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, "vm_fictitious", "Fictitious VM 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 vm_page_t vm_phys_alloc_domain_pages(int domain, int flind, int pool,
 125     int order);
 126 static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind,
 127     int domain);
 128 static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind);
 129 static int vm_phys_paddr_to_segind(vm_paddr_t pa);
 130 static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl,
 131     int order);
 132
 133 /*
 134  * Outputs the state of the physical memory allocator, specifically,
 135  * the amount of physical memory in each free list.
 136  */
 137 static int
 138 sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS)
 139 {
 140         struct sbuf sbuf;
 141         struct vm_freelist *fl;
 142         int error, flind, oind, pind;
 143
 144         error = sysctl_wire_old_buffer(req, 0);
 145         if (error != 0)
 146                 return (error);
 147         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 148         for (flind = 0; flind < vm_nfreelists; flind++) {
 149                 sbuf_printf(&sbuf, "\nFREE LIST %d:\n"
 150                     "\n  ORDER (SIZE)  |  NUMBER"
 151                     "\n              ", flind);
 152                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 153                         sbuf_printf(&sbuf, "  |  POOL %d", pind);
 154                 sbuf_printf(&sbuf, "\n--            ");
 155                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 156                         sbuf_printf(&sbuf, "-- --      ");
 157                 sbuf_printf(&sbuf, "--\n");
 158                 for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
 159                         sbuf_printf(&sbuf, "  %2d (%6dK)", oind,
 160                             1 << (PAGE_SHIFT - 10 + oind));
 161                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 162                                 fl = vm_phys_free_queues[flind][pind];
 163                                 sbuf_printf(&sbuf, "  |  %6d", fl[oind].lcnt);
 164                         }
 165                         sbuf_printf(&sbuf, "\n");
 166                 }
 167         }
 168         error = sbuf_finish(&sbuf);
 169         sbuf_delete(&sbuf);
 170         return (error);
 171 }
 172
 173 /*
 174  * Outputs the set of physical memory segments.
 175  */
 176 static int
 177 sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS)
 178 {
 179         struct sbuf sbuf;
 180         struct vm_phys_seg *seg;
 181         int error, segind;
 182
 183         error = sysctl_wire_old_buffer(req, 0);
 184         if (error != 0)
 185                 return (error);
 186         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 187         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 188                 sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind);
 189                 seg = &vm_phys_segs[segind];
 190                 sbuf_printf(&sbuf, "start:     %#jx\n",
 191                     (uintmax_t)seg->start);
 192                 sbuf_printf(&sbuf, "end:       %#jx\n",
 193                     (uintmax_t)seg->end);
 194                 sbuf_printf(&sbuf, "domain:    %d\n", seg->domain);
 195                 sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues);
 196         }
 197         error = sbuf_finish(&sbuf);
 198         sbuf_delete(&sbuf);
 199         return (error);
 200 }
 201
 202 #if VM_NDOMAIN > 1
 203 /*
 204  * Outputs the set of free list lookup lists.
 205  */
 206 static int
 207 sysctl_vm_phys_lookup_lists(SYSCTL_HANDLER_ARGS)
 208 {
 209         struct sbuf sbuf;
 210         int domain, error, flind, ndomains;
 211
 212         error = sysctl_wire_old_buffer(req, 0);
 213         if (error != 0)
 214                 return (error);
 215         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 216         ndomains = vm_nfreelists - VM_NFREELIST + 1;
 217         for (domain = 0; domain < ndomains; domain++) {
 218                 sbuf_printf(&sbuf, "\nDOMAIN %d:\n\n", domain);
 219                 for (flind = 0; flind < vm_nfreelists; flind++)
 220                         sbuf_printf(&sbuf, "  [%d]:\t%p\n", flind,
 221                             vm_phys_lookup_lists[domain][flind]);
 222         }
 223         error = sbuf_finish(&sbuf);
 224         sbuf_delete(&sbuf);
 225         return (error);
 226 }
 227 #endif
 228
 229 /*
 230  * Create a physical memory segment.
 231  */
 232 static void
 233 _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind, int domain)
 234 {
 235         struct vm_phys_seg *seg;
 236 #ifdef VM_PHYSSEG_SPARSE
 237         long pages;
 238         int segind;
 239
 240         pages = 0;
 241         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 242                 seg = &vm_phys_segs[segind];
 243                 pages += atop(seg->end - seg->start);
 244         }
 245 #endif
 246         KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX,
 247             ("vm_phys_create_seg: increase VM_PHYSSEG_MAX"));
 248         seg = &vm_phys_segs[vm_phys_nsegs++];
 249         seg->start = start;
 250         seg->end = end;
 251         seg->domain = domain;
 252 #ifdef VM_PHYSSEG_SPARSE
 253         seg->first_page = &vm_page_array[pages];
 254 #else
 255         seg->first_page = PHYS_TO_VM_PAGE(start);
 256 #endif
 257 #if VM_NDOMAIN > 1
 258         if (flind == VM_FREELIST_DEFAULT && domain != 0) {
 259                 flind = VM_NFREELIST + (domain - 1);
 260                 if (flind >= vm_nfreelists)
 261                         vm_nfreelists = flind + 1;
 262         }
 263 #endif
 264         seg->free_queues = &vm_phys_free_queues[flind];
 265 }
 266
 267 static void
 268 vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind)
 269 {
 270         int i;
 271
 272         if (mem_affinity == NULL) {
 273                 _vm_phys_create_seg(start, end, flind, 0);
 274                 return;
 275         }
 276
 277         for (i = 0;; i++) {
 278                 if (mem_affinity[i].end == 0)
 279                         panic("Reached end of affinity info");
 280                 if (mem_affinity[i].end <= start)
 281                         continue;
 282                 if (mem_affinity[i].start > start)
 283                         panic("No affinity info for start %jx",
 284                             (uintmax_t)start);
 285                 if (mem_affinity[i].end >= end) {
 286                         _vm_phys_create_seg(start, end, flind,
 287                             mem_affinity[i].domain);
 288                         break;
 289                 }
 290                 _vm_phys_create_seg(start, mem_affinity[i].end, flind,
 291                     mem_affinity[i].domain);
 292                 start = mem_affinity[i].end;
 293         }
 294 }
 295
 296 /*
 297  * Initialize the physical memory allocator.
 298  */
 299 void
 300 vm_phys_init(void)
 301 {
 302         struct vm_freelist *fl;
 303         int flind, i, oind, pind;
 304 #if VM_NDOMAIN > 1
 305         int ndomains, j;
 306 #endif
 307
 308         for (i = 0; phys_avail[i + 1] != 0; i += 2) {
 309 #ifdef  VM_FREELIST_ISADMA
 310                 if (phys_avail[i] < 16777216) {
 311                         if (phys_avail[i + 1] > 16777216) {
 312                                 vm_phys_create_seg(phys_avail[i], 16777216,
 313                                     VM_FREELIST_ISADMA);
 314                                 vm_phys_create_seg(16777216, phys_avail[i + 1],
 315                                     VM_FREELIST_DEFAULT);
 316                         } else {
 317                                 vm_phys_create_seg(phys_avail[i],
 318                                     phys_avail[i + 1], VM_FREELIST_ISADMA);
 319                         }
 320                         if (VM_FREELIST_ISADMA >= vm_nfreelists)
 321                                 vm_nfreelists = VM_FREELIST_ISADMA + 1;
 322                 } else
 323 #endif
 324 #ifdef  VM_FREELIST_HIGHMEM
 325                 if (phys_avail[i + 1] > VM_HIGHMEM_ADDRESS) {
 326                         if (phys_avail[i] < VM_HIGHMEM_ADDRESS) {
 327                                 vm_phys_create_seg(phys_avail[i],
 328                                     VM_HIGHMEM_ADDRESS, VM_FREELIST_DEFAULT);
 329                                 vm_phys_create_seg(VM_HIGHMEM_ADDRESS,
 330                                     phys_avail[i + 1], VM_FREELIST_HIGHMEM);
 331                         } else {
 332                                 vm_phys_create_seg(phys_avail[i],
 333                                     phys_avail[i + 1], VM_FREELIST_HIGHMEM);
 334                         }
 335                         if (VM_FREELIST_HIGHMEM >= vm_nfreelists)
 336                                 vm_nfreelists = VM_FREELIST_HIGHMEM + 1;
 337                 } else
 338 #endif
 339                 vm_phys_create_seg(phys_avail[i], phys_avail[i + 1],
 340                     VM_FREELIST_DEFAULT);
 341         }
 342         for (flind = 0; flind < vm_nfreelists; flind++) {
 343                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 344                         fl = vm_phys_free_queues[flind][pind];
 345                         for (oind = 0; oind < VM_NFREEORDER; oind++)
 346                                 TAILQ_INIT(&fl[oind].pl);
 347                 }
 348         }
 349 #if VM_NDOMAIN > 1
 350         /*
 351          * Build a free list lookup list for each domain.  All of the
 352          * memory domain lists are inserted at the VM_FREELIST_DEFAULT
 353          * index in a round-robin order starting with the current
 354          * domain.
 355          */
 356         ndomains = vm_nfreelists - VM_NFREELIST + 1;
 357         for (flind = 0; flind < VM_FREELIST_DEFAULT; flind++)
 358                 for (i = 0; i < ndomains; i++)
 359                         vm_phys_lookup_lists[i][flind] =
 360                             &vm_phys_free_queues[flind];
 361         for (i = 0; i < ndomains; i++)
 362                 for (j = 0; j < ndomains; j++) {
 363                         flind = (i + j) % ndomains;
 364                         if (flind == 0)
 365                                 flind = VM_FREELIST_DEFAULT;
 366                         else
 367                                 flind += VM_NFREELIST - 1;
 368                         vm_phys_lookup_lists[i][VM_FREELIST_DEFAULT + j] =
 369                             &vm_phys_free_queues[flind];
 370                 }
 371         for (flind = VM_FREELIST_DEFAULT + 1; flind < VM_NFREELIST;
 372              flind++)
 373                 for (i = 0; i < ndomains; i++)
 374                         vm_phys_lookup_lists[i][flind + ndomains - 1] =
 375                             &vm_phys_free_queues[flind];
 376 #else
 377         for (flind = 0; flind < vm_nfreelists; flind++)
 378                 vm_phys_lookup_lists[0][flind] = &vm_phys_free_queues[flind];
 379 #endif
 380
 381         mtx_init(&vm_phys_fictitious_reg_mtx, "vmfctr", NULL, MTX_DEF);
 382 }
 383
 384 /*
 385  * Split a contiguous, power of two-sized set of physical pages.
 386  */
 387 static __inline void
 388 vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order)
 389 {
 390         vm_page_t m_buddy;
 391
 392         while (oind > order) {
 393                 oind--;
 394                 m_buddy = &m[1 << oind];
 395                 KASSERT(m_buddy->order == VM_NFREEORDER,
 396                     ("vm_phys_split_pages: page %p has unexpected order %d",
 397                     m_buddy, m_buddy->order));
 398                 m_buddy->order = oind;
 399                 TAILQ_INSERT_HEAD(&fl[oind].pl, m_buddy, pageq);
 400                 fl[oind].lcnt++;
 401         }
 402 }
 403
 404 /*
 405  * Initialize a physical page and add it to the free lists.
 406  */
 407 void
 408 vm_phys_add_page(vm_paddr_t pa)
 409 {
 410         vm_page_t m;
 411
 412         cnt.v_page_count++;
 413         m = vm_phys_paddr_to_vm_page(pa);
 414         m->phys_addr = pa;
 415         m->queue = PQ_NONE;
 416         m->segind = vm_phys_paddr_to_segind(pa);
 417         m->flags = PG_FREE;
 418         KASSERT(m->order == VM_NFREEORDER,
 419             ("vm_phys_add_page: page %p has unexpected order %d",
 420             m, m->order));
 421         m->pool = VM_FREEPOOL_DEFAULT;
 422         pmap_page_init(m);
 423         mtx_lock(&vm_page_queue_free_mtx);
 424         cnt.v_free_count++;
 425         vm_phys_free_pages(m, 0);
 426         mtx_unlock(&vm_page_queue_free_mtx);
 427 }
 428
 429 /*
 430  * Allocate a contiguous, power of two-sized set of physical pages
 431  * from the free lists.
 432  *
 433  * The free page queues must be locked.
 434  */
 435 vm_page_t
 436 vm_phys_alloc_pages(int pool, int order)
 437 {
 438         vm_page_t m;
 439         int domain, flind;
 440
 441         KASSERT(pool < VM_NFREEPOOL,
 442             ("vm_phys_alloc_pages: pool %d is out of range", pool));
 443         KASSERT(order < VM_NFREEORDER,
 444             ("vm_phys_alloc_pages: order %d is out of range", order));
 445
 446 #if VM_NDOMAIN > 1
 447         domain = PCPU_GET(domain);
 448 #else
 449         domain = 0;
 450 #endif
 451         for (flind = 0; flind < vm_nfreelists; flind++) {
 452                 m = vm_phys_alloc_domain_pages(domain, flind, pool, order);
 453                 if (m != NULL)
 454                         return (m);
 455         }
 456         return (NULL);
 457 }
 458
 459 /*
 460  * Find and dequeue a free page on the given free list, with the
 461  * specified pool and order
 462  */
 463 vm_page_t
 464 vm_phys_alloc_freelist_pages(int flind, int pool, int order)
 465 {
 466 #if VM_NDOMAIN > 1
 467         vm_page_t m;
 468         int i, ndomains;
 469 #endif
 470         int domain;
 471
 472         KASSERT(flind < VM_NFREELIST,
 473             ("vm_phys_alloc_freelist_pages: freelist %d is out of range", flind));
 474         KASSERT(pool < VM_NFREEPOOL,
 475             ("vm_phys_alloc_freelist_pages: pool %d is out of range", pool));
 476         KASSERT(order < VM_NFREEORDER,
 477             ("vm_phys_alloc_freelist_pages: order %d is out of range", order));
 478
 479 #if VM_NDOMAIN > 1
 480         /*
 481          * This routine expects to be called with a VM_FREELIST_* constant.
 482          * On a system with multiple domains we need to adjust the flind
 483          * appropriately.  If it is for VM_FREELIST_DEFAULT we need to
 484          * iterate over the per-domain lists.
 485          */
 486         domain = PCPU_GET(domain);
 487         ndomains = vm_nfreelists - VM_NFREELIST + 1;
 488         if (flind == VM_FREELIST_DEFAULT) {
 489                 m = NULL;
 490                 for (i = 0; i < ndomains; i++, flind++) {
 491                         m = vm_phys_alloc_domain_pages(domain, flind, pool,
 492                             order);
 493                         if (m != NULL)
 494                                 break;
 495                 }
 496                 return (m);
 497         } else if (flind > VM_FREELIST_DEFAULT)
 498                 flind += ndomains - 1;
 499 #else
 500         domain = 0;
 501 #endif
 502         return (vm_phys_alloc_domain_pages(domain, flind, pool, order));
 503 }
 504
 505 static vm_page_t
 506 vm_phys_alloc_domain_pages(int domain, int flind, int pool, int order)
 507 {
 508         struct vm_freelist *fl;
 509         struct vm_freelist *alt;
 510         int oind, pind;
 511         vm_page_t m;
 512
 513         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 514         fl = (*vm_phys_lookup_lists[domain][flind])[pool];
 515         for (oind = order; oind < VM_NFREEORDER; oind++) {
 516                 m = TAILQ_FIRST(&fl[oind].pl);
 517                 if (m != NULL) {
 518                         TAILQ_REMOVE(&fl[oind].pl, m, pageq);
 519                         fl[oind].lcnt--;
 520                         m->order = VM_NFREEORDER;
 521                         vm_phys_split_pages(m, oind, fl, order);
 522                         return (m);
 523                 }
 524         }
 525
 526         /*
 527          * The given pool was empty.  Find the largest
 528          * contiguous, power-of-two-sized set of pages in any
 529          * pool.  Transfer these pages to the given pool, and
 530          * use them to satisfy the allocation.
 531          */
 532         for (oind = VM_NFREEORDER - 1; oind >= order; oind--) {
 533                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 534                         alt = (*vm_phys_lookup_lists[domain][flind])[pind];
 535                         m = TAILQ_FIRST(&alt[oind].pl);
 536                         if (m != NULL) {
 537                                 TAILQ_REMOVE(&alt[oind].pl, m, pageq);
 538                                 alt[oind].lcnt--;
 539                                 m->order = VM_NFREEORDER;
 540                                 vm_phys_set_pool(pool, m, oind);
 541                                 vm_phys_split_pages(m, oind, fl, order);
 542                                 return (m);
 543                         }
 544                 }
 545         }
 546         return (NULL);
 547 }
 548
 549 /*
 550  * Find the vm_page corresponding to the given physical address.
 551  */
 552 vm_page_t
 553 vm_phys_paddr_to_vm_page(vm_paddr_t pa)
 554 {
 555         struct vm_phys_seg *seg;
 556         int segind;
 557
 558         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 559                 seg = &vm_phys_segs[segind];
 560                 if (pa >= seg->start && pa < seg->end)
 561                         return (&seg->first_page[atop(pa - seg->start)]);
 562         }
 563         return (NULL);
 564 }
 565
 566 vm_page_t
 567 vm_phys_fictitious_to_vm_page(vm_paddr_t pa)
 568 {
 569         struct vm_phys_fictitious_seg *seg;
 570         vm_page_t m;
 571         int segind;
 572
 573         m = NULL;
 574         for (segind = 0; segind < VM_PHYS_FICTITIOUS_NSEGS; segind++) {
 575                 seg = &vm_phys_fictitious_segs[segind];
 576                 if (pa >= seg->start && pa < seg->end) {
 577                         m = &seg->first_page[atop(pa - seg->start)];
 578                         KASSERT((m->flags & PG_FICTITIOUS) != 0,
 579                             ("%p not fictitious", m));
 580                         break;
 581                 }
 582         }
 583         return (m);
 584 }
 585
 586 int
 587 vm_phys_fictitious_reg_range(vm_paddr_t start, vm_paddr_t end,
 588     vm_memattr_t memattr)
 589 {
 590         struct vm_phys_fictitious_seg *seg;
 591         vm_page_t fp;
 592         long i, page_count;
 593         int segind;
 594 #ifdef VM_PHYSSEG_DENSE
 595         long pi;
 596         boolean_t malloced;
 597 #endif
 598
 599         page_count = (end - start) / PAGE_SIZE;
 600
 601 #ifdef VM_PHYSSEG_DENSE
 602         pi = atop(start);
 603         if (pi >= first_page && pi < vm_page_array_size + first_page) {
 604                 if (atop(end) >= vm_page_array_size + first_page)
 605                         return (EINVAL);
 606                 fp = &vm_page_array[pi - first_page];
 607                 malloced = FALSE;
 608         } else
 609 #endif
 610         {
 611                 fp = malloc(page_count * sizeof(struct vm_page), M_FICT_PAGES,
 612                     M_WAITOK | M_ZERO);
 613 #ifdef VM_PHYSSEG_DENSE
 614                 malloced = TRUE;
 615 #endif
 616         }
 617         for (i = 0; i < page_count; i++) {
 618                 vm_page_initfake(&fp[i], start + PAGE_SIZE * i, memattr);
 619                 pmap_page_init(&fp[i]);
 620                 fp[i].oflags &= ~(VPO_BUSY | VPO_UNMANAGED);
 621         }
 622         mtx_lock(&vm_phys_fictitious_reg_mtx);
 623         for (segind = 0; segind < VM_PHYS_FICTITIOUS_NSEGS; segind++) {
 624                 seg = &vm_phys_fictitious_segs[segind];
 625                 if (seg->start == 0 && seg->end == 0) {
 626                         seg->start = start;
 627                         seg->end = end;
 628                         seg->first_page = fp;
 629                         mtx_unlock(&vm_phys_fictitious_reg_mtx);
 630                         return (0);
 631                 }
 632         }
 633         mtx_unlock(&vm_phys_fictitious_reg_mtx);
 634 #ifdef VM_PHYSSEG_DENSE
 635         if (malloced)
 636 #endif
 637                 free(fp, M_FICT_PAGES);
 638         return (EBUSY);
 639 }
 640
 641 void
 642 vm_phys_fictitious_unreg_range(vm_paddr_t start, vm_paddr_t end)
 643 {
 644         struct vm_phys_fictitious_seg *seg;
 645         vm_page_t fp;
 646         int segind;
 647 #ifdef VM_PHYSSEG_DENSE
 648         long pi;
 649 #endif
 650
 651 #ifdef VM_PHYSSEG_DENSE
 652         pi = atop(start);
 653 #endif
 654
 655         mtx_lock(&vm_phys_fictitious_reg_mtx);
 656         for (segind = 0; segind < VM_PHYS_FICTITIOUS_NSEGS; segind++) {
 657                 seg = &vm_phys_fictitious_segs[segind];
 658                 if (seg->start == start && seg->end == end) {
 659                         seg->start = seg->end = 0;
 660                         fp = seg->first_page;
 661                         seg->first_page = NULL;
 662                         mtx_unlock(&vm_phys_fictitious_reg_mtx);
 663 #ifdef VM_PHYSSEG_DENSE
 664                         if (pi < first_page || atop(end) >= vm_page_array_size)
 665 #endif
 666                                 free(fp, M_FICT_PAGES);
 667                         return;
 668                 }
 669         }
 670         mtx_unlock(&vm_phys_fictitious_reg_mtx);
 671         KASSERT(0, ("Unregistering not registered fictitious range"));
 672 }
 673
 674 /*
 675  * Find the segment containing the given physical address.
 676  */
 677 static int
 678 vm_phys_paddr_to_segind(vm_paddr_t pa)
 679 {
 680         struct vm_phys_seg *seg;
 681         int segind;
 682
 683         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 684                 seg = &vm_phys_segs[segind];
 685                 if (pa >= seg->start && pa < seg->end)
 686                         return (segind);
 687         }
 688         panic("vm_phys_paddr_to_segind: paddr %#jx is not in any segment" ,
 689             (uintmax_t)pa);
 690 }
 691
 692 /*
 693  * Free a contiguous, power of two-sized set of physical pages.
 694  *
 695  * The free page queues must be locked.
 696  */
 697 void
 698 vm_phys_free_pages(vm_page_t m, int order)
 699 {
 700         struct vm_freelist *fl;
 701         struct vm_phys_seg *seg;
 702         vm_paddr_t pa;
 703         vm_page_t m_buddy;
 704
 705         KASSERT(m->order == VM_NFREEORDER,
 706             ("vm_phys_free_pages: page %p has unexpected order %d",
 707             m, m->order));
 708         KASSERT(m->pool < VM_NFREEPOOL,
 709             ("vm_phys_free_pages: page %p has unexpected pool %d",
 710             m, m->pool));
 711         KASSERT(order < VM_NFREEORDER,
 712             ("vm_phys_free_pages: order %d is out of range", order));
 713         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 714         seg = &vm_phys_segs[m->segind];
 715         if (order < VM_NFREEORDER - 1) {
 716                 pa = VM_PAGE_TO_PHYS(m);
 717                 do {
 718                         pa ^= ((vm_paddr_t)1 << (PAGE_SHIFT + order));
 719                         if (pa < seg->start || pa >= seg->end)
 720                                 break;
 721                         m_buddy = &seg->first_page[atop(pa - seg->start)];
 722                         if (m_buddy->order != order)
 723                                 break;
 724                         fl = (*seg->free_queues)[m_buddy->pool];
 725                         TAILQ_REMOVE(&fl[order].pl, m_buddy, pageq);
 726                         fl[order].lcnt--;
 727                         m_buddy->order = VM_NFREEORDER;
 728                         if (m_buddy->pool != m->pool)
 729                                 vm_phys_set_pool(m->pool, m_buddy, order);
 730                         order++;
 731                         pa &= ~(((vm_paddr_t)1 << (PAGE_SHIFT + order)) - 1);
 732                         m = &seg->first_page[atop(pa - seg->start)];
 733                 } while (order < VM_NFREEORDER - 1);
 734         }
 735         m->order = order;
 736         fl = (*seg->free_queues)[m->pool];
 737         TAILQ_INSERT_TAIL(&fl[order].pl, m, pageq);
 738         fl[order].lcnt++;
 739 }
 740
 741 /*
 742  * Free a contiguous, arbitrarily sized set of physical pages.
 743  *
 744  * The free page queues must be locked.
 745  */
 746 void
 747 vm_phys_free_contig(vm_page_t m, u_long npages)
 748 {
 749         u_int n;
 750         int order;
 751
 752         /*
 753          * Avoid unnecessary coalescing by freeing the pages in the largest
 754          * possible power-of-two-sized subsets.
 755          */
 756         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 757         for (;; npages -= n) {
 758                 /*
 759                  * Unsigned "min" is used here so that "order" is assigned
 760                  * "VM_NFREEORDER - 1" when "m"'s physical address is zero
 761                  * or the low-order bits of its physical address are zero
 762                  * because the size of a physical address exceeds the size of
 763                  * a long.
 764                  */
 765                 order = min(ffsl(VM_PAGE_TO_PHYS(m) >> PAGE_SHIFT) - 1,
 766                     VM_NFREEORDER - 1);
 767                 n = 1 << order;
 768                 if (npages < n)
 769                         break;
 770                 vm_phys_free_pages(m, order);
 771                 m += n;
 772         }
 773         /* The residual "npages" is less than "1 << (VM_NFREEORDER - 1)". */
 774         for (; npages > 0; npages -= n) {
 775                 order = flsl(npages) - 1;
 776                 n = 1 << order;
 777                 vm_phys_free_pages(m, order);
 778                 m += n;
 779         }
 780 }
 781
 782 /*
 783  * Set the pool for a contiguous, power of two-sized set of physical pages.
 784  */
 785 void
 786 vm_phys_set_pool(int pool, vm_page_t m, int order)
 787 {
 788         vm_page_t m_tmp;
 789
 790         for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++)
 791                 m_tmp->pool = pool;
 792 }
 793
 794 /*
 795  * Search for the given physical page "m" in the free lists.  If the search
 796  * succeeds, remove "m" from the free lists and return TRUE.  Otherwise, return
 797  * FALSE, indicating that "m" is not in the free lists.
 798  *
 799  * The free page queues must be locked.
 800  */
 801 boolean_t
 802 vm_phys_unfree_page(vm_page_t m)
 803 {
 804         struct vm_freelist *fl;
 805         struct vm_phys_seg *seg;
 806         vm_paddr_t pa, pa_half;
 807         vm_page_t m_set, m_tmp;
 808         int order;
 809
 810         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 811
 812         /*
 813          * First, find the contiguous, power of two-sized set of free
 814          * physical pages containing the given physical page "m" and
 815          * assign it to "m_set".
 816          */
 817         seg = &vm_phys_segs[m->segind];
 818         for (m_set = m, order = 0; m_set->order == VM_NFREEORDER &&
 819             order < VM_NFREEORDER - 1; ) {
 820                 order++;
 821                 pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order));
 822                 if (pa >= seg->start)
 823                         m_set = &seg->first_page[atop(pa - seg->start)];
 824                 else
 825                         return (FALSE);
 826         }
 827         if (m_set->order < order)
 828                 return (FALSE);
 829         if (m_set->order == VM_NFREEORDER)
 830                 return (FALSE);
 831         KASSERT(m_set->order < VM_NFREEORDER,
 832             ("vm_phys_unfree_page: page %p has unexpected order %d",
 833             m_set, m_set->order));
 834
 835         /*
 836          * Next, remove "m_set" from the free lists.  Finally, extract
 837          * "m" from "m_set" using an iterative algorithm: While "m_set"
 838          * is larger than a page, shrink "m_set" by returning the half
 839          * of "m_set" that does not contain "m" to the free lists.
 840          */
 841         fl = (*seg->free_queues)[m_set->pool];
 842         order = m_set->order;
 843         TAILQ_REMOVE(&fl[order].pl, m_set, pageq);
 844         fl[order].lcnt--;
 845         m_set->order = VM_NFREEORDER;
 846         while (order > 0) {
 847                 order--;
 848                 pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order));
 849                 if (m->phys_addr < pa_half)
 850                         m_tmp = &seg->first_page[atop(pa_half - seg->start)];
 851                 else {
 852                         m_tmp = m_set;
 853                         m_set = &seg->first_page[atop(pa_half - seg->start)];
 854                 }
 855                 m_tmp->order = order;
 856                 TAILQ_INSERT_HEAD(&fl[order].pl, m_tmp, pageq);
 857                 fl[order].lcnt++;
 858         }
 859         KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency"));
 860         return (TRUE);
 861 }
 862
 863 /*
 864  * Try to zero one physical page.  Used by an idle priority thread.
 865  */
 866 boolean_t
 867 vm_phys_zero_pages_idle(void)
 868 {
 869         static struct vm_freelist *fl = vm_phys_free_queues[0][0];
 870         static int flind, oind, pind;
 871         vm_page_t m, m_tmp;
 872
 873         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 874         for (;;) {
 875                 TAILQ_FOREACH_REVERSE(m, &fl[oind].pl, pglist, pageq) {
 876                         for (m_tmp = m; m_tmp < &m[1 << oind]; m_tmp++) {
 877                                 if ((m_tmp->flags & (PG_CACHED | PG_ZERO)) == 0) {
 878                                         vm_phys_unfree_page(m_tmp);
 879                                         cnt.v_free_count--;
 880                                         mtx_unlock(&vm_page_queue_free_mtx);
 881                                         pmap_zero_page_idle(m_tmp);
 882                                         m_tmp->flags |= PG_ZERO;
 883                                         mtx_lock(&vm_page_queue_free_mtx);
 884                                         cnt.v_free_count++;
 885                                         vm_phys_free_pages(m_tmp, 0);
 886                                         vm_page_zero_count++;
 887                                         cnt_prezero++;
 888                                         return (TRUE);
 889                                 }
 890                         }
 891                 }
 892                 oind++;
 893                 if (oind == VM_NFREEORDER) {
 894                         oind = 0;
 895                         pind++;
 896                         if (pind == VM_NFREEPOOL) {
 897                                 pind = 0;
 898                                 flind++;
 899                                 if (flind == vm_nfreelists)
 900                                         flind = 0;
 901                         }
 902                         fl = vm_phys_free_queues[flind][pind];
 903                 }
 904         }
 905 }
 906
 907 /*
 908  * Allocate a contiguous set of physical pages of the given size
 909  * "npages" from the free lists.  All of the physical pages must be at
 910  * or above the given physical address "low" and below the given
 911  * physical address "high".  The given value "alignment" determines the
 912  * alignment of the first physical page in the set.  If the given value
 913  * "boundary" is non-zero, then the set of physical pages cannot cross
 914  * any physical address boundary that is a multiple of that value.  Both
 915  * "alignment" and "boundary" must be a power of two.
 916  */
 917 vm_page_t
 918 vm_phys_alloc_contig(u_long npages, vm_paddr_t low, vm_paddr_t high,
 919     u_long alignment, u_long boundary)
 920 {
 921         struct vm_freelist *fl;
 922         struct vm_phys_seg *seg;
 923         vm_paddr_t pa, pa_last, size;
 924         vm_page_t m, m_ret;
 925         u_long npages_end;
 926         int domain, flind, oind, order, pind;
 927
 928         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 929 #if VM_NDOMAIN > 1
 930         domain = PCPU_GET(domain);
 931 #else
 932         domain = 0;
 933 #endif
 934         size = npages << PAGE_SHIFT;
 935         KASSERT(size != 0,
 936             ("vm_phys_alloc_contig: size must not be 0"));
 937         KASSERT((alignment & (alignment - 1)) == 0,
 938             ("vm_phys_alloc_contig: alignment must be a power of 2"));
 939         KASSERT((boundary & (boundary - 1)) == 0,
 940             ("vm_phys_alloc_contig: boundary must be a power of 2"));
 941         /* Compute the queue that is the best fit for npages. */
 942         for (order = 0; (1 << order) < npages; order++);
 943         for (flind = 0; flind < vm_nfreelists; flind++) {
 944                 for (oind = min(order, VM_NFREEORDER - 1); oind < VM_NFREEORDER; oind++) {
 945                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 946                                 fl = (*vm_phys_lookup_lists[domain][flind])
 947                                     [pind];
 948                                 TAILQ_FOREACH(m_ret, &fl[oind].pl, pageq) {
 949                                         /*
 950                                          * A free list may contain physical pages
 951                                          * from one or more segments.
 952                                          */
 953                                         seg = &vm_phys_segs[m_ret->segind];
 954                                         if (seg->start > high ||
 955                                             low >= seg->end)
 956                                                 continue;
 957
 958                                         /*
 959                                          * Is the size of this allocation request
 960                                          * larger than the largest block size?
 961                                          */
 962                                         if (order >= VM_NFREEORDER) {
 963                                                 /*
 964                                                  * Determine if a sufficient number
 965                                                  * of subsequent blocks to satisfy
 966                                                  * the allocation request are free.
 967                                                  */
 968                                                 pa = VM_PAGE_TO_PHYS(m_ret);
 969                                                 pa_last = pa + size;
 970                                                 for (;;) {
 971                                                         pa += 1 << (PAGE_SHIFT + VM_NFREEORDER - 1);
 972                                                         if (pa >= pa_last)
 973                                                                 break;
 974                                                         if (pa < seg->start ||
 975                                                             pa >= seg->end)
 976                                                                 break;
 977                                                         m = &seg->first_page[atop(pa - seg->start)];
 978                                                         if (m->order != VM_NFREEORDER - 1)
 979                                                                 break;
 980                                                 }
 981                                                 /* If not, continue to the next block. */
 982                                                 if (pa < pa_last)
 983                                                         continue;
 984                                         }
 985
 986                                         /*
 987                                          * Determine if the blocks are within the given range,
 988                                          * satisfy the given alignment, and do not cross the
 989                                          * given boundary.
 990                                          */
 991                                         pa = VM_PAGE_TO_PHYS(m_ret);
 992                                         if (pa >= low &&
 993                                             pa + size <= high &&
 994                                             (pa & (alignment - 1)) == 0 &&
 995                                             ((pa ^ (pa + size - 1)) & ~(boundary - 1)) == 0)
 996                                                 goto done;
 997                                 }
 998                         }
 999                 }
1000         }
1001         return (NULL);
1002 done:
1003         for (m = m_ret; m < &m_ret[npages]; m = &m[1 << oind]) {
1004                 fl = (*seg->free_queues)[m->pool];
1005                 TAILQ_REMOVE(&fl[m->order].pl, m, pageq);
1006                 fl[m->order].lcnt--;
1007                 m->order = VM_NFREEORDER;
1008         }
1009         if (m_ret->pool != VM_FREEPOOL_DEFAULT)
1010                 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m_ret, oind);
1011         fl = (*seg->free_queues)[m_ret->pool];
1012         vm_phys_split_pages(m_ret, oind, fl, order);
1013         /* Return excess pages to the free lists. */
1014         npages_end = roundup2(npages, 1 << imin(oind, order));
1015         if (npages < npages_end)
1016                 vm_phys_free_contig(&m_ret[npages], npages_end - npages);
1017         return (m_ret);
1018 }
1019
1020 #ifdef DDB
1021 /*
1022  * Show the number of physical pages in each of the free lists.
1023  */
1024 DB_SHOW_COMMAND(freepages, db_show_freepages)
1025 {
1026         struct vm_freelist *fl;
1027         int flind, oind, pind;
1028
1029         for (flind = 0; flind < vm_nfreelists; flind++) {
1030                 db_printf("FREE LIST %d:\n"
1031                     "\n  ORDER (SIZE)  |  NUMBER"
1032                     "\n              ", flind);
1033                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
1034                         db_printf("  |  POOL %d", pind);
1035                 db_printf("\n--            ");
1036                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
1037                         db_printf("-- --      ");
1038                 db_printf("--\n");
1039                 for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
1040                         db_printf("  %2.2d (%6.6dK)", oind,
1041                             1 << (PAGE_SHIFT - 10 + oind));
1042                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1043                                 fl = vm_phys_free_queues[flind][pind];
1044                                 db_printf("  |  %6.6d", fl[oind].lcnt);
1045                         }
1046                         db_printf("\n");
1047                 }
1048                 db_printf("\n");
1049         }
1050 }
1051 #endif