xref: /freebsd/sys/vm/vm_glue.c (revision 74ca7bf1d4c7173d5575ba168bc4b5f6d181ff5a)
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/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(void *addr, int len, int rw)
117 {
118 	boolean_t rv;
119 	vm_offset_t saddr, eaddr;
120 	vm_prot_t prot;
121 
122 	KASSERT((rw & ~VM_PROT_ALL) == 0,
123 	    ("illegal ``rw'' argument to kernacc (%x)\n", rw));
124 
125 	if ((vm_offset_t)addr + len > kernel_map->max_offset ||
126 	    (vm_offset_t)addr + len < (vm_offset_t)addr)
127 		return (FALSE);
128 
129 	prot = rw;
130 	saddr = trunc_page((vm_offset_t)addr);
131 	eaddr = round_page((vm_offset_t)addr + len);
132 	vm_map_lock_read(kernel_map);
133 	rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot);
134 	vm_map_unlock_read(kernel_map);
135 	return (rv == TRUE);
136 }
137 
138 /*
139  * MPSAFE
140  *
141  * WARNING!  This code calls vm_map_check_protection() which only checks
142  * the associated vm_map_entry range.  It does not determine whether the
143  * contents of the memory is actually readable or writable.  vmapbuf(),
144  * vm_fault_quick(), or copyin()/copout()/su*()/fu*() functions should be
145  * used in conjunction with this call.
146  */
147 int
148 useracc(void *addr, int len, int rw)
149 {
150 	boolean_t rv;
151 	vm_prot_t prot;
152 	vm_map_t map;
153 
154 	KASSERT((rw & ~VM_PROT_ALL) == 0,
155 	    ("illegal ``rw'' argument to useracc (%x)\n", rw));
156 	prot = rw;
157 	map = &curproc->p_vmspace->vm_map;
158 	if ((vm_offset_t)addr + len > vm_map_max(map) ||
159 	    (vm_offset_t)addr + len < (vm_offset_t)addr) {
160 		return (FALSE);
161 	}
162 	vm_map_lock_read(map);
163 	rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr),
164 	    round_page((vm_offset_t)addr + len), prot);
165 	vm_map_unlock_read(map);
166 	return (rv == TRUE);
167 }
168 
169 int
170 vslock(void *addr, size_t len)
171 {
172 	vm_offset_t end, last, start;
173 	vm_size_t npages;
174 	int error;
175 
176 	last = (vm_offset_t)addr + len;
177 	start = trunc_page((vm_offset_t)addr);
178 	end = round_page(last);
179 	if (last < (vm_offset_t)addr || end < (vm_offset_t)addr)
180 		return (EINVAL);
181 	npages = atop(end - start);
182 	if (npages > vm_page_max_wired)
183 		return (ENOMEM);
184 #if 0
185 	/*
186 	 * XXX - not yet
187 	 *
188 	 * The limit for transient usage of wired pages should be
189 	 * larger than for "permanent" wired pages (mlock()).
190 	 *
191 	 * Also, the sysctl code, which is the only present user
192 	 * of vslock(), does a hard loop on EAGAIN.
193 	 */
194 	if (npages + vm_wire_count() > vm_page_max_wired)
195 		return (EAGAIN);
196 #endif
197 	error = vm_map_wire(&curproc->p_vmspace->vm_map, start, end,
198 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
199 	if (error == KERN_SUCCESS) {
200 		curthread->td_vslock_sz += len;
201 		return (0);
202 	}
203 
204 	/*
205 	 * Return EFAULT on error to match copy{in,out}() behaviour
206 	 * rather than returning ENOMEM like mlock() would.
207 	 */
208 	return (EFAULT);
209 }
210 
211 void
212 vsunlock(void *addr, size_t len)
213 {
214 
215 	/* Rely on the parameter sanity checks performed by vslock(). */
216 	MPASS(curthread->td_vslock_sz >= len);
217 	curthread->td_vslock_sz -= len;
218 	(void)vm_map_unwire(&curproc->p_vmspace->vm_map,
219 	    trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len),
220 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
221 }
222 
223 /*
224  * Pin the page contained within the given object at the given offset.  If the
225  * page is not resident, allocate and load it using the given object's pager.
226  * Return the pinned page if successful; otherwise, return NULL.
227  */
228 static vm_page_t
229 vm_imgact_hold_page(vm_object_t object, vm_ooffset_t offset)
230 {
231 	vm_page_t m;
232 	vm_pindex_t pindex;
233 	int rv;
234 
235 	VM_OBJECT_WLOCK(object);
236 	pindex = OFF_TO_IDX(offset);
237 	m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
238 	if (m->valid != VM_PAGE_BITS_ALL) {
239 		vm_page_xbusy(m);
240 		rv = vm_pager_get_pages(object, &m, 1, NULL, NULL);
241 		if (rv != VM_PAGER_OK) {
242 			vm_page_lock(m);
243 			vm_page_free(m);
244 			vm_page_unlock(m);
245 			m = NULL;
246 			goto out;
247 		}
248 		vm_page_xunbusy(m);
249 	}
250 	vm_page_lock(m);
251 	vm_page_hold(m);
252 	vm_page_activate(m);
253 	vm_page_unlock(m);
254 out:
255 	VM_OBJECT_WUNLOCK(object);
256 	return (m);
257 }
258 
259 /*
260  * Return a CPU private mapping to the page at the given offset within the
261  * given object.  The page is pinned before it is mapped.
262  */
263 struct sf_buf *
264 vm_imgact_map_page(vm_object_t object, vm_ooffset_t offset)
265 {
266 	vm_page_t m;
267 
268 	m = vm_imgact_hold_page(object, offset);
269 	if (m == NULL)
270 		return (NULL);
271 	sched_pin();
272 	return (sf_buf_alloc(m, SFB_CPUPRIVATE));
273 }
274 
275 /*
276  * Destroy the given CPU private mapping and unpin the page that it mapped.
277  */
278 void
279 vm_imgact_unmap_page(struct sf_buf *sf)
280 {
281 	vm_page_t m;
282 
283 	m = sf_buf_page(sf);
284 	sf_buf_free(sf);
285 	sched_unpin();
286 	vm_page_lock(m);
287 	vm_page_unhold(m);
288 	vm_page_unlock(m);
289 }
290 
291 void
292 vm_sync_icache(vm_map_t map, vm_offset_t va, vm_offset_t sz)
293 {
294 
295 	pmap_sync_icache(map->pmap, va, sz);
296 }
297 
298 struct kstack_cache_entry *kstack_cache;
299 static int kstack_cache_size = 128;
300 static int kstacks;
301 static struct mtx kstack_cache_mtx;
302 MTX_SYSINIT(kstack_cache, &kstack_cache_mtx, "kstkch", MTX_DEF);
303 
304 SYSCTL_INT(_vm, OID_AUTO, kstack_cache_size, CTLFLAG_RW, &kstack_cache_size, 0,
305     "");
306 SYSCTL_INT(_vm, OID_AUTO, kstacks, CTLFLAG_RD, &kstacks, 0,
307     "");
308 
309 /*
310  * Create the kernel stack (including pcb for i386) for a new thread.
311  * This routine directly affects the fork perf for a process and
312  * create performance for a thread.
313  */
314 int
315 vm_thread_new(struct thread *td, int pages)
316 {
317 	vm_object_t ksobj;
318 	vm_offset_t ks;
319 	vm_page_t ma[KSTACK_MAX_PAGES];
320 	struct kstack_cache_entry *ks_ce;
321 	int i;
322 
323 	/* Bounds check */
324 	if (pages <= 1)
325 		pages = kstack_pages;
326 	else if (pages > KSTACK_MAX_PAGES)
327 		pages = KSTACK_MAX_PAGES;
328 
329 	if (pages == kstack_pages) {
330 		mtx_lock(&kstack_cache_mtx);
331 		if (kstack_cache != NULL) {
332 			ks_ce = kstack_cache;
333 			kstack_cache = ks_ce->next_ks_entry;
334 			mtx_unlock(&kstack_cache_mtx);
335 
336 			td->td_kstack_obj = ks_ce->ksobj;
337 			td->td_kstack = (vm_offset_t)ks_ce;
338 			td->td_kstack_pages = kstack_pages;
339 			return (1);
340 		}
341 		mtx_unlock(&kstack_cache_mtx);
342 	}
343 
344 	/*
345 	 * Allocate an object for the kstack.
346 	 */
347 	ksobj = vm_object_allocate(OBJT_DEFAULT, pages);
348 
349 	/*
350 	 * Get a kernel virtual address for this thread's kstack.
351 	 */
352 #if defined(__mips__)
353 	/*
354 	 * We need to align the kstack's mapped address to fit within
355 	 * a single TLB entry.
356 	 */
357 	if (vmem_xalloc(kernel_arena, (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE,
358 	    PAGE_SIZE * 2, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX,
359 	    M_BESTFIT | M_NOWAIT, &ks)) {
360 		ks = 0;
361 	}
362 #else
363 	ks = kva_alloc((pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
364 #endif
365 	if (ks == 0) {
366 		printf("vm_thread_new: kstack allocation failed\n");
367 		vm_object_deallocate(ksobj);
368 		return (0);
369 	}
370 
371 	atomic_add_int(&kstacks, 1);
372 	if (KSTACK_GUARD_PAGES != 0) {
373 		pmap_qremove(ks, KSTACK_GUARD_PAGES);
374 		ks += KSTACK_GUARD_PAGES * PAGE_SIZE;
375 	}
376 	td->td_kstack_obj = ksobj;
377 	td->td_kstack = ks;
378 	/*
379 	 * Knowing the number of pages allocated is useful when you
380 	 * want to deallocate them.
381 	 */
382 	td->td_kstack_pages = pages;
383 	/*
384 	 * For the length of the stack, link in a real page of ram for each
385 	 * page of stack.
386 	 */
387 	VM_OBJECT_WLOCK(ksobj);
388 	(void)vm_page_grab_pages(ksobj, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY |
389 	    VM_ALLOC_WIRED, ma, pages);
390 	for (i = 0; i < pages; i++)
391 		ma[i]->valid = VM_PAGE_BITS_ALL;
392 	VM_OBJECT_WUNLOCK(ksobj);
393 	pmap_qenter(ks, ma, pages);
394 	return (1);
395 }
396 
397 static void
398 vm_thread_stack_dispose(vm_object_t ksobj, vm_offset_t ks, int pages)
399 {
400 	vm_page_t m;
401 	int i;
402 
403 	atomic_add_int(&kstacks, -1);
404 	pmap_qremove(ks, pages);
405 	VM_OBJECT_WLOCK(ksobj);
406 	for (i = 0; i < pages; i++) {
407 		m = vm_page_lookup(ksobj, i);
408 		if (m == NULL)
409 			panic("vm_thread_dispose: kstack already missing?");
410 		vm_page_lock(m);
411 		vm_page_unwire(m, PQ_NONE);
412 		vm_page_free(m);
413 		vm_page_unlock(m);
414 	}
415 	VM_OBJECT_WUNLOCK(ksobj);
416 	vm_object_deallocate(ksobj);
417 	kva_free(ks - (KSTACK_GUARD_PAGES * PAGE_SIZE),
418 	    (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
419 }
420 
421 /*
422  * Dispose of a thread's kernel stack.
423  */
424 void
425 vm_thread_dispose(struct thread *td)
426 {
427 	vm_object_t ksobj;
428 	vm_offset_t ks;
429 	struct kstack_cache_entry *ks_ce;
430 	int pages;
431 
432 	pages = td->td_kstack_pages;
433 	ksobj = td->td_kstack_obj;
434 	ks = td->td_kstack;
435 	td->td_kstack = 0;
436 	td->td_kstack_pages = 0;
437 	if (pages == kstack_pages && kstacks <= kstack_cache_size) {
438 		ks_ce = (struct kstack_cache_entry *)ks;
439 		ks_ce->ksobj = ksobj;
440 		mtx_lock(&kstack_cache_mtx);
441 		ks_ce->next_ks_entry = kstack_cache;
442 		kstack_cache = ks_ce;
443 		mtx_unlock(&kstack_cache_mtx);
444 		return;
445 	}
446 	vm_thread_stack_dispose(ksobj, ks, pages);
447 }
448 
449 static void
450 vm_thread_stack_lowmem(void *nulll)
451 {
452 	struct kstack_cache_entry *ks_ce, *ks_ce1;
453 
454 	mtx_lock(&kstack_cache_mtx);
455 	ks_ce = kstack_cache;
456 	kstack_cache = NULL;
457 	mtx_unlock(&kstack_cache_mtx);
458 
459 	while (ks_ce != NULL) {
460 		ks_ce1 = ks_ce;
461 		ks_ce = ks_ce->next_ks_entry;
462 
463 		vm_thread_stack_dispose(ks_ce1->ksobj, (vm_offset_t)ks_ce1,
464 		    kstack_pages);
465 	}
466 }
467 
468 static void
469 kstack_cache_init(void *nulll)
470 {
471 
472 	EVENTHANDLER_REGISTER(vm_lowmem, vm_thread_stack_lowmem, NULL,
473 	    EVENTHANDLER_PRI_ANY);
474 }
475 
476 SYSINIT(vm_kstacks, SI_SUB_KTHREAD_INIT, SI_ORDER_ANY, kstack_cache_init, NULL);
477 
478 #ifdef KSTACK_USAGE_PROF
479 /*
480  * Track maximum stack used by a thread in kernel.
481  */
482 static int max_kstack_used;
483 
484 SYSCTL_INT(_debug, OID_AUTO, max_kstack_used, CTLFLAG_RD,
485     &max_kstack_used, 0,
486     "Maxiumum stack depth used by a thread in kernel");
487 
488 void
489 intr_prof_stack_use(struct thread *td, struct trapframe *frame)
490 {
491 	vm_offset_t stack_top;
492 	vm_offset_t current;
493 	int used, prev_used;
494 
495 	/*
496 	 * Testing for interrupted kernel mode isn't strictly
497 	 * needed. It optimizes the execution, since interrupts from
498 	 * usermode will have only the trap frame on the stack.
499 	 */
500 	if (TRAPF_USERMODE(frame))
501 		return;
502 
503 	stack_top = td->td_kstack + td->td_kstack_pages * PAGE_SIZE;
504 	current = (vm_offset_t)(uintptr_t)&stack_top;
505 
506 	/*
507 	 * Try to detect if interrupt is using kernel thread stack.
508 	 * Hardware could use a dedicated stack for interrupt handling.
509 	 */
510 	if (stack_top <= current || current < td->td_kstack)
511 		return;
512 
513 	used = stack_top - current;
514 	for (;;) {
515 		prev_used = max_kstack_used;
516 		if (prev_used >= used)
517 			break;
518 		if (atomic_cmpset_int(&max_kstack_used, prev_used, used))
519 			break;
520 	}
521 }
522 #endif /* KSTACK_USAGE_PROF */
523 
524 /*
525  * Implement fork's actions on an address space.
526  * Here we arrange for the address space to be copied or referenced,
527  * allocate a user struct (pcb and kernel stack), then call the
528  * machine-dependent layer to fill those in and make the new process
529  * ready to run.  The new process is set up so that it returns directly
530  * to user mode to avoid stack copying and relocation problems.
531  */
532 int
533 vm_forkproc(struct thread *td, struct proc *p2, struct thread *td2,
534     struct vmspace *vm2, int flags)
535 {
536 	struct proc *p1 = td->td_proc;
537 	int error;
538 
539 	if ((flags & RFPROC) == 0) {
540 		/*
541 		 * Divorce the memory, if it is shared, essentially
542 		 * this changes shared memory amongst threads, into
543 		 * COW locally.
544 		 */
545 		if ((flags & RFMEM) == 0) {
546 			if (p1->p_vmspace->vm_refcnt > 1) {
547 				error = vmspace_unshare(p1);
548 				if (error)
549 					return (error);
550 			}
551 		}
552 		cpu_fork(td, p2, td2, flags);
553 		return (0);
554 	}
555 
556 	if (flags & RFMEM) {
557 		p2->p_vmspace = p1->p_vmspace;
558 		atomic_add_int(&p1->p_vmspace->vm_refcnt, 1);
559 	}
560 
561 	while (vm_page_count_severe()) {
562 		vm_wait_severe();
563 	}
564 
565 	if ((flags & RFMEM) == 0) {
566 		p2->p_vmspace = vm2;
567 		if (p1->p_vmspace->vm_shm)
568 			shmfork(p1, p2);
569 	}
570 
571 	/*
572 	 * cpu_fork will copy and update the pcb, set up the kernel stack,
573 	 * and make the child ready to run.
574 	 */
575 	cpu_fork(td, p2, td2, flags);
576 	return (0);
577 }
578 
579 /*
580  * Called after process has been wait(2)'ed upon and is being reaped.
581  * The idea is to reclaim resources that we could not reclaim while
582  * the process was still executing.
583  */
584 void
585 vm_waitproc(p)
586 	struct proc *p;
587 {
588 
589 	vmspace_exitfree(p);		/* and clean-out the vmspace */
590 }
591 
592 void
593 kick_proc0(void)
594 {
595 
596 	wakeup(&proc0);
597 }
598