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