sys/vm/vm_reserv.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2002-2006 Rice University
   5  * Copyright (c) 2007-2011 Alan L. Cox <alc@cs.rice.edu>
   6  * All rights reserved.
   7  *
   8  * This software was developed for the FreeBSD Project by Alan L. Cox,
   9  * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro.
  10  *
  11  * Redistribution and use in source and binary forms, with or without
  12  * modification, are permitted provided that the following conditions
  13  * are met:
  14  * 1. Redistributions of source code must retain the above copyright
  15  *    notice, this list of conditions and the following disclaimer.
  16  * 2. Redistributions in binary form must reproduce the above copyright
  17  *    notice, this list of conditions and the following disclaimer in the
  18  *    documentation and/or other materials provided with the distribution.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  23  * A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT
  24  * HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  25  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
  26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
  27  * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY
  30  * WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
  31  * POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 /*
  35  *      Superpage reservation management module
  36  *
  37  * Any external functions defined by this module are only to be used by the
  38  * virtual memory system.
  39  */
  40
  41 #include <sys/cdefs.h>
  42 __FBSDID("$FreeBSD$");
  43
  44 #include "opt_vm.h"
  45
  46 #include <sys/param.h>
  47 #include <sys/kernel.h>
  48 #include <sys/lock.h>
  49 #include <sys/malloc.h>
  50 #include <sys/mutex.h>
  51 #include <sys/queue.h>
  52 #include <sys/rwlock.h>
  53 #include <sys/sbuf.h>
  54 #include <sys/sysctl.h>
  55 #include <sys/systm.h>
  56 #include <sys/counter.h>
  57 #include <sys/ktr.h>
  58 #include <sys/vmmeter.h>
  59 #include <sys/smp.h>
  60
  61 #include <vm/vm.h>
  62 #include <vm/vm_param.h>
  63 #include <vm/vm_object.h>
  64 #include <vm/vm_page.h>
  65 #include <vm/vm_pageout.h>
  66 #include <vm/vm_phys.h>
  67 #include <vm/vm_pagequeue.h>
  68 #include <vm/vm_radix.h>
  69 #include <vm/vm_reserv.h>
  70
  71 /*
  72  * The reservation system supports the speculative allocation of large physical
  73  * pages ("superpages").  Speculative allocation enables the fully automatic
  74  * utilization of superpages by the virtual memory system.  In other words, no
  75  * programmatic directives are required to use superpages.
  76  */
  77
  78 #if VM_NRESERVLEVEL > 0
  79
  80 #ifndef VM_LEVEL_0_ORDER_MAX
  81 #define VM_LEVEL_0_ORDER_MAX    VM_LEVEL_0_ORDER
  82 #endif
  83
  84 /*
  85  * The number of small pages that are contained in a level 0 reservation
  86  */
  87 #define VM_LEVEL_0_NPAGES       (1 << VM_LEVEL_0_ORDER)
  88 #define VM_LEVEL_0_NPAGES_MAX   (1 << VM_LEVEL_0_ORDER_MAX)
  89
  90 /*
  91  * The number of bits by which a physical address is shifted to obtain the
  92  * reservation number
  93  */
  94 #define VM_LEVEL_0_SHIFT        (VM_LEVEL_0_ORDER + PAGE_SHIFT)
  95
  96 /*
  97  * The size of a level 0 reservation in bytes
  98  */
  99 #define VM_LEVEL_0_SIZE         (1 << VM_LEVEL_0_SHIFT)
 100
 101 /*
 102  * Computes the index of the small page underlying the given (object, pindex)
 103  * within the reservation's array of small pages.
 104  */
 105 #define VM_RESERV_INDEX(object, pindex) \
 106     (((object)->pg_color + (pindex)) & (VM_LEVEL_0_NPAGES - 1))
 107
 108 /*
 109  * The size of a population map entry
 110  */
 111 typedef u_long          popmap_t;
 112
 113 /*
 114  * The number of bits in a population map entry
 115  */
 116 #define NBPOPMAP        (NBBY * sizeof(popmap_t))
 117
 118 /*
 119  * The number of population map entries in a reservation
 120  */
 121 #define NPOPMAP         howmany(VM_LEVEL_0_NPAGES, NBPOPMAP)
 122 #define NPOPMAP_MAX     howmany(VM_LEVEL_0_NPAGES_MAX, NBPOPMAP)
 123
 124 /*
 125  * Number of elapsed ticks before we update the LRU queue position.  Used
 126  * to reduce contention and churn on the list.
 127  */
 128 #define PARTPOPSLOP     1
 129
 130 /*
 131  * Clear a bit in the population map.
 132  */
 133 static __inline void
 134 popmap_clear(popmap_t popmap[], int i)
 135 {
 136
 137         popmap[i / NBPOPMAP] &= ~(1UL << (i % NBPOPMAP));
 138 }
 139
 140 /*
 141  * Set a bit in the population map.
 142  */
 143 static __inline void
 144 popmap_set(popmap_t popmap[], int i)
 145 {
 146
 147         popmap[i / NBPOPMAP] |= 1UL << (i % NBPOPMAP);
 148 }
 149
 150 /*
 151  * Is a bit in the population map clear?
 152  */
 153 static __inline boolean_t
 154 popmap_is_clear(popmap_t popmap[], int i)
 155 {
 156
 157         return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) == 0);
 158 }
 159
 160 /*
 161  * Is a bit in the population map set?
 162  */
 163 static __inline boolean_t
 164 popmap_is_set(popmap_t popmap[], int i)
 165 {
 166
 167         return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) != 0);
 168 }
 169
 170 /*
 171  * The reservation structure
 172  *
 173  * A reservation structure is constructed whenever a large physical page is
 174  * speculatively allocated to an object.  The reservation provides the small
 175  * physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets
 176  * within that object.  The reservation's "popcnt" tracks the number of these
 177  * small physical pages that are in use at any given time.  When and if the
 178  * reservation is not fully utilized, it appears in the queue of partially
 179  * populated reservations.  The reservation always appears on the containing
 180  * object's list of reservations.
 181  *
 182  * A partially populated reservation can be broken and reclaimed at any time.
 183  *
 184  * c - constant after boot
 185  * d - vm_reserv_domain_lock
 186  * o - vm_reserv_object_lock
 187  * r - vm_reserv_lock
 188  * s - vm_reserv_domain_scan_lock
 189  */
 190 struct vm_reserv {
 191         struct mtx      lock;                   /* reservation lock. */
 192         TAILQ_ENTRY(vm_reserv) partpopq;        /* (d, r) per-domain queue. */
 193         LIST_ENTRY(vm_reserv) objq;             /* (o, r) object queue */
 194         vm_object_t     object;                 /* (o, r) containing object */
 195         vm_pindex_t     pindex;                 /* (o, r) offset in object */
 196         vm_page_t       pages;                  /* (c) first page  */
 197         uint16_t        popcnt;                 /* (r) # of pages in use */
 198         uint8_t         domain;                 /* (c) NUMA domain. */
 199         char            inpartpopq;             /* (d, r) */
 200         int             lasttick;               /* (r) last pop update tick. */
 201         popmap_t        popmap[NPOPMAP_MAX];    /* (r) bit vector, used pages */
 202 };
 203
 204 TAILQ_HEAD(vm_reserv_queue, vm_reserv);
 205
 206 #define vm_reserv_lockptr(rv)           (&(rv)->lock)
 207 #define vm_reserv_assert_locked(rv)                                     \
 208             mtx_assert(vm_reserv_lockptr(rv), MA_OWNED)
 209 #define vm_reserv_lock(rv)              mtx_lock(vm_reserv_lockptr(rv))
 210 #define vm_reserv_trylock(rv)           mtx_trylock(vm_reserv_lockptr(rv))
 211 #define vm_reserv_unlock(rv)            mtx_unlock(vm_reserv_lockptr(rv))
 212
 213 /*
 214  * The reservation array
 215  *
 216  * This array is analoguous in function to vm_page_array.  It differs in the
 217  * respect that it may contain a greater number of useful reservation
 218  * structures than there are (physical) superpages.  These "invalid"
 219  * reservation structures exist to trade-off space for time in the
 220  * implementation of vm_reserv_from_page().  Invalid reservation structures are
 221  * distinguishable from "valid" reservation structures by inspecting the
 222  * reservation's "pages" field.  Invalid reservation structures have a NULL
 223  * "pages" field.
 224  *
 225  * vm_reserv_from_page() maps a small (physical) page to an element of this
 226  * array by computing a physical reservation number from the page's physical
 227  * address.  The physical reservation number is used as the array index.
 228  *
 229  * An "active" reservation is a valid reservation structure that has a non-NULL
 230  * "object" field and a non-zero "popcnt" field.  In other words, every active
 231  * reservation belongs to a particular object.  Moreover, every active
 232  * reservation has an entry in the containing object's list of reservations.
 233  */
 234 static vm_reserv_t vm_reserv_array;
 235
 236 /*
 237  * The per-domain partially populated reservation queues
 238  *
 239  * These queues enable the fast recovery of an unused free small page from a
 240  * partially populated reservation.  The reservation at the head of a queue
 241  * is the least recently changed, partially populated reservation.
 242  *
 243  * Access to this queue is synchronized by the per-domain reservation lock.
 244  * Threads reclaiming free pages from the queue must hold the per-domain scan
 245  * lock.
 246  */
 247 struct vm_reserv_domain {
 248         struct mtx              lock;
 249         struct vm_reserv_queue  partpop;        /* (d) */
 250         struct vm_reserv        marker;         /* (d, s) scan marker/lock */
 251 } __aligned(CACHE_LINE_SIZE);
 252
 253 static struct vm_reserv_domain vm_rvd[MAXMEMDOM];
 254
 255 #define vm_reserv_domain_lockptr(d)     (&vm_rvd[(d)].lock)
 256 #define vm_reserv_domain_assert_locked(d)       \
 257         mtx_assert(vm_reserv_domain_lockptr(d), MA_OWNED)
 258 #define vm_reserv_domain_lock(d)        mtx_lock(vm_reserv_domain_lockptr(d))
 259 #define vm_reserv_domain_unlock(d)      mtx_unlock(vm_reserv_domain_lockptr(d))
 260
 261 #define vm_reserv_domain_scan_lock(d)   mtx_lock(&vm_rvd[(d)].marker.lock)
 262 #define vm_reserv_domain_scan_unlock(d) mtx_unlock(&vm_rvd[(d)].marker.lock)
 263
 264 static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD, 0, "Reservation Info");
 265
 266 static counter_u64_t vm_reserv_broken = EARLY_COUNTER;
 267 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD,
 268     &vm_reserv_broken, "Cumulative number of broken reservations");
 269
 270 static counter_u64_t vm_reserv_freed = EARLY_COUNTER;
 271 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD,
 272     &vm_reserv_freed, "Cumulative number of freed reservations");
 273
 274 static int sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS);
 275
 276 SYSCTL_PROC(_vm_reserv, OID_AUTO, fullpop, CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RD,
 277     NULL, 0, sysctl_vm_reserv_fullpop, "I", "Current number of full reservations");
 278
 279 static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);
 280
 281 SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
 282     sysctl_vm_reserv_partpopq, "A", "Partially populated reservation queues");
 283
 284 static counter_u64_t vm_reserv_reclaimed = EARLY_COUNTER;
 285 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
 286     &vm_reserv_reclaimed, "Cumulative number of reclaimed reservations");
 287
 288 /*
 289  * The object lock pool is used to synchronize the rvq.  We can not use a
 290  * pool mutex because it is required before malloc works.
 291  *
 292  * The "hash" function could be made faster without divide and modulo.
 293  */
 294 #define VM_RESERV_OBJ_LOCK_COUNT        MAXCPU
 295
 296 struct mtx_padalign vm_reserv_object_mtx[VM_RESERV_OBJ_LOCK_COUNT];
 297
 298 #define vm_reserv_object_lock_idx(object)                       \
 299             (((uintptr_t)object / sizeof(*object)) % VM_RESERV_OBJ_LOCK_COUNT)
 300 #define vm_reserv_object_lock_ptr(object)                       \
 301             &vm_reserv_object_mtx[vm_reserv_object_lock_idx((object))]
 302 #define vm_reserv_object_lock(object)                           \
 303             mtx_lock(vm_reserv_object_lock_ptr((object)))
 304 #define vm_reserv_object_unlock(object)                         \
 305             mtx_unlock(vm_reserv_object_lock_ptr((object)))
 306
 307 static void             vm_reserv_break(vm_reserv_t rv);
 308 static void             vm_reserv_depopulate(vm_reserv_t rv, int index);
 309 static vm_reserv_t      vm_reserv_from_page(vm_page_t m);
 310 static boolean_t        vm_reserv_has_pindex(vm_reserv_t rv,
 311                             vm_pindex_t pindex);
 312 static void             vm_reserv_populate(vm_reserv_t rv, int index);
 313 static void             vm_reserv_reclaim(vm_reserv_t rv);
 314
 315 /*
 316  * Returns the current number of full reservations.
 317  *
 318  * Since the number of full reservations is computed without acquiring any
 319  * locks, the returned value is inexact.
 320  */
 321 static int
 322 sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS)
 323 {
 324         vm_paddr_t paddr;
 325         struct vm_phys_seg *seg;
 326         vm_reserv_t rv;
 327         int fullpop, segind;
 328
 329         fullpop = 0;
 330         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 331                 seg = &vm_phys_segs[segind];
 332                 paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
 333                 while (paddr + VM_LEVEL_0_SIZE > paddr && paddr +
 334                     VM_LEVEL_0_SIZE <= seg->end) {
 335                         rv = &vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT];
 336                         fullpop += rv->popcnt == VM_LEVEL_0_NPAGES;
 337                         paddr += VM_LEVEL_0_SIZE;
 338                 }
 339         }
 340         return (sysctl_handle_int(oidp, &fullpop, 0, req));
 341 }
 342
 343 /*
 344  * Describes the current state of the partially populated reservation queue.
 345  */
 346 static int
 347 sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
 348 {
 349         struct sbuf sbuf;
 350         vm_reserv_t rv;
 351         int counter, error, domain, level, unused_pages;
 352
 353         error = sysctl_wire_old_buffer(req, 0);
 354         if (error != 0)
 355                 return (error);
 356         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 357         sbuf_printf(&sbuf, "\nDOMAIN    LEVEL     SIZE  NUMBER\n\n");
 358         for (domain = 0; domain < vm_ndomains; domain++) {
 359                 for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
 360                         counter = 0;
 361                         unused_pages = 0;
 362                         vm_reserv_domain_lock(domain);
 363                         TAILQ_FOREACH(rv, &vm_rvd[domain].partpop, partpopq) {
 364                                 if (rv == &vm_rvd[domain].marker)
 365                                         continue;
 366                                 counter++;
 367                                 unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
 368                         }
 369                         vm_reserv_domain_unlock(domain);
 370                         sbuf_printf(&sbuf, "%6d, %7d, %6dK, %6d\n",
 371                             domain, level,
 372                             unused_pages * ((int)PAGE_SIZE / 1024), counter);
 373                 }
 374         }
 375         error = sbuf_finish(&sbuf);
 376         sbuf_delete(&sbuf);
 377         return (error);
 378 }
 379
 380 /*
 381  * Remove a reservation from the object's objq.
 382  */
 383 static void
 384 vm_reserv_remove(vm_reserv_t rv)
 385 {
 386         vm_object_t object;
 387
 388         vm_reserv_assert_locked(rv);
 389         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
 390             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
 391         KASSERT(rv->object != NULL,
 392             ("vm_reserv_remove: reserv %p is free", rv));
 393         KASSERT(!rv->inpartpopq,
 394             ("vm_reserv_remove: reserv %p's inpartpopq is TRUE", rv));
 395         object = rv->object;
 396         vm_reserv_object_lock(object);
 397         LIST_REMOVE(rv, objq);
 398         rv->object = NULL;
 399         vm_reserv_object_unlock(object);
 400 }
 401
 402 /*
 403  * Insert a new reservation into the object's objq.
 404  */
 405 static void
 406 vm_reserv_insert(vm_reserv_t rv, vm_object_t object, vm_pindex_t pindex)
 407 {
 408         int i;
 409
 410         vm_reserv_assert_locked(rv);
 411         CTR6(KTR_VM,
 412             "%s: rv %p(%p) object %p new %p popcnt %d",
 413             __FUNCTION__, rv, rv->pages, rv->object, object,
 414            rv->popcnt);
 415         KASSERT(rv->object == NULL,
 416             ("vm_reserv_insert: reserv %p isn't free", rv));
 417         KASSERT(rv->popcnt == 0,
 418             ("vm_reserv_insert: reserv %p's popcnt is corrupted", rv));
 419         KASSERT(!rv->inpartpopq,
 420             ("vm_reserv_insert: reserv %p's inpartpopq is TRUE", rv));
 421         for (i = 0; i < NPOPMAP; i++)
 422                 KASSERT(rv->popmap[i] == 0,
 423                     ("vm_reserv_insert: reserv %p's popmap is corrupted", rv));
 424         vm_reserv_object_lock(object);
 425         rv->pindex = pindex;
 426         rv->object = object;
 427         rv->lasttick = ticks;
 428         LIST_INSERT_HEAD(&object->rvq, rv, objq);
 429         vm_reserv_object_unlock(object);
 430 }
 431
 432 /*
 433  * Reduces the given reservation's population count.  If the population count
 434  * becomes zero, the reservation is destroyed.  Additionally, moves the
 435  * reservation to the tail of the partially populated reservation queue if the
 436  * population count is non-zero.
 437  */
 438 static void
 439 vm_reserv_depopulate(vm_reserv_t rv, int index)
 440 {
 441         struct vm_domain *vmd;
 442
 443         vm_reserv_assert_locked(rv);
 444         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
 445             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
 446         KASSERT(rv->object != NULL,
 447             ("vm_reserv_depopulate: reserv %p is free", rv));
 448         KASSERT(popmap_is_set(rv->popmap, index),
 449             ("vm_reserv_depopulate: reserv %p's popmap[%d] is clear", rv,
 450             index));
 451         KASSERT(rv->popcnt > 0,
 452             ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
 453         KASSERT(rv->domain < vm_ndomains,
 454             ("vm_reserv_depopulate: reserv %p's domain is corrupted %d",
 455             rv, rv->domain));
 456         if (rv->popcnt == VM_LEVEL_0_NPAGES) {
 457                 KASSERT(rv->pages->psind == 1,
 458                     ("vm_reserv_depopulate: reserv %p is already demoted",
 459                     rv));
 460                 rv->pages->psind = 0;
 461         }
 462         popmap_clear(rv->popmap, index);
 463         rv->popcnt--;
 464         if ((unsigned)(ticks - rv->lasttick) >= PARTPOPSLOP ||
 465             rv->popcnt == 0) {
 466                 vm_reserv_domain_lock(rv->domain);
 467                 if (rv->inpartpopq) {
 468                         TAILQ_REMOVE(&vm_rvd[rv->domain].partpop, rv, partpopq);
 469                         rv->inpartpopq = FALSE;
 470                 }
 471                 if (rv->popcnt != 0) {
 472                         rv->inpartpopq = TRUE;
 473                         TAILQ_INSERT_TAIL(&vm_rvd[rv->domain].partpop, rv,
 474                             partpopq);
 475                 }
 476                 vm_reserv_domain_unlock(rv->domain);
 477                 rv->lasttick = ticks;
 478         }
 479         vmd = VM_DOMAIN(rv->domain);
 480         if (rv->popcnt == 0) {
 481                 vm_reserv_remove(rv);
 482                 vm_domain_free_lock(vmd);
 483                 vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
 484                 vm_domain_free_unlock(vmd);
 485                 counter_u64_add(vm_reserv_freed, 1);
 486         }
 487         vm_domain_freecnt_inc(vmd, 1);
 488 }
 489
 490 /*
 491  * Returns the reservation to which the given page might belong.
 492  */
 493 static __inline vm_reserv_t
 494 vm_reserv_from_page(vm_page_t m)
 495 {
 496
 497         return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
 498 }
 499
 500 /*
 501  * Returns an existing reservation or NULL and initialized successor pointer.
 502  */
 503 static vm_reserv_t
 504 vm_reserv_from_object(vm_object_t object, vm_pindex_t pindex,
 505     vm_page_t mpred, vm_page_t *msuccp)
 506 {
 507         vm_reserv_t rv;
 508         vm_page_t msucc;
 509
 510         msucc = NULL;
 511         if (mpred != NULL) {
 512                 KASSERT(mpred->object == object,
 513                     ("vm_reserv_from_object: object doesn't contain mpred"));
 514                 KASSERT(mpred->pindex < pindex,
 515                     ("vm_reserv_from_object: mpred doesn't precede pindex"));
 516                 rv = vm_reserv_from_page(mpred);
 517                 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
 518                         goto found;
 519                 msucc = TAILQ_NEXT(mpred, listq);
 520         } else
 521                 msucc = TAILQ_FIRST(&object->memq);
 522         if (msucc != NULL) {
 523                 KASSERT(msucc->pindex > pindex,
 524                     ("vm_reserv_from_object: msucc doesn't succeed pindex"));
 525                 rv = vm_reserv_from_page(msucc);
 526                 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
 527                         goto found;
 528         }
 529         rv = NULL;
 530
 531 found:
 532         *msuccp = msucc;
 533
 534         return (rv);
 535 }
 536
 537 /*
 538  * Returns TRUE if the given reservation contains the given page index and
 539  * FALSE otherwise.
 540  */
 541 static __inline boolean_t
 542 vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
 543 {
 544
 545         return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
 546 }
 547
 548 /*
 549  * Increases the given reservation's population count.  Moves the reservation
 550  * to the tail of the partially populated reservation queue.
 551  */
 552 static void
 553 vm_reserv_populate(vm_reserv_t rv, int index)
 554 {
 555
 556         vm_reserv_assert_locked(rv);
 557         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
 558             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
 559         KASSERT(rv->object != NULL,
 560             ("vm_reserv_populate: reserv %p is free", rv));
 561         KASSERT(popmap_is_clear(rv->popmap, index),
 562             ("vm_reserv_populate: reserv %p's popmap[%d] is set", rv,
 563             index));
 564         KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
 565             ("vm_reserv_populate: reserv %p is already full", rv));
 566         KASSERT(rv->pages->psind == 0,
 567             ("vm_reserv_populate: reserv %p is already promoted", rv));
 568         KASSERT(rv->domain < vm_ndomains,
 569             ("vm_reserv_populate: reserv %p's domain is corrupted %d",
 570             rv, rv->domain));
 571         popmap_set(rv->popmap, index);
 572         rv->popcnt++;
 573         if ((unsigned)(ticks - rv->lasttick) < PARTPOPSLOP &&
 574             rv->inpartpopq && rv->popcnt != VM_LEVEL_0_NPAGES)
 575                 return;
 576         rv->lasttick = ticks;
 577         vm_reserv_domain_lock(rv->domain);
 578         if (rv->inpartpopq) {
 579                 TAILQ_REMOVE(&vm_rvd[rv->domain].partpop, rv, partpopq);
 580                 rv->inpartpopq = FALSE;
 581         }
 582         if (rv->popcnt < VM_LEVEL_0_NPAGES) {
 583                 rv->inpartpopq = TRUE;
 584                 TAILQ_INSERT_TAIL(&vm_rvd[rv->domain].partpop, rv, partpopq);
 585         } else {
 586                 KASSERT(rv->pages->psind == 0,
 587                     ("vm_reserv_populate: reserv %p is already promoted",
 588                     rv));
 589                 rv->pages->psind = 1;
 590         }
 591         vm_reserv_domain_unlock(rv->domain);
 592 }
 593
 594 /*
 595  * Allocates a contiguous set of physical pages of the given size "npages"
 596  * from existing or newly created reservations.  All of the physical pages
 597  * must be at or above the given physical address "low" and below the given
 598  * physical address "high".  The given value "alignment" determines the
 599  * alignment of the first physical page in the set.  If the given value
 600  * "boundary" is non-zero, then the set of physical pages cannot cross any
 601  * physical address boundary that is a multiple of that value.  Both
 602  * "alignment" and "boundary" must be a power of two.
 603  *
 604  * The page "mpred" must immediately precede the offset "pindex" within the
 605  * specified object.
 606  *
 607  * The object must be locked.
 608  */
 609 vm_page_t
 610 vm_reserv_alloc_contig(vm_object_t object, vm_pindex_t pindex, int domain,
 611     int req, vm_page_t mpred, u_long npages, vm_paddr_t low, vm_paddr_t high,
 612     u_long alignment, vm_paddr_t boundary)
 613 {
 614         struct vm_domain *vmd;
 615         vm_paddr_t pa, size;
 616         vm_page_t m, m_ret, msucc;
 617         vm_pindex_t first, leftcap, rightcap;
 618         vm_reserv_t rv;
 619         u_long allocpages, maxpages, minpages;
 620         int i, index, n;
 621
 622         VM_OBJECT_ASSERT_WLOCKED(object);
 623         KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
 624
 625         /*
 626          * Is a reservation fundamentally impossible?
 627          */
 628         if (pindex < VM_RESERV_INDEX(object, pindex) ||
 629             pindex + npages > object->size)
 630                 return (NULL);
 631
 632         /*
 633          * All reservations of a particular size have the same alignment.
 634          * Assuming that the first page is allocated from a reservation, the
 635          * least significant bits of its physical address can be determined
 636          * from its offset from the beginning of the reservation and the size
 637          * of the reservation.
 638          *
 639          * Could the specified index within a reservation of the smallest
 640          * possible size satisfy the alignment and boundary requirements?
 641          */
 642         pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
 643         if ((pa & (alignment - 1)) != 0)
 644                 return (NULL);
 645         size = npages << PAGE_SHIFT;
 646         if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
 647                 return (NULL);
 648
 649         /*
 650          * Look for an existing reservation.
 651          */
 652         rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
 653         if (rv != NULL) {
 654                 KASSERT(object != kernel_object || rv->domain == domain,
 655                     ("vm_reserv_alloc_contig: domain mismatch"));
 656                 index = VM_RESERV_INDEX(object, pindex);
 657                 /* Does the allocation fit within the reservation? */
 658                 if (index + npages > VM_LEVEL_0_NPAGES)
 659                         return (NULL);
 660                 domain = rv->domain;
 661                 vmd = VM_DOMAIN(domain);
 662                 vm_reserv_lock(rv);
 663                 /* Handle reclaim race. */
 664                 if (rv->object != object)
 665                         goto out;
 666                 m = &rv->pages[index];
 667                 pa = VM_PAGE_TO_PHYS(m);
 668                 if (pa < low || pa + size > high ||
 669                     (pa & (alignment - 1)) != 0 ||
 670                     ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
 671                         goto out;
 672                 /* Handle vm_page_rename(m, new_object, ...). */
 673                 for (i = 0; i < npages; i++)
 674                         if (popmap_is_set(rv->popmap, index + i))
 675                                 goto out;
 676                 if (!vm_domain_allocate(vmd, req, npages))
 677                         goto out;
 678                 for (i = 0; i < npages; i++)
 679                         vm_reserv_populate(rv, index + i);
 680                 vm_reserv_unlock(rv);
 681                 return (m);
 682 out:
 683                 vm_reserv_unlock(rv);
 684                 return (NULL);
 685         }
 686
 687         /*
 688          * Could at least one reservation fit between the first index to the
 689          * left that can be used ("leftcap") and the first index to the right
 690          * that cannot be used ("rightcap")?
 691          *
 692          * We must synchronize with the reserv object lock to protect the
 693          * pindex/object of the resulting reservations against rename while
 694          * we are inspecting.
 695          */
 696         first = pindex - VM_RESERV_INDEX(object, pindex);
 697         minpages = VM_RESERV_INDEX(object, pindex) + npages;
 698         maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES);
 699         allocpages = maxpages;
 700         vm_reserv_object_lock(object);
 701         if (mpred != NULL) {
 702                 if ((rv = vm_reserv_from_page(mpred))->object != object)
 703                         leftcap = mpred->pindex + 1;
 704                 else
 705                         leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
 706                 if (leftcap > first) {
 707                         vm_reserv_object_unlock(object);
 708                         return (NULL);
 709                 }
 710         }
 711         if (msucc != NULL) {
 712                 if ((rv = vm_reserv_from_page(msucc))->object != object)
 713                         rightcap = msucc->pindex;
 714                 else
 715                         rightcap = rv->pindex;
 716                 if (first + maxpages > rightcap) {
 717                         if (maxpages == VM_LEVEL_0_NPAGES) {
 718                                 vm_reserv_object_unlock(object);
 719                                 return (NULL);
 720                         }
 721
 722                         /*
 723                          * At least one reservation will fit between "leftcap"
 724                          * and "rightcap".  However, a reservation for the
 725                          * last of the requested pages will not fit.  Reduce
 726                          * the size of the upcoming allocation accordingly.
 727                          */
 728                         allocpages = minpages;
 729                 }
 730         }
 731         vm_reserv_object_unlock(object);
 732
 733         /*
 734          * Would the last new reservation extend past the end of the object?
 735          *
 736          * If the object is unlikely to grow don't allocate a reservation for
 737          * the tail.
 738          */
 739         if ((object->flags & OBJ_ANON) == 0 &&
 740             first + maxpages > object->size) {
 741                 if (maxpages == VM_LEVEL_0_NPAGES)
 742                         return (NULL);
 743                 allocpages = minpages;
 744         }
 745
 746         /*
 747          * Allocate the physical pages.  The alignment and boundary specified
 748          * for this allocation may be different from the alignment and
 749          * boundary specified for the requested pages.  For instance, the
 750          * specified index may not be the first page within the first new
 751          * reservation.
 752          */
 753         m = NULL;
 754         vmd = VM_DOMAIN(domain);
 755         if (vm_domain_allocate(vmd, req, npages)) {
 756                 vm_domain_free_lock(vmd);
 757                 m = vm_phys_alloc_contig(domain, allocpages, low, high,
 758                     ulmax(alignment, VM_LEVEL_0_SIZE),
 759                     boundary > VM_LEVEL_0_SIZE ? boundary : 0);
 760                 vm_domain_free_unlock(vmd);
 761                 if (m == NULL) {
 762                         vm_domain_freecnt_inc(vmd, npages);
 763                         return (NULL);
 764                 }
 765         } else
 766                 return (NULL);
 767         KASSERT(vm_phys_domain(m) == domain,
 768             ("vm_reserv_alloc_contig: Page domain does not match requested."));
 769
 770         /*
 771          * The allocated physical pages always begin at a reservation
 772          * boundary, but they do not always end at a reservation boundary.
 773          * Initialize every reservation that is completely covered by the
 774          * allocated physical pages.
 775          */
 776         m_ret = NULL;
 777         index = VM_RESERV_INDEX(object, pindex);
 778         do {
 779                 rv = vm_reserv_from_page(m);
 780                 KASSERT(rv->pages == m,
 781                     ("vm_reserv_alloc_contig: reserv %p's pages is corrupted",
 782                     rv));
 783                 vm_reserv_lock(rv);
 784                 vm_reserv_insert(rv, object, first);
 785                 n = ulmin(VM_LEVEL_0_NPAGES - index, npages);
 786                 for (i = 0; i < n; i++)
 787                         vm_reserv_populate(rv, index + i);
 788                 npages -= n;
 789                 if (m_ret == NULL) {
 790                         m_ret = &rv->pages[index];
 791                         index = 0;
 792                 }
 793                 vm_reserv_unlock(rv);
 794                 m += VM_LEVEL_0_NPAGES;
 795                 first += VM_LEVEL_0_NPAGES;
 796                 allocpages -= VM_LEVEL_0_NPAGES;
 797         } while (allocpages >= VM_LEVEL_0_NPAGES);
 798         return (m_ret);
 799 }
 800
 801 /*
 802  * Allocate a physical page from an existing or newly created reservation.
 803  *
 804  * The page "mpred" must immediately precede the offset "pindex" within the
 805  * specified object.
 806  *
 807  * The object must be locked.
 808  */
 809 vm_page_t
 810 vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex, int domain,
 811     int req, vm_page_t mpred)
 812 {
 813         struct vm_domain *vmd;
 814         vm_page_t m, msucc;
 815         vm_pindex_t first, leftcap, rightcap;
 816         vm_reserv_t rv;
 817         int index;
 818
 819         VM_OBJECT_ASSERT_WLOCKED(object);
 820
 821         /*
 822          * Is a reservation fundamentally impossible?
 823          */
 824         if (pindex < VM_RESERV_INDEX(object, pindex) ||
 825             pindex >= object->size)
 826                 return (NULL);
 827
 828         /*
 829          * Look for an existing reservation.
 830          */
 831         rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
 832         if (rv != NULL) {
 833                 KASSERT(object != kernel_object || rv->domain == domain,
 834                     ("vm_reserv_alloc_page: domain mismatch"));
 835                 domain = rv->domain;
 836                 vmd = VM_DOMAIN(domain);
 837                 index = VM_RESERV_INDEX(object, pindex);
 838                 m = &rv->pages[index];
 839                 vm_reserv_lock(rv);
 840                 /* Handle reclaim race. */
 841                 if (rv->object != object ||
 842                     /* Handle vm_page_rename(m, new_object, ...). */
 843                     popmap_is_set(rv->popmap, index)) {
 844                         m = NULL;
 845                         goto out;
 846                 }
 847                 if (vm_domain_allocate(vmd, req, 1) == 0)
 848                         m = NULL;
 849                 else
 850                         vm_reserv_populate(rv, index);
 851 out:
 852                 vm_reserv_unlock(rv);
 853                 return (m);
 854         }
 855
 856         /*
 857          * Could a reservation fit between the first index to the left that
 858          * can be used and the first index to the right that cannot be used?
 859          *
 860          * We must synchronize with the reserv object lock to protect the
 861          * pindex/object of the resulting reservations against rename while
 862          * we are inspecting.
 863          */
 864         first = pindex - VM_RESERV_INDEX(object, pindex);
 865         vm_reserv_object_lock(object);
 866         if (mpred != NULL) {
 867                 if ((rv = vm_reserv_from_page(mpred))->object != object)
 868                         leftcap = mpred->pindex + 1;
 869                 else
 870                         leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
 871                 if (leftcap > first) {
 872                         vm_reserv_object_unlock(object);
 873                         return (NULL);
 874                 }
 875         }
 876         if (msucc != NULL) {
 877                 if ((rv = vm_reserv_from_page(msucc))->object != object)
 878                         rightcap = msucc->pindex;
 879                 else
 880                         rightcap = rv->pindex;
 881                 if (first + VM_LEVEL_0_NPAGES > rightcap) {
 882                         vm_reserv_object_unlock(object);
 883                         return (NULL);
 884                 }
 885         }
 886         vm_reserv_object_unlock(object);
 887
 888         /*
 889          * Would the last new reservation extend past the end of the object?
 890          *
 891          * If the object is unlikely to grow don't allocate a reservation for
 892          * the tail.
 893          */
 894         if ((object->flags & OBJ_ANON) == 0 &&
 895             first + VM_LEVEL_0_NPAGES > object->size)
 896                 return (NULL);
 897
 898         /*
 899          * Allocate and populate the new reservation.
 900          */
 901         m = NULL;
 902         vmd = VM_DOMAIN(domain);
 903         if (vm_domain_allocate(vmd, req, 1)) {
 904                 vm_domain_free_lock(vmd);
 905                 m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT,
 906                     VM_LEVEL_0_ORDER);
 907                 vm_domain_free_unlock(vmd);
 908                 if (m == NULL) {
 909                         vm_domain_freecnt_inc(vmd, 1);
 910                         return (NULL);
 911                 }
 912         } else
 913                 return (NULL);
 914         rv = vm_reserv_from_page(m);
 915         vm_reserv_lock(rv);
 916         KASSERT(rv->pages == m,
 917             ("vm_reserv_alloc_page: reserv %p's pages is corrupted", rv));
 918         vm_reserv_insert(rv, object, first);
 919         index = VM_RESERV_INDEX(object, pindex);
 920         vm_reserv_populate(rv, index);
 921         vm_reserv_unlock(rv);
 922
 923         return (&rv->pages[index]);
 924 }
 925
 926 /*
 927  * Breaks the given reservation.  All free pages in the reservation
 928  * are returned to the physical memory allocator.  The reservation's
 929  * population count and map are reset to their initial state.
 930  *
 931  * The given reservation must not be in the partially populated reservation
 932  * queue.
 933  */
 934 static void
 935 vm_reserv_break(vm_reserv_t rv)
 936 {
 937         u_long changes;
 938         int bitpos, hi, i, lo;
 939
 940         vm_reserv_assert_locked(rv);
 941         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
 942             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
 943         vm_reserv_remove(rv);
 944         rv->pages->psind = 0;
 945         hi = lo = -1;
 946         for (i = 0; i <= NPOPMAP; i++) {
 947                 /*
 948                  * "changes" is a bitmask that marks where a new sequence of
 949                  * 0s or 1s begins in popmap[i], with last bit in popmap[i-1]
 950                  * considered to be 1 if and only if lo == hi.  The bits of
 951                  * popmap[-1] and popmap[NPOPMAP] are considered all 1s.
 952                  */
 953                 if (i == NPOPMAP)
 954                         changes = lo != hi;
 955                 else {
 956                         changes = rv->popmap[i];
 957                         changes ^= (changes << 1) | (lo == hi);
 958                         rv->popmap[i] = 0;
 959                 }
 960                 while (changes != 0) {
 961                         /*
 962                          * If the next change marked begins a run of 0s, set
 963                          * lo to mark that position.  Otherwise set hi and
 964                          * free pages from lo up to hi.
 965                          */
 966                         bitpos = ffsl(changes) - 1;
 967                         changes ^= 1UL << bitpos;
 968                         if (lo == hi)
 969                                 lo = NBPOPMAP * i + bitpos;
 970                         else {
 971                                 hi = NBPOPMAP * i + bitpos;
 972                                 vm_domain_free_lock(VM_DOMAIN(rv->domain));
 973                                 vm_phys_enqueue_contig(&rv->pages[lo], hi - lo);
 974                                 vm_domain_free_unlock(VM_DOMAIN(rv->domain));
 975                                 lo = hi;
 976                         }
 977                 }
 978         }
 979         rv->popcnt = 0;
 980         counter_u64_add(vm_reserv_broken, 1);
 981 }
 982
 983 /*
 984  * Breaks all reservations belonging to the given object.
 985  */
 986 void
 987 vm_reserv_break_all(vm_object_t object)
 988 {
 989         vm_reserv_t rv;
 990
 991         /*
 992          * This access of object->rvq is unsynchronized so that the
 993          * object rvq lock can nest after the domain_free lock.  We
 994          * must check for races in the results.  However, the object
 995          * lock prevents new additions, so we are guaranteed that when
 996          * it returns NULL the object is properly empty.
 997          */
 998         while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
 999                 vm_reserv_lock(rv);
1000                 /* Reclaim race. */
1001                 if (rv->object != object) {
1002                         vm_reserv_unlock(rv);
1003                         continue;
1004                 }
1005                 vm_reserv_domain_lock(rv->domain);
1006                 if (rv->inpartpopq) {
1007                         TAILQ_REMOVE(&vm_rvd[rv->domain].partpop, rv, partpopq);
1008                         rv->inpartpopq = FALSE;
1009                 }
1010                 vm_reserv_domain_unlock(rv->domain);
1011                 vm_reserv_break(rv);
1012                 vm_reserv_unlock(rv);
1013         }
1014 }
1015
1016 /*
1017  * Frees the given page if it belongs to a reservation.  Returns TRUE if the
1018  * page is freed and FALSE otherwise.
1019  */
1020 boolean_t
1021 vm_reserv_free_page(vm_page_t m)
1022 {
1023         vm_reserv_t rv;
1024         boolean_t ret;
1025
1026         rv = vm_reserv_from_page(m);
1027         if (rv->object == NULL)
1028                 return (FALSE);
1029         vm_reserv_lock(rv);
1030         /* Re-validate after lock. */
1031         if (rv->object != NULL) {
1032                 vm_reserv_depopulate(rv, m - rv->pages);
1033                 ret = TRUE;
1034         } else
1035                 ret = FALSE;
1036         vm_reserv_unlock(rv);
1037
1038         return (ret);
1039 }
1040
1041 /*
1042  * Initializes the reservation management system.  Specifically, initializes
1043  * the reservation array.
1044  *
1045  * Requires that vm_page_array and first_page are initialized!
1046  */
1047 void
1048 vm_reserv_init(void)
1049 {
1050         vm_paddr_t paddr;
1051         struct vm_phys_seg *seg;
1052         struct vm_reserv *rv;
1053         struct vm_reserv_domain *rvd;
1054         int i, j, segind;
1055
1056         /*
1057          * Initialize the reservation array.  Specifically, initialize the
1058          * "pages" field for every element that has an underlying superpage.
1059          */
1060         for (segind = 0; segind < vm_phys_nsegs; segind++) {
1061                 seg = &vm_phys_segs[segind];
1062                 paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
1063                 while (paddr + VM_LEVEL_0_SIZE > paddr && paddr +
1064                     VM_LEVEL_0_SIZE <= seg->end) {
1065                         rv = &vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT];
1066                         rv->pages = PHYS_TO_VM_PAGE(paddr);
1067                         rv->domain = seg->domain;
1068                         mtx_init(&rv->lock, "vm reserv", NULL, MTX_DEF);
1069                         paddr += VM_LEVEL_0_SIZE;
1070                 }
1071         }
1072         for (i = 0; i < MAXMEMDOM; i++) {
1073                 rvd = &vm_rvd[i];
1074                 mtx_init(&rvd->lock, "vm reserv domain", NULL, MTX_DEF);
1075                 TAILQ_INIT(&rvd->partpop);
1076                 mtx_init(&rvd->marker.lock, "vm reserv marker", NULL, MTX_DEF);
1077
1078                 /*
1079                  * Fully populated reservations should never be present in the
1080                  * partially populated reservation queues.
1081                  */
1082                 rvd->marker.popcnt = VM_LEVEL_0_NPAGES;
1083                 for (j = 0; j < NBPOPMAP; j++)
1084                         popmap_set(rvd->marker.popmap, j);
1085         }
1086
1087         for (i = 0; i < VM_RESERV_OBJ_LOCK_COUNT; i++)
1088                 mtx_init(&vm_reserv_object_mtx[i], "resv obj lock", NULL,
1089                     MTX_DEF);
1090 }
1091
1092 /*
1093  * Returns true if the given page belongs to a reservation and that page is
1094  * free.  Otherwise, returns false.
1095  */
1096 bool
1097 vm_reserv_is_page_free(vm_page_t m)
1098 {
1099         vm_reserv_t rv;
1100
1101         rv = vm_reserv_from_page(m);
1102         if (rv->object == NULL)
1103                 return (false);
1104         return (popmap_is_clear(rv->popmap, m - rv->pages));
1105 }
1106
1107 /*
1108  * If the given page belongs to a reservation, returns the level of that
1109  * reservation.  Otherwise, returns -1.
1110  */
1111 int
1112 vm_reserv_level(vm_page_t m)
1113 {
1114         vm_reserv_t rv;
1115
1116         rv = vm_reserv_from_page(m);
1117         return (rv->object != NULL ? 0 : -1);
1118 }
1119
1120 /*
1121  * Returns a reservation level if the given page belongs to a fully populated
1122  * reservation and -1 otherwise.
1123  */
1124 int
1125 vm_reserv_level_iffullpop(vm_page_t m)
1126 {
1127         vm_reserv_t rv;
1128
1129         rv = vm_reserv_from_page(m);
1130         return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
1131 }
1132
1133 /*
1134  * Remove a partially populated reservation from the queue.
1135  */
1136 static void
1137 vm_reserv_dequeue(vm_reserv_t rv)
1138 {
1139
1140         vm_reserv_domain_assert_locked(rv->domain);
1141         vm_reserv_assert_locked(rv);
1142         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
1143             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
1144         KASSERT(rv->inpartpopq,
1145             ("vm_reserv_reclaim: reserv %p's inpartpopq is FALSE", rv));
1146
1147         TAILQ_REMOVE(&vm_rvd[rv->domain].partpop, rv, partpopq);
1148         rv->inpartpopq = FALSE;
1149 }
1150
1151 /*
1152  * Breaks the given partially populated reservation, releasing its free pages
1153  * to the physical memory allocator.
1154  */
1155 static void
1156 vm_reserv_reclaim(vm_reserv_t rv)
1157 {
1158
1159         vm_reserv_assert_locked(rv);
1160         CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
1161             __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
1162         if (rv->inpartpopq) {
1163                 vm_reserv_domain_lock(rv->domain);
1164                 vm_reserv_dequeue(rv);
1165                 vm_reserv_domain_unlock(rv->domain);
1166         }
1167         vm_reserv_break(rv);
1168         counter_u64_add(vm_reserv_reclaimed, 1);
1169 }
1170
1171 /*
1172  * Breaks a reservation near the head of the partially populated reservation
1173  * queue, releasing its free pages to the physical memory allocator.  Returns
1174  * TRUE if a reservation is broken and FALSE otherwise.
1175  */
1176 bool
1177 vm_reserv_reclaim_inactive(int domain)
1178 {
1179         vm_reserv_t rv;
1180
1181         vm_reserv_domain_lock(domain);
1182         TAILQ_FOREACH(rv, &vm_rvd[domain].partpop, partpopq) {
1183                 /*
1184                  * A locked reservation is likely being updated or reclaimed,
1185                  * so just skip ahead.
1186                  */
1187                 if (rv != &vm_rvd[domain].marker && vm_reserv_trylock(rv)) {
1188                         vm_reserv_dequeue(rv);
1189                         break;
1190                 }
1191         }
1192         vm_reserv_domain_unlock(domain);
1193         if (rv != NULL) {
1194                 vm_reserv_reclaim(rv);
1195                 vm_reserv_unlock(rv);
1196                 return (true);
1197         }
1198         return (false);
1199 }
1200
1201 /*
1202  * Determine whether this reservation has free pages that satisfy the given
1203  * request for contiguous physical memory.  Start searching from the lower
1204  * bound, defined by low_index.
1205  */
1206 static bool
1207 vm_reserv_test_contig(vm_reserv_t rv, u_long npages, vm_paddr_t low,
1208     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1209 {
1210         vm_paddr_t pa, size;
1211         u_long changes;
1212         int bitpos, bits_left, i, hi, lo, n;
1213
1214         vm_reserv_assert_locked(rv);
1215         size = npages << PAGE_SHIFT;
1216         pa = VM_PAGE_TO_PHYS(&rv->pages[0]);
1217         lo = (pa < low) ?
1218             ((low + PAGE_MASK - pa) >> PAGE_SHIFT) : 0;
1219         i = lo / NBPOPMAP;
1220         changes = rv->popmap[i] | ((1UL << (lo % NBPOPMAP)) - 1);
1221         hi = (pa + VM_LEVEL_0_SIZE > high) ?
1222             ((high + PAGE_MASK - pa) >> PAGE_SHIFT) : VM_LEVEL_0_NPAGES;
1223         n = hi / NBPOPMAP;
1224         bits_left = hi % NBPOPMAP;
1225         hi = lo = -1;
1226         for (;;) {
1227                 /*
1228                  * "changes" is a bitmask that marks where a new sequence of
1229                  * 0s or 1s begins in popmap[i], with last bit in popmap[i-1]
1230                  * considered to be 1 if and only if lo == hi.  The bits of
1231                  * popmap[-1] and popmap[NPOPMAP] are considered all 1s.
1232                  */
1233                 changes ^= (changes << 1) | (lo == hi);
1234                 while (changes != 0) {
1235                         /*
1236                          * If the next change marked begins a run of 0s, set
1237                          * lo to mark that position.  Otherwise set hi and
1238                          * look for a satisfactory first page from lo up to hi.
1239                          */
1240                         bitpos = ffsl(changes) - 1;
1241                         changes ^= 1UL << bitpos;
1242                         if (lo == hi) {
1243                                 lo = NBPOPMAP * i + bitpos;
1244                                 continue;
1245                         }
1246                         hi = NBPOPMAP * i + bitpos;
1247                         pa = VM_PAGE_TO_PHYS(&rv->pages[lo]);
1248                         if ((pa & (alignment - 1)) != 0) {
1249                                 /* Skip to next aligned page. */
1250                                 lo += (((pa - 1) | (alignment - 1)) + 1) >>
1251                                     PAGE_SHIFT;
1252                                 if (lo >= VM_LEVEL_0_NPAGES)
1253                                         return (false);
1254                                 pa = VM_PAGE_TO_PHYS(&rv->pages[lo]);
1255                         }
1256                         if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0) {
1257                                 /* Skip to next boundary-matching page. */
1258                                 lo += (((pa - 1) | (boundary - 1)) + 1) >>
1259                                     PAGE_SHIFT;
1260                                 if (lo >= VM_LEVEL_0_NPAGES)
1261                                         return (false);
1262                                 pa = VM_PAGE_TO_PHYS(&rv->pages[lo]);
1263                         }
1264                         if (lo * PAGE_SIZE + size <= hi * PAGE_SIZE)
1265                                 return (true);
1266                         lo = hi;
1267                 }
1268                 if (++i < n)
1269                         changes = rv->popmap[i];
1270                 else if (i == n)
1271                         changes = bits_left == 0 ? -1UL :
1272                             (rv->popmap[n] | (-1UL << bits_left));
1273                 else
1274                         return (false);
1275         }
1276 }
1277
1278 /*
1279  * Searches the partially populated reservation queue for the least recently
1280  * changed reservation with free pages that satisfy the given request for
1281  * contiguous physical memory.  If a satisfactory reservation is found, it is
1282  * broken.  Returns true if a reservation is broken and false otherwise.
1283  */
1284 bool
1285 vm_reserv_reclaim_contig(int domain, u_long npages, vm_paddr_t low,
1286     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1287 {
1288         struct vm_reserv_queue *queue;
1289         vm_paddr_t pa, size;
1290         vm_reserv_t marker, rv, rvn;
1291
1292         if (npages > VM_LEVEL_0_NPAGES - 1)
1293                 return (false);
1294         marker = &vm_rvd[domain].marker;
1295         queue = &vm_rvd[domain].partpop;
1296         size = npages << PAGE_SHIFT;
1297
1298         vm_reserv_domain_scan_lock(domain);
1299         vm_reserv_domain_lock(domain);
1300         TAILQ_FOREACH_SAFE(rv, queue, partpopq, rvn) {
1301                 pa = VM_PAGE_TO_PHYS(&rv->pages[0]);
1302                 if (pa + VM_LEVEL_0_SIZE - size < low) {
1303                         /* This entire reservation is too low; go to next. */
1304                         continue;
1305                 }
1306                 if (pa + size > high) {
1307                         /* This entire reservation is too high; go to next. */
1308                         continue;
1309                 }
1310
1311                 if (vm_reserv_trylock(rv) == 0) {
1312                         TAILQ_INSERT_AFTER(queue, rv, marker, partpopq);
1313                         vm_reserv_domain_unlock(domain);
1314                         vm_reserv_lock(rv);
1315                         if (!rv->inpartpopq ||
1316                             TAILQ_NEXT(rv, partpopq) != marker) {
1317                                 vm_reserv_unlock(rv);
1318                                 vm_reserv_domain_lock(domain);
1319                                 rvn = TAILQ_NEXT(marker, partpopq);
1320                                 TAILQ_REMOVE(queue, marker, partpopq);
1321                                 continue;
1322                         }
1323                         vm_reserv_domain_lock(domain);
1324                         TAILQ_REMOVE(queue, marker, partpopq);
1325                 }
1326                 vm_reserv_domain_unlock(domain);
1327                 if (vm_reserv_test_contig(rv, npages, low, high,
1328                     alignment, boundary)) {
1329                         vm_reserv_domain_scan_unlock(domain);
1330                         vm_reserv_reclaim(rv);
1331                         vm_reserv_unlock(rv);
1332                         return (true);
1333                 }
1334                 vm_reserv_unlock(rv);
1335                 vm_reserv_domain_lock(domain);
1336         }
1337         vm_reserv_domain_unlock(domain);
1338         vm_reserv_domain_scan_unlock(domain);
1339         return (false);
1340 }
1341
1342 /*
1343  * Transfers the reservation underlying the given page to a new object.
1344  *
1345  * The object must be locked.
1346  */
1347 void
1348 vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
1349     vm_pindex_t old_object_offset)
1350 {
1351         vm_reserv_t rv;
1352
1353         VM_OBJECT_ASSERT_WLOCKED(new_object);
1354         rv = vm_reserv_from_page(m);
1355         if (rv->object == old_object) {
1356                 vm_reserv_lock(rv);
1357                 CTR6(KTR_VM,
1358                     "%s: rv %p object %p new %p popcnt %d inpartpop %d",
1359                     __FUNCTION__, rv, rv->object, new_object, rv->popcnt,
1360                     rv->inpartpopq);
1361                 if (rv->object == old_object) {
1362                         vm_reserv_object_lock(old_object);
1363                         rv->object = NULL;
1364                         LIST_REMOVE(rv, objq);
1365                         vm_reserv_object_unlock(old_object);
1366                         vm_reserv_object_lock(new_object);
1367                         rv->object = new_object;
1368                         rv->pindex -= old_object_offset;
1369                         LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
1370                         vm_reserv_object_unlock(new_object);
1371                 }
1372                 vm_reserv_unlock(rv);
1373         }
1374 }
1375
1376 /*
1377  * Returns the size (in bytes) of a reservation of the specified level.
1378  */
1379 int
1380 vm_reserv_size(int level)
1381 {
1382
1383         switch (level) {
1384         case 0:
1385                 return (VM_LEVEL_0_SIZE);
1386         case -1:
1387                 return (PAGE_SIZE);
1388         default:
1389                 return (0);
1390         }
1391 }
1392
1393 /*
1394  * Allocates the virtual and physical memory required by the reservation
1395  * management system's data structures, in particular, the reservation array.
1396  */
1397 vm_paddr_t
1398 vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end)
1399 {
1400         vm_paddr_t new_end, high_water;
1401         size_t size;
1402         int i;
1403
1404         high_water = phys_avail[1];
1405         for (i = 0; i < vm_phys_nsegs; i++) {
1406                 if (vm_phys_segs[i].end > high_water)
1407                         high_water = vm_phys_segs[i].end;
1408         }
1409
1410         /* Skip the first chunk.  It is already accounted for. */
1411         for (i = 2; phys_avail[i + 1] != 0; i += 2) {
1412                 if (phys_avail[i + 1] > high_water)
1413                         high_water = phys_avail[i + 1];
1414         }
1415
1416         /*
1417          * Calculate the size (in bytes) of the reservation array.  Round up
1418          * from "high_water" because every small page is mapped to an element
1419          * in the reservation array based on its physical address.  Thus, the
1420          * number of elements in the reservation array can be greater than the
1421          * number of superpages.
1422          */
1423         size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv);
1424
1425         /*
1426          * Allocate and map the physical memory for the reservation array.  The
1427          * next available virtual address is returned by reference.
1428          */
1429         new_end = end - round_page(size);
1430         vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
1431             VM_PROT_READ | VM_PROT_WRITE);
1432         bzero(vm_reserv_array, size);
1433
1434         /*
1435          * Return the next available physical address.
1436          */
1437         return (new_end);
1438 }
1439
1440 /*
1441  * Initializes the reservation management system.  Specifically, initializes
1442  * the reservation counters.
1443  */
1444 static void
1445 vm_reserv_counter_init(void *unused)
1446 {
1447
1448         vm_reserv_freed = counter_u64_alloc(M_WAITOK);
1449         vm_reserv_broken = counter_u64_alloc(M_WAITOK);
1450         vm_reserv_reclaimed = counter_u64_alloc(M_WAITOK);
1451 }
1452 SYSINIT(vm_reserv_counter_init, SI_SUB_CPU, SI_ORDER_ANY,
1453     vm_reserv_counter_init, NULL);
1454
1455 /*
1456  * Returns the superpage containing the given page.
1457  */
1458 vm_page_t
1459 vm_reserv_to_superpage(vm_page_t m)
1460 {
1461         vm_reserv_t rv;
1462
1463         VM_OBJECT_ASSERT_LOCKED(m->object);
1464         rv = vm_reserv_from_page(m);
1465         if (rv->object == m->object && rv->popcnt == VM_LEVEL_0_NPAGES)
1466                 m = rv->pages;
1467         else
1468                 m = NULL;
1469
1470         return (m);
1471 }
1472
1473 #endif  /* VM_NRESERVLEVEL > 0 */