sys/vm/vm_phys.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2002-2006 Rice University
   5  * Copyright (c) 2007 Alan L. Cox <alc@cs.rice.edu>
   6  * All rights reserved.
   7  *
   8  * This software was developed for the FreeBSD Project by Alan L. Cox,
   9  * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro.
  10  *
  11  * Redistribution and use in source and binary forms, with or without
  12  * modification, are permitted provided that the following conditions
  13  * are met:
  14  * 1. Redistributions of source code must retain the above copyright
  15  *    notice, this list of conditions and the following disclaimer.
  16  * 2. Redistributions in binary form must reproduce the above copyright
  17  *    notice, this list of conditions and the following disclaimer in the
  18  *    documentation and/or other materials provided with the distribution.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  23  * A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT
  24  * HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  25  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
  26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
  27  * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY
  30  * WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
  31  * POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 /*
  35  *      Physical memory system implementation
  36  *
  37  * Any external functions defined by this module are only to be used by the
  38  * virtual memory system.
  39  */
  40
  41 #include <sys/cdefs.h>
  42 __FBSDID("$FreeBSD$");
  43
  44 #include "opt_ddb.h"
  45 #include "opt_vm.h"
  46
  47 #include <sys/param.h>
  48 #include <sys/systm.h>
  49 #include <sys/domainset.h>
  50 #include <sys/lock.h>
  51 #include <sys/kernel.h>
  52 #include <sys/malloc.h>
  53 #include <sys/mutex.h>
  54 #include <sys/proc.h>
  55 #include <sys/queue.h>
  56 #include <sys/rwlock.h>
  57 #include <sys/sbuf.h>
  58 #include <sys/sysctl.h>
  59 #include <sys/tree.h>
  60 #include <sys/vmmeter.h>
  61 #include <sys/seq.h>
  62
  63 #include <ddb/ddb.h>
  64
  65 #include <vm/vm.h>
  66 #include <vm/vm_param.h>
  67 #include <vm/vm_kern.h>
  68 #include <vm/vm_object.h>
  69 #include <vm/vm_page.h>
  70 #include <vm/vm_phys.h>
  71 #include <vm/vm_pagequeue.h>
  72
  73 _Static_assert(sizeof(long) * NBBY >= VM_PHYSSEG_MAX,
  74     "Too many physsegs.");
  75
  76 #ifdef NUMA
  77 struct mem_affinity __read_mostly *mem_affinity;
  78 int __read_mostly *mem_locality;
  79 #endif
  80
  81 int __read_mostly vm_ndomains = 1;
  82 domainset_t __read_mostly all_domains = DOMAINSET_T_INITIALIZER(0x1);
  83
  84 struct vm_phys_seg __read_mostly vm_phys_segs[VM_PHYSSEG_MAX];
  85 int __read_mostly vm_phys_nsegs;
  86
  87 struct vm_phys_fictitious_seg;
  88 static int vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *,
  89     struct vm_phys_fictitious_seg *);
  90
  91 RB_HEAD(fict_tree, vm_phys_fictitious_seg) vm_phys_fictitious_tree =
  92     RB_INITIALIZER(_vm_phys_fictitious_tree);
  93
  94 struct vm_phys_fictitious_seg {
  95         RB_ENTRY(vm_phys_fictitious_seg) node;
  96         /* Memory region data */
  97         vm_paddr_t      start;
  98         vm_paddr_t      end;
  99         vm_page_t       first_page;
 100 };
 101
 102 RB_GENERATE_STATIC(fict_tree, vm_phys_fictitious_seg, node,
 103     vm_phys_fictitious_cmp);
 104
 105 static struct rwlock_padalign vm_phys_fictitious_reg_lock;
 106 MALLOC_DEFINE(M_FICT_PAGES, "vm_fictitious", "Fictitious VM pages");
 107
 108 static struct vm_freelist __aligned(CACHE_LINE_SIZE)
 109     vm_phys_free_queues[MAXMEMDOM][VM_NFREELIST][VM_NFREEPOOL]
 110     [VM_NFREEORDER_MAX];
 111
 112 static int __read_mostly vm_nfreelists;
 113
 114 /*
 115  * Provides the mapping from VM_FREELIST_* to free list indices (flind).
 116  */
 117 static int __read_mostly vm_freelist_to_flind[VM_NFREELIST];
 118
 119 CTASSERT(VM_FREELIST_DEFAULT == 0);
 120
 121 #ifdef VM_FREELIST_DMA32
 122 #define VM_DMA32_BOUNDARY       ((vm_paddr_t)1 << 32)
 123 #endif
 124
 125 /*
 126  * Enforce the assumptions made by vm_phys_add_seg() and vm_phys_init() about
 127  * the ordering of the free list boundaries.
 128  */
 129 #if defined(VM_LOWMEM_BOUNDARY) && defined(VM_DMA32_BOUNDARY)
 130 CTASSERT(VM_LOWMEM_BOUNDARY < VM_DMA32_BOUNDARY);
 131 #endif
 132
 133 static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS);
 134 SYSCTL_OID(_vm, OID_AUTO, phys_free, CTLTYPE_STRING | CTLFLAG_RD,
 135     NULL, 0, sysctl_vm_phys_free, "A", "Phys Free Info");
 136
 137 static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS);
 138 SYSCTL_OID(_vm, OID_AUTO, phys_segs, CTLTYPE_STRING | CTLFLAG_RD,
 139     NULL, 0, sysctl_vm_phys_segs, "A", "Phys Seg Info");
 140
 141 #ifdef NUMA
 142 static int sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS);
 143 SYSCTL_OID(_vm, OID_AUTO, phys_locality, CTLTYPE_STRING | CTLFLAG_RD,
 144     NULL, 0, sysctl_vm_phys_locality, "A", "Phys Locality Info");
 145 #endif
 146
 147 SYSCTL_INT(_vm, OID_AUTO, ndomains, CTLFLAG_RD,
 148     &vm_ndomains, 0, "Number of physical memory domains available.");
 149
 150 static vm_page_t vm_phys_alloc_seg_contig(struct vm_phys_seg *seg,
 151     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
 152     vm_paddr_t boundary);
 153 static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain);
 154 static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end);
 155 static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl,
 156     int order, int tail);
 157
 158 /*
 159  * Red-black tree helpers for vm fictitious range management.
 160  */
 161 static inline int
 162 vm_phys_fictitious_in_range(struct vm_phys_fictitious_seg *p,
 163     struct vm_phys_fictitious_seg *range)
 164 {
 165
 166         KASSERT(range->start != 0 && range->end != 0,
 167             ("Invalid range passed on search for vm_fictitious page"));
 168         if (p->start >= range->end)
 169                 return (1);
 170         if (p->start < range->start)
 171                 return (-1);
 172
 173         return (0);
 174 }
 175
 176 static int
 177 vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *p1,
 178     struct vm_phys_fictitious_seg *p2)
 179 {
 180
 181         /* Check if this is a search for a page */
 182         if (p1->end == 0)
 183                 return (vm_phys_fictitious_in_range(p1, p2));
 184
 185         KASSERT(p2->end != 0,
 186     ("Invalid range passed as second parameter to vm fictitious comparison"));
 187
 188         /* Searching to add a new range */
 189         if (p1->end <= p2->start)
 190                 return (-1);
 191         if (p1->start >= p2->end)
 192                 return (1);
 193
 194         panic("Trying to add overlapping vm fictitious ranges:\n"
 195             "[%#jx:%#jx] and [%#jx:%#jx]", (uintmax_t)p1->start,
 196             (uintmax_t)p1->end, (uintmax_t)p2->start, (uintmax_t)p2->end);
 197 }
 198
 199 int
 200 vm_phys_domain_match(int prefer, vm_paddr_t low, vm_paddr_t high)
 201 {
 202 #ifdef NUMA
 203         domainset_t mask;
 204         int i;
 205
 206         if (vm_ndomains == 1 || mem_affinity == NULL)
 207                 return (0);
 208
 209         DOMAINSET_ZERO(&mask);
 210         /*
 211          * Check for any memory that overlaps low, high.
 212          */
 213         for (i = 0; mem_affinity[i].end != 0; i++)
 214                 if (mem_affinity[i].start <= high &&
 215                     mem_affinity[i].end >= low)
 216                         DOMAINSET_SET(mem_affinity[i].domain, &mask);
 217         if (prefer != -1 && DOMAINSET_ISSET(prefer, &mask))
 218                 return (prefer);
 219         if (DOMAINSET_EMPTY(&mask))
 220                 panic("vm_phys_domain_match:  Impossible constraint");
 221         return (DOMAINSET_FFS(&mask) - 1);
 222 #else
 223         return (0);
 224 #endif
 225 }
 226
 227 /*
 228  * Outputs the state of the physical memory allocator, specifically,
 229  * the amount of physical memory in each free list.
 230  */
 231 static int
 232 sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS)
 233 {
 234         struct sbuf sbuf;
 235         struct vm_freelist *fl;
 236         int dom, error, flind, oind, pind;
 237
 238         error = sysctl_wire_old_buffer(req, 0);
 239         if (error != 0)
 240                 return (error);
 241         sbuf_new_for_sysctl(&sbuf, NULL, 128 * vm_ndomains, req);
 242         for (dom = 0; dom < vm_ndomains; dom++) {
 243                 sbuf_printf(&sbuf,"\nDOMAIN %d:\n", dom);
 244                 for (flind = 0; flind < vm_nfreelists; flind++) {
 245                         sbuf_printf(&sbuf, "\nFREE LIST %d:\n"
 246                             "\n  ORDER (SIZE)  |  NUMBER"
 247                             "\n              ", flind);
 248                         for (pind = 0; pind < VM_NFREEPOOL; pind++)
 249                                 sbuf_printf(&sbuf, "  |  POOL %d", pind);
 250                         sbuf_printf(&sbuf, "\n--            ");
 251                         for (pind = 0; pind < VM_NFREEPOOL; pind++)
 252                                 sbuf_printf(&sbuf, "-- --      ");
 253                         sbuf_printf(&sbuf, "--\n");
 254                         for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
 255                                 sbuf_printf(&sbuf, "  %2d (%6dK)", oind,
 256                                     1 << (PAGE_SHIFT - 10 + oind));
 257                                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 258                                 fl = vm_phys_free_queues[dom][flind][pind];
 259                                         sbuf_printf(&sbuf, "  |  %6d",
 260                                             fl[oind].lcnt);
 261                                 }
 262                                 sbuf_printf(&sbuf, "\n");
 263                         }
 264                 }
 265         }
 266         error = sbuf_finish(&sbuf);
 267         sbuf_delete(&sbuf);
 268         return (error);
 269 }
 270
 271 /*
 272  * Outputs the set of physical memory segments.
 273  */
 274 static int
 275 sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS)
 276 {
 277         struct sbuf sbuf;
 278         struct vm_phys_seg *seg;
 279         int error, segind;
 280
 281         error = sysctl_wire_old_buffer(req, 0);
 282         if (error != 0)
 283                 return (error);
 284         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 285         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 286                 sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind);
 287                 seg = &vm_phys_segs[segind];
 288                 sbuf_printf(&sbuf, "start:     %#jx\n",
 289                     (uintmax_t)seg->start);
 290                 sbuf_printf(&sbuf, "end:       %#jx\n",
 291                     (uintmax_t)seg->end);
 292                 sbuf_printf(&sbuf, "domain:    %d\n", seg->domain);
 293                 sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues);
 294         }
 295         error = sbuf_finish(&sbuf);
 296         sbuf_delete(&sbuf);
 297         return (error);
 298 }
 299
 300 /*
 301  * Return affinity, or -1 if there's no affinity information.
 302  */
 303 int
 304 vm_phys_mem_affinity(int f, int t)
 305 {
 306
 307 #ifdef NUMA
 308         if (mem_locality == NULL)
 309                 return (-1);
 310         if (f >= vm_ndomains || t >= vm_ndomains)
 311                 return (-1);
 312         return (mem_locality[f * vm_ndomains + t]);
 313 #else
 314         return (-1);
 315 #endif
 316 }
 317
 318 #ifdef NUMA
 319 /*
 320  * Outputs the VM locality table.
 321  */
 322 static int
 323 sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS)
 324 {
 325         struct sbuf sbuf;
 326         int error, i, j;
 327
 328         error = sysctl_wire_old_buffer(req, 0);
 329         if (error != 0)
 330                 return (error);
 331         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 332
 333         sbuf_printf(&sbuf, "\n");
 334
 335         for (i = 0; i < vm_ndomains; i++) {
 336                 sbuf_printf(&sbuf, "%d: ", i);
 337                 for (j = 0; j < vm_ndomains; j++) {
 338                         sbuf_printf(&sbuf, "%d ", vm_phys_mem_affinity(i, j));
 339                 }
 340                 sbuf_printf(&sbuf, "\n");
 341         }
 342         error = sbuf_finish(&sbuf);
 343         sbuf_delete(&sbuf);
 344         return (error);
 345 }
 346 #endif
 347
 348 static void
 349 vm_freelist_add(struct vm_freelist *fl, vm_page_t m, int order, int tail)
 350 {
 351
 352         m->order = order;
 353         if (tail)
 354                 TAILQ_INSERT_TAIL(&fl[order].pl, m, listq);
 355         else
 356                 TAILQ_INSERT_HEAD(&fl[order].pl, m, listq);
 357         fl[order].lcnt++;
 358 }
 359
 360 static void
 361 vm_freelist_rem(struct vm_freelist *fl, vm_page_t m, int order)
 362 {
 363
 364         TAILQ_REMOVE(&fl[order].pl, m, listq);
 365         fl[order].lcnt--;
 366         m->order = VM_NFREEORDER;
 367 }
 368
 369 /*
 370  * Create a physical memory segment.
 371  */
 372 static void
 373 _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain)
 374 {
 375         struct vm_phys_seg *seg;
 376
 377         KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX,
 378             ("vm_phys_create_seg: increase VM_PHYSSEG_MAX"));
 379         KASSERT(domain >= 0 && domain < vm_ndomains,
 380             ("vm_phys_create_seg: invalid domain provided"));
 381         seg = &vm_phys_segs[vm_phys_nsegs++];
 382         while (seg > vm_phys_segs && (seg - 1)->start >= end) {
 383                 *seg = *(seg - 1);
 384                 seg--;
 385         }
 386         seg->start = start;
 387         seg->end = end;
 388         seg->domain = domain;
 389 }
 390
 391 static void
 392 vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end)
 393 {
 394 #ifdef NUMA
 395         int i;
 396
 397         if (mem_affinity == NULL) {
 398                 _vm_phys_create_seg(start, end, 0);
 399                 return;
 400         }
 401
 402         for (i = 0;; i++) {
 403                 if (mem_affinity[i].end == 0)
 404                         panic("Reached end of affinity info");
 405                 if (mem_affinity[i].end <= start)
 406                         continue;
 407                 if (mem_affinity[i].start > start)
 408                         panic("No affinity info for start %jx",
 409                             (uintmax_t)start);
 410                 if (mem_affinity[i].end >= end) {
 411                         _vm_phys_create_seg(start, end,
 412                             mem_affinity[i].domain);
 413                         break;
 414                 }
 415                 _vm_phys_create_seg(start, mem_affinity[i].end,
 416                     mem_affinity[i].domain);
 417                 start = mem_affinity[i].end;
 418         }
 419 #else
 420         _vm_phys_create_seg(start, end, 0);
 421 #endif
 422 }
 423
 424 /*
 425  * Add a physical memory segment.
 426  */
 427 void
 428 vm_phys_add_seg(vm_paddr_t start, vm_paddr_t end)
 429 {
 430         vm_paddr_t paddr;
 431
 432         KASSERT((start & PAGE_MASK) == 0,
 433             ("vm_phys_define_seg: start is not page aligned"));
 434         KASSERT((end & PAGE_MASK) == 0,
 435             ("vm_phys_define_seg: end is not page aligned"));
 436
 437         /*
 438          * Split the physical memory segment if it spans two or more free
 439          * list boundaries.
 440          */
 441         paddr = start;
 442 #ifdef  VM_FREELIST_LOWMEM
 443         if (paddr < VM_LOWMEM_BOUNDARY && end > VM_LOWMEM_BOUNDARY) {
 444                 vm_phys_create_seg(paddr, VM_LOWMEM_BOUNDARY);
 445                 paddr = VM_LOWMEM_BOUNDARY;
 446         }
 447 #endif
 448 #ifdef  VM_FREELIST_DMA32
 449         if (paddr < VM_DMA32_BOUNDARY && end > VM_DMA32_BOUNDARY) {
 450                 vm_phys_create_seg(paddr, VM_DMA32_BOUNDARY);
 451                 paddr = VM_DMA32_BOUNDARY;
 452         }
 453 #endif
 454         vm_phys_create_seg(paddr, end);
 455 }
 456
 457 /*
 458  * Initialize the physical memory allocator.
 459  *
 460  * Requires that vm_page_array is initialized!
 461  */
 462 void
 463 vm_phys_init(void)
 464 {
 465         struct vm_freelist *fl;
 466         struct vm_phys_seg *end_seg, *prev_seg, *seg, *tmp_seg;
 467         u_long npages;
 468         int dom, flind, freelist, oind, pind, segind;
 469
 470         /*
 471          * Compute the number of free lists, and generate the mapping from the
 472          * manifest constants VM_FREELIST_* to the free list indices.
 473          *
 474          * Initially, the entries of vm_freelist_to_flind[] are set to either
 475          * 0 or 1 to indicate which free lists should be created.
 476          */
 477         npages = 0;
 478         for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
 479                 seg = &vm_phys_segs[segind];
 480 #ifdef  VM_FREELIST_LOWMEM
 481                 if (seg->end <= VM_LOWMEM_BOUNDARY)
 482                         vm_freelist_to_flind[VM_FREELIST_LOWMEM] = 1;
 483                 else
 484 #endif
 485 #ifdef  VM_FREELIST_DMA32
 486                 if (
 487 #ifdef  VM_DMA32_NPAGES_THRESHOLD
 488                     /*
 489                      * Create the DMA32 free list only if the amount of
 490                      * physical memory above physical address 4G exceeds the
 491                      * given threshold.
 492                      */
 493                     npages > VM_DMA32_NPAGES_THRESHOLD &&
 494 #endif
 495                     seg->end <= VM_DMA32_BOUNDARY)
 496                         vm_freelist_to_flind[VM_FREELIST_DMA32] = 1;
 497                 else
 498 #endif
 499                 {
 500                         npages += atop(seg->end - seg->start);
 501                         vm_freelist_to_flind[VM_FREELIST_DEFAULT] = 1;
 502                 }
 503         }
 504         /* Change each entry into a running total of the free lists. */
 505         for (freelist = 1; freelist < VM_NFREELIST; freelist++) {
 506                 vm_freelist_to_flind[freelist] +=
 507                     vm_freelist_to_flind[freelist - 1];
 508         }
 509         vm_nfreelists = vm_freelist_to_flind[VM_NFREELIST - 1];
 510         KASSERT(vm_nfreelists > 0, ("vm_phys_init: no free lists"));
 511         /* Change each entry into a free list index. */
 512         for (freelist = 0; freelist < VM_NFREELIST; freelist++)
 513                 vm_freelist_to_flind[freelist]--;
 514
 515         /*
 516          * Initialize the first_page and free_queues fields of each physical
 517          * memory segment.
 518          */
 519 #ifdef VM_PHYSSEG_SPARSE
 520         npages = 0;
 521 #endif
 522         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 523                 seg = &vm_phys_segs[segind];
 524 #ifdef VM_PHYSSEG_SPARSE
 525                 seg->first_page = &vm_page_array[npages];
 526                 npages += atop(seg->end - seg->start);
 527 #else
 528                 seg->first_page = PHYS_TO_VM_PAGE(seg->start);
 529 #endif
 530 #ifdef  VM_FREELIST_LOWMEM
 531                 if (seg->end <= VM_LOWMEM_BOUNDARY) {
 532                         flind = vm_freelist_to_flind[VM_FREELIST_LOWMEM];
 533                         KASSERT(flind >= 0,
 534                             ("vm_phys_init: LOWMEM flind < 0"));
 535                 } else
 536 #endif
 537 #ifdef  VM_FREELIST_DMA32
 538                 if (seg->end <= VM_DMA32_BOUNDARY) {
 539                         flind = vm_freelist_to_flind[VM_FREELIST_DMA32];
 540                         KASSERT(flind >= 0,
 541                             ("vm_phys_init: DMA32 flind < 0"));
 542                 } else
 543 #endif
 544                 {
 545                         flind = vm_freelist_to_flind[VM_FREELIST_DEFAULT];
 546                         KASSERT(flind >= 0,
 547                             ("vm_phys_init: DEFAULT flind < 0"));
 548                 }
 549                 seg->free_queues = &vm_phys_free_queues[seg->domain][flind];
 550         }
 551
 552         /*
 553          * Coalesce physical memory segments that are contiguous and share the
 554          * same per-domain free queues.
 555          */
 556         prev_seg = vm_phys_segs;
 557         seg = &vm_phys_segs[1];
 558         end_seg = &vm_phys_segs[vm_phys_nsegs];
 559         while (seg < end_seg) {
 560                 if (prev_seg->end == seg->start &&
 561                     prev_seg->free_queues == seg->free_queues) {
 562                         prev_seg->end = seg->end;
 563                         KASSERT(prev_seg->domain == seg->domain,
 564                             ("vm_phys_init: free queues cannot span domains"));
 565                         vm_phys_nsegs--;
 566                         end_seg--;
 567                         for (tmp_seg = seg; tmp_seg < end_seg; tmp_seg++)
 568                                 *tmp_seg = *(tmp_seg + 1);
 569                 } else {
 570                         prev_seg = seg;
 571                         seg++;
 572                 }
 573         }
 574
 575         /*
 576          * Initialize the free queues.
 577          */
 578         for (dom = 0; dom < vm_ndomains; dom++) {
 579                 for (flind = 0; flind < vm_nfreelists; flind++) {
 580                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 581                                 fl = vm_phys_free_queues[dom][flind][pind];
 582                                 for (oind = 0; oind < VM_NFREEORDER; oind++)
 583                                         TAILQ_INIT(&fl[oind].pl);
 584                         }
 585                 }
 586         }
 587
 588         rw_init(&vm_phys_fictitious_reg_lock, "vmfctr");
 589 }
 590
 591 /*
 592  * Register info about the NUMA topology of the system.
 593  *
 594  * Invoked by platform-dependent code prior to vm_phys_init().
 595  */
 596 void
 597 vm_phys_register_domains(int ndomains, struct mem_affinity *affinity,
 598     int *locality)
 599 {
 600 #ifdef NUMA
 601         int d, i;
 602
 603         /*
 604          * For now the only override value that we support is 1, which
 605          * effectively disables NUMA-awareness in the allocators.
 606          */
 607         d = 0;
 608         TUNABLE_INT_FETCH("vm.numa.disabled", &d);
 609         if (d)
 610                 ndomains = 1;
 611
 612         if (ndomains > 1) {
 613                 vm_ndomains = ndomains;
 614                 mem_affinity = affinity;
 615                 mem_locality = locality;
 616         }
 617
 618         for (i = 0; i < vm_ndomains; i++)
 619                 DOMAINSET_SET(i, &all_domains);
 620 #else
 621         (void)ndomains;
 622         (void)affinity;
 623         (void)locality;
 624 #endif
 625 }
 626
 627 int
 628 _vm_phys_domain(vm_paddr_t pa)
 629 {
 630 #ifdef NUMA
 631         int i;
 632
 633         if (vm_ndomains == 1 || mem_affinity == NULL)
 634                 return (0);
 635
 636         /*
 637          * Check for any memory that overlaps.
 638          */
 639         for (i = 0; mem_affinity[i].end != 0; i++)
 640                 if (mem_affinity[i].start <= pa &&
 641                     mem_affinity[i].end >= pa)
 642                         return (mem_affinity[i].domain);
 643 #endif
 644         return (0);
 645 }
 646
 647 /*
 648  * Split a contiguous, power of two-sized set of physical pages.
 649  *
 650  * When this function is called by a page allocation function, the caller
 651  * should request insertion at the head unless the order [order, oind) queues
 652  * are known to be empty.  The objective being to reduce the likelihood of
 653  * long-term fragmentation by promoting contemporaneous allocation and
 654  * (hopefully) deallocation.
 655  */
 656 static __inline void
 657 vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order,
 658     int tail)
 659 {
 660         vm_page_t m_buddy;
 661
 662         while (oind > order) {
 663                 oind--;
 664                 m_buddy = &m[1 << oind];
 665                 KASSERT(m_buddy->order == VM_NFREEORDER,
 666                     ("vm_phys_split_pages: page %p has unexpected order %d",
 667                     m_buddy, m_buddy->order));
 668                 vm_freelist_add(fl, m_buddy, oind, tail);
 669         }
 670 }
 671
 672 /*
 673  * Add the physical pages [m, m + npages) at the end of a power-of-two aligned
 674  * and sized set to the specified free list.
 675  *
 676  * When this function is called by a page allocation function, the caller
 677  * should request insertion at the head unless the lower-order queues are
 678  * known to be empty.  The objective being to reduce the likelihood of long-
 679  * term fragmentation by promoting contemporaneous allocation and (hopefully)
 680  * deallocation.
 681  *
 682  * The physical page m's buddy must not be free.
 683  */
 684 static void
 685 vm_phys_enq_range(vm_page_t m, u_int npages, struct vm_freelist *fl, int tail)
 686 {
 687         u_int n;
 688         int order;
 689
 690         KASSERT(npages > 0, ("vm_phys_enq_range: npages is 0"));
 691         KASSERT(((VM_PAGE_TO_PHYS(m) + npages * PAGE_SIZE) &
 692             ((PAGE_SIZE << (fls(npages) - 1)) - 1)) == 0,
 693             ("vm_phys_enq_range: page %p and npages %u are misaligned",
 694             m, npages));
 695         do {
 696                 KASSERT(m->order == VM_NFREEORDER,
 697                     ("vm_phys_enq_range: page %p has unexpected order %d",
 698                     m, m->order));
 699                 order = ffs(npages) - 1;
 700                 KASSERT(order < VM_NFREEORDER,
 701                     ("vm_phys_enq_range: order %d is out of range", order));
 702                 vm_freelist_add(fl, m, order, tail);
 703                 n = 1 << order;
 704                 m += n;
 705                 npages -= n;
 706         } while (npages > 0);
 707 }
 708
 709 /*
 710  * Tries to allocate the specified number of pages from the specified pool
 711  * within the specified domain.  Returns the actual number of allocated pages
 712  * and a pointer to each page through the array ma[].
 713  *
 714  * The returned pages may not be physically contiguous.  However, in contrast
 715  * to performing multiple, back-to-back calls to vm_phys_alloc_pages(..., 0),
 716  * calling this function once to allocate the desired number of pages will
 717  * avoid wasted time in vm_phys_split_pages().
 718  *
 719  * The free page queues for the specified domain must be locked.
 720  */
 721 int
 722 vm_phys_alloc_npages(int domain, int pool, int npages, vm_page_t ma[])
 723 {
 724         struct vm_freelist *alt, *fl;
 725         vm_page_t m;
 726         int avail, end, flind, freelist, i, need, oind, pind;
 727
 728         KASSERT(domain >= 0 && domain < vm_ndomains,
 729             ("vm_phys_alloc_npages: domain %d is out of range", domain));
 730         KASSERT(pool < VM_NFREEPOOL,
 731             ("vm_phys_alloc_npages: pool %d is out of range", pool));
 732         KASSERT(npages <= 1 << (VM_NFREEORDER - 1),
 733             ("vm_phys_alloc_npages: npages %d is out of range", npages));
 734         vm_domain_free_assert_locked(VM_DOMAIN(domain));
 735         i = 0;
 736         for (freelist = 0; freelist < VM_NFREELIST; freelist++) {
 737                 flind = vm_freelist_to_flind[freelist];
 738                 if (flind < 0)
 739                         continue;
 740                 fl = vm_phys_free_queues[domain][flind][pool];
 741                 for (oind = 0; oind < VM_NFREEORDER; oind++) {
 742                         while ((m = TAILQ_FIRST(&fl[oind].pl)) != NULL) {
 743                                 vm_freelist_rem(fl, m, oind);
 744                                 avail = 1 << oind;
 745                                 need = imin(npages - i, avail);
 746                                 for (end = i + need; i < end;)
 747                                         ma[i++] = m++;
 748                                 if (need < avail) {
 749                                         /*
 750                                          * Return excess pages to fl.  Its
 751                                          * order [0, oind) queues are empty.
 752                                          */
 753                                         vm_phys_enq_range(m, avail - need, fl,
 754                                             1);
 755                                         return (npages);
 756                                 } else if (i == npages)
 757                                         return (npages);
 758                         }
 759                 }
 760                 for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
 761                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 762                                 alt = vm_phys_free_queues[domain][flind][pind];
 763                                 while ((m = TAILQ_FIRST(&alt[oind].pl)) !=
 764                                     NULL) {
 765                                         vm_freelist_rem(alt, m, oind);
 766                                         vm_phys_set_pool(pool, m, oind);
 767                                         avail = 1 << oind;
 768                                         need = imin(npages - i, avail);
 769                                         for (end = i + need; i < end;)
 770                                                 ma[i++] = m++;
 771                                         if (need < avail) {
 772                                                 /*
 773                                                  * Return excess pages to fl.
 774                                                  * Its order [0, oind) queues
 775                                                  * are empty.
 776                                                  */
 777                                                 vm_phys_enq_range(m, avail -
 778                                                     need, fl, 1);
 779                                                 return (npages);
 780                                         } else if (i == npages)
 781                                                 return (npages);
 782                                 }
 783                         }
 784                 }
 785         }
 786         return (i);
 787 }
 788
 789 /*
 790  * Allocate a contiguous, power of two-sized set of physical pages
 791  * from the free lists.
 792  *
 793  * The free page queues must be locked.
 794  */
 795 vm_page_t
 796 vm_phys_alloc_pages(int domain, int pool, int order)
 797 {
 798         vm_page_t m;
 799         int freelist;
 800
 801         for (freelist = 0; freelist < VM_NFREELIST; freelist++) {
 802                 m = vm_phys_alloc_freelist_pages(domain, freelist, pool, order);
 803                 if (m != NULL)
 804                         return (m);
 805         }
 806         return (NULL);
 807 }
 808
 809 /*
 810  * Allocate a contiguous, power of two-sized set of physical pages from the
 811  * specified free list.  The free list must be specified using one of the
 812  * manifest constants VM_FREELIST_*.
 813  *
 814  * The free page queues must be locked.
 815  */
 816 vm_page_t
 817 vm_phys_alloc_freelist_pages(int domain, int freelist, int pool, int order)
 818 {
 819         struct vm_freelist *alt, *fl;
 820         vm_page_t m;
 821         int oind, pind, flind;
 822
 823         KASSERT(domain >= 0 && domain < vm_ndomains,
 824             ("vm_phys_alloc_freelist_pages: domain %d is out of range",
 825             domain));
 826         KASSERT(freelist < VM_NFREELIST,
 827             ("vm_phys_alloc_freelist_pages: freelist %d is out of range",
 828             freelist));
 829         KASSERT(pool < VM_NFREEPOOL,
 830             ("vm_phys_alloc_freelist_pages: pool %d is out of range", pool));
 831         KASSERT(order < VM_NFREEORDER,
 832             ("vm_phys_alloc_freelist_pages: order %d is out of range", order));
 833
 834         flind = vm_freelist_to_flind[freelist];
 835         /* Check if freelist is present */
 836         if (flind < 0)
 837                 return (NULL);
 838
 839         vm_domain_free_assert_locked(VM_DOMAIN(domain));
 840         fl = &vm_phys_free_queues[domain][flind][pool][0];
 841         for (oind = order; oind < VM_NFREEORDER; oind++) {
 842                 m = TAILQ_FIRST(&fl[oind].pl);
 843                 if (m != NULL) {
 844                         vm_freelist_rem(fl, m, oind);
 845                         /* The order [order, oind) queues are empty. */
 846                         vm_phys_split_pages(m, oind, fl, order, 1);
 847                         return (m);
 848                 }
 849         }
 850
 851         /*
 852          * The given pool was empty.  Find the largest
 853          * contiguous, power-of-two-sized set of pages in any
 854          * pool.  Transfer these pages to the given pool, and
 855          * use them to satisfy the allocation.
 856          */
 857         for (oind = VM_NFREEORDER - 1; oind >= order; oind--) {
 858                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 859                         alt = &vm_phys_free_queues[domain][flind][pind][0];
 860                         m = TAILQ_FIRST(&alt[oind].pl);
 861                         if (m != NULL) {
 862                                 vm_freelist_rem(alt, m, oind);
 863                                 vm_phys_set_pool(pool, m, oind);
 864                                 /* The order [order, oind) queues are empty. */
 865                                 vm_phys_split_pages(m, oind, fl, order, 1);
 866                                 return (m);
 867                         }
 868                 }
 869         }
 870         return (NULL);
 871 }
 872
 873 /*
 874  * Find the vm_page corresponding to the given physical address.
 875  */
 876 vm_page_t
 877 vm_phys_paddr_to_vm_page(vm_paddr_t pa)
 878 {
 879         struct vm_phys_seg *seg;
 880         int segind;
 881
 882         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 883                 seg = &vm_phys_segs[segind];
 884                 if (pa >= seg->start && pa < seg->end)
 885                         return (&seg->first_page[atop(pa - seg->start)]);
 886         }
 887         return (NULL);
 888 }
 889
 890 vm_page_t
 891 vm_phys_fictitious_to_vm_page(vm_paddr_t pa)
 892 {
 893         struct vm_phys_fictitious_seg tmp, *seg;
 894         vm_page_t m;
 895
 896         m = NULL;
 897         tmp.start = pa;
 898         tmp.end = 0;
 899
 900         rw_rlock(&vm_phys_fictitious_reg_lock);
 901         seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
 902         rw_runlock(&vm_phys_fictitious_reg_lock);
 903         if (seg == NULL)
 904                 return (NULL);
 905
 906         m = &seg->first_page[atop(pa - seg->start)];
 907         KASSERT((m->flags & PG_FICTITIOUS) != 0, ("%p not fictitious", m));
 908
 909         return (m);
 910 }
 911
 912 static inline void
 913 vm_phys_fictitious_init_range(vm_page_t range, vm_paddr_t start,
 914     long page_count, vm_memattr_t memattr)
 915 {
 916         long i;
 917
 918         bzero(range, page_count * sizeof(*range));
 919         for (i = 0; i < page_count; i++) {
 920                 vm_page_initfake(&range[i], start + PAGE_SIZE * i, memattr);
 921                 range[i].oflags &= ~VPO_UNMANAGED;
 922                 range[i].busy_lock = VPB_UNBUSIED;
 923         }
 924 }
 925
 926 int
 927 vm_phys_fictitious_reg_range(vm_paddr_t start, vm_paddr_t end,
 928     vm_memattr_t memattr)
 929 {
 930         struct vm_phys_fictitious_seg *seg;
 931         vm_page_t fp;
 932         long page_count;
 933 #ifdef VM_PHYSSEG_DENSE
 934         long pi, pe;
 935         long dpage_count;
 936 #endif
 937
 938         KASSERT(start < end,
 939             ("Start of segment isn't less than end (start: %jx end: %jx)",
 940             (uintmax_t)start, (uintmax_t)end));
 941
 942         page_count = (end - start) / PAGE_SIZE;
 943
 944 #ifdef VM_PHYSSEG_DENSE
 945         pi = atop(start);
 946         pe = atop(end);
 947         if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
 948                 fp = &vm_page_array[pi - first_page];
 949                 if ((pe - first_page) > vm_page_array_size) {
 950                         /*
 951                          * We have a segment that starts inside
 952                          * of vm_page_array, but ends outside of it.
 953                          *
 954                          * Use vm_page_array pages for those that are
 955                          * inside of the vm_page_array range, and
 956                          * allocate the remaining ones.
 957                          */
 958                         dpage_count = vm_page_array_size - (pi - first_page);
 959                         vm_phys_fictitious_init_range(fp, start, dpage_count,
 960                             memattr);
 961                         page_count -= dpage_count;
 962                         start += ptoa(dpage_count);
 963                         goto alloc;
 964                 }
 965                 /*
 966                  * We can allocate the full range from vm_page_array,
 967                  * so there's no need to register the range in the tree.
 968                  */
 969                 vm_phys_fictitious_init_range(fp, start, page_count, memattr);
 970                 return (0);
 971         } else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
 972                 /*
 973                  * We have a segment that ends inside of vm_page_array,
 974                  * but starts outside of it.
 975                  */
 976                 fp = &vm_page_array[0];
 977                 dpage_count = pe - first_page;
 978                 vm_phys_fictitious_init_range(fp, ptoa(first_page), dpage_count,
 979                     memattr);
 980                 end -= ptoa(dpage_count);
 981                 page_count -= dpage_count;
 982                 goto alloc;
 983         } else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
 984                 /*
 985                  * Trying to register a fictitious range that expands before
 986                  * and after vm_page_array.
 987                  */
 988                 return (EINVAL);
 989         } else {
 990 alloc:
 991 #endif
 992                 fp = malloc(page_count * sizeof(struct vm_page), M_FICT_PAGES,
 993                     M_WAITOK);
 994 #ifdef VM_PHYSSEG_DENSE
 995         }
 996 #endif
 997         vm_phys_fictitious_init_range(fp, start, page_count, memattr);
 998
 999         seg = malloc(sizeof(*seg), M_FICT_PAGES, M_WAITOK | M_ZERO);
1000         seg->start = start;
1001         seg->end = end;
1002         seg->first_page = fp;
1003
1004         rw_wlock(&vm_phys_fictitious_reg_lock);
1005         RB_INSERT(fict_tree, &vm_phys_fictitious_tree, seg);
1006         rw_wunlock(&vm_phys_fictitious_reg_lock);
1007
1008         return (0);
1009 }
1010
1011 void
1012 vm_phys_fictitious_unreg_range(vm_paddr_t start, vm_paddr_t end)
1013 {
1014         struct vm_phys_fictitious_seg *seg, tmp;
1015 #ifdef VM_PHYSSEG_DENSE
1016         long pi, pe;
1017 #endif
1018
1019         KASSERT(start < end,
1020             ("Start of segment isn't less than end (start: %jx end: %jx)",
1021             (uintmax_t)start, (uintmax_t)end));
1022
1023 #ifdef VM_PHYSSEG_DENSE
1024         pi = atop(start);
1025         pe = atop(end);
1026         if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1027                 if ((pe - first_page) <= vm_page_array_size) {
1028                         /*
1029                          * This segment was allocated using vm_page_array
1030                          * only, there's nothing to do since those pages
1031                          * were never added to the tree.
1032                          */
1033                         return;
1034                 }
1035                 /*
1036                  * We have a segment that starts inside
1037                  * of vm_page_array, but ends outside of it.
1038                  *
1039                  * Calculate how many pages were added to the
1040                  * tree and free them.
1041                  */
1042                 start = ptoa(first_page + vm_page_array_size);
1043         } else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
1044                 /*
1045                  * We have a segment that ends inside of vm_page_array,
1046                  * but starts outside of it.
1047                  */
1048                 end = ptoa(first_page);
1049         } else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
1050                 /* Since it's not possible to register such a range, panic. */
1051                 panic(
1052                     "Unregistering not registered fictitious range [%#jx:%#jx]",
1053                     (uintmax_t)start, (uintmax_t)end);
1054         }
1055 #endif
1056         tmp.start = start;
1057         tmp.end = 0;
1058
1059         rw_wlock(&vm_phys_fictitious_reg_lock);
1060         seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
1061         if (seg->start != start || seg->end != end) {
1062                 rw_wunlock(&vm_phys_fictitious_reg_lock);
1063                 panic(
1064                     "Unregistering not registered fictitious range [%#jx:%#jx]",
1065                     (uintmax_t)start, (uintmax_t)end);
1066         }
1067         RB_REMOVE(fict_tree, &vm_phys_fictitious_tree, seg);
1068         rw_wunlock(&vm_phys_fictitious_reg_lock);
1069         free(seg->first_page, M_FICT_PAGES);
1070         free(seg, M_FICT_PAGES);
1071 }
1072
1073 /*
1074  * Free a contiguous, power of two-sized set of physical pages.
1075  *
1076  * The free page queues must be locked.
1077  */
1078 void
1079 vm_phys_free_pages(vm_page_t m, int order)
1080 {
1081         struct vm_freelist *fl;
1082         struct vm_phys_seg *seg;
1083         vm_paddr_t pa;
1084         vm_page_t m_buddy;
1085
1086         KASSERT(m->order == VM_NFREEORDER,
1087             ("vm_phys_free_pages: page %p has unexpected order %d",
1088             m, m->order));
1089         KASSERT(m->pool < VM_NFREEPOOL,
1090             ("vm_phys_free_pages: page %p has unexpected pool %d",
1091             m, m->pool));
1092         KASSERT(order < VM_NFREEORDER,
1093             ("vm_phys_free_pages: order %d is out of range", order));
1094         seg = &vm_phys_segs[m->segind];
1095         vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1096         if (order < VM_NFREEORDER - 1) {
1097                 pa = VM_PAGE_TO_PHYS(m);
1098                 do {
1099                         pa ^= ((vm_paddr_t)1 << (PAGE_SHIFT + order));
1100                         if (pa < seg->start || pa >= seg->end)
1101                                 break;
1102                         m_buddy = &seg->first_page[atop(pa - seg->start)];
1103                         if (m_buddy->order != order)
1104                                 break;
1105                         fl = (*seg->free_queues)[m_buddy->pool];
1106                         vm_freelist_rem(fl, m_buddy, order);
1107                         if (m_buddy->pool != m->pool)
1108                                 vm_phys_set_pool(m->pool, m_buddy, order);
1109                         order++;
1110                         pa &= ~(((vm_paddr_t)1 << (PAGE_SHIFT + order)) - 1);
1111                         m = &seg->first_page[atop(pa - seg->start)];
1112                 } while (order < VM_NFREEORDER - 1);
1113         }
1114         fl = (*seg->free_queues)[m->pool];
1115         vm_freelist_add(fl, m, order, 1);
1116 }
1117
1118 /*
1119  * Free a contiguous, arbitrarily sized set of physical pages.
1120  *
1121  * The free page queues must be locked.
1122  */
1123 void
1124 vm_phys_free_contig(vm_page_t m, u_long npages)
1125 {
1126         u_int n;
1127         int order;
1128
1129         /*
1130          * Avoid unnecessary coalescing by freeing the pages in the largest
1131          * possible power-of-two-sized subsets.
1132          */
1133         vm_domain_free_assert_locked(vm_pagequeue_domain(m));
1134         for (;; npages -= n) {
1135                 /*
1136                  * Unsigned "min" is used here so that "order" is assigned
1137                  * "VM_NFREEORDER - 1" when "m"'s physical address is zero
1138                  * or the low-order bits of its physical address are zero
1139                  * because the size of a physical address exceeds the size of
1140                  * a long.
1141                  */
1142                 order = min(ffsl(VM_PAGE_TO_PHYS(m) >> PAGE_SHIFT) - 1,
1143                     VM_NFREEORDER - 1);
1144                 n = 1 << order;
1145                 if (npages < n)
1146                         break;
1147                 vm_phys_free_pages(m, order);
1148                 m += n;
1149         }
1150         /* The residual "npages" is less than "1 << (VM_NFREEORDER - 1)". */
1151         for (; npages > 0; npages -= n) {
1152                 order = flsl(npages) - 1;
1153                 n = 1 << order;
1154                 vm_phys_free_pages(m, order);
1155                 m += n;
1156         }
1157 }
1158
1159 /*
1160  * Scan physical memory between the specified addresses "low" and "high" for a
1161  * run of contiguous physical pages that satisfy the specified conditions, and
1162  * return the lowest page in the run.  The specified "alignment" determines
1163  * the alignment of the lowest physical page in the run.  If the specified
1164  * "boundary" is non-zero, then the run of physical pages cannot span a
1165  * physical address that is a multiple of "boundary".
1166  *
1167  * "npages" must be greater than zero.  Both "alignment" and "boundary" must
1168  * be a power of two.
1169  */
1170 vm_page_t
1171 vm_phys_scan_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
1172     u_long alignment, vm_paddr_t boundary, int options)
1173 {
1174         vm_paddr_t pa_end;
1175         vm_page_t m_end, m_run, m_start;
1176         struct vm_phys_seg *seg;
1177         int segind;
1178
1179         KASSERT(npages > 0, ("npages is 0"));
1180         KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1181         KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1182         if (low >= high)
1183                 return (NULL);
1184         for (segind = 0; segind < vm_phys_nsegs; segind++) {
1185                 seg = &vm_phys_segs[segind];
1186                 if (seg->domain != domain)
1187                         continue;
1188                 if (seg->start >= high)
1189                         break;
1190                 if (low >= seg->end)
1191                         continue;
1192                 if (low <= seg->start)
1193                         m_start = seg->first_page;
1194                 else
1195                         m_start = &seg->first_page[atop(low - seg->start)];
1196                 if (high < seg->end)
1197                         pa_end = high;
1198                 else
1199                         pa_end = seg->end;
1200                 if (pa_end - VM_PAGE_TO_PHYS(m_start) < ptoa(npages))
1201                         continue;
1202                 m_end = &seg->first_page[atop(pa_end - seg->start)];
1203                 m_run = vm_page_scan_contig(npages, m_start, m_end,
1204                     alignment, boundary, options);
1205                 if (m_run != NULL)
1206                         return (m_run);
1207         }
1208         return (NULL);
1209 }
1210
1211 /*
1212  * Set the pool for a contiguous, power of two-sized set of physical pages.
1213  */
1214 void
1215 vm_phys_set_pool(int pool, vm_page_t m, int order)
1216 {
1217         vm_page_t m_tmp;
1218
1219         for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++)
1220                 m_tmp->pool = pool;
1221 }
1222
1223 /*
1224  * Search for the given physical page "m" in the free lists.  If the search
1225  * succeeds, remove "m" from the free lists and return TRUE.  Otherwise, return
1226  * FALSE, indicating that "m" is not in the free lists.
1227  *
1228  * The free page queues must be locked.
1229  */
1230 boolean_t
1231 vm_phys_unfree_page(vm_page_t m)
1232 {
1233         struct vm_freelist *fl;
1234         struct vm_phys_seg *seg;
1235         vm_paddr_t pa, pa_half;
1236         vm_page_t m_set, m_tmp;
1237         int order;
1238
1239         /*
1240          * First, find the contiguous, power of two-sized set of free
1241          * physical pages containing the given physical page "m" and
1242          * assign it to "m_set".
1243          */
1244         seg = &vm_phys_segs[m->segind];
1245         vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1246         for (m_set = m, order = 0; m_set->order == VM_NFREEORDER &&
1247             order < VM_NFREEORDER - 1; ) {
1248                 order++;
1249                 pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order));
1250                 if (pa >= seg->start)
1251                         m_set = &seg->first_page[atop(pa - seg->start)];
1252                 else
1253                         return (FALSE);
1254         }
1255         if (m_set->order < order)
1256                 return (FALSE);
1257         if (m_set->order == VM_NFREEORDER)
1258                 return (FALSE);
1259         KASSERT(m_set->order < VM_NFREEORDER,
1260             ("vm_phys_unfree_page: page %p has unexpected order %d",
1261             m_set, m_set->order));
1262
1263         /*
1264          * Next, remove "m_set" from the free lists.  Finally, extract
1265          * "m" from "m_set" using an iterative algorithm: While "m_set"
1266          * is larger than a page, shrink "m_set" by returning the half
1267          * of "m_set" that does not contain "m" to the free lists.
1268          */
1269         fl = (*seg->free_queues)[m_set->pool];
1270         order = m_set->order;
1271         vm_freelist_rem(fl, m_set, order);
1272         while (order > 0) {
1273                 order--;
1274                 pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order));
1275                 if (m->phys_addr < pa_half)
1276                         m_tmp = &seg->first_page[atop(pa_half - seg->start)];
1277                 else {
1278                         m_tmp = m_set;
1279                         m_set = &seg->first_page[atop(pa_half - seg->start)];
1280                 }
1281                 vm_freelist_add(fl, m_tmp, order, 0);
1282         }
1283         KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency"));
1284         return (TRUE);
1285 }
1286
1287 /*
1288  * Allocate a contiguous set of physical pages of the given size
1289  * "npages" from the free lists.  All of the physical pages must be at
1290  * or above the given physical address "low" and below the given
1291  * physical address "high".  The given value "alignment" determines the
1292  * alignment of the first physical page in the set.  If the given value
1293  * "boundary" is non-zero, then the set of physical pages cannot cross
1294  * any physical address boundary that is a multiple of that value.  Both
1295  * "alignment" and "boundary" must be a power of two.
1296  */
1297 vm_page_t
1298 vm_phys_alloc_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
1299     u_long alignment, vm_paddr_t boundary)
1300 {
1301         vm_paddr_t pa_end, pa_start;
1302         vm_page_t m_run;
1303         struct vm_phys_seg *seg;
1304         int segind;
1305
1306         KASSERT(npages > 0, ("npages is 0"));
1307         KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1308         KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1309         vm_domain_free_assert_locked(VM_DOMAIN(domain));
1310         if (low >= high)
1311                 return (NULL);
1312         m_run = NULL;
1313         for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
1314                 seg = &vm_phys_segs[segind];
1315                 if (seg->start >= high || seg->domain != domain)
1316                         continue;
1317                 if (low >= seg->end)
1318                         break;
1319                 if (low <= seg->start)
1320                         pa_start = seg->start;
1321                 else
1322                         pa_start = low;
1323                 if (high < seg->end)
1324                         pa_end = high;
1325                 else
1326                         pa_end = seg->end;
1327                 if (pa_end - pa_start < ptoa(npages))
1328                         continue;
1329                 m_run = vm_phys_alloc_seg_contig(seg, npages, low, high,
1330                     alignment, boundary);
1331                 if (m_run != NULL)
1332                         break;
1333         }
1334         return (m_run);
1335 }
1336
1337 /*
1338  * Allocate a run of contiguous physical pages from the free list for the
1339  * specified segment.
1340  */
1341 static vm_page_t
1342 vm_phys_alloc_seg_contig(struct vm_phys_seg *seg, u_long npages,
1343     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1344 {
1345         struct vm_freelist *fl;
1346         vm_paddr_t pa, pa_end, size;
1347         vm_page_t m, m_ret;
1348         u_long npages_end;
1349         int oind, order, pind;
1350
1351         KASSERT(npages > 0, ("npages is 0"));
1352         KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1353         KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1354         vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1355         /* Compute the queue that is the best fit for npages. */
1356         order = flsl(npages - 1);
1357         /* Search for a run satisfying the specified conditions. */
1358         size = npages << PAGE_SHIFT;
1359         for (oind = min(order, VM_NFREEORDER - 1); oind < VM_NFREEORDER;
1360             oind++) {
1361                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1362                         fl = (*seg->free_queues)[pind];
1363                         TAILQ_FOREACH(m_ret, &fl[oind].pl, listq) {
1364                                 /*
1365                                  * Is the size of this allocation request
1366                                  * larger than the largest block size?
1367                                  */
1368                                 if (order >= VM_NFREEORDER) {
1369                                         /*
1370                                          * Determine if a sufficient number of
1371                                          * subsequent blocks to satisfy the
1372                                          * allocation request are free.
1373                                          */
1374                                         pa = VM_PAGE_TO_PHYS(m_ret);
1375                                         pa_end = pa + size;
1376                                         if (pa_end < pa)
1377                                                 continue;
1378                                         for (;;) {
1379                                                 pa += 1 << (PAGE_SHIFT +
1380                                                     VM_NFREEORDER - 1);
1381                                                 if (pa >= pa_end ||
1382                                                     pa < seg->start ||
1383                                                     pa >= seg->end)
1384                                                         break;
1385                                                 m = &seg->first_page[atop(pa -
1386                                                     seg->start)];
1387                                                 if (m->order != VM_NFREEORDER -
1388                                                     1)
1389                                                         break;
1390                                         }
1391                                         /* If not, go to the next block. */
1392                                         if (pa < pa_end)
1393                                                 continue;
1394                                 }
1395
1396                                 /*
1397                                  * Determine if the blocks are within the
1398                                  * given range, satisfy the given alignment,
1399                                  * and do not cross the given boundary.
1400                                  */
1401                                 pa = VM_PAGE_TO_PHYS(m_ret);
1402                                 pa_end = pa + size;
1403                                 if (pa >= low && pa_end <= high &&
1404                                     (pa & (alignment - 1)) == 0 &&
1405                                     rounddown2(pa ^ (pa_end - 1), boundary) == 0)
1406                                         goto done;
1407                         }
1408                 }
1409         }
1410         return (NULL);
1411 done:
1412         for (m = m_ret; m < &m_ret[npages]; m = &m[1 << oind]) {
1413                 fl = (*seg->free_queues)[m->pool];
1414                 vm_freelist_rem(fl, m, oind);
1415                 if (m->pool != VM_FREEPOOL_DEFAULT)
1416                         vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, oind);
1417         }
1418         /* Return excess pages to the free lists. */
1419         npages_end = roundup2(npages, 1 << oind);
1420         if (npages < npages_end) {
1421                 fl = (*seg->free_queues)[VM_FREEPOOL_DEFAULT];
1422                 vm_phys_enq_range(&m_ret[npages], npages_end - npages, fl, 0);
1423         }
1424         return (m_ret);
1425 }
1426
1427 #ifdef DDB
1428 /*
1429  * Show the number of physical pages in each of the free lists.
1430  */
1431 DB_SHOW_COMMAND(freepages, db_show_freepages)
1432 {
1433         struct vm_freelist *fl;
1434         int flind, oind, pind, dom;
1435
1436         for (dom = 0; dom < vm_ndomains; dom++) {
1437                 db_printf("DOMAIN: %d\n", dom);
1438                 for (flind = 0; flind < vm_nfreelists; flind++) {
1439                         db_printf("FREE LIST %d:\n"
1440                             "\n  ORDER (SIZE)  |  NUMBER"
1441                             "\n              ", flind);
1442                         for (pind = 0; pind < VM_NFREEPOOL; pind++)
1443                                 db_printf("  |  POOL %d", pind);
1444                         db_printf("\n--            ");
1445                         for (pind = 0; pind < VM_NFREEPOOL; pind++)
1446                                 db_printf("-- --      ");
1447                         db_printf("--\n");
1448                         for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
1449                                 db_printf("  %2.2d (%6.6dK)", oind,
1450                                     1 << (PAGE_SHIFT - 10 + oind));
1451                                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1452                                 fl = vm_phys_free_queues[dom][flind][pind];
1453                                         db_printf("  |  %6.6d", fl[oind].lcnt);
1454                                 }
1455                                 db_printf("\n");
1456                         }
1457                         db_printf("\n");
1458                 }
1459                 db_printf("\n");
1460         }
1461 }
1462 #endif