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