(SYSCTL_HANDLER_ARGS)1024 debug_uma_reclaim(SYSCTL_HANDLER_ARGS)
1025 {
1026 int error, i;
1027
1028 i = 0;
1029 error = sysctl_handle_int(oidp, &i, 0, req);
1030 if (error != 0 || req->newptr == NULL)
1031 return (error);
1032 if (i != UMA_RECLAIM_TRIM && i != UMA_RECLAIM_DRAIN &&
1033 i != UMA_RECLAIM_DRAIN_CPU)
1034 return (EINVAL);
1035 uma_reclaim(i);
1036 return (0);
1037 }
1038 SYSCTL_PROC(_debug, OID_AUTO, uma_reclaim,
1039 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, debug_uma_reclaim, "I",
1040 "set to generate request to reclaim uma caches");
1041
1042 static int
debug_uma_reclaim_domain(SYSCTL_HANDLER_ARGS)1043 debug_uma_reclaim_domain(SYSCTL_HANDLER_ARGS)
1044 {
1045 int domain, error, request;
1046
1047 request = 0;
1048 error = sysctl_handle_int(oidp, &request, 0, req);
1049 if (error != 0 || req->newptr == NULL)
1050 return (error);
1051
1052 domain = request >> 4;
1053 request &= 0xf;
1054 if (request != UMA_RECLAIM_TRIM && request != UMA_RECLAIM_DRAIN &&
1055 request != UMA_RECLAIM_DRAIN_CPU)
1056 return (EINVAL);
1057 if (domain < 0 || domain >= vm_ndomains)
1058 return (EINVAL);
1059 uma_reclaim_domain(request, domain);
1060 return (0);
1061 }
1062 SYSCTL_PROC(_debug, OID_AUTO, uma_reclaim_domain,
1063 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0,
1064 debug_uma_reclaim_domain, "I",
1065 "");
1066