xref: /freebsd/sys/vm/vm_glue.c (revision 718cf2ccb9956613756ab15d7a0e28f2c8e91cab)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
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/limits.h>
72 #include <sys/lock.h>
73 #include <sys/malloc.h>
74 #include <sys/mutex.h>
75 #include <sys/proc.h>
76 #include <sys/racct.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/vmmeter.h>
83 #include <sys/vmem.h>
84 #include <sys/sx.h>
85 #include <sys/sysctl.h>
86 #include <sys/_kstack_cache.h>
87 #include <sys/eventhandler.h>
88 #include <sys/kernel.h>
89 #include <sys/ktr.h>
90 #include <sys/unistd.h>
91 
92 #include <vm/vm.h>
93 #include <vm/vm_param.h>
94 #include <vm/pmap.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_page.h>
97 #include <vm/vm_pageout.h>
98 #include <vm/vm_object.h>
99 #include <vm/vm_kern.h>
100 #include <vm/vm_extern.h>
101 #include <vm/vm_pager.h>
102 #include <vm/swap_pager.h>
103 
104 #include <machine/cpu.h>
105 
106 /*
107  * MPSAFE
108  *
109  * WARNING!  This code calls vm_map_check_protection() which only checks
110  * the associated vm_map_entry range.  It does not determine whether the
111  * contents of the memory is actually readable or writable.  In most cases
112  * just checking the vm_map_entry is sufficient within the kernel's address
113  * space.
114  */
115 int
116 kernacc(addr, len, rw)
117 	void *addr;
118 	int len, rw;
119 {
120 	boolean_t rv;
121 	vm_offset_t saddr, eaddr;
122 	vm_prot_t prot;
123 
124 	KASSERT((rw & ~VM_PROT_ALL) == 0,
125 	    ("illegal ``rw'' argument to kernacc (%x)\n", rw));
126 
127 	if ((vm_offset_t)addr + len > kernel_map->max_offset ||
128 	    (vm_offset_t)addr + len < (vm_offset_t)addr)
129 		return (FALSE);
130 
131 	prot = rw;
132 	saddr = trunc_page((vm_offset_t)addr);
133 	eaddr = round_page((vm_offset_t)addr + len);
134 	vm_map_lock_read(kernel_map);
135 	rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot);
136 	vm_map_unlock_read(kernel_map);
137 	return (rv == TRUE);
138 }
139 
140 /*
141  * MPSAFE
142  *
143  * WARNING!  This code calls vm_map_check_protection() which only checks
144  * the associated vm_map_entry range.  It does not determine whether the
145  * contents of the memory is actually readable or writable.  vmapbuf(),
146  * vm_fault_quick(), or copyin()/copout()/su*()/fu*() functions should be
147  * used in conjunction with this call.
148  */
149 int
150 useracc(addr, len, rw)
151 	void *addr;
152 	int len, rw;
153 {
154 	boolean_t rv;
155 	vm_prot_t prot;
156 	vm_map_t map;
157 
158 	KASSERT((rw & ~VM_PROT_ALL) == 0,
159 	    ("illegal ``rw'' argument to useracc (%x)\n", rw));
160 	prot = rw;
161 	map = &curproc->p_vmspace->vm_map;
162 	if ((vm_offset_t)addr + len > vm_map_max(map) ||
163 	    (vm_offset_t)addr + len < (vm_offset_t)addr) {
164 		return (FALSE);
165 	}
166 	vm_map_lock_read(map);
167 	rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr),
168 	    round_page((vm_offset_t)addr + len), prot);
169 	vm_map_unlock_read(map);
170 	return (rv == TRUE);
171 }
172 
173 int
174 vslock(void *addr, size_t len)
175 {
176 	vm_offset_t end, last, start;
177 	vm_size_t npages;
178 	int error;
179 
180 	last = (vm_offset_t)addr + len;
181 	start = trunc_page((vm_offset_t)addr);
182 	end = round_page(last);
183 	if (last < (vm_offset_t)addr || end < (vm_offset_t)addr)
184 		return (EINVAL);
185 	npages = atop(end - start);
186 	if (npages > vm_page_max_wired)
187 		return (ENOMEM);
188 #if 0
189 	/*
190 	 * XXX - not yet
191 	 *
192 	 * The limit for transient usage of wired pages should be
193 	 * larger than for "permanent" wired pages (mlock()).
194 	 *
195 	 * Also, the sysctl code, which is the only present user
196 	 * of vslock(), does a hard loop on EAGAIN.
197 	 */
198 	if (npages + vm_cnt.v_wire_count > vm_page_max_wired)
199 		return (EAGAIN);
200 #endif
201 	error = vm_map_wire(&curproc->p_vmspace->vm_map, start, end,
202 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
203 	/*
204 	 * Return EFAULT on error to match copy{in,out}() behaviour
205 	 * rather than returning ENOMEM like mlock() would.
206 	 */
207 	return (error == KERN_SUCCESS ? 0 : EFAULT);
208 }
209 
210 void
211 vsunlock(void *addr, size_t len)
212 {
213 
214 	/* Rely on the parameter sanity checks performed by vslock(). */
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 	int rv;
231 
232 	VM_OBJECT_WLOCK(object);
233 	pindex = OFF_TO_IDX(offset);
234 	m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
235 	if (m->valid != VM_PAGE_BITS_ALL) {
236 		vm_page_xbusy(m);
237 		rv = vm_pager_get_pages(object, &m, 1, NULL, NULL);
238 		if (rv != VM_PAGER_OK) {
239 			vm_page_lock(m);
240 			vm_page_free(m);
241 			vm_page_unlock(m);
242 			m = NULL;
243 			goto out;
244 		}
245 		vm_page_xunbusy(m);
246 	}
247 	vm_page_lock(m);
248 	vm_page_hold(m);
249 	vm_page_activate(m);
250 	vm_page_unlock(m);
251 out:
252 	VM_OBJECT_WUNLOCK(object);
253 	return (m);
254 }
255 
256 /*
257  * Return a CPU private mapping to the page at the given offset within the
258  * given object.  The page is pinned before it is mapped.
259  */
260 struct sf_buf *
261 vm_imgact_map_page(vm_object_t object, vm_ooffset_t offset)
262 {
263 	vm_page_t m;
264 
265 	m = vm_imgact_hold_page(object, offset);
266 	if (m == NULL)
267 		return (NULL);
268 	sched_pin();
269 	return (sf_buf_alloc(m, SFB_CPUPRIVATE));
270 }
271 
272 /*
273  * Destroy the given CPU private mapping and unpin the page that it mapped.
274  */
275 void
276 vm_imgact_unmap_page(struct sf_buf *sf)
277 {
278 	vm_page_t m;
279 
280 	m = sf_buf_page(sf);
281 	sf_buf_free(sf);
282 	sched_unpin();
283 	vm_page_lock(m);
284 	vm_page_unhold(m);
285 	vm_page_unlock(m);
286 }
287 
288 void
289 vm_sync_icache(vm_map_t map, vm_offset_t va, vm_offset_t sz)
290 {
291 
292 	pmap_sync_icache(map->pmap, va, sz);
293 }
294 
295 struct kstack_cache_entry *kstack_cache;
296 static int kstack_cache_size = 128;
297 static int kstacks;
298 static struct mtx kstack_cache_mtx;
299 MTX_SYSINIT(kstack_cache, &kstack_cache_mtx, "kstkch", MTX_DEF);
300 
301 SYSCTL_INT(_vm, OID_AUTO, kstack_cache_size, CTLFLAG_RW, &kstack_cache_size, 0,
302     "");
303 SYSCTL_INT(_vm, OID_AUTO, kstacks, CTLFLAG_RD, &kstacks, 0,
304     "");
305 
306 /*
307  * Create the kernel stack (including pcb for i386) for a new thread.
308  * This routine directly affects the fork perf for a process and
309  * create performance for a thread.
310  */
311 int
312 vm_thread_new(struct thread *td, int pages)
313 {
314 	vm_object_t ksobj;
315 	vm_offset_t ks;
316 	vm_page_t ma[KSTACK_MAX_PAGES];
317 	struct kstack_cache_entry *ks_ce;
318 	int i;
319 
320 	/* Bounds check */
321 	if (pages <= 1)
322 		pages = kstack_pages;
323 	else if (pages > KSTACK_MAX_PAGES)
324 		pages = KSTACK_MAX_PAGES;
325 
326 	if (pages == kstack_pages) {
327 		mtx_lock(&kstack_cache_mtx);
328 		if (kstack_cache != NULL) {
329 			ks_ce = kstack_cache;
330 			kstack_cache = ks_ce->next_ks_entry;
331 			mtx_unlock(&kstack_cache_mtx);
332 
333 			td->td_kstack_obj = ks_ce->ksobj;
334 			td->td_kstack = (vm_offset_t)ks_ce;
335 			td->td_kstack_pages = kstack_pages;
336 			return (1);
337 		}
338 		mtx_unlock(&kstack_cache_mtx);
339 	}
340 
341 	/*
342 	 * Allocate an object for the kstack.
343 	 */
344 	ksobj = vm_object_allocate(OBJT_DEFAULT, pages);
345 
346 	/*
347 	 * Get a kernel virtual address for this thread's kstack.
348 	 */
349 #if defined(__mips__)
350 	/*
351 	 * We need to align the kstack's mapped address to fit within
352 	 * a single TLB entry.
353 	 */
354 	if (vmem_xalloc(kernel_arena, (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE,
355 	    PAGE_SIZE * 2, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX,
356 	    M_BESTFIT | M_NOWAIT, &ks)) {
357 		ks = 0;
358 	}
359 #else
360 	ks = kva_alloc((pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
361 #endif
362 	if (ks == 0) {
363 		printf("vm_thread_new: kstack allocation failed\n");
364 		vm_object_deallocate(ksobj);
365 		return (0);
366 	}
367 
368 	atomic_add_int(&kstacks, 1);
369 	if (KSTACK_GUARD_PAGES != 0) {
370 		pmap_qremove(ks, KSTACK_GUARD_PAGES);
371 		ks += KSTACK_GUARD_PAGES * PAGE_SIZE;
372 	}
373 	td->td_kstack_obj = ksobj;
374 	td->td_kstack = ks;
375 	/*
376 	 * Knowing the number of pages allocated is useful when you
377 	 * want to deallocate them.
378 	 */
379 	td->td_kstack_pages = pages;
380 	/*
381 	 * For the length of the stack, link in a real page of ram for each
382 	 * page of stack.
383 	 */
384 	VM_OBJECT_WLOCK(ksobj);
385 	(void)vm_page_grab_pages(ksobj, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY |
386 	    VM_ALLOC_WIRED, ma, pages);
387 	for (i = 0; i < pages; i++)
388 		ma[i]->valid = VM_PAGE_BITS_ALL;
389 	VM_OBJECT_WUNLOCK(ksobj);
390 	pmap_qenter(ks, ma, pages);
391 	return (1);
392 }
393 
394 static void
395 vm_thread_stack_dispose(vm_object_t ksobj, vm_offset_t ks, int pages)
396 {
397 	vm_page_t m;
398 	int i;
399 
400 	atomic_add_int(&kstacks, -1);
401 	pmap_qremove(ks, pages);
402 	VM_OBJECT_WLOCK(ksobj);
403 	for (i = 0; i < pages; i++) {
404 		m = vm_page_lookup(ksobj, i);
405 		if (m == NULL)
406 			panic("vm_thread_dispose: kstack already missing?");
407 		vm_page_lock(m);
408 		vm_page_unwire(m, PQ_NONE);
409 		vm_page_free(m);
410 		vm_page_unlock(m);
411 	}
412 	VM_OBJECT_WUNLOCK(ksobj);
413 	vm_object_deallocate(ksobj);
414 	kva_free(ks - (KSTACK_GUARD_PAGES * PAGE_SIZE),
415 	    (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
416 }
417 
418 /*
419  * Dispose of a thread's kernel stack.
420  */
421 void
422 vm_thread_dispose(struct thread *td)
423 {
424 	vm_object_t ksobj;
425 	vm_offset_t ks;
426 	struct kstack_cache_entry *ks_ce;
427 	int pages;
428 
429 	pages = td->td_kstack_pages;
430 	ksobj = td->td_kstack_obj;
431 	ks = td->td_kstack;
432 	td->td_kstack = 0;
433 	td->td_kstack_pages = 0;
434 	if (pages == kstack_pages && kstacks <= kstack_cache_size) {
435 		ks_ce = (struct kstack_cache_entry *)ks;
436 		ks_ce->ksobj = ksobj;
437 		mtx_lock(&kstack_cache_mtx);
438 		ks_ce->next_ks_entry = kstack_cache;
439 		kstack_cache = ks_ce;
440 		mtx_unlock(&kstack_cache_mtx);
441 		return;
442 	}
443 	vm_thread_stack_dispose(ksobj, ks, pages);
444 }
445 
446 static void
447 vm_thread_stack_lowmem(void *nulll)
448 {
449 	struct kstack_cache_entry *ks_ce, *ks_ce1;
450 
451 	mtx_lock(&kstack_cache_mtx);
452 	ks_ce = kstack_cache;
453 	kstack_cache = NULL;
454 	mtx_unlock(&kstack_cache_mtx);
455 
456 	while (ks_ce != NULL) {
457 		ks_ce1 = ks_ce;
458 		ks_ce = ks_ce->next_ks_entry;
459 
460 		vm_thread_stack_dispose(ks_ce1->ksobj, (vm_offset_t)ks_ce1,
461 		    kstack_pages);
462 	}
463 }
464 
465 static void
466 kstack_cache_init(void *nulll)
467 {
468 
469 	EVENTHANDLER_REGISTER(vm_lowmem, vm_thread_stack_lowmem, NULL,
470 	    EVENTHANDLER_PRI_ANY);
471 }
472 
473 SYSINIT(vm_kstacks, SI_SUB_KTHREAD_INIT, SI_ORDER_ANY, kstack_cache_init, NULL);
474 
475 #ifdef KSTACK_USAGE_PROF
476 /*
477  * Track maximum stack used by a thread in kernel.
478  */
479 static int max_kstack_used;
480 
481 SYSCTL_INT(_debug, OID_AUTO, max_kstack_used, CTLFLAG_RD,
482     &max_kstack_used, 0,
483     "Maxiumum stack depth used by a thread in kernel");
484 
485 void
486 intr_prof_stack_use(struct thread *td, struct trapframe *frame)
487 {
488 	vm_offset_t stack_top;
489 	vm_offset_t current;
490 	int used, prev_used;
491 
492 	/*
493 	 * Testing for interrupted kernel mode isn't strictly
494 	 * needed. It optimizes the execution, since interrupts from
495 	 * usermode will have only the trap frame on the stack.
496 	 */
497 	if (TRAPF_USERMODE(frame))
498 		return;
499 
500 	stack_top = td->td_kstack + td->td_kstack_pages * PAGE_SIZE;
501 	current = (vm_offset_t)(uintptr_t)&stack_top;
502 
503 	/*
504 	 * Try to detect if interrupt is using kernel thread stack.
505 	 * Hardware could use a dedicated stack for interrupt handling.
506 	 */
507 	if (stack_top <= current || current < td->td_kstack)
508 		return;
509 
510 	used = stack_top - current;
511 	for (;;) {
512 		prev_used = max_kstack_used;
513 		if (prev_used >= used)
514 			break;
515 		if (atomic_cmpset_int(&max_kstack_used, prev_used, used))
516 			break;
517 	}
518 }
519 #endif /* KSTACK_USAGE_PROF */
520 
521 /*
522  * Implement fork's actions on an address space.
523  * Here we arrange for the address space to be copied or referenced,
524  * allocate a user struct (pcb and kernel stack), then call the
525  * machine-dependent layer to fill those in and make the new process
526  * ready to run.  The new process is set up so that it returns directly
527  * to user mode to avoid stack copying and relocation problems.
528  */
529 int
530 vm_forkproc(td, p2, td2, vm2, flags)
531 	struct thread *td;
532 	struct proc *p2;
533 	struct thread *td2;
534 	struct vmspace *vm2;
535 	int flags;
536 {
537 	struct proc *p1 = td->td_proc;
538 	int error;
539 
540 	if ((flags & RFPROC) == 0) {
541 		/*
542 		 * Divorce the memory, if it is shared, essentially
543 		 * this changes shared memory amongst threads, into
544 		 * COW locally.
545 		 */
546 		if ((flags & RFMEM) == 0) {
547 			if (p1->p_vmspace->vm_refcnt > 1) {
548 				error = vmspace_unshare(p1);
549 				if (error)
550 					return (error);
551 			}
552 		}
553 		cpu_fork(td, p2, td2, flags);
554 		return (0);
555 	}
556 
557 	if (flags & RFMEM) {
558 		p2->p_vmspace = p1->p_vmspace;
559 		atomic_add_int(&p1->p_vmspace->vm_refcnt, 1);
560 	}
561 
562 	while (vm_page_count_severe()) {
563 		VM_WAIT;
564 	}
565 
566 	if ((flags & RFMEM) == 0) {
567 		p2->p_vmspace = vm2;
568 		if (p1->p_vmspace->vm_shm)
569 			shmfork(p1, p2);
570 	}
571 
572 	/*
573 	 * cpu_fork will copy and update the pcb, set up the kernel stack,
574 	 * and make the child ready to run.
575 	 */
576 	cpu_fork(td, p2, td2, flags);
577 	return (0);
578 }
579 
580 /*
581  * Called after process has been wait(2)'ed upon and is being reaped.
582  * The idea is to reclaim resources that we could not reclaim while
583  * the process was still executing.
584  */
585 void
586 vm_waitproc(p)
587 	struct proc *p;
588 {
589 
590 	vmspace_exitfree(p);		/* and clean-out the vmspace */
591 }
592 
593 void
594 kick_proc0(void)
595 {
596 
597 	wakeup(&proc0);
598 }
599