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