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/lock.h>
  68 #include <sys/malloc.h>
  69 #include <sys/mount.h>
  70 #include <sys/mutex.h>
  71 #include <sys/proc.h>
  72 #include <sys/kernel.h>
  73 #include <sys/sysctl.h>
  74 #include <sys/vmmeter.h>
  75 #include <sys/vnode.h>
  76
  77 #include <vm/vm.h>
  78 #include <vm/vm_param.h>
  79 #include <vm/vm_kern.h>
  80 #include <vm/pmap.h>
  81 #include <vm/vm_map.h>
  82 #include <vm/vm_object.h>
  83 #include <vm/vm_page.h>
  84 #include <vm/vm_pageout.h>
  85 #include <vm/vm_pager.h>
  86 #include <vm/vm_phys.h>
  87 #include <vm/vm_extern.h>
  88
  89 static int
  90 vm_contig_launder_page(vm_page_t m, vm_page_t *next)
  91 {
  92         vm_object_t object;
  93         vm_page_t m_tmp;
  94         struct vnode *vp;
  95         struct mount *mp;
  96         int vfslocked;
  97
  98         mtx_assert(&vm_page_queue_mtx, MA_OWNED);
  99         vm_page_lock_assert(m, MA_OWNED);
 100         object = m->object;
 101         if (!VM_OBJECT_TRYLOCK(object) &&
 102             (!vm_pageout_fallback_object_lock(m, next) || m->hold_count != 0)) {
 103                 vm_page_unlock(m);
 104                 VM_OBJECT_UNLOCK(object);
 105                 return (EAGAIN);
 106         }
 107         if (vm_page_sleep_if_busy(m, TRUE, "vpctw0")) {
 108                 VM_OBJECT_UNLOCK(object);
 109                 vm_page_lock_queues();
 110                 return (EBUSY);
 111         }
 112         vm_page_test_dirty(m);
 113         if (m->dirty == 0)
 114                 pmap_remove_all(m);
 115         if (m->dirty != 0) {
 116                 vm_page_unlock(m);
 117                 if ((object->flags & OBJ_DEAD) != 0) {
 118                         VM_OBJECT_UNLOCK(object);
 119                         return (EAGAIN);
 120                 }
 121                 if (object->type == OBJT_VNODE) {
 122                         vm_page_unlock_queues();
 123                         vp = object->handle;
 124                         vm_object_reference_locked(object);
 125                         VM_OBJECT_UNLOCK(object);
 126                         (void) vn_start_write(vp, &mp, V_WAIT);
 127                         vfslocked = VFS_LOCK_GIANT(vp->v_mount);
 128                         vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 129                         VM_OBJECT_LOCK(object);
 130                         vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
 131                         VM_OBJECT_UNLOCK(object);
 132                         VOP_UNLOCK(vp, 0);
 133                         VFS_UNLOCK_GIANT(vfslocked);
 134                         vm_object_deallocate(object);
 135                         vn_finished_write(mp);
 136                         vm_page_lock_queues();
 137                         return (0);
 138                 } else if (object->type == OBJT_SWAP ||
 139                            object->type == OBJT_DEFAULT) {
 140                         vm_page_unlock_queues();
 141                         m_tmp = m;
 142                         vm_pageout_flush(&m_tmp, 1, VM_PAGER_PUT_SYNC, 0,
 143                             NULL, NULL);
 144                         VM_OBJECT_UNLOCK(object);
 145                         vm_page_lock_queues();
 146                         return (0);
 147                 }
 148         } else {
 149                 vm_page_cache(m);
 150                 vm_page_unlock(m);
 151         }
 152         VM_OBJECT_UNLOCK(object);
 153         return (0);
 154 }
 155
 156 static int
 157 vm_contig_launder(int queue, vm_paddr_t low, vm_paddr_t high)
 158 {
 159         vm_page_t m, next;
 160         vm_paddr_t pa;
 161         int error;
 162
 163         TAILQ_FOREACH_SAFE(m, &vm_page_queues[queue].pl, pageq, next) {
 164
 165                 /* Skip marker pages */
 166                 if ((m->flags & PG_MARKER) != 0)
 167                         continue;
 168
 169                 pa = VM_PAGE_TO_PHYS(m);
 170                 if (pa < low || pa + PAGE_SIZE > high)
 171                         continue;
 172
 173                 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) {
 174                         vm_page_unlock(m);
 175                         continue;
 176                 }
 177                 KASSERT(m->queue == queue,
 178                     ("vm_contig_launder: page %p's queue is not %d", m, queue));
 179                 error = vm_contig_launder_page(m, &next);
 180                 vm_page_lock_assert(m, MA_NOTOWNED);
 181                 if (error == 0)
 182                         return (TRUE);
 183                 if (error == EBUSY)
 184                         return (FALSE);
 185         }
 186         return (FALSE);
 187 }
 188
 189 /*
 190  *      Frees the given physically contiguous pages.
 191  *
 192  *      N.B.: Any pages with PG_ZERO set must, in fact, be zero filled.
 193  */
 194 static void
 195 vm_page_release_contig(vm_page_t m, vm_pindex_t count)
 196 {
 197
 198         while (count--) {
 199                 /* Leave PG_ZERO unchanged. */
 200                 vm_page_free_toq(m);
 201                 m++;
 202         }
 203 }
 204
 205 /*
 206  * Increase the number of cached pages.
 207  */
 208 void
 209 vm_contig_grow_cache(int tries, vm_paddr_t low, vm_paddr_t high)
 210 {
 211         int actl, actmax, inactl, inactmax;
 212
 213         vm_page_lock_queues();
 214         inactl = 0;
 215         inactmax = tries < 1 ? 0 : cnt.v_inactive_count;
 216         actl = 0;
 217         actmax = tries < 2 ? 0 : cnt.v_active_count;
 218 again:
 219         if (inactl < inactmax && vm_contig_launder(PQ_INACTIVE, low, high)) {
 220                 inactl++;
 221                 goto again;
 222         }
 223         if (actl < actmax && vm_contig_launder(PQ_ACTIVE, low, high)) {
 224                 actl++;
 225                 goto again;
 226         }
 227         vm_page_unlock_queues();
 228 }
 229
 230 /*
 231  * Allocates a region from the kernel address map and pages within the
 232  * specified physical address range to the kernel object, creates a wired
 233  * mapping from the region to these pages, and returns the region's starting
 234  * virtual address.  The allocated pages are not necessarily physically
 235  * contiguous.  If M_ZERO is specified through the given flags, then the pages
 236  * are zeroed before they are mapped.
 237  */
 238 vm_offset_t
 239 kmem_alloc_attr(vm_map_t map, vm_size_t size, int flags, vm_paddr_t low,
 240     vm_paddr_t high, vm_memattr_t memattr)
 241 {
 242         vm_object_t object = kernel_object;
 243         vm_offset_t addr, i, offset;
 244         vm_page_t m;
 245         int tries;
 246
 247         size = round_page(size);
 248         vm_map_lock(map);
 249         if (vm_map_findspace(map, vm_map_min(map), size, &addr)) {
 250                 vm_map_unlock(map);
 251                 return (0);
 252         }
 253         offset = addr - VM_MIN_KERNEL_ADDRESS;
 254         vm_object_reference(object);
 255         vm_map_insert(map, object, offset, addr, addr + size, VM_PROT_ALL,
 256             VM_PROT_ALL, 0);
 257         VM_OBJECT_LOCK(object);
 258         for (i = 0; i < size; i += PAGE_SIZE) {
 259                 tries = 0;
 260 retry:
 261                 m = vm_phys_alloc_contig(1, low, high, PAGE_SIZE, 0);
 262                 if (m == NULL) {
 263                         if (tries < ((flags & M_NOWAIT) != 0 ? 1 : 3)) {
 264                                 VM_OBJECT_UNLOCK(object);
 265                                 vm_map_unlock(map);
 266                                 vm_contig_grow_cache(tries, low, high);
 267                                 vm_map_lock(map);
 268                                 VM_OBJECT_LOCK(object);
 269                                 tries++;
 270                                 goto retry;
 271                         }
 272                         while (i != 0) {
 273                                 i -= PAGE_SIZE;
 274                                 m = vm_page_lookup(object, OFF_TO_IDX(offset +
 275                                     i));
 276                                 vm_page_free(m);
 277                         }
 278                         VM_OBJECT_UNLOCK(object);
 279                         vm_map_delete(map, addr, addr + size);
 280                         vm_map_unlock(map);
 281                         return (0);
 282                 }
 283                 if (memattr != VM_MEMATTR_DEFAULT)
 284                         pmap_page_set_memattr(m, memattr);
 285                 vm_page_insert(m, object, OFF_TO_IDX(offset + i));
 286                 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
 287                         pmap_zero_page(m);
 288                 m->valid = VM_PAGE_BITS_ALL;
 289         }
 290         VM_OBJECT_UNLOCK(object);
 291         vm_map_unlock(map);
 292         vm_map_wire(map, addr, addr + size, VM_MAP_WIRE_SYSTEM |
 293             VM_MAP_WIRE_NOHOLES);
 294         return (addr);
 295 }
 296
 297 /*
 298  *      Allocates a region from the kernel address map, inserts the
 299  *      given physically contiguous pages into the kernel object,
 300  *      creates a wired mapping from the region to the pages, and
 301  *      returns the region's starting virtual address.  If M_ZERO is
 302  *      specified through the given flags, then the pages are zeroed
 303  *      before they are mapped.
 304  */
 305 static vm_offset_t
 306 contigmapping(vm_map_t map, vm_size_t size, vm_page_t m, vm_memattr_t memattr,
 307     int flags)
 308 {
 309         vm_object_t object = kernel_object;
 310         vm_offset_t addr, tmp_addr;
 311
 312         vm_map_lock(map);
 313         if (vm_map_findspace(map, vm_map_min(map), size, &addr)) {
 314                 vm_map_unlock(map);
 315                 return (0);
 316         }
 317         vm_object_reference(object);
 318         vm_map_insert(map, object, addr - VM_MIN_KERNEL_ADDRESS,
 319             addr, addr + size, VM_PROT_ALL, VM_PROT_ALL, 0);
 320         vm_map_unlock(map);
 321         VM_OBJECT_LOCK(object);
 322         for (tmp_addr = addr; tmp_addr < addr + size; tmp_addr += PAGE_SIZE) {
 323                 if (memattr != VM_MEMATTR_DEFAULT)
 324                         pmap_page_set_memattr(m, memattr);
 325                 vm_page_insert(m, object,
 326                     OFF_TO_IDX(tmp_addr - VM_MIN_KERNEL_ADDRESS));
 327                 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
 328                         pmap_zero_page(m);
 329                 m->valid = VM_PAGE_BITS_ALL;
 330                 m++;
 331         }
 332         VM_OBJECT_UNLOCK(object);
 333         vm_map_wire(map, addr, addr + size,
 334             VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
 335         return (addr);
 336 }
 337
 338 void *
 339 contigmalloc(
 340         unsigned long size,     /* should be size_t here and for malloc() */
 341         struct malloc_type *type,
 342         int flags,
 343         vm_paddr_t low,
 344         vm_paddr_t high,
 345         unsigned long alignment,
 346         unsigned long boundary)
 347 {
 348         void *ret;
 349
 350         ret = (void *)kmem_alloc_contig(kernel_map, size, flags, low, high,
 351             alignment, boundary, VM_MEMATTR_DEFAULT);
 352         if (ret != NULL)
 353                 malloc_type_allocated(type, round_page(size));
 354         return (ret);
 355 }
 356
 357 vm_offset_t
 358 kmem_alloc_contig(vm_map_t map, vm_size_t size, int flags, vm_paddr_t low,
 359     vm_paddr_t high, unsigned long alignment, unsigned long boundary,
 360     vm_memattr_t memattr)
 361 {
 362         vm_offset_t ret;
 363         vm_page_t pages;
 364         unsigned long npgs;
 365         int tries;
 366
 367         size = round_page(size);
 368         npgs = size >> PAGE_SHIFT;
 369         tries = 0;
 370 retry:
 371         pages = vm_phys_alloc_contig(npgs, low, high, alignment, boundary);
 372         if (pages == NULL) {
 373                 if (tries < ((flags & M_NOWAIT) != 0 ? 1 : 3)) {
 374                         vm_contig_grow_cache(tries, low, high);
 375                         tries++;
 376                         goto retry;
 377                 }
 378                 ret = 0;
 379         } else {
 380                 ret = contigmapping(map, size, pages, memattr, flags);
 381                 if (ret == 0)
 382                         vm_page_release_contig(pages, npgs);
 383         }
 384         return (ret);
 385 }
 386
 387 void
 388 contigfree(void *addr, unsigned long size, struct malloc_type *type)
 389 {
 390
 391         kmem_free(kernel_map, (vm_offset_t)addr, size);
 392         malloc_type_freed(type, round_page(size));
 393 }