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, NULL);
 143                         VM_OBJECT_UNLOCK(object);
 144                         vm_page_lock_queues();
 145                         return (0);
 146                 }
 147         } else {
 148                 vm_page_cache(m);
 149                 vm_page_unlock(m);
 150         }
 151         VM_OBJECT_UNLOCK(object);
 152         return (0);
 153 }
 154
 155 static int
 156 vm_contig_launder(int queue, vm_paddr_t low, vm_paddr_t high)
 157 {
 158         vm_page_t m, next;
 159         vm_paddr_t pa;
 160         int error;
 161
 162         TAILQ_FOREACH_SAFE(m, &vm_page_queues[queue].pl, pageq, next) {
 163
 164                 /* Skip marker pages */
 165                 if ((m->flags & PG_MARKER) != 0)
 166                         continue;
 167
 168                 pa = VM_PAGE_TO_PHYS(m);
 169                 if (pa < low || pa + PAGE_SIZE > high)
 170                         continue;
 171
 172                 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) {
 173                         vm_page_unlock(m);
 174                         continue;
 175                 }
 176                 KASSERT(m->queue == queue,
 177                     ("vm_contig_launder: page %p's queue is not %d", m, queue));
 178                 error = vm_contig_launder_page(m, &next);
 179                 vm_page_lock_assert(m, MA_NOTOWNED);
 180                 if (error == 0)
 181                         return (TRUE);
 182                 if (error == EBUSY)
 183                         return (FALSE);
 184         }
 185         return (FALSE);
 186 }
 187
 188 /*
 189  *      Frees the given physically contiguous pages.
 190  *
 191  *      N.B.: Any pages with PG_ZERO set must, in fact, be zero filled.
 192  */
 193 static void
 194 vm_page_release_contig(vm_page_t m, vm_pindex_t count)
 195 {
 196
 197         while (count--) {
 198                 /* Leave PG_ZERO unchanged. */
 199                 vm_page_free_toq(m);
 200                 m++;
 201         }
 202 }
 203
 204 /*
 205  * Increase the number of cached pages.
 206  */
 207 void
 208 vm_contig_grow_cache(int tries, vm_paddr_t low, vm_paddr_t high)
 209 {
 210         int actl, actmax, inactl, inactmax;
 211
 212         vm_page_lock_queues();
 213         inactl = 0;
 214         inactmax = tries < 1 ? 0 : cnt.v_inactive_count;
 215         actl = 0;
 216         actmax = tries < 2 ? 0 : cnt.v_active_count;
 217 again:
 218         if (inactl < inactmax && vm_contig_launder(PQ_INACTIVE, low, high)) {
 219                 inactl++;
 220                 goto again;
 221         }
 222         if (actl < actmax && vm_contig_launder(PQ_ACTIVE, low, high)) {
 223                 actl++;
 224                 goto again;
 225         }
 226         vm_page_unlock_queues();
 227 }
 228
 229 /*
 230  * Allocates a region from the kernel address map and pages within the
 231  * specified physical address range to the kernel object, creates a wired
 232  * mapping from the region to these pages, and returns the region's starting
 233  * virtual address.  The allocated pages are not necessarily physically
 234  * contiguous.  If M_ZERO is specified through the given flags, then the pages
 235  * are zeroed before they are mapped.
 236  */
 237 vm_offset_t
 238 kmem_alloc_attr(vm_map_t map, vm_size_t size, int flags, vm_paddr_t low,
 239     vm_paddr_t high, vm_memattr_t memattr)
 240 {
 241         vm_object_t object = kernel_object;
 242         vm_offset_t addr, i, offset;
 243         vm_page_t m;
 244         int tries;
 245
 246         size = round_page(size);
 247         vm_map_lock(map);
 248         if (vm_map_findspace(map, vm_map_min(map), size, &addr)) {
 249                 vm_map_unlock(map);
 250                 return (0);
 251         }
 252         offset = addr - VM_MIN_KERNEL_ADDRESS;
 253         vm_object_reference(object);
 254         vm_map_insert(map, object, offset, addr, addr + size, VM_PROT_ALL,
 255             VM_PROT_ALL, 0);
 256         VM_OBJECT_LOCK(object);
 257         for (i = 0; i < size; i += PAGE_SIZE) {
 258                 tries = 0;
 259 retry:
 260                 m = vm_phys_alloc_contig(1, low, high, PAGE_SIZE, 0);
 261                 if (m == NULL) {
 262                         if (tries < ((flags & M_NOWAIT) != 0 ? 1 : 3)) {
 263                                 VM_OBJECT_UNLOCK(object);
 264                                 vm_map_unlock(map);
 265                                 vm_contig_grow_cache(tries, low, high);
 266                                 vm_map_lock(map);
 267                                 VM_OBJECT_LOCK(object);
 268                                 tries++;
 269                                 goto retry;
 270                         }
 271                         while (i != 0) {
 272                                 i -= PAGE_SIZE;
 273                                 m = vm_page_lookup(object, OFF_TO_IDX(offset +
 274                                     i));
 275                                 vm_page_free(m);
 276                         }
 277                         VM_OBJECT_UNLOCK(object);
 278                         vm_map_delete(map, addr, addr + size);
 279                         vm_map_unlock(map);
 280                         return (0);
 281                 }
 282                 if (memattr != VM_MEMATTR_DEFAULT)
 283                         pmap_page_set_memattr(m, memattr);
 284                 vm_page_insert(m, object, OFF_TO_IDX(offset + i));
 285                 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
 286                         pmap_zero_page(m);
 287                 m->valid = VM_PAGE_BITS_ALL;
 288         }
 289         VM_OBJECT_UNLOCK(object);
 290         vm_map_unlock(map);
 291         vm_map_wire(map, addr, addr + size, VM_MAP_WIRE_SYSTEM |
 292             VM_MAP_WIRE_NOHOLES);
 293         return (addr);
 294 }
 295
 296 /*
 297  *      Allocates a region from the kernel address map, inserts the
 298  *      given physically contiguous pages into the kernel object,
 299  *      creates a wired mapping from the region to the pages, and
 300  *      returns the region's starting virtual address.  If M_ZERO is
 301  *      specified through the given flags, then the pages are zeroed
 302  *      before they are mapped.
 303  */
 304 static vm_offset_t
 305 contigmapping(vm_map_t map, vm_size_t size, vm_page_t m, vm_memattr_t memattr,
 306     int flags)
 307 {
 308         vm_object_t object = kernel_object;
 309         vm_offset_t addr, tmp_addr;
 310
 311         vm_map_lock(map);
 312         if (vm_map_findspace(map, vm_map_min(map), size, &addr)) {
 313                 vm_map_unlock(map);
 314                 return (0);
 315         }
 316         vm_object_reference(object);
 317         vm_map_insert(map, object, addr - VM_MIN_KERNEL_ADDRESS,
 318             addr, addr + size, VM_PROT_ALL, VM_PROT_ALL, 0);
 319         vm_map_unlock(map);
 320         VM_OBJECT_LOCK(object);
 321         for (tmp_addr = addr; tmp_addr < addr + size; tmp_addr += PAGE_SIZE) {
 322                 if (memattr != VM_MEMATTR_DEFAULT)
 323                         pmap_page_set_memattr(m, memattr);
 324                 vm_page_insert(m, object,
 325                     OFF_TO_IDX(tmp_addr - VM_MIN_KERNEL_ADDRESS));
 326                 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
 327                         pmap_zero_page(m);
 328                 m->valid = VM_PAGE_BITS_ALL;
 329                 m++;
 330         }
 331         VM_OBJECT_UNLOCK(object);
 332         vm_map_wire(map, addr, addr + size,
 333             VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
 334         return (addr);
 335 }
 336
 337 void *
 338 contigmalloc(
 339         unsigned long size,     /* should be size_t here and for malloc() */
 340         struct malloc_type *type,
 341         int flags,
 342         vm_paddr_t low,
 343         vm_paddr_t high,
 344         unsigned long alignment,
 345         unsigned long boundary)
 346 {
 347         void *ret;
 348
 349         ret = (void *)kmem_alloc_contig(kernel_map, size, flags, low, high,
 350             alignment, boundary, VM_MEMATTR_DEFAULT);
 351         if (ret != NULL)
 352                 malloc_type_allocated(type, round_page(size));
 353         return (ret);
 354 }
 355
 356 vm_offset_t
 357 kmem_alloc_contig(vm_map_t map, vm_size_t size, int flags, vm_paddr_t low,
 358     vm_paddr_t high, unsigned long alignment, unsigned long boundary,
 359     vm_memattr_t memattr)
 360 {
 361         vm_offset_t ret;
 362         vm_page_t pages;
 363         unsigned long npgs;
 364         int tries;
 365
 366         size = round_page(size);
 367         npgs = size >> PAGE_SHIFT;
 368         tries = 0;
 369 retry:
 370         pages = vm_phys_alloc_contig(npgs, low, high, alignment, boundary);
 371         if (pages == NULL) {
 372                 if (tries < ((flags & M_NOWAIT) != 0 ? 1 : 3)) {
 373                         vm_contig_grow_cache(tries, low, high);
 374                         tries++;
 375                         goto retry;
 376                 }
 377                 ret = 0;
 378         } else {
 379                 ret = contigmapping(map, size, pages, memattr, flags);
 380                 if (ret == 0)
 381                         vm_page_release_contig(pages, npgs);
 382         }
 383         return (ret);
 384 }
 385
 386 void
 387 contigfree(void *addr, unsigned long size, struct malloc_type *type)
 388 {
 389
 390         kmem_free(kernel_map, (vm_offset_t)addr, size);
 391         malloc_type_freed(type, round_page(size));
 392 }