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