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