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/linker_set.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)
  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         object = m->object;
 101         if (!VM_OBJECT_TRYLOCK(object) &&
 102             !vm_pageout_fallback_object_lock(m, next)) {
 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 && m->hold_count == 0)
 113                 pmap_remove_all(m);
 114         if (m->dirty) {
 115                 if ((object->flags & OBJ_DEAD) != 0) {
 116                         VM_OBJECT_UNLOCK(object);
 117                         return (EAGAIN);
 118                 }
 119                 if (object->type == OBJT_VNODE) {
 120                         vm_page_unlock_queues();
 121                         vp = object->handle;
 122                         vm_object_reference_locked(object);
 123                         VM_OBJECT_UNLOCK(object);
 124                         (void) vn_start_write(vp, &mp, V_WAIT);
 125                         vfslocked = VFS_LOCK_GIANT(vp->v_mount);
 126                         vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curthread);
 127                         VM_OBJECT_LOCK(object);
 128                         vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
 129                         VM_OBJECT_UNLOCK(object);
 130                         VOP_UNLOCK(vp, 0, curthread);
 131                         VFS_UNLOCK_GIANT(vfslocked);
 132                         vm_object_deallocate(object);
 133                         vn_finished_write(mp);
 134                         vm_page_lock_queues();
 135                         return (0);
 136                 } else if (object->type == OBJT_SWAP ||
 137                            object->type == OBJT_DEFAULT) {
 138                         m_tmp = m;
 139                         vm_pageout_flush(&m_tmp, 1, VM_PAGER_PUT_SYNC);
 140                         VM_OBJECT_UNLOCK(object);
 141                         return (0);
 142                 }
 143         } else if (m->hold_count == 0)
 144                 vm_page_cache(m);
 145         VM_OBJECT_UNLOCK(object);
 146         return (0);
 147 }
 148
 149 static int
 150 vm_contig_launder(int queue)
 151 {
 152         vm_page_t m, next;
 153         int error;
 154
 155         TAILQ_FOREACH_SAFE(m, &vm_page_queues[queue].pl, pageq, next) {
 156
 157                 /* Skip marker pages */
 158                 if ((m->flags & PG_MARKER) != 0)
 159                         continue;
 160
 161                 KASSERT(VM_PAGE_INQUEUE2(m, queue),
 162                     ("vm_contig_launder: page %p's queue is not %d", m, queue));
 163                 error = vm_contig_launder_page(m, &next);
 164                 if (error == 0)
 165                         return (TRUE);
 166                 if (error == EBUSY)
 167                         return (FALSE);
 168         }
 169         return (FALSE);
 170 }
 171
 172 /*
 173  *      Frees the given physically contiguous pages.
 174  *
 175  *      N.B.: Any pages with PG_ZERO set must, in fact, be zero filled.
 176  */
 177 static void
 178 vm_page_release_contig(vm_page_t m, vm_pindex_t count)
 179 {
 180
 181         while (count--) {
 182                 /* Leave PG_ZERO unchanged. */
 183                 vm_page_free_toq(m);
 184                 m++;
 185         }
 186 }
 187
 188 /*
 189  *      Allocates a region from the kernel address map, inserts the
 190  *      given physically contiguous pages into the kernel object,
 191  *      creates a wired mapping from the region to the pages, and
 192  *      returns the region's starting virtual address.  If M_ZERO is
 193  *      specified through the given flags, then the pages are zeroed
 194  *      before they are mapped.
 195  */
 196 static void *
 197 contigmapping(vm_page_t m, vm_pindex_t npages, int flags)
 198 {
 199         vm_object_t object = kernel_object;
 200         vm_map_t map = kernel_map;
 201         vm_offset_t addr, tmp_addr;
 202         vm_pindex_t i;
 203
 204         vm_map_lock(map);
 205         if (vm_map_findspace(map, vm_map_min(map), npages << PAGE_SHIFT, &addr)
 206             != KERN_SUCCESS) {
 207                 vm_map_unlock(map);
 208                 return (NULL);
 209         }
 210         vm_object_reference(object);
 211         vm_map_insert(map, object, addr - VM_MIN_KERNEL_ADDRESS,
 212             addr, addr + (npages << PAGE_SHIFT), VM_PROT_ALL, VM_PROT_ALL, 0);
 213         vm_map_unlock(map);
 214         tmp_addr = addr;
 215         VM_OBJECT_LOCK(object);
 216         for (i = 0; i < npages; i++) {
 217                 vm_page_insert(&m[i], object,
 218                     OFF_TO_IDX(tmp_addr - VM_MIN_KERNEL_ADDRESS));
 219                 if ((flags & M_ZERO) && !(m[i].flags & PG_ZERO))
 220                         pmap_zero_page(&m[i]);
 221                 tmp_addr += PAGE_SIZE;
 222         }
 223         VM_OBJECT_UNLOCK(object);
 224         vm_map_wire(map, addr, addr + (npages << PAGE_SHIFT),
 225             VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
 226         return ((void *)addr);
 227 }
 228
 229 void *
 230 contigmalloc(
 231         unsigned long size,     /* should be size_t here and for malloc() */
 232         struct malloc_type *type,
 233         int flags,
 234         vm_paddr_t low,
 235         vm_paddr_t high,
 236         unsigned long alignment,
 237         unsigned long boundary)
 238 {
 239         void *ret;
 240         vm_page_t pages;
 241         unsigned long npgs;
 242         int actl, actmax, inactl, inactmax, tries;
 243
 244         npgs = round_page(size) >> PAGE_SHIFT;
 245         tries = 0;
 246 retry:
 247         pages = vm_phys_alloc_contig(npgs, low, high, alignment, boundary);
 248         if (pages == NULL) {
 249                 if (tries < ((flags & M_NOWAIT) != 0 ? 1 : 3)) {
 250                         vm_page_lock_queues();
 251                         inactl = 0;
 252                         inactmax = tries < 1 ? 0 : cnt.v_inactive_count;
 253                         actl = 0;
 254                         actmax = tries < 2 ? 0 : cnt.v_active_count;
 255 again:
 256                         if (inactl < inactmax &&
 257                             vm_contig_launder(PQ_INACTIVE)) {
 258                                 inactl++;
 259                                 goto again;
 260                         }
 261                         if (actl < actmax &&
 262                             vm_contig_launder(PQ_ACTIVE)) {
 263                                 actl++;
 264                                 goto again;
 265                         }
 266                         vm_page_unlock_queues();
 267                         tries++;
 268                         goto retry;
 269                 }
 270                 ret = NULL;
 271         } else {
 272                 ret = contigmapping(pages, npgs, flags);
 273                 if (ret == NULL)
 274                         vm_page_release_contig(pages, npgs);
 275                 else
 276                         malloc_type_allocated(type, npgs << PAGE_SHIFT);
 277         }
 278         return (ret);
 279 }
 280
 281 void
 282 contigfree(void *addr, unsigned long size, struct malloc_type *type)
 283 {
 284         vm_pindex_t npgs;
 285
 286         npgs = round_page(size) >> PAGE_SHIFT;
 287         kmem_free(kernel_map, (vm_offset_t)addr, size);
 288         malloc_type_freed(type, npgs << PAGE_SHIFT);
 289 }