xref: /freebsd/sys/vm/vm_glue.c (revision efe014e6b177ba1b0d70f7df751f5066d4eafe9b)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	from: @(#)vm_glue.c	8.6 (Berkeley) 1/5/94
35  *
36  *
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
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 #include <sys/cdefs.h>
62 __FBSDID("$FreeBSD$");
63 
64 #include "opt_vm.h"
65 #include "opt_kstack_pages.h"
66 #include "opt_kstack_max_pages.h"
67 #include "opt_kstack_usage_prof.h"
68 
69 #include <sys/param.h>
70 #include <sys/systm.h>
71 #include <sys/asan.h>
72 #include <sys/domainset.h>
73 #include <sys/limits.h>
74 #include <sys/lock.h>
75 #include <sys/malloc.h>
76 #include <sys/msan.h>
77 #include <sys/mutex.h>
78 #include <sys/proc.h>
79 #include <sys/racct.h>
80 #include <sys/refcount.h>
81 #include <sys/resourcevar.h>
82 #include <sys/rwlock.h>
83 #include <sys/sched.h>
84 #include <sys/sf_buf.h>
85 #include <sys/shm.h>
86 #include <sys/smp.h>
87 #include <sys/vmmeter.h>
88 #include <sys/vmem.h>
89 #include <sys/sx.h>
90 #include <sys/sysctl.h>
91 #include <sys/kernel.h>
92 #include <sys/ktr.h>
93 #include <sys/unistd.h>
94 
95 #include <vm/uma.h>
96 #include <vm/vm.h>
97 #include <vm/vm_param.h>
98 #include <vm/pmap.h>
99 #include <vm/vm_domainset.h>
100 #include <vm/vm_map.h>
101 #include <vm/vm_page.h>
102 #include <vm/vm_pageout.h>
103 #include <vm/vm_object.h>
104 #include <vm/vm_kern.h>
105 #include <vm/vm_extern.h>
106 #include <vm/vm_pager.h>
107 #include <vm/swap_pager.h>
108 
109 #include <machine/cpu.h>
110 
111 /*
112  * MPSAFE
113  *
114  * WARNING!  This code calls vm_map_check_protection() which only checks
115  * the associated vm_map_entry range.  It does not determine whether the
116  * contents of the memory is actually readable or writable.  In most cases
117  * just checking the vm_map_entry is sufficient within the kernel's address
118  * space.
119  */
120 int
121 kernacc(void *addr, int len, int rw)
122 {
123 	boolean_t rv;
124 	vm_offset_t saddr, eaddr;
125 	vm_prot_t prot;
126 
127 	KASSERT((rw & ~VM_PROT_ALL) == 0,
128 	    ("illegal ``rw'' argument to kernacc (%x)\n", rw));
129 
130 	if ((vm_offset_t)addr + len > vm_map_max(kernel_map) ||
131 	    (vm_offset_t)addr + len < (vm_offset_t)addr)
132 		return (FALSE);
133 
134 	prot = rw;
135 	saddr = trunc_page((vm_offset_t)addr);
136 	eaddr = round_page((vm_offset_t)addr + len);
137 	vm_map_lock_read(kernel_map);
138 	rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot);
139 	vm_map_unlock_read(kernel_map);
140 	return (rv == TRUE);
141 }
142 
143 /*
144  * MPSAFE
145  *
146  * WARNING!  This code calls vm_map_check_protection() which only checks
147  * the associated vm_map_entry range.  It does not determine whether the
148  * contents of the memory is actually readable or writable.  vmapbuf(),
149  * vm_fault_quick(), or copyin()/copout()/su*()/fu*() functions should be
150  * used in conjunction with this call.
151  */
152 int
153 useracc(void *addr, int len, int rw)
154 {
155 	boolean_t rv;
156 	vm_prot_t prot;
157 	vm_map_t map;
158 
159 	KASSERT((rw & ~VM_PROT_ALL) == 0,
160 	    ("illegal ``rw'' argument to useracc (%x)\n", rw));
161 	prot = rw;
162 	map = &curproc->p_vmspace->vm_map;
163 	if ((vm_offset_t)addr + len > vm_map_max(map) ||
164 	    (vm_offset_t)addr + len < (vm_offset_t)addr) {
165 		return (FALSE);
166 	}
167 	vm_map_lock_read(map);
168 	rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr),
169 	    round_page((vm_offset_t)addr + len), prot);
170 	vm_map_unlock_read(map);
171 	return (rv == TRUE);
172 }
173 
174 int
175 vslock(void *addr, size_t len)
176 {
177 	vm_offset_t end, last, start;
178 	vm_size_t npages;
179 	int error;
180 
181 	last = (vm_offset_t)addr + len;
182 	start = trunc_page((vm_offset_t)addr);
183 	end = round_page(last);
184 	if (last < (vm_offset_t)addr || end < (vm_offset_t)addr)
185 		return (EINVAL);
186 	npages = atop(end - start);
187 	if (npages > vm_page_max_user_wired)
188 		return (ENOMEM);
189 	error = vm_map_wire(&curproc->p_vmspace->vm_map, start, end,
190 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
191 	if (error == KERN_SUCCESS) {
192 		curthread->td_vslock_sz += len;
193 		return (0);
194 	}
195 
196 	/*
197 	 * Return EFAULT on error to match copy{in,out}() behaviour
198 	 * rather than returning ENOMEM like mlock() would.
199 	 */
200 	return (EFAULT);
201 }
202 
203 void
204 vsunlock(void *addr, size_t len)
205 {
206 
207 	/* Rely on the parameter sanity checks performed by vslock(). */
208 	MPASS(curthread->td_vslock_sz >= len);
209 	curthread->td_vslock_sz -= len;
210 	(void)vm_map_unwire(&curproc->p_vmspace->vm_map,
211 	    trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len),
212 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
213 }
214 
215 /*
216  * Pin the page contained within the given object at the given offset.  If the
217  * page is not resident, allocate and load it using the given object's pager.
218  * Return the pinned page if successful; otherwise, return NULL.
219  */
220 static vm_page_t
221 vm_imgact_hold_page(vm_object_t object, vm_ooffset_t offset)
222 {
223 	vm_page_t m;
224 	vm_pindex_t pindex;
225 
226 	pindex = OFF_TO_IDX(offset);
227 	(void)vm_page_grab_valid_unlocked(&m, object, pindex,
228 	    VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
229 	return (m);
230 }
231 
232 /*
233  * Return a CPU private mapping to the page at the given offset within the
234  * given object.  The page is pinned before it is mapped.
235  */
236 struct sf_buf *
237 vm_imgact_map_page(vm_object_t object, vm_ooffset_t offset)
238 {
239 	vm_page_t m;
240 
241 	m = vm_imgact_hold_page(object, offset);
242 	if (m == NULL)
243 		return (NULL);
244 	sched_pin();
245 	return (sf_buf_alloc(m, SFB_CPUPRIVATE));
246 }
247 
248 /*
249  * Destroy the given CPU private mapping and unpin the page that it mapped.
250  */
251 void
252 vm_imgact_unmap_page(struct sf_buf *sf)
253 {
254 	vm_page_t m;
255 
256 	m = sf_buf_page(sf);
257 	sf_buf_free(sf);
258 	sched_unpin();
259 	vm_page_unwire(m, PQ_ACTIVE);
260 }
261 
262 void
263 vm_sync_icache(vm_map_t map, vm_offset_t va, vm_offset_t sz)
264 {
265 
266 	pmap_sync_icache(map->pmap, va, sz);
267 }
268 
269 vm_object_t kstack_object;
270 static uma_zone_t kstack_cache;
271 static int kstack_cache_size;
272 
273 static int
274 sysctl_kstack_cache_size(SYSCTL_HANDLER_ARGS)
275 {
276 	int error, oldsize;
277 
278 	oldsize = kstack_cache_size;
279 	error = sysctl_handle_int(oidp, arg1, arg2, req);
280 	if (error == 0 && req->newptr && oldsize != kstack_cache_size)
281 		uma_zone_set_maxcache(kstack_cache, kstack_cache_size);
282 	return (error);
283 }
284 SYSCTL_PROC(_vm, OID_AUTO, kstack_cache_size,
285     CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &kstack_cache_size, 0,
286     sysctl_kstack_cache_size, "IU", "Maximum number of cached kernel stacks");
287 
288 /*
289  * Create the kernel stack (including pcb for i386) for a new thread.
290  */
291 static vm_offset_t
292 vm_thread_stack_create(struct domainset *ds, int pages)
293 {
294 	vm_page_t ma[KSTACK_MAX_PAGES];
295 	vm_offset_t ks;
296 	int i;
297 
298 	/*
299 	 * Get a kernel virtual address for this thread's kstack.
300 	 */
301 #if defined(__mips__)
302 	/*
303 	 * We need to align the kstack's mapped address to fit within
304 	 * a single TLB entry.
305 	 */
306 	if (vmem_xalloc(kernel_arena, (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE,
307 	    PAGE_SIZE * 2, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX,
308 	    M_BESTFIT | M_NOWAIT, &ks)) {
309 		ks = 0;
310 	}
311 #else
312 	ks = kva_alloc((pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
313 #endif
314 	if (ks == 0) {
315 		printf("%s: kstack allocation failed\n", __func__);
316 		return (0);
317 	}
318 
319 	if (KSTACK_GUARD_PAGES != 0) {
320 		pmap_qremove(ks, KSTACK_GUARD_PAGES);
321 		ks += KSTACK_GUARD_PAGES * PAGE_SIZE;
322 	}
323 
324 	/*
325 	 * Allocate physical pages to back the stack.
326 	 */
327 	vm_thread_stack_back(ds, ks, ma, pages, VM_ALLOC_NORMAL);
328 	for (i = 0; i < pages; i++)
329 		vm_page_valid(ma[i]);
330 	pmap_qenter(ks, ma, pages);
331 
332 	return (ks);
333 }
334 
335 static void
336 vm_thread_stack_dispose(vm_offset_t ks, int pages)
337 {
338 	vm_page_t m;
339 	vm_pindex_t pindex;
340 	int i;
341 
342 	pindex = atop(ks - VM_MIN_KERNEL_ADDRESS);
343 
344 	pmap_qremove(ks, pages);
345 	VM_OBJECT_WLOCK(kstack_object);
346 	for (i = 0; i < pages; i++) {
347 		m = vm_page_lookup(kstack_object, pindex + i);
348 		if (m == NULL)
349 			panic("%s: kstack already missing?", __func__);
350 		vm_page_xbusy_claim(m);
351 		vm_page_unwire_noq(m);
352 		vm_page_free(m);
353 	}
354 	VM_OBJECT_WUNLOCK(kstack_object);
355 	kasan_mark((void *)ks, ptoa(pages), ptoa(pages), 0);
356 	kva_free(ks - (KSTACK_GUARD_PAGES * PAGE_SIZE),
357 	    (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
358 }
359 
360 /*
361  * Allocate the kernel stack for a new thread.
362  */
363 int
364 vm_thread_new(struct thread *td, int pages)
365 {
366 	vm_offset_t ks;
367 
368 	/* Bounds check */
369 	if (pages <= 1)
370 		pages = kstack_pages;
371 	else if (pages > KSTACK_MAX_PAGES)
372 		pages = KSTACK_MAX_PAGES;
373 
374 	ks = 0;
375 	if (pages == kstack_pages && kstack_cache != NULL)
376 		ks = (vm_offset_t)uma_zalloc(kstack_cache, M_NOWAIT);
377 
378 	/*
379 	 * Ensure that kstack objects can draw pages from any memory
380 	 * domain.  Otherwise a local memory shortage can block a process
381 	 * swap-in.
382 	 */
383 	if (ks == 0)
384 		ks = vm_thread_stack_create(DOMAINSET_PREF(PCPU_GET(domain)),
385 		    pages);
386 	if (ks == 0)
387 		return (0);
388 	td->td_kstack = ks;
389 	td->td_kstack_pages = pages;
390 	kasan_mark((void *)ks, ptoa(pages), ptoa(pages), 0);
391 	kmsan_mark((void *)ks, ptoa(pages), KMSAN_STATE_UNINIT);
392 	return (1);
393 }
394 
395 /*
396  * Dispose of a thread's kernel stack.
397  */
398 void
399 vm_thread_dispose(struct thread *td)
400 {
401 	vm_offset_t ks;
402 	int pages;
403 
404 	pages = td->td_kstack_pages;
405 	ks = td->td_kstack;
406 	td->td_kstack = 0;
407 	td->td_kstack_pages = 0;
408 	kasan_mark((void *)ks, 0, ptoa(pages), KASAN_KSTACK_FREED);
409 	if (pages == kstack_pages)
410 		uma_zfree(kstack_cache, (void *)ks);
411 	else
412 		vm_thread_stack_dispose(ks, pages);
413 }
414 
415 /*
416  * Allocate physical pages, following the specified NUMA policy, to back a
417  * kernel stack.
418  */
419 void
420 vm_thread_stack_back(struct domainset *ds, vm_offset_t ks, vm_page_t ma[],
421     int npages, int req_class)
422 {
423 	vm_pindex_t pindex;
424 	int n;
425 
426 	pindex = atop(ks - VM_MIN_KERNEL_ADDRESS);
427 
428 	VM_OBJECT_WLOCK(kstack_object);
429 	for (n = 0; n < npages;) {
430 		if (vm_ndomains > 1)
431 			kstack_object->domain.dr_policy = ds;
432 
433 		/*
434 		 * Use WAITFAIL to force a reset of the domain selection policy
435 		 * if we had to sleep for pages.
436 		 */
437 		n += vm_page_grab_pages(kstack_object, pindex + n,
438 		    req_class | VM_ALLOC_WIRED | VM_ALLOC_WAITFAIL,
439 		    &ma[n], npages - n);
440 	}
441 	VM_OBJECT_WUNLOCK(kstack_object);
442 }
443 
444 static int
445 kstack_import(void *arg, void **store, int cnt, int domain, int flags)
446 {
447 	struct domainset *ds;
448 	int i;
449 
450 	if (domain == UMA_ANYDOMAIN)
451 		ds = DOMAINSET_RR();
452 	else
453 		ds = DOMAINSET_PREF(domain);
454 
455 	for (i = 0; i < cnt; i++) {
456 		store[i] = (void *)vm_thread_stack_create(ds, kstack_pages);
457 		if (store[i] == NULL)
458 			break;
459 	}
460 	return (i);
461 }
462 
463 static void
464 kstack_release(void *arg, void **store, int cnt)
465 {
466 	vm_offset_t ks;
467 	int i;
468 
469 	for (i = 0; i < cnt; i++) {
470 		ks = (vm_offset_t)store[i];
471 		vm_thread_stack_dispose(ks, kstack_pages);
472 	}
473 }
474 
475 static void
476 kstack_cache_init(void *null)
477 {
478 	kstack_object = vm_object_allocate(OBJT_SWAP,
479 	    atop(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS));
480 	kstack_cache = uma_zcache_create("kstack_cache",
481 	    kstack_pages * PAGE_SIZE, NULL, NULL, NULL, NULL,
482 	    kstack_import, kstack_release, NULL,
483 	    UMA_ZONE_FIRSTTOUCH);
484 	kstack_cache_size = imax(128, mp_ncpus * 4);
485 	uma_zone_set_maxcache(kstack_cache, kstack_cache_size);
486 }
487 SYSINIT(vm_kstacks, SI_SUB_KMEM, SI_ORDER_ANY, kstack_cache_init, NULL);
488 
489 #ifdef KSTACK_USAGE_PROF
490 /*
491  * Track maximum stack used by a thread in kernel.
492  */
493 static int max_kstack_used;
494 
495 SYSCTL_INT(_debug, OID_AUTO, max_kstack_used, CTLFLAG_RD,
496     &max_kstack_used, 0,
497     "Maxiumum stack depth used by a thread in kernel");
498 
499 void
500 intr_prof_stack_use(struct thread *td, struct trapframe *frame)
501 {
502 	vm_offset_t stack_top;
503 	vm_offset_t current;
504 	int used, prev_used;
505 
506 	/*
507 	 * Testing for interrupted kernel mode isn't strictly
508 	 * needed. It optimizes the execution, since interrupts from
509 	 * usermode will have only the trap frame on the stack.
510 	 */
511 	if (TRAPF_USERMODE(frame))
512 		return;
513 
514 	stack_top = td->td_kstack + td->td_kstack_pages * PAGE_SIZE;
515 	current = (vm_offset_t)(uintptr_t)&stack_top;
516 
517 	/*
518 	 * Try to detect if interrupt is using kernel thread stack.
519 	 * Hardware could use a dedicated stack for interrupt handling.
520 	 */
521 	if (stack_top <= current || current < td->td_kstack)
522 		return;
523 
524 	used = stack_top - current;
525 	for (;;) {
526 		prev_used = max_kstack_used;
527 		if (prev_used >= used)
528 			break;
529 		if (atomic_cmpset_int(&max_kstack_used, prev_used, used))
530 			break;
531 	}
532 }
533 #endif /* KSTACK_USAGE_PROF */
534 
535 /*
536  * Implement fork's actions on an address space.
537  * Here we arrange for the address space to be copied or referenced,
538  * allocate a user struct (pcb and kernel stack), then call the
539  * machine-dependent layer to fill those in and make the new process
540  * ready to run.  The new process is set up so that it returns directly
541  * to user mode to avoid stack copying and relocation problems.
542  */
543 int
544 vm_forkproc(struct thread *td, struct proc *p2, struct thread *td2,
545     struct vmspace *vm2, int flags)
546 {
547 	struct proc *p1 = td->td_proc;
548 	struct domainset *dset;
549 	int error;
550 
551 	if ((flags & RFPROC) == 0) {
552 		/*
553 		 * Divorce the memory, if it is shared, essentially
554 		 * this changes shared memory amongst threads, into
555 		 * COW locally.
556 		 */
557 		if ((flags & RFMEM) == 0) {
558 			error = vmspace_unshare(p1);
559 			if (error)
560 				return (error);
561 		}
562 		cpu_fork(td, p2, td2, flags);
563 		return (0);
564 	}
565 
566 	if (flags & RFMEM) {
567 		p2->p_vmspace = p1->p_vmspace;
568 		refcount_acquire(&p1->p_vmspace->vm_refcnt);
569 	}
570 	dset = td2->td_domain.dr_policy;
571 	while (vm_page_count_severe_set(&dset->ds_mask)) {
572 		vm_wait_doms(&dset->ds_mask, 0);
573 	}
574 
575 	if ((flags & RFMEM) == 0) {
576 		p2->p_vmspace = vm2;
577 		if (p1->p_vmspace->vm_shm)
578 			shmfork(p1, p2);
579 	}
580 
581 	/*
582 	 * cpu_fork will copy and update the pcb, set up the kernel stack,
583 	 * and make the child ready to run.
584 	 */
585 	cpu_fork(td, p2, td2, flags);
586 	return (0);
587 }
588 
589 /*
590  * Called after process has been wait(2)'ed upon and is being reaped.
591  * The idea is to reclaim resources that we could not reclaim while
592  * the process was still executing.
593  */
594 void
595 vm_waitproc(p)
596 	struct proc *p;
597 {
598 
599 	vmspace_exitfree(p);		/* and clean-out the vmspace */
600 }
601 
602 void
603 kick_proc0(void)
604 {
605 
606 	wakeup(&proc0);
607 }
608