sys/vm/vm_reserv.c

   1 /*-
   2  * Copyright (c) 2002-2006 Rice University
   3  * Copyright (c) 2007-2008 Alan L. Cox <alc@cs.rice.edu>
   4  * All rights reserved.
   5  *
   6  * This software was developed for the FreeBSD Project by Alan L. Cox,
   7  * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  19  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  20  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  21  * A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT
  22  * HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  23  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
  24  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
  25  * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  26  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY
  28  * WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
  29  * POSSIBILITY OF SUCH DAMAGE.
  30  */
  31
  32 /*
  33  *      Superpage reservation management module
  34  *
  35  * Any external functions defined by this module are only to be used by the
  36  * virtual memory system.
  37  */
  38
  39 #include <sys/cdefs.h>
  40 __FBSDID("$FreeBSD$");
  41
  42 #include "opt_vm.h"
  43
  44 #include <sys/param.h>
  45 #include <sys/kernel.h>
  46 #include <sys/lock.h>
  47 #include <sys/malloc.h>
  48 #include <sys/mutex.h>
  49 #include <sys/queue.h>
  50 #include <sys/rwlock.h>
  51 #include <sys/sbuf.h>
  52 #include <sys/sysctl.h>
  53 #include <sys/systm.h>
  54
  55 #include <vm/vm.h>
  56 #include <vm/vm_param.h>
  57 #include <vm/vm_object.h>
  58 #include <vm/vm_page.h>
  59 #include <vm/vm_phys.h>
  60 #include <vm/vm_radix.h>
  61 #include <vm/vm_reserv.h>
  62
  63 /*
  64  * The reservation system supports the speculative allocation of large physical
  65  * pages ("superpages").  Speculative allocation enables the fully-automatic
  66  * utilization of superpages by the virtual memory system.  In other words, no
  67  * programmatic directives are required to use superpages.
  68  */
  69
  70 #if VM_NRESERVLEVEL > 0
  71
  72 /*
  73  * The number of small pages that are contained in a level 0 reservation
  74  */
  75 #define VM_LEVEL_0_NPAGES       (1 << VM_LEVEL_0_ORDER)
  76
  77 /*
  78  * The number of bits by which a physical address is shifted to obtain the
  79  * reservation number
  80  */
  81 #define VM_LEVEL_0_SHIFT        (VM_LEVEL_0_ORDER + PAGE_SHIFT)
  82
  83 /*
  84  * The size of a level 0 reservation in bytes
  85  */
  86 #define VM_LEVEL_0_SIZE         (1 << VM_LEVEL_0_SHIFT)
  87
  88 /*
  89  * Computes the index of the small page underlying the given (object, pindex)
  90  * within the reservation's array of small pages.
  91  */
  92 #define VM_RESERV_INDEX(object, pindex) \
  93     (((object)->pg_color + (pindex)) & (VM_LEVEL_0_NPAGES - 1))
  94
  95 /*
  96  * The reservation structure
  97  *
  98  * A reservation structure is constructed whenever a large physical page is
  99  * speculatively allocated to an object.  The reservation provides the small
 100  * physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets
 101  * within that object.  The reservation's "popcnt" tracks the number of these
 102  * small physical pages that are in use at any given time.  When and if the
 103  * reservation is not fully utilized, it appears in the queue of partially-
 104  * populated reservations.  The reservation always appears on the containing
 105  * object's list of reservations.
 106  *
 107  * A partially-populated reservation can be broken and reclaimed at any time.
 108  */
 109 struct vm_reserv {
 110         TAILQ_ENTRY(vm_reserv) partpopq;
 111         LIST_ENTRY(vm_reserv) objq;
 112         vm_object_t     object;                 /* containing object */
 113         vm_pindex_t     pindex;                 /* offset within object */
 114         vm_page_t       pages;                  /* first page of a superpage */
 115         int             popcnt;                 /* # of pages in use */
 116         char            inpartpopq;
 117 };
 118
 119 /*
 120  * The reservation array
 121  *
 122  * This array is analoguous in function to vm_page_array.  It differs in the
 123  * respect that it may contain a greater number of useful reservation
 124  * structures than there are (physical) superpages.  These "invalid"
 125  * reservation structures exist to trade-off space for time in the
 126  * implementation of vm_reserv_from_page().  Invalid reservation structures are
 127  * distinguishable from "valid" reservation structures by inspecting the
 128  * reservation's "pages" field.  Invalid reservation structures have a NULL
 129  * "pages" field.
 130  *
 131  * vm_reserv_from_page() maps a small (physical) page to an element of this
 132  * array by computing a physical reservation number from the page's physical
 133  * address.  The physical reservation number is used as the array index.
 134  *
 135  * An "active" reservation is a valid reservation structure that has a non-NULL
 136  * "object" field and a non-zero "popcnt" field.  In other words, every active
 137  * reservation belongs to a particular object.  Moreover, every active
 138  * reservation has an entry in the containing object's list of reservations.
 139  */
 140 static vm_reserv_t vm_reserv_array;
 141
 142 /*
 143  * The partially-populated reservation queue
 144  *
 145  * This queue enables the fast recovery of an unused cached or free small page
 146  * from a partially-populated reservation.  The reservation at the head of
 147  * this queue is the least-recently-changed, partially-populated reservation.
 148  *
 149  * Access to this queue is synchronized by the free page queue lock.
 150  */
 151 static TAILQ_HEAD(, vm_reserv) vm_rvq_partpop =
 152                             TAILQ_HEAD_INITIALIZER(vm_rvq_partpop);
 153
 154 static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD, 0, "Reservation Info");
 155
 156 static long vm_reserv_broken;
 157 SYSCTL_LONG(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD,
 158     &vm_reserv_broken, 0, "Cumulative number of broken reservations");
 159
 160 static long vm_reserv_freed;
 161 SYSCTL_LONG(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD,
 162     &vm_reserv_freed, 0, "Cumulative number of freed reservations");
 163
 164 static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);
 165
 166 SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
 167     sysctl_vm_reserv_partpopq, "A", "Partially-populated reservation queues");
 168
 169 static long vm_reserv_reclaimed;
 170 SYSCTL_LONG(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
 171     &vm_reserv_reclaimed, 0, "Cumulative number of reclaimed reservations");
 172
 173 static void             vm_reserv_depopulate(vm_reserv_t rv);
 174 static vm_reserv_t      vm_reserv_from_page(vm_page_t m);
 175 static boolean_t        vm_reserv_has_pindex(vm_reserv_t rv,
 176                             vm_pindex_t pindex);
 177 static void             vm_reserv_populate(vm_reserv_t rv);
 178 static void             vm_reserv_reclaim(vm_reserv_t rv);
 179
 180 /*
 181  * Describes the current state of the partially-populated reservation queue.
 182  */
 183 static int
 184 sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
 185 {
 186         struct sbuf sbuf;
 187         vm_reserv_t rv;
 188         int counter, error, level, unused_pages;
 189
 190         error = sysctl_wire_old_buffer(req, 0);
 191         if (error != 0)
 192                 return (error);
 193         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 194         sbuf_printf(&sbuf, "\nLEVEL     SIZE  NUMBER\n\n");
 195         for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
 196                 counter = 0;
 197                 unused_pages = 0;
 198                 mtx_lock(&vm_page_queue_free_mtx);
 199                 TAILQ_FOREACH(rv, &vm_rvq_partpop/*[level]*/, partpopq) {
 200                         counter++;
 201                         unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
 202                 }
 203                 mtx_unlock(&vm_page_queue_free_mtx);
 204                 sbuf_printf(&sbuf, "%5d: %6dK, %6d\n", level,
 205                     unused_pages * ((int)PAGE_SIZE / 1024), counter);
 206         }
 207         error = sbuf_finish(&sbuf);
 208         sbuf_delete(&sbuf);
 209         return (error);
 210 }
 211
 212 /*
 213  * Reduces the given reservation's population count.  If the population count
 214  * becomes zero, the reservation is destroyed.  Additionally, moves the
 215  * reservation to the tail of the partially-populated reservations queue if the
 216  * population count is non-zero.
 217  *
 218  * The free page queue lock must be held.
 219  */
 220 static void
 221 vm_reserv_depopulate(vm_reserv_t rv)
 222 {
 223
 224         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 225         KASSERT(rv->object != NULL,
 226             ("vm_reserv_depopulate: reserv %p is free", rv));
 227         KASSERT(rv->popcnt > 0,
 228             ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
 229         if (rv->inpartpopq) {
 230                 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
 231                 rv->inpartpopq = FALSE;
 232         } else {
 233                 KASSERT(rv->pages->psind == 1,
 234                     ("vm_reserv_depopulate: reserv %p is already demoted",
 235                     rv));
 236                 rv->pages->psind = 0;
 237         }
 238         rv->popcnt--;
 239         if (rv->popcnt == 0) {
 240                 LIST_REMOVE(rv, objq);
 241                 rv->object = NULL;
 242                 vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
 243                 vm_reserv_freed++;
 244         } else {
 245                 rv->inpartpopq = TRUE;
 246                 TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
 247         }
 248 }
 249
 250 /*
 251  * Returns the reservation to which the given page might belong.
 252  */
 253 static __inline vm_reserv_t
 254 vm_reserv_from_page(vm_page_t m)
 255 {
 256
 257         return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
 258 }
 259
 260 /*
 261  * Returns TRUE if the given reservation contains the given page index and
 262  * FALSE otherwise.
 263  */
 264 static __inline boolean_t
 265 vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
 266 {
 267
 268         return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
 269 }
 270
 271 /*
 272  * Increases the given reservation's population count.  Moves the reservation
 273  * to the tail of the partially-populated reservation queue.
 274  *
 275  * The free page queue must be locked.
 276  */
 277 static void
 278 vm_reserv_populate(vm_reserv_t rv)
 279 {
 280
 281         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 282         KASSERT(rv->object != NULL,
 283             ("vm_reserv_populate: reserv %p is free", rv));
 284         KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
 285             ("vm_reserv_populate: reserv %p is already full", rv));
 286         KASSERT(rv->pages->psind == 0,
 287             ("vm_reserv_populate: reserv %p is already promoted", rv));
 288         if (rv->inpartpopq) {
 289                 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
 290                 rv->inpartpopq = FALSE;
 291         }
 292         rv->popcnt++;
 293         if (rv->popcnt < VM_LEVEL_0_NPAGES) {
 294                 rv->inpartpopq = TRUE;
 295                 TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
 296         } else
 297                 rv->pages->psind = 1;
 298 }
 299
 300 /*
 301  * Allocates a contiguous set of physical pages of the given size "npages"
 302  * from existing or newly created reservations.  All of the physical pages
 303  * must be at or above the given physical address "low" and below the given
 304  * physical address "high".  The given value "alignment" determines the
 305  * alignment of the first physical page in the set.  If the given value
 306  * "boundary" is non-zero, then the set of physical pages cannot cross any
 307  * physical address boundary that is a multiple of that value.  Both
 308  * "alignment" and "boundary" must be a power of two.
 309  *
 310  * The object and free page queue must be locked.
 311  */
 312 vm_page_t
 313 vm_reserv_alloc_contig(vm_object_t object, vm_pindex_t pindex, u_long npages,
 314     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
 315 {
 316         vm_paddr_t pa, size;
 317         vm_page_t m, m_ret, mpred, msucc;
 318         vm_pindex_t first, leftcap, rightcap;
 319         vm_reserv_t rv;
 320         u_long allocpages, maxpages, minpages;
 321         int i, index, n;
 322
 323         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 324         VM_OBJECT_ASSERT_WLOCKED(object);
 325         KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
 326
 327         /*
 328          * Is a reservation fundamentally impossible?
 329          */
 330         if (pindex < VM_RESERV_INDEX(object, pindex) ||
 331             pindex + npages > object->size)
 332                 return (NULL);
 333
 334         /*
 335          * All reservations of a particular size have the same alignment.
 336          * Assuming that the first page is allocated from a reservation, the
 337          * least significant bits of its physical address can be determined
 338          * from its offset from the beginning of the reservation and the size
 339          * of the reservation.
 340          *
 341          * Could the specified index within a reservation of the smallest
 342          * possible size satisfy the alignment and boundary requirements?
 343          */
 344         pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
 345         if ((pa & (alignment - 1)) != 0)
 346                 return (NULL);
 347         size = npages << PAGE_SHIFT;
 348         if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
 349                 return (NULL);
 350
 351         /*
 352          * Look for an existing reservation.
 353          */
 354         mpred = vm_radix_lookup_le(&object->rtree, pindex);
 355         if (mpred != NULL) {
 356                 KASSERT(mpred->pindex < pindex,
 357                     ("vm_reserv_alloc_contig: pindex already allocated"));
 358                 rv = vm_reserv_from_page(mpred);
 359                 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
 360                         goto found;
 361                 msucc = TAILQ_NEXT(mpred, listq);
 362         } else
 363                 msucc = TAILQ_FIRST(&object->memq);
 364         if (msucc != NULL) {
 365                 KASSERT(msucc->pindex > pindex,
 366                     ("vm_reserv_alloc_contig: pindex already allocated"));
 367                 rv = vm_reserv_from_page(msucc);
 368                 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
 369                         goto found;
 370         }
 371
 372         /*
 373          * Could at least one reservation fit between the first index to the
 374          * left that can be used ("leftcap") and the first index to the right
 375          * that cannot be used ("rightcap")?
 376          */
 377         first = pindex - VM_RESERV_INDEX(object, pindex);
 378         if (mpred != NULL) {
 379                 if ((rv = vm_reserv_from_page(mpred))->object != object)
 380                         leftcap = mpred->pindex + 1;
 381                 else
 382                         leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
 383                 if (leftcap > first)
 384                         return (NULL);
 385         }
 386         minpages = VM_RESERV_INDEX(object, pindex) + npages;
 387         maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES);
 388         allocpages = maxpages;
 389         if (msucc != NULL) {
 390                 if ((rv = vm_reserv_from_page(msucc))->object != object)
 391                         rightcap = msucc->pindex;
 392                 else
 393                         rightcap = rv->pindex;
 394                 if (first + maxpages > rightcap) {
 395                         if (maxpages == VM_LEVEL_0_NPAGES)
 396                                 return (NULL);
 397
 398                         /*
 399                          * At least one reservation will fit between "leftcap"
 400                          * and "rightcap".  However, a reservation for the
 401                          * last of the requested pages will not fit.  Reduce
 402                          * the size of the upcoming allocation accordingly.
 403                          */
 404                         allocpages = minpages;
 405                 }
 406         }
 407
 408         /*
 409          * Would the last new reservation extend past the end of the object?
 410          */
 411         if (first + maxpages > object->size) {
 412                 /*
 413                  * Don't allocate the last new reservation if the object is a
 414                  * vnode or backed by another object that is a vnode.
 415                  */
 416                 if (object->type == OBJT_VNODE ||
 417                     (object->backing_object != NULL &&
 418                     object->backing_object->type == OBJT_VNODE)) {
 419                         if (maxpages == VM_LEVEL_0_NPAGES)
 420                                 return (NULL);
 421                         allocpages = minpages;
 422                 }
 423                 /* Speculate that the object may grow. */
 424         }
 425
 426         /*
 427          * Allocate the physical pages.  The alignment and boundary specified
 428          * for this allocation may be different from the alignment and
 429          * boundary specified for the requested pages.  For instance, the
 430          * specified index may not be the first page within the first new
 431          * reservation.
 432          */
 433         m = vm_phys_alloc_contig(allocpages, low, high, ulmax(alignment,
 434             VM_LEVEL_0_SIZE), boundary > VM_LEVEL_0_SIZE ? boundary : 0);
 435         if (m == NULL)
 436                 return (NULL);
 437
 438         /*
 439          * The allocated physical pages always begin at a reservation
 440          * boundary, but they do not always end at a reservation boundary.
 441          * Initialize every reservation that is completely covered by the
 442          * allocated physical pages.
 443          */
 444         m_ret = NULL;
 445         index = VM_RESERV_INDEX(object, pindex);
 446         do {
 447                 rv = vm_reserv_from_page(m);
 448                 KASSERT(rv->pages == m,
 449                     ("vm_reserv_alloc_contig: reserv %p's pages is corrupted",
 450                     rv));
 451                 KASSERT(rv->object == NULL,
 452                     ("vm_reserv_alloc_contig: reserv %p isn't free", rv));
 453                 LIST_INSERT_HEAD(&object->rvq, rv, objq);
 454                 rv->object = object;
 455                 rv->pindex = first;
 456                 KASSERT(rv->popcnt == 0,
 457                     ("vm_reserv_alloc_contig: reserv %p's popcnt is corrupted",
 458                     rv));
 459                 KASSERT(!rv->inpartpopq,
 460                     ("vm_reserv_alloc_contig: reserv %p's inpartpopq is TRUE",
 461                     rv));
 462                 n = ulmin(VM_LEVEL_0_NPAGES - index, npages);
 463                 for (i = 0; i < n; i++)
 464                         vm_reserv_populate(rv);
 465                 npages -= n;
 466                 if (m_ret == NULL) {
 467                         m_ret = &rv->pages[index];
 468                         index = 0;
 469                 }
 470                 m += VM_LEVEL_0_NPAGES;
 471                 first += VM_LEVEL_0_NPAGES;
 472                 allocpages -= VM_LEVEL_0_NPAGES;
 473         } while (allocpages >= VM_LEVEL_0_NPAGES);
 474         return (m_ret);
 475
 476         /*
 477          * Found a matching reservation.
 478          */
 479 found:
 480         index = VM_RESERV_INDEX(object, pindex);
 481         /* Does the allocation fit within the reservation? */
 482         if (index + npages > VM_LEVEL_0_NPAGES)
 483                 return (NULL);
 484         m = &rv->pages[index];
 485         pa = VM_PAGE_TO_PHYS(m);
 486         if (pa < low || pa + size > high || (pa & (alignment - 1)) != 0 ||
 487             ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
 488                 return (NULL);
 489         /* Handle vm_page_rename(m, new_object, ...). */
 490         for (i = 0; i < npages; i++)
 491                 if ((rv->pages[index + i].flags & (PG_CACHED | PG_FREE)) == 0)
 492                         return (NULL);
 493         for (i = 0; i < npages; i++)
 494                 vm_reserv_populate(rv);
 495         return (m);
 496 }
 497
 498 /*
 499  * Allocates a page from an existing or newly-created reservation.
 500  *
 501  * The page "mpred" must immediately precede the offset "pindex" within the
 502  * specified object.
 503  *
 504  * The object and free page queue must be locked.
 505  */
 506 vm_page_t
 507 vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex, vm_page_t mpred)
 508 {
 509         vm_page_t m, msucc;
 510         vm_pindex_t first, leftcap, rightcap;
 511         vm_reserv_t rv;
 512
 513         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 514         VM_OBJECT_ASSERT_WLOCKED(object);
 515
 516         /*
 517          * Is a reservation fundamentally impossible?
 518          */
 519         if (pindex < VM_RESERV_INDEX(object, pindex) ||
 520             pindex >= object->size)
 521                 return (NULL);
 522
 523         /*
 524          * Look for an existing reservation.
 525          */
 526         if (mpred != NULL) {
 527                 KASSERT(mpred->object == object,
 528                     ("vm_reserv_alloc_page: object doesn't contain mpred"));
 529                 KASSERT(mpred->pindex < pindex,
 530                     ("vm_reserv_alloc_page: mpred doesn't precede pindex"));
 531                 rv = vm_reserv_from_page(mpred);
 532                 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
 533                         goto found;
 534                 msucc = TAILQ_NEXT(mpred, listq);
 535         } else
 536                 msucc = TAILQ_FIRST(&object->memq);
 537         if (msucc != NULL) {
 538                 KASSERT(msucc->pindex > pindex,
 539                     ("vm_reserv_alloc_page: msucc doesn't succeed pindex"));
 540                 rv = vm_reserv_from_page(msucc);
 541                 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
 542                         goto found;
 543         }
 544
 545         /*
 546          * Could a reservation fit between the first index to the left that
 547          * can be used and the first index to the right that cannot be used?
 548          */
 549         first = pindex - VM_RESERV_INDEX(object, pindex);
 550         if (mpred != NULL) {
 551                 if ((rv = vm_reserv_from_page(mpred))->object != object)
 552                         leftcap = mpred->pindex + 1;
 553                 else
 554                         leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
 555                 if (leftcap > first)
 556                         return (NULL);
 557         }
 558         if (msucc != NULL) {
 559                 if ((rv = vm_reserv_from_page(msucc))->object != object)
 560                         rightcap = msucc->pindex;
 561                 else
 562                         rightcap = rv->pindex;
 563                 if (first + VM_LEVEL_0_NPAGES > rightcap)
 564                         return (NULL);
 565         }
 566
 567         /*
 568          * Would a new reservation extend past the end of the object?
 569          */
 570         if (first + VM_LEVEL_0_NPAGES > object->size) {
 571                 /*
 572                  * Don't allocate a new reservation if the object is a vnode or
 573                  * backed by another object that is a vnode.
 574                  */
 575                 if (object->type == OBJT_VNODE ||
 576                     (object->backing_object != NULL &&
 577                     object->backing_object->type == OBJT_VNODE))
 578                         return (NULL);
 579                 /* Speculate that the object may grow. */
 580         }
 581
 582         /*
 583          * Allocate and populate the new reservation.
 584          */
 585         m = vm_phys_alloc_pages(VM_FREEPOOL_DEFAULT, VM_LEVEL_0_ORDER);
 586         if (m == NULL)
 587                 return (NULL);
 588         rv = vm_reserv_from_page(m);
 589         KASSERT(rv->pages == m,
 590             ("vm_reserv_alloc_page: reserv %p's pages is corrupted", rv));
 591         KASSERT(rv->object == NULL,
 592             ("vm_reserv_alloc_page: reserv %p isn't free", rv));
 593         LIST_INSERT_HEAD(&object->rvq, rv, objq);
 594         rv->object = object;
 595         rv->pindex = first;
 596         KASSERT(rv->popcnt == 0,
 597             ("vm_reserv_alloc_page: reserv %p's popcnt is corrupted", rv));
 598         KASSERT(!rv->inpartpopq,
 599             ("vm_reserv_alloc_page: reserv %p's inpartpopq is TRUE", rv));
 600         vm_reserv_populate(rv);
 601         return (&rv->pages[VM_RESERV_INDEX(object, pindex)]);
 602
 603         /*
 604          * Found a matching reservation.
 605          */
 606 found:
 607         m = &rv->pages[VM_RESERV_INDEX(object, pindex)];
 608         /* Handle vm_page_rename(m, new_object, ...). */
 609         if ((m->flags & (PG_CACHED | PG_FREE)) == 0)
 610                 return (NULL);
 611         vm_reserv_populate(rv);
 612         return (m);
 613 }
 614
 615 /*
 616  * Breaks all reservations belonging to the given object.
 617  */
 618 void
 619 vm_reserv_break_all(vm_object_t object)
 620 {
 621         vm_reserv_t rv;
 622         int i;
 623
 624         mtx_lock(&vm_page_queue_free_mtx);
 625         while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
 626                 KASSERT(rv->object == object,
 627                     ("vm_reserv_break_all: reserv %p is corrupted", rv));
 628                 if (rv->inpartpopq) {
 629                         TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
 630                         rv->inpartpopq = FALSE;
 631                 }
 632                 LIST_REMOVE(rv, objq);
 633                 rv->object = NULL;
 634                 for (i = 0; i < VM_LEVEL_0_NPAGES; i++) {
 635                         if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
 636                                 vm_phys_free_pages(&rv->pages[i], 0);
 637                         else
 638                                 rv->popcnt--;
 639                 }
 640                 KASSERT(rv->popcnt == 0,
 641                     ("vm_reserv_break_all: reserv %p's popcnt is corrupted",
 642                     rv));
 643                 vm_reserv_broken++;
 644         }
 645         mtx_unlock(&vm_page_queue_free_mtx);
 646 }
 647
 648 /*
 649  * Frees the given page if it belongs to a reservation.  Returns TRUE if the
 650  * page is freed and FALSE otherwise.
 651  *
 652  * The free page queue lock must be held.
 653  */
 654 boolean_t
 655 vm_reserv_free_page(vm_page_t m)
 656 {
 657         vm_reserv_t rv;
 658
 659         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 660         rv = vm_reserv_from_page(m);
 661         if (rv->object == NULL)
 662                 return (FALSE);
 663         if ((m->flags & PG_CACHED) != 0 && m->pool != VM_FREEPOOL_CACHE)
 664                 vm_phys_set_pool(VM_FREEPOOL_CACHE, rv->pages,
 665                     VM_LEVEL_0_ORDER);
 666         vm_reserv_depopulate(rv);
 667         return (TRUE);
 668 }
 669
 670 /*
 671  * Initializes the reservation management system.  Specifically, initializes
 672  * the reservation array.
 673  *
 674  * Requires that vm_page_array and first_page are initialized!
 675  */
 676 void
 677 vm_reserv_init(void)
 678 {
 679         vm_paddr_t paddr;
 680         int i;
 681
 682         /*
 683          * Initialize the reservation array.  Specifically, initialize the
 684          * "pages" field for every element that has an underlying superpage.
 685          */
 686         for (i = 0; phys_avail[i + 1] != 0; i += 2) {
 687                 paddr = roundup2(phys_avail[i], VM_LEVEL_0_SIZE);
 688                 while (paddr + VM_LEVEL_0_SIZE <= phys_avail[i + 1]) {
 689                         vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT].pages =
 690                             PHYS_TO_VM_PAGE(paddr);
 691                         paddr += VM_LEVEL_0_SIZE;
 692                 }
 693         }
 694 }
 695
 696 /*
 697  * Returns a reservation level if the given page belongs to a fully-populated
 698  * reservation and -1 otherwise.
 699  */
 700 int
 701 vm_reserv_level_iffullpop(vm_page_t m)
 702 {
 703         vm_reserv_t rv;
 704
 705         rv = vm_reserv_from_page(m);
 706         return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
 707 }
 708
 709 /*
 710  * Prepare for the reactivation of a cached page.
 711  *
 712  * First, suppose that the given page "m" was allocated individually, i.e., not
 713  * as part of a reservation, and cached.  Then, suppose a reservation
 714  * containing "m" is allocated by the same object.  Although "m" and the
 715  * reservation belong to the same object, "m"'s pindex may not match the
 716  * reservation's.
 717  *
 718  * The free page queue must be locked.
 719  */
 720 boolean_t
 721 vm_reserv_reactivate_page(vm_page_t m)
 722 {
 723         vm_reserv_t rv;
 724         int i, m_index;
 725
 726         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 727         rv = vm_reserv_from_page(m);
 728         if (rv->object == NULL)
 729                 return (FALSE);
 730         KASSERT((m->flags & PG_CACHED) != 0,
 731             ("vm_reserv_uncache_page: page %p is not cached", m));
 732         if (m->object == rv->object &&
 733             m->pindex - rv->pindex == VM_RESERV_INDEX(m->object, m->pindex))
 734                 vm_reserv_populate(rv);
 735         else {
 736                 KASSERT(rv->inpartpopq,
 737                     ("vm_reserv_uncache_page: reserv %p's inpartpopq is FALSE",
 738                     rv));
 739                 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
 740                 rv->inpartpopq = FALSE;
 741                 LIST_REMOVE(rv, objq);
 742                 rv->object = NULL;
 743                 /* Don't vm_phys_free_pages(m, 0). */
 744                 m_index = m - rv->pages;
 745                 for (i = 0; i < m_index; i++) {
 746                         if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
 747                                 vm_phys_free_pages(&rv->pages[i], 0);
 748                         else
 749                                 rv->popcnt--;
 750                 }
 751                 for (i++; i < VM_LEVEL_0_NPAGES; i++) {
 752                         if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
 753                                 vm_phys_free_pages(&rv->pages[i], 0);
 754                         else
 755                                 rv->popcnt--;
 756                 }
 757                 KASSERT(rv->popcnt == 0,
 758                     ("vm_reserv_uncache_page: reserv %p's popcnt is corrupted",
 759                     rv));
 760                 vm_reserv_broken++;
 761         }
 762         return (TRUE);
 763 }
 764
 765 /*
 766  * Breaks the given partially-populated reservation, releasing its cached and
 767  * free pages to the physical memory allocator.
 768  *
 769  * The free page queue lock must be held.
 770  */
 771 static void
 772 vm_reserv_reclaim(vm_reserv_t rv)
 773 {
 774         int i;
 775
 776         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 777         KASSERT(rv->inpartpopq,
 778             ("vm_reserv_reclaim: reserv %p's inpartpopq is corrupted", rv));
 779         TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
 780         rv->inpartpopq = FALSE;
 781         KASSERT(rv->object != NULL,
 782             ("vm_reserv_reclaim: reserv %p is free", rv));
 783         LIST_REMOVE(rv, objq);
 784         rv->object = NULL;
 785         for (i = 0; i < VM_LEVEL_0_NPAGES; i++) {
 786                 if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
 787                         vm_phys_free_pages(&rv->pages[i], 0);
 788                 else
 789                         rv->popcnt--;
 790         }
 791         KASSERT(rv->popcnt == 0,
 792             ("vm_reserv_reclaim: reserv %p's popcnt is corrupted", rv));
 793         vm_reserv_reclaimed++;
 794 }
 795
 796 /*
 797  * Breaks the reservation at the head of the partially-populated reservation
 798  * queue, releasing its cached and free pages to the physical memory
 799  * allocator.  Returns TRUE if a reservation is broken and FALSE otherwise.
 800  *
 801  * The free page queue lock must be held.
 802  */
 803 boolean_t
 804 vm_reserv_reclaim_inactive(void)
 805 {
 806         vm_reserv_t rv;
 807
 808         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 809         if ((rv = TAILQ_FIRST(&vm_rvq_partpop)) != NULL) {
 810                 vm_reserv_reclaim(rv);
 811                 return (TRUE);
 812         }
 813         return (FALSE);
 814 }
 815
 816 /*
 817  * Searches the partially-populated reservation queue for the least recently
 818  * active reservation with unused pages, i.e., cached or free, that satisfy the
 819  * given request for contiguous physical memory.  If a satisfactory reservation
 820  * is found, it is broken.  Returns TRUE if a reservation is broken and FALSE
 821  * otherwise.
 822  *
 823  * The free page queue lock must be held.
 824  */
 825 boolean_t
 826 vm_reserv_reclaim_contig(u_long npages, vm_paddr_t low, vm_paddr_t high,
 827     u_long alignment, vm_paddr_t boundary)
 828 {
 829         vm_paddr_t pa, pa_length, size;
 830         vm_reserv_t rv;
 831         int i;
 832
 833         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 834         if (npages > VM_LEVEL_0_NPAGES - 1)
 835                 return (FALSE);
 836         size = npages << PAGE_SHIFT;
 837         TAILQ_FOREACH(rv, &vm_rvq_partpop, partpopq) {
 838                 pa = VM_PAGE_TO_PHYS(&rv->pages[VM_LEVEL_0_NPAGES - 1]);
 839                 if (pa + PAGE_SIZE - size < low) {
 840                         /* this entire reservation is too low; go to next */
 841                         continue;
 842                 }
 843                 pa_length = 0;
 844                 for (i = 0; i < VM_LEVEL_0_NPAGES; i++)
 845                         if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0) {
 846                                 pa_length += PAGE_SIZE;
 847                                 if (pa_length == PAGE_SIZE) {
 848                                         pa = VM_PAGE_TO_PHYS(&rv->pages[i]);
 849                                         if (pa + size > high) {
 850                                                 /* skip to next reservation */
 851                                                 break;
 852                                         } else if (pa < low ||
 853                                             (pa & (alignment - 1)) != 0 ||
 854                                             ((pa ^ (pa + size - 1)) &
 855                                             ~(boundary - 1)) != 0)
 856                                                 pa_length = 0;
 857                                 }
 858                                 if (pa_length >= size) {
 859                                         vm_reserv_reclaim(rv);
 860                                         return (TRUE);
 861                                 }
 862                         } else
 863                                 pa_length = 0;
 864         }
 865         return (FALSE);
 866 }
 867
 868 /*
 869  * Transfers the reservation underlying the given page to a new object.
 870  *
 871  * The object must be locked.
 872  */
 873 void
 874 vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
 875     vm_pindex_t old_object_offset)
 876 {
 877         vm_reserv_t rv;
 878
 879         VM_OBJECT_ASSERT_WLOCKED(new_object);
 880         rv = vm_reserv_from_page(m);
 881         if (rv->object == old_object) {
 882                 mtx_lock(&vm_page_queue_free_mtx);
 883                 if (rv->object == old_object) {
 884                         LIST_REMOVE(rv, objq);
 885                         LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
 886                         rv->object = new_object;
 887                         rv->pindex -= old_object_offset;
 888                 }
 889                 mtx_unlock(&vm_page_queue_free_mtx);
 890         }
 891 }
 892
 893 /*
 894  * Allocates the virtual and physical memory required by the reservation
 895  * management system's data structures, in particular, the reservation array.
 896  */
 897 vm_paddr_t
 898 vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t high_water)
 899 {
 900         vm_paddr_t new_end;
 901         size_t size;
 902
 903         /*
 904          * Calculate the size (in bytes) of the reservation array.  Round up
 905          * from "high_water" because every small page is mapped to an element
 906          * in the reservation array based on its physical address.  Thus, the
 907          * number of elements in the reservation array can be greater than the
 908          * number of superpages.
 909          */
 910         size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv);
 911
 912         /*
 913          * Allocate and map the physical memory for the reservation array.  The
 914          * next available virtual address is returned by reference.
 915          */
 916         new_end = end - round_page(size);
 917         vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
 918             VM_PROT_READ | VM_PROT_WRITE);
 919         bzero(vm_reserv_array, size);
 920
 921         /*
 922          * Return the next available physical address.
 923          */
 924         return (new_end);
 925 }
 926
 927 #endif  /* VM_NRESERVLEVEL > 0 */