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 #ifndef VM_LEVEL_0_ORDER_MAX
  81 #define VM_LEVEL_0_ORDER_MAX    VM_LEVEL_0_ORDER
  82 #endif
  83
  84 /*
  85  * The number of small pages that are contained in a level 0 reservation
  86  */
  87 #define VM_LEVEL_0_NPAGES       (1 << VM_LEVEL_0_ORDER)
  88 #define VM_LEVEL_0_NPAGES_MAX   (1 << VM_LEVEL_0_ORDER_MAX)
  89
  90 /*
  91  * The number of bits by which a physical address is shifted to obtain the
  92  * reservation number
  93  */
  94 #define VM_LEVEL_0_SHIFT        (VM_LEVEL_0_ORDER + PAGE_SHIFT)
  95
  96 /*
  97  * The size of a level 0 reservation in bytes
  98  */
  99 #define VM_LEVEL_0_SIZE         (1 << VM_LEVEL_0_SHIFT)
 100
 101 /*
 102  * Computes the index of the small page underlying the given (object, pindex)
 103  * within the reservation's array of small pages.
 104  */
 105 #define VM_RESERV_INDEX(object, pindex) \
 106     (((object)->pg_color + (pindex)) & (VM_LEVEL_0_NPAGES - 1))
 107
 108 /*
 109  * The size of a population map entry
 110  */
 111 typedef u_long          popmap_t;
 112
 113 /*
 114  * The number of bits in a population map entry
 115  */
 116 #define NBPOPMAP        (NBBY * sizeof(popmap_t))
 117
 118 /*
 119  * The number of population map entries in a reservation
 120  */
 121 #define NPOPMAP         howmany(VM_LEVEL_0_NPAGES, NBPOPMAP)
 122 #define NPOPMAP_MAX     howmany(VM_LEVEL_0_NPAGES_MAX, NBPOPMAP)
 123
 124 /*
 125  * Number of elapsed ticks before we update the LRU queue position.  Used
 126  * to reduce contention and churn on the list.
 127  */
 128 #define PARTPOPSLOP     1
 129
 130 /*
 131  * Clear a bit in the population map.
 132  */
 133 static __inline void
 134 popmap_clear(popmap_t popmap[], int i)
 135 {
 136
 137         popmap[i / NBPOPMAP] &= ~(1UL << (i % NBPOPMAP));
 138 }
 139
 140 /*
 141  * Set a bit in the population map.
 142  */
 143 static __inline void
 144 popmap_set(popmap_t popmap[], int i)
 145 {
 146
 147         popmap[i / NBPOPMAP] |= 1UL << (i % NBPOPMAP);
 148 }
 149
 150 /*
 151  * Is a bit in the population map clear?
 152  */
 153 static __inline boolean_t
 154 popmap_is_clear(popmap_t popmap[], int i)
 155 {
 156
 157         return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) == 0);
 158 }
 159
 160 /*
 161  * Is a bit in the population map set?
 162  */
 163 static __inline boolean_t
 164 popmap_is_set(popmap_t popmap[], int i)
 165 {
 166
 167         return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) != 0);
 168 }
 169
 170 /*
 171  * The reservation structure
 172  *
 173  * A reservation structure is constructed whenever a large physical page is
 174  * speculatively allocated to an object.  The reservation provides the small
 175  * physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets
 176  * within that object.  The reservation's "popcnt" tracks the number of these
 177  * small physical pages that are in use at any given time.  When and if the
 178  * reservation is not fully utilized, it appears in the queue of partially
 179  * populated reservations.  The reservation always appears on the containing
 180  * object's list of reservations.
 181  *
 182  * A partially populated reservation can be broken and reclaimed at any time.
 183  *
 184  * c - constant after boot
 185  * d - vm_reserv_domain_lock
 186  * o - vm_reserv_object_lock
 187  * r - vm_reserv_lock
 188  * s - vm_reserv_domain_scan_lock
 189  */
 190 struct vm_reserv {
 191         struct mtx      lock;                   /* reservation lock. */
 192         TAILQ_ENTRY(vm_reserv) partpopq;        /* (d, r) per-domain queue. */
 193         LIST_ENTRY(vm_reserv) objq;             /* (o, r) object queue */
 194         vm_object_t     object;                 /* (o, r) containing object */
 195         vm_pindex_t     pindex;                 /* (o, r) offset in object */
 196         vm_page_t       pages;                  /* (c) first page  */
 197         uint16_t        popcnt;                 /* (r) # of pages in use */
 198         uint8_t         domain;                 /* (c) NUMA domain. */
 199         char            inpartpopq;             /* (d, r) */
 200         int             lasttick;               /* (r) last pop update tick. */
 201         popmap_t        popmap[NPOPMAP_MAX];    /* (r) bit vector, used pages */
 202 };
 203
 204 TAILQ_HEAD(vm_reserv_queue, vm_reserv);
 205
 206 #define vm_reserv_lockptr(rv)           (&(rv)->lock)
 207 #define vm_reserv_assert_locked(rv)                                     \
 208             mtx_assert(vm_reserv_lockptr(rv), MA_OWNED)
 209 #define vm_reserv_lock(rv)              mtx_lock(vm_reserv_lockptr(rv))
 210 #define vm_reserv_trylock(rv)           mtx_trylock(vm_reserv_lockptr(rv))
 211 #define vm_reserv_unlock(rv)            mtx_unlock(vm_reserv_lockptr(rv))
 212
 213 /*
 214  * The reservation array
 215  *
 216  * This array is analoguous in function to vm_page_array.  It differs in the
 217  * respect that it may contain a greater number of useful reservation
 218  * structures than there are (physical) superpages.  These "invalid"
 219  * reservation structures exist to trade-off space for time in the
 220  * implementation of vm_reserv_from_page().  Invalid reservation structures are
 221  * distinguishable from "valid" reservation structures by inspecting the
 222  * reservation's "pages" field.  Invalid reservation structures have a NULL
 223  * "pages" field.
 224  *
 225  * vm_reserv_from_page() maps a small (physical) page to an element of this
 226  * array by computing a physical reservation number from the page's physical
 227  * address.  The physical reservation number is used as the array index.
 228  *
 229  * An "active" reservation is a valid reservation structure that has a non-NULL
 230  * "object" field and a non-zero "popcnt" field.  In other words, every active
 231  * reservation belongs to a particular object.  Moreover, every active
 232  * reservation has an entry in the containing object's list of reservations.
 233  */
 234 static vm_reserv_t vm_reserv_array;
 235
 236 /*
 237  * The per-domain partially populated reservation queues
 238  *
 239  * These queues enable the fast recovery of an unused free small page from a
 240  * partially populated reservation.  The reservation at the head of a queue
 241  * is the least recently changed, partially populated reservation.
 242  *
 243  * Access to this queue is synchronized by the per-domain reservation lock.
 244  * Threads reclaiming free pages from the queue must hold the per-domain scan
 245  * lock.
 246  */
 247 struct vm_reserv_domain {
 248         struct mtx              lock;
 249         struct vm_reserv_queue  partpop;        /* (d) */
 250         struct vm_reserv        marker;         /* (d, s) scan marker/lock */
 251 } __aligned(CACHE_LINE_SIZE);
 252
 253 static struct vm_reserv_domain vm_rvd[MAXMEMDOM];
 254
 255 #define vm_reserv_domain_lockptr(d)     (&vm_rvd[(d)].lock)
 256 #define vm_reserv_domain_assert_locked(d)       \
 257         mtx_assert(vm_reserv_domain_lockptr(d), MA_OWNED)
 258 #define vm_reserv_domain_lock(d)        mtx_lock(vm_reserv_domain_lockptr(d))
 259 #define vm_reserv_domain_unlock(d)      mtx_unlock(vm_reserv_domain_lockptr(d))
 260
 261 #define vm_reserv_domain_scan_lock(d)   mtx_lock(&vm_rvd[(d)].marker.lock)
 262 #define vm_reserv_domain_scan_unlock(d) mtx_unlock(&vm_rvd[(d)].marker.lock)
 263
 264 static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
 265     "Reservation Info");
 266
 267 static COUNTER_U64_DEFINE_EARLY(vm_reserv_broken);
 268 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD,
 269     &vm_reserv_broken, "Cumulative number of broken reservations");
 270
 271 static COUNTER_U64_DEFINE_EARLY(vm_reserv_freed);
 272 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD,
 273     &vm_reserv_freed, "Cumulative number of freed reservations");
 274
 275 static int sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS);
 276
 277 SYSCTL_PROC(_vm_reserv, OID_AUTO, fullpop, CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RD,
 278     NULL, 0, sysctl_vm_reserv_fullpop, "I", "Current number of full reservations");
 279
 280 static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);
 281
 282 SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq,
 283     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, NULL, 0,
 284     sysctl_vm_reserv_partpopq, "A",
 285     "Partially populated reservation queues");
 286
 287 static COUNTER_U64_DEFINE_EARLY(vm_reserv_reclaimed);
 288 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
 289     &vm_reserv_reclaimed, "Cumulative number of reclaimed reservations");
 290
 291 /*
 292  * The object lock pool is used to synchronize the rvq.  We can not use a
 293  * pool mutex because it is required before malloc works.
 294  *
 295  * The "hash" function could be made faster without divide and modulo.
 296  */
 297 #define VM_RESERV_OBJ_LOCK_COUNT        MAXCPU
 298
 299 struct mtx_padalign vm_reserv_object_mtx[VM_RESERV_OBJ_LOCK_COUNT];
 300
 301 #define vm_reserv_object_lock_idx(object)                       \
 302             (((uintptr_t)object / sizeof(*object)) % VM_RESERV_OBJ_LOCK_COUNT)
 303 #define vm_reserv_object_lock_ptr(object)                       \
 304             &vm_reserv_object_mtx[vm_reserv_object_lock_idx((object))]
 305 #define vm_reserv_object_lock(object)                           \
 306             mtx_lock(vm_reserv_object_lock_ptr((object)))
 307 #define vm_reserv_object_unlock(object)                         \
 308             mtx_unlock(vm_reserv_object_lock_ptr((object)))
 309
 310 static void             vm_reserv_break(vm_reserv_t rv);
 311 static void             vm_reserv_depopulate(vm_reserv_t rv, int index);
 312 static vm_reserv_t      vm_reserv_from_page(vm_page_t m);
 313 static boolean_t        vm_reserv_has_pindex(vm_reserv_t rv,
 314                             vm_pindex_t pindex);
 315 static void             vm_reserv_populate(vm_reserv_t rv, int index);
 316 static void             vm_reserv_reclaim(vm_reserv_t rv);
 317
 318 /*
 319  * Returns the current number of full reservations.
 320  *
 321  * Since the number of full reservations is computed without acquiring any
 322  * locks, the returned value is inexact.
 323  */
 324 static int
 325 sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS)
 326 {
 327         vm_paddr_t paddr;
 328         struct vm_phys_seg *seg;
 329         vm_reserv_t rv;
 330         int fullpop, segind;
 331
 332         fullpop = 0;
 333         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 334                 seg = &vm_phys_segs[segind];
 335                 paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
 336                 while (paddr + VM_LEVEL_0_SIZE > paddr && paddr +
 337                     VM_LEVEL_0_SIZE <= seg->end) {
 338                         rv = &vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT];
 339                         fullpop += rv->popcnt == VM_LEVEL_0_NPAGES;
 340                         paddr += VM_LEVEL_0_SIZE;
 341                 }
 342         }
 343         return (sysctl_handle_int(oidp, &fullpop, 0, req));
 344 }
 345
 346 /*
 347  * Describes the current state of the partially populated reservation queue.
 348  */
 349 static int
 350 sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
 351 {
 352         struct sbuf sbuf;
 353         vm_reserv_t rv;
 354         int counter, error, domain, level, unused_pages;
 355
 356         error = sysctl_wire_old_buffer(req, 0);
 357         if (error != 0)
 358                 return (error);
 359         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 360         sbuf_printf(&sbuf, "\nDOMAIN    LEVEL     SIZE  NUMBER\n\n");
 361         for (domain = 0; domain < vm_ndomains; domain++) {
 362                 for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
 363                         counter = 0;
 364                         unused_pages = 0;
 365                         vm_reserv_domain_lock(domain);
 366                         TAILQ_FOREACH(rv, &vm_rvd[domain].partpop, partpopq) {
 367                                 if (rv == &vm_rvd[domain].marker)
 368                                         continue;
 369                                 counter++;
 370                                 unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
 371                         }
 372                         vm_reserv_domain_unlock(domain);
 373                         sbuf_printf(&sbuf, "%6d, %7d, %6dK, %6d\n",
 374                             domain, level,
 375                             unused_pages * ((int)PAGE_SIZE / 1024), counter);
 376                 }
 377         }
 378         error = sbuf_finish(&sbuf);
 379         sbuf_delete(&sbuf);
 380         return (error);
 381 }
 382
 383 /*
 384  * Remove a reservation from the object's objq.
 385  */
 386 static void
 387 vm_reserv_remove(vm_reserv_t rv)
 388 {
 389         vm_object_t object;
 390
 391         vm_reserv_assert_locked(rv);
 392         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
 393             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
 394         KASSERT(rv->object != NULL,
 395             ("vm_reserv_remove: reserv %p is free", rv));
 396         KASSERT(!rv->inpartpopq,
 397             ("vm_reserv_remove: reserv %p's inpartpopq is TRUE", rv));
 398         object = rv->object;
 399         vm_reserv_object_lock(object);
 400         LIST_REMOVE(rv, objq);
 401         rv->object = NULL;
 402         vm_reserv_object_unlock(object);
 403 }
 404
 405 /*
 406  * Insert a new reservation into the object's objq.
 407  */
 408 static void
 409 vm_reserv_insert(vm_reserv_t rv, vm_object_t object, vm_pindex_t pindex)
 410 {
 411         int i;
 412
 413         vm_reserv_assert_locked(rv);
 414         CTR6(KTR_VM,
 415             "%s: rv %p(%p) object %p new %p popcnt %d",
 416             __FUNCTION__, rv, rv->pages, rv->object, object,
 417            rv->popcnt);
 418         KASSERT(rv->object == NULL,
 419             ("vm_reserv_insert: reserv %p isn't free", rv));
 420         KASSERT(rv->popcnt == 0,
 421             ("vm_reserv_insert: reserv %p's popcnt is corrupted", rv));
 422         KASSERT(!rv->inpartpopq,
 423             ("vm_reserv_insert: reserv %p's inpartpopq is TRUE", rv));
 424         for (i = 0; i < NPOPMAP; i++)
 425                 KASSERT(rv->popmap[i] == 0,
 426                     ("vm_reserv_insert: reserv %p's popmap is corrupted", rv));
 427         vm_reserv_object_lock(object);
 428         rv->pindex = pindex;
 429         rv->object = object;
 430         rv->lasttick = ticks;
 431         LIST_INSERT_HEAD(&object->rvq, rv, objq);
 432         vm_reserv_object_unlock(object);
 433 }
 434
 435 /*
 436  * Reduces the given reservation's population count.  If the population count
 437  * becomes zero, the reservation is destroyed.  Additionally, moves the
 438  * reservation to the tail of the partially populated reservation queue if the
 439  * population count is non-zero.
 440  */
 441 static void
 442 vm_reserv_depopulate(vm_reserv_t rv, int index)
 443 {
 444         struct vm_domain *vmd;
 445
 446         vm_reserv_assert_locked(rv);
 447         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
 448             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
 449         KASSERT(rv->object != NULL,
 450             ("vm_reserv_depopulate: reserv %p is free", rv));
 451         KASSERT(popmap_is_set(rv->popmap, index),
 452             ("vm_reserv_depopulate: reserv %p's popmap[%d] is clear", rv,
 453             index));
 454         KASSERT(rv->popcnt > 0,
 455             ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
 456         KASSERT(rv->domain < vm_ndomains,
 457             ("vm_reserv_depopulate: reserv %p's domain is corrupted %d",
 458             rv, rv->domain));
 459         if (rv->popcnt == VM_LEVEL_0_NPAGES) {
 460                 KASSERT(rv->pages->psind == 1,
 461                     ("vm_reserv_depopulate: reserv %p is already demoted",
 462                     rv));
 463                 rv->pages->psind = 0;
 464         }
 465         popmap_clear(rv->popmap, index);
 466         rv->popcnt--;
 467         if ((unsigned)(ticks - rv->lasttick) >= PARTPOPSLOP ||
 468             rv->popcnt == 0) {
 469                 vm_reserv_domain_lock(rv->domain);
 470                 if (rv->inpartpopq) {
 471                         TAILQ_REMOVE(&vm_rvd[rv->domain].partpop, rv, partpopq);
 472                         rv->inpartpopq = FALSE;
 473                 }
 474                 if (rv->popcnt != 0) {
 475                         rv->inpartpopq = TRUE;
 476                         TAILQ_INSERT_TAIL(&vm_rvd[rv->domain].partpop, rv,
 477                             partpopq);
 478                 }
 479                 vm_reserv_domain_unlock(rv->domain);
 480                 rv->lasttick = ticks;
 481         }
 482         vmd = VM_DOMAIN(rv->domain);
 483         if (rv->popcnt == 0) {
 484                 vm_reserv_remove(rv);
 485                 vm_domain_free_lock(vmd);
 486                 vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
 487                 vm_domain_free_unlock(vmd);
 488                 counter_u64_add(vm_reserv_freed, 1);
 489         }
 490         vm_domain_freecnt_inc(vmd, 1);
 491 }
 492
 493 /*
 494  * Returns the reservation to which the given page might belong.
 495  */
 496 static __inline vm_reserv_t
 497 vm_reserv_from_page(vm_page_t m)
 498 {
 499
 500         return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
 501 }
 502
 503 /*
 504  * Returns an existing reservation or NULL and initialized successor pointer.
 505  */
 506 static vm_reserv_t
 507 vm_reserv_from_object(vm_object_t object, vm_pindex_t pindex,
 508     vm_page_t mpred, vm_page_t *msuccp)
 509 {
 510         vm_reserv_t rv;
 511         vm_page_t msucc;
 512
 513         msucc = NULL;
 514         if (mpred != NULL) {
 515                 KASSERT(mpred->object == object,
 516                     ("vm_reserv_from_object: object doesn't contain mpred"));
 517                 KASSERT(mpred->pindex < pindex,
 518                     ("vm_reserv_from_object: mpred doesn't precede pindex"));
 519                 rv = vm_reserv_from_page(mpred);
 520                 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
 521                         goto found;
 522                 msucc = TAILQ_NEXT(mpred, listq);
 523         } else
 524                 msucc = TAILQ_FIRST(&object->memq);
 525         if (msucc != NULL) {
 526                 KASSERT(msucc->pindex > pindex,
 527                     ("vm_reserv_from_object: msucc doesn't succeed pindex"));
 528                 rv = vm_reserv_from_page(msucc);
 529                 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
 530                         goto found;
 531         }
 532         rv = NULL;
 533
 534 found:
 535         *msuccp = msucc;
 536
 537         return (rv);
 538 }
 539
 540 /*
 541  * Returns TRUE if the given reservation contains the given page index and
 542  * FALSE otherwise.
 543  */
 544 static __inline boolean_t
 545 vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
 546 {
 547
 548         return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
 549 }
 550
 551 /*
 552  * Increases the given reservation's population count.  Moves the reservation
 553  * to the tail of the partially populated reservation queue.
 554  */
 555 static void
 556 vm_reserv_populate(vm_reserv_t rv, int index)
 557 {
 558
 559         vm_reserv_assert_locked(rv);
 560         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
 561             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
 562         KASSERT(rv->object != NULL,
 563             ("vm_reserv_populate: reserv %p is free", rv));
 564         KASSERT(popmap_is_clear(rv->popmap, index),
 565             ("vm_reserv_populate: reserv %p's popmap[%d] is set", rv,
 566             index));
 567         KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
 568             ("vm_reserv_populate: reserv %p is already full", rv));
 569         KASSERT(rv->pages->psind == 0,
 570             ("vm_reserv_populate: reserv %p is already promoted", rv));
 571         KASSERT(rv->domain < vm_ndomains,
 572             ("vm_reserv_populate: reserv %p's domain is corrupted %d",
 573             rv, rv->domain));
 574         popmap_set(rv->popmap, index);
 575         rv->popcnt++;
 576         if ((unsigned)(ticks - rv->lasttick) < PARTPOPSLOP &&
 577             rv->inpartpopq && rv->popcnt != VM_LEVEL_0_NPAGES)
 578                 return;
 579         rv->lasttick = ticks;
 580         vm_reserv_domain_lock(rv->domain);
 581         if (rv->inpartpopq) {
 582                 TAILQ_REMOVE(&vm_rvd[rv->domain].partpop, rv, partpopq);
 583                 rv->inpartpopq = FALSE;
 584         }
 585         if (rv->popcnt < VM_LEVEL_0_NPAGES) {
 586                 rv->inpartpopq = TRUE;
 587                 TAILQ_INSERT_TAIL(&vm_rvd[rv->domain].partpop, rv, partpopq);
 588         } else {
 589                 KASSERT(rv->pages->psind == 0,
 590                     ("vm_reserv_populate: reserv %p is already promoted",
 591                     rv));
 592                 rv->pages->psind = 1;
 593         }
 594         vm_reserv_domain_unlock(rv->domain);
 595 }
 596
 597 /*
 598  * Allocates a contiguous set of physical pages of the given size "npages"
 599  * from existing or newly created reservations.  All of the physical pages
 600  * must be at or above the given physical address "low" and below the given
 601  * physical address "high".  The given value "alignment" determines the
 602  * alignment of the first physical page in the set.  If the given value
 603  * "boundary" is non-zero, then the set of physical pages cannot cross any
 604  * physical address boundary that is a multiple of that value.  Both
 605  * "alignment" and "boundary" must be a power of two.
 606  *
 607  * The page "mpred" must immediately precede the offset "pindex" within the
 608  * specified object.
 609  *
 610  * The object must be locked.
 611  */
 612 vm_page_t
 613 vm_reserv_alloc_contig(vm_object_t object, vm_pindex_t pindex, int domain,
 614     int req, vm_page_t mpred, u_long npages, vm_paddr_t low, vm_paddr_t high,
 615     u_long alignment, vm_paddr_t boundary)
 616 {
 617         struct vm_domain *vmd;
 618         vm_paddr_t pa, size;
 619         vm_page_t m, m_ret, msucc;
 620         vm_pindex_t first, leftcap, rightcap;
 621         vm_reserv_t rv;
 622         u_long allocpages, maxpages, minpages;
 623         int i, index, n;
 624
 625         VM_OBJECT_ASSERT_WLOCKED(object);
 626         KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
 627
 628         /*
 629          * Is a reservation fundamentally impossible?
 630          */
 631         if (pindex < VM_RESERV_INDEX(object, pindex) ||
 632             pindex + npages > object->size)
 633                 return (NULL);
 634
 635         /*
 636          * All reservations of a particular size have the same alignment.
 637          * Assuming that the first page is allocated from a reservation, the
 638          * least significant bits of its physical address can be determined
 639          * from its offset from the beginning of the reservation and the size
 640          * of the reservation.
 641          *
 642          * Could the specified index within a reservation of the smallest
 643          * possible size satisfy the alignment and boundary requirements?
 644          */
 645         pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
 646         if ((pa & (alignment - 1)) != 0)
 647                 return (NULL);
 648         size = npages << PAGE_SHIFT;
 649         if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
 650                 return (NULL);
 651
 652         /*
 653          * Look for an existing reservation.
 654          */
 655         rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
 656         if (rv != NULL) {
 657                 KASSERT(object != kernel_object || rv->domain == domain,
 658                     ("vm_reserv_alloc_contig: domain mismatch"));
 659                 index = VM_RESERV_INDEX(object, pindex);
 660                 /* Does the allocation fit within the reservation? */
 661                 if (index + npages > VM_LEVEL_0_NPAGES)
 662                         return (NULL);
 663                 domain = rv->domain;
 664                 vmd = VM_DOMAIN(domain);
 665                 vm_reserv_lock(rv);
 666                 /* Handle reclaim race. */
 667                 if (rv->object != object)
 668                         goto out;
 669                 m = &rv->pages[index];
 670                 pa = VM_PAGE_TO_PHYS(m);
 671                 if (pa < low || pa + size > high ||
 672                     (pa & (alignment - 1)) != 0 ||
 673                     ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
 674                         goto out;
 675                 /* Handle vm_page_rename(m, new_object, ...). */
 676                 for (i = 0; i < npages; i++)
 677                         if (popmap_is_set(rv->popmap, index + i))
 678                                 goto out;
 679                 if (!vm_domain_allocate(vmd, req, npages))
 680                         goto out;
 681                 for (i = 0; i < npages; i++)
 682                         vm_reserv_populate(rv, index + i);
 683                 vm_reserv_unlock(rv);
 684                 return (m);
 685 out:
 686                 vm_reserv_unlock(rv);
 687                 return (NULL);
 688         }
 689
 690         /*
 691          * Could at least one reservation fit between the first index to the
 692          * left that can be used ("leftcap") and the first index to the right
 693          * that cannot be used ("rightcap")?
 694          *
 695          * We must synchronize with the reserv object lock to protect the
 696          * pindex/object of the resulting reservations against rename while
 697          * we are inspecting.
 698          */
 699         first = pindex - VM_RESERV_INDEX(object, pindex);
 700         minpages = VM_RESERV_INDEX(object, pindex) + npages;
 701         maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES);
 702         allocpages = maxpages;
 703         vm_reserv_object_lock(object);
 704         if (mpred != NULL) {
 705                 if ((rv = vm_reserv_from_page(mpred))->object != object)
 706                         leftcap = mpred->pindex + 1;
 707                 else
 708                         leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
 709                 if (leftcap > first) {
 710                         vm_reserv_object_unlock(object);
 711                         return (NULL);
 712                 }
 713         }
 714         if (msucc != NULL) {
 715                 if ((rv = vm_reserv_from_page(msucc))->object != object)
 716                         rightcap = msucc->pindex;
 717                 else
 718                         rightcap = rv->pindex;
 719                 if (first + maxpages > rightcap) {
 720                         if (maxpages == VM_LEVEL_0_NPAGES) {
 721                                 vm_reserv_object_unlock(object);
 722                                 return (NULL);
 723                         }
 724
 725                         /*
 726                          * At least one reservation will fit between "leftcap"
 727                          * and "rightcap".  However, a reservation for the
 728                          * last of the requested pages will not fit.  Reduce
 729                          * the size of the upcoming allocation accordingly.
 730                          */
 731                         allocpages = minpages;
 732                 }
 733         }
 734         vm_reserv_object_unlock(object);
 735
 736         /*
 737          * Would the last new reservation extend past the end of the object?
 738          *
 739          * If the object is unlikely to grow don't allocate a reservation for
 740          * the tail.
 741          */
 742         if ((object->flags & OBJ_ANON) == 0 &&
 743             first + maxpages > object->size) {
 744                 if (maxpages == VM_LEVEL_0_NPAGES)
 745                         return (NULL);
 746                 allocpages = minpages;
 747         }
 748
 749         /*
 750          * Allocate the physical pages.  The alignment and boundary specified
 751          * for this allocation may be different from the alignment and
 752          * boundary specified for the requested pages.  For instance, the
 753          * specified index may not be the first page within the first new
 754          * reservation.
 755          */
 756         m = NULL;
 757         vmd = VM_DOMAIN(domain);
 758         if (vm_domain_allocate(vmd, req, npages)) {
 759                 vm_domain_free_lock(vmd);
 760                 m = vm_phys_alloc_contig(domain, allocpages, low, high,
 761                     ulmax(alignment, VM_LEVEL_0_SIZE),
 762                     boundary > VM_LEVEL_0_SIZE ? boundary : 0);
 763                 vm_domain_free_unlock(vmd);
 764                 if (m == NULL) {
 765                         vm_domain_freecnt_inc(vmd, npages);
 766                         return (NULL);
 767                 }
 768         } else
 769                 return (NULL);
 770         KASSERT(vm_phys_domain(m) == domain,
 771             ("vm_reserv_alloc_contig: Page domain does not match requested."));
 772
 773         /*
 774          * The allocated physical pages always begin at a reservation
 775          * boundary, but they do not always end at a reservation boundary.
 776          * Initialize every reservation that is completely covered by the
 777          * allocated physical pages.
 778          */
 779         m_ret = NULL;
 780         index = VM_RESERV_INDEX(object, pindex);
 781         do {
 782                 rv = vm_reserv_from_page(m);
 783                 KASSERT(rv->pages == m,
 784                     ("vm_reserv_alloc_contig: reserv %p's pages is corrupted",
 785                     rv));
 786                 vm_reserv_lock(rv);
 787                 vm_reserv_insert(rv, object, first);
 788                 n = ulmin(VM_LEVEL_0_NPAGES - index, npages);
 789                 for (i = 0; i < n; i++)
 790                         vm_reserv_populate(rv, index + i);
 791                 npages -= n;
 792                 if (m_ret == NULL) {
 793                         m_ret = &rv->pages[index];
 794                         index = 0;
 795                 }
 796                 vm_reserv_unlock(rv);
 797                 m += VM_LEVEL_0_NPAGES;
 798                 first += VM_LEVEL_0_NPAGES;
 799                 allocpages -= VM_LEVEL_0_NPAGES;
 800         } while (allocpages >= VM_LEVEL_0_NPAGES);
 801         return (m_ret);
 802 }
 803
 804 /*
 805  * Allocate a physical page from an existing or newly created reservation.
 806  *
 807  * The page "mpred" must immediately precede the offset "pindex" within the
 808  * specified object.
 809  *
 810  * The object must be locked.
 811  */
 812 vm_page_t
 813 vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex, int domain,
 814     int req, vm_page_t mpred)
 815 {
 816         struct vm_domain *vmd;
 817         vm_page_t m, msucc;
 818         vm_pindex_t first, leftcap, rightcap;
 819         vm_reserv_t rv;
 820         int index;
 821
 822         VM_OBJECT_ASSERT_WLOCKED(object);
 823
 824         /*
 825          * Is a reservation fundamentally impossible?
 826          */
 827         if (pindex < VM_RESERV_INDEX(object, pindex) ||
 828             pindex >= object->size)
 829                 return (NULL);
 830
 831         /*
 832          * Look for an existing reservation.
 833          */
 834         rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
 835         if (rv != NULL) {
 836                 KASSERT(object != kernel_object || rv->domain == domain,
 837                     ("vm_reserv_alloc_page: domain mismatch"));
 838                 domain = rv->domain;
 839                 vmd = VM_DOMAIN(domain);
 840                 index = VM_RESERV_INDEX(object, pindex);
 841                 m = &rv->pages[index];
 842                 vm_reserv_lock(rv);
 843                 /* Handle reclaim race. */
 844                 if (rv->object != object ||
 845                     /* Handle vm_page_rename(m, new_object, ...). */
 846                     popmap_is_set(rv->popmap, index)) {
 847                         m = NULL;
 848                         goto out;
 849                 }
 850                 if (vm_domain_allocate(vmd, req, 1) == 0)
 851                         m = NULL;
 852                 else
 853                         vm_reserv_populate(rv, index);
 854 out:
 855                 vm_reserv_unlock(rv);
 856                 return (m);
 857         }
 858
 859         /*
 860          * Could a reservation fit between the first index to the left that
 861          * can be used and the first index to the right that cannot be used?
 862          *
 863          * We must synchronize with the reserv object lock to protect the
 864          * pindex/object of the resulting reservations against rename while
 865          * we are inspecting.
 866          */
 867         first = pindex - VM_RESERV_INDEX(object, pindex);
 868         vm_reserv_object_lock(object);
 869         if (mpred != NULL) {
 870                 if ((rv = vm_reserv_from_page(mpred))->object != object)
 871                         leftcap = mpred->pindex + 1;
 872                 else
 873                         leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
 874                 if (leftcap > first) {
 875                         vm_reserv_object_unlock(object);
 876                         return (NULL);
 877                 }
 878         }
 879         if (msucc != NULL) {
 880                 if ((rv = vm_reserv_from_page(msucc))->object != object)
 881                         rightcap = msucc->pindex;
 882                 else
 883                         rightcap = rv->pindex;
 884                 if (first + VM_LEVEL_0_NPAGES > rightcap) {
 885                         vm_reserv_object_unlock(object);
 886                         return (NULL);
 887                 }
 888         }
 889         vm_reserv_object_unlock(object);
 890
 891         /*
 892          * Would the last new reservation extend past the end of the object?
 893          *
 894          * If the object is unlikely to grow don't allocate a reservation for
 895          * the tail.
 896          */
 897         if ((object->flags & OBJ_ANON) == 0 &&
 898             first + VM_LEVEL_0_NPAGES > object->size)
 899                 return (NULL);
 900
 901         /*
 902          * Allocate and populate the new reservation.
 903          */
 904         m = NULL;
 905         vmd = VM_DOMAIN(domain);
 906         if (vm_domain_allocate(vmd, req, 1)) {
 907                 vm_domain_free_lock(vmd);
 908                 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT,
 909                     VM_LEVEL_0_ORDER);
 910                 vm_domain_free_unlock(vmd);
 911                 if (m == NULL) {
 912                         vm_domain_freecnt_inc(vmd, 1);
 913                         return (NULL);
 914                 }
 915         } else
 916                 return (NULL);
 917         rv = vm_reserv_from_page(m);
 918         vm_reserv_lock(rv);
 919         KASSERT(rv->pages == m,
 920             ("vm_reserv_alloc_page: reserv %p's pages is corrupted", rv));
 921         vm_reserv_insert(rv, object, first);
 922         index = VM_RESERV_INDEX(object, pindex);
 923         vm_reserv_populate(rv, index);
 924         vm_reserv_unlock(rv);
 925
 926         return (&rv->pages[index]);
 927 }
 928
 929 /*
 930  * Breaks the given reservation.  All free pages in the reservation
 931  * are returned to the physical memory allocator.  The reservation's
 932  * population count and map are reset to their initial state.
 933  *
 934  * The given reservation must not be in the partially populated reservation
 935  * queue.
 936  */
 937 static void
 938 vm_reserv_break(vm_reserv_t rv)
 939 {
 940         u_long changes;
 941         int bitpos, hi, i, lo;
 942
 943         vm_reserv_assert_locked(rv);
 944         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
 945             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
 946         vm_reserv_remove(rv);
 947         rv->pages->psind = 0;
 948         hi = lo = -1;
 949         for (i = 0; i <= NPOPMAP; i++) {
 950                 /*
 951                  * "changes" is a bitmask that marks where a new sequence of
 952                  * 0s or 1s begins in popmap[i], with last bit in popmap[i-1]
 953                  * considered to be 1 if and only if lo == hi.  The bits of
 954                  * popmap[-1] and popmap[NPOPMAP] are considered all 1s.
 955                  */
 956                 if (i == NPOPMAP)
 957                         changes = lo != hi;
 958                 else {
 959                         changes = rv->popmap[i];
 960                         changes ^= (changes << 1) | (lo == hi);
 961                         rv->popmap[i] = 0;
 962                 }
 963                 while (changes != 0) {
 964                         /*
 965                          * If the next change marked begins a run of 0s, set
 966                          * lo to mark that position.  Otherwise set hi and
 967                          * free pages from lo up to hi.
 968                          */
 969                         bitpos = ffsl(changes) - 1;
 970                         changes ^= 1UL << bitpos;
 971                         if (lo == hi)
 972                                 lo = NBPOPMAP * i + bitpos;
 973                         else {
 974                                 hi = NBPOPMAP * i + bitpos;
 975                                 vm_domain_free_lock(VM_DOMAIN(rv->domain));
 976                                 vm_phys_enqueue_contig(&rv->pages[lo], hi - lo);
 977                                 vm_domain_free_unlock(VM_DOMAIN(rv->domain));
 978                                 lo = hi;
 979                         }
 980                 }
 981         }
 982         rv->popcnt = 0;
 983         counter_u64_add(vm_reserv_broken, 1);
 984 }
 985
 986 /*
 987  * Breaks all reservations belonging to the given object.
 988  */
 989 void
 990 vm_reserv_break_all(vm_object_t object)
 991 {
 992         vm_reserv_t rv;
 993
 994         /*
 995          * This access of object->rvq is unsynchronized so that the
 996          * object rvq lock can nest after the domain_free lock.  We
 997          * must check for races in the results.  However, the object
 998          * lock prevents new additions, so we are guaranteed that when
 999          * it returns NULL the object is properly empty.
1000          */
1001         while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
1002                 vm_reserv_lock(rv);
1003                 /* Reclaim race. */
1004                 if (rv->object != object) {
1005                         vm_reserv_unlock(rv);
1006                         continue;
1007                 }
1008                 vm_reserv_domain_lock(rv->domain);
1009                 if (rv->inpartpopq) {
1010                         TAILQ_REMOVE(&vm_rvd[rv->domain].partpop, rv, partpopq);
1011                         rv->inpartpopq = FALSE;
1012                 }
1013                 vm_reserv_domain_unlock(rv->domain);
1014                 vm_reserv_break(rv);
1015                 vm_reserv_unlock(rv);
1016         }
1017 }
1018
1019 /*
1020  * Frees the given page if it belongs to a reservation.  Returns TRUE if the
1021  * page is freed and FALSE otherwise.
1022  */
1023 boolean_t
1024 vm_reserv_free_page(vm_page_t m)
1025 {
1026         vm_reserv_t rv;
1027         boolean_t ret;
1028
1029         rv = vm_reserv_from_page(m);
1030         if (rv->object == NULL)
1031                 return (FALSE);
1032         vm_reserv_lock(rv);
1033         /* Re-validate after lock. */
1034         if (rv->object != NULL) {
1035                 vm_reserv_depopulate(rv, m - rv->pages);
1036                 ret = TRUE;
1037         } else
1038                 ret = FALSE;
1039         vm_reserv_unlock(rv);
1040
1041         return (ret);
1042 }
1043
1044 /*
1045  * Initializes the reservation management system.  Specifically, initializes
1046  * the reservation array.
1047  *
1048  * Requires that vm_page_array and first_page are initialized!
1049  */
1050 void
1051 vm_reserv_init(void)
1052 {
1053         vm_paddr_t paddr;
1054         struct vm_phys_seg *seg;
1055         struct vm_reserv *rv;
1056         struct vm_reserv_domain *rvd;
1057         int i, j, segind;
1058
1059         /*
1060          * Initialize the reservation array.  Specifically, initialize the
1061          * "pages" field for every element that has an underlying superpage.
1062          */
1063         for (segind = 0; segind < vm_phys_nsegs; segind++) {
1064                 seg = &vm_phys_segs[segind];
1065                 paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
1066                 while (paddr + VM_LEVEL_0_SIZE > paddr && paddr +
1067                     VM_LEVEL_0_SIZE <= seg->end) {
1068                         rv = &vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT];
1069                         rv->pages = PHYS_TO_VM_PAGE(paddr);
1070                         rv->domain = seg->domain;
1071                         mtx_init(&rv->lock, "vm reserv", NULL, MTX_DEF);
1072                         paddr += VM_LEVEL_0_SIZE;
1073                 }
1074         }
1075         for (i = 0; i < MAXMEMDOM; i++) {
1076                 rvd = &vm_rvd[i];
1077                 mtx_init(&rvd->lock, "vm reserv domain", NULL, MTX_DEF);
1078                 TAILQ_INIT(&rvd->partpop);
1079                 mtx_init(&rvd->marker.lock, "vm reserv marker", NULL, MTX_DEF);
1080
1081                 /*
1082                  * Fully populated reservations should never be present in the
1083                  * partially populated reservation queues.
1084                  */
1085                 rvd->marker.popcnt = VM_LEVEL_0_NPAGES;
1086                 for (j = 0; j < NBPOPMAP; j++)
1087                         popmap_set(rvd->marker.popmap, j);
1088         }
1089
1090         for (i = 0; i < VM_RESERV_OBJ_LOCK_COUNT; i++)
1091                 mtx_init(&vm_reserv_object_mtx[i], "resv obj lock", NULL,
1092                     MTX_DEF);
1093 }
1094
1095 /*
1096  * Returns true if the given page belongs to a reservation and that page is
1097  * free.  Otherwise, returns false.
1098  */
1099 bool
1100 vm_reserv_is_page_free(vm_page_t m)
1101 {
1102         vm_reserv_t rv;
1103
1104         rv = vm_reserv_from_page(m);
1105         if (rv->object == NULL)
1106                 return (false);
1107         return (popmap_is_clear(rv->popmap, m - rv->pages));
1108 }
1109
1110 /*
1111  * If the given page belongs to a reservation, returns the level of that
1112  * reservation.  Otherwise, returns -1.
1113  */
1114 int
1115 vm_reserv_level(vm_page_t m)
1116 {
1117         vm_reserv_t rv;
1118
1119         rv = vm_reserv_from_page(m);
1120         return (rv->object != NULL ? 0 : -1);
1121 }
1122
1123 /*
1124  * Returns a reservation level if the given page belongs to a fully populated
1125  * reservation and -1 otherwise.
1126  */
1127 int
1128 vm_reserv_level_iffullpop(vm_page_t m)
1129 {
1130         vm_reserv_t rv;
1131
1132         rv = vm_reserv_from_page(m);
1133         return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
1134 }
1135
1136 /*
1137  * Remove a partially populated reservation from the queue.
1138  */
1139 static void
1140 vm_reserv_dequeue(vm_reserv_t rv)
1141 {
1142
1143         vm_reserv_domain_assert_locked(rv->domain);
1144         vm_reserv_assert_locked(rv);
1145         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
1146             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
1147         KASSERT(rv->inpartpopq,
1148             ("vm_reserv_reclaim: reserv %p's inpartpopq is FALSE", rv));
1149
1150         TAILQ_REMOVE(&vm_rvd[rv->domain].partpop, rv, partpopq);
1151         rv->inpartpopq = FALSE;
1152 }
1153
1154 /*
1155  * Breaks the given partially populated reservation, releasing its free pages
1156  * to the physical memory allocator.
1157  */
1158 static void
1159 vm_reserv_reclaim(vm_reserv_t rv)
1160 {
1161
1162         vm_reserv_assert_locked(rv);
1163         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
1164             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
1165         if (rv->inpartpopq) {
1166                 vm_reserv_domain_lock(rv->domain);
1167                 vm_reserv_dequeue(rv);
1168                 vm_reserv_domain_unlock(rv->domain);
1169         }
1170         vm_reserv_break(rv);
1171         counter_u64_add(vm_reserv_reclaimed, 1);
1172 }
1173
1174 /*
1175  * Breaks a reservation near the head of the partially populated reservation
1176  * queue, releasing its free pages to the physical memory allocator.  Returns
1177  * TRUE if a reservation is broken and FALSE otherwise.
1178  */
1179 bool
1180 vm_reserv_reclaim_inactive(int domain)
1181 {
1182         vm_reserv_t rv;
1183
1184         vm_reserv_domain_lock(domain);
1185         TAILQ_FOREACH(rv, &vm_rvd[domain].partpop, partpopq) {
1186                 /*
1187                  * A locked reservation is likely being updated or reclaimed,
1188                  * so just skip ahead.
1189                  */
1190                 if (rv != &vm_rvd[domain].marker && vm_reserv_trylock(rv)) {
1191                         vm_reserv_dequeue(rv);
1192                         break;
1193                 }
1194         }
1195         vm_reserv_domain_unlock(domain);
1196         if (rv != NULL) {
1197                 vm_reserv_reclaim(rv);
1198                 vm_reserv_unlock(rv);
1199                 return (true);
1200         }
1201         return (false);
1202 }
1203
1204 /*
1205  * Determine whether this reservation has free pages that satisfy the given
1206  * request for contiguous physical memory.  Start searching from the lower
1207  * bound, defined by low_index.
1208  */
1209 static bool
1210 vm_reserv_test_contig(vm_reserv_t rv, u_long npages, vm_paddr_t low,
1211     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1212 {
1213         vm_paddr_t pa, size;
1214         u_long changes;
1215         int bitpos, bits_left, i, hi, lo, n;
1216
1217         vm_reserv_assert_locked(rv);
1218         size = npages << PAGE_SHIFT;
1219         pa = VM_PAGE_TO_PHYS(&rv->pages[0]);
1220         lo = (pa < low) ?
1221             ((low + PAGE_MASK - pa) >> PAGE_SHIFT) : 0;
1222         i = lo / NBPOPMAP;
1223         changes = rv->popmap[i] | ((1UL << (lo % NBPOPMAP)) - 1);
1224         hi = (pa + VM_LEVEL_0_SIZE > high) ?
1225             ((high + PAGE_MASK - pa) >> PAGE_SHIFT) : VM_LEVEL_0_NPAGES;
1226         n = hi / NBPOPMAP;
1227         bits_left = hi % NBPOPMAP;
1228         hi = lo = -1;
1229         for (;;) {
1230                 /*
1231                  * "changes" is a bitmask that marks where a new sequence of
1232                  * 0s or 1s begins in popmap[i], with last bit in popmap[i-1]
1233                  * considered to be 1 if and only if lo == hi.  The bits of
1234                  * popmap[-1] and popmap[NPOPMAP] are considered all 1s.
1235                  */
1236                 changes ^= (changes << 1) | (lo == hi);
1237                 while (changes != 0) {
1238                         /*
1239                          * If the next change marked begins a run of 0s, set
1240                          * lo to mark that position.  Otherwise set hi and
1241                          * look for a satisfactory first page from lo up to hi.
1242                          */
1243                         bitpos = ffsl(changes) - 1;
1244                         changes ^= 1UL << bitpos;
1245                         if (lo == hi) {
1246                                 lo = NBPOPMAP * i + bitpos;
1247                                 continue;
1248                         }
1249                         hi = NBPOPMAP * i + bitpos;
1250                         pa = VM_PAGE_TO_PHYS(&rv->pages[lo]);
1251                         if ((pa & (alignment - 1)) != 0) {
1252                                 /* Skip to next aligned page. */
1253                                 lo += (((pa - 1) | (alignment - 1)) + 1) >>
1254                                     PAGE_SHIFT;
1255                                 if (lo >= VM_LEVEL_0_NPAGES)
1256                                         return (false);
1257                                 pa = VM_PAGE_TO_PHYS(&rv->pages[lo]);
1258                         }
1259                         if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0) {
1260                                 /* Skip to next boundary-matching page. */
1261                                 lo += (((pa - 1) | (boundary - 1)) + 1) >>
1262                                     PAGE_SHIFT;
1263                                 if (lo >= VM_LEVEL_0_NPAGES)
1264                                         return (false);
1265                                 pa = VM_PAGE_TO_PHYS(&rv->pages[lo]);
1266                         }
1267                         if (lo * PAGE_SIZE + size <= hi * PAGE_SIZE)
1268                                 return (true);
1269                         lo = hi;
1270                 }
1271                 if (++i < n)
1272                         changes = rv->popmap[i];
1273                 else if (i == n)
1274                         changes = bits_left == 0 ? -1UL :
1275                             (rv->popmap[n] | (-1UL << bits_left));
1276                 else
1277                         return (false);
1278         }
1279 }
1280
1281 /*
1282  * Searches the partially populated reservation queue for the least recently
1283  * changed reservation with free pages that satisfy the given request for
1284  * contiguous physical memory.  If a satisfactory reservation is found, it is
1285  * broken.  Returns true if a reservation is broken and false otherwise.
1286  */
1287 bool
1288 vm_reserv_reclaim_contig(int domain, u_long npages, vm_paddr_t low,
1289     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1290 {
1291         struct vm_reserv_queue *queue;
1292         vm_paddr_t pa, size;
1293         vm_reserv_t marker, rv, rvn;
1294
1295         if (npages > VM_LEVEL_0_NPAGES - 1)
1296                 return (false);
1297         marker = &vm_rvd[domain].marker;
1298         queue = &vm_rvd[domain].partpop;
1299         size = npages << PAGE_SHIFT;
1300
1301         vm_reserv_domain_scan_lock(domain);
1302         vm_reserv_domain_lock(domain);
1303         TAILQ_FOREACH_SAFE(rv, queue, partpopq, rvn) {
1304                 pa = VM_PAGE_TO_PHYS(&rv->pages[0]);
1305                 if (pa + VM_LEVEL_0_SIZE - size < low) {
1306                         /* This entire reservation is too low; go to next. */
1307                         continue;
1308                 }
1309                 if (pa + size > high) {
1310                         /* This entire reservation is too high; go to next. */
1311                         continue;
1312                 }
1313
1314                 if (vm_reserv_trylock(rv) == 0) {
1315                         TAILQ_INSERT_AFTER(queue, rv, marker, partpopq);
1316                         vm_reserv_domain_unlock(domain);
1317                         vm_reserv_lock(rv);
1318                         if (!rv->inpartpopq ||
1319                             TAILQ_NEXT(rv, partpopq) != marker) {
1320                                 vm_reserv_unlock(rv);
1321                                 vm_reserv_domain_lock(domain);
1322                                 rvn = TAILQ_NEXT(marker, partpopq);
1323                                 TAILQ_REMOVE(queue, marker, partpopq);
1324                                 continue;
1325                         }
1326                         vm_reserv_domain_lock(domain);
1327                         TAILQ_REMOVE(queue, marker, partpopq);
1328                 }
1329                 vm_reserv_domain_unlock(domain);
1330                 if (vm_reserv_test_contig(rv, npages, low, high,
1331                     alignment, boundary)) {
1332                         vm_reserv_domain_scan_unlock(domain);
1333                         vm_reserv_reclaim(rv);
1334                         vm_reserv_unlock(rv);
1335                         return (true);
1336                 }
1337                 vm_reserv_unlock(rv);
1338                 vm_reserv_domain_lock(domain);
1339         }
1340         vm_reserv_domain_unlock(domain);
1341         vm_reserv_domain_scan_unlock(domain);
1342         return (false);
1343 }
1344
1345 /*
1346  * Transfers the reservation underlying the given page to a new object.
1347  *
1348  * The object must be locked.
1349  */
1350 void
1351 vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
1352     vm_pindex_t old_object_offset)
1353 {
1354         vm_reserv_t rv;
1355
1356         VM_OBJECT_ASSERT_WLOCKED(new_object);
1357         rv = vm_reserv_from_page(m);
1358         if (rv->object == old_object) {
1359                 vm_reserv_lock(rv);
1360                 CTR6(KTR_VM,
1361                     "%s: rv %p object %p new %p popcnt %d inpartpop %d",
1362                     __FUNCTION__, rv, rv->object, new_object, rv->popcnt,
1363                     rv->inpartpopq);
1364                 if (rv->object == old_object) {
1365                         vm_reserv_object_lock(old_object);
1366                         rv->object = NULL;
1367                         LIST_REMOVE(rv, objq);
1368                         vm_reserv_object_unlock(old_object);
1369                         vm_reserv_object_lock(new_object);
1370                         rv->object = new_object;
1371                         rv->pindex -= old_object_offset;
1372                         LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
1373                         vm_reserv_object_unlock(new_object);
1374                 }
1375                 vm_reserv_unlock(rv);
1376         }
1377 }
1378
1379 /*
1380  * Returns the size (in bytes) of a reservation of the specified level.
1381  */
1382 int
1383 vm_reserv_size(int level)
1384 {
1385
1386         switch (level) {
1387         case 0:
1388                 return (VM_LEVEL_0_SIZE);
1389         case -1:
1390                 return (PAGE_SIZE);
1391         default:
1392                 return (0);
1393         }
1394 }
1395
1396 /*
1397  * Allocates the virtual and physical memory required by the reservation
1398  * management system's data structures, in particular, the reservation array.
1399  */
1400 vm_paddr_t
1401 vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end)
1402 {
1403         vm_paddr_t new_end, high_water;
1404         size_t size;
1405         int i;
1406
1407         high_water = phys_avail[1];
1408         for (i = 0; i < vm_phys_nsegs; i++) {
1409                 if (vm_phys_segs[i].end > high_water)
1410                         high_water = vm_phys_segs[i].end;
1411         }
1412
1413         /* Skip the first chunk.  It is already accounted for. */
1414         for (i = 2; phys_avail[i + 1] != 0; i += 2) {
1415                 if (phys_avail[i + 1] > high_water)
1416                         high_water = phys_avail[i + 1];
1417         }
1418
1419         /*
1420          * Calculate the size (in bytes) of the reservation array.  Round up
1421          * from "high_water" because every small page is mapped to an element
1422          * in the reservation array based on its physical address.  Thus, the
1423          * number of elements in the reservation array can be greater than the
1424          * number of superpages.
1425          */
1426         size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv);
1427
1428         /*
1429          * Allocate and map the physical memory for the reservation array.  The
1430          * next available virtual address is returned by reference.
1431          */
1432         new_end = end - round_page(size);
1433         vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
1434             VM_PROT_READ | VM_PROT_WRITE);
1435         bzero(vm_reserv_array, size);
1436
1437         /*
1438          * Return the next available physical address.
1439          */
1440         return (new_end);
1441 }
1442
1443 /*
1444  * Returns the superpage containing the given page.
1445  */
1446 vm_page_t
1447 vm_reserv_to_superpage(vm_page_t m)
1448 {
1449         vm_reserv_t rv;
1450
1451         VM_OBJECT_ASSERT_LOCKED(m->object);
1452         rv = vm_reserv_from_page(m);
1453         if (rv->object == m->object && rv->popcnt == VM_LEVEL_0_NPAGES)
1454                 m = rv->pages;
1455         else
1456                 m = NULL;
1457
1458         return (m);
1459 }
1460
1461 #endif  /* VM_NRESERVLEVEL > 0 */