sys/vm/vm_reserv.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2002-2006 Rice University
   5  * Copyright (c) 2007-2011 Alan L. Cox <alc@cs.rice.edu>
   6  * All rights reserved.
   7  *
   8  * This software was developed for the FreeBSD Project by Alan L. Cox,
   9  * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro.
  10  *
  11  * Redistribution and use in source and binary forms, with or without
  12  * modification, are permitted provided that the following conditions
  13  * are met:
  14  * 1. Redistributions of source code must retain the above copyright
  15  *    notice, this list of conditions and the following disclaimer.
  16  * 2. Redistributions in binary form must reproduce the above copyright
  17  *    notice, this list of conditions and the following disclaimer in the
  18  *    documentation and/or other materials provided with the distribution.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  23  * A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT
  24  * HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  25  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
  26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
  27  * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY
  30  * WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
  31  * POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 /*
  35  *      Superpage reservation management module
  36  *
  37  * Any external functions defined by this module are only to be used by the
  38  * virtual memory system.
  39  */
  40
  41 #include <sys/cdefs.h>
  42 __FBSDID("$FreeBSD$");
  43
  44 #include "opt_vm.h"
  45
  46 #include <sys/param.h>
  47 #include <sys/kernel.h>
  48 #include <sys/lock.h>
  49 #include <sys/malloc.h>
  50 #include <sys/mutex.h>
  51 #include <sys/queue.h>
  52 #include <sys/rwlock.h>
  53 #include <sys/sbuf.h>
  54 #include <sys/sysctl.h>
  55 #include <sys/systm.h>
  56 #include <sys/counter.h>
  57 #include <sys/ktr.h>
  58 #include <sys/vmmeter.h>
  59 #include <sys/smp.h>
  60
  61 #include <vm/vm.h>
  62 #include <vm/vm_param.h>
  63 #include <vm/vm_object.h>
  64 #include <vm/vm_page.h>
  65 #include <vm/vm_pageout.h>
  66 #include <vm/vm_phys.h>
  67 #include <vm/vm_pagequeue.h>
  68 #include <vm/vm_radix.h>
  69 #include <vm/vm_reserv.h>
  70
  71 /*
  72  * The reservation system supports the speculative allocation of large physical
  73  * pages ("superpages").  Speculative allocation enables the fully automatic
  74  * utilization of superpages by the virtual memory system.  In other words, no
  75  * programmatic directives are required to use superpages.
  76  */
  77
  78 #if VM_NRESERVLEVEL > 0
  79
  80 /*
  81  * The number of small pages that are contained in a level 0 reservation
  82  */
  83 #define VM_LEVEL_0_NPAGES       (1 << VM_LEVEL_0_ORDER)
  84
  85 /*
  86  * The number of bits by which a physical address is shifted to obtain the
  87  * reservation number
  88  */
  89 #define VM_LEVEL_0_SHIFT        (VM_LEVEL_0_ORDER + PAGE_SHIFT)
  90
  91 /*
  92  * The size of a level 0 reservation in bytes
  93  */
  94 #define VM_LEVEL_0_SIZE         (1 << VM_LEVEL_0_SHIFT)
  95
  96 /*
  97  * Computes the index of the small page underlying the given (object, pindex)
  98  * within the reservation's array of small pages.
  99  */
 100 #define VM_RESERV_INDEX(object, pindex) \
 101     (((object)->pg_color + (pindex)) & (VM_LEVEL_0_NPAGES - 1))
 102
 103 /*
 104  * The size of a population map entry
 105  */
 106 typedef u_long          popmap_t;
 107
 108 /*
 109  * The number of bits in a population map entry
 110  */
 111 #define NBPOPMAP        (NBBY * sizeof(popmap_t))
 112
 113 /*
 114  * The number of population map entries in a reservation
 115  */
 116 #define NPOPMAP         howmany(VM_LEVEL_0_NPAGES, NBPOPMAP)
 117
 118 /*
 119  * Clear a bit in the population map.
 120  */
 121 static __inline void
 122 popmap_clear(popmap_t popmap[], int i)
 123 {
 124
 125         popmap[i / NBPOPMAP] &= ~(1UL << (i % NBPOPMAP));
 126 }
 127
 128 /*
 129  * Set a bit in the population map.
 130  */
 131 static __inline void
 132 popmap_set(popmap_t popmap[], int i)
 133 {
 134
 135         popmap[i / NBPOPMAP] |= 1UL << (i % NBPOPMAP);
 136 }
 137
 138 /*
 139  * Is a bit in the population map clear?
 140  */
 141 static __inline boolean_t
 142 popmap_is_clear(popmap_t popmap[], int i)
 143 {
 144
 145         return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) == 0);
 146 }
 147
 148 /*
 149  * Is a bit in the population map set?
 150  */
 151 static __inline boolean_t
 152 popmap_is_set(popmap_t popmap[], int i)
 153 {
 154
 155         return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) != 0);
 156 }
 157
 158 /*
 159  * The reservation structure
 160  *
 161  * A reservation structure is constructed whenever a large physical page is
 162  * speculatively allocated to an object.  The reservation provides the small
 163  * physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets
 164  * within that object.  The reservation's "popcnt" tracks the number of these
 165  * small physical pages that are in use at any given time.  When and if the
 166  * reservation is not fully utilized, it appears in the queue of partially
 167  * populated reservations.  The reservation always appears on the containing
 168  * object's list of reservations.
 169  *
 170  * A partially populated reservation can be broken and reclaimed at any time.
 171  *
 172  * r - vm_reserv_lock
 173  * d - vm_reserv_domain_lock
 174  * o - vm_reserv_object_lock
 175  * c - constant after boot
 176  */
 177 struct vm_reserv {
 178         struct mtx      lock;                   /* reservation lock. */
 179         TAILQ_ENTRY(vm_reserv) partpopq;        /* (d) per-domain queue. */
 180         LIST_ENTRY(vm_reserv) objq;             /* (o, r) object queue */
 181         vm_object_t     object;                 /* (o, r) containing object */
 182         vm_pindex_t     pindex;                 /* (o, r) offset in object */
 183         vm_page_t       pages;                  /* (c) first page  */
 184         uint16_t        domain;                 /* (c) NUMA domain. */
 185         uint16_t        popcnt;                 /* (r) # of pages in use */
 186         char            inpartpopq;             /* (d) */
 187         popmap_t        popmap[NPOPMAP];        /* (r) bit vector, used pages */
 188 };
 189
 190 #define vm_reserv_lockptr(rv)           (&(rv)->lock)
 191 #define vm_reserv_assert_locked(rv)                                     \
 192             mtx_assert(vm_reserv_lockptr(rv), MA_OWNED)
 193 #define vm_reserv_lock(rv)              mtx_lock(vm_reserv_lockptr(rv))
 194 #define vm_reserv_trylock(rv)           mtx_trylock(vm_reserv_lockptr(rv))
 195 #define vm_reserv_unlock(rv)            mtx_unlock(vm_reserv_lockptr(rv))
 196
 197 static struct mtx_padalign vm_reserv_domain_locks[MAXMEMDOM];
 198
 199 #define vm_reserv_domain_lockptr(d)     &vm_reserv_domain_locks[(d)]
 200 #define vm_reserv_domain_lock(d)        mtx_lock(vm_reserv_domain_lockptr(d))
 201 #define vm_reserv_domain_unlock(d)      mtx_unlock(vm_reserv_domain_lockptr(d))
 202
 203 /*
 204  * The reservation array
 205  *
 206  * This array is analoguous in function to vm_page_array.  It differs in the
 207  * respect that it may contain a greater number of useful reservation
 208  * structures than there are (physical) superpages.  These "invalid"
 209  * reservation structures exist to trade-off space for time in the
 210  * implementation of vm_reserv_from_page().  Invalid reservation structures are
 211  * distinguishable from "valid" reservation structures by inspecting the
 212  * reservation's "pages" field.  Invalid reservation structures have a NULL
 213  * "pages" field.
 214  *
 215  * vm_reserv_from_page() maps a small (physical) page to an element of this
 216  * array by computing a physical reservation number from the page's physical
 217  * address.  The physical reservation number is used as the array index.
 218  *
 219  * An "active" reservation is a valid reservation structure that has a non-NULL
 220  * "object" field and a non-zero "popcnt" field.  In other words, every active
 221  * reservation belongs to a particular object.  Moreover, every active
 222  * reservation has an entry in the containing object's list of reservations.
 223  */
 224 static vm_reserv_t vm_reserv_array;
 225
 226 /*
 227  * The partially populated reservation queue
 228  *
 229  * This queue enables the fast recovery of an unused free small page from a
 230  * partially populated reservation.  The reservation at the head of this queue
 231  * is the least recently changed, partially populated reservation.
 232  *
 233  * Access to this queue is synchronized by the free page queue lock.
 234  */
 235 static TAILQ_HEAD(, vm_reserv) vm_rvq_partpop[MAXMEMDOM];
 236
 237 static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD, 0, "Reservation Info");
 238
 239 static counter_u64_t vm_reserv_broken = EARLY_COUNTER;
 240 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD,
 241     &vm_reserv_broken, "Cumulative number of broken reservations");
 242
 243 static counter_u64_t vm_reserv_freed = EARLY_COUNTER;
 244 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD,
 245     &vm_reserv_freed, "Cumulative number of freed reservations");
 246
 247 static int sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS);
 248
 249 SYSCTL_PROC(_vm_reserv, OID_AUTO, fullpop, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
 250     sysctl_vm_reserv_fullpop, "I", "Current number of full reservations");
 251
 252 static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);
 253
 254 SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
 255     sysctl_vm_reserv_partpopq, "A", "Partially populated reservation queues");
 256
 257 static counter_u64_t vm_reserv_reclaimed = EARLY_COUNTER;
 258 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
 259     &vm_reserv_reclaimed, "Cumulative number of reclaimed reservations");
 260
 261 /*
 262  * The object lock pool is used to synchronize the rvq.  We can not use a
 263  * pool mutex because it is required before malloc works.
 264  *
 265  * The "hash" function could be made faster without divide and modulo.
 266  */
 267 #define VM_RESERV_OBJ_LOCK_COUNT        MAXCPU
 268
 269 struct mtx_padalign vm_reserv_object_mtx[VM_RESERV_OBJ_LOCK_COUNT];
 270
 271 #define vm_reserv_object_lock_idx(object)                       \
 272             (((uintptr_t)object / sizeof(*object)) % VM_RESERV_OBJ_LOCK_COUNT)
 273 #define vm_reserv_object_lock_ptr(object)                       \
 274             &vm_reserv_object_mtx[vm_reserv_object_lock_idx((object))]
 275 #define vm_reserv_object_lock(object)                           \
 276             mtx_lock(vm_reserv_object_lock_ptr((object)))
 277 #define vm_reserv_object_unlock(object)                         \
 278             mtx_unlock(vm_reserv_object_lock_ptr((object)))
 279
 280 static void             vm_reserv_break(vm_reserv_t rv);
 281 static void             vm_reserv_depopulate(vm_reserv_t rv, int index);
 282 static vm_reserv_t      vm_reserv_from_page(vm_page_t m);
 283 static boolean_t        vm_reserv_has_pindex(vm_reserv_t rv,
 284                             vm_pindex_t pindex);
 285 static void             vm_reserv_populate(vm_reserv_t rv, int index);
 286 static void             vm_reserv_reclaim(vm_reserv_t rv);
 287
 288 /*
 289  * Returns the current number of full reservations.
 290  *
 291  * Since the number of full reservations is computed without acquiring the
 292  * free page queue lock, the returned value may be inexact.
 293  */
 294 static int
 295 sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS)
 296 {
 297         vm_paddr_t paddr;
 298         struct vm_phys_seg *seg;
 299         vm_reserv_t rv;
 300         int fullpop, segind;
 301
 302         fullpop = 0;
 303         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 304                 seg = &vm_phys_segs[segind];
 305                 paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
 306                 while (paddr + VM_LEVEL_0_SIZE <= seg->end) {
 307                         rv = &vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT];
 308                         fullpop += rv->popcnt == VM_LEVEL_0_NPAGES;
 309                         paddr += VM_LEVEL_0_SIZE;
 310                 }
 311         }
 312         return (sysctl_handle_int(oidp, &fullpop, 0, req));
 313 }
 314
 315 /*
 316  * Describes the current state of the partially populated reservation queue.
 317  */
 318 static int
 319 sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
 320 {
 321         struct sbuf sbuf;
 322         vm_reserv_t rv;
 323         int counter, error, domain, level, unused_pages;
 324
 325         error = sysctl_wire_old_buffer(req, 0);
 326         if (error != 0)
 327                 return (error);
 328         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 329         sbuf_printf(&sbuf, "\nDOMAIN    LEVEL     SIZE  NUMBER\n\n");
 330         for (domain = 0; domain < vm_ndomains; domain++) {
 331                 for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
 332                         counter = 0;
 333                         unused_pages = 0;
 334                         vm_reserv_domain_lock(domain);
 335                         TAILQ_FOREACH(rv, &vm_rvq_partpop[domain], partpopq) {
 336                                 counter++;
 337                                 unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
 338                         }
 339                         vm_reserv_domain_unlock(domain);
 340                         sbuf_printf(&sbuf, "%6d, %7d, %6dK, %6d\n",
 341                             domain, level,
 342                             unused_pages * ((int)PAGE_SIZE / 1024), counter);
 343                 }
 344         }
 345         error = sbuf_finish(&sbuf);
 346         sbuf_delete(&sbuf);
 347         return (error);
 348 }
 349
 350 /*
 351  * Remove a reservation from the object's objq.
 352  */
 353 static void
 354 vm_reserv_remove(vm_reserv_t rv)
 355 {
 356         vm_object_t object;
 357
 358         vm_reserv_assert_locked(rv);
 359         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
 360             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
 361         KASSERT(rv->object != NULL,
 362             ("vm_reserv_remove: reserv %p is free", rv));
 363         KASSERT(!rv->inpartpopq,
 364             ("vm_reserv_remove: reserv %p's inpartpopq is TRUE", rv));
 365         object = rv->object;
 366         vm_reserv_object_lock(object);
 367         LIST_REMOVE(rv, objq);
 368         rv->object = NULL;
 369         vm_reserv_object_unlock(object);
 370 }
 371
 372 /*
 373  * Insert a new reservation into the object's objq.
 374  */
 375 static void
 376 vm_reserv_insert(vm_reserv_t rv, vm_object_t object, vm_pindex_t pindex)
 377 {
 378         int i;
 379
 380         vm_reserv_assert_locked(rv);
 381         CTR6(KTR_VM,
 382             "%s: rv %p(%p) object %p new %p popcnt %d",
 383             __FUNCTION__, rv, rv->pages, rv->object, object,
 384            rv->popcnt);
 385         KASSERT(rv->object == NULL,
 386             ("vm_reserv_insert: reserv %p isn't free", rv));
 387         KASSERT(rv->popcnt == 0,
 388             ("vm_reserv_insert: reserv %p's popcnt is corrupted", rv));
 389         KASSERT(!rv->inpartpopq,
 390             ("vm_reserv_insert: reserv %p's inpartpopq is TRUE", rv));
 391         for (i = 0; i < NPOPMAP; i++)
 392                 KASSERT(rv->popmap[i] == 0,
 393                     ("vm_reserv_insert: reserv %p's popmap is corrupted", rv));
 394         vm_reserv_object_lock(object);
 395         rv->pindex = pindex;
 396         rv->object = object;
 397         LIST_INSERT_HEAD(&object->rvq, rv, objq);
 398         vm_reserv_object_unlock(object);
 399 }
 400
 401 /*
 402  * Reduces the given reservation's population count.  If the population count
 403  * becomes zero, the reservation is destroyed.  Additionally, moves the
 404  * reservation to the tail of the partially populated reservation queue if the
 405  * population count is non-zero.
 406  */
 407 static void
 408 vm_reserv_depopulate(vm_reserv_t rv, int index)
 409 {
 410         struct vm_domain *vmd;
 411
 412         vm_reserv_assert_locked(rv);
 413         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
 414             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
 415         KASSERT(rv->object != NULL,
 416             ("vm_reserv_depopulate: reserv %p is free", rv));
 417         KASSERT(popmap_is_set(rv->popmap, index),
 418             ("vm_reserv_depopulate: reserv %p's popmap[%d] is clear", rv,
 419             index));
 420         KASSERT(rv->popcnt > 0,
 421             ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
 422         KASSERT(rv->domain < vm_ndomains,
 423             ("vm_reserv_depopulate: reserv %p's domain is corrupted %d",
 424             rv, rv->domain));
 425         if (rv->popcnt == VM_LEVEL_0_NPAGES) {
 426                 KASSERT(rv->pages->psind == 1,
 427                     ("vm_reserv_depopulate: reserv %p is already demoted",
 428                     rv));
 429                 rv->pages->psind = 0;
 430         }
 431         popmap_clear(rv->popmap, index);
 432         rv->popcnt--;
 433         vm_reserv_domain_lock(rv->domain);
 434         if (rv->inpartpopq) {
 435                 TAILQ_REMOVE(&vm_rvq_partpop[rv->domain], rv, partpopq);
 436                 rv->inpartpopq = FALSE;
 437         }
 438         if (rv->popcnt != 0) {
 439                 rv->inpartpopq = TRUE;
 440                 TAILQ_INSERT_TAIL(&vm_rvq_partpop[rv->domain], rv, partpopq);
 441         }
 442         vm_reserv_domain_unlock(rv->domain);
 443         vmd = VM_DOMAIN(rv->domain);
 444         if (rv->popcnt == 0) {
 445                 vm_reserv_remove(rv);
 446                 vm_domain_free_lock(vmd);
 447                 vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
 448                 vm_domain_free_unlock(vmd);
 449                 counter_u64_add(vm_reserv_freed, 1);
 450         }
 451         vm_domain_freecnt_inc(vmd, 1);
 452 }
 453
 454 /*
 455  * Returns the reservation to which the given page might belong.
 456  */
 457 static __inline vm_reserv_t
 458 vm_reserv_from_page(vm_page_t m)
 459 {
 460
 461         return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
 462 }
 463
 464 /*
 465  * Returns an existing reservation or NULL and initialized successor pointer.
 466  */
 467 static vm_reserv_t
 468 vm_reserv_from_object(vm_object_t object, vm_pindex_t pindex,
 469     vm_page_t mpred, vm_page_t *msuccp)
 470 {
 471         vm_reserv_t rv;
 472         vm_page_t msucc;
 473
 474         msucc = NULL;
 475         if (mpred != NULL) {
 476                 KASSERT(mpred->object == object,
 477                     ("vm_reserv_from_object: object doesn't contain mpred"));
 478                 KASSERT(mpred->pindex < pindex,
 479                     ("vm_reserv_from_object: mpred doesn't precede pindex"));
 480                 rv = vm_reserv_from_page(mpred);
 481                 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
 482                         goto found;
 483                 msucc = TAILQ_NEXT(mpred, listq);
 484         } else
 485                 msucc = TAILQ_FIRST(&object->memq);
 486         if (msucc != NULL) {
 487                 KASSERT(msucc->pindex > pindex,
 488                     ("vm_reserv_from_object: msucc doesn't succeed pindex"));
 489                 rv = vm_reserv_from_page(msucc);
 490                 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
 491                         goto found;
 492         }
 493         rv = NULL;
 494
 495 found:
 496         *msuccp = msucc;
 497
 498         return (rv);
 499 }
 500
 501 /*
 502  * Returns TRUE if the given reservation contains the given page index and
 503  * FALSE otherwise.
 504  */
 505 static __inline boolean_t
 506 vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
 507 {
 508
 509         return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
 510 }
 511
 512 /*
 513  * Increases the given reservation's population count.  Moves the reservation
 514  * to the tail of the partially populated reservation queue.
 515  *
 516  * The free page queue must be locked.
 517  */
 518 static void
 519 vm_reserv_populate(vm_reserv_t rv, int index)
 520 {
 521
 522         vm_reserv_assert_locked(rv);
 523         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
 524             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
 525         KASSERT(rv->object != NULL,
 526             ("vm_reserv_populate: reserv %p is free", rv));
 527         KASSERT(popmap_is_clear(rv->popmap, index),
 528             ("vm_reserv_populate: reserv %p's popmap[%d] is set", rv,
 529             index));
 530         KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
 531             ("vm_reserv_populate: reserv %p is already full", rv));
 532         KASSERT(rv->pages->psind == 0,
 533             ("vm_reserv_populate: reserv %p is already promoted", rv));
 534         KASSERT(rv->domain < vm_ndomains,
 535             ("vm_reserv_populate: reserv %p's domain is corrupted %d",
 536             rv, rv->domain));
 537         popmap_set(rv->popmap, index);
 538         rv->popcnt++;
 539         vm_reserv_domain_lock(rv->domain);
 540         if (rv->inpartpopq) {
 541                 TAILQ_REMOVE(&vm_rvq_partpop[rv->domain], rv, partpopq);
 542                 rv->inpartpopq = FALSE;
 543         }
 544         if (rv->popcnt < VM_LEVEL_0_NPAGES) {
 545                 rv->inpartpopq = TRUE;
 546                 TAILQ_INSERT_TAIL(&vm_rvq_partpop[rv->domain], rv, partpopq);
 547         } else {
 548                 KASSERT(rv->pages->psind == 0,
 549                     ("vm_reserv_populate: reserv %p is already promoted",
 550                     rv));
 551                 rv->pages->psind = 1;
 552         }
 553         vm_reserv_domain_unlock(rv->domain);
 554 }
 555
 556 /*
 557  * Attempts to allocate a contiguous set of physical pages from existing
 558  * reservations.  See vm_reserv_alloc_contig() for a description of the
 559  * function's parameters.
 560  *
 561  * The page "mpred" must immediately precede the offset "pindex" within the
 562  * specified object.
 563  *
 564  * The object must be locked.
 565  */
 566 vm_page_t
 567 vm_reserv_extend_contig(int req, vm_object_t object, vm_pindex_t pindex,
 568     int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
 569     u_long alignment, vm_paddr_t boundary, vm_page_t mpred)
 570 {
 571         struct vm_domain *vmd;
 572         vm_paddr_t pa, size;
 573         vm_page_t m, msucc;
 574         vm_reserv_t rv;
 575         int i, index;
 576
 577         VM_OBJECT_ASSERT_WLOCKED(object);
 578         KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
 579
 580         /*
 581          * Is a reservation fundamentally impossible?
 582          */
 583         if (pindex < VM_RESERV_INDEX(object, pindex) ||
 584             pindex + npages > object->size || object->resident_page_count == 0)
 585                 return (NULL);
 586
 587         /*
 588          * All reservations of a particular size have the same alignment.
 589          * Assuming that the first page is allocated from a reservation, the
 590          * least significant bits of its physical address can be determined
 591          * from its offset from the beginning of the reservation and the size
 592          * of the reservation.
 593          *
 594          * Could the specified index within a reservation of the smallest
 595          * possible size satisfy the alignment and boundary requirements?
 596          */
 597         pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
 598         if ((pa & (alignment - 1)) != 0)
 599                 return (NULL);
 600         size = npages << PAGE_SHIFT;
 601         if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
 602                 return (NULL);
 603
 604         /*
 605          * Look for an existing reservation.
 606          */
 607         rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
 608         if (rv == NULL)
 609                 return (NULL);
 610         KASSERT(object != kernel_object || rv->domain == domain,
 611             ("vm_reserv_extend_contig: Domain mismatch from reservation."));
 612         index = VM_RESERV_INDEX(object, pindex);
 613         /* Does the allocation fit within the reservation? */
 614         if (index + npages > VM_LEVEL_0_NPAGES)
 615                 return (NULL);
 616         domain = rv->domain;
 617         vmd = VM_DOMAIN(domain);
 618         vm_reserv_lock(rv);
 619         if (rv->object != object)
 620                 goto out;
 621         m = &rv->pages[index];
 622         pa = VM_PAGE_TO_PHYS(m);
 623         if (pa < low || pa + size > high || (pa & (alignment - 1)) != 0 ||
 624             ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
 625                 goto out;
 626         /* Handle vm_page_rename(m, new_object, ...). */
 627         for (i = 0; i < npages; i++) {
 628                 if (popmap_is_set(rv->popmap, index + i))
 629                         goto out;
 630         }
 631         if (!vm_domain_allocate(vmd, req, npages))
 632                 goto out;
 633         for (i = 0; i < npages; i++)
 634                 vm_reserv_populate(rv, index + i);
 635         vm_reserv_unlock(rv);
 636         return (m);
 637
 638 out:
 639         vm_reserv_unlock(rv);
 640         return (NULL);
 641 }
 642
 643 /*
 644  * Allocates a contiguous set of physical pages of the given size "npages"
 645  * from newly created reservations.  All of the physical pages
 646  * must be at or above the given physical address "low" and below the given
 647  * physical address "high".  The given value "alignment" determines the
 648  * alignment of the first physical page in the set.  If the given value
 649  * "boundary" is non-zero, then the set of physical pages cannot cross any
 650  * physical address boundary that is a multiple of that value.  Both
 651  * "alignment" and "boundary" must be a power of two.
 652  *
 653  * Callers should first invoke vm_reserv_extend_contig() to attempt an
 654  * allocation from existing reservations.
 655  *
 656  * The page "mpred" must immediately precede the offset "pindex" within the
 657  * specified object.
 658  *
 659  * The object and free page queue must be locked.
 660  */
 661 vm_page_t
 662 vm_reserv_alloc_contig(int req, vm_object_t object, vm_pindex_t pindex, int domain,
 663     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
 664     vm_paddr_t boundary, vm_page_t mpred)
 665 {
 666         struct vm_domain *vmd;
 667         vm_paddr_t pa, size;
 668         vm_page_t m, m_ret, msucc;
 669         vm_pindex_t first, leftcap, rightcap;
 670         vm_reserv_t rv;
 671         u_long allocpages, maxpages, minpages;
 672         int i, index, n;
 673
 674         VM_OBJECT_ASSERT_WLOCKED(object);
 675         KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
 676
 677         /*
 678          * Is a reservation fundamentally impossible?
 679          */
 680         if (pindex < VM_RESERV_INDEX(object, pindex) ||
 681             pindex + npages > object->size)
 682                 return (NULL);
 683
 684         /*
 685          * All reservations of a particular size have the same alignment.
 686          * Assuming that the first page is allocated from a reservation, the
 687          * least significant bits of its physical address can be determined
 688          * from its offset from the beginning of the reservation and the size
 689          * of the reservation.
 690          *
 691          * Could the specified index within a reservation of the smallest
 692          * possible size satisfy the alignment and boundary requirements?
 693          */
 694         pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
 695         if ((pa & (alignment - 1)) != 0)
 696                 return (NULL);
 697         size = npages << PAGE_SHIFT;
 698         if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
 699                 return (NULL);
 700
 701         /*
 702          * Callers should've extended an existing reservation prior to
 703          * calling this function.  If a reservation exists it is
 704          * incompatible with the allocation.
 705          */
 706         rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
 707         if (rv != NULL)
 708                 return (NULL);
 709
 710         /*
 711          * Could at least one reservation fit between the first index to the
 712          * left that can be used ("leftcap") and the first index to the right
 713          * that cannot be used ("rightcap")?
 714          *
 715          * We must synchronize with the reserv object lock to protect the
 716          * pindex/object of the resulting reservations against rename while
 717          * we are inspecting.
 718          */
 719         first = pindex - VM_RESERV_INDEX(object, pindex);
 720         minpages = VM_RESERV_INDEX(object, pindex) + npages;
 721         maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES);
 722         allocpages = maxpages;
 723         vm_reserv_object_lock(object);
 724         if (mpred != NULL) {
 725                 if ((rv = vm_reserv_from_page(mpred))->object != object)
 726                         leftcap = mpred->pindex + 1;
 727                 else
 728                         leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
 729                 if (leftcap > first) {
 730                         vm_reserv_object_unlock(object);
 731                         return (NULL);
 732                 }
 733         }
 734         if (msucc != NULL) {
 735                 if ((rv = vm_reserv_from_page(msucc))->object != object)
 736                         rightcap = msucc->pindex;
 737                 else
 738                         rightcap = rv->pindex;
 739                 if (first + maxpages > rightcap) {
 740                         if (maxpages == VM_LEVEL_0_NPAGES) {
 741                                 vm_reserv_object_unlock(object);
 742                                 return (NULL);
 743                         }
 744
 745                         /*
 746                          * At least one reservation will fit between "leftcap"
 747                          * and "rightcap".  However, a reservation for the
 748                          * last of the requested pages will not fit.  Reduce
 749                          * the size of the upcoming allocation accordingly.
 750                          */
 751                         allocpages = minpages;
 752                 }
 753         }
 754         vm_reserv_object_unlock(object);
 755
 756         /*
 757          * Would the last new reservation extend past the end of the object?
 758          */
 759         if (first + maxpages > object->size) {
 760                 /*
 761                  * Don't allocate the last new reservation if the object is a
 762                  * vnode or backed by another object that is a vnode.
 763                  */
 764                 if (object->type == OBJT_VNODE ||
 765                     (object->backing_object != NULL &&
 766                     object->backing_object->type == OBJT_VNODE)) {
 767                         if (maxpages == VM_LEVEL_0_NPAGES)
 768                                 return (NULL);
 769                         allocpages = minpages;
 770                 }
 771                 /* Speculate that the object may grow. */
 772         }
 773
 774         /*
 775          * Allocate the physical pages.  The alignment and boundary specified
 776          * for this allocation may be different from the alignment and
 777          * boundary specified for the requested pages.  For instance, the
 778          * specified index may not be the first page within the first new
 779          * reservation.
 780          */
 781         m = NULL;
 782         vmd = VM_DOMAIN(domain);
 783         if (vm_domain_allocate(vmd, req, npages)) {
 784                 vm_domain_free_lock(vmd);
 785                 m = vm_phys_alloc_contig(domain, allocpages, low, high,
 786                     ulmax(alignment, VM_LEVEL_0_SIZE),
 787                     boundary > VM_LEVEL_0_SIZE ? boundary : 0);
 788                 vm_domain_free_unlock(vmd);
 789                 if (m == NULL) {
 790                         vm_domain_freecnt_inc(vmd, npages);
 791                         return (NULL);
 792                 }
 793         } else
 794                 return (NULL);
 795         KASSERT(vm_phys_domain(m) == domain,
 796             ("vm_reserv_alloc_contig: Page domain does not match requested."));
 797
 798         /*
 799          * The allocated physical pages always begin at a reservation
 800          * boundary, but they do not always end at a reservation boundary.
 801          * Initialize every reservation that is completely covered by the
 802          * allocated physical pages.
 803          */
 804         m_ret = NULL;
 805         index = VM_RESERV_INDEX(object, pindex);
 806         do {
 807                 rv = vm_reserv_from_page(m);
 808                 KASSERT(rv->pages == m,
 809                     ("vm_reserv_alloc_contig: reserv %p's pages is corrupted",
 810                     rv));
 811                 vm_reserv_lock(rv);
 812                 vm_reserv_insert(rv, object, first);
 813                 n = ulmin(VM_LEVEL_0_NPAGES - index, npages);
 814                 for (i = 0; i < n; i++)
 815                         vm_reserv_populate(rv, index + i);
 816                 npages -= n;
 817                 if (m_ret == NULL) {
 818                         m_ret = &rv->pages[index];
 819                         index = 0;
 820                 }
 821                 vm_reserv_unlock(rv);
 822                 m += VM_LEVEL_0_NPAGES;
 823                 first += VM_LEVEL_0_NPAGES;
 824                 allocpages -= VM_LEVEL_0_NPAGES;
 825         } while (allocpages >= VM_LEVEL_0_NPAGES);
 826         return (m_ret);
 827 }
 828
 829 /*
 830  * Attempts to extend an existing reservation and allocate the page to the
 831  * object.
 832  *
 833  * The page "mpred" must immediately precede the offset "pindex" within the
 834  * specified object.
 835  *
 836  * The object must be locked.
 837  */
 838 vm_page_t
 839 vm_reserv_extend(int req, vm_object_t object, vm_pindex_t pindex, int domain,
 840     vm_page_t mpred)
 841 {
 842         struct vm_domain *vmd;
 843         vm_page_t m, msucc;
 844         vm_reserv_t rv;
 845         int index;
 846
 847         VM_OBJECT_ASSERT_WLOCKED(object);
 848
 849         /*
 850          * Could a reservation currently exist?
 851          */
 852         if (pindex < VM_RESERV_INDEX(object, pindex) ||
 853             pindex >= object->size || object->resident_page_count == 0)
 854                 return (NULL);
 855
 856         /*
 857          * Look for an existing reservation.
 858          */
 859         rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
 860         if (rv == NULL)
 861                 return (NULL);
 862
 863         KASSERT(object != kernel_object || rv->domain == domain,
 864             ("vm_reserv_extend: Domain mismatch from reservation."));
 865         domain = rv->domain;
 866         vmd = VM_DOMAIN(domain);
 867         index = VM_RESERV_INDEX(object, pindex);
 868         m = &rv->pages[index];
 869         vm_reserv_lock(rv);
 870         /* Handle reclaim race. */
 871         if (rv->object != object ||
 872             /* Handle vm_page_rename(m, new_object, ...). */
 873             popmap_is_set(rv->popmap, index)) {
 874                 m = NULL;
 875                 goto out;
 876         }
 877         if (vm_domain_allocate(vmd, req, 1) == 0)
 878                 m = NULL;
 879         else
 880                 vm_reserv_populate(rv, index);
 881 out:
 882         vm_reserv_unlock(rv);
 883
 884         return (m);
 885 }
 886
 887 /*
 888  * Attempts to allocate a new reservation for the object, and allocates a
 889  * page from that reservation.  Callers should first invoke vm_reserv_extend()
 890  * to attempt an allocation from an existing reservation.
 891  *
 892  * The page "mpred" must immediately precede the offset "pindex" within the
 893  * specified object.
 894  *
 895  * The object and free page queue must be locked.
 896  */
 897 vm_page_t
 898 vm_reserv_alloc_page(int req, vm_object_t object, vm_pindex_t pindex, int domain,
 899     vm_page_t mpred)
 900 {
 901         struct vm_domain *vmd;
 902         vm_page_t m, msucc;
 903         vm_pindex_t first, leftcap, rightcap;
 904         vm_reserv_t rv;
 905         int index;
 906
 907         VM_OBJECT_ASSERT_WLOCKED(object);
 908
 909         /*
 910          * Is a reservation fundamentally impossible?
 911          */
 912         if (pindex < VM_RESERV_INDEX(object, pindex) ||
 913             pindex >= object->size)
 914                 return (NULL);
 915
 916         /*
 917          * Callers should've extended an existing reservation prior to
 918          * calling this function.  If a reservation exists it is
 919          * incompatible with the allocation.
 920          */
 921         rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
 922         if (rv != NULL)
 923                 return (NULL);
 924
 925         /*
 926          * Could a reservation fit between the first index to the left that
 927          * can be used and the first index to the right that cannot be used?
 928          *
 929          * We must synchronize with the reserv object lock to protect the
 930          * pindex/object of the resulting reservations against rename while
 931          * we are inspecting.
 932          */
 933         first = pindex - VM_RESERV_INDEX(object, pindex);
 934         vm_reserv_object_lock(object);
 935         if (mpred != NULL) {
 936                 if ((rv = vm_reserv_from_page(mpred))->object != object)
 937                         leftcap = mpred->pindex + 1;
 938                 else
 939                         leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
 940                 if (leftcap > first) {
 941                         vm_reserv_object_unlock(object);
 942                         return (NULL);
 943                 }
 944         }
 945         if (msucc != NULL) {
 946                 if ((rv = vm_reserv_from_page(msucc))->object != object)
 947                         rightcap = msucc->pindex;
 948                 else
 949                         rightcap = rv->pindex;
 950                 if (first + VM_LEVEL_0_NPAGES > rightcap) {
 951                         vm_reserv_object_unlock(object);
 952                         return (NULL);
 953                 }
 954         }
 955         vm_reserv_object_unlock(object);
 956
 957         /*
 958          * Would a new reservation extend past the end of the object?
 959          */
 960         if (first + VM_LEVEL_0_NPAGES > object->size) {
 961                 /*
 962                  * Don't allocate a new reservation if the object is a vnode or
 963                  * backed by another object that is a vnode.
 964                  */
 965                 if (object->type == OBJT_VNODE ||
 966                     (object->backing_object != NULL &&
 967                     object->backing_object->type == OBJT_VNODE))
 968                         return (NULL);
 969                 /* Speculate that the object may grow. */
 970         }
 971
 972         /*
 973          * Allocate and populate the new reservation.
 974          */
 975         m = NULL;
 976         vmd = VM_DOMAIN(domain);
 977         if (vm_domain_allocate(vmd, req, 1)) {
 978                 vm_domain_free_lock(vmd);
 979                 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT,
 980                     VM_LEVEL_0_ORDER);
 981                 vm_domain_free_unlock(vmd);
 982                 if (m == NULL) {
 983                         vm_domain_freecnt_inc(vmd, 1);
 984                         return (NULL);
 985                 }
 986         } else
 987                 return (NULL);
 988         rv = vm_reserv_from_page(m);
 989         vm_reserv_lock(rv);
 990         KASSERT(rv->pages == m,
 991             ("vm_reserv_alloc_page: reserv %p's pages is corrupted", rv));
 992         vm_reserv_insert(rv, object, first);
 993         index = VM_RESERV_INDEX(object, pindex);
 994         vm_reserv_populate(rv, index);
 995         vm_reserv_unlock(rv);
 996
 997         return (&rv->pages[index]);
 998 }
 999
1000 /*
1001  * Breaks the given reservation.  All free pages in the reservation
1002  * are returned to the physical memory allocator.  The reservation's
1003  * population count and map are reset to their initial state.
1004  *
1005  * The given reservation must not be in the partially populated reservation
1006  * queue.  The free page queue lock must be held.
1007  */
1008 static void
1009 vm_reserv_break(vm_reserv_t rv)
1010 {
1011         int begin_zeroes, hi, i, lo;
1012
1013         vm_reserv_assert_locked(rv);
1014         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
1015             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
1016         vm_reserv_remove(rv);
1017         rv->pages->psind = 0;
1018         i = hi = 0;
1019         do {
1020                 /* Find the next 0 bit.  Any previous 0 bits are < "hi". */
1021                 lo = ffsl(~(((1UL << hi) - 1) | rv->popmap[i]));
1022                 if (lo == 0) {
1023                         /* Redundantly clears bits < "hi". */
1024                         rv->popmap[i] = 0;
1025                         rv->popcnt -= NBPOPMAP - hi;
1026                         while (++i < NPOPMAP) {
1027                                 lo = ffsl(~rv->popmap[i]);
1028                                 if (lo == 0) {
1029                                         rv->popmap[i] = 0;
1030                                         rv->popcnt -= NBPOPMAP;
1031                                 } else
1032                                         break;
1033                         }
1034                         if (i == NPOPMAP)
1035                                 break;
1036                         hi = 0;
1037                 }
1038                 KASSERT(lo > 0, ("vm_reserv_break: lo is %d", lo));
1039                 /* Convert from ffsl() to ordinary bit numbering. */
1040                 lo--;
1041                 if (lo > 0) {
1042                         /* Redundantly clears bits < "hi". */
1043                         rv->popmap[i] &= ~((1UL << lo) - 1);
1044                         rv->popcnt -= lo - hi;
1045                 }
1046                 begin_zeroes = NBPOPMAP * i + lo;
1047                 /* Find the next 1 bit. */
1048                 do
1049                         hi = ffsl(rv->popmap[i]);
1050                 while (hi == 0 && ++i < NPOPMAP);
1051                 if (i != NPOPMAP)
1052                         /* Convert from ffsl() to ordinary bit numbering. */
1053                         hi--;
1054                 vm_domain_free_lock(VM_DOMAIN(rv->domain));
1055                 vm_phys_free_contig(&rv->pages[begin_zeroes], NBPOPMAP * i +
1056                     hi - begin_zeroes);
1057                 vm_domain_free_unlock(VM_DOMAIN(rv->domain));
1058         } while (i < NPOPMAP);
1059         KASSERT(rv->popcnt == 0,
1060             ("vm_reserv_break: reserv %p's popcnt is corrupted", rv));
1061         counter_u64_add(vm_reserv_broken, 1);
1062 }
1063
1064 /*
1065  * Breaks all reservations belonging to the given object.
1066  */
1067 void
1068 vm_reserv_break_all(vm_object_t object)
1069 {
1070         vm_reserv_t rv;
1071
1072         /*
1073          * This access of object->rvq is unsynchronized so that the
1074          * object rvq lock can nest after the domain_free lock.  We
1075          * must check for races in the results.  However, the object
1076          * lock prevents new additions, so we are guaranteed that when
1077          * it returns NULL the object is properly empty.
1078          */
1079         while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
1080                 vm_reserv_lock(rv);
1081                 /* Reclaim race. */
1082                 if (rv->object != object) {
1083                         vm_reserv_unlock(rv);
1084                         continue;
1085                 }
1086                 vm_reserv_domain_lock(rv->domain);
1087                 if (rv->inpartpopq) {
1088                         TAILQ_REMOVE(&vm_rvq_partpop[rv->domain], rv, partpopq);
1089                         rv->inpartpopq = FALSE;
1090                 }
1091                 vm_reserv_domain_unlock(rv->domain);
1092                 vm_reserv_break(rv);
1093                 vm_reserv_unlock(rv);
1094         }
1095 }
1096
1097 /*
1098  * Frees the given page if it belongs to a reservation.  Returns TRUE if the
1099  * page is freed and FALSE otherwise.
1100  *
1101  * The free page queue lock must be held.
1102  */
1103 boolean_t
1104 vm_reserv_free_page(vm_page_t m)
1105 {
1106         vm_reserv_t rv;
1107         boolean_t ret;
1108
1109         rv = vm_reserv_from_page(m);
1110         if (rv->object == NULL)
1111                 return (FALSE);
1112         vm_reserv_lock(rv);
1113         /* Re-validate after lock. */
1114         if (rv->object != NULL) {
1115                 vm_reserv_depopulate(rv, m - rv->pages);
1116                 ret = TRUE;
1117         } else
1118                 ret = FALSE;
1119         vm_reserv_unlock(rv);
1120
1121         return (ret);
1122 }
1123
1124 /*
1125  * Initializes the reservation management system.  Specifically, initializes
1126  * the reservation array.
1127  *
1128  * Requires that vm_page_array and first_page are initialized!
1129  */
1130 void
1131 vm_reserv_init(void)
1132 {
1133         vm_paddr_t paddr;
1134         struct vm_phys_seg *seg;
1135         struct vm_reserv *rv;
1136         int i, segind;
1137
1138         /*
1139          * Initialize the reservation array.  Specifically, initialize the
1140          * "pages" field for every element that has an underlying superpage.
1141          */
1142         for (segind = 0; segind < vm_phys_nsegs; segind++) {
1143                 seg = &vm_phys_segs[segind];
1144                 paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
1145                 while (paddr + VM_LEVEL_0_SIZE <= seg->end) {
1146                         rv = &vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT];
1147                         rv->pages = PHYS_TO_VM_PAGE(paddr);
1148                         rv->domain = seg->domain;
1149                         mtx_init(&rv->lock, "vm reserv", NULL, MTX_DEF);
1150                         paddr += VM_LEVEL_0_SIZE;
1151                 }
1152         }
1153         for (i = 0; i < MAXMEMDOM; i++) {
1154                 mtx_init(&vm_reserv_domain_locks[i], "VM reserv domain", NULL,
1155                     MTX_DEF);
1156                 TAILQ_INIT(&vm_rvq_partpop[i]);
1157         }
1158
1159         for (i = 0; i < VM_RESERV_OBJ_LOCK_COUNT; i++)
1160                 mtx_init(&vm_reserv_object_mtx[i], "resv obj lock", NULL,
1161                     MTX_DEF);
1162 }
1163
1164 /*
1165  * Returns true if the given page belongs to a reservation and that page is
1166  * free.  Otherwise, returns false.
1167  */
1168 bool
1169 vm_reserv_is_page_free(vm_page_t m)
1170 {
1171         vm_reserv_t rv;
1172
1173         rv = vm_reserv_from_page(m);
1174         if (rv->object == NULL)
1175                 return (false);
1176         return (popmap_is_clear(rv->popmap, m - rv->pages));
1177 }
1178
1179 /*
1180  * If the given page belongs to a reservation, returns the level of that
1181  * reservation.  Otherwise, returns -1.
1182  */
1183 int
1184 vm_reserv_level(vm_page_t m)
1185 {
1186         vm_reserv_t rv;
1187
1188         rv = vm_reserv_from_page(m);
1189         return (rv->object != NULL ? 0 : -1);
1190 }
1191
1192 /*
1193  * Returns a reservation level if the given page belongs to a fully populated
1194  * reservation and -1 otherwise.
1195  */
1196 int
1197 vm_reserv_level_iffullpop(vm_page_t m)
1198 {
1199         vm_reserv_t rv;
1200
1201         rv = vm_reserv_from_page(m);
1202         return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
1203 }
1204
1205 /*
1206  * Breaks the given partially populated reservation, releasing its free pages
1207  * to the physical memory allocator.
1208  *
1209  * The free page queue lock must be held.
1210  */
1211 static void
1212 vm_reserv_reclaim(vm_reserv_t rv)
1213 {
1214
1215         vm_reserv_assert_locked(rv);
1216         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
1217             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
1218         vm_reserv_domain_lock(rv->domain);
1219         KASSERT(rv->inpartpopq,
1220             ("vm_reserv_reclaim: reserv %p's inpartpopq is FALSE", rv));
1221         KASSERT(rv->domain < vm_ndomains,
1222             ("vm_reserv_reclaim: reserv %p's domain is corrupted %d",
1223             rv, rv->domain));
1224         TAILQ_REMOVE(&vm_rvq_partpop[rv->domain], rv, partpopq);
1225         rv->inpartpopq = FALSE;
1226         vm_reserv_domain_unlock(rv->domain);
1227         vm_reserv_break(rv);
1228         counter_u64_add(vm_reserv_reclaimed, 1);
1229 }
1230
1231 /*
1232  * Breaks the reservation at the head of the partially populated reservation
1233  * queue, releasing its free pages to the physical memory allocator.  Returns
1234  * TRUE if a reservation is broken and FALSE otherwise.
1235  *
1236  * The free page queue lock must be held.
1237  */
1238 boolean_t
1239 vm_reserv_reclaim_inactive(int domain)
1240 {
1241         vm_reserv_t rv;
1242
1243         while ((rv = TAILQ_FIRST(&vm_rvq_partpop[domain])) != NULL) {
1244                 vm_reserv_lock(rv);
1245                 if (rv != TAILQ_FIRST(&vm_rvq_partpop[domain])) {
1246                         vm_reserv_unlock(rv);
1247                         continue;
1248                 }
1249                 vm_reserv_reclaim(rv);
1250                 vm_reserv_unlock(rv);
1251                 return (TRUE);
1252         }
1253         return (FALSE);
1254 }
1255
1256 /*
1257  * Searches the partially populated reservation queue for the least recently
1258  * changed reservation with free pages that satisfy the given request for
1259  * contiguous physical memory.  If a satisfactory reservation is found, it is
1260  * broken.  Returns TRUE if a reservation is broken and FALSE otherwise.
1261  *
1262  * The free page queue lock must be held.
1263  */
1264 boolean_t
1265 vm_reserv_reclaim_contig(int domain, u_long npages, vm_paddr_t low,
1266     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1267 {
1268         vm_paddr_t pa, size;
1269         vm_reserv_t rv, rvn;
1270         int hi, i, lo, low_index, next_free;
1271
1272         if (npages > VM_LEVEL_0_NPAGES - 1)
1273                 return (FALSE);
1274         size = npages << PAGE_SHIFT;
1275         vm_reserv_domain_lock(domain);
1276 again:
1277         for (rv = TAILQ_FIRST(&vm_rvq_partpop[domain]); rv != NULL; rv = rvn) {
1278                 rvn = TAILQ_NEXT(rv, partpopq);
1279                 pa = VM_PAGE_TO_PHYS(&rv->pages[VM_LEVEL_0_NPAGES - 1]);
1280                 if (pa + PAGE_SIZE - size < low) {
1281                         /* This entire reservation is too low; go to next. */
1282                         continue;
1283                 }
1284                 pa = VM_PAGE_TO_PHYS(&rv->pages[0]);
1285                 if (pa + size > high) {
1286                         /* This entire reservation is too high; go to next. */
1287                         continue;
1288                 }
1289                 if (vm_reserv_trylock(rv) == 0) {
1290                         vm_reserv_domain_unlock(domain);
1291                         vm_reserv_lock(rv);
1292                         if (!rv->inpartpopq) {
1293                                 vm_reserv_domain_lock(domain);
1294                                 if (!rvn->inpartpopq)
1295                                         goto again;
1296                                 continue;
1297                         }
1298                 } else
1299                         vm_reserv_domain_unlock(domain);
1300                 if (pa < low) {
1301                         /* Start the search for free pages at "low". */
1302                         low_index = (low + PAGE_MASK - pa) >> PAGE_SHIFT;
1303                         i = low_index / NBPOPMAP;
1304                         hi = low_index % NBPOPMAP;
1305                 } else
1306                         i = hi = 0;
1307                 do {
1308                         /* Find the next free page. */
1309                         lo = ffsl(~(((1UL << hi) - 1) | rv->popmap[i]));
1310                         while (lo == 0 && ++i < NPOPMAP)
1311                                 lo = ffsl(~rv->popmap[i]);
1312                         if (i == NPOPMAP)
1313                                 break;
1314                         /* Convert from ffsl() to ordinary bit numbering. */
1315                         lo--;
1316                         next_free = NBPOPMAP * i + lo;
1317                         pa = VM_PAGE_TO_PHYS(&rv->pages[next_free]);
1318                         KASSERT(pa >= low,
1319                             ("vm_reserv_reclaim_contig: pa is too low"));
1320                         if (pa + size > high) {
1321                                 /* The rest of this reservation is too high. */
1322                                 break;
1323                         } else if ((pa & (alignment - 1)) != 0 ||
1324                             ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0) {
1325                                 /*
1326                                  * The current page doesn't meet the alignment
1327                                  * and/or boundary requirements.  Continue
1328                                  * searching this reservation until the rest
1329                                  * of its free pages are either excluded or
1330                                  * exhausted.
1331                                  */
1332                                 hi = lo + 1;
1333                                 if (hi >= NBPOPMAP) {
1334                                         hi = 0;
1335                                         i++;
1336                                 }
1337                                 continue;
1338                         }
1339                         /* Find the next used page. */
1340                         hi = ffsl(rv->popmap[i] & ~((1UL << lo) - 1));
1341                         while (hi == 0 && ++i < NPOPMAP) {
1342                                 if ((NBPOPMAP * i - next_free) * PAGE_SIZE >=
1343                                     size) {
1344                                         vm_reserv_reclaim(rv);
1345                                         vm_reserv_unlock(rv);
1346                                         return (TRUE);
1347                                 }
1348                                 hi = ffsl(rv->popmap[i]);
1349                         }
1350                         /* Convert from ffsl() to ordinary bit numbering. */
1351                         if (i != NPOPMAP)
1352                                 hi--;
1353                         if ((NBPOPMAP * i + hi - next_free) * PAGE_SIZE >=
1354                             size) {
1355                                 vm_reserv_reclaim(rv);
1356                                 vm_reserv_unlock(rv);
1357                                 return (TRUE);
1358                         }
1359                 } while (i < NPOPMAP);
1360                 vm_reserv_unlock(rv);
1361                 vm_reserv_domain_lock(domain);
1362                 if (rvn != NULL && !rvn->inpartpopq)
1363                         goto again;
1364         }
1365         vm_reserv_domain_unlock(domain);
1366         return (FALSE);
1367 }
1368
1369 /*
1370  * Transfers the reservation underlying the given page to a new object.
1371  *
1372  * The object must be locked.
1373  */
1374 void
1375 vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
1376     vm_pindex_t old_object_offset)
1377 {
1378         vm_reserv_t rv;
1379
1380         VM_OBJECT_ASSERT_WLOCKED(new_object);
1381         rv = vm_reserv_from_page(m);
1382         if (rv->object == old_object) {
1383                 vm_reserv_lock(rv);
1384                 CTR6(KTR_VM,
1385                     "%s: rv %p object %p new %p popcnt %d inpartpop %d",
1386                     __FUNCTION__, rv, rv->object, new_object, rv->popcnt,
1387                     rv->inpartpopq);
1388                 if (rv->object == old_object) {
1389                         vm_reserv_object_lock(old_object);
1390                         rv->object = NULL;
1391                         LIST_REMOVE(rv, objq);
1392                         vm_reserv_object_unlock(old_object);
1393                         vm_reserv_object_lock(new_object);
1394                         rv->object = new_object;
1395                         rv->pindex -= old_object_offset;
1396                         LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
1397                         vm_reserv_object_unlock(new_object);
1398                 }
1399                 vm_reserv_unlock(rv);
1400         }
1401 }
1402
1403 /*
1404  * Returns the size (in bytes) of a reservation of the specified level.
1405  */
1406 int
1407 vm_reserv_size(int level)
1408 {
1409
1410         switch (level) {
1411         case 0:
1412                 return (VM_LEVEL_0_SIZE);
1413         case -1:
1414                 return (PAGE_SIZE);
1415         default:
1416                 return (0);
1417         }
1418 }
1419
1420 /*
1421  * Allocates the virtual and physical memory required by the reservation
1422  * management system's data structures, in particular, the reservation array.
1423  */
1424 vm_paddr_t
1425 vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t high_water)
1426 {
1427         vm_paddr_t new_end;
1428         size_t size;
1429
1430         /*
1431          * Calculate the size (in bytes) of the reservation array.  Round up
1432          * from "high_water" because every small page is mapped to an element
1433          * in the reservation array based on its physical address.  Thus, the
1434          * number of elements in the reservation array can be greater than the
1435          * number of superpages.
1436          */
1437         size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv);
1438
1439         /*
1440          * Allocate and map the physical memory for the reservation array.  The
1441          * next available virtual address is returned by reference.
1442          */
1443         new_end = end - round_page(size);
1444         vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
1445             VM_PROT_READ | VM_PROT_WRITE);
1446         bzero(vm_reserv_array, size);
1447
1448         /*
1449          * Return the next available physical address.
1450          */
1451         return (new_end);
1452 }
1453
1454 /*
1455  * Initializes the reservation management system.  Specifically, initializes
1456  * the reservation counters.
1457  */
1458 static void
1459 vm_reserv_counter_init(void *unused)
1460 {
1461
1462         vm_reserv_freed = counter_u64_alloc(M_WAITOK);
1463         vm_reserv_broken = counter_u64_alloc(M_WAITOK);
1464         vm_reserv_reclaimed = counter_u64_alloc(M_WAITOK);
1465 }
1466 SYSINIT(vm_reserv_counter_init, SI_SUB_CPU, SI_ORDER_ANY,
1467     vm_reserv_counter_init, NULL);
1468
1469 /*
1470  * Returns the superpage containing the given page.
1471  */
1472 vm_page_t
1473 vm_reserv_to_superpage(vm_page_t m)
1474 {
1475         vm_reserv_t rv;
1476
1477         VM_OBJECT_ASSERT_LOCKED(m->object);
1478         rv = vm_reserv_from_page(m);
1479         if (rv->object == m->object && rv->popcnt == VM_LEVEL_0_NPAGES)
1480                 m = rv->pages;
1481         else
1482                 m = NULL;
1483
1484         return (m);
1485 }
1486
1487 #endif  /* VM_NRESERVLEVEL > 0 */