sys/vm/phys_pager.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2000 Peter Wemm
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice, this list of conditions and the following disclaimer.
  11  * 2. Redistributions in binary form must reproduce the above copyright
  12  *    notice, this list of conditions and the following disclaimer in the
  13  *    documentation and/or other materials provided with the distribution.
  14  *
  15  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
  16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
  19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  25  * SUCH DAMAGE.
  26  */
  27
  28 #include <sys/cdefs.h>
  29 __FBSDID("$FreeBSD$");
  30
  31 #include <sys/param.h>
  32 #include <sys/systm.h>
  33 #include <sys/conf.h>
  34 #include <sys/kernel.h>
  35 #include <sys/lock.h>
  36 #include <sys/proc.h>
  37 #include <sys/mutex.h>
  38 #include <sys/mman.h>
  39 #include <sys/rwlock.h>
  40 #include <sys/sysctl.h>
  41
  42 #include <vm/vm.h>
  43 #include <vm/vm_param.h>
  44 #include <vm/vm_object.h>
  45 #include <vm/vm_page.h>
  46 #include <vm/vm_pageout.h>
  47 #include <vm/vm_pager.h>
  48
  49 /* list of phys pager objects */
  50 static struct pagerlst phys_pager_object_list;
  51 /* protect access to phys_pager_object_list */
  52 static struct mtx phys_pager_mtx;
  53
  54 static int default_phys_pager_getpages(vm_object_t object, vm_page_t *m,
  55     int count, int *rbehind, int *rahead);
  56 static int default_phys_pager_populate(vm_object_t object, vm_pindex_t pidx,
  57     int fault_type, vm_prot_t max_prot, vm_pindex_t *first, vm_pindex_t *last);
  58 static boolean_t default_phys_pager_haspage(vm_object_t object,
  59     vm_pindex_t pindex, int *before, int *after);
  60 struct phys_pager_ops default_phys_pg_ops = {
  61         .phys_pg_getpages = default_phys_pager_getpages,
  62         .phys_pg_populate = default_phys_pager_populate,
  63         .phys_pg_haspage = default_phys_pager_haspage,
  64         .phys_pg_ctor = NULL,
  65         .phys_pg_dtor = NULL,
  66 };
  67
  68 static void
  69 phys_pager_init(void)
  70 {
  71
  72         TAILQ_INIT(&phys_pager_object_list);
  73         mtx_init(&phys_pager_mtx, "phys_pager list", NULL, MTX_DEF);
  74 }
  75
  76 vm_object_t
  77 phys_pager_allocate(void *handle, struct phys_pager_ops *ops, void *data,
  78     vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred)
  79 {
  80         vm_object_t object, object1;
  81         vm_pindex_t pindex;
  82         bool init;
  83
  84         /*
  85          * Offset should be page aligned.
  86          */
  87         if (foff & PAGE_MASK)
  88                 return (NULL);
  89
  90         pindex = OFF_TO_IDX(foff + PAGE_MASK + size);
  91         init = true;
  92
  93         if (handle != NULL) {
  94                 mtx_lock(&phys_pager_mtx);
  95                 /*
  96                  * Look up pager, creating as necessary.
  97                  */
  98                 object1 = NULL;
  99                 object = vm_pager_object_lookup(&phys_pager_object_list, handle);
 100                 if (object == NULL) {
 101                         /*
 102                          * Allocate object and associate it with the pager.
 103                          */
 104                         mtx_unlock(&phys_pager_mtx);
 105                         object1 = vm_object_allocate(OBJT_PHYS, pindex);
 106                         mtx_lock(&phys_pager_mtx);
 107                         object = vm_pager_object_lookup(&phys_pager_object_list,
 108                             handle);
 109                         if (object != NULL) {
 110                                 /*
 111                                  * We raced with other thread while
 112                                  * allocating object.
 113                                  */
 114                                 if (pindex > object->size)
 115                                         object->size = pindex;
 116                                 init = false;
 117                         } else {
 118                                 object = object1;
 119                                 object1 = NULL;
 120                                 object->handle = handle;
 121                                 object->un_pager.phys.ops = ops;
 122                                 object->un_pager.phys.data_ptr = data;
 123                                 if (ops->phys_pg_populate != NULL)
 124                                         vm_object_set_flag(object, OBJ_POPULATE);
 125                                 TAILQ_INSERT_TAIL(&phys_pager_object_list,
 126                                     object, pager_object_list);
 127                         }
 128                 } else {
 129                         if (pindex > object->size)
 130                                 object->size = pindex;
 131                 }
 132                 mtx_unlock(&phys_pager_mtx);
 133                 vm_object_deallocate(object1);
 134         } else {
 135                 object = vm_object_allocate(OBJT_PHYS, pindex);
 136                 object->un_pager.phys.ops = ops;
 137                 object->un_pager.phys.data_ptr = data;
 138                 if (ops->phys_pg_populate != NULL)
 139                         vm_object_set_flag(object, OBJ_POPULATE);
 140         }
 141         if (init && ops->phys_pg_ctor != NULL)
 142                 ops->phys_pg_ctor(object, prot, foff, cred);
 143
 144         return (object);
 145 }
 146
 147 static vm_object_t
 148 phys_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
 149     vm_ooffset_t foff, struct ucred *ucred)
 150 {
 151         return (phys_pager_allocate(handle, &default_phys_pg_ops, NULL,
 152             size, prot, foff, ucred));
 153 }
 154
 155 static void
 156 phys_pager_dealloc(vm_object_t object)
 157 {
 158
 159         if (object->handle != NULL) {
 160                 VM_OBJECT_WUNLOCK(object);
 161                 mtx_lock(&phys_pager_mtx);
 162                 TAILQ_REMOVE(&phys_pager_object_list, object, pager_object_list);
 163                 mtx_unlock(&phys_pager_mtx);
 164                 VM_OBJECT_WLOCK(object);
 165         }
 166         object->type = OBJT_DEAD;
 167         if (object->un_pager.phys.ops->phys_pg_dtor != NULL)
 168                 object->un_pager.phys.ops->phys_pg_dtor(object);
 169         object->handle = NULL;
 170 }
 171
 172 /*
 173  * Fill as many pages as vm_fault has allocated for us.
 174  */
 175 static int
 176 default_phys_pager_getpages(vm_object_t object, vm_page_t *m, int count,
 177     int *rbehind, int *rahead)
 178 {
 179         int i;
 180
 181         VM_OBJECT_ASSERT_WLOCKED(object);
 182         for (i = 0; i < count; i++) {
 183                 if (m[i]->valid == 0) {
 184                         if ((m[i]->flags & PG_ZERO) == 0)
 185                                 pmap_zero_page(m[i]);
 186                         m[i]->valid = VM_PAGE_BITS_ALL;
 187                 }
 188                 KASSERT(m[i]->valid == VM_PAGE_BITS_ALL,
 189                     ("phys_pager_getpages: partially valid page %p", m[i]));
 190                 KASSERT(m[i]->dirty == 0,
 191                     ("phys_pager_getpages: dirty page %p", m[i]));
 192         }
 193         if (rbehind)
 194                 *rbehind = 0;
 195         if (rahead)
 196                 *rahead = 0;
 197         return (VM_PAGER_OK);
 198 }
 199
 200 static int
 201 phys_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
 202     int *rahead)
 203 {
 204         return (object->un_pager.phys.ops->phys_pg_getpages(object, m,
 205             count, rbehind, rahead));
 206 }
 207
 208 /*
 209  * Implement a pretty aggressive clustered getpages strategy.  Hint that
 210  * everything in an entire 4MB window should be prefaulted at once.
 211  *
 212  * 4MB (1024 slots per page table page) is convenient for x86,
 213  * but may not be for other arches.
 214  */
 215 #ifndef PHYSCLUSTER
 216 #define PHYSCLUSTER 1024
 217 #endif
 218 static int phys_pager_cluster = PHYSCLUSTER;
 219 SYSCTL_INT(_vm, OID_AUTO, phys_pager_cluster, CTLFLAG_RWTUN,
 220     &phys_pager_cluster, 0,
 221     "prefault window size for phys pager");
 222
 223 /*
 224  * Max hint to vm_page_alloc() about the further allocation needs
 225  * inside the phys_pager_populate() loop.  The number of bits used to
 226  * implement VM_ALLOC_COUNT() determines the hard limit on this value.
 227  * That limit is currently 65535.
 228  */
 229 #define PHYSALLOC       16
 230
 231 static int
 232 default_phys_pager_populate(vm_object_t object, vm_pindex_t pidx,
 233     int fault_type __unused, vm_prot_t max_prot __unused, vm_pindex_t *first,
 234     vm_pindex_t *last)
 235 {
 236         vm_page_t m;
 237         vm_pindex_t base, end, i;
 238         int ahead;
 239
 240         base = rounddown(pidx, phys_pager_cluster);
 241         end = base + phys_pager_cluster - 1;
 242         if (end >= object->size)
 243                 end = object->size - 1;
 244         if (*first > base)
 245                 base = *first;
 246         if (end > *last)
 247                 end = *last;
 248         *first = base;
 249         *last = end;
 250
 251         for (i = base; i <= end; i++) {
 252 retry:
 253                 m = vm_page_lookup(object, i);
 254                 if (m == NULL) {
 255                         ahead = MIN(end - i, PHYSALLOC);
 256                         m = vm_page_alloc(object, i, VM_ALLOC_NORMAL |
 257                             VM_ALLOC_ZERO | VM_ALLOC_WAITFAIL |
 258                             VM_ALLOC_COUNT(ahead));
 259                         if (m == NULL)
 260                                 goto retry;
 261                         if ((m->flags & PG_ZERO) == 0)
 262                                 pmap_zero_page(m);
 263                         m->valid = VM_PAGE_BITS_ALL;
 264                 } else if (vm_page_xbusied(m)) {
 265                         vm_page_lock(m);
 266                         VM_OBJECT_WUNLOCK(object);
 267                         vm_page_busy_sleep(m, "physb", true);
 268                         VM_OBJECT_WLOCK(object);
 269                         goto retry;
 270                 } else {
 271                         vm_page_xbusy(m);
 272                         if (m->valid != VM_PAGE_BITS_ALL)
 273                                 vm_page_zero_invalid(m, TRUE);
 274                 }
 275
 276                 KASSERT(m->valid == VM_PAGE_BITS_ALL,
 277                     ("phys_pager_populate: partially valid page %p", m));
 278                 KASSERT(m->dirty == 0,
 279                     ("phys_pager_populate: dirty page %p", m));
 280         }
 281         return (VM_PAGER_OK);
 282 }
 283
 284 static int
 285 phys_pager_populate(vm_object_t object, vm_pindex_t pidx, int fault_type,
 286     vm_prot_t max_prot, vm_pindex_t *first, vm_pindex_t *last)
 287 {
 288         return (object->un_pager.phys.ops->phys_pg_populate(object, pidx,
 289             fault_type, max_prot, first, last));
 290 }
 291
 292 static void
 293 phys_pager_putpages(vm_object_t object, vm_page_t *m, int count, boolean_t sync,
 294     int *rtvals)
 295 {
 296
 297         panic("phys_pager_putpage called");
 298 }
 299
 300 static boolean_t
 301 default_phys_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
 302     int *after)
 303 {
 304         vm_pindex_t base, end;
 305
 306         base = rounddown(pindex, phys_pager_cluster);
 307         end = base + phys_pager_cluster - 1;
 308         if (before != NULL)
 309                 *before = pindex - base;
 310         if (after != NULL)
 311                 *after = end - pindex;
 312         return (TRUE);
 313 }
 314
 315 static boolean_t
 316 phys_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
 317     int *after)
 318 {
 319         return (object->un_pager.phys.ops->phys_pg_haspage(object, pindex,
 320             before, after));
 321 }
 322
 323 struct pagerops physpagerops = {
 324         .pgo_init =     phys_pager_init,
 325         .pgo_alloc =    phys_pager_alloc,
 326         .pgo_dealloc =  phys_pager_dealloc,
 327         .pgo_getpages = phys_pager_getpages,
 328         .pgo_putpages = phys_pager_putpages,
 329         .pgo_haspage =  phys_pager_haspage,
 330         .pgo_populate = phys_pager_populate,
 331 };