sys/vm/vm_phys.c

   1 /*-
   2  * Copyright (c) 2002-2006 Rice University
   3  * Copyright (c) 2007 Alan L. Cox <alc@cs.rice.edu>
   4  * All rights reserved.
   5  *
   6  * This software was developed for the FreeBSD Project by Alan L. Cox,
   7  * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  19  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  20  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  21  * A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT
  22  * HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  23  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
  24  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
  25  * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  26  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY
  28  * WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
  29  * POSSIBILITY OF SUCH DAMAGE.
  30  */
  31
  32 /*
  33  *      Physical memory system implementation
  34  *
  35  * Any external functions defined by this module are only to be used by the
  36  * virtual memory system.
  37  */
  38
  39 #include <sys/cdefs.h>
  40 __FBSDID("$FreeBSD$");
  41
  42 #include "opt_ddb.h"
  43 #include "opt_vm.h"
  44
  45 #include <sys/param.h>
  46 #include <sys/systm.h>
  47 #include <sys/lock.h>
  48 #include <sys/kernel.h>
  49 #include <sys/malloc.h>
  50 #include <sys/mutex.h>
  51 #include <sys/proc.h>
  52 #include <sys/queue.h>
  53 #include <sys/rwlock.h>
  54 #include <sys/sbuf.h>
  55 #include <sys/sysctl.h>
  56 #include <sys/tree.h>
  57 #include <sys/vmmeter.h>
  58 #include <sys/seq.h>
  59
  60 #include <ddb/ddb.h>
  61
  62 #include <vm/vm.h>
  63 #include <vm/vm_param.h>
  64 #include <vm/vm_kern.h>
  65 #include <vm/vm_object.h>
  66 #include <vm/vm_page.h>
  67 #include <vm/vm_phys.h>
  68
  69 #include <vm/vm_domain.h>
  70
  71 _Static_assert(sizeof(long) * NBBY >= VM_PHYSSEG_MAX,
  72     "Too many physsegs.");
  73
  74 #ifdef VM_NUMA_ALLOC
  75 struct mem_affinity *mem_affinity;
  76 int *mem_locality;
  77 #endif
  78
  79 int vm_ndomains = 1;
  80
  81 struct vm_phys_seg vm_phys_segs[VM_PHYSSEG_MAX];
  82 int vm_phys_nsegs;
  83
  84 struct vm_phys_fictitious_seg;
  85 static int vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *,
  86     struct vm_phys_fictitious_seg *);
  87
  88 RB_HEAD(fict_tree, vm_phys_fictitious_seg) vm_phys_fictitious_tree =
  89     RB_INITIALIZER(_vm_phys_fictitious_tree);
  90
  91 struct vm_phys_fictitious_seg {
  92         RB_ENTRY(vm_phys_fictitious_seg) node;
  93         /* Memory region data */
  94         vm_paddr_t      start;
  95         vm_paddr_t      end;
  96         vm_page_t       first_page;
  97 };
  98
  99 RB_GENERATE_STATIC(fict_tree, vm_phys_fictitious_seg, node,
 100     vm_phys_fictitious_cmp);
 101
 102 static struct rwlock vm_phys_fictitious_reg_lock;
 103 MALLOC_DEFINE(M_FICT_PAGES, "vm_fictitious", "Fictitious VM pages");
 104
 105 static struct vm_freelist
 106     vm_phys_free_queues[MAXMEMDOM][VM_NFREELIST][VM_NFREEPOOL][VM_NFREEORDER];
 107
 108 static int vm_nfreelists;
 109
 110 /*
 111  * Provides the mapping from VM_FREELIST_* to free list indices (flind).
 112  */
 113 static int vm_freelist_to_flind[VM_NFREELIST];
 114
 115 CTASSERT(VM_FREELIST_DEFAULT == 0);
 116
 117 #ifdef VM_FREELIST_ISADMA
 118 #define VM_ISADMA_BOUNDARY      16777216
 119 #endif
 120 #ifdef VM_FREELIST_DMA32
 121 #define VM_DMA32_BOUNDARY       ((vm_paddr_t)1 << 32)
 122 #endif
 123
 124 /*
 125  * Enforce the assumptions made by vm_phys_add_seg() and vm_phys_init() about
 126  * the ordering of the free list boundaries.
 127  */
 128 #if defined(VM_ISADMA_BOUNDARY) && defined(VM_LOWMEM_BOUNDARY)
 129 CTASSERT(VM_ISADMA_BOUNDARY < VM_LOWMEM_BOUNDARY);
 130 #endif
 131 #if defined(VM_LOWMEM_BOUNDARY) && defined(VM_DMA32_BOUNDARY)
 132 CTASSERT(VM_LOWMEM_BOUNDARY < VM_DMA32_BOUNDARY);
 133 #endif
 134
 135 static int cnt_prezero;
 136 SYSCTL_INT(_vm_stats_misc, OID_AUTO, cnt_prezero, CTLFLAG_RD,
 137     &cnt_prezero, 0, "The number of physical pages prezeroed at idle time");
 138
 139 static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS);
 140 SYSCTL_OID(_vm, OID_AUTO, phys_free, CTLTYPE_STRING | CTLFLAG_RD,
 141     NULL, 0, sysctl_vm_phys_free, "A", "Phys Free Info");
 142
 143 static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS);
 144 SYSCTL_OID(_vm, OID_AUTO, phys_segs, CTLTYPE_STRING | CTLFLAG_RD,
 145     NULL, 0, sysctl_vm_phys_segs, "A", "Phys Seg Info");
 146
 147 #ifdef VM_NUMA_ALLOC
 148 static int sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS);
 149 SYSCTL_OID(_vm, OID_AUTO, phys_locality, CTLTYPE_STRING | CTLFLAG_RD,
 150     NULL, 0, sysctl_vm_phys_locality, "A", "Phys Locality Info");
 151 #endif
 152
 153 SYSCTL_INT(_vm, OID_AUTO, ndomains, CTLFLAG_RD,
 154     &vm_ndomains, 0, "Number of physical memory domains available.");
 155
 156 /*
 157  * Default to first-touch + round-robin.
 158  */
 159 static struct mtx vm_default_policy_mtx;
 160 MTX_SYSINIT(vm_default_policy, &vm_default_policy_mtx, "default policy mutex",
 161     MTX_DEF);
 162 #ifdef VM_NUMA_ALLOC
 163 static struct vm_domain_policy vm_default_policy =
 164     VM_DOMAIN_POLICY_STATIC_INITIALISER(VM_POLICY_FIRST_TOUCH_ROUND_ROBIN, 0);
 165 #else
 166 /* Use round-robin so the domain policy code will only try once per allocation */
 167 static struct vm_domain_policy vm_default_policy =
 168     VM_DOMAIN_POLICY_STATIC_INITIALISER(VM_POLICY_ROUND_ROBIN, 0);
 169 #endif
 170
 171 static vm_page_t vm_phys_alloc_domain_pages(int domain, int flind, int pool,
 172     int order);
 173 static vm_page_t vm_phys_alloc_seg_contig(struct vm_phys_seg *seg,
 174     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
 175     vm_paddr_t boundary);
 176 static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain);
 177 static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end);
 178 static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl,
 179     int order);
 180
 181 static int
 182 sysctl_vm_default_policy(SYSCTL_HANDLER_ARGS)
 183 {
 184         char policy_name[32];
 185         int error;
 186
 187         mtx_lock(&vm_default_policy_mtx);
 188
 189         /* Map policy to output string */
 190         switch (vm_default_policy.p.policy) {
 191         case VM_POLICY_FIRST_TOUCH:
 192                 strcpy(policy_name, "first-touch");
 193                 break;
 194         case VM_POLICY_FIRST_TOUCH_ROUND_ROBIN:
 195                 strcpy(policy_name, "first-touch-rr");
 196                 break;
 197         case VM_POLICY_ROUND_ROBIN:
 198         default:
 199                 strcpy(policy_name, "rr");
 200                 break;
 201         }
 202         mtx_unlock(&vm_default_policy_mtx);
 203
 204         error = sysctl_handle_string(oidp, &policy_name[0],
 205             sizeof(policy_name), req);
 206         if (error != 0 || req->newptr == NULL)
 207                 return (error);
 208
 209         mtx_lock(&vm_default_policy_mtx);
 210         /* Set: match on the subset of policies that make sense as a default */
 211         if (strcmp("first-touch-rr", policy_name) == 0) {
 212                 vm_domain_policy_set(&vm_default_policy,
 213                     VM_POLICY_FIRST_TOUCH_ROUND_ROBIN, 0);
 214         } else if (strcmp("first-touch", policy_name) == 0) {
 215                 vm_domain_policy_set(&vm_default_policy,
 216                     VM_POLICY_FIRST_TOUCH, 0);
 217         } else if (strcmp("rr", policy_name) == 0) {
 218                 vm_domain_policy_set(&vm_default_policy,
 219                     VM_POLICY_ROUND_ROBIN, 0);
 220         } else {
 221                 error = EINVAL;
 222                 goto finish;
 223         }
 224
 225         error = 0;
 226 finish:
 227         mtx_unlock(&vm_default_policy_mtx);
 228         return (error);
 229 }
 230
 231 SYSCTL_PROC(_vm, OID_AUTO, default_policy, CTLTYPE_STRING | CTLFLAG_RW,
 232     0, 0, sysctl_vm_default_policy, "A",
 233     "Default policy (rr, first-touch, first-touch-rr");
 234
 235 /*
 236  * Red-black tree helpers for vm fictitious range management.
 237  */
 238 static inline int
 239 vm_phys_fictitious_in_range(struct vm_phys_fictitious_seg *p,
 240     struct vm_phys_fictitious_seg *range)
 241 {
 242
 243         KASSERT(range->start != 0 && range->end != 0,
 244             ("Invalid range passed on search for vm_fictitious page"));
 245         if (p->start >= range->end)
 246                 return (1);
 247         if (p->start < range->start)
 248                 return (-1);
 249
 250         return (0);
 251 }
 252
 253 static int
 254 vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *p1,
 255     struct vm_phys_fictitious_seg *p2)
 256 {
 257
 258         /* Check if this is a search for a page */
 259         if (p1->end == 0)
 260                 return (vm_phys_fictitious_in_range(p1, p2));
 261
 262         KASSERT(p2->end != 0,
 263     ("Invalid range passed as second parameter to vm fictitious comparison"));
 264
 265         /* Searching to add a new range */
 266         if (p1->end <= p2->start)
 267                 return (-1);
 268         if (p1->start >= p2->end)
 269                 return (1);
 270
 271         panic("Trying to add overlapping vm fictitious ranges:\n"
 272             "[%#jx:%#jx] and [%#jx:%#jx]", (uintmax_t)p1->start,
 273             (uintmax_t)p1->end, (uintmax_t)p2->start, (uintmax_t)p2->end);
 274 }
 275
 276 static __inline int
 277 vm_rr_selectdomain(void)
 278 {
 279 #ifdef VM_NUMA_ALLOC
 280         struct thread *td;
 281
 282         td = curthread;
 283
 284         td->td_dom_rr_idx++;
 285         td->td_dom_rr_idx %= vm_ndomains;
 286         return (td->td_dom_rr_idx);
 287 #else
 288         return (0);
 289 #endif
 290 }
 291
 292 /*
 293  * Initialise a VM domain iterator.
 294  *
 295  * Check the thread policy, then the proc policy,
 296  * then default to the system policy.
 297  *
 298  * Later on the various layers will have this logic
 299  * plumbed into them and the phys code will be explicitly
 300  * handed a VM domain policy to use.
 301  */
 302 static void
 303 vm_policy_iterator_init(struct vm_domain_iterator *vi)
 304 {
 305 #ifdef VM_NUMA_ALLOC
 306         struct vm_domain_policy lcl;
 307 #endif
 308
 309         vm_domain_iterator_init(vi);
 310
 311 #ifdef VM_NUMA_ALLOC
 312         /* Copy out the thread policy */
 313         vm_domain_policy_localcopy(&lcl, &curthread->td_vm_dom_policy);
 314         if (lcl.p.policy != VM_POLICY_NONE) {
 315                 /* Thread policy is present; use it */
 316                 vm_domain_iterator_set_policy(vi, &lcl);
 317                 return;
 318         }
 319
 320         vm_domain_policy_localcopy(&lcl,
 321             &curthread->td_proc->p_vm_dom_policy);
 322         if (lcl.p.policy != VM_POLICY_NONE) {
 323                 /* Process policy is present; use it */
 324                 vm_domain_iterator_set_policy(vi, &lcl);
 325                 return;
 326         }
 327 #endif
 328         /* Use system default policy */
 329         vm_domain_iterator_set_policy(vi, &vm_default_policy);
 330 }
 331
 332 static void
 333 vm_policy_iterator_finish(struct vm_domain_iterator *vi)
 334 {
 335
 336         vm_domain_iterator_cleanup(vi);
 337 }
 338
 339 boolean_t
 340 vm_phys_domain_intersects(long mask, vm_paddr_t low, vm_paddr_t high)
 341 {
 342         struct vm_phys_seg *s;
 343         int idx;
 344
 345         while ((idx = ffsl(mask)) != 0) {
 346                 idx--;  /* ffsl counts from 1 */
 347                 mask &= ~(1UL << idx);
 348                 s = &vm_phys_segs[idx];
 349                 if (low < s->end && high > s->start)
 350                         return (TRUE);
 351         }
 352         return (FALSE);
 353 }
 354
 355 /*
 356  * Outputs the state of the physical memory allocator, specifically,
 357  * the amount of physical memory in each free list.
 358  */
 359 static int
 360 sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS)
 361 {
 362         struct sbuf sbuf;
 363         struct vm_freelist *fl;
 364         int dom, error, flind, oind, pind;
 365
 366         error = sysctl_wire_old_buffer(req, 0);
 367         if (error != 0)
 368                 return (error);
 369         sbuf_new_for_sysctl(&sbuf, NULL, 128 * vm_ndomains, req);
 370         for (dom = 0; dom < vm_ndomains; dom++) {
 371                 sbuf_printf(&sbuf,"\nDOMAIN %d:\n", dom);
 372                 for (flind = 0; flind < vm_nfreelists; flind++) {
 373                         sbuf_printf(&sbuf, "\nFREE LIST %d:\n"
 374                             "\n  ORDER (SIZE)  |  NUMBER"
 375                             "\n              ", flind);
 376                         for (pind = 0; pind < VM_NFREEPOOL; pind++)
 377                                 sbuf_printf(&sbuf, "  |  POOL %d", pind);
 378                         sbuf_printf(&sbuf, "\n--            ");
 379                         for (pind = 0; pind < VM_NFREEPOOL; pind++)
 380                                 sbuf_printf(&sbuf, "-- --      ");
 381                         sbuf_printf(&sbuf, "--\n");
 382                         for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
 383                                 sbuf_printf(&sbuf, "  %2d (%6dK)", oind,
 384                                     1 << (PAGE_SHIFT - 10 + oind));
 385                                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 386                                 fl = vm_phys_free_queues[dom][flind][pind];
 387                                         sbuf_printf(&sbuf, "  |  %6d",
 388                                             fl[oind].lcnt);
 389                                 }
 390                                 sbuf_printf(&sbuf, "\n");
 391                         }
 392                 }
 393         }
 394         error = sbuf_finish(&sbuf);
 395         sbuf_delete(&sbuf);
 396         return (error);
 397 }
 398
 399 /*
 400  * Outputs the set of physical memory segments.
 401  */
 402 static int
 403 sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS)
 404 {
 405         struct sbuf sbuf;
 406         struct vm_phys_seg *seg;
 407         int error, segind;
 408
 409         error = sysctl_wire_old_buffer(req, 0);
 410         if (error != 0)
 411                 return (error);
 412         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 413         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 414                 sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind);
 415                 seg = &vm_phys_segs[segind];
 416                 sbuf_printf(&sbuf, "start:     %#jx\n",
 417                     (uintmax_t)seg->start);
 418                 sbuf_printf(&sbuf, "end:       %#jx\n",
 419                     (uintmax_t)seg->end);
 420                 sbuf_printf(&sbuf, "domain:    %d\n", seg->domain);
 421                 sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues);
 422         }
 423         error = sbuf_finish(&sbuf);
 424         sbuf_delete(&sbuf);
 425         return (error);
 426 }
 427
 428 /*
 429  * Return affinity, or -1 if there's no affinity information.
 430  */
 431 int
 432 vm_phys_mem_affinity(int f, int t)
 433 {
 434
 435 #ifdef VM_NUMA_ALLOC
 436         if (mem_locality == NULL)
 437                 return (-1);
 438         if (f >= vm_ndomains || t >= vm_ndomains)
 439                 return (-1);
 440         return (mem_locality[f * vm_ndomains + t]);
 441 #else
 442         return (-1);
 443 #endif
 444 }
 445
 446 #ifdef VM_NUMA_ALLOC
 447 /*
 448  * Outputs the VM locality table.
 449  */
 450 static int
 451 sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS)
 452 {
 453         struct sbuf sbuf;
 454         int error, i, j;
 455
 456         error = sysctl_wire_old_buffer(req, 0);
 457         if (error != 0)
 458                 return (error);
 459         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 460
 461         sbuf_printf(&sbuf, "\n");
 462
 463         for (i = 0; i < vm_ndomains; i++) {
 464                 sbuf_printf(&sbuf, "%d: ", i);
 465                 for (j = 0; j < vm_ndomains; j++) {
 466                         sbuf_printf(&sbuf, "%d ", vm_phys_mem_affinity(i, j));
 467                 }
 468                 sbuf_printf(&sbuf, "\n");
 469         }
 470         error = sbuf_finish(&sbuf);
 471         sbuf_delete(&sbuf);
 472         return (error);
 473 }
 474 #endif
 475
 476 static void
 477 vm_freelist_add(struct vm_freelist *fl, vm_page_t m, int order, int tail)
 478 {
 479
 480         m->order = order;
 481         if (tail)
 482                 TAILQ_INSERT_TAIL(&fl[order].pl, m, plinks.q);
 483         else
 484                 TAILQ_INSERT_HEAD(&fl[order].pl, m, plinks.q);
 485         fl[order].lcnt++;
 486 }
 487
 488 static void
 489 vm_freelist_rem(struct vm_freelist *fl, vm_page_t m, int order)
 490 {
 491
 492         TAILQ_REMOVE(&fl[order].pl, m, plinks.q);
 493         fl[order].lcnt--;
 494         m->order = VM_NFREEORDER;
 495 }
 496
 497 /*
 498  * Create a physical memory segment.
 499  */
 500 static void
 501 _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain)
 502 {
 503         struct vm_phys_seg *seg;
 504
 505         KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX,
 506             ("vm_phys_create_seg: increase VM_PHYSSEG_MAX"));
 507         KASSERT(domain < vm_ndomains,
 508             ("vm_phys_create_seg: invalid domain provided"));
 509         seg = &vm_phys_segs[vm_phys_nsegs++];
 510         while (seg > vm_phys_segs && (seg - 1)->start >= end) {
 511                 *seg = *(seg - 1);
 512                 seg--;
 513         }
 514         seg->start = start;
 515         seg->end = end;
 516         seg->domain = domain;
 517 }
 518
 519 static void
 520 vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end)
 521 {
 522 #ifdef VM_NUMA_ALLOC
 523         int i;
 524
 525         if (mem_affinity == NULL) {
 526                 _vm_phys_create_seg(start, end, 0);
 527                 return;
 528         }
 529
 530         for (i = 0;; i++) {
 531                 if (mem_affinity[i].end == 0)
 532                         panic("Reached end of affinity info");
 533                 if (mem_affinity[i].end <= start)
 534                         continue;
 535                 if (mem_affinity[i].start > start)
 536                         panic("No affinity info for start %jx",
 537                             (uintmax_t)start);
 538                 if (mem_affinity[i].end >= end) {
 539                         _vm_phys_create_seg(start, end,
 540                             mem_affinity[i].domain);
 541                         break;
 542                 }
 543                 _vm_phys_create_seg(start, mem_affinity[i].end,
 544                     mem_affinity[i].domain);
 545                 start = mem_affinity[i].end;
 546         }
 547 #else
 548         _vm_phys_create_seg(start, end, 0);
 549 #endif
 550 }
 551
 552 /*
 553  * Add a physical memory segment.
 554  */
 555 void
 556 vm_phys_add_seg(vm_paddr_t start, vm_paddr_t end)
 557 {
 558         vm_paddr_t paddr;
 559
 560         KASSERT((start & PAGE_MASK) == 0,
 561             ("vm_phys_define_seg: start is not page aligned"));
 562         KASSERT((end & PAGE_MASK) == 0,
 563             ("vm_phys_define_seg: end is not page aligned"));
 564
 565         /*
 566          * Split the physical memory segment if it spans two or more free
 567          * list boundaries.
 568          */
 569         paddr = start;
 570 #ifdef  VM_FREELIST_ISADMA
 571         if (paddr < VM_ISADMA_BOUNDARY && end > VM_ISADMA_BOUNDARY) {
 572                 vm_phys_create_seg(paddr, VM_ISADMA_BOUNDARY);
 573                 paddr = VM_ISADMA_BOUNDARY;
 574         }
 575 #endif
 576 #ifdef  VM_FREELIST_LOWMEM
 577         if (paddr < VM_LOWMEM_BOUNDARY && end > VM_LOWMEM_BOUNDARY) {
 578                 vm_phys_create_seg(paddr, VM_LOWMEM_BOUNDARY);
 579                 paddr = VM_LOWMEM_BOUNDARY;
 580         }
 581 #endif
 582 #ifdef  VM_FREELIST_DMA32
 583         if (paddr < VM_DMA32_BOUNDARY && end > VM_DMA32_BOUNDARY) {
 584                 vm_phys_create_seg(paddr, VM_DMA32_BOUNDARY);
 585                 paddr = VM_DMA32_BOUNDARY;
 586         }
 587 #endif
 588         vm_phys_create_seg(paddr, end);
 589 }
 590
 591 /*
 592  * Initialize the physical memory allocator.
 593  *
 594  * Requires that vm_page_array is initialized!
 595  */
 596 void
 597 vm_phys_init(void)
 598 {
 599         struct vm_freelist *fl;
 600         struct vm_phys_seg *seg;
 601         u_long npages;
 602         int dom, flind, freelist, oind, pind, segind;
 603
 604         /*
 605          * Compute the number of free lists, and generate the mapping from the
 606          * manifest constants VM_FREELIST_* to the free list indices.
 607          *
 608          * Initially, the entries of vm_freelist_to_flind[] are set to either
 609          * 0 or 1 to indicate which free lists should be created.
 610          */
 611         npages = 0;
 612         for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
 613                 seg = &vm_phys_segs[segind];
 614 #ifdef  VM_FREELIST_ISADMA
 615                 if (seg->end <= VM_ISADMA_BOUNDARY)
 616                         vm_freelist_to_flind[VM_FREELIST_ISADMA] = 1;
 617                 else
 618 #endif
 619 #ifdef  VM_FREELIST_LOWMEM
 620                 if (seg->end <= VM_LOWMEM_BOUNDARY)
 621                         vm_freelist_to_flind[VM_FREELIST_LOWMEM] = 1;
 622                 else
 623 #endif
 624 #ifdef  VM_FREELIST_DMA32
 625                 if (
 626 #ifdef  VM_DMA32_NPAGES_THRESHOLD
 627                     /*
 628                      * Create the DMA32 free list only if the amount of
 629                      * physical memory above physical address 4G exceeds the
 630                      * given threshold.
 631                      */
 632                     npages > VM_DMA32_NPAGES_THRESHOLD &&
 633 #endif
 634                     seg->end <= VM_DMA32_BOUNDARY)
 635                         vm_freelist_to_flind[VM_FREELIST_DMA32] = 1;
 636                 else
 637 #endif
 638                 {
 639                         npages += atop(seg->end - seg->start);
 640                         vm_freelist_to_flind[VM_FREELIST_DEFAULT] = 1;
 641                 }
 642         }
 643         /* Change each entry into a running total of the free lists. */
 644         for (freelist = 1; freelist < VM_NFREELIST; freelist++) {
 645                 vm_freelist_to_flind[freelist] +=
 646                     vm_freelist_to_flind[freelist - 1];
 647         }
 648         vm_nfreelists = vm_freelist_to_flind[VM_NFREELIST - 1];
 649         KASSERT(vm_nfreelists > 0, ("vm_phys_init: no free lists"));
 650         /* Change each entry into a free list index. */
 651         for (freelist = 0; freelist < VM_NFREELIST; freelist++)
 652                 vm_freelist_to_flind[freelist]--;
 653
 654         /*
 655          * Initialize the first_page and free_queues fields of each physical
 656          * memory segment.
 657          */
 658 #ifdef VM_PHYSSEG_SPARSE
 659         npages = 0;
 660 #endif
 661         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 662                 seg = &vm_phys_segs[segind];
 663 #ifdef VM_PHYSSEG_SPARSE
 664                 seg->first_page = &vm_page_array[npages];
 665                 npages += atop(seg->end - seg->start);
 666 #else
 667                 seg->first_page = PHYS_TO_VM_PAGE(seg->start);
 668 #endif
 669 #ifdef  VM_FREELIST_ISADMA
 670                 if (seg->end <= VM_ISADMA_BOUNDARY) {
 671                         flind = vm_freelist_to_flind[VM_FREELIST_ISADMA];
 672                         KASSERT(flind >= 0,
 673                             ("vm_phys_init: ISADMA flind < 0"));
 674                 } else
 675 #endif
 676 #ifdef  VM_FREELIST_LOWMEM
 677                 if (seg->end <= VM_LOWMEM_BOUNDARY) {
 678                         flind = vm_freelist_to_flind[VM_FREELIST_LOWMEM];
 679                         KASSERT(flind >= 0,
 680                             ("vm_phys_init: LOWMEM flind < 0"));
 681                 } else
 682 #endif
 683 #ifdef  VM_FREELIST_DMA32
 684                 if (seg->end <= VM_DMA32_BOUNDARY) {
 685                         flind = vm_freelist_to_flind[VM_FREELIST_DMA32];
 686                         KASSERT(flind >= 0,
 687                             ("vm_phys_init: DMA32 flind < 0"));
 688                 } else
 689 #endif
 690                 {
 691                         flind = vm_freelist_to_flind[VM_FREELIST_DEFAULT];
 692                         KASSERT(flind >= 0,
 693                             ("vm_phys_init: DEFAULT flind < 0"));
 694                 }
 695                 seg->free_queues = &vm_phys_free_queues[seg->domain][flind];
 696         }
 697
 698         /*
 699          * Initialize the free queues.
 700          */
 701         for (dom = 0; dom < vm_ndomains; dom++) {
 702                 for (flind = 0; flind < vm_nfreelists; flind++) {
 703                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 704                                 fl = vm_phys_free_queues[dom][flind][pind];
 705                                 for (oind = 0; oind < VM_NFREEORDER; oind++)
 706                                         TAILQ_INIT(&fl[oind].pl);
 707                         }
 708                 }
 709         }
 710
 711         rw_init(&vm_phys_fictitious_reg_lock, "vmfctr");
 712 }
 713
 714 /*
 715  * Split a contiguous, power of two-sized set of physical pages.
 716  */
 717 static __inline void
 718 vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order)
 719 {
 720         vm_page_t m_buddy;
 721
 722         while (oind > order) {
 723                 oind--;
 724                 m_buddy = &m[1 << oind];
 725                 KASSERT(m_buddy->order == VM_NFREEORDER,
 726                     ("vm_phys_split_pages: page %p has unexpected order %d",
 727                     m_buddy, m_buddy->order));
 728                 vm_freelist_add(fl, m_buddy, oind, 0);
 729         }
 730 }
 731
 732 /*
 733  * Allocate a contiguous, power of two-sized set of physical pages
 734  * from the free lists.
 735  *
 736  * The free page queues must be locked.
 737  */
 738 vm_page_t
 739 vm_phys_alloc_pages(int pool, int order)
 740 {
 741         vm_page_t m;
 742         int domain, flind;
 743         struct vm_domain_iterator vi;
 744
 745         KASSERT(pool < VM_NFREEPOOL,
 746             ("vm_phys_alloc_pages: pool %d is out of range", pool));
 747         KASSERT(order < VM_NFREEORDER,
 748             ("vm_phys_alloc_pages: order %d is out of range", order));
 749
 750         vm_policy_iterator_init(&vi);
 751
 752         while ((vm_domain_iterator_run(&vi, &domain)) == 0) {
 753                 for (flind = 0; flind < vm_nfreelists; flind++) {
 754                         m = vm_phys_alloc_domain_pages(domain, flind, pool,
 755                             order);
 756                         if (m != NULL)
 757                                 return (m);
 758                 }
 759         }
 760
 761         vm_policy_iterator_finish(&vi);
 762         return (NULL);
 763 }
 764
 765 /*
 766  * Allocate a contiguous, power of two-sized set of physical pages from the
 767  * specified free list.  The free list must be specified using one of the
 768  * manifest constants VM_FREELIST_*.
 769  *
 770  * The free page queues must be locked.
 771  */
 772 vm_page_t
 773 vm_phys_alloc_freelist_pages(int freelist, int pool, int order)
 774 {
 775         vm_page_t m;
 776         struct vm_domain_iterator vi;
 777         int domain;
 778
 779         KASSERT(freelist < VM_NFREELIST,
 780             ("vm_phys_alloc_freelist_pages: freelist %d is out of range",
 781             freelist));
 782         KASSERT(pool < VM_NFREEPOOL,
 783             ("vm_phys_alloc_freelist_pages: pool %d is out of range", pool));
 784         KASSERT(order < VM_NFREEORDER,
 785             ("vm_phys_alloc_freelist_pages: order %d is out of range", order));
 786
 787         vm_policy_iterator_init(&vi);
 788
 789         while ((vm_domain_iterator_run(&vi, &domain)) == 0) {
 790                 m = vm_phys_alloc_domain_pages(domain,
 791                     vm_freelist_to_flind[freelist], pool, order);
 792                 if (m != NULL)
 793                         return (m);
 794         }
 795
 796         vm_policy_iterator_finish(&vi);
 797         return (NULL);
 798 }
 799
 800 static vm_page_t
 801 vm_phys_alloc_domain_pages(int domain, int flind, int pool, int order)
 802 {
 803         struct vm_freelist *fl;
 804         struct vm_freelist *alt;
 805         int oind, pind;
 806         vm_page_t m;
 807
 808         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 809         fl = &vm_phys_free_queues[domain][flind][pool][0];
 810         for (oind = order; oind < VM_NFREEORDER; oind++) {
 811                 m = TAILQ_FIRST(&fl[oind].pl);
 812                 if (m != NULL) {
 813                         vm_freelist_rem(fl, m, oind);
 814                         vm_phys_split_pages(m, oind, fl, order);
 815                         return (m);
 816                 }
 817         }
 818
 819         /*
 820          * The given pool was empty.  Find the largest
 821          * contiguous, power-of-two-sized set of pages in any
 822          * pool.  Transfer these pages to the given pool, and
 823          * use them to satisfy the allocation.
 824          */
 825         for (oind = VM_NFREEORDER - 1; oind >= order; oind--) {
 826                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 827                         alt = &vm_phys_free_queues[domain][flind][pind][0];
 828                         m = TAILQ_FIRST(&alt[oind].pl);
 829                         if (m != NULL) {
 830                                 vm_freelist_rem(alt, m, oind);
 831                                 vm_phys_set_pool(pool, m, oind);
 832                                 vm_phys_split_pages(m, oind, fl, order);
 833                                 return (m);
 834                         }
 835                 }
 836         }
 837         return (NULL);
 838 }
 839
 840 /*
 841  * Find the vm_page corresponding to the given physical address.
 842  */
 843 vm_page_t
 844 vm_phys_paddr_to_vm_page(vm_paddr_t pa)
 845 {
 846         struct vm_phys_seg *seg;
 847         int segind;
 848
 849         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 850                 seg = &vm_phys_segs[segind];
 851                 if (pa >= seg->start && pa < seg->end)
 852                         return (&seg->first_page[atop(pa - seg->start)]);
 853         }
 854         return (NULL);
 855 }
 856
 857 vm_page_t
 858 vm_phys_fictitious_to_vm_page(vm_paddr_t pa)
 859 {
 860         struct vm_phys_fictitious_seg tmp, *seg;
 861         vm_page_t m;
 862
 863         m = NULL;
 864         tmp.start = pa;
 865         tmp.end = 0;
 866
 867         rw_rlock(&vm_phys_fictitious_reg_lock);
 868         seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
 869         rw_runlock(&vm_phys_fictitious_reg_lock);
 870         if (seg == NULL)
 871                 return (NULL);
 872
 873         m = &seg->first_page[atop(pa - seg->start)];
 874         KASSERT((m->flags & PG_FICTITIOUS) != 0, ("%p not fictitious", m));
 875
 876         return (m);
 877 }
 878
 879 static inline void
 880 vm_phys_fictitious_init_range(vm_page_t range, vm_paddr_t start,
 881     long page_count, vm_memattr_t memattr)
 882 {
 883         long i;
 884
 885         bzero(range, page_count * sizeof(*range));
 886         for (i = 0; i < page_count; i++) {
 887                 vm_page_initfake(&range[i], start + PAGE_SIZE * i, memattr);
 888                 range[i].oflags &= ~VPO_UNMANAGED;
 889                 range[i].busy_lock = VPB_UNBUSIED;
 890         }
 891 }
 892
 893 int
 894 vm_phys_fictitious_reg_range(vm_paddr_t start, vm_paddr_t end,
 895     vm_memattr_t memattr)
 896 {
 897         struct vm_phys_fictitious_seg *seg;
 898         vm_page_t fp;
 899         long page_count;
 900 #ifdef VM_PHYSSEG_DENSE
 901         long pi, pe;
 902         long dpage_count;
 903 #endif
 904
 905         KASSERT(start < end,
 906             ("Start of segment isn't less than end (start: %jx end: %jx)",
 907             (uintmax_t)start, (uintmax_t)end));
 908
 909         page_count = (end - start) / PAGE_SIZE;
 910
 911 #ifdef VM_PHYSSEG_DENSE
 912         pi = atop(start);
 913         pe = atop(end);
 914         if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
 915                 fp = &vm_page_array[pi - first_page];
 916                 if ((pe - first_page) > vm_page_array_size) {
 917                         /*
 918                          * We have a segment that starts inside
 919                          * of vm_page_array, but ends outside of it.
 920                          *
 921                          * Use vm_page_array pages for those that are
 922                          * inside of the vm_page_array range, and
 923                          * allocate the remaining ones.
 924                          */
 925                         dpage_count = vm_page_array_size - (pi - first_page);
 926                         vm_phys_fictitious_init_range(fp, start, dpage_count,
 927                             memattr);
 928                         page_count -= dpage_count;
 929                         start += ptoa(dpage_count);
 930                         goto alloc;
 931                 }
 932                 /*
 933                  * We can allocate the full range from vm_page_array,
 934                  * so there's no need to register the range in the tree.
 935                  */
 936                 vm_phys_fictitious_init_range(fp, start, page_count, memattr);
 937                 return (0);
 938         } else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
 939                 /*
 940                  * We have a segment that ends inside of vm_page_array,
 941                  * but starts outside of it.
 942                  */
 943                 fp = &vm_page_array[0];
 944                 dpage_count = pe - first_page;
 945                 vm_phys_fictitious_init_range(fp, ptoa(first_page), dpage_count,
 946                     memattr);
 947                 end -= ptoa(dpage_count);
 948                 page_count -= dpage_count;
 949                 goto alloc;
 950         } else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
 951                 /*
 952                  * Trying to register a fictitious range that expands before
 953                  * and after vm_page_array.
 954                  */
 955                 return (EINVAL);
 956         } else {
 957 alloc:
 958 #endif
 959                 fp = malloc(page_count * sizeof(struct vm_page), M_FICT_PAGES,
 960                     M_WAITOK);
 961 #ifdef VM_PHYSSEG_DENSE
 962         }
 963 #endif
 964         vm_phys_fictitious_init_range(fp, start, page_count, memattr);
 965
 966         seg = malloc(sizeof(*seg), M_FICT_PAGES, M_WAITOK | M_ZERO);
 967         seg->start = start;
 968         seg->end = end;
 969         seg->first_page = fp;
 970
 971         rw_wlock(&vm_phys_fictitious_reg_lock);
 972         RB_INSERT(fict_tree, &vm_phys_fictitious_tree, seg);
 973         rw_wunlock(&vm_phys_fictitious_reg_lock);
 974
 975         return (0);
 976 }
 977
 978 void
 979 vm_phys_fictitious_unreg_range(vm_paddr_t start, vm_paddr_t end)
 980 {
 981         struct vm_phys_fictitious_seg *seg, tmp;
 982 #ifdef VM_PHYSSEG_DENSE
 983         long pi, pe;
 984 #endif
 985
 986         KASSERT(start < end,
 987             ("Start of segment isn't less than end (start: %jx end: %jx)",
 988             (uintmax_t)start, (uintmax_t)end));
 989
 990 #ifdef VM_PHYSSEG_DENSE
 991         pi = atop(start);
 992         pe = atop(end);
 993         if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
 994                 if ((pe - first_page) <= vm_page_array_size) {
 995                         /*
 996                          * This segment was allocated using vm_page_array
 997                          * only, there's nothing to do since those pages
 998                          * were never added to the tree.
 999                          */
1000                         return;
1001                 }
1002                 /*
1003                  * We have a segment that starts inside
1004                  * of vm_page_array, but ends outside of it.
1005                  *
1006                  * Calculate how many pages were added to the
1007                  * tree and free them.
1008                  */
1009                 start = ptoa(first_page + vm_page_array_size);
1010         } else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
1011                 /*
1012                  * We have a segment that ends inside of vm_page_array,
1013                  * but starts outside of it.
1014                  */
1015                 end = ptoa(first_page);
1016         } else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
1017                 /* Since it's not possible to register such a range, panic. */
1018                 panic(
1019                     "Unregistering not registered fictitious range [%#jx:%#jx]",
1020                     (uintmax_t)start, (uintmax_t)end);
1021         }
1022 #endif
1023         tmp.start = start;
1024         tmp.end = 0;
1025
1026         rw_wlock(&vm_phys_fictitious_reg_lock);
1027         seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
1028         if (seg->start != start || seg->end != end) {
1029                 rw_wunlock(&vm_phys_fictitious_reg_lock);
1030                 panic(
1031                     "Unregistering not registered fictitious range [%#jx:%#jx]",
1032                     (uintmax_t)start, (uintmax_t)end);
1033         }
1034         RB_REMOVE(fict_tree, &vm_phys_fictitious_tree, seg);
1035         rw_wunlock(&vm_phys_fictitious_reg_lock);
1036         free(seg->first_page, M_FICT_PAGES);
1037         free(seg, M_FICT_PAGES);
1038 }
1039
1040 /*
1041  * Free a contiguous, power of two-sized set of physical pages.
1042  *
1043  * The free page queues must be locked.
1044  */
1045 void
1046 vm_phys_free_pages(vm_page_t m, int order)
1047 {
1048         struct vm_freelist *fl;
1049         struct vm_phys_seg *seg;
1050         vm_paddr_t pa;
1051         vm_page_t m_buddy;
1052
1053         KASSERT(m->order == VM_NFREEORDER,
1054             ("vm_phys_free_pages: page %p has unexpected order %d",
1055             m, m->order));
1056         KASSERT(m->pool < VM_NFREEPOOL,
1057             ("vm_phys_free_pages: page %p has unexpected pool %d",
1058             m, m->pool));
1059         KASSERT(order < VM_NFREEORDER,
1060             ("vm_phys_free_pages: order %d is out of range", order));
1061         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1062         seg = &vm_phys_segs[m->segind];
1063         if (order < VM_NFREEORDER - 1) {
1064                 pa = VM_PAGE_TO_PHYS(m);
1065                 do {
1066                         pa ^= ((vm_paddr_t)1 << (PAGE_SHIFT + order));
1067                         if (pa < seg->start || pa >= seg->end)
1068                                 break;
1069                         m_buddy = &seg->first_page[atop(pa - seg->start)];
1070                         if (m_buddy->order != order)
1071                                 break;
1072                         fl = (*seg->free_queues)[m_buddy->pool];
1073                         vm_freelist_rem(fl, m_buddy, order);
1074                         if (m_buddy->pool != m->pool)
1075                                 vm_phys_set_pool(m->pool, m_buddy, order);
1076                         order++;
1077                         pa &= ~(((vm_paddr_t)1 << (PAGE_SHIFT + order)) - 1);
1078                         m = &seg->first_page[atop(pa - seg->start)];
1079                 } while (order < VM_NFREEORDER - 1);
1080         }
1081         fl = (*seg->free_queues)[m->pool];
1082         vm_freelist_add(fl, m, order, 1);
1083 }
1084
1085 /*
1086  * Free a contiguous, arbitrarily sized set of physical pages.
1087  *
1088  * The free page queues must be locked.
1089  */
1090 void
1091 vm_phys_free_contig(vm_page_t m, u_long npages)
1092 {
1093         u_int n;
1094         int order;
1095
1096         /*
1097          * Avoid unnecessary coalescing by freeing the pages in the largest
1098          * possible power-of-two-sized subsets.
1099          */
1100         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1101         for (;; npages -= n) {
1102                 /*
1103                  * Unsigned "min" is used here so that "order" is assigned
1104                  * "VM_NFREEORDER - 1" when "m"'s physical address is zero
1105                  * or the low-order bits of its physical address are zero
1106                  * because the size of a physical address exceeds the size of
1107                  * a long.
1108                  */
1109                 order = min(ffsl(VM_PAGE_TO_PHYS(m) >> PAGE_SHIFT) - 1,
1110                     VM_NFREEORDER - 1);
1111                 n = 1 << order;
1112                 if (npages < n)
1113                         break;
1114                 vm_phys_free_pages(m, order);
1115                 m += n;
1116         }
1117         /* The residual "npages" is less than "1 << (VM_NFREEORDER - 1)". */
1118         for (; npages > 0; npages -= n) {
1119                 order = flsl(npages) - 1;
1120                 n = 1 << order;
1121                 vm_phys_free_pages(m, order);
1122                 m += n;
1123         }
1124 }
1125
1126 /*
1127  * Scan physical memory between the specified addresses "low" and "high" for a
1128  * run of contiguous physical pages that satisfy the specified conditions, and
1129  * return the lowest page in the run.  The specified "alignment" determines
1130  * the alignment of the lowest physical page in the run.  If the specified
1131  * "boundary" is non-zero, then the run of physical pages cannot span a
1132  * physical address that is a multiple of "boundary".
1133  *
1134  * "npages" must be greater than zero.  Both "alignment" and "boundary" must
1135  * be a power of two.
1136  */
1137 vm_page_t
1138 vm_phys_scan_contig(u_long npages, vm_paddr_t low, vm_paddr_t high,
1139     u_long alignment, vm_paddr_t boundary, int options)
1140 {
1141         vm_paddr_t pa_end;
1142         vm_page_t m_end, m_run, m_start;
1143         struct vm_phys_seg *seg;
1144         int segind;
1145
1146         KASSERT(npages > 0, ("npages is 0"));
1147         KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1148         KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1149         if (low >= high)
1150                 return (NULL);
1151         for (segind = 0; segind < vm_phys_nsegs; segind++) {
1152                 seg = &vm_phys_segs[segind];
1153                 if (seg->start >= high)
1154                         break;
1155                 if (low >= seg->end)
1156                         continue;
1157                 if (low <= seg->start)
1158                         m_start = seg->first_page;
1159                 else
1160                         m_start = &seg->first_page[atop(low - seg->start)];
1161                 if (high < seg->end)
1162                         pa_end = high;
1163                 else
1164                         pa_end = seg->end;
1165                 if (pa_end - VM_PAGE_TO_PHYS(m_start) < ptoa(npages))
1166                         continue;
1167                 m_end = &seg->first_page[atop(pa_end - seg->start)];
1168                 m_run = vm_page_scan_contig(npages, m_start, m_end,
1169                     alignment, boundary, options);
1170                 if (m_run != NULL)
1171                         return (m_run);
1172         }
1173         return (NULL);
1174 }
1175
1176 /*
1177  * Set the pool for a contiguous, power of two-sized set of physical pages.
1178  */
1179 void
1180 vm_phys_set_pool(int pool, vm_page_t m, int order)
1181 {
1182         vm_page_t m_tmp;
1183
1184         for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++)
1185                 m_tmp->pool = pool;
1186 }
1187
1188 /*
1189  * Search for the given physical page "m" in the free lists.  If the search
1190  * succeeds, remove "m" from the free lists and return TRUE.  Otherwise, return
1191  * FALSE, indicating that "m" is not in the free lists.
1192  *
1193  * The free page queues must be locked.
1194  */
1195 boolean_t
1196 vm_phys_unfree_page(vm_page_t m)
1197 {
1198         struct vm_freelist *fl;
1199         struct vm_phys_seg *seg;
1200         vm_paddr_t pa, pa_half;
1201         vm_page_t m_set, m_tmp;
1202         int order;
1203
1204         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1205
1206         /*
1207          * First, find the contiguous, power of two-sized set of free
1208          * physical pages containing the given physical page "m" and
1209          * assign it to "m_set".
1210          */
1211         seg = &vm_phys_segs[m->segind];
1212         for (m_set = m, order = 0; m_set->order == VM_NFREEORDER &&
1213             order < VM_NFREEORDER - 1; ) {
1214                 order++;
1215                 pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order));
1216                 if (pa >= seg->start)
1217                         m_set = &seg->first_page[atop(pa - seg->start)];
1218                 else
1219                         return (FALSE);
1220         }
1221         if (m_set->order < order)
1222                 return (FALSE);
1223         if (m_set->order == VM_NFREEORDER)
1224                 return (FALSE);
1225         KASSERT(m_set->order < VM_NFREEORDER,
1226             ("vm_phys_unfree_page: page %p has unexpected order %d",
1227             m_set, m_set->order));
1228
1229         /*
1230          * Next, remove "m_set" from the free lists.  Finally, extract
1231          * "m" from "m_set" using an iterative algorithm: While "m_set"
1232          * is larger than a page, shrink "m_set" by returning the half
1233          * of "m_set" that does not contain "m" to the free lists.
1234          */
1235         fl = (*seg->free_queues)[m_set->pool];
1236         order = m_set->order;
1237         vm_freelist_rem(fl, m_set, order);
1238         while (order > 0) {
1239                 order--;
1240                 pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order));
1241                 if (m->phys_addr < pa_half)
1242                         m_tmp = &seg->first_page[atop(pa_half - seg->start)];
1243                 else {
1244                         m_tmp = m_set;
1245                         m_set = &seg->first_page[atop(pa_half - seg->start)];
1246                 }
1247                 vm_freelist_add(fl, m_tmp, order, 0);
1248         }
1249         KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency"));
1250         return (TRUE);
1251 }
1252
1253 /*
1254  * Try to zero one physical page.  Used by an idle priority thread.
1255  */
1256 boolean_t
1257 vm_phys_zero_pages_idle(void)
1258 {
1259         static struct vm_freelist *fl;
1260         static int flind, oind, pind;
1261         vm_page_t m, m_tmp;
1262         int domain;
1263
1264         domain = vm_rr_selectdomain();
1265         fl = vm_phys_free_queues[domain][0][0];
1266         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1267         for (;;) {
1268                 TAILQ_FOREACH_REVERSE(m, &fl[oind].pl, pglist, plinks.q) {
1269                         for (m_tmp = m; m_tmp < &m[1 << oind]; m_tmp++) {
1270                                 if ((m_tmp->flags & PG_ZERO) == 0) {
1271                                         vm_phys_unfree_page(m_tmp);
1272                                         vm_phys_freecnt_adj(m, -1);
1273                                         mtx_unlock(&vm_page_queue_free_mtx);
1274                                         pmap_zero_page_idle(m_tmp);
1275                                         m_tmp->flags |= PG_ZERO;
1276                                         mtx_lock(&vm_page_queue_free_mtx);
1277                                         vm_phys_freecnt_adj(m, 1);
1278                                         vm_phys_free_pages(m_tmp, 0);
1279                                         vm_page_zero_count++;
1280                                         cnt_prezero++;
1281                                         return (TRUE);
1282                                 }
1283                         }
1284                 }
1285                 oind++;
1286                 if (oind == VM_NFREEORDER) {
1287                         oind = 0;
1288                         pind++;
1289                         if (pind == VM_NFREEPOOL) {
1290                                 pind = 0;
1291                                 flind++;
1292                                 if (flind == vm_nfreelists)
1293                                         flind = 0;
1294                         }
1295                         fl = vm_phys_free_queues[domain][flind][pind];
1296                 }
1297         }
1298 }
1299
1300 /*
1301  * Allocate a contiguous set of physical pages of the given size
1302  * "npages" from the free lists.  All of the physical pages must be at
1303  * or above the given physical address "low" and below the given
1304  * physical address "high".  The given value "alignment" determines the
1305  * alignment of the first physical page in the set.  If the given value
1306  * "boundary" is non-zero, then the set of physical pages cannot cross
1307  * any physical address boundary that is a multiple of that value.  Both
1308  * "alignment" and "boundary" must be a power of two.
1309  */
1310 vm_page_t
1311 vm_phys_alloc_contig(u_long npages, vm_paddr_t low, vm_paddr_t high,
1312     u_long alignment, vm_paddr_t boundary)
1313 {
1314         vm_paddr_t pa_end, pa_start;
1315         vm_page_t m_run;
1316         struct vm_domain_iterator vi;
1317         struct vm_phys_seg *seg;
1318         int domain, segind;
1319
1320         KASSERT(npages > 0, ("npages is 0"));
1321         KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1322         KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1323         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1324         if (low >= high)
1325                 return (NULL);
1326         vm_policy_iterator_init(&vi);
1327 restartdom:
1328         if (vm_domain_iterator_run(&vi, &domain) != 0) {
1329                 vm_policy_iterator_finish(&vi);
1330                 return (NULL);
1331         }
1332         m_run = NULL;
1333         for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
1334                 seg = &vm_phys_segs[segind];
1335                 if (seg->start >= high || seg->domain != domain)
1336                         continue;
1337                 if (low >= seg->end)
1338                         break;
1339                 if (low <= seg->start)
1340                         pa_start = seg->start;
1341                 else
1342                         pa_start = low;
1343                 if (high < seg->end)
1344                         pa_end = high;
1345                 else
1346                         pa_end = seg->end;
1347                 if (pa_end - pa_start < ptoa(npages))
1348                         continue;
1349                 m_run = vm_phys_alloc_seg_contig(seg, npages, low, high,
1350                     alignment, boundary);
1351                 if (m_run != NULL)
1352                         break;
1353         }
1354         if (m_run == NULL && !vm_domain_iterator_isdone(&vi))
1355                 goto restartdom;
1356         vm_policy_iterator_finish(&vi);
1357         return (m_run);
1358 }
1359
1360 /*
1361  * Allocate a run of contiguous physical pages from the free list for the
1362  * specified segment.
1363  */
1364 static vm_page_t
1365 vm_phys_alloc_seg_contig(struct vm_phys_seg *seg, u_long npages,
1366     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1367 {
1368         struct vm_freelist *fl;
1369         vm_paddr_t pa, pa_end, size;
1370         vm_page_t m, m_ret;
1371         u_long npages_end;
1372         int oind, order, pind;
1373
1374         KASSERT(npages > 0, ("npages is 0"));
1375         KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1376         KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1377         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1378         /* Compute the queue that is the best fit for npages. */
1379         for (order = 0; (1 << order) < npages; order++);
1380         /* Search for a run satisfying the specified conditions. */
1381         size = npages << PAGE_SHIFT;
1382         for (oind = min(order, VM_NFREEORDER - 1); oind < VM_NFREEORDER;
1383             oind++) {
1384                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1385                         fl = (*seg->free_queues)[pind];
1386                         TAILQ_FOREACH(m_ret, &fl[oind].pl, plinks.q) {
1387                                 /*
1388                                  * Is the size of this allocation request
1389                                  * larger than the largest block size?
1390                                  */
1391                                 if (order >= VM_NFREEORDER) {
1392                                         /*
1393                                          * Determine if a sufficient number of
1394                                          * subsequent blocks to satisfy the
1395                                          * allocation request are free.
1396                                          */
1397                                         pa = VM_PAGE_TO_PHYS(m_ret);
1398                                         pa_end = pa + size;
1399                                         if (pa_end < pa)
1400                                                 continue;
1401                                         for (;;) {
1402                                                 pa += 1 << (PAGE_SHIFT +
1403                                                     VM_NFREEORDER - 1);
1404                                                 if (pa >= pa_end ||
1405                                                     pa < seg->start ||
1406                                                     pa >= seg->end)
1407                                                         break;
1408                                                 m = &seg->first_page[atop(pa -
1409                                                     seg->start)];
1410                                                 if (m->order != VM_NFREEORDER -
1411                                                     1)
1412                                                         break;
1413                                         }
1414                                         /* If not, go to the next block. */
1415                                         if (pa < pa_end)
1416                                                 continue;
1417                                 }
1418
1419                                 /*
1420                                  * Determine if the blocks are within the
1421                                  * given range, satisfy the given alignment,
1422                                  * and do not cross the given boundary.
1423                                  */
1424                                 pa = VM_PAGE_TO_PHYS(m_ret);
1425                                 pa_end = pa + size;
1426                                 if (pa >= low && pa_end <= high &&
1427                                     (pa & (alignment - 1)) == 0 &&
1428                                     rounddown2(pa ^ (pa_end - 1), boundary) == 0)
1429                                         goto done;
1430                         }
1431                 }
1432         }
1433         return (NULL);
1434 done:
1435         for (m = m_ret; m < &m_ret[npages]; m = &m[1 << oind]) {
1436                 fl = (*seg->free_queues)[m->pool];
1437                 vm_freelist_rem(fl, m, m->order);
1438         }
1439         if (m_ret->pool != VM_FREEPOOL_DEFAULT)
1440                 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m_ret, oind);
1441         fl = (*seg->free_queues)[m_ret->pool];
1442         vm_phys_split_pages(m_ret, oind, fl, order);
1443         /* Return excess pages to the free lists. */
1444         npages_end = roundup2(npages, 1 << imin(oind, order));
1445         if (npages < npages_end)
1446                 vm_phys_free_contig(&m_ret[npages], npages_end - npages);
1447         return (m_ret);
1448 }
1449
1450 #ifdef DDB
1451 /*
1452  * Show the number of physical pages in each of the free lists.
1453  */
1454 DB_SHOW_COMMAND(freepages, db_show_freepages)
1455 {
1456         struct vm_freelist *fl;
1457         int flind, oind, pind, dom;
1458
1459         for (dom = 0; dom < vm_ndomains; dom++) {
1460                 db_printf("DOMAIN: %d\n", dom);
1461                 for (flind = 0; flind < vm_nfreelists; flind++) {
1462                         db_printf("FREE LIST %d:\n"
1463                             "\n  ORDER (SIZE)  |  NUMBER"
1464                             "\n              ", flind);
1465                         for (pind = 0; pind < VM_NFREEPOOL; pind++)
1466                                 db_printf("  |  POOL %d", pind);
1467                         db_printf("\n--            ");
1468                         for (pind = 0; pind < VM_NFREEPOOL; pind++)
1469                                 db_printf("-- --      ");
1470                         db_printf("--\n");
1471                         for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
1472                                 db_printf("  %2.2d (%6.6dK)", oind,
1473                                     1 << (PAGE_SHIFT - 10 + oind));
1474                                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1475                                 fl = vm_phys_free_queues[dom][flind][pind];
1476                                         db_printf("  |  %6.6d", fl[oind].lcnt);
1477                                 }
1478                                 db_printf("\n");
1479                         }
1480                         db_printf("\n");
1481                 }
1482                 db_printf("\n");
1483         }
1484 }
1485 #endif