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 #include <sys/cdefs.h>
  33 __FBSDID("$FreeBSD$");
  34
  35 #include "opt_ddb.h"
  36
  37 #include <sys/param.h>
  38 #include <sys/systm.h>
  39 #include <sys/lock.h>
  40 #include <sys/kernel.h>
  41 #include <sys/malloc.h>
  42 #include <sys/mutex.h>
  43 #include <sys/queue.h>
  44 #include <sys/sbuf.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/vmmeter.h>
  47 #include <sys/vnode.h>
  48
  49 #include <ddb/ddb.h>
  50
  51 #include <vm/vm.h>
  52 #include <vm/vm_param.h>
  53 #include <vm/vm_kern.h>
  54 #include <vm/vm_object.h>
  55 #include <vm/vm_page.h>
  56 #include <vm/vm_phys.h>
  57 #include <vm/vm_reserv.h>
  58
  59 struct vm_freelist {
  60         struct pglist pl;
  61         int lcnt;
  62 };
  63
  64 struct vm_phys_seg {
  65         vm_paddr_t      start;
  66         vm_paddr_t      end;
  67         vm_page_t       first_page;
  68         struct vm_freelist (*free_queues)[VM_NFREEPOOL][VM_NFREEORDER];
  69 };
  70
  71 static struct vm_phys_seg vm_phys_segs[VM_PHYSSEG_MAX];
  72
  73 static int vm_phys_nsegs;
  74
  75 static struct vm_freelist
  76     vm_phys_free_queues[VM_NFREELIST][VM_NFREEPOOL][VM_NFREEORDER];
  77
  78 static int vm_nfreelists = VM_FREELIST_DEFAULT + 1;
  79
  80 static int cnt_prezero;
  81 SYSCTL_INT(_vm_stats_misc, OID_AUTO, cnt_prezero, CTLFLAG_RD,
  82     &cnt_prezero, 0, "The number of physical pages prezeroed at idle time");
  83
  84 static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS);
  85 SYSCTL_OID(_vm, OID_AUTO, phys_free, CTLTYPE_STRING | CTLFLAG_RD,
  86     NULL, 0, sysctl_vm_phys_free, "A", "Phys Free Info");
  87
  88 static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS);
  89 SYSCTL_OID(_vm, OID_AUTO, phys_segs, CTLTYPE_STRING | CTLFLAG_RD,
  90     NULL, 0, sysctl_vm_phys_segs, "A", "Phys Seg Info");
  91
  92 static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind);
  93 static int vm_phys_paddr_to_segind(vm_paddr_t pa);
  94 static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl,
  95     int order);
  96
  97 /*
  98  * Outputs the state of the physical memory allocator, specifically,
  99  * the amount of physical memory in each free list.
 100  */
 101 static int
 102 sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS)
 103 {
 104         struct sbuf sbuf;
 105         struct vm_freelist *fl;
 106         char *cbuf;
 107         const int cbufsize = vm_nfreelists*(VM_NFREEORDER + 1)*81;
 108         int error, flind, oind, pind;
 109
 110         cbuf = malloc(cbufsize, M_TEMP, M_WAITOK | M_ZERO);
 111         sbuf_new(&sbuf, cbuf, cbufsize, SBUF_FIXEDLEN);
 112         for (flind = 0; flind < vm_nfreelists; flind++) {
 113                 sbuf_printf(&sbuf, "\nFREE LIST %d:\n"
 114                     "\n  ORDER (SIZE)  |  NUMBER"
 115                     "\n              ", flind);
 116                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 117                         sbuf_printf(&sbuf, "  |  POOL %d", pind);
 118                 sbuf_printf(&sbuf, "\n--            ");
 119                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 120                         sbuf_printf(&sbuf, "-- --      ");
 121                 sbuf_printf(&sbuf, "--\n");
 122                 for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
 123                         sbuf_printf(&sbuf, "  %2d (%6dK)", oind,
 124                             1 << (PAGE_SHIFT - 10 + oind));
 125                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 126                                 fl = vm_phys_free_queues[flind][pind];
 127                                 sbuf_printf(&sbuf, "  |  %6d", fl[oind].lcnt);
 128                         }
 129                         sbuf_printf(&sbuf, "\n");
 130                 }
 131         }
 132         sbuf_finish(&sbuf);
 133         error = SYSCTL_OUT(req, sbuf_data(&sbuf), sbuf_len(&sbuf));
 134         sbuf_delete(&sbuf);
 135         free(cbuf, M_TEMP);
 136         return (error);
 137 }
 138
 139 /*
 140  * Outputs the set of physical memory segments.
 141  */
 142 static int
 143 sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS)
 144 {
 145         struct sbuf sbuf;
 146         struct vm_phys_seg *seg;
 147         char *cbuf;
 148         const int cbufsize = VM_PHYSSEG_MAX*(VM_NFREEORDER + 1)*81;
 149         int error, segind;
 150
 151         cbuf = malloc(cbufsize, M_TEMP, M_WAITOK | M_ZERO);
 152         sbuf_new(&sbuf, cbuf, cbufsize, SBUF_FIXEDLEN);
 153         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 154                 sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind);
 155                 seg = &vm_phys_segs[segind];
 156                 sbuf_printf(&sbuf, "start:     %#jx\n",
 157                     (uintmax_t)seg->start);
 158                 sbuf_printf(&sbuf, "end:       %#jx\n",
 159                     (uintmax_t)seg->end);
 160                 sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues);
 161         }
 162         sbuf_finish(&sbuf);
 163         error = SYSCTL_OUT(req, sbuf_data(&sbuf), sbuf_len(&sbuf));
 164         sbuf_delete(&sbuf);
 165         free(cbuf, M_TEMP);
 166         return (error);
 167 }
 168
 169 /*
 170  * Create a physical memory segment.
 171  */
 172 static void
 173 vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind)
 174 {
 175         struct vm_phys_seg *seg;
 176 #ifdef VM_PHYSSEG_SPARSE
 177         long pages;
 178         int segind;
 179
 180         pages = 0;
 181         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 182                 seg = &vm_phys_segs[segind];
 183                 pages += atop(seg->end - seg->start);
 184         }
 185 #endif
 186         KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX,
 187             ("vm_phys_create_seg: increase VM_PHYSSEG_MAX"));
 188         seg = &vm_phys_segs[vm_phys_nsegs++];
 189         seg->start = start;
 190         seg->end = end;
 191 #ifdef VM_PHYSSEG_SPARSE
 192         seg->first_page = &vm_page_array[pages];
 193 #else
 194         seg->first_page = PHYS_TO_VM_PAGE(start);
 195 #endif
 196         seg->free_queues = &vm_phys_free_queues[flind];
 197 }
 198
 199 /*
 200  * Initialize the physical memory allocator.
 201  */
 202 void
 203 vm_phys_init(void)
 204 {
 205         struct vm_freelist *fl;
 206         int flind, i, oind, pind;
 207
 208         for (i = 0; phys_avail[i + 1] != 0; i += 2) {
 209 #ifdef  VM_FREELIST_ISADMA
 210                 if (phys_avail[i] < 16777216) {
 211                         if (phys_avail[i + 1] > 16777216) {
 212                                 vm_phys_create_seg(phys_avail[i], 16777216,
 213                                     VM_FREELIST_ISADMA);
 214                                 vm_phys_create_seg(16777216, phys_avail[i + 1],
 215                                     VM_FREELIST_DEFAULT);
 216                         } else {
 217                                 vm_phys_create_seg(phys_avail[i],
 218                                     phys_avail[i + 1], VM_FREELIST_ISADMA);
 219                         }
 220                         if (VM_FREELIST_ISADMA >= vm_nfreelists)
 221                                 vm_nfreelists = VM_FREELIST_ISADMA + 1;
 222                 } else
 223 #endif
 224 #ifdef  VM_FREELIST_HIGHMEM
 225                 if (phys_avail[i + 1] > VM_HIGHMEM_ADDRESS) {
 226                         if (phys_avail[i] < VM_HIGHMEM_ADDRESS) {
 227                                 vm_phys_create_seg(phys_avail[i],
 228                                     VM_HIGHMEM_ADDRESS, VM_FREELIST_DEFAULT);
 229                                 vm_phys_create_seg(VM_HIGHMEM_ADDRESS,
 230                                     phys_avail[i + 1], VM_FREELIST_HIGHMEM);
 231                         } else {
 232                                 vm_phys_create_seg(phys_avail[i],
 233                                     phys_avail[i + 1], VM_FREELIST_HIGHMEM);
 234                         }
 235                         if (VM_FREELIST_HIGHMEM >= vm_nfreelists)
 236                                 vm_nfreelists = VM_FREELIST_HIGHMEM + 1;
 237                 } else
 238 #endif
 239                 vm_phys_create_seg(phys_avail[i], phys_avail[i + 1],
 240                     VM_FREELIST_DEFAULT);
 241         }
 242         for (flind = 0; flind < vm_nfreelists; flind++) {
 243                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 244                         fl = vm_phys_free_queues[flind][pind];
 245                         for (oind = 0; oind < VM_NFREEORDER; oind++)
 246                                 TAILQ_INIT(&fl[oind].pl);
 247                 }
 248         }
 249 }
 250
 251 /*
 252  * Split a contiguous, power of two-sized set of physical pages.
 253  */
 254 static __inline void
 255 vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order)
 256 {
 257         vm_page_t m_buddy;
 258
 259         while (oind > order) {
 260                 oind--;
 261                 m_buddy = &m[1 << oind];
 262                 KASSERT(m_buddy->order == VM_NFREEORDER,
 263                     ("vm_phys_split_pages: page %p has unexpected order %d",
 264                     m_buddy, m_buddy->order));
 265                 m_buddy->order = oind;
 266                 TAILQ_INSERT_HEAD(&fl[oind].pl, m_buddy, pageq);
 267                 fl[oind].lcnt++;
 268         }
 269 }
 270
 271 /*
 272  * Initialize a physical page and add it to the free lists.
 273  */
 274 void
 275 vm_phys_add_page(vm_paddr_t pa)
 276 {
 277         vm_page_t m;
 278
 279         cnt.v_page_count++;
 280         m = vm_phys_paddr_to_vm_page(pa);
 281         m->phys_addr = pa;
 282         m->segind = vm_phys_paddr_to_segind(pa);
 283         m->flags = PG_FREE;
 284         KASSERT(m->order == VM_NFREEORDER,
 285             ("vm_phys_add_page: page %p has unexpected order %d",
 286             m, m->order));
 287         m->pool = VM_FREEPOOL_DEFAULT;
 288         pmap_page_init(m);
 289         mtx_lock(&vm_page_queue_free_mtx);
 290         cnt.v_free_count++;
 291         vm_phys_free_pages(m, 0);
 292         mtx_unlock(&vm_page_queue_free_mtx);
 293 }
 294
 295 /*
 296  * Allocate a contiguous, power of two-sized set of physical pages
 297  * from the free lists.
 298  *
 299  * The free page queues must be locked.
 300  */
 301 vm_page_t
 302 vm_phys_alloc_pages(int pool, int order)
 303 {
 304         vm_page_t m;
 305         int flind;
 306
 307         for (flind = 0; flind < vm_nfreelists; flind++) {
 308                 m = vm_phys_alloc_freelist_pages(flind, pool, order);
 309                 if (m != NULL)
 310                         return (m);
 311         }
 312         return (NULL);
 313 }
 314
 315 /*
 316  * Find and dequeue a free page on the given free list, with the
 317  * specified pool and order
 318  */
 319 vm_page_t
 320 vm_phys_alloc_freelist_pages(int flind, int pool, int order)
 321 {
 322         struct vm_freelist *fl;
 323         struct vm_freelist *alt;
 324         int oind, pind;
 325         vm_page_t m;
 326
 327         KASSERT(flind < VM_NFREELIST,
 328             ("vm_phys_alloc_freelist_pages: freelist %d is out of range", flind));
 329         KASSERT(pool < VM_NFREEPOOL,
 330             ("vm_phys_alloc_freelist_pages: pool %d is out of range", pool));
 331         KASSERT(order < VM_NFREEORDER,
 332             ("vm_phys_alloc_freelist_pages: order %d is out of range", order));
 333         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 334         fl = vm_phys_free_queues[flind][pool];
 335         for (oind = order; oind < VM_NFREEORDER; oind++) {
 336                 m = TAILQ_FIRST(&fl[oind].pl);
 337                 if (m != NULL) {
 338                         TAILQ_REMOVE(&fl[oind].pl, m, pageq);
 339                         fl[oind].lcnt--;
 340                         m->order = VM_NFREEORDER;
 341                         vm_phys_split_pages(m, oind, fl, order);
 342                         return (m);
 343                 }
 344         }
 345
 346         /*
 347          * The given pool was empty.  Find the largest
 348          * contiguous, power-of-two-sized set of pages in any
 349          * pool.  Transfer these pages to the given pool, and
 350          * use them to satisfy the allocation.
 351          */
 352         for (oind = VM_NFREEORDER - 1; oind >= order; oind--) {
 353                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 354                         alt = vm_phys_free_queues[flind][pind];
 355                         m = TAILQ_FIRST(&alt[oind].pl);
 356                         if (m != NULL) {
 357                                 TAILQ_REMOVE(&alt[oind].pl, m, pageq);
 358                                 alt[oind].lcnt--;
 359                                 m->order = VM_NFREEORDER;
 360                                 vm_phys_set_pool(pool, m, oind);
 361                                 vm_phys_split_pages(m, oind, fl, order);
 362                                 return (m);
 363                         }
 364                 }
 365         }
 366         return (NULL);
 367 }
 368
 369 /*
 370  * Allocate physical memory from phys_avail[].
 371  */
 372 vm_paddr_t
 373 vm_phys_bootstrap_alloc(vm_size_t size, unsigned long alignment)
 374 {
 375         vm_paddr_t pa;
 376         int i;
 377
 378         size = round_page(size);
 379         for (i = 0; phys_avail[i + 1] != 0; i += 2) {
 380                 if (phys_avail[i + 1] - phys_avail[i] < size)
 381                         continue;
 382                 pa = phys_avail[i];
 383                 phys_avail[i] += size;
 384                 return (pa);
 385         }
 386         panic("vm_phys_bootstrap_alloc");
 387 }
 388
 389 /*
 390  * Find the vm_page corresponding to the given physical address.
 391  */
 392 vm_page_t
 393 vm_phys_paddr_to_vm_page(vm_paddr_t pa)
 394 {
 395         struct vm_phys_seg *seg;
 396         int segind;
 397
 398         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 399                 seg = &vm_phys_segs[segind];
 400                 if (pa >= seg->start && pa < seg->end)
 401                         return (&seg->first_page[atop(pa - seg->start)]);
 402         }
 403         return (NULL);
 404 }
 405
 406 /*
 407  * Find the segment containing the given physical address.
 408  */
 409 static int
 410 vm_phys_paddr_to_segind(vm_paddr_t pa)
 411 {
 412         struct vm_phys_seg *seg;
 413         int segind;
 414
 415         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 416                 seg = &vm_phys_segs[segind];
 417                 if (pa >= seg->start && pa < seg->end)
 418                         return (segind);
 419         }
 420         panic("vm_phys_paddr_to_segind: paddr %#jx is not in any segment" ,
 421             (uintmax_t)pa);
 422 }
 423
 424 /*
 425  * Free a contiguous, power of two-sized set of physical pages.
 426  *
 427  * The free page queues must be locked.
 428  */
 429 void
 430 vm_phys_free_pages(vm_page_t m, int order)
 431 {
 432         struct vm_freelist *fl;
 433         struct vm_phys_seg *seg;
 434         vm_paddr_t pa, pa_buddy;
 435         vm_page_t m_buddy;
 436
 437         KASSERT(m->order == VM_NFREEORDER,
 438             ("vm_phys_free_pages: page %p has unexpected order %d",
 439             m, m->order));
 440         KASSERT(m->pool < VM_NFREEPOOL,
 441             ("vm_phys_free_pages: page %p has unexpected pool %d",
 442             m, m->pool));
 443         KASSERT(order < VM_NFREEORDER,
 444             ("vm_phys_free_pages: order %d is out of range", order));
 445         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 446         pa = VM_PAGE_TO_PHYS(m);
 447         seg = &vm_phys_segs[m->segind];
 448         while (order < VM_NFREEORDER - 1) {
 449                 pa_buddy = pa ^ (1 << (PAGE_SHIFT + order));
 450                 if (pa_buddy < seg->start ||
 451                     pa_buddy >= seg->end)
 452                         break;
 453                 m_buddy = &seg->first_page[atop(pa_buddy - seg->start)];
 454                 if (m_buddy->order != order)
 455                         break;
 456                 fl = (*seg->free_queues)[m_buddy->pool];
 457                 TAILQ_REMOVE(&fl[m_buddy->order].pl, m_buddy, pageq);
 458                 fl[m_buddy->order].lcnt--;
 459                 m_buddy->order = VM_NFREEORDER;
 460                 if (m_buddy->pool != m->pool)
 461                         vm_phys_set_pool(m->pool, m_buddy, order);
 462                 order++;
 463                 pa &= ~((1 << (PAGE_SHIFT + order)) - 1);
 464                 m = &seg->first_page[atop(pa - seg->start)];
 465         }
 466         m->order = order;
 467         fl = (*seg->free_queues)[m->pool];
 468         TAILQ_INSERT_TAIL(&fl[order].pl, m, pageq);
 469         fl[order].lcnt++;
 470 }
 471
 472 /*
 473  * Set the pool for a contiguous, power of two-sized set of physical pages.
 474  */
 475 void
 476 vm_phys_set_pool(int pool, vm_page_t m, int order)
 477 {
 478         vm_page_t m_tmp;
 479
 480         for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++)
 481                 m_tmp->pool = pool;
 482 }
 483
 484 /*
 485  * Search for the given physical page "m" in the free lists.  If the search
 486  * succeeds, remove "m" from the free lists and return TRUE.  Otherwise, return
 487  * FALSE, indicating that "m" is not in the free lists.
 488  *
 489  * The free page queues must be locked.
 490  */
 491 boolean_t
 492 vm_phys_unfree_page(vm_page_t m)
 493 {
 494         struct vm_freelist *fl;
 495         struct vm_phys_seg *seg;
 496         vm_paddr_t pa, pa_half;
 497         vm_page_t m_set, m_tmp;
 498         int order;
 499
 500         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 501
 502         /*
 503          * First, find the contiguous, power of two-sized set of free
 504          * physical pages containing the given physical page "m" and
 505          * assign it to "m_set".
 506          */
 507         seg = &vm_phys_segs[m->segind];
 508         for (m_set = m, order = 0; m_set->order == VM_NFREEORDER &&
 509             order < VM_NFREEORDER - 1; ) {
 510                 order++;
 511                 pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order));
 512                 if (pa >= seg->start)
 513                         m_set = &seg->first_page[atop(pa - seg->start)];
 514                 else
 515                         return (FALSE);
 516         }
 517         if (m_set->order < order)
 518                 return (FALSE);
 519         if (m_set->order == VM_NFREEORDER)
 520                 return (FALSE);
 521         KASSERT(m_set->order < VM_NFREEORDER,
 522             ("vm_phys_unfree_page: page %p has unexpected order %d",
 523             m_set, m_set->order));
 524
 525         /*
 526          * Next, remove "m_set" from the free lists.  Finally, extract
 527          * "m" from "m_set" using an iterative algorithm: While "m_set"
 528          * is larger than a page, shrink "m_set" by returning the half
 529          * of "m_set" that does not contain "m" to the free lists.
 530          */
 531         fl = (*seg->free_queues)[m_set->pool];
 532         order = m_set->order;
 533         TAILQ_REMOVE(&fl[order].pl, m_set, pageq);
 534         fl[order].lcnt--;
 535         m_set->order = VM_NFREEORDER;
 536         while (order > 0) {
 537                 order--;
 538                 pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order));
 539                 if (m->phys_addr < pa_half)
 540                         m_tmp = &seg->first_page[atop(pa_half - seg->start)];
 541                 else {
 542                         m_tmp = m_set;
 543                         m_set = &seg->first_page[atop(pa_half - seg->start)];
 544                 }
 545                 m_tmp->order = order;
 546                 TAILQ_INSERT_HEAD(&fl[order].pl, m_tmp, pageq);
 547                 fl[order].lcnt++;
 548         }
 549         KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency"));
 550         return (TRUE);
 551 }
 552
 553 /*
 554  * Try to zero one physical page.  Used by an idle priority thread.
 555  */
 556 boolean_t
 557 vm_phys_zero_pages_idle(void)
 558 {
 559         static struct vm_freelist *fl = vm_phys_free_queues[0][0];
 560         static int flind, oind, pind;
 561         vm_page_t m, m_tmp;
 562
 563         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 564         for (;;) {
 565                 TAILQ_FOREACH_REVERSE(m, &fl[oind].pl, pglist, pageq) {
 566                         for (m_tmp = m; m_tmp < &m[1 << oind]; m_tmp++) {
 567                                 if ((m_tmp->flags & (PG_CACHED | PG_ZERO)) == 0) {
 568                                         vm_phys_unfree_page(m_tmp);
 569                                         cnt.v_free_count--;
 570                                         mtx_unlock(&vm_page_queue_free_mtx);
 571                                         pmap_zero_page_idle(m_tmp);
 572                                         m_tmp->flags |= PG_ZERO;
 573                                         mtx_lock(&vm_page_queue_free_mtx);
 574                                         cnt.v_free_count++;
 575                                         vm_phys_free_pages(m_tmp, 0);
 576                                         vm_page_zero_count++;
 577                                         cnt_prezero++;
 578                                         return (TRUE);
 579                                 }
 580                         }
 581                 }
 582                 oind++;
 583                 if (oind == VM_NFREEORDER) {
 584                         oind = 0;
 585                         pind++;
 586                         if (pind == VM_NFREEPOOL) {
 587                                 pind = 0;
 588                                 flind++;
 589                                 if (flind == vm_nfreelists)
 590                                         flind = 0;
 591                         }
 592                         fl = vm_phys_free_queues[flind][pind];
 593                 }
 594         }
 595 }
 596
 597 /*
 598  * Allocate a contiguous set of physical pages of the given size
 599  * "npages" from the free lists.  All of the physical pages must be at
 600  * or above the given physical address "low" and below the given
 601  * physical address "high".  The given value "alignment" determines the
 602  * alignment of the first physical page in the set.  If the given value
 603  * "boundary" is non-zero, then the set of physical pages cannot cross
 604  * any physical address boundary that is a multiple of that value.  Both
 605  * "alignment" and "boundary" must be a power of two.
 606  */
 607 vm_page_t
 608 vm_phys_alloc_contig(unsigned long npages, vm_paddr_t low, vm_paddr_t high,
 609     unsigned long alignment, unsigned long boundary)
 610 {
 611         struct vm_freelist *fl;
 612         struct vm_phys_seg *seg;
 613         struct vnode *vp;
 614         vm_paddr_t pa, pa_last, size;
 615         vm_page_t deferred_vdrop_list, m, m_ret;
 616         int flind, i, oind, order, pind;
 617
 618         size = npages << PAGE_SHIFT;
 619         KASSERT(size != 0,
 620             ("vm_phys_alloc_contig: size must not be 0"));
 621         KASSERT((alignment & (alignment - 1)) == 0,
 622             ("vm_phys_alloc_contig: alignment must be a power of 2"));
 623         KASSERT((boundary & (boundary - 1)) == 0,
 624             ("vm_phys_alloc_contig: boundary must be a power of 2"));
 625         deferred_vdrop_list = NULL;
 626         /* Compute the queue that is the best fit for npages. */
 627         for (order = 0; (1 << order) < npages; order++);
 628         mtx_lock(&vm_page_queue_free_mtx);
 629 #if VM_NRESERVLEVEL > 0
 630 retry:
 631 #endif
 632         for (flind = 0; flind < vm_nfreelists; flind++) {
 633                 for (oind = min(order, VM_NFREEORDER - 1); oind < VM_NFREEORDER; oind++) {
 634                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 635                                 fl = vm_phys_free_queues[flind][pind];
 636                                 TAILQ_FOREACH(m_ret, &fl[oind].pl, pageq) {
 637                                         /*
 638                                          * A free list may contain physical pages
 639                                          * from one or more segments.
 640                                          */
 641                                         seg = &vm_phys_segs[m_ret->segind];
 642                                         if (seg->start > high ||
 643                                             low >= seg->end)
 644                                                 continue;
 645
 646                                         /*
 647                                          * Is the size of this allocation request
 648                                          * larger than the largest block size?
 649                                          */
 650                                         if (order >= VM_NFREEORDER) {
 651                                                 /*
 652                                                  * Determine if a sufficient number
 653                                                  * of subsequent blocks to satisfy
 654                                                  * the allocation request are free.
 655                                                  */
 656                                                 pa = VM_PAGE_TO_PHYS(m_ret);
 657                                                 pa_last = pa + size;
 658                                                 for (;;) {
 659                                                         pa += 1 << (PAGE_SHIFT + VM_NFREEORDER - 1);
 660                                                         if (pa >= pa_last)
 661                                                                 break;
 662                                                         if (pa < seg->start ||
 663                                                             pa >= seg->end)
 664                                                                 break;
 665                                                         m = &seg->first_page[atop(pa - seg->start)];
 666                                                         if (m->order != VM_NFREEORDER - 1)
 667                                                                 break;
 668                                                 }
 669                                                 /* If not, continue to the next block. */
 670                                                 if (pa < pa_last)
 671                                                         continue;
 672                                         }
 673
 674                                         /*
 675                                          * Determine if the blocks are within the given range,
 676                                          * satisfy the given alignment, and do not cross the
 677                                          * given boundary.
 678                                          */
 679                                         pa = VM_PAGE_TO_PHYS(m_ret);
 680                                         if (pa >= low &&
 681                                             pa + size <= high &&
 682                                             (pa & (alignment - 1)) == 0 &&
 683                                             ((pa ^ (pa + size - 1)) & ~(boundary - 1)) == 0)
 684                                                 goto done;
 685                                 }
 686                         }
 687                 }
 688         }
 689 #if VM_NRESERVLEVEL > 0
 690         if (vm_reserv_reclaim_contig(size, low, high, alignment, boundary))
 691                 goto retry;
 692 #endif
 693         mtx_unlock(&vm_page_queue_free_mtx);
 694         return (NULL);
 695 done:
 696         for (m = m_ret; m < &m_ret[npages]; m = &m[1 << oind]) {
 697                 fl = (*seg->free_queues)[m->pool];
 698                 TAILQ_REMOVE(&fl[m->order].pl, m, pageq);
 699                 fl[m->order].lcnt--;
 700                 m->order = VM_NFREEORDER;
 701         }
 702         if (m_ret->pool != VM_FREEPOOL_DEFAULT)
 703                 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m_ret, oind);
 704         fl = (*seg->free_queues)[m_ret->pool];
 705         vm_phys_split_pages(m_ret, oind, fl, order);
 706         for (i = 0; i < npages; i++) {
 707                 m = &m_ret[i];
 708                 vp = vm_page_alloc_init(m);
 709                 if (vp != NULL) {
 710                         /*
 711                          * Enqueue the vnode for deferred vdrop().
 712                          *
 713                          * Unmanaged pages don't use "pageq", so it
 714                          * can be safely abused to construct a short-
 715                          * lived queue of vnodes.
 716                          */
 717                         m->pageq.tqe_prev = (void *)vp;
 718                         m->pageq.tqe_next = deferred_vdrop_list;
 719                         deferred_vdrop_list = m;
 720                 }
 721         }
 722         for (; i < roundup2(npages, 1 << imin(oind, order)); i++) {
 723                 m = &m_ret[i];
 724                 KASSERT(m->order == VM_NFREEORDER,
 725                     ("vm_phys_alloc_contig: page %p has unexpected order %d",
 726                     m, m->order));
 727                 vm_phys_free_pages(m, 0);
 728         }
 729         mtx_unlock(&vm_page_queue_free_mtx);
 730         while (deferred_vdrop_list != NULL) {
 731                 vdrop((struct vnode *)deferred_vdrop_list->pageq.tqe_prev);
 732                 deferred_vdrop_list = deferred_vdrop_list->pageq.tqe_next;
 733         }
 734         return (m_ret);
 735 }
 736
 737 #ifdef DDB
 738 /*
 739  * Show the number of physical pages in each of the free lists.
 740  */
 741 DB_SHOW_COMMAND(freepages, db_show_freepages)
 742 {
 743         struct vm_freelist *fl;
 744         int flind, oind, pind;
 745
 746         for (flind = 0; flind < vm_nfreelists; flind++) {
 747                 db_printf("FREE LIST %d:\n"
 748                     "\n  ORDER (SIZE)  |  NUMBER"
 749                     "\n              ", flind);
 750                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 751                         db_printf("  |  POOL %d", pind);
 752                 db_printf("\n--            ");
 753                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 754                         db_printf("-- --      ");
 755                 db_printf("--\n");
 756                 for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
 757                         db_printf("  %2.2d (%6.6dK)", oind,
 758                             1 << (PAGE_SHIFT - 10 + oind));
 759                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 760                                 fl = vm_phys_free_queues[flind][pind];
 761                                 db_printf("  |  %6.6d", fl[oind].lcnt);
 762                         }
 763                         db_printf("\n");
 764                 }
 765                 db_printf("\n");
 766         }
 767 }
 768 #endif