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 > paddr && paddr +
 320                     VM_LEVEL_0_SIZE <= seg->end) {
 321                         rv = &vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT];
 322                         fullpop += rv->popcnt == VM_LEVEL_0_NPAGES;
 323                         paddr += VM_LEVEL_0_SIZE;
 324                 }
 325         }
 326         return (sysctl_handle_int(oidp, &fullpop, 0, req));
 327 }
 328
 329 /*
 330  * Describes the current state of the partially populated reservation queue.
 331  */
 332 static int
 333 sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
 334 {
 335         struct sbuf sbuf;
 336         vm_reserv_t rv;
 337         int counter, error, domain, level, unused_pages;
 338
 339         error = sysctl_wire_old_buffer(req, 0);
 340         if (error != 0)
 341                 return (error);
 342         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 343         sbuf_printf(&sbuf, "\nDOMAIN    LEVEL     SIZE  NUMBER\n\n");
 344         for (domain = 0; domain < vm_ndomains; domain++) {
 345                 for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
 346                         counter = 0;
 347                         unused_pages = 0;
 348                         vm_reserv_domain_lock(domain);
 349                         TAILQ_FOREACH(rv, &vm_rvq_partpop[domain], partpopq) {
 350                                 counter++;
 351                                 unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
 352                         }
 353                         vm_reserv_domain_unlock(domain);
 354                         sbuf_printf(&sbuf, "%6d, %7d, %6dK, %6d\n",
 355                             domain, level,
 356                             unused_pages * ((int)PAGE_SIZE / 1024), counter);
 357                 }
 358         }
 359         error = sbuf_finish(&sbuf);
 360         sbuf_delete(&sbuf);
 361         return (error);
 362 }
 363
 364 /*
 365  * Remove a reservation from the object's objq.
 366  */
 367 static void
 368 vm_reserv_remove(vm_reserv_t rv)
 369 {
 370         vm_object_t object;
 371
 372         vm_reserv_assert_locked(rv);
 373         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
 374             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
 375         KASSERT(rv->object != NULL,
 376             ("vm_reserv_remove: reserv %p is free", rv));
 377         KASSERT(!rv->inpartpopq,
 378             ("vm_reserv_remove: reserv %p's inpartpopq is TRUE", rv));
 379         object = rv->object;
 380         vm_reserv_object_lock(object);
 381         LIST_REMOVE(rv, objq);
 382         rv->object = NULL;
 383         vm_reserv_object_unlock(object);
 384 }
 385
 386 /*
 387  * Insert a new reservation into the object's objq.
 388  */
 389 static void
 390 vm_reserv_insert(vm_reserv_t rv, vm_object_t object, vm_pindex_t pindex)
 391 {
 392         int i;
 393
 394         vm_reserv_assert_locked(rv);
 395         CTR6(KTR_VM,
 396             "%s: rv %p(%p) object %p new %p popcnt %d",
 397             __FUNCTION__, rv, rv->pages, rv->object, object,
 398            rv->popcnt);
 399         KASSERT(rv->object == NULL,
 400             ("vm_reserv_insert: reserv %p isn't free", rv));
 401         KASSERT(rv->popcnt == 0,
 402             ("vm_reserv_insert: reserv %p's popcnt is corrupted", rv));
 403         KASSERT(!rv->inpartpopq,
 404             ("vm_reserv_insert: reserv %p's inpartpopq is TRUE", rv));
 405         for (i = 0; i < NPOPMAP; i++)
 406                 KASSERT(rv->popmap[i] == 0,
 407                     ("vm_reserv_insert: reserv %p's popmap is corrupted", rv));
 408         vm_reserv_object_lock(object);
 409         rv->pindex = pindex;
 410         rv->object = object;
 411         rv->lasttick = ticks;
 412         LIST_INSERT_HEAD(&object->rvq, rv, objq);
 413         vm_reserv_object_unlock(object);
 414 }
 415
 416 /*
 417  * Reduces the given reservation's population count.  If the population count
 418  * becomes zero, the reservation is destroyed.  Additionally, moves the
 419  * reservation to the tail of the partially populated reservation queue if the
 420  * population count is non-zero.
 421  */
 422 static void
 423 vm_reserv_depopulate(vm_reserv_t rv, int index)
 424 {
 425         struct vm_domain *vmd;
 426
 427         vm_reserv_assert_locked(rv);
 428         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
 429             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
 430         KASSERT(rv->object != NULL,
 431             ("vm_reserv_depopulate: reserv %p is free", rv));
 432         KASSERT(popmap_is_set(rv->popmap, index),
 433             ("vm_reserv_depopulate: reserv %p's popmap[%d] is clear", rv,
 434             index));
 435         KASSERT(rv->popcnt > 0,
 436             ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
 437         KASSERT(rv->domain < vm_ndomains,
 438             ("vm_reserv_depopulate: reserv %p's domain is corrupted %d",
 439             rv, rv->domain));
 440         if (rv->popcnt == VM_LEVEL_0_NPAGES) {
 441                 KASSERT(rv->pages->psind == 1,
 442                     ("vm_reserv_depopulate: reserv %p is already demoted",
 443                     rv));
 444                 rv->pages->psind = 0;
 445         }
 446         popmap_clear(rv->popmap, index);
 447         rv->popcnt--;
 448         if ((unsigned)(ticks - rv->lasttick) >= PARTPOPSLOP ||
 449             rv->popcnt == 0) {
 450                 vm_reserv_domain_lock(rv->domain);
 451                 if (rv->inpartpopq) {
 452                         TAILQ_REMOVE(&vm_rvq_partpop[rv->domain], rv, partpopq);
 453                         rv->inpartpopq = FALSE;
 454                 }
 455                 if (rv->popcnt != 0) {
 456                         rv->inpartpopq = TRUE;
 457                         TAILQ_INSERT_TAIL(&vm_rvq_partpop[rv->domain], rv, partpopq);
 458                 }
 459                 vm_reserv_domain_unlock(rv->domain);
 460                 rv->lasttick = ticks;
 461         }
 462         vmd = VM_DOMAIN(rv->domain);
 463         if (rv->popcnt == 0) {
 464                 vm_reserv_remove(rv);
 465                 vm_domain_free_lock(vmd);
 466                 vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
 467                 vm_domain_free_unlock(vmd);
 468                 counter_u64_add(vm_reserv_freed, 1);
 469         }
 470         vm_domain_freecnt_inc(vmd, 1);
 471 }
 472
 473 /*
 474  * Returns the reservation to which the given page might belong.
 475  */
 476 static __inline vm_reserv_t
 477 vm_reserv_from_page(vm_page_t m)
 478 {
 479
 480         return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
 481 }
 482
 483 /*
 484  * Returns an existing reservation or NULL and initialized successor pointer.
 485  */
 486 static vm_reserv_t
 487 vm_reserv_from_object(vm_object_t object, vm_pindex_t pindex,
 488     vm_page_t mpred, vm_page_t *msuccp)
 489 {
 490         vm_reserv_t rv;
 491         vm_page_t msucc;
 492
 493         msucc = NULL;
 494         if (mpred != NULL) {
 495                 KASSERT(mpred->object == object,
 496                     ("vm_reserv_from_object: object doesn't contain mpred"));
 497                 KASSERT(mpred->pindex < pindex,
 498                     ("vm_reserv_from_object: mpred doesn't precede pindex"));
 499                 rv = vm_reserv_from_page(mpred);
 500                 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
 501                         goto found;
 502                 msucc = TAILQ_NEXT(mpred, listq);
 503         } else
 504                 msucc = TAILQ_FIRST(&object->memq);
 505         if (msucc != NULL) {
 506                 KASSERT(msucc->pindex > pindex,
 507                     ("vm_reserv_from_object: msucc doesn't succeed pindex"));
 508                 rv = vm_reserv_from_page(msucc);
 509                 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
 510                         goto found;
 511         }
 512         rv = NULL;
 513
 514 found:
 515         *msuccp = msucc;
 516
 517         return (rv);
 518 }
 519
 520 /*
 521  * Returns TRUE if the given reservation contains the given page index and
 522  * FALSE otherwise.
 523  */
 524 static __inline boolean_t
 525 vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
 526 {
 527
 528         return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
 529 }
 530
 531 /*
 532  * Increases the given reservation's population count.  Moves the reservation
 533  * to the tail of the partially populated reservation queue.
 534  *
 535  * The free page queue must be locked.
 536  */
 537 static void
 538 vm_reserv_populate(vm_reserv_t rv, int index)
 539 {
 540
 541         vm_reserv_assert_locked(rv);
 542         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
 543             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
 544         KASSERT(rv->object != NULL,
 545             ("vm_reserv_populate: reserv %p is free", rv));
 546         KASSERT(popmap_is_clear(rv->popmap, index),
 547             ("vm_reserv_populate: reserv %p's popmap[%d] is set", rv,
 548             index));
 549         KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
 550             ("vm_reserv_populate: reserv %p is already full", rv));
 551         KASSERT(rv->pages->psind == 0,
 552             ("vm_reserv_populate: reserv %p is already promoted", rv));
 553         KASSERT(rv->domain < vm_ndomains,
 554             ("vm_reserv_populate: reserv %p's domain is corrupted %d",
 555             rv, rv->domain));
 556         popmap_set(rv->popmap, index);
 557         rv->popcnt++;
 558         if ((unsigned)(ticks - rv->lasttick) < PARTPOPSLOP &&
 559             rv->inpartpopq && rv->popcnt != VM_LEVEL_0_NPAGES)
 560                 return;
 561         rv->lasttick = ticks;
 562         vm_reserv_domain_lock(rv->domain);
 563         if (rv->inpartpopq) {
 564                 TAILQ_REMOVE(&vm_rvq_partpop[rv->domain], rv, partpopq);
 565                 rv->inpartpopq = FALSE;
 566         }
 567         if (rv->popcnt < VM_LEVEL_0_NPAGES) {
 568                 rv->inpartpopq = TRUE;
 569                 TAILQ_INSERT_TAIL(&vm_rvq_partpop[rv->domain], rv, partpopq);
 570         } else {
 571                 KASSERT(rv->pages->psind == 0,
 572                     ("vm_reserv_populate: reserv %p is already promoted",
 573                     rv));
 574                 rv->pages->psind = 1;
 575         }
 576         vm_reserv_domain_unlock(rv->domain);
 577 }
 578
 579 /*
 580  * Attempts to allocate a contiguous set of physical pages from existing
 581  * reservations.  See vm_reserv_alloc_contig() for a description of the
 582  * function's parameters.
 583  *
 584  * The page "mpred" must immediately precede the offset "pindex" within the
 585  * specified object.
 586  *
 587  * The object must be locked.
 588  */
 589 vm_page_t
 590 vm_reserv_extend_contig(int req, vm_object_t object, vm_pindex_t pindex,
 591     int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
 592     u_long alignment, vm_paddr_t boundary, vm_page_t mpred)
 593 {
 594         struct vm_domain *vmd;
 595         vm_paddr_t pa, size;
 596         vm_page_t m, msucc;
 597         vm_reserv_t rv;
 598         int i, index;
 599
 600         VM_OBJECT_ASSERT_WLOCKED(object);
 601         KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
 602
 603         /*
 604          * Is a reservation fundamentally impossible?
 605          */
 606         if (pindex < VM_RESERV_INDEX(object, pindex) ||
 607             pindex + npages > object->size || object->resident_page_count == 0)
 608                 return (NULL);
 609
 610         /*
 611          * All reservations of a particular size have the same alignment.
 612          * Assuming that the first page is allocated from a reservation, the
 613          * least significant bits of its physical address can be determined
 614          * from its offset from the beginning of the reservation and the size
 615          * of the reservation.
 616          *
 617          * Could the specified index within a reservation of the smallest
 618          * possible size satisfy the alignment and boundary requirements?
 619          */
 620         pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
 621         if ((pa & (alignment - 1)) != 0)
 622                 return (NULL);
 623         size = npages << PAGE_SHIFT;
 624         if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
 625                 return (NULL);
 626
 627         /*
 628          * Look for an existing reservation.
 629          */
 630         rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
 631         if (rv == NULL)
 632                 return (NULL);
 633         KASSERT(object != kernel_object || rv->domain == domain,
 634             ("vm_reserv_extend_contig: Domain mismatch from reservation."));
 635         index = VM_RESERV_INDEX(object, pindex);
 636         /* Does the allocation fit within the reservation? */
 637         if (index + npages > VM_LEVEL_0_NPAGES)
 638                 return (NULL);
 639         domain = rv->domain;
 640         vmd = VM_DOMAIN(domain);
 641         vm_reserv_lock(rv);
 642         if (rv->object != object)
 643                 goto out;
 644         m = &rv->pages[index];
 645         pa = VM_PAGE_TO_PHYS(m);
 646         if (pa < low || pa + size > high || (pa & (alignment - 1)) != 0 ||
 647             ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
 648                 goto out;
 649         /* Handle vm_page_rename(m, new_object, ...). */
 650         for (i = 0; i < npages; i++) {
 651                 if (popmap_is_set(rv->popmap, index + i))
 652                         goto out;
 653         }
 654         if (!vm_domain_allocate(vmd, req, npages))
 655                 goto out;
 656         for (i = 0; i < npages; i++)
 657                 vm_reserv_populate(rv, index + i);
 658         vm_reserv_unlock(rv);
 659         return (m);
 660
 661 out:
 662         vm_reserv_unlock(rv);
 663         return (NULL);
 664 }
 665
 666 /*
 667  * Allocates a contiguous set of physical pages of the given size "npages"
 668  * from newly created reservations.  All of the physical pages
 669  * must be at or above the given physical address "low" and below the given
 670  * physical address "high".  The given value "alignment" determines the
 671  * alignment of the first physical page in the set.  If the given value
 672  * "boundary" is non-zero, then the set of physical pages cannot cross any
 673  * physical address boundary that is a multiple of that value.  Both
 674  * "alignment" and "boundary" must be a power of two.
 675  *
 676  * Callers should first invoke vm_reserv_extend_contig() to attempt an
 677  * allocation from existing reservations.
 678  *
 679  * The page "mpred" must immediately precede the offset "pindex" within the
 680  * specified object.
 681  *
 682  * The object and free page queue must be locked.
 683  */
 684 vm_page_t
 685 vm_reserv_alloc_contig(int req, vm_object_t object, vm_pindex_t pindex, int domain,
 686     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
 687     vm_paddr_t boundary, vm_page_t mpred)
 688 {
 689         struct vm_domain *vmd;
 690         vm_paddr_t pa, size;
 691         vm_page_t m, m_ret, msucc;
 692         vm_pindex_t first, leftcap, rightcap;
 693         vm_reserv_t rv;
 694         u_long allocpages, maxpages, minpages;
 695         int i, index, n;
 696
 697         VM_OBJECT_ASSERT_WLOCKED(object);
 698         KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
 699
 700         /*
 701          * Is a reservation fundamentally impossible?
 702          */
 703         if (pindex < VM_RESERV_INDEX(object, pindex) ||
 704             pindex + npages > object->size)
 705                 return (NULL);
 706
 707         /*
 708          * All reservations of a particular size have the same alignment.
 709          * Assuming that the first page is allocated from a reservation, the
 710          * least significant bits of its physical address can be determined
 711          * from its offset from the beginning of the reservation and the size
 712          * of the reservation.
 713          *
 714          * Could the specified index within a reservation of the smallest
 715          * possible size satisfy the alignment and boundary requirements?
 716          */
 717         pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
 718         if ((pa & (alignment - 1)) != 0)
 719                 return (NULL);
 720         size = npages << PAGE_SHIFT;
 721         if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
 722                 return (NULL);
 723
 724         /*
 725          * Callers should've extended an existing reservation prior to
 726          * calling this function.  If a reservation exists it is
 727          * incompatible with the allocation.
 728          */
 729         rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
 730         if (rv != NULL)
 731                 return (NULL);
 732
 733         /*
 734          * Could at least one reservation fit between the first index to the
 735          * left that can be used ("leftcap") and the first index to the right
 736          * that cannot be used ("rightcap")?
 737          *
 738          * We must synchronize with the reserv object lock to protect the
 739          * pindex/object of the resulting reservations against rename while
 740          * we are inspecting.
 741          */
 742         first = pindex - VM_RESERV_INDEX(object, pindex);
 743         minpages = VM_RESERV_INDEX(object, pindex) + npages;
 744         maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES);
 745         allocpages = maxpages;
 746         vm_reserv_object_lock(object);
 747         if (mpred != NULL) {
 748                 if ((rv = vm_reserv_from_page(mpred))->object != object)
 749                         leftcap = mpred->pindex + 1;
 750                 else
 751                         leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
 752                 if (leftcap > first) {
 753                         vm_reserv_object_unlock(object);
 754                         return (NULL);
 755                 }
 756         }
 757         if (msucc != NULL) {
 758                 if ((rv = vm_reserv_from_page(msucc))->object != object)
 759                         rightcap = msucc->pindex;
 760                 else
 761                         rightcap = rv->pindex;
 762                 if (first + maxpages > rightcap) {
 763                         if (maxpages == VM_LEVEL_0_NPAGES) {
 764                                 vm_reserv_object_unlock(object);
 765                                 return (NULL);
 766                         }
 767
 768                         /*
 769                          * At least one reservation will fit between "leftcap"
 770                          * and "rightcap".  However, a reservation for the
 771                          * last of the requested pages will not fit.  Reduce
 772                          * the size of the upcoming allocation accordingly.
 773                          */
 774                         allocpages = minpages;
 775                 }
 776         }
 777         vm_reserv_object_unlock(object);
 778
 779         /*
 780          * Would the last new reservation extend past the end of the object?
 781          */
 782         if (first + maxpages > object->size) {
 783                 /*
 784                  * Don't allocate the last new reservation if the object is a
 785                  * vnode or backed by another object that is a vnode.
 786                  */
 787                 if (object->type == OBJT_VNODE ||
 788                     (object->backing_object != NULL &&
 789                     object->backing_object->type == OBJT_VNODE)) {
 790                         if (maxpages == VM_LEVEL_0_NPAGES)
 791                                 return (NULL);
 792                         allocpages = minpages;
 793                 }
 794                 /* Speculate that the object may grow. */
 795         }
 796
 797         /*
 798          * Allocate the physical pages.  The alignment and boundary specified
 799          * for this allocation may be different from the alignment and
 800          * boundary specified for the requested pages.  For instance, the
 801          * specified index may not be the first page within the first new
 802          * reservation.
 803          */
 804         m = NULL;
 805         vmd = VM_DOMAIN(domain);
 806         if (vm_domain_allocate(vmd, req, npages)) {
 807                 vm_domain_free_lock(vmd);
 808                 m = vm_phys_alloc_contig(domain, allocpages, low, high,
 809                     ulmax(alignment, VM_LEVEL_0_SIZE),
 810                     boundary > VM_LEVEL_0_SIZE ? boundary : 0);
 811                 vm_domain_free_unlock(vmd);
 812                 if (m == NULL) {
 813                         vm_domain_freecnt_inc(vmd, npages);
 814                         return (NULL);
 815                 }
 816         } else
 817                 return (NULL);
 818         KASSERT(vm_phys_domain(m) == domain,
 819             ("vm_reserv_alloc_contig: Page domain does not match requested."));
 820
 821         /*
 822          * The allocated physical pages always begin at a reservation
 823          * boundary, but they do not always end at a reservation boundary.
 824          * Initialize every reservation that is completely covered by the
 825          * allocated physical pages.
 826          */
 827         m_ret = NULL;
 828         index = VM_RESERV_INDEX(object, pindex);
 829         do {
 830                 rv = vm_reserv_from_page(m);
 831                 KASSERT(rv->pages == m,
 832                     ("vm_reserv_alloc_contig: reserv %p's pages is corrupted",
 833                     rv));
 834                 vm_reserv_lock(rv);
 835                 vm_reserv_insert(rv, object, first);
 836                 n = ulmin(VM_LEVEL_0_NPAGES - index, npages);
 837                 for (i = 0; i < n; i++)
 838                         vm_reserv_populate(rv, index + i);
 839                 npages -= n;
 840                 if (m_ret == NULL) {
 841                         m_ret = &rv->pages[index];
 842                         index = 0;
 843                 }
 844                 vm_reserv_unlock(rv);
 845                 m += VM_LEVEL_0_NPAGES;
 846                 first += VM_LEVEL_0_NPAGES;
 847                 allocpages -= VM_LEVEL_0_NPAGES;
 848         } while (allocpages >= VM_LEVEL_0_NPAGES);
 849         return (m_ret);
 850 }
 851
 852 /*
 853  * Attempts to extend an existing reservation and allocate the page to the
 854  * object.
 855  *
 856  * The page "mpred" must immediately precede the offset "pindex" within the
 857  * specified object.
 858  *
 859  * The object must be locked.
 860  */
 861 vm_page_t
 862 vm_reserv_extend(int req, vm_object_t object, vm_pindex_t pindex, int domain,
 863     vm_page_t mpred)
 864 {
 865         struct vm_domain *vmd;
 866         vm_page_t m, msucc;
 867         vm_reserv_t rv;
 868         int index;
 869
 870         VM_OBJECT_ASSERT_WLOCKED(object);
 871
 872         /*
 873          * Could a reservation currently exist?
 874          */
 875         if (pindex < VM_RESERV_INDEX(object, pindex) ||
 876             pindex >= object->size || object->resident_page_count == 0)
 877                 return (NULL);
 878
 879         /*
 880          * Look for an existing reservation.
 881          */
 882         rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
 883         if (rv == NULL)
 884                 return (NULL);
 885
 886         KASSERT(object != kernel_object || rv->domain == domain,
 887             ("vm_reserv_extend: Domain mismatch from reservation."));
 888         domain = rv->domain;
 889         vmd = VM_DOMAIN(domain);
 890         index = VM_RESERV_INDEX(object, pindex);
 891         m = &rv->pages[index];
 892         vm_reserv_lock(rv);
 893         /* Handle reclaim race. */
 894         if (rv->object != object ||
 895             /* Handle vm_page_rename(m, new_object, ...). */
 896             popmap_is_set(rv->popmap, index)) {
 897                 m = NULL;
 898                 goto out;
 899         }
 900         if (vm_domain_allocate(vmd, req, 1) == 0)
 901                 m = NULL;
 902         else
 903                 vm_reserv_populate(rv, index);
 904 out:
 905         vm_reserv_unlock(rv);
 906
 907         return (m);
 908 }
 909
 910 /*
 911  * Attempts to allocate a new reservation for the object, and allocates a
 912  * page from that reservation.  Callers should first invoke vm_reserv_extend()
 913  * to attempt an allocation from an existing reservation.
 914  *
 915  * The page "mpred" must immediately precede the offset "pindex" within the
 916  * specified object.
 917  *
 918  * The object and free page queue must be locked.
 919  */
 920 vm_page_t
 921 vm_reserv_alloc_page(int req, vm_object_t object, vm_pindex_t pindex, int domain,
 922     vm_page_t mpred)
 923 {
 924         struct vm_domain *vmd;
 925         vm_page_t m, msucc;
 926         vm_pindex_t first, leftcap, rightcap;
 927         vm_reserv_t rv;
 928         int index;
 929
 930         VM_OBJECT_ASSERT_WLOCKED(object);
 931
 932         /*
 933          * Is a reservation fundamentally impossible?
 934          */
 935         if (pindex < VM_RESERV_INDEX(object, pindex) ||
 936             pindex >= object->size)
 937                 return (NULL);
 938
 939         /*
 940          * Callers should've extended an existing reservation prior to
 941          * calling this function.  If a reservation exists it is
 942          * incompatible with the allocation.
 943          */
 944         rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
 945         if (rv != NULL)
 946                 return (NULL);
 947
 948         /*
 949          * Could a reservation fit between the first index to the left that
 950          * can be used and the first index to the right that cannot be used?
 951          *
 952          * We must synchronize with the reserv object lock to protect the
 953          * pindex/object of the resulting reservations against rename while
 954          * we are inspecting.
 955          */
 956         first = pindex - VM_RESERV_INDEX(object, pindex);
 957         vm_reserv_object_lock(object);
 958         if (mpred != NULL) {
 959                 if ((rv = vm_reserv_from_page(mpred))->object != object)
 960                         leftcap = mpred->pindex + 1;
 961                 else
 962                         leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
 963                 if (leftcap > first) {
 964                         vm_reserv_object_unlock(object);
 965                         return (NULL);
 966                 }
 967         }
 968         if (msucc != NULL) {
 969                 if ((rv = vm_reserv_from_page(msucc))->object != object)
 970                         rightcap = msucc->pindex;
 971                 else
 972                         rightcap = rv->pindex;
 973                 if (first + VM_LEVEL_0_NPAGES > rightcap) {
 974                         vm_reserv_object_unlock(object);
 975                         return (NULL);
 976                 }
 977         }
 978         vm_reserv_object_unlock(object);
 979
 980         /*
 981          * Would a new reservation extend past the end of the object?
 982          */
 983         if (first + VM_LEVEL_0_NPAGES > object->size) {
 984                 /*
 985                  * Don't allocate a new reservation if the object is a vnode or
 986                  * backed by another object that is a vnode.
 987                  */
 988                 if (object->type == OBJT_VNODE ||
 989                     (object->backing_object != NULL &&
 990                     object->backing_object->type == OBJT_VNODE))
 991                         return (NULL);
 992                 /* Speculate that the object may grow. */
 993         }
 994
 995         /*
 996          * Allocate and populate the new reservation.
 997          */
 998         m = NULL;
 999         vmd = VM_DOMAIN(domain);
1000         if (vm_domain_allocate(vmd, req, 1)) {
1001                 vm_domain_free_lock(vmd);
1002                 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT,
1003                     VM_LEVEL_0_ORDER);
1004                 vm_domain_free_unlock(vmd);
1005                 if (m == NULL) {
1006                         vm_domain_freecnt_inc(vmd, 1);
1007                         return (NULL);
1008                 }
1009         } else
1010                 return (NULL);
1011         rv = vm_reserv_from_page(m);
1012         vm_reserv_lock(rv);
1013         KASSERT(rv->pages == m,
1014             ("vm_reserv_alloc_page: reserv %p's pages is corrupted", rv));
1015         vm_reserv_insert(rv, object, first);
1016         index = VM_RESERV_INDEX(object, pindex);
1017         vm_reserv_populate(rv, index);
1018         vm_reserv_unlock(rv);
1019
1020         return (&rv->pages[index]);
1021 }
1022
1023 /*
1024  * Breaks the given reservation.  All free pages in the reservation
1025  * are returned to the physical memory allocator.  The reservation's
1026  * population count and map are reset to their initial state.
1027  *
1028  * The given reservation must not be in the partially populated reservation
1029  * queue.  The free page queue lock must be held.
1030  */
1031 static void
1032 vm_reserv_break(vm_reserv_t rv)
1033 {
1034         int begin_zeroes, 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         i = hi = 0;
1042         do {
1043                 /* Find the next 0 bit.  Any previous 0 bits are < "hi". */
1044                 lo = ffsl(~(((1UL << hi) - 1) | rv->popmap[i]));
1045                 if (lo == 0) {
1046                         /* Redundantly clears bits < "hi". */
1047                         rv->popmap[i] = 0;
1048                         rv->popcnt -= NBPOPMAP - hi;
1049                         while (++i < NPOPMAP) {
1050                                 lo = ffsl(~rv->popmap[i]);
1051                                 if (lo == 0) {
1052                                         rv->popmap[i] = 0;
1053                                         rv->popcnt -= NBPOPMAP;
1054                                 } else
1055                                         break;
1056                         }
1057                         if (i == NPOPMAP)
1058                                 break;
1059                         hi = 0;
1060                 }
1061                 KASSERT(lo > 0, ("vm_reserv_break: lo is %d", lo));
1062                 /* Convert from ffsl() to ordinary bit numbering. */
1063                 lo--;
1064                 if (lo > 0) {
1065                         /* Redundantly clears bits < "hi". */
1066                         rv->popmap[i] &= ~((1UL << lo) - 1);
1067                         rv->popcnt -= lo - hi;
1068                 }
1069                 begin_zeroes = NBPOPMAP * i + lo;
1070                 /* Find the next 1 bit. */
1071                 do
1072                         hi = ffsl(rv->popmap[i]);
1073                 while (hi == 0 && ++i < NPOPMAP);
1074                 if (i != NPOPMAP)
1075                         /* Convert from ffsl() to ordinary bit numbering. */
1076                         hi--;
1077                 vm_domain_free_lock(VM_DOMAIN(rv->domain));
1078                 vm_phys_free_contig(&rv->pages[begin_zeroes], NBPOPMAP * i +
1079                     hi - begin_zeroes);
1080                 vm_domain_free_unlock(VM_DOMAIN(rv->domain));
1081         } while (i < NPOPMAP);
1082         KASSERT(rv->popcnt == 0,
1083             ("vm_reserv_break: reserv %p's popcnt is corrupted", rv));
1084         counter_u64_add(vm_reserv_broken, 1);
1085 }
1086
1087 /*
1088  * Breaks all reservations belonging to the given object.
1089  */
1090 void
1091 vm_reserv_break_all(vm_object_t object)
1092 {
1093         vm_reserv_t rv;
1094
1095         /*
1096          * This access of object->rvq is unsynchronized so that the
1097          * object rvq lock can nest after the domain_free lock.  We
1098          * must check for races in the results.  However, the object
1099          * lock prevents new additions, so we are guaranteed that when
1100          * it returns NULL the object is properly empty.
1101          */
1102         while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
1103                 vm_reserv_lock(rv);
1104                 /* Reclaim race. */
1105                 if (rv->object != object) {
1106                         vm_reserv_unlock(rv);
1107                         continue;
1108                 }
1109                 vm_reserv_domain_lock(rv->domain);
1110                 if (rv->inpartpopq) {
1111                         TAILQ_REMOVE(&vm_rvq_partpop[rv->domain], rv, partpopq);
1112                         rv->inpartpopq = FALSE;
1113                 }
1114                 vm_reserv_domain_unlock(rv->domain);
1115                 vm_reserv_break(rv);
1116                 vm_reserv_unlock(rv);
1117         }
1118 }
1119
1120 /*
1121  * Frees the given page if it belongs to a reservation.  Returns TRUE if the
1122  * page is freed and FALSE otherwise.
1123  *
1124  * The free page queue lock must be held.
1125  */
1126 boolean_t
1127 vm_reserv_free_page(vm_page_t m)
1128 {
1129         vm_reserv_t rv;
1130         boolean_t ret;
1131
1132         rv = vm_reserv_from_page(m);
1133         if (rv->object == NULL)
1134                 return (FALSE);
1135         vm_reserv_lock(rv);
1136         /* Re-validate after lock. */
1137         if (rv->object != NULL) {
1138                 vm_reserv_depopulate(rv, m - rv->pages);
1139                 ret = TRUE;
1140         } else
1141                 ret = FALSE;
1142         vm_reserv_unlock(rv);
1143
1144         return (ret);
1145 }
1146
1147 /*
1148  * Initializes the reservation management system.  Specifically, initializes
1149  * the reservation array.
1150  *
1151  * Requires that vm_page_array and first_page are initialized!
1152  */
1153 void
1154 vm_reserv_init(void)
1155 {
1156         vm_paddr_t paddr;
1157         struct vm_phys_seg *seg;
1158         struct vm_reserv *rv;
1159         int i, segind;
1160
1161         /*
1162          * Initialize the reservation array.  Specifically, initialize the
1163          * "pages" field for every element that has an underlying superpage.
1164          */
1165         for (segind = 0; segind < vm_phys_nsegs; segind++) {
1166                 seg = &vm_phys_segs[segind];
1167                 paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
1168                 while (paddr + VM_LEVEL_0_SIZE > paddr && paddr +
1169                     VM_LEVEL_0_SIZE <= seg->end) {
1170                         rv = &vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT];
1171                         rv->pages = PHYS_TO_VM_PAGE(paddr);
1172                         rv->domain = seg->domain;
1173                         mtx_init(&rv->lock, "vm reserv", NULL, MTX_DEF);
1174                         paddr += VM_LEVEL_0_SIZE;
1175                 }
1176         }
1177         for (i = 0; i < MAXMEMDOM; i++) {
1178                 mtx_init(&vm_reserv_domain_locks[i], "VM reserv domain", NULL,
1179                     MTX_DEF);
1180                 TAILQ_INIT(&vm_rvq_partpop[i]);
1181         }
1182
1183         for (i = 0; i < VM_RESERV_OBJ_LOCK_COUNT; i++)
1184                 mtx_init(&vm_reserv_object_mtx[i], "resv obj lock", NULL,
1185                     MTX_DEF);
1186 }
1187
1188 /*
1189  * Returns true if the given page belongs to a reservation and that page is
1190  * free.  Otherwise, returns false.
1191  */
1192 bool
1193 vm_reserv_is_page_free(vm_page_t m)
1194 {
1195         vm_reserv_t rv;
1196
1197         rv = vm_reserv_from_page(m);
1198         if (rv->object == NULL)
1199                 return (false);
1200         return (popmap_is_clear(rv->popmap, m - rv->pages));
1201 }
1202
1203 /*
1204  * If the given page belongs to a reservation, returns the level of that
1205  * reservation.  Otherwise, returns -1.
1206  */
1207 int
1208 vm_reserv_level(vm_page_t m)
1209 {
1210         vm_reserv_t rv;
1211
1212         rv = vm_reserv_from_page(m);
1213         return (rv->object != NULL ? 0 : -1);
1214 }
1215
1216 /*
1217  * Returns a reservation level if the given page belongs to a fully populated
1218  * reservation and -1 otherwise.
1219  */
1220 int
1221 vm_reserv_level_iffullpop(vm_page_t m)
1222 {
1223         vm_reserv_t rv;
1224
1225         rv = vm_reserv_from_page(m);
1226         return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
1227 }
1228
1229 /*
1230  * Breaks the given partially populated reservation, releasing its free pages
1231  * to the physical memory allocator.
1232  *
1233  * The free page queue lock must be held.
1234  */
1235 static void
1236 vm_reserv_reclaim(vm_reserv_t rv)
1237 {
1238
1239         vm_reserv_assert_locked(rv);
1240         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
1241             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
1242         vm_reserv_domain_lock(rv->domain);
1243         KASSERT(rv->inpartpopq,
1244             ("vm_reserv_reclaim: reserv %p's inpartpopq is FALSE", rv));
1245         KASSERT(rv->domain < vm_ndomains,
1246             ("vm_reserv_reclaim: reserv %p's domain is corrupted %d",
1247             rv, rv->domain));
1248         TAILQ_REMOVE(&vm_rvq_partpop[rv->domain], rv, partpopq);
1249         rv->inpartpopq = FALSE;
1250         vm_reserv_domain_unlock(rv->domain);
1251         vm_reserv_break(rv);
1252         counter_u64_add(vm_reserv_reclaimed, 1);
1253 }
1254
1255 /*
1256  * Breaks the reservation at the head of the partially populated reservation
1257  * queue, releasing its free pages to the physical memory allocator.  Returns
1258  * TRUE if a reservation is broken and FALSE otherwise.
1259  *
1260  * The free page queue lock must be held.
1261  */
1262 boolean_t
1263 vm_reserv_reclaim_inactive(int domain)
1264 {
1265         vm_reserv_t rv;
1266
1267         while ((rv = TAILQ_FIRST(&vm_rvq_partpop[domain])) != NULL) {
1268                 vm_reserv_lock(rv);
1269                 if (rv != TAILQ_FIRST(&vm_rvq_partpop[domain])) {
1270                         vm_reserv_unlock(rv);
1271                         continue;
1272                 }
1273                 vm_reserv_reclaim(rv);
1274                 vm_reserv_unlock(rv);
1275                 return (TRUE);
1276         }
1277         return (FALSE);
1278 }
1279
1280 /*
1281  * Searches the partially populated reservation queue for the least recently
1282  * changed reservation with free pages that satisfy the given request for
1283  * contiguous physical memory.  If a satisfactory reservation is found, it is
1284  * broken.  Returns TRUE if a reservation is broken and FALSE otherwise.
1285  *
1286  * The free page queue lock must be held.
1287  */
1288 boolean_t
1289 vm_reserv_reclaim_contig(int domain, u_long npages, vm_paddr_t low,
1290     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1291 {
1292         vm_paddr_t pa, size;
1293         vm_reserv_t rv, rvn;
1294         int hi, i, lo, low_index, next_free;
1295
1296         if (npages > VM_LEVEL_0_NPAGES - 1)
1297                 return (FALSE);
1298         size = npages << PAGE_SHIFT;
1299         vm_reserv_domain_lock(domain);
1300 again:
1301         for (rv = TAILQ_FIRST(&vm_rvq_partpop[domain]); rv != NULL; rv = rvn) {
1302                 rvn = TAILQ_NEXT(rv, partpopq);
1303                 pa = VM_PAGE_TO_PHYS(&rv->pages[VM_LEVEL_0_NPAGES - 1]);
1304                 if (pa + PAGE_SIZE - size < low) {
1305                         /* This entire reservation is too low; go to next. */
1306                         continue;
1307                 }
1308                 pa = VM_PAGE_TO_PHYS(&rv->pages[0]);
1309                 if (pa + size > high) {
1310                         /* This entire reservation is too high; go to next. */
1311                         continue;
1312                 }
1313                 if (vm_reserv_trylock(rv) == 0) {
1314                         vm_reserv_domain_unlock(domain);
1315                         vm_reserv_lock(rv);
1316                         if (!rv->inpartpopq) {
1317                                 vm_reserv_domain_lock(domain);
1318                                 if (!rvn->inpartpopq)
1319                                         goto again;
1320                                 continue;
1321                         }
1322                 } else
1323                         vm_reserv_domain_unlock(domain);
1324                 if (pa < low) {
1325                         /* Start the search for free pages at "low". */
1326                         low_index = (low + PAGE_MASK - pa) >> PAGE_SHIFT;
1327                         i = low_index / NBPOPMAP;
1328                         hi = low_index % NBPOPMAP;
1329                 } else
1330                         i = hi = 0;
1331                 do {
1332                         /* Find the next free page. */
1333                         lo = ffsl(~(((1UL << hi) - 1) | rv->popmap[i]));
1334                         while (lo == 0 && ++i < NPOPMAP)
1335                                 lo = ffsl(~rv->popmap[i]);
1336                         if (i == NPOPMAP)
1337                                 break;
1338                         /* Convert from ffsl() to ordinary bit numbering. */
1339                         lo--;
1340                         next_free = NBPOPMAP * i + lo;
1341                         pa = VM_PAGE_TO_PHYS(&rv->pages[next_free]);
1342                         KASSERT(pa >= low,
1343                             ("vm_reserv_reclaim_contig: pa is too low"));
1344                         if (pa + size > high) {
1345                                 /* The rest of this reservation is too high. */
1346                                 break;
1347                         } else if ((pa & (alignment - 1)) != 0 ||
1348                             ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0) {
1349                                 /*
1350                                  * The current page doesn't meet the alignment
1351                                  * and/or boundary requirements.  Continue
1352                                  * searching this reservation until the rest
1353                                  * of its free pages are either excluded or
1354                                  * exhausted.
1355                                  */
1356                                 hi = lo + 1;
1357                                 if (hi >= NBPOPMAP) {
1358                                         hi = 0;
1359                                         i++;
1360                                 }
1361                                 continue;
1362                         }
1363                         /* Find the next used page. */
1364                         hi = ffsl(rv->popmap[i] & ~((1UL << lo) - 1));
1365                         while (hi == 0 && ++i < NPOPMAP) {
1366                                 if ((NBPOPMAP * i - next_free) * PAGE_SIZE >=
1367                                     size) {
1368                                         vm_reserv_reclaim(rv);
1369                                         vm_reserv_unlock(rv);
1370                                         return (TRUE);
1371                                 }
1372                                 hi = ffsl(rv->popmap[i]);
1373                         }
1374                         /* Convert from ffsl() to ordinary bit numbering. */
1375                         if (i != NPOPMAP)
1376                                 hi--;
1377                         if ((NBPOPMAP * i + hi - next_free) * PAGE_SIZE >=
1378                             size) {
1379                                 vm_reserv_reclaim(rv);
1380                                 vm_reserv_unlock(rv);
1381                                 return (TRUE);
1382                         }
1383                 } while (i < NPOPMAP);
1384                 vm_reserv_unlock(rv);
1385                 vm_reserv_domain_lock(domain);
1386                 if (rvn != NULL && !rvn->inpartpopq)
1387                         goto again;
1388         }
1389         vm_reserv_domain_unlock(domain);
1390         return (FALSE);
1391 }
1392
1393 /*
1394  * Transfers the reservation underlying the given page to a new object.
1395  *
1396  * The object must be locked.
1397  */
1398 void
1399 vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
1400     vm_pindex_t old_object_offset)
1401 {
1402         vm_reserv_t rv;
1403
1404         VM_OBJECT_ASSERT_WLOCKED(new_object);
1405         rv = vm_reserv_from_page(m);
1406         if (rv->object == old_object) {
1407                 vm_reserv_lock(rv);
1408                 CTR6(KTR_VM,
1409                     "%s: rv %p object %p new %p popcnt %d inpartpop %d",
1410                     __FUNCTION__, rv, rv->object, new_object, rv->popcnt,
1411                     rv->inpartpopq);
1412                 if (rv->object == old_object) {
1413                         vm_reserv_object_lock(old_object);
1414                         rv->object = NULL;
1415                         LIST_REMOVE(rv, objq);
1416                         vm_reserv_object_unlock(old_object);
1417                         vm_reserv_object_lock(new_object);
1418                         rv->object = new_object;
1419                         rv->pindex -= old_object_offset;
1420                         LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
1421                         vm_reserv_object_unlock(new_object);
1422                 }
1423                 vm_reserv_unlock(rv);
1424         }
1425 }
1426
1427 /*
1428  * Returns the size (in bytes) of a reservation of the specified level.
1429  */
1430 int
1431 vm_reserv_size(int level)
1432 {
1433
1434         switch (level) {
1435         case 0:
1436                 return (VM_LEVEL_0_SIZE);
1437         case -1:
1438                 return (PAGE_SIZE);
1439         default:
1440                 return (0);
1441         }
1442 }
1443
1444 /*
1445  * Allocates the virtual and physical memory required by the reservation
1446  * management system's data structures, in particular, the reservation array.
1447  */
1448 vm_paddr_t
1449 vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t high_water)
1450 {
1451         vm_paddr_t new_end;
1452         size_t size;
1453
1454         /*
1455          * Calculate the size (in bytes) of the reservation array.  Round up
1456          * from "high_water" because every small page is mapped to an element
1457          * in the reservation array based on its physical address.  Thus, the
1458          * number of elements in the reservation array can be greater than the
1459          * number of superpages.
1460          */
1461         size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv);
1462
1463         /*
1464          * Allocate and map the physical memory for the reservation array.  The
1465          * next available virtual address is returned by reference.
1466          */
1467         new_end = end - round_page(size);
1468         vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
1469             VM_PROT_READ | VM_PROT_WRITE);
1470         bzero(vm_reserv_array, size);
1471
1472         /*
1473          * Return the next available physical address.
1474          */
1475         return (new_end);
1476 }
1477
1478 /*
1479  * Initializes the reservation management system.  Specifically, initializes
1480  * the reservation counters.
1481  */
1482 static void
1483 vm_reserv_counter_init(void *unused)
1484 {
1485
1486         vm_reserv_freed = counter_u64_alloc(M_WAITOK);
1487         vm_reserv_broken = counter_u64_alloc(M_WAITOK);
1488         vm_reserv_reclaimed = counter_u64_alloc(M_WAITOK);
1489 }
1490 SYSINIT(vm_reserv_counter_init, SI_SUB_CPU, SI_ORDER_ANY,
1491     vm_reserv_counter_init, NULL);
1492
1493 /*
1494  * Returns the superpage containing the given page.
1495  */
1496 vm_page_t
1497 vm_reserv_to_superpage(vm_page_t m)
1498 {
1499         vm_reserv_t rv;
1500
1501         VM_OBJECT_ASSERT_LOCKED(m->object);
1502         rv = vm_reserv_from_page(m);
1503         if (rv->object == m->object && rv->popcnt == VM_LEVEL_0_NPAGES)
1504                 m = rv->pages;
1505         else
1506                 m = NULL;
1507
1508         return (m);
1509 }
1510
1511 #endif  /* VM_NRESERVLEVEL > 0 */