]> CyberLeo.Net >> Repos - FreeBSD/releng/9.2.git/blob - sys/vm/vm_contig.c
- Copy stable/9 to releng/9.2 as part of the 9.2-RELEASE cycle.
[FreeBSD/releng/9.2.git] / sys / vm / vm_contig.c
1 /*-
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * The Mach Operating System project at Carnegie-Mellon University.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
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 the
15  *    documentation and/or other materials provided with the distribution.
16  * 4. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  *      from: @(#)vm_page.c     7.4 (Berkeley) 5/7/91
33  */
34
35 /*-
36  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
37  * All rights reserved.
38  *
39  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
40  *
41  * Permission to use, copy, modify and distribute this software and
42  * its documentation is hereby granted, provided that both the copyright
43  * notice and this permission notice appear in all copies of the
44  * software, derivative works or modified versions, and any portions
45  * thereof, and that both notices appear in supporting documentation.
46  *
47  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
48  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
49  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
50  *
51  * Carnegie Mellon requests users of this software to return to
52  *
53  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
54  *  School of Computer Science
55  *  Carnegie Mellon University
56  *  Pittsburgh PA 15213-3890
57  *
58  * any improvements or extensions that they make and grant Carnegie the
59  * rights to redistribute these changes.
60  */
61
62 #include <sys/cdefs.h>
63 __FBSDID("$FreeBSD$");
64
65 #include <sys/param.h>
66 #include <sys/systm.h>
67 #include <sys/eventhandler.h>
68 #include <sys/lock.h>
69 #include <sys/malloc.h>
70 #include <sys/mount.h>
71 #include <sys/mutex.h>
72 #include <sys/proc.h>
73 #include <sys/kernel.h>
74 #include <sys/sysctl.h>
75 #include <sys/vmmeter.h>
76 #include <sys/vnode.h>
77
78 #include <vm/vm.h>
79 #include <vm/vm_param.h>
80 #include <vm/vm_kern.h>
81 #include <vm/pmap.h>
82 #include <vm/vm_map.h>
83 #include <vm/vm_object.h>
84 #include <vm/vm_page.h>
85 #include <vm/vm_pageout.h>
86 #include <vm/vm_pager.h>
87 #include <vm/vm_phys.h>
88 #include <vm/vm_extern.h>
89
90 static int
91 vm_contig_launder_page(vm_page_t m, vm_page_t *next, int tries)
92 {
93         vm_object_t object;
94         vm_page_t m_tmp;
95         struct vnode *vp;
96         struct mount *mp;
97         int vfslocked;
98
99         mtx_assert(&vm_page_queue_mtx, MA_OWNED);
100         if (!vm_pageout_page_lock(m, next) || m->hold_count != 0) {
101                 vm_page_unlock(m);
102                 return (EAGAIN);
103         }
104         object = m->object;
105         if (!VM_OBJECT_TRYLOCK(object) &&
106             (!vm_pageout_fallback_object_lock(m, next) || m->hold_count != 0)) {
107                 vm_page_unlock(m);
108                 VM_OBJECT_UNLOCK(object);
109                 return (EAGAIN);
110         }
111         if ((m->oflags & VPO_BUSY) != 0 || m->busy != 0) {
112                 if (tries == 0) {
113                         vm_page_unlock(m);
114                         VM_OBJECT_UNLOCK(object);
115                         return (EAGAIN);
116                 }
117                 vm_page_sleep(m, "vpctw0");
118                 VM_OBJECT_UNLOCK(object);
119                 vm_page_lock_queues();
120                 return (EBUSY);
121         }
122         vm_page_test_dirty(m);
123         if (m->dirty == 0)
124                 pmap_remove_all(m);
125         if (m->dirty != 0) {
126                 vm_page_unlock(m);
127                 if (tries == 0 || (object->flags & OBJ_DEAD) != 0) {
128                         VM_OBJECT_UNLOCK(object);
129                         return (EAGAIN);
130                 }
131                 if (object->type == OBJT_VNODE) {
132                         vm_page_unlock_queues();
133                         vp = object->handle;
134                         vm_object_reference_locked(object);
135                         VM_OBJECT_UNLOCK(object);
136                         (void) vn_start_write(vp, &mp, V_WAIT);
137                         vfslocked = VFS_LOCK_GIANT(vp->v_mount);
138                         vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
139                         VM_OBJECT_LOCK(object);
140                         vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
141                         VM_OBJECT_UNLOCK(object);
142                         VOP_UNLOCK(vp, 0);
143                         VFS_UNLOCK_GIANT(vfslocked);
144                         vm_object_deallocate(object);
145                         vn_finished_write(mp);
146                         vm_page_lock_queues();
147                         return (0);
148                 } else if (object->type == OBJT_SWAP ||
149                            object->type == OBJT_DEFAULT) {
150                         vm_page_unlock_queues();
151                         m_tmp = m;
152                         vm_pageout_flush(&m_tmp, 1, VM_PAGER_PUT_SYNC, 0,
153                             NULL, NULL);
154                         VM_OBJECT_UNLOCK(object);
155                         vm_page_lock_queues();
156                         return (0);
157                 }
158         } else {
159                 vm_page_cache(m);
160                 vm_page_unlock(m);
161         }
162         VM_OBJECT_UNLOCK(object);
163         return (EAGAIN);
164 }
165
166 static int
167 vm_contig_launder(int queue, int tries, vm_paddr_t low, vm_paddr_t high)
168 {
169         vm_page_t m, next;
170         vm_paddr_t pa;
171         int error;
172
173         TAILQ_FOREACH_SAFE(m, &vm_page_queues[queue].pl, pageq, next) {
174                 KASSERT(m->queue == queue,
175                     ("vm_contig_launder: page %p's queue is not %d", m, queue));
176                 if ((m->flags & PG_MARKER) != 0)
177                         continue;
178                 pa = VM_PAGE_TO_PHYS(m);
179                 if (pa < low || pa + PAGE_SIZE > high)
180                         continue;
181                 error = vm_contig_launder_page(m, &next, tries);
182                 if (error == 0)
183                         return (TRUE);
184                 if (error == EBUSY)
185                         return (FALSE);
186         }
187         return (FALSE);
188 }
189
190 /*
191  *      Frees the given physically contiguous pages.
192  *
193  *      N.B.: Any pages with PG_ZERO set must, in fact, be zero filled.
194  */
195 static void
196 vm_page_release_contig(vm_page_t m, vm_pindex_t count)
197 {
198
199         while (count--) {
200                 /* Leave PG_ZERO unchanged. */
201                 vm_page_free_toq(m);
202                 m++;
203         }
204 }
205
206 /*
207  * Increase the number of cached pages.  The specified value, "tries",
208  * determines which categories of pages are cached:
209  *
210  *  0: All clean, inactive pages within the specified physical address range
211  *     are cached.  Will not sleep.
212  *  1: The vm_lowmem handlers are called.  All inactive pages within
213  *     the specified physical address range are cached.  May sleep.
214  *  2: The vm_lowmem handlers are called.  All inactive and active pages
215  *     within the specified physical address range are cached.  May sleep.
216  */
217 void
218 vm_contig_grow_cache(int tries, vm_paddr_t low, vm_paddr_t high)
219 {
220         int actl, actmax, inactl, inactmax;
221
222         if (tries > 0) {
223                 /*
224                  * Decrease registered cache sizes.  The vm_lowmem handlers
225                  * may acquire locks and/or sleep, so they can only be invoked
226                  * when "tries" is greater than zero.
227                  */
228                 EVENTHANDLER_INVOKE(vm_lowmem, 0);
229
230                 /*
231                  * We do this explicitly after the caches have been drained
232                  * above.
233                  */
234                 uma_reclaim();
235         }
236         vm_page_lock_queues();
237         inactl = 0;
238         inactmax = cnt.v_inactive_count;
239         actl = 0;
240         actmax = tries < 2 ? 0 : cnt.v_active_count;
241 again:
242         if (inactl < inactmax && vm_contig_launder(PQ_INACTIVE, tries, low,
243             high)) {
244                 inactl++;
245                 goto again;
246         }
247         if (actl < actmax && vm_contig_launder(PQ_ACTIVE, tries, low, high)) {
248                 actl++;
249                 goto again;
250         }
251         vm_page_unlock_queues();
252 }
253
254 /*
255  * Allocates a region from the kernel address map and pages within the
256  * specified physical address range to the kernel object, creates a wired
257  * mapping from the region to these pages, and returns the region's starting
258  * virtual address.  The allocated pages are not necessarily physically
259  * contiguous.  If M_ZERO is specified through the given flags, then the pages
260  * are zeroed before they are mapped.
261  */
262 vm_offset_t
263 kmem_alloc_attr(vm_map_t map, vm_size_t size, int flags, vm_paddr_t low,
264     vm_paddr_t high, vm_memattr_t memattr)
265 {
266         vm_object_t object = kernel_object;
267         vm_offset_t addr, i, offset;
268         vm_page_t m;
269         int tries;
270
271         size = round_page(size);
272         vm_map_lock(map);
273         if (vm_map_findspace(map, vm_map_min(map), size, &addr)) {
274                 vm_map_unlock(map);
275                 return (0);
276         }
277         offset = addr - VM_MIN_KERNEL_ADDRESS;
278         vm_object_reference(object);
279         vm_map_insert(map, object, offset, addr, addr + size, VM_PROT_ALL,
280             VM_PROT_ALL, 0);
281         VM_OBJECT_LOCK(object);
282         for (i = 0; i < size; i += PAGE_SIZE) {
283                 tries = 0;
284 retry:
285                 m = vm_phys_alloc_contig(1, low, high, PAGE_SIZE, 0);
286                 if (m == NULL) {
287                         if (tries < ((flags & M_NOWAIT) != 0 ? 1 : 3)) {
288                                 VM_OBJECT_UNLOCK(object);
289                                 vm_map_unlock(map);
290                                 vm_contig_grow_cache(tries, low, high);
291                                 vm_map_lock(map);
292                                 VM_OBJECT_LOCK(object);
293                                 tries++;
294                                 goto retry;
295                         }
296                         while (i != 0) {
297                                 i -= PAGE_SIZE;
298                                 m = vm_page_lookup(object, OFF_TO_IDX(offset +
299                                     i));
300                                 vm_page_free(m);
301                         }
302                         VM_OBJECT_UNLOCK(object);
303                         vm_map_delete(map, addr, addr + size);
304                         vm_map_unlock(map);
305                         return (0);
306                 }
307                 if (memattr != VM_MEMATTR_DEFAULT)
308                         pmap_page_set_memattr(m, memattr);
309                 vm_page_insert(m, object, OFF_TO_IDX(offset + i));
310                 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
311                         pmap_zero_page(m);
312                 m->valid = VM_PAGE_BITS_ALL;
313         }
314         VM_OBJECT_UNLOCK(object);
315         vm_map_unlock(map);
316         vm_map_wire(map, addr, addr + size, VM_MAP_WIRE_SYSTEM |
317             VM_MAP_WIRE_NOHOLES);
318         return (addr);
319 }
320
321 /*
322  *      Allocates a region from the kernel address map, inserts the
323  *      given physically contiguous pages into the kernel object,
324  *      creates a wired mapping from the region to the pages, and
325  *      returns the region's starting virtual address.  If M_ZERO is
326  *      specified through the given flags, then the pages are zeroed
327  *      before they are mapped.
328  */
329 static vm_offset_t
330 contigmapping(vm_map_t map, vm_size_t size, vm_page_t m, vm_memattr_t memattr,
331     int flags)
332 {
333         vm_object_t object = kernel_object;
334         vm_offset_t addr, tmp_addr;
335  
336         vm_map_lock(map);
337         if (vm_map_findspace(map, vm_map_min(map), size, &addr)) {
338                 vm_map_unlock(map);
339                 return (0);
340         }
341         vm_object_reference(object);
342         vm_map_insert(map, object, addr - VM_MIN_KERNEL_ADDRESS,
343             addr, addr + size, VM_PROT_ALL, VM_PROT_ALL, 0);
344         vm_map_unlock(map);
345         VM_OBJECT_LOCK(object);
346         for (tmp_addr = addr; tmp_addr < addr + size; tmp_addr += PAGE_SIZE) {
347                 if (memattr != VM_MEMATTR_DEFAULT)
348                         pmap_page_set_memattr(m, memattr);
349                 vm_page_insert(m, object,
350                     OFF_TO_IDX(tmp_addr - VM_MIN_KERNEL_ADDRESS));
351                 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
352                         pmap_zero_page(m);
353                 m->valid = VM_PAGE_BITS_ALL;
354                 m++;
355         }
356         VM_OBJECT_UNLOCK(object);
357         vm_map_wire(map, addr, addr + size,
358             VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
359         return (addr);
360 }
361
362 void *
363 contigmalloc(
364         unsigned long size,     /* should be size_t here and for malloc() */
365         struct malloc_type *type,
366         int flags,
367         vm_paddr_t low,
368         vm_paddr_t high,
369         unsigned long alignment,
370         unsigned long boundary)
371 {
372         void *ret;
373
374         ret = (void *)kmem_alloc_contig(kernel_map, size, flags, low, high,
375             alignment, boundary, VM_MEMATTR_DEFAULT);
376         if (ret != NULL)
377                 malloc_type_allocated(type, round_page(size));
378         return (ret);
379 }
380
381 vm_offset_t
382 kmem_alloc_contig(vm_map_t map, vm_size_t size, int flags, vm_paddr_t low,
383     vm_paddr_t high, unsigned long alignment, unsigned long boundary,
384     vm_memattr_t memattr)
385 {
386         vm_offset_t ret;
387         vm_page_t pages;
388         unsigned long npgs;
389         int tries;
390
391         size = round_page(size);
392         npgs = size >> PAGE_SHIFT;
393         tries = 0;
394 retry:
395         pages = vm_phys_alloc_contig(npgs, low, high, alignment, boundary);
396         if (pages == NULL) {
397                 if (tries < ((flags & M_NOWAIT) != 0 ? 1 : 3)) {
398                         vm_contig_grow_cache(tries, low, high);
399                         tries++;
400                         goto retry;
401                 }
402                 ret = 0;
403         } else {
404                 ret = contigmapping(map, size, pages, memattr, flags);
405                 if (ret == 0)
406                         vm_page_release_contig(pages, npgs);
407         }
408         return (ret);
409 }
410
411 void
412 contigfree(void *addr, unsigned long size, struct malloc_type *type)
413 {
414
415         kmem_free(kernel_map, (vm_offset_t)addr, size);
416         malloc_type_freed(type, round_page(size));
417 }