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  * r - vm_reserv_lock
 185  * d - vm_reserv_domain_lock
 186  * o - vm_reserv_object_lock
 187  * c - constant after boot
 188  */
 189 struct vm_reserv {
 190         struct mtx      lock;                   /* reservation lock. */
 191         TAILQ_ENTRY(vm_reserv) partpopq;        /* (d) per-domain queue. */
 192         LIST_ENTRY(vm_reserv) objq;             /* (o, r) object queue */
 193         vm_object_t     object;                 /* (o, r) containing object */
 194         vm_pindex_t     pindex;                 /* (o, r) offset in object */
 195         vm_page_t       pages;                  /* (c) first page  */
 196         uint16_t        domain;                 /* (c) NUMA domain. */
 197         uint16_t        popcnt;                 /* (r) # of pages in use */
 198         int             lasttick;               /* (r) last pop update tick. */
 199         char            inpartpopq;             /* (d) */
 200         popmap_t        popmap[NPOPMAP_MAX];    /* (r) bit vector, used pages */
 201 };
 202
 203 #define vm_reserv_lockptr(rv)           (&(rv)->lock)
 204 #define vm_reserv_assert_locked(rv)                                     \
 205             mtx_assert(vm_reserv_lockptr(rv), MA_OWNED)
 206 #define vm_reserv_lock(rv)              mtx_lock(vm_reserv_lockptr(rv))
 207 #define vm_reserv_trylock(rv)           mtx_trylock(vm_reserv_lockptr(rv))
 208 #define vm_reserv_unlock(rv)            mtx_unlock(vm_reserv_lockptr(rv))
 209
 210 static struct mtx_padalign vm_reserv_domain_locks[MAXMEMDOM];
 211
 212 #define vm_reserv_domain_lockptr(d)     &vm_reserv_domain_locks[(d)]
 213 #define vm_reserv_domain_lock(d)        mtx_lock(vm_reserv_domain_lockptr(d))
 214 #define vm_reserv_domain_unlock(d)      mtx_unlock(vm_reserv_domain_lockptr(d))
 215
 216 /*
 217  * The reservation array
 218  *
 219  * This array is analoguous in function to vm_page_array.  It differs in the
 220  * respect that it may contain a greater number of useful reservation
 221  * structures than there are (physical) superpages.  These "invalid"
 222  * reservation structures exist to trade-off space for time in the
 223  * implementation of vm_reserv_from_page().  Invalid reservation structures are
 224  * distinguishable from "valid" reservation structures by inspecting the
 225  * reservation's "pages" field.  Invalid reservation structures have a NULL
 226  * "pages" field.
 227  *
 228  * vm_reserv_from_page() maps a small (physical) page to an element of this
 229  * array by computing a physical reservation number from the page's physical
 230  * address.  The physical reservation number is used as the array index.
 231  *
 232  * An "active" reservation is a valid reservation structure that has a non-NULL
 233  * "object" field and a non-zero "popcnt" field.  In other words, every active
 234  * reservation belongs to a particular object.  Moreover, every active
 235  * reservation has an entry in the containing object's list of reservations.
 236  */
 237 static vm_reserv_t vm_reserv_array;
 238
 239 /*
 240  * The partially populated reservation queue
 241  *
 242  * This queue enables the fast recovery of an unused free small page from a
 243  * partially populated reservation.  The reservation at the head of this queue
 244  * is the least recently changed, partially populated reservation.
 245  *
 246  * Access to this queue is synchronized by the free page queue lock.
 247  */
 248 static TAILQ_HEAD(, vm_reserv) vm_rvq_partpop[MAXMEMDOM];
 249
 250 static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD, 0, "Reservation Info");
 251
 252 static counter_u64_t vm_reserv_broken = EARLY_COUNTER;
 253 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD,
 254     &vm_reserv_broken, "Cumulative number of broken reservations");
 255
 256 static counter_u64_t vm_reserv_freed = EARLY_COUNTER;
 257 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD,
 258     &vm_reserv_freed, "Cumulative number of freed reservations");
 259
 260 static int sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS);
 261
 262 SYSCTL_PROC(_vm_reserv, OID_AUTO, fullpop, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
 263     sysctl_vm_reserv_fullpop, "I", "Current number of full reservations");
 264
 265 static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);
 266
 267 SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
 268     sysctl_vm_reserv_partpopq, "A", "Partially populated reservation queues");
 269
 270 static counter_u64_t vm_reserv_reclaimed = EARLY_COUNTER;
 271 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
 272     &vm_reserv_reclaimed, "Cumulative number of reclaimed reservations");
 273
 274 /*
 275  * The object lock pool is used to synchronize the rvq.  We can not use a
 276  * pool mutex because it is required before malloc works.
 277  *
 278  * The "hash" function could be made faster without divide and modulo.
 279  */
 280 #define VM_RESERV_OBJ_LOCK_COUNT        MAXCPU
 281
 282 struct mtx_padalign vm_reserv_object_mtx[VM_RESERV_OBJ_LOCK_COUNT];
 283
 284 #define vm_reserv_object_lock_idx(object)                       \
 285             (((uintptr_t)object / sizeof(*object)) % VM_RESERV_OBJ_LOCK_COUNT)
 286 #define vm_reserv_object_lock_ptr(object)                       \
 287             &vm_reserv_object_mtx[vm_reserv_object_lock_idx((object))]
 288 #define vm_reserv_object_lock(object)                           \
 289             mtx_lock(vm_reserv_object_lock_ptr((object)))
 290 #define vm_reserv_object_unlock(object)                         \
 291             mtx_unlock(vm_reserv_object_lock_ptr((object)))
 292
 293 static void             vm_reserv_break(vm_reserv_t rv);
 294 static void             vm_reserv_depopulate(vm_reserv_t rv, int index);
 295 static vm_reserv_t      vm_reserv_from_page(vm_page_t m);
 296 static boolean_t        vm_reserv_has_pindex(vm_reserv_t rv,
 297                             vm_pindex_t pindex);
 298 static void             vm_reserv_populate(vm_reserv_t rv, int index);
 299 static void             vm_reserv_reclaim(vm_reserv_t rv);
 300
 301 /*
 302  * Returns the current number of full reservations.
 303  *
 304  * Since the number of full reservations is computed without acquiring the
 305  * free page queue lock, the returned value may be inexact.
 306  */
 307 static int
 308 sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS)
 309 {
 310         vm_paddr_t paddr;
 311         struct vm_phys_seg *seg;
 312         vm_reserv_t rv;
 313         int fullpop, segind;
 314
 315         fullpop = 0;
 316         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 317                 seg = &vm_phys_segs[segind];
 318                 paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
 319                 while (paddr + VM_LEVEL_0_SIZE <= seg->end) {
 320                         rv = &vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT];
 321                         fullpop += rv->popcnt == VM_LEVEL_0_NPAGES;
 322                         paddr += VM_LEVEL_0_SIZE;
 323                 }
 324         }
 325         return (sysctl_handle_int(oidp, &fullpop, 0, req));
 326 }
 327
 328 /*
 329  * Describes the current state of the partially populated reservation queue.
 330  */
 331 static int
 332 sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
 333 {
 334         struct sbuf sbuf;
 335         vm_reserv_t rv;
 336         int counter, error, domain, level, unused_pages;
 337
 338         error = sysctl_wire_old_buffer(req, 0);
 339         if (error != 0)
 340                 return (error);
 341         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 342         sbuf_printf(&sbuf, "\nDOMAIN    LEVEL     SIZE  NUMBER\n\n");
 343         for (domain = 0; domain < vm_ndomains; domain++) {
 344                 for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
 345                         counter = 0;
 346                         unused_pages = 0;
 347                         vm_reserv_domain_lock(domain);
 348                         TAILQ_FOREACH(rv, &vm_rvq_partpop[domain], partpopq) {
 349                                 counter++;
 350                                 unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
 351                         }
 352                         vm_reserv_domain_unlock(domain);
 353                         sbuf_printf(&sbuf, "%6d, %7d, %6dK, %6d\n",
 354                             domain, level,
 355                             unused_pages * ((int)PAGE_SIZE / 1024), counter);
 356                 }
 357         }
 358         error = sbuf_finish(&sbuf);
 359         sbuf_delete(&sbuf);
 360         return (error);
 361 }
 362
 363 /*
 364  * Remove a reservation from the object's objq.
 365  */
 366 static void
 367 vm_reserv_remove(vm_reserv_t rv)
 368 {
 369         vm_object_t object;
 370
 371         vm_reserv_assert_locked(rv);
 372         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
 373             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
 374         KASSERT(rv->object != NULL,
 375             ("vm_reserv_remove: reserv %p is free", rv));
 376         KASSERT(!rv->inpartpopq,
 377             ("vm_reserv_remove: reserv %p's inpartpopq is TRUE", rv));
 378         object = rv->object;
 379         vm_reserv_object_lock(object);
 380         LIST_REMOVE(rv, objq);
 381         rv->object = NULL;
 382         vm_reserv_object_unlock(object);
 383 }
 384
 385 /*
 386  * Insert a new reservation into the object's objq.
 387  */
 388 static void
 389 vm_reserv_insert(vm_reserv_t rv, vm_object_t object, vm_pindex_t pindex)
 390 {
 391         int i;
 392
 393         vm_reserv_assert_locked(rv);
 394         CTR6(KTR_VM,
 395             "%s: rv %p(%p) object %p new %p popcnt %d",
 396             __FUNCTION__, rv, rv->pages, rv->object, object,
 397            rv->popcnt);
 398         KASSERT(rv->object == NULL,
 399             ("vm_reserv_insert: reserv %p isn't free", rv));
 400         KASSERT(rv->popcnt == 0,
 401             ("vm_reserv_insert: reserv %p's popcnt is corrupted", rv));
 402         KASSERT(!rv->inpartpopq,
 403             ("vm_reserv_insert: reserv %p's inpartpopq is TRUE", rv));
 404         for (i = 0; i < NPOPMAP; i++)
 405                 KASSERT(rv->popmap[i] == 0,
 406                     ("vm_reserv_insert: reserv %p's popmap is corrupted", rv));
 407         vm_reserv_object_lock(object);
 408         rv->pindex = pindex;
 409         rv->object = object;
 410         rv->lasttick = ticks;
 411         LIST_INSERT_HEAD(&object->rvq, rv, objq);
 412         vm_reserv_object_unlock(object);
 413 }
 414
 415 /*
 416  * Reduces the given reservation's population count.  If the population count
 417  * becomes zero, the reservation is destroyed.  Additionally, moves the
 418  * reservation to the tail of the partially populated reservation queue if the
 419  * population count is non-zero.
 420  */
 421 static void
 422 vm_reserv_depopulate(vm_reserv_t rv, int index)
 423 {
 424         struct vm_domain *vmd;
 425
 426         vm_reserv_assert_locked(rv);
 427         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
 428             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
 429         KASSERT(rv->object != NULL,
 430             ("vm_reserv_depopulate: reserv %p is free", rv));
 431         KASSERT(popmap_is_set(rv->popmap, index),
 432             ("vm_reserv_depopulate: reserv %p's popmap[%d] is clear", rv,
 433             index));
 434         KASSERT(rv->popcnt > 0,
 435             ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
 436         KASSERT(rv->domain < vm_ndomains,
 437             ("vm_reserv_depopulate: reserv %p's domain is corrupted %d",
 438             rv, rv->domain));
 439         if (rv->popcnt == VM_LEVEL_0_NPAGES) {
 440                 KASSERT(rv->pages->psind == 1,
 441                     ("vm_reserv_depopulate: reserv %p is already demoted",
 442                     rv));
 443                 rv->pages->psind = 0;
 444         }
 445         popmap_clear(rv->popmap, index);
 446         rv->popcnt--;
 447         if ((unsigned)(ticks - rv->lasttick) >= PARTPOPSLOP ||
 448             rv->popcnt == 0) {
 449                 vm_reserv_domain_lock(rv->domain);
 450                 if (rv->inpartpopq) {
 451                         TAILQ_REMOVE(&vm_rvq_partpop[rv->domain], rv, partpopq);
 452                         rv->inpartpopq = FALSE;
 453                 }
 454                 if (rv->popcnt != 0) {
 455                         rv->inpartpopq = TRUE;
 456                         TAILQ_INSERT_TAIL(&vm_rvq_partpop[rv->domain], rv, partpopq);
 457                 }
 458                 vm_reserv_domain_unlock(rv->domain);
 459                 rv->lasttick = ticks;
 460         }
 461         vmd = VM_DOMAIN(rv->domain);
 462         if (rv->popcnt == 0) {
 463                 vm_reserv_remove(rv);
 464                 vm_domain_free_lock(vmd);
 465                 vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
 466                 vm_domain_free_unlock(vmd);
 467                 counter_u64_add(vm_reserv_freed, 1);
 468         }
 469         vm_domain_freecnt_inc(vmd, 1);
 470 }
 471
 472 /*
 473  * Returns the reservation to which the given page might belong.
 474  */
 475 static __inline vm_reserv_t
 476 vm_reserv_from_page(vm_page_t m)
 477 {
 478
 479         return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
 480 }
 481
 482 /*
 483  * Returns an existing reservation or NULL and initialized successor pointer.
 484  */
 485 static vm_reserv_t
 486 vm_reserv_from_object(vm_object_t object, vm_pindex_t pindex,
 487     vm_page_t mpred, vm_page_t *msuccp)
 488 {
 489         vm_reserv_t rv;
 490         vm_page_t msucc;
 491
 492         msucc = NULL;
 493         if (mpred != NULL) {
 494                 KASSERT(mpred->object == object,
 495                     ("vm_reserv_from_object: object doesn't contain mpred"));
 496                 KASSERT(mpred->pindex < pindex,
 497                     ("vm_reserv_from_object: mpred doesn't precede pindex"));
 498                 rv = vm_reserv_from_page(mpred);
 499                 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
 500                         goto found;
 501                 msucc = TAILQ_NEXT(mpred, listq);
 502         } else
 503                 msucc = TAILQ_FIRST(&object->memq);
 504         if (msucc != NULL) {
 505                 KASSERT(msucc->pindex > pindex,
 506                     ("vm_reserv_from_object: msucc doesn't succeed pindex"));
 507                 rv = vm_reserv_from_page(msucc);
 508                 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
 509                         goto found;
 510         }
 511         rv = NULL;
 512
 513 found:
 514         *msuccp = msucc;
 515
 516         return (rv);
 517 }
 518
 519 /*
 520  * Returns TRUE if the given reservation contains the given page index and
 521  * FALSE otherwise.
 522  */
 523 static __inline boolean_t
 524 vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
 525 {
 526
 527         return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
 528 }
 529
 530 /*
 531  * Increases the given reservation's population count.  Moves the reservation
 532  * to the tail of the partially populated reservation queue.
 533  *
 534  * The free page queue must be locked.
 535  */
 536 static void
 537 vm_reserv_populate(vm_reserv_t rv, int index)
 538 {
 539
 540         vm_reserv_assert_locked(rv);
 541         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
 542             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
 543         KASSERT(rv->object != NULL,
 544             ("vm_reserv_populate: reserv %p is free", rv));
 545         KASSERT(popmap_is_clear(rv->popmap, index),
 546             ("vm_reserv_populate: reserv %p's popmap[%d] is set", rv,
 547             index));
 548         KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
 549             ("vm_reserv_populate: reserv %p is already full", rv));
 550         KASSERT(rv->pages->psind == 0,
 551             ("vm_reserv_populate: reserv %p is already promoted", rv));
 552         KASSERT(rv->domain < vm_ndomains,
 553             ("vm_reserv_populate: reserv %p's domain is corrupted %d",
 554             rv, rv->domain));
 555         popmap_set(rv->popmap, index);
 556         rv->popcnt++;
 557         if ((unsigned)(ticks - rv->lasttick) < PARTPOPSLOP &&
 558             rv->inpartpopq && rv->popcnt != VM_LEVEL_0_NPAGES)
 559                 return;
 560         rv->lasttick = ticks;
 561         vm_reserv_domain_lock(rv->domain);
 562         if (rv->inpartpopq) {
 563                 TAILQ_REMOVE(&vm_rvq_partpop[rv->domain], rv, partpopq);
 564                 rv->inpartpopq = FALSE;
 565         }
 566         if (rv->popcnt < VM_LEVEL_0_NPAGES) {
 567                 rv->inpartpopq = TRUE;
 568                 TAILQ_INSERT_TAIL(&vm_rvq_partpop[rv->domain], rv, partpopq);
 569         } else {
 570                 KASSERT(rv->pages->psind == 0,
 571                     ("vm_reserv_populate: reserv %p is already promoted",
 572                     rv));
 573                 rv->pages->psind = 1;
 574         }
 575         vm_reserv_domain_unlock(rv->domain);
 576 }
 577
 578 /*
 579  * Attempts to allocate a contiguous set of physical pages from existing
 580  * reservations.  See vm_reserv_alloc_contig() for a description of the
 581  * function's parameters.
 582  *
 583  * The page "mpred" must immediately precede the offset "pindex" within the
 584  * specified object.
 585  *
 586  * The object must be locked.
 587  */
 588 vm_page_t
 589 vm_reserv_extend_contig(int req, vm_object_t object, vm_pindex_t pindex,
 590     int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
 591     u_long alignment, vm_paddr_t boundary, vm_page_t mpred)
 592 {
 593         struct vm_domain *vmd;
 594         vm_paddr_t pa, size;
 595         vm_page_t m, msucc;
 596         vm_reserv_t rv;
 597         int i, index;
 598
 599         VM_OBJECT_ASSERT_WLOCKED(object);
 600         KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
 601
 602         /*
 603          * Is a reservation fundamentally impossible?
 604          */
 605         if (pindex < VM_RESERV_INDEX(object, pindex) ||
 606             pindex + npages > object->size || object->resident_page_count == 0)
 607                 return (NULL);
 608
 609         /*
 610          * All reservations of a particular size have the same alignment.
 611          * Assuming that the first page is allocated from a reservation, the
 612          * least significant bits of its physical address can be determined
 613          * from its offset from the beginning of the reservation and the size
 614          * of the reservation.
 615          *
 616          * Could the specified index within a reservation of the smallest
 617          * possible size satisfy the alignment and boundary requirements?
 618          */
 619         pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
 620         if ((pa & (alignment - 1)) != 0)
 621                 return (NULL);
 622         size = npages << PAGE_SHIFT;
 623         if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
 624                 return (NULL);
 625
 626         /*
 627          * Look for an existing reservation.
 628          */
 629         rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
 630         if (rv == NULL)
 631                 return (NULL);
 632         KASSERT(object != kernel_object || rv->domain == domain,
 633             ("vm_reserv_extend_contig: Domain mismatch from reservation."));
 634         index = VM_RESERV_INDEX(object, pindex);
 635         /* Does the allocation fit within the reservation? */
 636         if (index + npages > VM_LEVEL_0_NPAGES)
 637                 return (NULL);
 638         domain = rv->domain;
 639         vmd = VM_DOMAIN(domain);
 640         vm_reserv_lock(rv);
 641         if (rv->object != object)
 642                 goto out;
 643         m = &rv->pages[index];
 644         pa = VM_PAGE_TO_PHYS(m);
 645         if (pa < low || pa + size > high || (pa & (alignment - 1)) != 0 ||
 646             ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
 647                 goto out;
 648         /* Handle vm_page_rename(m, new_object, ...). */
 649         for (i = 0; i < npages; i++) {
 650                 if (popmap_is_set(rv->popmap, index + i))
 651                         goto out;
 652         }
 653         if (!vm_domain_allocate(vmd, req, npages))
 654                 goto out;
 655         for (i = 0; i < npages; i++)
 656                 vm_reserv_populate(rv, index + i);
 657         vm_reserv_unlock(rv);
 658         return (m);
 659
 660 out:
 661         vm_reserv_unlock(rv);
 662         return (NULL);
 663 }
 664
 665 /*
 666  * Allocates a contiguous set of physical pages of the given size "npages"
 667  * from newly created reservations.  All of the physical pages
 668  * must be at or above the given physical address "low" and below the given
 669  * physical address "high".  The given value "alignment" determines the
 670  * alignment of the first physical page in the set.  If the given value
 671  * "boundary" is non-zero, then the set of physical pages cannot cross any
 672  * physical address boundary that is a multiple of that value.  Both
 673  * "alignment" and "boundary" must be a power of two.
 674  *
 675  * Callers should first invoke vm_reserv_extend_contig() to attempt an
 676  * allocation from existing reservations.
 677  *
 678  * The page "mpred" must immediately precede the offset "pindex" within the
 679  * specified object.
 680  *
 681  * The object and free page queue must be locked.
 682  */
 683 vm_page_t
 684 vm_reserv_alloc_contig(int req, vm_object_t object, vm_pindex_t pindex, int domain,
 685     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
 686     vm_paddr_t boundary, vm_page_t mpred)
 687 {
 688         struct vm_domain *vmd;
 689         vm_paddr_t pa, size;
 690         vm_page_t m, m_ret, msucc;
 691         vm_pindex_t first, leftcap, rightcap;
 692         vm_reserv_t rv;
 693         u_long allocpages, maxpages, minpages;
 694         int i, index, n;
 695
 696         VM_OBJECT_ASSERT_WLOCKED(object);
 697         KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
 698
 699         /*
 700          * Is a reservation fundamentally impossible?
 701          */
 702         if (pindex < VM_RESERV_INDEX(object, pindex) ||
 703             pindex + npages > object->size)
 704                 return (NULL);
 705
 706         /*
 707          * All reservations of a particular size have the same alignment.
 708          * Assuming that the first page is allocated from a reservation, the
 709          * least significant bits of its physical address can be determined
 710          * from its offset from the beginning of the reservation and the size
 711          * of the reservation.
 712          *
 713          * Could the specified index within a reservation of the smallest
 714          * possible size satisfy the alignment and boundary requirements?
 715          */
 716         pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
 717         if ((pa & (alignment - 1)) != 0)
 718                 return (NULL);
 719         size = npages << PAGE_SHIFT;
 720         if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
 721                 return (NULL);
 722
 723         /*
 724          * Callers should've extended an existing reservation prior to
 725          * calling this function.  If a reservation exists it is
 726          * incompatible with the allocation.
 727          */
 728         rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
 729         if (rv != NULL)
 730                 return (NULL);
 731
 732         /*
 733          * Could at least one reservation fit between the first index to the
 734          * left that can be used ("leftcap") and the first index to the right
 735          * that cannot be used ("rightcap")?
 736          *
 737          * We must synchronize with the reserv object lock to protect the
 738          * pindex/object of the resulting reservations against rename while
 739          * we are inspecting.
 740          */
 741         first = pindex - VM_RESERV_INDEX(object, pindex);
 742         minpages = VM_RESERV_INDEX(object, pindex) + npages;
 743         maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES);
 744         allocpages = maxpages;
 745         vm_reserv_object_lock(object);
 746         if (mpred != NULL) {
 747                 if ((rv = vm_reserv_from_page(mpred))->object != object)
 748                         leftcap = mpred->pindex + 1;
 749                 else
 750                         leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
 751                 if (leftcap > first) {
 752                         vm_reserv_object_unlock(object);
 753                         return (NULL);
 754                 }
 755         }
 756         if (msucc != NULL) {
 757                 if ((rv = vm_reserv_from_page(msucc))->object != object)
 758                         rightcap = msucc->pindex;
 759                 else
 760                         rightcap = rv->pindex;
 761                 if (first + maxpages > rightcap) {
 762                         if (maxpages == VM_LEVEL_0_NPAGES) {
 763                                 vm_reserv_object_unlock(object);
 764                                 return (NULL);
 765                         }
 766
 767                         /*
 768                          * At least one reservation will fit between "leftcap"
 769                          * and "rightcap".  However, a reservation for the
 770                          * last of the requested pages will not fit.  Reduce
 771                          * the size of the upcoming allocation accordingly.
 772                          */
 773                         allocpages = minpages;
 774                 }
 775         }
 776         vm_reserv_object_unlock(object);
 777
 778         /*
 779          * Would the last new reservation extend past the end of the object?
 780          */
 781         if (first + maxpages > object->size) {
 782                 /*
 783                  * Don't allocate the last new reservation if the object is a
 784                  * vnode or backed by another object that is a vnode.
 785                  */
 786                 if (object->type == OBJT_VNODE ||
 787                     (object->backing_object != NULL &&
 788                     object->backing_object->type == OBJT_VNODE)) {
 789                         if (maxpages == VM_LEVEL_0_NPAGES)
 790                                 return (NULL);
 791                         allocpages = minpages;
 792                 }
 793                 /* Speculate that the object may grow. */
 794         }
 795
 796         /*
 797          * Allocate the physical pages.  The alignment and boundary specified
 798          * for this allocation may be different from the alignment and
 799          * boundary specified for the requested pages.  For instance, the
 800          * specified index may not be the first page within the first new
 801          * reservation.
 802          */
 803         m = NULL;
 804         vmd = VM_DOMAIN(domain);
 805         if (vm_domain_allocate(vmd, req, npages)) {
 806                 vm_domain_free_lock(vmd);
 807                 m = vm_phys_alloc_contig(domain, allocpages, low, high,
 808                     ulmax(alignment, VM_LEVEL_0_SIZE),
 809                     boundary > VM_LEVEL_0_SIZE ? boundary : 0);
 810                 vm_domain_free_unlock(vmd);
 811                 if (m == NULL) {
 812                         vm_domain_freecnt_inc(vmd, npages);
 813                         return (NULL);
 814                 }
 815         } else
 816                 return (NULL);
 817         KASSERT(vm_phys_domain(m) == domain,
 818             ("vm_reserv_alloc_contig: Page domain does not match requested."));
 819
 820         /*
 821          * The allocated physical pages always begin at a reservation
 822          * boundary, but they do not always end at a reservation boundary.
 823          * Initialize every reservation that is completely covered by the
 824          * allocated physical pages.
 825          */
 826         m_ret = NULL;
 827         index = VM_RESERV_INDEX(object, pindex);
 828         do {
 829                 rv = vm_reserv_from_page(m);
 830                 KASSERT(rv->pages == m,
 831                     ("vm_reserv_alloc_contig: reserv %p's pages is corrupted",
 832                     rv));
 833                 vm_reserv_lock(rv);
 834                 vm_reserv_insert(rv, object, first);
 835                 n = ulmin(VM_LEVEL_0_NPAGES - index, npages);
 836                 for (i = 0; i < n; i++)
 837                         vm_reserv_populate(rv, index + i);
 838                 npages -= n;
 839                 if (m_ret == NULL) {
 840                         m_ret = &rv->pages[index];
 841                         index = 0;
 842                 }
 843                 vm_reserv_unlock(rv);
 844                 m += VM_LEVEL_0_NPAGES;
 845                 first += VM_LEVEL_0_NPAGES;
 846                 allocpages -= VM_LEVEL_0_NPAGES;
 847         } while (allocpages >= VM_LEVEL_0_NPAGES);
 848         return (m_ret);
 849 }
 850
 851 /*
 852  * Attempts to extend an existing reservation and allocate the page to the
 853  * object.
 854  *
 855  * The page "mpred" must immediately precede the offset "pindex" within the
 856  * specified object.
 857  *
 858  * The object must be locked.
 859  */
 860 vm_page_t
 861 vm_reserv_extend(int req, vm_object_t object, vm_pindex_t pindex, int domain,
 862     vm_page_t mpred)
 863 {
 864         struct vm_domain *vmd;
 865         vm_page_t m, msucc;
 866         vm_reserv_t rv;
 867         int index;
 868
 869         VM_OBJECT_ASSERT_WLOCKED(object);
 870
 871         /*
 872          * Could a reservation currently exist?
 873          */
 874         if (pindex < VM_RESERV_INDEX(object, pindex) ||
 875             pindex >= object->size || object->resident_page_count == 0)
 876                 return (NULL);
 877
 878         /*
 879          * Look for an existing reservation.
 880          */
 881         rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
 882         if (rv == NULL)
 883                 return (NULL);
 884
 885         KASSERT(object != kernel_object || rv->domain == domain,
 886             ("vm_reserv_extend: Domain mismatch from reservation."));
 887         domain = rv->domain;
 888         vmd = VM_DOMAIN(domain);
 889         index = VM_RESERV_INDEX(object, pindex);
 890         m = &rv->pages[index];
 891         vm_reserv_lock(rv);
 892         /* Handle reclaim race. */
 893         if (rv->object != object ||
 894             /* Handle vm_page_rename(m, new_object, ...). */
 895             popmap_is_set(rv->popmap, index)) {
 896                 m = NULL;
 897                 goto out;
 898         }
 899         if (vm_domain_allocate(vmd, req, 1) == 0)
 900                 m = NULL;
 901         else
 902                 vm_reserv_populate(rv, index);
 903 out:
 904         vm_reserv_unlock(rv);
 905
 906         return (m);
 907 }
 908
 909 /*
 910  * Attempts to allocate a new reservation for the object, and allocates a
 911  * page from that reservation.  Callers should first invoke vm_reserv_extend()
 912  * to attempt an allocation from an existing reservation.
 913  *
 914  * The page "mpred" must immediately precede the offset "pindex" within the
 915  * specified object.
 916  *
 917  * The object and free page queue must be locked.
 918  */
 919 vm_page_t
 920 vm_reserv_alloc_page(int req, vm_object_t object, vm_pindex_t pindex, int domain,
 921     vm_page_t mpred)
 922 {
 923         struct vm_domain *vmd;
 924         vm_page_t m, msucc;
 925         vm_pindex_t first, leftcap, rightcap;
 926         vm_reserv_t rv;
 927         int index;
 928
 929         VM_OBJECT_ASSERT_WLOCKED(object);
 930
 931         /*
 932          * Is a reservation fundamentally impossible?
 933          */
 934         if (pindex < VM_RESERV_INDEX(object, pindex) ||
 935             pindex >= object->size)
 936                 return (NULL);
 937
 938         /*
 939          * Callers should've extended an existing reservation prior to
 940          * calling this function.  If a reservation exists it is
 941          * incompatible with the allocation.
 942          */
 943         rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
 944         if (rv != NULL)
 945                 return (NULL);
 946
 947         /*
 948          * Could a reservation fit between the first index to the left that
 949          * can be used and the first index to the right that cannot be used?
 950          *
 951          * We must synchronize with the reserv object lock to protect the
 952          * pindex/object of the resulting reservations against rename while
 953          * we are inspecting.
 954          */
 955         first = pindex - VM_RESERV_INDEX(object, pindex);
 956         vm_reserv_object_lock(object);
 957         if (mpred != NULL) {
 958                 if ((rv = vm_reserv_from_page(mpred))->object != object)
 959                         leftcap = mpred->pindex + 1;
 960                 else
 961                         leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
 962                 if (leftcap > first) {
 963                         vm_reserv_object_unlock(object);
 964                         return (NULL);
 965                 }
 966         }
 967         if (msucc != NULL) {
 968                 if ((rv = vm_reserv_from_page(msucc))->object != object)
 969                         rightcap = msucc->pindex;
 970                 else
 971                         rightcap = rv->pindex;
 972                 if (first + VM_LEVEL_0_NPAGES > rightcap) {
 973                         vm_reserv_object_unlock(object);
 974                         return (NULL);
 975                 }
 976         }
 977         vm_reserv_object_unlock(object);
 978
 979         /*
 980          * Would a new reservation extend past the end of the object?
 981          */
 982         if (first + VM_LEVEL_0_NPAGES > object->size) {
 983                 /*
 984                  * Don't allocate a new reservation if the object is a vnode or
 985                  * backed by another object that is a vnode.
 986                  */
 987                 if (object->type == OBJT_VNODE ||
 988                     (object->backing_object != NULL &&
 989                     object->backing_object->type == OBJT_VNODE))
 990                         return (NULL);
 991                 /* Speculate that the object may grow. */
 992         }
 993
 994         /*
 995          * Allocate and populate the new reservation.
 996          */
 997         m = NULL;
 998         vmd = VM_DOMAIN(domain);
 999         if (vm_domain_allocate(vmd, req, 1)) {
1000                 vm_domain_free_lock(vmd);
1001                 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT,
1002                     VM_LEVEL_0_ORDER);
1003                 vm_domain_free_unlock(vmd);
1004                 if (m == NULL) {
1005                         vm_domain_freecnt_inc(vmd, 1);
1006                         return (NULL);
1007                 }
1008         } else
1009                 return (NULL);
1010         rv = vm_reserv_from_page(m);
1011         vm_reserv_lock(rv);
1012         KASSERT(rv->pages == m,
1013             ("vm_reserv_alloc_page: reserv %p's pages is corrupted", rv));
1014         vm_reserv_insert(rv, object, first);
1015         index = VM_RESERV_INDEX(object, pindex);
1016         vm_reserv_populate(rv, index);
1017         vm_reserv_unlock(rv);
1018
1019         return (&rv->pages[index]);
1020 }
1021
1022 /*
1023  * Breaks the given reservation.  All free pages in the reservation
1024  * are returned to the physical memory allocator.  The reservation's
1025  * population count and map are reset to their initial state.
1026  *
1027  * The given reservation must not be in the partially populated reservation
1028  * queue.  The free page queue lock must be held.
1029  */
1030 static void
1031 vm_reserv_break(vm_reserv_t rv)
1032 {
1033         u_long changes;
1034         int bitpos, hi, i, lo;
1035
1036         vm_reserv_assert_locked(rv);
1037         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
1038             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
1039         vm_reserv_remove(rv);
1040         rv->pages->psind = 0;
1041         hi = lo = -1;
1042         for (i = 0; i <= NPOPMAP; i++) {
1043                 /*
1044                  * "changes" is a bitmask that marks where a new sequence of
1045                  * 0s or 1s begins in popmap[i], with last bit in popmap[i-1]
1046                  * considered to be 1 if and only if lo == hi.  The bits of
1047                  * popmap[-1] and popmap[NPOPMAP] are considered all 1s.
1048                  */
1049                 if (i == NPOPMAP)
1050                         changes = lo != hi;
1051                 else {
1052                         changes = rv->popmap[i];
1053                         changes ^= (changes << 1) | (lo == hi);
1054                         rv->popmap[i] = 0;
1055                 }
1056                 while (changes != 0) {
1057                         /*
1058                          * If the next change marked begins a run of 0s, set
1059                          * lo to mark that position.  Otherwise set hi and
1060                          * free pages from lo up to hi.
1061                          */
1062                         bitpos = ffsl(changes) - 1;
1063                         changes ^= 1UL << bitpos;
1064                         if (lo == hi)
1065                                 lo = NBPOPMAP * i + bitpos;
1066                         else {
1067                                 hi = NBPOPMAP * i + bitpos;
1068                                 vm_domain_free_lock(VM_DOMAIN(rv->domain));
1069                                 vm_phys_enqueue_contig(&rv->pages[lo], hi - lo);
1070                                 vm_domain_free_unlock(VM_DOMAIN(rv->domain));
1071                                 lo = hi;
1072                         }
1073                 }
1074         }
1075         rv->popcnt = 0;
1076         counter_u64_add(vm_reserv_broken, 1);
1077 }
1078
1079 /*
1080  * Breaks all reservations belonging to the given object.
1081  */
1082 void
1083 vm_reserv_break_all(vm_object_t object)
1084 {
1085         vm_reserv_t rv;
1086
1087         /*
1088          * This access of object->rvq is unsynchronized so that the
1089          * object rvq lock can nest after the domain_free lock.  We
1090          * must check for races in the results.  However, the object
1091          * lock prevents new additions, so we are guaranteed that when
1092          * it returns NULL the object is properly empty.
1093          */
1094         while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
1095                 vm_reserv_lock(rv);
1096                 /* Reclaim race. */
1097                 if (rv->object != object) {
1098                         vm_reserv_unlock(rv);
1099                         continue;
1100                 }
1101                 vm_reserv_domain_lock(rv->domain);
1102                 if (rv->inpartpopq) {
1103                         TAILQ_REMOVE(&vm_rvq_partpop[rv->domain], rv, partpopq);
1104                         rv->inpartpopq = FALSE;
1105                 }
1106                 vm_reserv_domain_unlock(rv->domain);
1107                 vm_reserv_break(rv);
1108                 vm_reserv_unlock(rv);
1109         }
1110 }
1111
1112 /*
1113  * Frees the given page if it belongs to a reservation.  Returns TRUE if the
1114  * page is freed and FALSE otherwise.
1115  *
1116  * The free page queue lock must be held.
1117  */
1118 boolean_t
1119 vm_reserv_free_page(vm_page_t m)
1120 {
1121         vm_reserv_t rv;
1122         boolean_t ret;
1123
1124         rv = vm_reserv_from_page(m);
1125         if (rv->object == NULL)
1126                 return (FALSE);
1127         vm_reserv_lock(rv);
1128         /* Re-validate after lock. */
1129         if (rv->object != NULL) {
1130                 vm_reserv_depopulate(rv, m - rv->pages);
1131                 ret = TRUE;
1132         } else
1133                 ret = FALSE;
1134         vm_reserv_unlock(rv);
1135
1136         return (ret);
1137 }
1138
1139 /*
1140  * Initializes the reservation management system.  Specifically, initializes
1141  * the reservation array.
1142  *
1143  * Requires that vm_page_array and first_page are initialized!
1144  */
1145 void
1146 vm_reserv_init(void)
1147 {
1148         vm_paddr_t paddr;
1149         struct vm_phys_seg *seg;
1150         struct vm_reserv *rv;
1151         int i, segind;
1152
1153         /*
1154          * Initialize the reservation array.  Specifically, initialize the
1155          * "pages" field for every element that has an underlying superpage.
1156          */
1157         for (segind = 0; segind < vm_phys_nsegs; segind++) {
1158                 seg = &vm_phys_segs[segind];
1159                 paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
1160                 while (paddr + VM_LEVEL_0_SIZE <= seg->end) {
1161                         rv = &vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT];
1162                         rv->pages = PHYS_TO_VM_PAGE(paddr);
1163                         rv->domain = seg->domain;
1164                         mtx_init(&rv->lock, "vm reserv", NULL, MTX_DEF);
1165                         paddr += VM_LEVEL_0_SIZE;
1166                 }
1167         }
1168         for (i = 0; i < MAXMEMDOM; i++) {
1169                 mtx_init(&vm_reserv_domain_locks[i], "VM reserv domain", NULL,
1170                     MTX_DEF);
1171                 TAILQ_INIT(&vm_rvq_partpop[i]);
1172         }
1173
1174         for (i = 0; i < VM_RESERV_OBJ_LOCK_COUNT; i++)
1175                 mtx_init(&vm_reserv_object_mtx[i], "resv obj lock", NULL,
1176                     MTX_DEF);
1177 }
1178
1179 /*
1180  * Returns true if the given page belongs to a reservation and that page is
1181  * free.  Otherwise, returns false.
1182  */
1183 bool
1184 vm_reserv_is_page_free(vm_page_t m)
1185 {
1186         vm_reserv_t rv;
1187
1188         rv = vm_reserv_from_page(m);
1189         if (rv->object == NULL)
1190                 return (false);
1191         return (popmap_is_clear(rv->popmap, m - rv->pages));
1192 }
1193
1194 /*
1195  * If the given page belongs to a reservation, returns the level of that
1196  * reservation.  Otherwise, returns -1.
1197  */
1198 int
1199 vm_reserv_level(vm_page_t m)
1200 {
1201         vm_reserv_t rv;
1202
1203         rv = vm_reserv_from_page(m);
1204         return (rv->object != NULL ? 0 : -1);
1205 }
1206
1207 /*
1208  * Returns a reservation level if the given page belongs to a fully populated
1209  * reservation and -1 otherwise.
1210  */
1211 int
1212 vm_reserv_level_iffullpop(vm_page_t m)
1213 {
1214         vm_reserv_t rv;
1215
1216         rv = vm_reserv_from_page(m);
1217         return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
1218 }
1219
1220 /*
1221  * Breaks the given partially populated reservation, releasing its free pages
1222  * to the physical memory allocator.
1223  *
1224  * The free page queue lock must be held.
1225  */
1226 static void
1227 vm_reserv_reclaim(vm_reserv_t rv)
1228 {
1229
1230         vm_reserv_assert_locked(rv);
1231         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
1232             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
1233         vm_reserv_domain_lock(rv->domain);
1234         KASSERT(rv->inpartpopq,
1235             ("vm_reserv_reclaim: reserv %p's inpartpopq is FALSE", rv));
1236         KASSERT(rv->domain < vm_ndomains,
1237             ("vm_reserv_reclaim: reserv %p's domain is corrupted %d",
1238             rv, rv->domain));
1239         TAILQ_REMOVE(&vm_rvq_partpop[rv->domain], rv, partpopq);
1240         rv->inpartpopq = FALSE;
1241         vm_reserv_domain_unlock(rv->domain);
1242         vm_reserv_break(rv);
1243         counter_u64_add(vm_reserv_reclaimed, 1);
1244 }
1245
1246 /*
1247  * Breaks the reservation at the head of the partially populated reservation
1248  * queue, releasing its free pages to the physical memory allocator.  Returns
1249  * TRUE if a reservation is broken and FALSE otherwise.
1250  *
1251  * The free page queue lock must be held.
1252  */
1253 boolean_t
1254 vm_reserv_reclaim_inactive(int domain)
1255 {
1256         vm_reserv_t rv;
1257
1258         while ((rv = TAILQ_FIRST(&vm_rvq_partpop[domain])) != NULL) {
1259                 vm_reserv_lock(rv);
1260                 if (rv != TAILQ_FIRST(&vm_rvq_partpop[domain])) {
1261                         vm_reserv_unlock(rv);
1262                         continue;
1263                 }
1264                 vm_reserv_reclaim(rv);
1265                 vm_reserv_unlock(rv);
1266                 return (TRUE);
1267         }
1268         return (FALSE);
1269 }
1270
1271 /*
1272  * Searches the partially populated reservation queue for the least recently
1273  * changed reservation with free pages that satisfy the given request for
1274  * contiguous physical memory.  If a satisfactory reservation is found, it is
1275  * broken.  Returns TRUE if a reservation is broken and FALSE otherwise.
1276  *
1277  * The free page queue lock must be held.
1278  */
1279 boolean_t
1280 vm_reserv_reclaim_contig(int domain, u_long npages, vm_paddr_t low,
1281     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1282 {
1283         vm_paddr_t pa, size;
1284         vm_reserv_t rv, rvn;
1285         int hi, i, lo, low_index, next_free;
1286
1287         if (npages > VM_LEVEL_0_NPAGES - 1)
1288                 return (FALSE);
1289         size = npages << PAGE_SHIFT;
1290         vm_reserv_domain_lock(domain);
1291 again:
1292         for (rv = TAILQ_FIRST(&vm_rvq_partpop[domain]); rv != NULL; rv = rvn) {
1293                 rvn = TAILQ_NEXT(rv, partpopq);
1294                 pa = VM_PAGE_TO_PHYS(&rv->pages[VM_LEVEL_0_NPAGES - 1]);
1295                 if (pa + PAGE_SIZE - size < low) {
1296                         /* This entire reservation is too low; go to next. */
1297                         continue;
1298                 }
1299                 pa = VM_PAGE_TO_PHYS(&rv->pages[0]);
1300                 if (pa + size > high) {
1301                         /* This entire reservation is too high; go to next. */
1302                         continue;
1303                 }
1304                 if (vm_reserv_trylock(rv) == 0) {
1305                         vm_reserv_domain_unlock(domain);
1306                         vm_reserv_lock(rv);
1307                         if (!rv->inpartpopq) {
1308                                 vm_reserv_domain_lock(domain);
1309                                 if (!rvn->inpartpopq)
1310                                         goto again;
1311                                 continue;
1312                         }
1313                 } else
1314                         vm_reserv_domain_unlock(domain);
1315                 if (pa < low) {
1316                         /* Start the search for free pages at "low". */
1317                         low_index = (low + PAGE_MASK - pa) >> PAGE_SHIFT;
1318                         i = low_index / NBPOPMAP;
1319                         hi = low_index % NBPOPMAP;
1320                 } else
1321                         i = hi = 0;
1322                 do {
1323                         /* Find the next free page. */
1324                         lo = ffsl(~(((1UL << hi) - 1) | rv->popmap[i]));
1325                         while (lo == 0 && ++i < NPOPMAP)
1326                                 lo = ffsl(~rv->popmap[i]);
1327                         if (i == NPOPMAP)
1328                                 break;
1329                         /* Convert from ffsl() to ordinary bit numbering. */
1330                         lo--;
1331                         next_free = NBPOPMAP * i + lo;
1332                         pa = VM_PAGE_TO_PHYS(&rv->pages[next_free]);
1333                         KASSERT(pa >= low,
1334                             ("vm_reserv_reclaim_contig: pa is too low"));
1335                         if (pa + size > high) {
1336                                 /* The rest of this reservation is too high. */
1337                                 break;
1338                         } else if ((pa & (alignment - 1)) != 0 ||
1339                             ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0) {
1340                                 /*
1341                                  * The current page doesn't meet the alignment
1342                                  * and/or boundary requirements.  Continue
1343                                  * searching this reservation until the rest
1344                                  * of its free pages are either excluded or
1345                                  * exhausted.
1346                                  */
1347                                 hi = lo + 1;
1348                                 if (hi >= NBPOPMAP) {
1349                                         hi = 0;
1350                                         i++;
1351                                 }
1352                                 continue;
1353                         }
1354                         /* Find the next used page. */
1355                         hi = ffsl(rv->popmap[i] & ~((1UL << lo) - 1));
1356                         while (hi == 0 && ++i < NPOPMAP) {
1357                                 if ((NBPOPMAP * i - next_free) * PAGE_SIZE >=
1358                                     size) {
1359                                         vm_reserv_reclaim(rv);
1360                                         vm_reserv_unlock(rv);
1361                                         return (TRUE);
1362                                 }
1363                                 hi = ffsl(rv->popmap[i]);
1364                         }
1365                         /* Convert from ffsl() to ordinary bit numbering. */
1366                         if (i != NPOPMAP)
1367                                 hi--;
1368                         if ((NBPOPMAP * i + hi - next_free) * PAGE_SIZE >=
1369                             size) {
1370                                 vm_reserv_reclaim(rv);
1371                                 vm_reserv_unlock(rv);
1372                                 return (TRUE);
1373                         }
1374                 } while (i < NPOPMAP);
1375                 vm_reserv_unlock(rv);
1376                 vm_reserv_domain_lock(domain);
1377                 if (rvn != NULL && !rvn->inpartpopq)
1378                         goto again;
1379         }
1380         vm_reserv_domain_unlock(domain);
1381         return (FALSE);
1382 }
1383
1384 /*
1385  * Transfers the reservation underlying the given page to a new object.
1386  *
1387  * The object must be locked.
1388  */
1389 void
1390 vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
1391     vm_pindex_t old_object_offset)
1392 {
1393         vm_reserv_t rv;
1394
1395         VM_OBJECT_ASSERT_WLOCKED(new_object);
1396         rv = vm_reserv_from_page(m);
1397         if (rv->object == old_object) {
1398                 vm_reserv_lock(rv);
1399                 CTR6(KTR_VM,
1400                     "%s: rv %p object %p new %p popcnt %d inpartpop %d",
1401                     __FUNCTION__, rv, rv->object, new_object, rv->popcnt,
1402                     rv->inpartpopq);
1403                 if (rv->object == old_object) {
1404                         vm_reserv_object_lock(old_object);
1405                         rv->object = NULL;
1406                         LIST_REMOVE(rv, objq);
1407                         vm_reserv_object_unlock(old_object);
1408                         vm_reserv_object_lock(new_object);
1409                         rv->object = new_object;
1410                         rv->pindex -= old_object_offset;
1411                         LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
1412                         vm_reserv_object_unlock(new_object);
1413                 }
1414                 vm_reserv_unlock(rv);
1415         }
1416 }
1417
1418 /*
1419  * Returns the size (in bytes) of a reservation of the specified level.
1420  */
1421 int
1422 vm_reserv_size(int level)
1423 {
1424
1425         switch (level) {
1426         case 0:
1427                 return (VM_LEVEL_0_SIZE);
1428         case -1:
1429                 return (PAGE_SIZE);
1430         default:
1431                 return (0);
1432         }
1433 }
1434
1435 /*
1436  * Allocates the virtual and physical memory required by the reservation
1437  * management system's data structures, in particular, the reservation array.
1438  */
1439 vm_paddr_t
1440 vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t high_water)
1441 {
1442         vm_paddr_t new_end;
1443         size_t size;
1444
1445         /*
1446          * Calculate the size (in bytes) of the reservation array.  Round up
1447          * from "high_water" because every small page is mapped to an element
1448          * in the reservation array based on its physical address.  Thus, the
1449          * number of elements in the reservation array can be greater than the
1450          * number of superpages.
1451          */
1452         size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv);
1453
1454         /*
1455          * Allocate and map the physical memory for the reservation array.  The
1456          * next available virtual address is returned by reference.
1457          */
1458         new_end = end - round_page(size);
1459         vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
1460             VM_PROT_READ | VM_PROT_WRITE);
1461         bzero(vm_reserv_array, size);
1462
1463         /*
1464          * Return the next available physical address.
1465          */
1466         return (new_end);
1467 }
1468
1469 /*
1470  * Initializes the reservation management system.  Specifically, initializes
1471  * the reservation counters.
1472  */
1473 static void
1474 vm_reserv_counter_init(void *unused)
1475 {
1476
1477         vm_reserv_freed = counter_u64_alloc(M_WAITOK);
1478         vm_reserv_broken = counter_u64_alloc(M_WAITOK);
1479         vm_reserv_reclaimed = counter_u64_alloc(M_WAITOK);
1480 }
1481 SYSINIT(vm_reserv_counter_init, SI_SUB_CPU, SI_ORDER_ANY,
1482     vm_reserv_counter_init, NULL);
1483
1484 /*
1485  * Returns the superpage containing the given page.
1486  */
1487 vm_page_t
1488 vm_reserv_to_superpage(vm_page_t m)
1489 {
1490         vm_reserv_t rv;
1491
1492         VM_OBJECT_ASSERT_LOCKED(m->object);
1493         rv = vm_reserv_from_page(m);
1494         if (rv->object == m->object && rv->popcnt == VM_LEVEL_0_NPAGES)
1495                 m = rv->pages;
1496         else
1497                 m = NULL;
1498
1499         return (m);
1500 }
1501
1502 #endif  /* VM_NRESERVLEVEL > 0 */