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/queue.h>
  52 #include <sys/sbuf.h>
  53 #include <sys/sysctl.h>
  54 #include <sys/vmmeter.h>
  55
  56 #include <ddb/ddb.h>
  57
  58 #include <vm/vm.h>
  59 #include <vm/vm_param.h>
  60 #include <vm/vm_kern.h>
  61 #include <vm/vm_object.h>
  62 #include <vm/vm_page.h>
  63 #include <vm/vm_phys.h>
  64
  65 /*
  66  * VM_FREELIST_DEFAULT is split into VM_NDOMAIN lists, one for each
  67  * domain.  These extra lists are stored at the end of the regular
  68  * free lists starting with VM_NFREELIST.
  69  */
  70 #define VM_RAW_NFREELIST        (VM_NFREELIST + VM_NDOMAIN - 1)
  71
  72 struct vm_freelist {
  73         struct pglist pl;
  74         int lcnt;
  75 };
  76
  77 struct vm_phys_seg {
  78         vm_paddr_t      start;
  79         vm_paddr_t      end;
  80         vm_page_t       first_page;
  81         int             domain;
  82         struct vm_freelist (*free_queues)[VM_NFREEPOOL][VM_NFREEORDER];
  83 };
  84
  85 struct mem_affinity *mem_affinity;
  86
  87 static struct vm_phys_seg vm_phys_segs[VM_PHYSSEG_MAX];
  88
  89 static int vm_phys_nsegs;
  90
  91 #define VM_PHYS_FICTITIOUS_NSEGS        8
  92 static struct vm_phys_fictitious_seg {
  93         vm_paddr_t      start;
  94         vm_paddr_t      end;
  95         vm_page_t       first_page;
  96 } vm_phys_fictitious_segs[VM_PHYS_FICTITIOUS_NSEGS];
  97 static struct mtx vm_phys_fictitious_reg_mtx;
  98 MALLOC_DEFINE(M_FICT_PAGES, "", "");
  99
 100 static struct vm_freelist
 101     vm_phys_free_queues[VM_RAW_NFREELIST][VM_NFREEPOOL][VM_NFREEORDER];
 102 static struct vm_freelist
 103 (*vm_phys_lookup_lists[VM_NDOMAIN][VM_RAW_NFREELIST])[VM_NFREEPOOL][VM_NFREEORDER];
 104
 105 static int vm_nfreelists = VM_FREELIST_DEFAULT + 1;
 106
 107 static int cnt_prezero;
 108 SYSCTL_INT(_vm_stats_misc, OID_AUTO, cnt_prezero, CTLFLAG_RD,
 109     &cnt_prezero, 0, "The number of physical pages prezeroed at idle time");
 110
 111 static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS);
 112 SYSCTL_OID(_vm, OID_AUTO, phys_free, CTLTYPE_STRING | CTLFLAG_RD,
 113     NULL, 0, sysctl_vm_phys_free, "A", "Phys Free Info");
 114
 115 static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS);
 116 SYSCTL_OID(_vm, OID_AUTO, phys_segs, CTLTYPE_STRING | CTLFLAG_RD,
 117     NULL, 0, sysctl_vm_phys_segs, "A", "Phys Seg Info");
 118
 119 #if VM_NDOMAIN > 1
 120 static int sysctl_vm_phys_lookup_lists(SYSCTL_HANDLER_ARGS);
 121 SYSCTL_OID(_vm, OID_AUTO, phys_lookup_lists, CTLTYPE_STRING | CTLFLAG_RD,
 122     NULL, 0, sysctl_vm_phys_lookup_lists, "A", "Phys Lookup Lists");
 123 #endif
 124
 125 static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind,
 126     int domain);
 127 static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind);
 128 static int vm_phys_paddr_to_segind(vm_paddr_t pa);
 129 static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl,
 130     int order);
 131
 132 /*
 133  * Outputs the state of the physical memory allocator, specifically,
 134  * the amount of physical memory in each free list.
 135  */
 136 static int
 137 sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS)
 138 {
 139         struct sbuf sbuf;
 140         struct vm_freelist *fl;
 141         int error, flind, oind, pind;
 142
 143         error = sysctl_wire_old_buffer(req, 0);
 144         if (error != 0)
 145                 return (error);
 146         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 147         for (flind = 0; flind < vm_nfreelists; flind++) {
 148                 sbuf_printf(&sbuf, "\nFREE LIST %d:\n"
 149                     "\n  ORDER (SIZE)  |  NUMBER"
 150                     "\n              ", flind);
 151                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 152                         sbuf_printf(&sbuf, "  |  POOL %d", pind);
 153                 sbuf_printf(&sbuf, "\n--            ");
 154                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 155                         sbuf_printf(&sbuf, "-- --      ");
 156                 sbuf_printf(&sbuf, "--\n");
 157                 for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
 158                         sbuf_printf(&sbuf, "  %2d (%6dK)", oind,
 159                             1 << (PAGE_SHIFT - 10 + oind));
 160                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 161                                 fl = vm_phys_free_queues[flind][pind];
 162                                 sbuf_printf(&sbuf, "  |  %6d", fl[oind].lcnt);
 163                         }
 164                         sbuf_printf(&sbuf, "\n");
 165                 }
 166         }
 167         error = sbuf_finish(&sbuf);
 168         sbuf_delete(&sbuf);
 169         return (error);
 170 }
 171
 172 /*
 173  * Outputs the set of physical memory segments.
 174  */
 175 static int
 176 sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS)
 177 {
 178         struct sbuf sbuf;
 179         struct vm_phys_seg *seg;
 180         int error, segind;
 181
 182         error = sysctl_wire_old_buffer(req, 0);
 183         if (error != 0)
 184                 return (error);
 185         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 186         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 187                 sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind);
 188                 seg = &vm_phys_segs[segind];
 189                 sbuf_printf(&sbuf, "start:     %#jx\n",
 190                     (uintmax_t)seg->start);
 191                 sbuf_printf(&sbuf, "end:       %#jx\n",
 192                     (uintmax_t)seg->end);
 193                 sbuf_printf(&sbuf, "domain:    %d\n", seg->domain);
 194                 sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues);
 195         }
 196         error = sbuf_finish(&sbuf);
 197         sbuf_delete(&sbuf);
 198         return (error);
 199 }
 200
 201 #if VM_NDOMAIN > 1
 202 /*
 203  * Outputs the set of free list lookup lists.
 204  */
 205 static int
 206 sysctl_vm_phys_lookup_lists(SYSCTL_HANDLER_ARGS)
 207 {
 208         struct sbuf sbuf;
 209         int domain, error, flind, ndomains;
 210
 211         error = sysctl_wire_old_buffer(req, 0);
 212         if (error != 0)
 213                 return (error);
 214         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 215         ndomains = vm_nfreelists - VM_NFREELIST + 1;
 216         for (domain = 0; domain < ndomains; domain++) {
 217                 sbuf_printf(&sbuf, "\nDOMAIN %d:\n\n", domain);
 218                 for (flind = 0; flind < vm_nfreelists; flind++)
 219                         sbuf_printf(&sbuf, "  [%d]:\t%p\n", flind,
 220                             vm_phys_lookup_lists[domain][flind]);
 221         }
 222         error = sbuf_finish(&sbuf);
 223         sbuf_delete(&sbuf);
 224         return (error);
 225 }
 226 #endif
 227
 228 /*
 229  * Create a physical memory segment.
 230  */
 231 static void
 232 _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind, int domain)
 233 {
 234         struct vm_phys_seg *seg;
 235 #ifdef VM_PHYSSEG_SPARSE
 236         long pages;
 237         int segind;
 238
 239         pages = 0;
 240         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 241                 seg = &vm_phys_segs[segind];
 242                 pages += atop(seg->end - seg->start);
 243         }
 244 #endif
 245         KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX,
 246             ("vm_phys_create_seg: increase VM_PHYSSEG_MAX"));
 247         seg = &vm_phys_segs[vm_phys_nsegs++];
 248         seg->start = start;
 249         seg->end = end;
 250         seg->domain = domain;
 251 #ifdef VM_PHYSSEG_SPARSE
 252         seg->first_page = &vm_page_array[pages];
 253 #else
 254         seg->first_page = PHYS_TO_VM_PAGE(start);
 255 #endif
 256 #if VM_NDOMAIN > 1
 257         if (flind == VM_FREELIST_DEFAULT && domain != 0) {
 258                 flind = VM_NFREELIST + (domain - 1);
 259                 if (flind >= vm_nfreelists)
 260                         vm_nfreelists = flind + 1;
 261         }
 262 #endif
 263         seg->free_queues = &vm_phys_free_queues[flind];
 264 }
 265
 266 static void
 267 vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind)
 268 {
 269         int i;
 270
 271         if (mem_affinity == NULL) {
 272                 _vm_phys_create_seg(start, end, flind, 0);
 273                 return;
 274         }
 275
 276         for (i = 0;; i++) {
 277                 if (mem_affinity[i].end == 0)
 278                         panic("Reached end of affinity info");
 279                 if (mem_affinity[i].end <= start)
 280                         continue;
 281                 if (mem_affinity[i].start > start)
 282                         panic("No affinity info for start %jx",
 283                             (uintmax_t)start);
 284                 if (mem_affinity[i].end >= end) {
 285                         _vm_phys_create_seg(start, end, flind,
 286                             mem_affinity[i].domain);
 287                         break;
 288                 }
 289                 _vm_phys_create_seg(start, mem_affinity[i].end, flind,
 290                     mem_affinity[i].domain);
 291                 start = mem_affinity[i].end;
 292         }
 293 }
 294
 295 /*
 296  * Initialize the physical memory allocator.
 297  */
 298 void
 299 vm_phys_init(void)
 300 {
 301         struct vm_freelist *fl;
 302         int flind, i, oind, pind;
 303 #if VM_NDOMAIN > 1
 304         int ndomains, j;
 305 #endif
 306
 307         for (i = 0; phys_avail[i + 1] != 0; i += 2) {
 308 #ifdef  VM_FREELIST_ISADMA
 309                 if (phys_avail[i] < 16777216) {
 310                         if (phys_avail[i + 1] > 16777216) {
 311                                 vm_phys_create_seg(phys_avail[i], 16777216,
 312                                     VM_FREELIST_ISADMA);
 313                                 vm_phys_create_seg(16777216, phys_avail[i + 1],
 314                                     VM_FREELIST_DEFAULT);
 315                         } else {
 316                                 vm_phys_create_seg(phys_avail[i],
 317                                     phys_avail[i + 1], VM_FREELIST_ISADMA);
 318                         }
 319                         if (VM_FREELIST_ISADMA >= vm_nfreelists)
 320                                 vm_nfreelists = VM_FREELIST_ISADMA + 1;
 321                 } else
 322 #endif
 323 #ifdef  VM_FREELIST_HIGHMEM
 324                 if (phys_avail[i + 1] > VM_HIGHMEM_ADDRESS) {
 325                         if (phys_avail[i] < VM_HIGHMEM_ADDRESS) {
 326                                 vm_phys_create_seg(phys_avail[i],
 327                                     VM_HIGHMEM_ADDRESS, VM_FREELIST_DEFAULT);
 328                                 vm_phys_create_seg(VM_HIGHMEM_ADDRESS,
 329                                     phys_avail[i + 1], VM_FREELIST_HIGHMEM);
 330                         } else {
 331                                 vm_phys_create_seg(phys_avail[i],
 332                                     phys_avail[i + 1], VM_FREELIST_HIGHMEM);
 333                         }
 334                         if (VM_FREELIST_HIGHMEM >= vm_nfreelists)
 335                                 vm_nfreelists = VM_FREELIST_HIGHMEM + 1;
 336                 } else
 337 #endif
 338                 vm_phys_create_seg(phys_avail[i], phys_avail[i + 1],
 339                     VM_FREELIST_DEFAULT);
 340         }
 341         for (flind = 0; flind < vm_nfreelists; flind++) {
 342                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 343                         fl = vm_phys_free_queues[flind][pind];
 344                         for (oind = 0; oind < VM_NFREEORDER; oind++)
 345                                 TAILQ_INIT(&fl[oind].pl);
 346                 }
 347         }
 348 #if VM_NDOMAIN > 1
 349         /*
 350          * Build a free list lookup list for each domain.  All of the
 351          * memory domain lists are inserted at the VM_FREELIST_DEFAULT
 352          * index in a round-robin order starting with the current
 353          * domain.
 354          */
 355         ndomains = vm_nfreelists - VM_NFREELIST + 1;
 356         for (flind = 0; flind < VM_FREELIST_DEFAULT; flind++)
 357                 for (i = 0; i < ndomains; i++)
 358                         vm_phys_lookup_lists[i][flind] =
 359                             &vm_phys_free_queues[flind];
 360         for (i = 0; i < ndomains; i++)
 361                 for (j = 0; j < ndomains; j++) {
 362                         flind = (i + j) % ndomains;
 363                         if (flind == 0)
 364                                 flind = VM_FREELIST_DEFAULT;
 365                         else
 366                                 flind += VM_NFREELIST - 1;
 367                         vm_phys_lookup_lists[i][VM_FREELIST_DEFAULT + j] =
 368                             &vm_phys_free_queues[flind];
 369                 }
 370         for (flind = VM_FREELIST_DEFAULT + 1; flind < VM_NFREELIST;
 371              flind++)
 372                 for (i = 0; i < ndomains; i++)
 373                         vm_phys_lookup_lists[i][flind + ndomains - 1] =
 374                             &vm_phys_free_queues[flind];
 375 #else
 376         for (flind = 0; flind < vm_nfreelists; flind++)
 377                 vm_phys_lookup_lists[0][flind] = &vm_phys_free_queues[flind];
 378 #endif
 379
 380         mtx_init(&vm_phys_fictitious_reg_mtx, "vmfctr", NULL, MTX_DEF);
 381 }
 382
 383 /*
 384  * Split a contiguous, power of two-sized set of physical pages.
 385  */
 386 static __inline void
 387 vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order)
 388 {
 389         vm_page_t m_buddy;
 390
 391         while (oind > order) {
 392                 oind--;
 393                 m_buddy = &m[1 << oind];
 394                 KASSERT(m_buddy->order == VM_NFREEORDER,
 395                     ("vm_phys_split_pages: page %p has unexpected order %d",
 396                     m_buddy, m_buddy->order));
 397                 m_buddy->order = oind;
 398                 TAILQ_INSERT_HEAD(&fl[oind].pl, m_buddy, pageq);
 399                 fl[oind].lcnt++;
 400         }
 401 }
 402
 403 /*
 404  * Initialize a physical page and add it to the free lists.
 405  */
 406 void
 407 vm_phys_add_page(vm_paddr_t pa)
 408 {
 409         vm_page_t m;
 410
 411         cnt.v_page_count++;
 412         m = vm_phys_paddr_to_vm_page(pa);
 413         m->phys_addr = pa;
 414         m->queue = PQ_NONE;
 415         m->segind = vm_phys_paddr_to_segind(pa);
 416         m->flags = PG_FREE;
 417         KASSERT(m->order == VM_NFREEORDER,
 418             ("vm_phys_add_page: page %p has unexpected order %d",
 419             m, m->order));
 420         m->pool = VM_FREEPOOL_DEFAULT;
 421         pmap_page_init(m);
 422         mtx_lock(&vm_page_queue_free_mtx);
 423         cnt.v_free_count++;
 424         vm_phys_free_pages(m, 0);
 425         mtx_unlock(&vm_page_queue_free_mtx);
 426 }
 427
 428 /*
 429  * Allocate a contiguous, power of two-sized set of physical pages
 430  * from the free lists.
 431  *
 432  * The free page queues must be locked.
 433  */
 434 vm_page_t
 435 vm_phys_alloc_pages(int pool, int order)
 436 {
 437         vm_page_t m;
 438         int flind;
 439
 440         for (flind = 0; flind < vm_nfreelists; flind++) {
 441                 m = vm_phys_alloc_freelist_pages(flind, pool, order);
 442                 if (m != NULL)
 443                         return (m);
 444         }
 445         return (NULL);
 446 }
 447
 448 /*
 449  * Find and dequeue a free page on the given free list, with the
 450  * specified pool and order
 451  */
 452 vm_page_t
 453 vm_phys_alloc_freelist_pages(int flind, int pool, int order)
 454 {
 455         struct vm_freelist *fl;
 456         struct vm_freelist *alt;
 457         int domain, oind, pind;
 458         vm_page_t m;
 459
 460         KASSERT(flind < VM_NFREELIST,
 461             ("vm_phys_alloc_freelist_pages: freelist %d is out of range", flind));
 462         KASSERT(pool < VM_NFREEPOOL,
 463             ("vm_phys_alloc_freelist_pages: pool %d is out of range", pool));
 464         KASSERT(order < VM_NFREEORDER,
 465             ("vm_phys_alloc_freelist_pages: order %d is out of range", order));
 466
 467 #if VM_NDOMAIN > 1
 468         domain = PCPU_GET(domain);
 469 #else
 470         domain = 0;
 471 #endif
 472         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 473         fl = (*vm_phys_lookup_lists[domain][flind])[pool];
 474         for (oind = order; oind < VM_NFREEORDER; oind++) {
 475                 m = TAILQ_FIRST(&fl[oind].pl);
 476                 if (m != NULL) {
 477                         TAILQ_REMOVE(&fl[oind].pl, m, pageq);
 478                         fl[oind].lcnt--;
 479                         m->order = VM_NFREEORDER;
 480                         vm_phys_split_pages(m, oind, fl, order);
 481                         return (m);
 482                 }
 483         }
 484
 485         /*
 486          * The given pool was empty.  Find the largest
 487          * contiguous, power-of-two-sized set of pages in any
 488          * pool.  Transfer these pages to the given pool, and
 489          * use them to satisfy the allocation.
 490          */
 491         for (oind = VM_NFREEORDER - 1; oind >= order; oind--) {
 492                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 493                         alt = (*vm_phys_lookup_lists[domain][flind])[pind];
 494                         m = TAILQ_FIRST(&alt[oind].pl);
 495                         if (m != NULL) {
 496                                 TAILQ_REMOVE(&alt[oind].pl, m, pageq);
 497                                 alt[oind].lcnt--;
 498                                 m->order = VM_NFREEORDER;
 499                                 vm_phys_set_pool(pool, m, oind);
 500                                 vm_phys_split_pages(m, oind, fl, order);
 501                                 return (m);
 502                         }
 503                 }
 504         }
 505         return (NULL);
 506 }
 507
 508 /*
 509  * Find the vm_page corresponding to the given physical address.
 510  */
 511 vm_page_t
 512 vm_phys_paddr_to_vm_page(vm_paddr_t pa)
 513 {
 514         struct vm_phys_seg *seg;
 515         int segind;
 516
 517         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 518                 seg = &vm_phys_segs[segind];
 519                 if (pa >= seg->start && pa < seg->end)
 520                         return (&seg->first_page[atop(pa - seg->start)]);
 521         }
 522         return (NULL);
 523 }
 524
 525 vm_page_t
 526 vm_phys_fictitious_to_vm_page(vm_paddr_t pa)
 527 {
 528         struct vm_phys_fictitious_seg *seg;
 529         vm_page_t m;
 530         int segind;
 531
 532         m = NULL;
 533         for (segind = 0; segind < VM_PHYS_FICTITIOUS_NSEGS; segind++) {
 534                 seg = &vm_phys_fictitious_segs[segind];
 535                 if (pa >= seg->start && pa < seg->end) {
 536                         m = &seg->first_page[atop(pa - seg->start)];
 537                         KASSERT((m->flags & PG_FICTITIOUS) != 0,
 538                             ("%p not fictitious", m));
 539                         break;
 540                 }
 541         }
 542         return (m);
 543 }
 544
 545 int
 546 vm_phys_fictitious_reg_range(vm_paddr_t start, vm_paddr_t end,
 547     vm_memattr_t memattr)
 548 {
 549         struct vm_phys_fictitious_seg *seg;
 550         vm_page_t fp;
 551         long i, page_count;
 552         int segind;
 553 #ifdef VM_PHYSSEG_DENSE
 554         long pi;
 555         boolean_t malloced;
 556 #endif
 557
 558         page_count = (end - start) / PAGE_SIZE;
 559
 560 #ifdef VM_PHYSSEG_DENSE
 561         pi = atop(start);
 562         if (pi >= first_page && atop(end) < vm_page_array_size) {
 563                 fp = &vm_page_array[pi - first_page];
 564                 malloced = FALSE;
 565         } else
 566 #endif
 567         {
 568                 fp = malloc(page_count * sizeof(struct vm_page), M_FICT_PAGES,
 569                     M_WAITOK | M_ZERO);
 570 #ifdef VM_PHYSSEG_DENSE
 571                 malloced = TRUE;
 572 #endif
 573         }
 574         for (i = 0; i < page_count; i++) {
 575                 vm_page_initfake(&fp[i], start + PAGE_SIZE * i, memattr);
 576                 pmap_page_init(&fp[i]);
 577                 fp[i].oflags &= ~(VPO_BUSY | VPO_UNMANAGED);
 578         }
 579         mtx_lock(&vm_phys_fictitious_reg_mtx);
 580         for (segind = 0; segind < VM_PHYS_FICTITIOUS_NSEGS; segind++) {
 581                 seg = &vm_phys_fictitious_segs[segind];
 582                 if (seg->start == 0 && seg->end == 0) {
 583                         seg->start = start;
 584                         seg->end = end;
 585                         seg->first_page = fp;
 586                         mtx_unlock(&vm_phys_fictitious_reg_mtx);
 587                         return (0);
 588                 }
 589         }
 590         mtx_unlock(&vm_phys_fictitious_reg_mtx);
 591 #ifdef VM_PHYSSEG_DENSE
 592         if (malloced)
 593 #endif
 594                 free(fp, M_FICT_PAGES);
 595         return (EBUSY);
 596 }
 597
 598 void
 599 vm_phys_fictitious_unreg_range(vm_paddr_t start, vm_paddr_t end)
 600 {
 601         struct vm_phys_fictitious_seg *seg;
 602         vm_page_t fp;
 603         int segind;
 604 #ifdef VM_PHYSSEG_DENSE
 605         long pi;
 606 #endif
 607
 608 #ifdef VM_PHYSSEG_DENSE
 609         pi = atop(start);
 610 #endif
 611
 612         mtx_lock(&vm_phys_fictitious_reg_mtx);
 613         for (segind = 0; segind < VM_PHYS_FICTITIOUS_NSEGS; segind++) {
 614                 seg = &vm_phys_fictitious_segs[segind];
 615                 if (seg->start == start && seg->end == end) {
 616                         seg->start = seg->end = 0;
 617                         fp = seg->first_page;
 618                         seg->first_page = NULL;
 619                         mtx_unlock(&vm_phys_fictitious_reg_mtx);
 620 #ifdef VM_PHYSSEG_DENSE
 621                         if (pi < first_page || atop(end) >= vm_page_array_size)
 622 #endif
 623                                 free(fp, M_FICT_PAGES);
 624                         return;
 625                 }
 626         }
 627         mtx_unlock(&vm_phys_fictitious_reg_mtx);
 628         KASSERT(0, ("Unregistering not registered fictitious range"));
 629 }
 630
 631 /*
 632  * Find the segment containing the given physical address.
 633  */
 634 static int
 635 vm_phys_paddr_to_segind(vm_paddr_t pa)
 636 {
 637         struct vm_phys_seg *seg;
 638         int segind;
 639
 640         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 641                 seg = &vm_phys_segs[segind];
 642                 if (pa >= seg->start && pa < seg->end)
 643                         return (segind);
 644         }
 645         panic("vm_phys_paddr_to_segind: paddr %#jx is not in any segment" ,
 646             (uintmax_t)pa);
 647 }
 648
 649 /*
 650  * Free a contiguous, power of two-sized set of physical pages.
 651  *
 652  * The free page queues must be locked.
 653  */
 654 void
 655 vm_phys_free_pages(vm_page_t m, int order)
 656 {
 657         struct vm_freelist *fl;
 658         struct vm_phys_seg *seg;
 659         vm_paddr_t pa;
 660         vm_page_t m_buddy;
 661
 662         KASSERT(m->order == VM_NFREEORDER,
 663             ("vm_phys_free_pages: page %p has unexpected order %d",
 664             m, m->order));
 665         KASSERT(m->pool < VM_NFREEPOOL,
 666             ("vm_phys_free_pages: page %p has unexpected pool %d",
 667             m, m->pool));
 668         KASSERT(order < VM_NFREEORDER,
 669             ("vm_phys_free_pages: order %d is out of range", order));
 670         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 671         seg = &vm_phys_segs[m->segind];
 672         if (order < VM_NFREEORDER - 1) {
 673                 pa = VM_PAGE_TO_PHYS(m);
 674                 do {
 675                         pa ^= ((vm_paddr_t)1 << (PAGE_SHIFT + order));
 676                         if (pa < seg->start || pa >= seg->end)
 677                                 break;
 678                         m_buddy = &seg->first_page[atop(pa - seg->start)];
 679                         if (m_buddy->order != order)
 680                                 break;
 681                         fl = (*seg->free_queues)[m_buddy->pool];
 682                         TAILQ_REMOVE(&fl[order].pl, m_buddy, pageq);
 683                         fl[order].lcnt--;
 684                         m_buddy->order = VM_NFREEORDER;
 685                         if (m_buddy->pool != m->pool)
 686                                 vm_phys_set_pool(m->pool, m_buddy, order);
 687                         order++;
 688                         pa &= ~(((vm_paddr_t)1 << (PAGE_SHIFT + order)) - 1);
 689                         m = &seg->first_page[atop(pa - seg->start)];
 690                 } while (order < VM_NFREEORDER - 1);
 691         }
 692         m->order = order;
 693         fl = (*seg->free_queues)[m->pool];
 694         TAILQ_INSERT_TAIL(&fl[order].pl, m, pageq);
 695         fl[order].lcnt++;
 696 }
 697
 698 /*
 699  * Free a contiguous, arbitrarily sized set of physical pages.
 700  *
 701  * The free page queues must be locked.
 702  */
 703 void
 704 vm_phys_free_contig(vm_page_t m, u_long npages)
 705 {
 706         u_int n;
 707         int order;
 708
 709         /*
 710          * Avoid unnecessary coalescing by freeing the pages in the largest
 711          * possible power-of-two-sized subsets.
 712          */
 713         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 714         for (;; npages -= n) {
 715                 /*
 716                  * Unsigned "min" is used here so that "order" is assigned
 717                  * "VM_NFREEORDER - 1" when "m"'s physical address is zero
 718                  * or the low-order bits of its physical address are zero
 719                  * because the size of a physical address exceeds the size of
 720                  * a long.
 721                  */
 722                 order = min(ffsl(VM_PAGE_TO_PHYS(m) >> PAGE_SHIFT) - 1,
 723                     VM_NFREEORDER - 1);
 724                 n = 1 << order;
 725                 if (npages < n)
 726                         break;
 727                 vm_phys_free_pages(m, order);
 728                 m += n;
 729         }
 730         /* The residual "npages" is less than "1 << (VM_NFREEORDER - 1)". */
 731         for (; npages > 0; npages -= n) {
 732                 order = flsl(npages) - 1;
 733                 n = 1 << order;
 734                 vm_phys_free_pages(m, order);
 735                 m += n;
 736         }
 737 }
 738
 739 /*
 740  * Set the pool for a contiguous, power of two-sized set of physical pages.
 741  */
 742 void
 743 vm_phys_set_pool(int pool, vm_page_t m, int order)
 744 {
 745         vm_page_t m_tmp;
 746
 747         for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++)
 748                 m_tmp->pool = pool;
 749 }
 750
 751 /*
 752  * Search for the given physical page "m" in the free lists.  If the search
 753  * succeeds, remove "m" from the free lists and return TRUE.  Otherwise, return
 754  * FALSE, indicating that "m" is not in the free lists.
 755  *
 756  * The free page queues must be locked.
 757  */
 758 boolean_t
 759 vm_phys_unfree_page(vm_page_t m)
 760 {
 761         struct vm_freelist *fl;
 762         struct vm_phys_seg *seg;
 763         vm_paddr_t pa, pa_half;
 764         vm_page_t m_set, m_tmp;
 765         int order;
 766
 767         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 768
 769         /*
 770          * First, find the contiguous, power of two-sized set of free
 771          * physical pages containing the given physical page "m" and
 772          * assign it to "m_set".
 773          */
 774         seg = &vm_phys_segs[m->segind];
 775         for (m_set = m, order = 0; m_set->order == VM_NFREEORDER &&
 776             order < VM_NFREEORDER - 1; ) {
 777                 order++;
 778                 pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order));
 779                 if (pa >= seg->start)
 780                         m_set = &seg->first_page[atop(pa - seg->start)];
 781                 else
 782                         return (FALSE);
 783         }
 784         if (m_set->order < order)
 785                 return (FALSE);
 786         if (m_set->order == VM_NFREEORDER)
 787                 return (FALSE);
 788         KASSERT(m_set->order < VM_NFREEORDER,
 789             ("vm_phys_unfree_page: page %p has unexpected order %d",
 790             m_set, m_set->order));
 791
 792         /*
 793          * Next, remove "m_set" from the free lists.  Finally, extract
 794          * "m" from "m_set" using an iterative algorithm: While "m_set"
 795          * is larger than a page, shrink "m_set" by returning the half
 796          * of "m_set" that does not contain "m" to the free lists.
 797          */
 798         fl = (*seg->free_queues)[m_set->pool];
 799         order = m_set->order;
 800         TAILQ_REMOVE(&fl[order].pl, m_set, pageq);
 801         fl[order].lcnt--;
 802         m_set->order = VM_NFREEORDER;
 803         while (order > 0) {
 804                 order--;
 805                 pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order));
 806                 if (m->phys_addr < pa_half)
 807                         m_tmp = &seg->first_page[atop(pa_half - seg->start)];
 808                 else {
 809                         m_tmp = m_set;
 810                         m_set = &seg->first_page[atop(pa_half - seg->start)];
 811                 }
 812                 m_tmp->order = order;
 813                 TAILQ_INSERT_HEAD(&fl[order].pl, m_tmp, pageq);
 814                 fl[order].lcnt++;
 815         }
 816         KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency"));
 817         return (TRUE);
 818 }
 819
 820 /*
 821  * Try to zero one physical page.  Used by an idle priority thread.
 822  */
 823 boolean_t
 824 vm_phys_zero_pages_idle(void)
 825 {
 826         static struct vm_freelist *fl = vm_phys_free_queues[0][0];
 827         static int flind, oind, pind;
 828         vm_page_t m, m_tmp;
 829
 830         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 831         for (;;) {
 832                 TAILQ_FOREACH_REVERSE(m, &fl[oind].pl, pglist, pageq) {
 833                         for (m_tmp = m; m_tmp < &m[1 << oind]; m_tmp++) {
 834                                 if ((m_tmp->flags & (PG_CACHED | PG_ZERO)) == 0) {
 835                                         vm_phys_unfree_page(m_tmp);
 836                                         cnt.v_free_count--;
 837                                         mtx_unlock(&vm_page_queue_free_mtx);
 838                                         pmap_zero_page_idle(m_tmp);
 839                                         m_tmp->flags |= PG_ZERO;
 840                                         mtx_lock(&vm_page_queue_free_mtx);
 841                                         cnt.v_free_count++;
 842                                         vm_phys_free_pages(m_tmp, 0);
 843                                         vm_page_zero_count++;
 844                                         cnt_prezero++;
 845                                         return (TRUE);
 846                                 }
 847                         }
 848                 }
 849                 oind++;
 850                 if (oind == VM_NFREEORDER) {
 851                         oind = 0;
 852                         pind++;
 853                         if (pind == VM_NFREEPOOL) {
 854                                 pind = 0;
 855                                 flind++;
 856                                 if (flind == vm_nfreelists)
 857                                         flind = 0;
 858                         }
 859                         fl = vm_phys_free_queues[flind][pind];
 860                 }
 861         }
 862 }
 863
 864 /*
 865  * Allocate a contiguous set of physical pages of the given size
 866  * "npages" from the free lists.  All of the physical pages must be at
 867  * or above the given physical address "low" and below the given
 868  * physical address "high".  The given value "alignment" determines the
 869  * alignment of the first physical page in the set.  If the given value
 870  * "boundary" is non-zero, then the set of physical pages cannot cross
 871  * any physical address boundary that is a multiple of that value.  Both
 872  * "alignment" and "boundary" must be a power of two.
 873  */
 874 vm_page_t
 875 vm_phys_alloc_contig(u_long npages, vm_paddr_t low, vm_paddr_t high,
 876     u_long alignment, vm_paddr_t boundary)
 877 {
 878         struct vm_freelist *fl;
 879         struct vm_phys_seg *seg;
 880         vm_paddr_t pa, pa_last, size;
 881         vm_page_t m, m_ret;
 882         u_long npages_end;
 883         int domain, flind, oind, order, pind;
 884
 885         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 886 #if VM_NDOMAIN > 1
 887         domain = PCPU_GET(domain);
 888 #else
 889         domain = 0;
 890 #endif
 891         size = npages << PAGE_SHIFT;
 892         KASSERT(size != 0,
 893             ("vm_phys_alloc_contig: size must not be 0"));
 894         KASSERT((alignment & (alignment - 1)) == 0,
 895             ("vm_phys_alloc_contig: alignment must be a power of 2"));
 896         KASSERT((boundary & (boundary - 1)) == 0,
 897             ("vm_phys_alloc_contig: boundary must be a power of 2"));
 898         /* Compute the queue that is the best fit for npages. */
 899         for (order = 0; (1 << order) < npages; order++);
 900         for (flind = 0; flind < vm_nfreelists; flind++) {
 901                 for (oind = min(order, VM_NFREEORDER - 1); oind < VM_NFREEORDER; oind++) {
 902                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 903                                 fl = (*vm_phys_lookup_lists[domain][flind])
 904                                     [pind];
 905                                 TAILQ_FOREACH(m_ret, &fl[oind].pl, pageq) {
 906                                         /*
 907                                          * A free list may contain physical pages
 908                                          * from one or more segments.
 909                                          */
 910                                         seg = &vm_phys_segs[m_ret->segind];
 911                                         if (seg->start > high ||
 912                                             low >= seg->end)
 913                                                 continue;
 914
 915                                         /*
 916                                          * Is the size of this allocation request
 917                                          * larger than the largest block size?
 918                                          */
 919                                         if (order >= VM_NFREEORDER) {
 920                                                 /*
 921                                                  * Determine if a sufficient number
 922                                                  * of subsequent blocks to satisfy
 923                                                  * the allocation request are free.
 924                                                  */
 925                                                 pa = VM_PAGE_TO_PHYS(m_ret);
 926                                                 pa_last = pa + size;
 927                                                 for (;;) {
 928                                                         pa += 1 << (PAGE_SHIFT + VM_NFREEORDER - 1);
 929                                                         if (pa >= pa_last)
 930                                                                 break;
 931                                                         if (pa < seg->start ||
 932                                                             pa >= seg->end)
 933                                                                 break;
 934                                                         m = &seg->first_page[atop(pa - seg->start)];
 935                                                         if (m->order != VM_NFREEORDER - 1)
 936                                                                 break;
 937                                                 }
 938                                                 /* If not, continue to the next block. */
 939                                                 if (pa < pa_last)
 940                                                         continue;
 941                                         }
 942
 943                                         /*
 944                                          * Determine if the blocks are within the given range,
 945                                          * satisfy the given alignment, and do not cross the
 946                                          * given boundary.
 947                                          */
 948                                         pa = VM_PAGE_TO_PHYS(m_ret);
 949                                         if (pa >= low &&
 950                                             pa + size <= high &&
 951                                             (pa & (alignment - 1)) == 0 &&
 952                                             ((pa ^ (pa + size - 1)) & ~(boundary - 1)) == 0)
 953                                                 goto done;
 954                                 }
 955                         }
 956                 }
 957         }
 958         return (NULL);
 959 done:
 960         for (m = m_ret; m < &m_ret[npages]; m = &m[1 << oind]) {
 961                 fl = (*seg->free_queues)[m->pool];
 962                 TAILQ_REMOVE(&fl[m->order].pl, m, pageq);
 963                 fl[m->order].lcnt--;
 964                 m->order = VM_NFREEORDER;
 965         }
 966         if (m_ret->pool != VM_FREEPOOL_DEFAULT)
 967                 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m_ret, oind);
 968         fl = (*seg->free_queues)[m_ret->pool];
 969         vm_phys_split_pages(m_ret, oind, fl, order);
 970         /* Return excess pages to the free lists. */
 971         npages_end = roundup2(npages, 1 << imin(oind, order));
 972         if (npages < npages_end)
 973                 vm_phys_free_contig(&m_ret[npages], npages_end - npages);
 974         return (m_ret);
 975 }
 976
 977 #ifdef DDB
 978 /*
 979  * Show the number of physical pages in each of the free lists.
 980  */
 981 DB_SHOW_COMMAND(freepages, db_show_freepages)
 982 {
 983         struct vm_freelist *fl;
 984         int flind, oind, pind;
 985
 986         for (flind = 0; flind < vm_nfreelists; flind++) {
 987                 db_printf("FREE LIST %d:\n"
 988                     "\n  ORDER (SIZE)  |  NUMBER"
 989                     "\n              ", flind);
 990                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 991                         db_printf("  |  POOL %d", pind);
 992                 db_printf("\n--            ");
 993                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 994                         db_printf("-- --      ");
 995                 db_printf("--\n");
 996                 for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
 997                         db_printf("  %2.2d (%6.6dK)", oind,
 998                             1 << (PAGE_SHIFT - 10 + oind));
 999                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1000                                 fl = vm_phys_free_queues[flind][pind];
1001                                 db_printf("  |  %6.6d", fl[oind].lcnt);
1002                         }
1003                         db_printf("\n");
1004                 }
1005                 db_printf("\n");
1006         }
1007 }
1008 #endif