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 if (vm_domainset_iter_policy_init(&di, ds, &domain, &flags) != 0)
327 return (NULL);
328
329 do {
330 addr = kmem_alloc_attr_domain(domain, size, flags, low, high,
331 memattr);
332 if (addr != NULL)
333 break;
334 if (start_segind == -1)
335 start_segind = vm_phys_lookup_segind(low);
336 if (vm_phys_find_range(bounds, start_segind, domain,
337 atop(round_page(size)), low, high) == -1) {
338 vm_domainset_iter_ignore(&di, domain);
339 }
340 } while (vm_domainset_iter_policy(&di, &domain) == 0);
341
342 return (addr);
343 }
344
345 /*
346 * Allocates a region from the kernel address map and physically
347 * contiguous pages within the specified address range to the kernel
348 * object. Creates a wired mapping from this region to these pages, and
349 * returns the region's starting virtual address. If M_ZERO is specified
350 * through the given flags, then the pages are zeroed before they are
351 * mapped.
352 */
353 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)354 kmem_alloc_contig_domain(int domain, vm_size_t size, int flags, vm_paddr_t low,
355 vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
356 vm_memattr_t memattr)
357 {
358 vmem_t *vmem;
359 vm_object_t object;
360 vm_offset_t addr, offset, tmp;
361 vm_page_t end_m, m;
362 vm_size_t asize;
363 u_long npages;
364 int pflags;
365
366 object = kernel_object;
367 asize = round_page(size);
368 vmem = vm_dom[domain].vmd_kernel_arena;
369 if (vmem_alloc(vmem, asize, flags | M_BESTFIT, &addr))
370 return (NULL);
371 offset = addr - VM_MIN_KERNEL_ADDRESS;
372 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
373 npages = atop(asize);
374 VM_OBJECT_WLOCK(object);
375 m = kmem_alloc_contig_pages(object, atop(offset), domain,
376 pflags, npages, low, high, alignment, boundary, memattr);
377 if (m == NULL) {
378 VM_OBJECT_WUNLOCK(object);
379 vmem_free(vmem, addr, asize);
380 return (NULL);
381 }
382 KASSERT(vm_page_domain(m) == domain,
383 ("kmem_alloc_contig_domain: Domain mismatch %d != %d",
384 vm_page_domain(m), domain));
385 end_m = m + npages;
386 tmp = addr;
387 for (; m < end_m; m++) {
388 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
389 pmap_zero_page(m);
390 vm_page_valid(m);
391 pmap_enter(kernel_pmap, tmp, m, VM_PROT_RW,
392 VM_PROT_RW | PMAP_ENTER_WIRED, 0);
393 tmp += PAGE_SIZE;
394 }
395 VM_OBJECT_WUNLOCK(object);
396 kmem_alloc_san(addr, size, asize, flags);
397 return ((void *)addr);
398 }
399
400 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)401 kmem_alloc_contig(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high,
402 u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr)
403 {
404
405 return (kmem_alloc_contig_domainset(DOMAINSET_RR(), size, flags, low,
406 high, alignment, boundary, memattr));
407 }
408
409 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)410 kmem_alloc_contig_domainset(struct domainset *ds, vm_size_t size, int flags,
411 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
412 vm_memattr_t memattr)
413 {
414 struct vm_domainset_iter di;
415 vm_page_t bounds[2];
416 void *addr;
417 int domain;
418 int start_segind;
419
420 start_segind = -1;
421
422 if (vm_domainset_iter_policy_init(&di, ds, &domain, &flags))
423 return (NULL);
424
425 do {
426 addr = kmem_alloc_contig_domain(domain, size, flags, low, high,
427 alignment, boundary, memattr);
428 if (addr != NULL)
429 break;
430 if (start_segind == -1)
431 start_segind = vm_phys_lookup_segind(low);
432 if (vm_phys_find_range(bounds, start_segind, domain,
433 atop(round_page(size)), low, high) == -1) {
434 vm_domainset_iter_ignore(&di, domain);
435 }
436 } while (vm_domainset_iter_policy(&di, &domain) == 0);
437
438 return (addr);
439 }
440
441 /*
442 * kmem_subinit:
443 *
444 * Initializes a map to manage a subrange
445 * of the kernel virtual address space.
446 *
447 * Arguments are as follows:
448 *
449 * parent Map to take range from
450 * min, max Returned endpoints of map
451 * size Size of range to find
452 * superpage_align Request that min is superpage aligned
453 */
454 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)455 kmem_subinit(vm_map_t map, vm_map_t parent, vm_offset_t *min, vm_offset_t *max,
456 vm_size_t size, bool superpage_align)
457 {
458 int ret;
459
460 size = round_page(size);
461
462 *min = vm_map_min(parent);
463 ret = vm_map_find(parent, NULL, 0, min, size, 0, superpage_align ?
464 VMFS_SUPER_SPACE : VMFS_ANY_SPACE, VM_PROT_ALL, VM_PROT_ALL,
465 MAP_ACC_NO_CHARGE);
466 if (ret != KERN_SUCCESS)
467 panic("kmem_subinit: bad status return of %d", ret);
468 *max = *min + size;
469 vm_map_init(map, vm_map_pmap(parent), *min, *max);
470 if (vm_map_submap(parent, *min, *max, map) != KERN_SUCCESS)
471 panic("kmem_subinit: unable to change range to submap");
472 }
473
474 /*
475 * kmem_malloc_domain:
476 *
477 * Allocate wired-down pages in the kernel's address space.
478 */
479 static void *
kmem_malloc_domain(int domain,vm_size_t size,int flags)480 kmem_malloc_domain(int domain, vm_size_t size, int flags)
481 {
482 vmem_t *arena;
483 vm_offset_t addr;
484 vm_size_t asize;
485 int rv;
486
487 if (__predict_true((flags & (M_EXEC | M_NEVERFREED)) == 0))
488 arena = vm_dom[domain].vmd_kernel_arena;
489 else if ((flags & M_EXEC) != 0)
490 arena = vm_dom[domain].vmd_kernel_rwx_arena;
491 else
492 arena = vm_dom[domain].vmd_kernel_nofree_arena;
493 asize = round_page(size);
494 if (vmem_alloc(arena, asize, flags | M_BESTFIT, &addr))
495 return (0);
496
497 rv = kmem_back_domain(domain, kernel_object, addr, asize, flags);
498 if (rv != KERN_SUCCESS) {
499 vmem_free(arena, addr, asize);
500 return (0);
501 }
502 kasan_mark((void *)addr, size, asize, KASAN_KMEM_REDZONE);
503 return ((void *)addr);
504 }
505
506 void *
kmem_malloc(vm_size_t size,int flags)507 kmem_malloc(vm_size_t size, int flags)
508 {
509 void * p;
510
511 TSENTER();
512 p = kmem_malloc_domainset(DOMAINSET_RR(), size, flags);
513 TSEXIT();
514 return (p);
515 }
516
517 void *
kmem_malloc_domainset(struct domainset * ds,vm_size_t size,int flags)518 kmem_malloc_domainset(struct domainset *ds, vm_size_t size, int flags)
519 {
520 struct vm_domainset_iter di;
521 void *addr;
522 int domain;
523
524 if (vm_domainset_iter_policy_init(&di, ds, &domain, &flags) != 0)
525 return (NULL);
526
527 do {
528 addr = kmem_malloc_domain(domain, size, flags);
529 if (addr != NULL)
530 break;
531 } while (vm_domainset_iter_policy(&di, &domain) == 0);
532
533 return (addr);
534 }
535
536 /*
537 * kmem_back_domain:
538 *
539 * Allocate physical pages from the specified domain for the specified
540 * virtual address range.
541 */
542 int
kmem_back_domain(int domain,vm_object_t object,vm_offset_t addr,vm_size_t size,int flags)543 kmem_back_domain(int domain, vm_object_t object, vm_offset_t addr,
544 vm_size_t size, int flags)
545 {
546 struct pctrie_iter pages;
547 vm_offset_t offset, i;
548 vm_page_t m;
549 vm_prot_t prot;
550 int pflags;
551
552 KASSERT(object == kernel_object,
553 ("kmem_back_domain: only supports kernel object."));
554
555 offset = addr - VM_MIN_KERNEL_ADDRESS;
556 pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
557 pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
558 if (flags & M_WAITOK)
559 pflags |= VM_ALLOC_WAITFAIL;
560 prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW;
561
562 i = 0;
563 vm_page_iter_init(&pages, object);
564 VM_OBJECT_WLOCK(object);
565 retry:
566 for (; i < size; i += PAGE_SIZE) {
567 m = vm_page_alloc_domain_iter(object, atop(offset + i),
568 domain, pflags, &pages);
569
570 /*
571 * Ran out of space, free everything up and return. Don't need
572 * to lock page queues here as we know that the pages we got
573 * aren't on any queues.
574 */
575 if (m == NULL) {
576 if ((flags & M_NOWAIT) == 0)
577 goto retry;
578 VM_OBJECT_WUNLOCK(object);
579 kmem_unback(object, addr, i);
580 return (KERN_NO_SPACE);
581 }
582 KASSERT(vm_page_domain(m) == domain,
583 ("kmem_back_domain: Domain mismatch %d != %d",
584 vm_page_domain(m), domain));
585 if (flags & M_ZERO && (m->flags & PG_ZERO) == 0)
586 pmap_zero_page(m);
587 KASSERT((m->oflags & VPO_UNMANAGED) != 0,
588 ("kmem_malloc: page %p is managed", m));
589 vm_page_valid(m);
590 pmap_enter(kernel_pmap, addr + i, m, prot,
591 prot | PMAP_ENTER_WIRED, 0);
592 if (__predict_false((prot & VM_PROT_EXECUTE) != 0))
593 m->oflags |= VPO_KMEM_EXEC;
594 }
595 VM_OBJECT_WUNLOCK(object);
596 kmem_alloc_san(addr, size, size, flags);
597 return (KERN_SUCCESS);
598 }
599
600 /*
601 * kmem_back:
602 *
603 * Allocate physical pages for the specified virtual address range.
604 */
605 int
kmem_back(vm_object_t object,vm_offset_t addr,vm_size_t size,int flags)606 kmem_back(vm_object_t object, vm_offset_t addr, vm_size_t size, int flags)
607 {
608 vm_offset_t end, next, start;
609 int domain, rv;
610
611 KASSERT(object == kernel_object,
612 ("kmem_back: only supports kernel object."));
613
614 for (start = addr, end = addr + size; addr < end; addr = next) {
615 /*
616 * We must ensure that pages backing a given large virtual page
617 * all come from the same physical domain.
618 */
619 if (vm_ndomains > 1) {
620 domain = (addr >> KVA_QUANTUM_SHIFT) % vm_ndomains;
621 while (VM_DOMAIN_EMPTY(domain))
622 domain++;
623 next = roundup2(addr + 1, KVA_QUANTUM);
624 if (next > end || next < start)
625 next = end;
626 } else {
627 domain = 0;
628 next = end;
629 }
630 rv = kmem_back_domain(domain, object, addr, next - addr, flags);
631 if (rv != KERN_SUCCESS) {
632 kmem_unback(object, start, addr - start);
633 break;
634 }
635 }
636 return (rv);
637 }
638
639 /*
640 * kmem_unback:
641 *
642 * Unmap and free the physical pages underlying the specified virtual
643 * address range.
644 *
645 * A physical page must exist within the specified object at each index
646 * that is being unmapped.
647 */
648 static struct vmem *
_kmem_unback(vm_object_t object,vm_offset_t addr,vm_size_t size)649 _kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size)
650 {
651 struct pctrie_iter pages;
652 struct vmem *arena;
653 vm_page_t m;
654 vm_offset_t end, offset;
655 int domain;
656
657 KASSERT(object == kernel_object,
658 ("kmem_unback: only supports kernel object."));
659
660 if (size == 0)
661 return (NULL);
662 pmap_remove(kernel_pmap, addr, addr + size);
663 offset = addr - VM_MIN_KERNEL_ADDRESS;
664 end = offset + size;
665 vm_page_iter_init(&pages, object);
666 VM_OBJECT_WLOCK(object);
667 m = vm_radix_iter_lookup(&pages, atop(offset));
668 domain = vm_page_domain(m);
669 if (__predict_true((m->oflags & VPO_KMEM_EXEC) == 0))
670 arena = vm_dom[domain].vmd_kernel_arena;
671 else
672 arena = vm_dom[domain].vmd_kernel_rwx_arena;
673 for (; offset < end; offset += PAGE_SIZE,
674 m = vm_radix_iter_lookup(&pages, atop(offset))) {
675 vm_page_xbusy_claim(m);
676 vm_page_unwire_noq(m);
677 vm_page_iter_free(&pages, m);
678 }
679 VM_OBJECT_WUNLOCK(object);
680
681 return (arena);
682 }
683
684 void
kmem_unback(vm_object_t object,vm_offset_t addr,vm_size_t size)685 kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size)
686 {
687
688 (void)_kmem_unback(object, addr, size);
689 }
690
691 /*
692 * kmem_free:
693 *
694 * Free memory allocated with kmem_malloc. The size must match the
695 * original allocation.
696 */
697 void
kmem_free(void * addr,vm_size_t size)698 kmem_free(void *addr, vm_size_t size)
699 {
700 struct vmem *arena;
701
702 size = round_page(size);
703 kasan_mark(addr, size, size, 0);
704 arena = _kmem_unback(kernel_object, (uintptr_t)addr, size);
705 if (arena != NULL)
706 vmem_free(arena, (uintptr_t)addr, size);
707 }
708
709 /*
710 * kmap_alloc_wait:
711 *
712 * Allocates pageable memory from a sub-map of the kernel. If the submap
713 * has no room, the caller sleeps waiting for more memory in the submap.
714 *
715 * This routine may block.
716 */
717 vm_offset_t
kmap_alloc_wait(vm_map_t map,vm_size_t size)718 kmap_alloc_wait(vm_map_t map, vm_size_t size)
719 {
720 vm_offset_t addr;
721
722 size = round_page(size);
723 if (!swap_reserve(size))
724 return (0);
725
726 for (;;) {
727 /*
728 * To make this work for more than one map, use the map's lock
729 * to lock out sleepers/wakers.
730 */
731 vm_map_lock(map);
732 addr = vm_map_findspace(map, vm_map_min(map), size);
733 if (addr + size <= vm_map_max(map))
734 break;
735 /* no space now; see if we can ever get space */
736 if (vm_map_max(map) - vm_map_min(map) < size) {
737 vm_map_unlock(map);
738 swap_release(size);
739 return (0);
740 }
741 vm_map_modflags(map, MAP_NEEDS_WAKEUP, 0);
742 vm_map_unlock_and_wait(map, 0);
743 }
744 vm_map_insert(map, NULL, 0, addr, addr + size, VM_PROT_RW, VM_PROT_RW,
745 MAP_ACC_CHARGED);
746 vm_map_unlock(map);
747 return (addr);
748 }
749
750 /*
751 * kmap_free_wakeup:
752 *
753 * Returns memory to a submap of the kernel, and wakes up any processes
754 * waiting for memory in that map.
755 */
756 void
kmap_free_wakeup(vm_map_t map,vm_offset_t addr,vm_size_t size)757 kmap_free_wakeup(vm_map_t map, vm_offset_t addr, vm_size_t size)
758 {
759
760 vm_map_lock(map);
761 (void) vm_map_delete(map, trunc_page(addr), round_page(addr + size));
762 if ((map->flags & MAP_NEEDS_WAKEUP) != 0) {
763 vm_map_modflags(map, 0, MAP_NEEDS_WAKEUP);
764 vm_map_wakeup(map);
765 }
766 vm_map_unlock(map);
767 }
768
769 void
kmem_init_zero_region(void)770 kmem_init_zero_region(void)
771 {
772 vm_offset_t addr, i;
773 vm_page_t m;
774
775 /*
776 * Map a single physical page of zeros to a larger virtual range.
777 * This requires less looping in places that want large amounts of
778 * zeros, while not using much more physical resources.
779 */
780 addr = kva_alloc(ZERO_REGION_SIZE);
781 m = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_ZERO |
782 VM_ALLOC_NOFREE);
783 for (i = 0; i < ZERO_REGION_SIZE; i += PAGE_SIZE)
784 pmap_qenter(addr + i, &m, 1);
785 pmap_protect(kernel_pmap, addr, addr + ZERO_REGION_SIZE, VM_PROT_READ);
786
787 zero_region = (const void *)addr;
788 }
789
790 /*
791 * Import KVA from the kernel map into the kernel arena.
792 */
793 static int
kva_import(void * unused,vmem_size_t size,int flags,vmem_addr_t * addrp)794 kva_import(void *unused, vmem_size_t size, int flags, vmem_addr_t *addrp)
795 {
796 vm_offset_t addr;
797 int result;
798
799 TSENTER();
800 KASSERT((size % KVA_QUANTUM) == 0,
801 ("kva_import: Size %jd is not a multiple of %d",
802 (intmax_t)size, (int)KVA_QUANTUM));
803 addr = vm_map_min(kernel_map);
804 result = vm_map_find(kernel_map, NULL, 0, &addr, size, 0,
805 VMFS_SUPER_SPACE, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
806 if (result != KERN_SUCCESS) {
807 TSEXIT();
808 return (ENOMEM);
809 }
810
811 *addrp = addr;
812
813 TSEXIT();
814 return (0);
815 }
816
817 /*
818 * Import KVA from a parent arena into a per-domain arena. Imports must be
819 * KVA_QUANTUM-aligned and a multiple of KVA_QUANTUM in size.
820 */
821 static int
kva_import_domain(void * arena,vmem_size_t size,int flags,vmem_addr_t * addrp)822 kva_import_domain(void *arena, vmem_size_t size, int flags, vmem_addr_t *addrp)
823 {
824
825 KASSERT((size % KVA_QUANTUM) == 0,
826 ("kva_import_domain: Size %jd is not a multiple of %d",
827 (intmax_t)size, (int)KVA_QUANTUM));
828 return (vmem_xalloc(arena, size, KVA_QUANTUM, 0, 0, VMEM_ADDR_MIN,
829 VMEM_ADDR_MAX, flags, addrp));
830 }
831
832 /*
833 * kmem_init:
834 *
835 * Create the kernel map; insert a mapping covering kernel text,
836 * data, bss, and all space allocated thus far (`boostrap' data). The
837 * new map will thus map the range between VM_MIN_KERNEL_ADDRESS and
838 * `start' as allocated, and the range between `start' and `end' as free.
839 * Create the kernel vmem arena and its per-domain children.
840 */
841 void
kmem_init(vm_offset_t start,vm_offset_t end)842 kmem_init(vm_offset_t start, vm_offset_t end)
843 {
844 vm_size_t quantum;
845 int domain;
846
847 vm_map_init_system(kernel_map, kernel_pmap, VM_MIN_KERNEL_ADDRESS, end);
848 vm_map_lock(kernel_map);
849 /* N.B.: cannot use kgdb to debug, starting with this assignment ... */
850 (void)vm_map_insert(kernel_map, NULL, 0,
851 #ifdef __amd64__
852 KERNBASE,
853 #else
854 VM_MIN_KERNEL_ADDRESS,
855 #endif
856 start, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
857 /* ... and ending with the completion of the above `insert' */
858
859 #ifdef __amd64__
860 /*
861 * Mark KVA used for the page array as allocated. Other platforms
862 * that handle vm_page_array allocation can simply adjust virtual_avail
863 * instead.
864 */
865 (void)vm_map_insert(kernel_map, NULL, 0, (vm_offset_t)vm_page_array,
866 (vm_offset_t)vm_page_array + round_2mpage(vm_page_array_size *
867 sizeof(struct vm_page)),
868 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
869 #endif
870 vm_map_unlock(kernel_map);
871
872 /*
873 * Use a large import quantum on NUMA systems. This helps minimize
874 * interleaving of superpages, reducing internal fragmentation within
875 * the per-domain arenas.
876 */
877 if (vm_ndomains > 1 && PMAP_HAS_DMAP)
878 quantum = KVA_NUMA_IMPORT_QUANTUM;
879 else
880 quantum = KVA_QUANTUM;
881
882 /*
883 * Initialize the kernel_arena. This can grow on demand.
884 */
885 vmem_init(kernel_arena, "kernel arena", 0, 0, PAGE_SIZE, 0, 0);
886 vmem_set_import(kernel_arena, kva_import, NULL, NULL, quantum);
887
888 for (domain = 0; domain < vm_ndomains; domain++) {
889 /*
890 * Initialize the per-domain arenas. These are used to color
891 * the KVA space in a way that ensures that virtual large pages
892 * are backed by memory from the same physical domain,
893 * maximizing the potential for superpage promotion.
894 */
895 vm_dom[domain].vmd_kernel_arena = vmem_create(
896 "kernel arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
897 vmem_set_import(vm_dom[domain].vmd_kernel_arena,
898 kva_import_domain, NULL, kernel_arena, quantum);
899
900 /*
901 * In architectures with superpages, maintain separate arenas
902 * for allocations with permissions that differ from the
903 * "standard" read/write permissions used for kernel memory
904 * and pages that are never released, so as not to inhibit
905 * superpage promotion.
906 *
907 * Use the base import quantum since these arenas are rarely
908 * used.
909 */
910 #if VM_NRESERVLEVEL > 0
911 vm_dom[domain].vmd_kernel_rwx_arena = vmem_create(
912 "kernel rwx arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
913 vm_dom[domain].vmd_kernel_nofree_arena = vmem_create(
914 "kernel NOFREE arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
915 vmem_set_import(vm_dom[domain].vmd_kernel_rwx_arena,
916 kva_import_domain, (vmem_release_t *)vmem_xfree,
917 kernel_arena, KVA_QUANTUM);
918 vmem_set_import(vm_dom[domain].vmd_kernel_nofree_arena,
919 kva_import_domain, (vmem_release_t *)vmem_xfree,
920 kernel_arena, KVA_QUANTUM);
921 #else
922 vm_dom[domain].vmd_kernel_rwx_arena =
923 vm_dom[domain].vmd_kernel_arena;
924 vm_dom[domain].vmd_kernel_nofree_arena =
925 vm_dom[domain].vmd_kernel_arena;
926 #endif
927 }
928
929 /*
930 * This must be the very first call so that the virtual address
931 * space used for early allocations is properly marked used in
932 * the map.
933 */
934 uma_startup2();
935 }
936
937 /*
938 * kmem_bootstrap_free:
939 *
940 * Free pages backing preloaded data (e.g., kernel modules) to the
941 * system. Currently only supported on platforms that create a
942 * vm_phys segment for preloaded data.
943 */
944 void
kmem_bootstrap_free(vm_offset_t start,vm_size_t size)945 kmem_bootstrap_free(vm_offset_t start, vm_size_t size)
946 {
947 #if defined(__i386__) || defined(__amd64__)
948 struct vm_domain *vmd;
949 vm_offset_t end, va;
950 vm_paddr_t pa;
951 vm_page_t m;
952
953 end = trunc_page(start + size);
954 start = round_page(start);
955
956 #ifdef __amd64__
957 /*
958 * Preloaded files do not have execute permissions by default on amd64.
959 * Restore the default permissions to ensure that the direct map alias
960 * is updated.
961 */
962 pmap_change_prot(start, end - start, VM_PROT_RW);
963 #endif
964 for (va = start; va < end; va += PAGE_SIZE) {
965 pa = pmap_kextract(va);
966 m = PHYS_TO_VM_PAGE(pa);
967
968 vmd = vm_pagequeue_domain(m);
969 vm_domain_free_lock(vmd);
970 vm_phys_free_pages(m, m->pool, 0);
971 vm_domain_free_unlock(vmd);
972
973 vm_domain_freecnt_inc(vmd, 1);
974 vm_cnt.v_page_count++;
975 }
976 pmap_remove(kernel_pmap, start, end);
977 (void)vmem_add(kernel_arena, start, end - start, M_WAITOK);
978 #endif
979 }
980
981 #ifdef PMAP_WANT_ACTIVE_CPUS_NAIVE
982 void
pmap_active_cpus(pmap_t pmap,cpuset_t * res)983 pmap_active_cpus(pmap_t pmap, cpuset_t *res)
984 {
985 struct thread *td;
986 struct proc *p;
987 struct vmspace *vm;
988 int c;
989
990 CPU_ZERO(res);
991 CPU_FOREACH(c) {
992 td = cpuid_to_pcpu[c]->pc_curthread;
993 p = td->td_proc;
994 if (p == NULL)
995 continue;
996 vm = vmspace_acquire_ref(p);
997 if (vm == NULL)
998 continue;
999 if (pmap == vmspace_pmap(vm))
1000 CPU_SET(c, res);
1001 vmspace_free(vm);
1002 }
1003 }
1004 #endif
1005
1006 /*
1007 * Allow userspace to directly trigger the VM drain routine for testing
1008 * purposes.
1009 */
1010 static int
debug_vm_lowmem(SYSCTL_HANDLER_ARGS)1011 debug_vm_lowmem(SYSCTL_HANDLER_ARGS)
1012 {
1013 int error, i;
1014
1015 i = 0;
1016 error = sysctl_handle_int(oidp, &i, 0, req);
1017 if (error != 0)
1018 return (error);
1019 if ((i & ~(VM_LOW_KMEM | VM_LOW_PAGES)) != 0)
1020 return (EINVAL);
1021 if (i != 0)
1022 EVENTHANDLER_INVOKE(vm_lowmem, i);
1023 return (0);
1024 }
1025 SYSCTL_PROC(_debug, OID_AUTO, vm_lowmem,
1026 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, debug_vm_lowmem, "I",
1027 "set to trigger vm_lowmem event with given flags");
1028
1029 static int
debug_uma_reclaim(SYSCTL_HANDLER_ARGS)1030 debug_uma_reclaim(SYSCTL_HANDLER_ARGS)
1031 {
1032 int error, i;
1033
1034 i = 0;
1035 error = sysctl_handle_int(oidp, &i, 0, req);
1036 if (error != 0 || req->newptr == NULL)
1037 return (error);
1038 if (i != UMA_RECLAIM_TRIM && i != UMA_RECLAIM_DRAIN &&
1039 i != UMA_RECLAIM_DRAIN_CPU)
1040 return (EINVAL);
1041 uma_reclaim(i);
1042 return (0);
1043 }
1044 SYSCTL_PROC(_debug, OID_AUTO, uma_reclaim,
1045 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, debug_uma_reclaim, "I",
1046 "set to generate request to reclaim uma caches");
1047
1048 static int
debug_uma_reclaim_domain(SYSCTL_HANDLER_ARGS)1049 debug_uma_reclaim_domain(SYSCTL_HANDLER_ARGS)
1050 {
1051 int domain, error, request;
1052
1053 request = 0;
1054 error = sysctl_handle_int(oidp, &request, 0, req);
1055 if (error != 0 || req->newptr == NULL)
1056 return (error);
1057
1058 domain = request >> 4;
1059 request &= 0xf;
1060 if (request != UMA_RECLAIM_TRIM && request != UMA_RECLAIM_DRAIN &&
1061 request != UMA_RECLAIM_DRAIN_CPU)
1062 return (EINVAL);
1063 if (domain < 0 || domain >= vm_ndomains)
1064 return (EINVAL);
1065 uma_reclaim_domain(request, domain);
1066 return (0);
1067 }
1068 SYSCTL_PROC(_debug, OID_AUTO, uma_reclaim_domain,
1069 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0,
1070 debug_uma_reclaim_domain, "I",
1071 "");
1072