xref: /freebsd/sys/vm/vm_glue.c (revision 8aac90f18aef7c9eea906c3ff9a001ca7b94f375)
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  * Permission to use, copy, modify and distribute this software and
39  * its documentation is hereby granted, provided that both the copyright
40  * notice and this permission notice appear in all copies of the
41  * software, derivative works or modified versions, and any portions
42  * thereof, and that both notices appear in supporting documentation.
43  *
44  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
45  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
46  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
47  *
48  * Carnegie Mellon requests users of this software to return to
49  *
50  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
51  *  School of Computer Science
52  *  Carnegie Mellon University
53  *  Pittsburgh PA 15213-3890
54  *
55  * any improvements or extensions that they make and grant Carnegie the
56  * rights to redistribute these changes.
57  */
58 
59 #include <sys/cdefs.h>
60 #include "opt_vm.h"
61 #include "opt_kstack_pages.h"
62 #include "opt_kstack_max_pages.h"
63 #include "opt_kstack_usage_prof.h"
64 
65 #include <sys/param.h>
66 #include <sys/systm.h>
67 #include <sys/asan.h>
68 #include <sys/domainset.h>
69 #include <sys/limits.h>
70 #include <sys/lock.h>
71 #include <sys/malloc.h>
72 #include <sys/msan.h>
73 #include <sys/mutex.h>
74 #include <sys/proc.h>
75 #include <sys/racct.h>
76 #include <sys/refcount.h>
77 #include <sys/resourcevar.h>
78 #include <sys/rwlock.h>
79 #include <sys/sched.h>
80 #include <sys/sf_buf.h>
81 #include <sys/shm.h>
82 #include <sys/smp.h>
83 #include <sys/vmmeter.h>
84 #include <sys/vmem.h>
85 #include <sys/sx.h>
86 #include <sys/sysctl.h>
87 #include <sys/kernel.h>
88 #include <sys/ktr.h>
89 #include <sys/unistd.h>
90 
91 #include <vm/uma.h>
92 #include <vm/vm.h>
93 #include <vm/vm_param.h>
94 #include <vm/pmap.h>
95 #include <vm/vm_domainset.h>
96 #include <vm/vm_map.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_pageout.h>
99 #include <vm/vm_pagequeue.h>
100 #include <vm/vm_object.h>
101 #include <vm/vm_kern.h>
102 #include <vm/vm_extern.h>
103 #include <vm/vm_pager.h>
104 #include <vm/swap_pager.h>
105 #include <vm/vm_phys.h>
106 
107 #include <machine/cpu.h>
108 
109 #if VM_NRESERVLEVEL > 0
110 #define KVA_KSTACK_QUANTUM_SHIFT (VM_LEVEL_0_ORDER + PAGE_SHIFT)
111 #else
112 #define KVA_KSTACK_QUANTUM_SHIFT (8 + PAGE_SHIFT)
113 #endif
114 #define KVA_KSTACK_QUANTUM (1ul << KVA_KSTACK_QUANTUM_SHIFT)
115 
116 /*
117  * MPSAFE
118  *
119  * WARNING!  This code calls vm_map_check_protection() which only checks
120  * the associated vm_map_entry range.  It does not determine whether the
121  * contents of the memory is actually readable or writable.  In most cases
122  * just checking the vm_map_entry is sufficient within the kernel's address
123  * space.
124  */
125 bool
126 kernacc(void *addr, int len, int rw)
127 {
128 	boolean_t rv;
129 	vm_offset_t saddr, eaddr;
130 	vm_prot_t prot;
131 
132 	KASSERT((rw & ~VM_PROT_ALL) == 0,
133 	    ("illegal ``rw'' argument to kernacc (%x)\n", rw));
134 
135 	if ((vm_offset_t)addr + len > vm_map_max(kernel_map) ||
136 	    (vm_offset_t)addr + len < (vm_offset_t)addr)
137 		return (false);
138 
139 	prot = rw;
140 	saddr = trunc_page((vm_offset_t)addr);
141 	eaddr = round_page((vm_offset_t)addr + len);
142 	vm_map_lock_read(kernel_map);
143 	rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot);
144 	vm_map_unlock_read(kernel_map);
145 	return (rv == TRUE);
146 }
147 
148 /*
149  * MPSAFE
150  *
151  * WARNING!  This code calls vm_map_check_protection() which only checks
152  * the associated vm_map_entry range.  It does not determine whether the
153  * contents of the memory is actually readable or writable.  vmapbuf(),
154  * vm_fault_quick(), or copyin()/copout()/su*()/fu*() functions should be
155  * used in conjunction with this call.
156  */
157 bool
158 useracc(void *addr, int len, int rw)
159 {
160 	boolean_t rv;
161 	vm_prot_t prot;
162 	vm_map_t map;
163 
164 	KASSERT((rw & ~VM_PROT_ALL) == 0,
165 	    ("illegal ``rw'' argument to useracc (%x)\n", rw));
166 	prot = rw;
167 	map = &curproc->p_vmspace->vm_map;
168 	if ((vm_offset_t)addr + len > vm_map_max(map) ||
169 	    (vm_offset_t)addr + len < (vm_offset_t)addr) {
170 		return (false);
171 	}
172 	vm_map_lock_read(map);
173 	rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr),
174 	    round_page((vm_offset_t)addr + len), prot);
175 	vm_map_unlock_read(map);
176 	return (rv == TRUE);
177 }
178 
179 int
180 vslock(void *addr, size_t len)
181 {
182 	vm_offset_t end, last, start;
183 	vm_size_t npages;
184 	int error;
185 
186 	last = (vm_offset_t)addr + len;
187 	start = trunc_page((vm_offset_t)addr);
188 	end = round_page(last);
189 	if (last < (vm_offset_t)addr || end < (vm_offset_t)addr)
190 		return (EINVAL);
191 	npages = atop(end - start);
192 	if (npages > vm_page_max_user_wired)
193 		return (ENOMEM);
194 	error = vm_map_wire(&curproc->p_vmspace->vm_map, start, end,
195 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
196 	if (error == KERN_SUCCESS) {
197 		curthread->td_vslock_sz += len;
198 		return (0);
199 	}
200 
201 	/*
202 	 * Return EFAULT on error to match copy{in,out}() behaviour
203 	 * rather than returning ENOMEM like mlock() would.
204 	 */
205 	return (EFAULT);
206 }
207 
208 void
209 vsunlock(void *addr, size_t len)
210 {
211 
212 	/* Rely on the parameter sanity checks performed by vslock(). */
213 	MPASS(curthread->td_vslock_sz >= len);
214 	curthread->td_vslock_sz -= len;
215 	(void)vm_map_unwire(&curproc->p_vmspace->vm_map,
216 	    trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len),
217 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
218 }
219 
220 /*
221  * Pin the page contained within the given object at the given offset.  If the
222  * page is not resident, allocate and load it using the given object's pager.
223  * Return the pinned page if successful; otherwise, return NULL.
224  */
225 static vm_page_t
226 vm_imgact_hold_page(vm_object_t object, vm_ooffset_t offset)
227 {
228 	vm_page_t m;
229 	vm_pindex_t pindex;
230 
231 	pindex = OFF_TO_IDX(offset);
232 	(void)vm_page_grab_valid_unlocked(&m, object, pindex,
233 	    VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
234 	return (m);
235 }
236 
237 /*
238  * Return a CPU private mapping to the page at the given offset within the
239  * given object.  The page is pinned before it is mapped.
240  */
241 struct sf_buf *
242 vm_imgact_map_page(vm_object_t object, vm_ooffset_t offset)
243 {
244 	vm_page_t m;
245 
246 	m = vm_imgact_hold_page(object, offset);
247 	if (m == NULL)
248 		return (NULL);
249 	sched_pin();
250 	return (sf_buf_alloc(m, SFB_CPUPRIVATE));
251 }
252 
253 /*
254  * Destroy the given CPU private mapping and unpin the page that it mapped.
255  */
256 void
257 vm_imgact_unmap_page(struct sf_buf *sf)
258 {
259 	vm_page_t m;
260 
261 	m = sf_buf_page(sf);
262 	sf_buf_free(sf);
263 	sched_unpin();
264 	vm_page_unwire(m, PQ_ACTIVE);
265 }
266 
267 void
268 vm_sync_icache(vm_map_t map, vm_offset_t va, vm_offset_t sz)
269 {
270 
271 	pmap_sync_icache(map->pmap, va, sz);
272 }
273 
274 static vm_object_t kstack_object;
275 static vm_object_t kstack_alt_object;
276 static uma_zone_t kstack_cache;
277 static int kstack_cache_size;
278 static vmem_t *vmd_kstack_arena[MAXMEMDOM];
279 
280 static int
281 sysctl_kstack_cache_size(SYSCTL_HANDLER_ARGS)
282 {
283 	int error, oldsize;
284 
285 	oldsize = kstack_cache_size;
286 	error = sysctl_handle_int(oidp, arg1, arg2, req);
287 	if (error == 0 && req->newptr && oldsize != kstack_cache_size)
288 		uma_zone_set_maxcache(kstack_cache, kstack_cache_size);
289 	return (error);
290 }
291 SYSCTL_PROC(_vm, OID_AUTO, kstack_cache_size,
292     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &kstack_cache_size, 0,
293     sysctl_kstack_cache_size, "IU", "Maximum number of cached kernel stacks");
294 
295 /*
296  *	Allocate a virtual address range from a domain kstack arena, following
297  *	the specified NUMA policy.
298  */
299 static vm_offset_t
300 vm_thread_alloc_kstack_kva(vm_size_t size, int domain)
301 {
302 #ifndef __ILP32__
303 	int rv;
304 	vmem_t *arena;
305 	vm_offset_t addr = 0;
306 
307 	size = round_page(size);
308 	/* Allocate from the kernel arena for non-standard kstack sizes. */
309 	if (size != ptoa(kstack_pages + KSTACK_GUARD_PAGES)) {
310 		arena = vm_dom[domain].vmd_kernel_arena;
311 	} else {
312 		arena = vmd_kstack_arena[domain];
313 	}
314 	rv = vmem_alloc(arena, size, M_BESTFIT | M_NOWAIT, &addr);
315 	if (rv == ENOMEM)
316 		return (0);
317 	KASSERT(atop(addr - VM_MIN_KERNEL_ADDRESS) %
318 	    (kstack_pages + KSTACK_GUARD_PAGES) == 0,
319 	    ("%s: allocated kstack KVA not aligned to multiple of kstack size",
320 	    __func__));
321 
322 	return (addr);
323 #else
324 	return (kva_alloc(size));
325 #endif
326 }
327 
328 /*
329  *	Release a region of kernel virtual memory
330  *	allocated from the kstack arena.
331  */
332 static __noinline void
333 vm_thread_free_kstack_kva(vm_offset_t addr, vm_size_t size, int domain)
334 {
335 	vmem_t *arena;
336 
337 	size = round_page(size);
338 #ifdef __ILP32__
339 	arena = kernel_arena;
340 #else
341 	arena = vmd_kstack_arena[domain];
342 	if (size != ptoa(kstack_pages + KSTACK_GUARD_PAGES)) {
343 		arena = vm_dom[domain].vmd_kernel_arena;
344 	}
345 #endif
346 	vmem_free(arena, addr, size);
347 }
348 
349 static vmem_size_t
350 vm_thread_kstack_import_quantum(void)
351 {
352 #ifndef __ILP32__
353 	/*
354 	 * The kstack_quantum is larger than KVA_QUANTUM to account
355 	 * for holes induced by guard pages.
356 	 */
357 	return (KVA_KSTACK_QUANTUM * (kstack_pages + KSTACK_GUARD_PAGES));
358 #else
359 	return (KVA_KSTACK_QUANTUM);
360 #endif
361 }
362 
363 /*
364  * Import KVA from a parent arena into the kstack arena. Imports must be
365  * a multiple of kernel stack pages + guard pages in size.
366  *
367  * Kstack VA allocations need to be aligned so that the linear KVA pindex
368  * is divisible by the total number of kstack VA pages. This is necessary to
369  * make vm_kstack_pindex work properly.
370  *
371  * We import a multiple of KVA_KSTACK_QUANTUM-sized region from the parent
372  * arena. The actual size used by the kstack arena is one kstack smaller to
373  * allow for the necessary alignment adjustments to be made.
374  */
375 static int
376 vm_thread_kstack_arena_import(void *arena, vmem_size_t size, int flags,
377     vmem_addr_t *addrp)
378 {
379 	int error, rem;
380 	size_t kpages = kstack_pages + KSTACK_GUARD_PAGES;
381 
382 	KASSERT(atop(size) % kpages == 0,
383 	    ("%s: Size %jd is not a multiple of kstack pages (%d)", __func__,
384 	    (intmax_t)size, (int)kpages));
385 
386 	error = vmem_xalloc(arena, vm_thread_kstack_import_quantum(),
387 	    KVA_KSTACK_QUANTUM, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX, flags,
388 	    addrp);
389 	if (error) {
390 		return (error);
391 	}
392 
393 	rem = atop(*addrp - VM_MIN_KERNEL_ADDRESS) % kpages;
394 	if (rem != 0) {
395 		/* Bump addr to next aligned address */
396 		*addrp = *addrp + (kpages - rem) * PAGE_SIZE;
397 	}
398 
399 	return (0);
400 }
401 
402 /*
403  * Release KVA from a parent arena into the kstack arena. Released imports must
404  * be a multiple of kernel stack pages + guard pages in size.
405  */
406 static void
407 vm_thread_kstack_arena_release(void *arena, vmem_addr_t addr, vmem_size_t size)
408 {
409 	int rem;
410 	size_t kpages __diagused = kstack_pages + KSTACK_GUARD_PAGES;
411 
412 	KASSERT(size % kpages == 0,
413 	    ("%s: Size %jd is not a multiple of kstack pages (%d)", __func__,
414 	    (intmax_t)size, (int)kpages));
415 
416 	KASSERT((addr - VM_MIN_KERNEL_ADDRESS) % kpages == 0,
417 	    ("%s: Address %p is not properly aligned (%p)", __func__,
418 		(void *)addr, (void *)VM_MIN_KERNEL_ADDRESS));
419 	/*
420 	 * If the address is not KVA_KSTACK_QUANTUM-aligned we have to decrement
421 	 * it to account for the shift in kva_import_kstack.
422 	 */
423 	rem = addr % KVA_KSTACK_QUANTUM;
424 	if (rem) {
425 		KASSERT(rem <= ptoa(kpages),
426 		    ("%s: rem > kpages (%d), (%d)", __func__, rem,
427 			(int)kpages));
428 		addr -= rem;
429 	}
430 	vmem_xfree(arena, addr, vm_thread_kstack_import_quantum());
431 }
432 
433 /*
434  * Create the kernel stack for a new thread.
435  */
436 static vm_offset_t
437 vm_thread_stack_create(struct domainset *ds, int pages)
438 {
439 	vm_page_t ma[KSTACK_MAX_PAGES];
440 	struct vm_domainset_iter di;
441 	int req = VM_ALLOC_NORMAL;
442 	vm_object_t obj;
443 	vm_offset_t ks;
444 	int domain, i;
445 
446 	obj = vm_thread_kstack_size_to_obj(pages);
447 	if (vm_ndomains > 1)
448 		obj->domain.dr_policy = ds;
449 	vm_domainset_iter_page_init(&di, obj, 0, &domain, &req);
450 	do {
451 		/*
452 		 * Get a kernel virtual address for this thread's kstack.
453 		 */
454 		ks = vm_thread_alloc_kstack_kva(ptoa(pages + KSTACK_GUARD_PAGES),
455 		    domain);
456 		if (ks == 0)
457 			continue;
458 		ks += ptoa(KSTACK_GUARD_PAGES);
459 
460 		/*
461 		 * Allocate physical pages to back the stack.
462 		 */
463 		if (vm_thread_stack_back(ks, ma, pages, req, domain) != 0) {
464 			vm_thread_free_kstack_kva(ks - ptoa(KSTACK_GUARD_PAGES),
465 			    ptoa(pages + KSTACK_GUARD_PAGES), domain);
466 			continue;
467 		}
468 		if (KSTACK_GUARD_PAGES != 0) {
469 			pmap_qremove(ks - ptoa(KSTACK_GUARD_PAGES),
470 			    KSTACK_GUARD_PAGES);
471 		}
472 		for (i = 0; i < pages; i++)
473 			vm_page_valid(ma[i]);
474 		pmap_qenter(ks, ma, pages);
475 		return (ks);
476 	} while (vm_domainset_iter_page(&di, obj, &domain) == 0);
477 
478 	return (0);
479 }
480 
481 static __noinline void
482 vm_thread_stack_dispose(vm_offset_t ks, int pages)
483 {
484 	vm_page_t m;
485 	vm_pindex_t pindex;
486 	int i, domain;
487 	vm_object_t obj = vm_thread_kstack_size_to_obj(pages);
488 
489 	pindex = vm_kstack_pindex(ks, pages);
490 	domain = vm_phys_domain(vtophys(ks));
491 	pmap_qremove(ks, pages);
492 	VM_OBJECT_WLOCK(obj);
493 	for (i = 0; i < pages; i++) {
494 		m = vm_page_lookup(obj, pindex + i);
495 		if (m == NULL)
496 			panic("%s: kstack already missing?", __func__);
497 		KASSERT(vm_page_domain(m) == domain,
498 		    ("%s: page %p domain mismatch, expected %d got %d",
499 		    __func__, m, domain, vm_page_domain(m)));
500 		vm_page_xbusy_claim(m);
501 		vm_page_unwire_noq(m);
502 		vm_page_free(m);
503 	}
504 	VM_OBJECT_WUNLOCK(obj);
505 	kasan_mark((void *)ks, ptoa(pages), ptoa(pages), 0);
506 	vm_thread_free_kstack_kva(ks - (KSTACK_GUARD_PAGES * PAGE_SIZE),
507 	    ptoa(pages + KSTACK_GUARD_PAGES), domain);
508 }
509 
510 /*
511  * Allocate the kernel stack for a new thread.
512  */
513 int
514 vm_thread_new(struct thread *td, int pages)
515 {
516 	vm_offset_t ks;
517 	u_short ks_domain;
518 
519 	/* Bounds check */
520 	if (pages <= 1)
521 		pages = kstack_pages;
522 	else if (pages > KSTACK_MAX_PAGES)
523 		pages = KSTACK_MAX_PAGES;
524 
525 	ks = 0;
526 	if (pages == kstack_pages && kstack_cache != NULL)
527 		ks = (vm_offset_t)uma_zalloc(kstack_cache, M_NOWAIT);
528 
529 	/*
530 	 * Ensure that kstack objects can draw pages from any memory
531 	 * domain.  Otherwise a local memory shortage can block a process
532 	 * swap-in.
533 	 */
534 	if (ks == 0)
535 		ks = vm_thread_stack_create(DOMAINSET_PREF(PCPU_GET(domain)),
536 		    pages);
537 	if (ks == 0)
538 		return (0);
539 
540 	ks_domain = vm_phys_domain(vtophys(ks));
541 	KASSERT(ks_domain >= 0 && ks_domain < vm_ndomains,
542 	    ("%s: invalid domain for kstack %p", __func__, (void *)ks));
543 	td->td_kstack = ks;
544 	td->td_kstack_pages = pages;
545 	td->td_kstack_domain = ks_domain;
546 	return (1);
547 }
548 
549 /*
550  * Dispose of a thread's kernel stack.
551  */
552 void
553 vm_thread_dispose(struct thread *td)
554 {
555 	vm_offset_t ks;
556 	int pages;
557 
558 	pages = td->td_kstack_pages;
559 	ks = td->td_kstack;
560 	td->td_kstack = 0;
561 	td->td_kstack_pages = 0;
562 	td->td_kstack_domain = MAXMEMDOM;
563 	if (pages == kstack_pages) {
564 		kasan_mark((void *)ks, 0, ptoa(pages), KASAN_KSTACK_FREED);
565 		uma_zfree(kstack_cache, (void *)ks);
566 	} else {
567 		vm_thread_stack_dispose(ks, pages);
568 	}
569 }
570 
571 /*
572  * Calculate kstack pindex.
573  *
574  * Uses a non-identity mapping if guard pages are
575  * active to avoid pindex holes in the kstack object.
576  */
577 vm_pindex_t
578 vm_kstack_pindex(vm_offset_t ks, int kpages)
579 {
580 	vm_pindex_t pindex = atop(ks - VM_MIN_KERNEL_ADDRESS);
581 
582 #ifdef __ILP32__
583 	return (pindex);
584 #else
585 	/*
586 	 * Return the linear pindex if guard pages aren't active or if we are
587 	 * allocating a non-standard kstack size.
588 	 */
589 	if (KSTACK_GUARD_PAGES == 0 || kpages != kstack_pages) {
590 		return (pindex);
591 	}
592 	KASSERT(pindex % (kpages + KSTACK_GUARD_PAGES) >= KSTACK_GUARD_PAGES,
593 	    ("%s: Attempting to calculate kstack guard page pindex", __func__));
594 
595 	return (pindex -
596 	    (pindex / (kpages + KSTACK_GUARD_PAGES) + 1) * KSTACK_GUARD_PAGES);
597 #endif
598 }
599 
600 /*
601  * Allocate physical pages, following the specified NUMA policy, to back a
602  * kernel stack.
603  */
604 int
605 vm_thread_stack_back(vm_offset_t ks, vm_page_t ma[], int npages, int req_class,
606     int domain)
607 {
608 	vm_object_t obj = vm_thread_kstack_size_to_obj(npages);
609 	vm_pindex_t pindex;
610 	vm_page_t m;
611 	int n;
612 
613 	pindex = vm_kstack_pindex(ks, npages);
614 
615 	VM_OBJECT_WLOCK(obj);
616 	for (n = 0; n < npages;) {
617 		m = vm_page_grab(obj, pindex + n,
618 		    VM_ALLOC_NOCREAT | VM_ALLOC_WIRED);
619 		if (m == NULL) {
620 			m = vm_page_alloc_domain(obj, pindex + n, domain,
621 			    req_class | VM_ALLOC_WIRED);
622 		}
623 		if (m == NULL)
624 			break;
625 		ma[n++] = m;
626 	}
627 	if (n < npages)
628 		goto cleanup;
629 	VM_OBJECT_WUNLOCK(obj);
630 
631 	return (0);
632 cleanup:
633 	for (int i = 0; i < n; i++) {
634 		m = ma[i];
635 		(void)vm_page_unwire_noq(m);
636 		vm_page_free(m);
637 	}
638 	VM_OBJECT_WUNLOCK(obj);
639 
640 	return (ENOMEM);
641 }
642 
643 vm_object_t
644 vm_thread_kstack_size_to_obj(int npages)
645 {
646 	return (npages == kstack_pages ? kstack_object : kstack_alt_object);
647 }
648 
649 static int
650 kstack_import(void *arg, void **store, int cnt, int domain, int flags)
651 {
652 	struct domainset *ds;
653 	int i;
654 
655 	if (domain == UMA_ANYDOMAIN)
656 		ds = DOMAINSET_RR();
657 	else
658 		ds = DOMAINSET_PREF(domain);
659 
660 	for (i = 0; i < cnt; i++) {
661 		store[i] = (void *)vm_thread_stack_create(ds, kstack_pages);
662 		if (store[i] == NULL)
663 			break;
664 	}
665 	return (i);
666 }
667 
668 static void
669 kstack_release(void *arg, void **store, int cnt)
670 {
671 	vm_offset_t ks;
672 	int i;
673 
674 	for (i = 0; i < cnt; i++) {
675 		ks = (vm_offset_t)store[i];
676 		vm_thread_stack_dispose(ks, kstack_pages);
677 	}
678 }
679 
680 static void
681 kstack_cache_init(void *null)
682 {
683 	vm_size_t kstack_quantum;
684 	int domain;
685 
686 	kstack_object = vm_object_allocate(OBJT_SWAP,
687 	    atop(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS));
688 	kstack_cache = uma_zcache_create("kstack_cache",
689 	    kstack_pages * PAGE_SIZE, NULL, NULL, NULL, NULL,
690 	    kstack_import, kstack_release, NULL,
691 	    UMA_ZONE_FIRSTTOUCH);
692 	kstack_cache_size = imax(128, mp_ncpus * 4);
693 	uma_zone_set_maxcache(kstack_cache, kstack_cache_size);
694 
695 	kstack_alt_object = vm_object_allocate(OBJT_SWAP,
696 	    atop(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS));
697 
698 	kstack_quantum = vm_thread_kstack_import_quantum();
699 	/*
700 	 * Reduce size used by the kstack arena to allow for
701 	 * alignment adjustments in vm_thread_kstack_arena_import.
702 	 */
703 	kstack_quantum -= (kstack_pages + KSTACK_GUARD_PAGES) * PAGE_SIZE;
704 	/*
705 	 * Create the kstack_arena for each domain and set kernel_arena as
706 	 * parent.
707 	 */
708 	for (domain = 0; domain < vm_ndomains; domain++) {
709 		vmd_kstack_arena[domain] = vmem_create("kstack arena", 0, 0,
710 		    PAGE_SIZE, 0, M_WAITOK);
711 		KASSERT(vmd_kstack_arena[domain] != NULL,
712 		    ("%s: failed to create domain %d kstack_arena", __func__,
713 		    domain));
714 		vmem_set_import(vmd_kstack_arena[domain],
715 		    vm_thread_kstack_arena_import,
716 		    vm_thread_kstack_arena_release,
717 		    vm_dom[domain].vmd_kernel_arena, kstack_quantum);
718 	}
719 }
720 SYSINIT(vm_kstacks, SI_SUB_KMEM, SI_ORDER_ANY, kstack_cache_init, NULL);
721 
722 #ifdef KSTACK_USAGE_PROF
723 /*
724  * Track maximum stack used by a thread in kernel.
725  */
726 static int max_kstack_used;
727 
728 SYSCTL_INT(_debug, OID_AUTO, max_kstack_used, CTLFLAG_RD,
729     &max_kstack_used, 0,
730     "Maximum stack depth used by a thread in kernel");
731 
732 void
733 intr_prof_stack_use(struct thread *td, struct trapframe *frame)
734 {
735 	vm_offset_t stack_top;
736 	vm_offset_t current;
737 	int used, prev_used;
738 
739 	/*
740 	 * Testing for interrupted kernel mode isn't strictly
741 	 * needed. It optimizes the execution, since interrupts from
742 	 * usermode will have only the trap frame on the stack.
743 	 */
744 	if (TRAPF_USERMODE(frame))
745 		return;
746 
747 	stack_top = td->td_kstack + td->td_kstack_pages * PAGE_SIZE;
748 	current = (vm_offset_t)(uintptr_t)&stack_top;
749 
750 	/*
751 	 * Try to detect if interrupt is using kernel thread stack.
752 	 * Hardware could use a dedicated stack for interrupt handling.
753 	 */
754 	if (stack_top <= current || current < td->td_kstack)
755 		return;
756 
757 	used = stack_top - current;
758 	for (;;) {
759 		prev_used = max_kstack_used;
760 		if (prev_used >= used)
761 			break;
762 		if (atomic_cmpset_int(&max_kstack_used, prev_used, used))
763 			break;
764 	}
765 }
766 #endif /* KSTACK_USAGE_PROF */
767 
768 /*
769  * Implement fork's actions on an address space.
770  * Here we arrange for the address space to be copied or referenced,
771  * allocate a user struct (pcb and kernel stack), then call the
772  * machine-dependent layer to fill those in and make the new process
773  * ready to run.  The new process is set up so that it returns directly
774  * to user mode to avoid stack copying and relocation problems.
775  */
776 int
777 vm_forkproc(struct thread *td, struct proc *p2, struct thread *td2,
778     struct vmspace *vm2, int flags)
779 {
780 	struct proc *p1 = td->td_proc;
781 	struct domainset *dset;
782 	int error;
783 
784 	if ((flags & RFPROC) == 0) {
785 		/*
786 		 * Divorce the memory, if it is shared, essentially
787 		 * this changes shared memory amongst threads, into
788 		 * COW locally.
789 		 */
790 		if ((flags & RFMEM) == 0) {
791 			error = vmspace_unshare(p1);
792 			if (error)
793 				return (error);
794 		}
795 		cpu_fork(td, p2, td2, flags);
796 		return (0);
797 	}
798 
799 	if (flags & RFMEM) {
800 		p2->p_vmspace = p1->p_vmspace;
801 		refcount_acquire(&p1->p_vmspace->vm_refcnt);
802 	}
803 	dset = td2->td_domain.dr_policy;
804 	while (vm_page_count_severe_set(&dset->ds_mask)) {
805 		vm_wait_doms(&dset->ds_mask, 0);
806 	}
807 
808 	if ((flags & RFMEM) == 0) {
809 		p2->p_vmspace = vm2;
810 		if (p1->p_vmspace->vm_shm)
811 			shmfork(p1, p2);
812 	}
813 
814 	/*
815 	 * cpu_fork will copy and update the pcb, set up the kernel stack,
816 	 * and make the child ready to run.
817 	 */
818 	cpu_fork(td, p2, td2, flags);
819 	return (0);
820 }
821 
822 /*
823  * Called after process has been wait(2)'ed upon and is being reaped.
824  * The idea is to reclaim resources that we could not reclaim while
825  * the process was still executing.
826  */
827 void
828 vm_waitproc(struct proc *p)
829 {
830 
831 	vmspace_exitfree(p);		/* and clean-out the vmspace */
832 }
833 
834 void
835 kick_proc0(void)
836 {
837 
838 	wakeup(&proc0);
839 }
840