xref: /dragonfly/sys/vm/vm_contig.c (revision c3762235)
1 /*
2  * Copyright (c) 2003, 2004 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Hiten Pandya <hmp@backplane.com>.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  */
35 /*
36  * Copyright (c) 1991 Regents of the University of California.
37  * All rights reserved.
38  *
39  * This code is derived from software contributed to Berkeley by
40  * The Mach Operating System project at Carnegie-Mellon University.
41  *
42  * Redistribution and use in source and binary forms, with or without
43  * modification, are permitted provided that the following conditions
44  * are met:
45  * 1. Redistributions of source code must retain the above copyright
46  *    notice, this list of conditions and the following disclaimer.
47  * 2. Redistributions in binary form must reproduce the above copyright
48  *    notice, this list of conditions and the following disclaimer in the
49  *    documentation and/or other materials provided with the distribution.
50  * 3. Neither the name of the University nor the names of its contributors
51  *    may be used to endorse or promote products derived from this software
52  *    without specific prior written permission.
53  *
54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64  * SUCH DAMAGE.
65  *
66  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
67  * $DragonFly: src/sys/vm/vm_contig.c,v 1.21 2006/12/28 21:24:02 dillon Exp $
68  */
69 
70 /*
71  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
72  * All rights reserved.
73  *
74  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
75  *
76  * Permission to use, copy, modify and distribute this software and
77  * its documentation is hereby granted, provided that both the copyright
78  * notice and this permission notice appear in all copies of the
79  * software, derivative works or modified versions, and any portions
80  * thereof, and that both notices appear in supporting documentation.
81  *
82  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
83  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
84  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
85  *
86  * Carnegie Mellon requests users of this software to return to
87  *
88  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
89  *  School of Computer Science
90  *  Carnegie Mellon University
91  *  Pittsburgh PA 15213-3890
92  *
93  * any improvements or extensions that they make and grant Carnegie the
94  * rights to redistribute these changes.
95  */
96 
97 /*
98  * Contiguous memory allocation API.
99  */
100 
101 #include <sys/param.h>
102 #include <sys/systm.h>
103 #include <sys/malloc.h>
104 #include <sys/proc.h>
105 #include <sys/lock.h>
106 #include <sys/vmmeter.h>
107 #include <sys/vnode.h>
108 
109 #include <vm/vm.h>
110 #include <vm/vm_param.h>
111 #include <vm/vm_kern.h>
112 #include <vm/pmap.h>
113 #include <vm/vm_map.h>
114 #include <vm/vm_object.h>
115 #include <vm/vm_page.h>
116 #include <vm/vm_pageout.h>
117 #include <vm/vm_pager.h>
118 #include <vm/vm_extern.h>
119 
120 #include <sys/thread2.h>
121 #include <sys/spinlock2.h>
122 #include <vm/vm_page2.h>
123 
124 static void vm_contig_pg_free(int start, u_long size);
125 
126 /*
127  * vm_contig_pg_clean:
128  *
129  * Do a thorough cleanup of the specified 'queue', which can be either
130  * PQ_ACTIVE or PQ_INACTIVE by doing a walkthrough.  If the page is not
131  * marked dirty, it is shoved into the page cache, provided no one has
132  * currently aqcuired it, otherwise localized action per object type
133  * is taken for cleanup:
134  *
135  * 	In the OBJT_VNODE case, the whole page range is cleaned up
136  * 	using the vm_object_page_clean() routine, by specyfing a
137  * 	start and end of '0'.
138  *
139  * 	Otherwise if the object is of any other type, the generic
140  * 	pageout (daemon) flush routine is invoked.
141  */
142 static void
143 vm_contig_pg_clean(int queue, int count)
144 {
145 	vm_object_t object;
146 	vm_page_t m, m_tmp;
147 	struct vm_page marker;
148 	struct vpgqueues *pq = &vm_page_queues[queue];
149 
150 	/*
151 	 * Setup a local marker
152 	 */
153 	bzero(&marker, sizeof(marker));
154 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
155 	marker.queue = queue;
156 	marker.wire_count = 1;
157 
158 	vm_page_queues_spin_lock(queue);
159 	TAILQ_INSERT_HEAD(&pq->pl, &marker, pageq);
160 	vm_page_queues_spin_unlock(queue);
161 
162 	/*
163 	 * Iterate the queue.  Note that the vm_page spinlock must be
164 	 * acquired before the pageq spinlock so it's easiest to simply
165 	 * not hold it in the loop iteration.
166 	 */
167 	while (count-- > 0 && (m = TAILQ_NEXT(&marker, pageq)) != NULL) {
168 		vm_page_and_queue_spin_lock(m);
169 		if (m != TAILQ_NEXT(&marker, pageq)) {
170 			vm_page_and_queue_spin_unlock(m);
171 			++count;
172 			continue;
173 		}
174 		KKASSERT(m->queue == queue);
175 
176 		TAILQ_REMOVE(&pq->pl, &marker, pageq);
177 		TAILQ_INSERT_AFTER(&pq->pl, m, &marker, pageq);
178 
179 		if (m->flags & PG_MARKER) {
180 			vm_page_and_queue_spin_unlock(m);
181 			continue;
182 		}
183 		if (vm_page_busy_try(m, TRUE)) {
184 			vm_page_and_queue_spin_unlock(m);
185 			continue;
186 		}
187 		vm_page_and_queue_spin_unlock(m);
188 
189 		/*
190 		 * We've successfully busied the page
191 		 */
192 		if (m->queue - m->pc != queue) {
193 			vm_page_wakeup(m);
194 			continue;
195 		}
196 		if (m->wire_count || m->hold_count) {
197 			vm_page_wakeup(m);
198 			continue;
199 		}
200 		if ((object = m->object) == NULL) {
201 			vm_page_wakeup(m);
202 			continue;
203 		}
204 		vm_page_test_dirty(m);
205 		if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
206 			vm_object_hold(object);
207 			KKASSERT(m->object == object);
208 
209 			if (object->type == OBJT_VNODE) {
210 				vm_page_wakeup(m);
211 				vn_lock(object->handle, LK_EXCLUSIVE|LK_RETRY);
212 				vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
213 				vn_unlock(((struct vnode *)object->handle));
214 			} else if (object->type == OBJT_SWAP ||
215 					object->type == OBJT_DEFAULT) {
216 				m_tmp = m;
217 				vm_pageout_flush(&m_tmp, 1, 0);
218 			} else {
219 				vm_page_wakeup(m);
220 			}
221 			vm_object_drop(object);
222 		} else if (m->hold_count == 0) {
223 			vm_page_cache(m);
224 		} else {
225 			vm_page_wakeup(m);
226 		}
227 	}
228 
229 	/*
230 	 * Scrap our local marker
231 	 */
232 	vm_page_queues_spin_lock(queue);
233 	TAILQ_REMOVE(&pq->pl, &marker, pageq);
234 	vm_page_queues_spin_unlock(queue);
235 }
236 
237 /*
238  * vm_contig_pg_alloc:
239  *
240  * Allocate contiguous pages from the VM.  This function does not
241  * map the allocated pages into the kernel map, otherwise it is
242  * impossible to make large allocations (i.e. >2G).
243  *
244  * Malloc()'s data structures have been used for collection of
245  * statistics and for allocations of less than a page.
246  */
247 static int
248 vm_contig_pg_alloc(unsigned long size, vm_paddr_t low, vm_paddr_t high,
249 		   unsigned long alignment, unsigned long boundary, int mflags)
250 {
251 	int i, q, start, pass;
252 	vm_offset_t phys;
253 	vm_page_t pga = vm_page_array;
254 	vm_page_t m;
255 	int pqtype;
256 
257 	size = round_page(size);
258 	if (size == 0)
259 		panic("vm_contig_pg_alloc: size must not be 0");
260 	if ((alignment & (alignment - 1)) != 0)
261 		panic("vm_contig_pg_alloc: alignment must be a power of 2");
262 	if ((boundary & (boundary - 1)) != 0)
263 		panic("vm_contig_pg_alloc: boundary must be a power of 2");
264 
265 	/*
266 	 * See if we can get the pages from the contiguous page reserve
267 	 * alist.  The returned pages will be allocated and wired but not
268 	 * busied.
269 	 */
270 	m = vm_page_alloc_contig(
271 		low, high, alignment, boundary, size, VM_MEMATTR_DEFAULT);
272 	if (m)
273 		return (m - &pga[0]);
274 
275 	/*
276 	 * Three passes (0, 1, 2).  Each pass scans the VM page list for
277 	 * free or cached pages.  After each pass if the entire scan failed
278 	 * we attempt to flush inactive pages and reset the start index back
279 	 * to 0.  For passes 1 and 2 we also attempt to flush active pages.
280 	 */
281 	start = 0;
282 	for (pass = 0; pass < 3; pass++) {
283 		/*
284 		 * Find first page in array that is free, within range,
285 		 * aligned, and such that the boundary won't be crossed.
286 		 */
287 again:
288 		for (i = start; i < vmstats.v_page_count; i++) {
289 			m = &pga[i];
290 			phys = VM_PAGE_TO_PHYS(m);
291 			pqtype = m->queue - m->pc;
292 			if (((pqtype == PQ_FREE) || (pqtype == PQ_CACHE)) &&
293 			    (phys >= low) && (phys < high) &&
294 			    ((phys & (alignment - 1)) == 0) &&
295 			    (((phys ^ (phys + size - 1)) & ~(boundary - 1)) == 0) &&
296 			    m->busy == 0 && m->wire_count == 0 &&
297 			    m->hold_count == 0 &&
298 			    (m->flags & (PG_BUSY | PG_NEED_COMMIT)) == 0)
299 			{
300 				break;
301 			}
302 		}
303 
304 		/*
305 		 * If we cannot find the page in the given range, or we have
306 		 * crossed the boundary, call the vm_contig_pg_clean() function
307 		 * for flushing out the queues, and returning it back to
308 		 * normal state.
309 		 */
310 		if ((i == vmstats.v_page_count) ||
311 		    ((VM_PAGE_TO_PHYS(&pga[i]) + size) > high)) {
312 
313 			/*
314 			 * Best effort flush of all inactive pages.
315 			 * This is quite quick, for now stall all
316 			 * callers, even if they've specified M_NOWAIT.
317 			 */
318 			for (q = 0; q < PQ_L2_SIZE; ++q) {
319 				vm_contig_pg_clean(PQ_INACTIVE + q,
320 						   vmstats.v_inactive_count);
321 				lwkt_yield();
322 			}
323 
324 			/*
325 			 * Best effort flush of active pages.
326 			 *
327 			 * This is very, very slow.
328 			 * Only do this if the caller has agreed to M_WAITOK.
329 			 *
330 			 * If enough pages are flushed, we may succeed on
331 			 * next (final) pass, if not the caller, contigmalloc(),
332 			 * will fail in the index < 0 case.
333 			 */
334 			if (pass > 0 && (mflags & M_WAITOK)) {
335 				for (q = 0; q < PQ_L2_SIZE; ++q) {
336 					vm_contig_pg_clean(PQ_ACTIVE + q,
337 						       vmstats.v_active_count);
338 				}
339 				lwkt_yield();
340 			}
341 
342 			/*
343 			 * We're already too high in the address space
344 			 * to succeed, reset to 0 for the next iteration.
345 			 */
346 			start = 0;
347 			continue;	/* next pass */
348 		}
349 		start = i;
350 
351 		/*
352 		 * Check successive pages for contiguous and free.
353 		 *
354 		 * (still in critical section)
355 		 */
356 		for (i = start + 1; i < (start + size / PAGE_SIZE); i++) {
357 			m = &pga[i];
358 			pqtype = m->queue - m->pc;
359 			if ((VM_PAGE_TO_PHYS(&m[0]) !=
360 			    (VM_PAGE_TO_PHYS(&m[-1]) + PAGE_SIZE)) ||
361 			    ((pqtype != PQ_FREE) && (pqtype != PQ_CACHE)) ||
362 			    m->busy || m->wire_count ||
363 			    m->hold_count ||
364 			    (m->flags & (PG_BUSY | PG_NEED_COMMIT)))
365 			{
366 				start++;
367 				goto again;
368 			}
369 		}
370 
371 		/*
372 		 * Try to allocate the pages, wiring them as we go.
373 		 *
374 		 * (still in critical section)
375 		 */
376 		for (i = start; i < (start + size / PAGE_SIZE); i++) {
377 			m = &pga[i];
378 
379 			if (vm_page_busy_try(m, TRUE)) {
380 				vm_contig_pg_free(start,
381 						  (i - start) * PAGE_SIZE);
382 				start++;
383 				goto again;
384 			}
385 			pqtype = m->queue - m->pc;
386 			if (pqtype == PQ_CACHE &&
387 			    m->hold_count == 0 &&
388 			    m->wire_count == 0 &&
389 			    (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0) {
390 				vm_page_protect(m, VM_PROT_NONE);
391 				KKASSERT((m->flags & PG_MAPPED) == 0);
392 				KKASSERT(m->dirty == 0);
393 				vm_page_free(m);
394 				--i;
395 				continue;	/* retry the page */
396 			}
397 			if (pqtype != PQ_FREE || m->hold_count) {
398 				vm_page_wakeup(m);
399 				vm_contig_pg_free(start,
400 						  (i - start) * PAGE_SIZE);
401 				start++;
402 				goto again;
403 			}
404 			KKASSERT((m->valid & m->dirty) == 0);
405 			KKASSERT(m->wire_count == 0);
406 			KKASSERT(m->object == NULL);
407 			vm_page_unqueue_nowakeup(m);
408 			m->valid = VM_PAGE_BITS_ALL;
409 			if (m->flags & PG_ZERO)
410 				vm_page_zero_count--;
411 			KASSERT(m->dirty == 0,
412 				("vm_contig_pg_alloc: page %p was dirty", m));
413 			KKASSERT(m->wire_count == 0);
414 			KKASSERT(m->busy == 0);
415 
416 			/*
417 			 * Clear all flags except PG_BUSY, PG_ZERO, and
418 			 * PG_WANTED, then unbusy the now allocated page.
419 			 */
420 			vm_page_flag_clear(m, ~(PG_BUSY | PG_SBUSY |
421 						PG_ZERO | PG_WANTED));
422 			vm_page_wire(m);
423 			vm_page_wakeup(m);
424 		}
425 
426 		/*
427 		 * Our job is done, return the index page of vm_page_array.
428 		 */
429 		return (start); /* aka &pga[start] */
430 	}
431 
432 	/*
433 	 * Failed.
434 	 */
435 	return (-1);
436 }
437 
438 /*
439  * vm_contig_pg_free:
440  *
441  * Remove pages previously allocated by vm_contig_pg_alloc, and
442  * assume all references to the pages have been removed, and that
443  * it is OK to add them back to the free list.
444  *
445  * Caller must ensure no races on the page range in question.
446  * No other requirements.
447  */
448 static void
449 vm_contig_pg_free(int start, u_long size)
450 {
451 	vm_page_t pga = vm_page_array;
452 
453 	size = round_page(size);
454 	if (size == 0)
455 		panic("vm_contig_pg_free: size must not be 0");
456 
457 	/*
458 	 * The pages are wired, vm_page_free_contig() determines whether they
459 	 * belong to the contig space or not and either frees them to that
460 	 * space (leaving them wired), or unwires the page and frees it to the
461 	 * normal PQ_FREE queue.
462 	 */
463 	vm_page_free_contig(&pga[start], size);
464 }
465 
466 /*
467  * vm_contig_pg_kmap:
468  *
469  * Map previously allocated (vm_contig_pg_alloc) range of pages from
470  * vm_page_array[] into the KVA.  Once mapped, the pages are part of
471  * the Kernel, and are to free'ed with kmem_free(&kernel_map, addr, size).
472  *
473  * No requirements.
474  */
475 static vm_offset_t
476 vm_contig_pg_kmap(int start, u_long size, vm_map_t map, int flags)
477 {
478 	vm_offset_t addr;
479 	vm_paddr_t pa;
480 	vm_page_t pga = vm_page_array;
481 	u_long offset;
482 
483 	if (size == 0)
484 		panic("vm_contig_pg_kmap: size must not be 0");
485 	size = round_page(size);
486 	addr = kmem_alloc_pageable(&kernel_map, size);
487 	if (addr) {
488 		pa = VM_PAGE_TO_PHYS(&pga[start]);
489 		for (offset = 0; offset < size; offset += PAGE_SIZE)
490 			pmap_kenter_quick(addr + offset, pa + offset);
491 		smp_invltlb();
492 		if (flags & M_ZERO)
493 			bzero((void *)addr, size);
494 	}
495 	return(addr);
496 }
497 
498 /*
499  * No requirements.
500  */
501 void *
502 contigmalloc(
503 	unsigned long size,	/* should be size_t here and for malloc() */
504 	struct malloc_type *type,
505 	int flags,
506 	vm_paddr_t low,
507 	vm_paddr_t high,
508 	unsigned long alignment,
509 	unsigned long boundary)
510 {
511 	return contigmalloc_map(size, type, flags, low, high, alignment,
512 			boundary, &kernel_map);
513 }
514 
515 /*
516  * No requirements.
517  */
518 void *
519 contigmalloc_map(unsigned long size, struct malloc_type *type,
520 		 int flags, vm_paddr_t low, vm_paddr_t high,
521 		 unsigned long alignment, unsigned long boundary,
522 		 vm_map_t map)
523 {
524 	int index;
525 	void *rv;
526 
527 	index = vm_contig_pg_alloc(size, low, high, alignment, boundary, flags);
528 	if (index < 0) {
529 		kprintf("contigmalloc_map: failed size %lu low=%llx "
530 			"high=%llx align=%lu boundary=%lu flags=%08x\n",
531 			size, (long long)low, (long long)high,
532 			alignment, boundary, flags);
533 		return NULL;
534 	}
535 
536 	rv = (void *)vm_contig_pg_kmap(index, size, map, flags);
537 	if (rv == NULL)
538 		vm_contig_pg_free(index, size);
539 
540 	return rv;
541 }
542 
543 /*
544  * No requirements.
545  */
546 void
547 contigfree(void *addr, unsigned long size, struct malloc_type *type)
548 {
549 	vm_paddr_t pa;
550 	vm_page_t m;
551 
552 	if (size == 0)
553 		panic("vm_contig_pg_kmap: size must not be 0");
554 	size = round_page(size);
555 
556 	pa = pmap_extract(&kernel_pmap, (vm_offset_t)addr);
557 	pmap_qremove((vm_offset_t)addr, size / PAGE_SIZE);
558 	kmem_free(&kernel_map, (vm_offset_t)addr, size);
559 
560 	m = PHYS_TO_VM_PAGE(pa);
561 	vm_page_free_contig(m, size);
562 }
563 
564 /*
565  * No requirements.
566  */
567 vm_offset_t
568 kmem_alloc_contig(vm_offset_t size, vm_paddr_t low, vm_paddr_t high,
569 		  vm_offset_t alignment)
570 {
571 	return ((vm_offset_t)contigmalloc_map(size, M_DEVBUF, M_NOWAIT, low,
572 				high, alignment, 0ul, &kernel_map));
573 }
574