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, "  %2.2d (%6.6dK)", 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, "  |  %6.6d", 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         struct vm_freelist *fl;
 305         struct vm_freelist *alt;
 306         int flind, oind, pind;
 307         vm_page_t m;
 308
 309         KASSERT(pool < VM_NFREEPOOL,
 310             ("vm_phys_alloc_pages: pool %d is out of range", pool));
 311         KASSERT(order < VM_NFREEORDER,
 312             ("vm_phys_alloc_pages: order %d is out of range", order));
 313         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 314         for (flind = 0; flind < vm_nfreelists; flind++) {
 315                 fl = vm_phys_free_queues[flind][pool];
 316                 for (oind = order; oind < VM_NFREEORDER; oind++) {
 317                         m = TAILQ_FIRST(&fl[oind].pl);
 318                         if (m != NULL) {
 319                                 TAILQ_REMOVE(&fl[oind].pl, m, pageq);
 320                                 fl[oind].lcnt--;
 321                                 m->order = VM_NFREEORDER;
 322                                 vm_phys_split_pages(m, oind, fl, order);
 323                                 return (m);
 324                         }
 325                 }
 326
 327                 /*
 328                  * The given pool was empty.  Find the largest
 329                  * contiguous, power-of-two-sized set of pages in any
 330                  * pool.  Transfer these pages to the given pool, and
 331                  * use them to satisfy the allocation.
 332                  */
 333                 for (oind = VM_NFREEORDER - 1; oind >= order; oind--) {
 334                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 335                                 alt = vm_phys_free_queues[flind][pind];
 336                                 m = TAILQ_FIRST(&alt[oind].pl);
 337                                 if (m != NULL) {
 338                                         TAILQ_REMOVE(&alt[oind].pl, m, pageq);
 339                                         alt[oind].lcnt--;
 340                                         m->order = VM_NFREEORDER;
 341                                         vm_phys_set_pool(pool, m, oind);
 342                                         vm_phys_split_pages(m, oind, fl, order);
 343                                         return (m);
 344                                 }
 345                         }
 346                 }
 347         }
 348         return (NULL);
 349 }
 350
 351 /*
 352  * Allocate physical memory from phys_avail[].
 353  */
 354 vm_paddr_t
 355 vm_phys_bootstrap_alloc(vm_size_t size, unsigned long alignment)
 356 {
 357         vm_paddr_t pa;
 358         int i;
 359
 360         size = round_page(size);
 361         for (i = 0; phys_avail[i + 1] != 0; i += 2) {
 362                 if (phys_avail[i + 1] - phys_avail[i] < size)
 363                         continue;
 364                 pa = phys_avail[i];
 365                 phys_avail[i] += size;
 366                 return (pa);
 367         }
 368         panic("vm_phys_bootstrap_alloc");
 369 }
 370
 371 /*
 372  * Find the vm_page corresponding to the given physical address.
 373  */
 374 vm_page_t
 375 vm_phys_paddr_to_vm_page(vm_paddr_t pa)
 376 {
 377         struct vm_phys_seg *seg;
 378         int segind;
 379
 380         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 381                 seg = &vm_phys_segs[segind];
 382                 if (pa >= seg->start && pa < seg->end)
 383                         return (&seg->first_page[atop(pa - seg->start)]);
 384         }
 385         return (NULL);
 386 }
 387
 388 /*
 389  * Find the segment containing the given physical address.
 390  */
 391 static int
 392 vm_phys_paddr_to_segind(vm_paddr_t pa)
 393 {
 394         struct vm_phys_seg *seg;
 395         int segind;
 396
 397         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 398                 seg = &vm_phys_segs[segind];
 399                 if (pa >= seg->start && pa < seg->end)
 400                         return (segind);
 401         }
 402         panic("vm_phys_paddr_to_segind: paddr %#jx is not in any segment" ,
 403             (uintmax_t)pa);
 404 }
 405
 406 /*
 407  * Free a contiguous, power of two-sized set of physical pages.
 408  *
 409  * The free page queues must be locked.
 410  */
 411 void
 412 vm_phys_free_pages(vm_page_t m, int order)
 413 {
 414         struct vm_freelist *fl;
 415         struct vm_phys_seg *seg;
 416         vm_paddr_t pa, pa_buddy;
 417         vm_page_t m_buddy;
 418
 419         KASSERT(m->order == VM_NFREEORDER,
 420             ("vm_phys_free_pages: page %p has unexpected order %d",
 421             m, m->order));
 422         KASSERT(m->pool < VM_NFREEPOOL,
 423             ("vm_phys_free_pages: page %p has unexpected pool %d",
 424             m, m->pool));
 425         KASSERT(order < VM_NFREEORDER,
 426             ("vm_phys_free_pages: order %d is out of range", order));
 427         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 428         pa = VM_PAGE_TO_PHYS(m);
 429         seg = &vm_phys_segs[m->segind];
 430         while (order < VM_NFREEORDER - 1) {
 431                 pa_buddy = pa ^ (1 << (PAGE_SHIFT + order));
 432                 if (pa_buddy < seg->start ||
 433                     pa_buddy >= seg->end)
 434                         break;
 435                 m_buddy = &seg->first_page[atop(pa_buddy - seg->start)];
 436                 if (m_buddy->order != order)
 437                         break;
 438                 fl = (*seg->free_queues)[m_buddy->pool];
 439                 TAILQ_REMOVE(&fl[m_buddy->order].pl, m_buddy, pageq);
 440                 fl[m_buddy->order].lcnt--;
 441                 m_buddy->order = VM_NFREEORDER;
 442                 if (m_buddy->pool != m->pool)
 443                         vm_phys_set_pool(m->pool, m_buddy, order);
 444                 order++;
 445                 pa &= ~((1 << (PAGE_SHIFT + order)) - 1);
 446                 m = &seg->first_page[atop(pa - seg->start)];
 447         }
 448         m->order = order;
 449         fl = (*seg->free_queues)[m->pool];
 450         TAILQ_INSERT_TAIL(&fl[order].pl, m, pageq);
 451         fl[order].lcnt++;
 452 }
 453
 454 /*
 455  * Set the pool for a contiguous, power of two-sized set of physical pages.
 456  */
 457 void
 458 vm_phys_set_pool(int pool, vm_page_t m, int order)
 459 {
 460         vm_page_t m_tmp;
 461
 462         for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++)
 463                 m_tmp->pool = pool;
 464 }
 465
 466 /*
 467  * Search for the given physical page "m" in the free lists.  If the search
 468  * succeeds, remove "m" from the free lists and return TRUE.  Otherwise, return
 469  * FALSE, indicating that "m" is not in the free lists.
 470  *
 471  * The free page queues must be locked.
 472  */
 473 boolean_t
 474 vm_phys_unfree_page(vm_page_t m)
 475 {
 476         struct vm_freelist *fl;
 477         struct vm_phys_seg *seg;
 478         vm_paddr_t pa, pa_half;
 479         vm_page_t m_set, m_tmp;
 480         int order;
 481
 482         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 483
 484         /*
 485          * First, find the contiguous, power of two-sized set of free
 486          * physical pages containing the given physical page "m" and
 487          * assign it to "m_set".
 488          */
 489         seg = &vm_phys_segs[m->segind];
 490         for (m_set = m, order = 0; m_set->order == VM_NFREEORDER &&
 491             order < VM_NFREEORDER - 1; ) {
 492                 order++;
 493                 pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order));
 494                 if (pa >= seg->start)
 495                         m_set = &seg->first_page[atop(pa - seg->start)];
 496                 else
 497                         return (FALSE);
 498         }
 499         if (m_set->order < order)
 500                 return (FALSE);
 501         if (m_set->order == VM_NFREEORDER)
 502                 return (FALSE);
 503         KASSERT(m_set->order < VM_NFREEORDER,
 504             ("vm_phys_unfree_page: page %p has unexpected order %d",
 505             m_set, m_set->order));
 506
 507         /*
 508          * Next, remove "m_set" from the free lists.  Finally, extract
 509          * "m" from "m_set" using an iterative algorithm: While "m_set"
 510          * is larger than a page, shrink "m_set" by returning the half
 511          * of "m_set" that does not contain "m" to the free lists.
 512          */
 513         fl = (*seg->free_queues)[m_set->pool];
 514         order = m_set->order;
 515         TAILQ_REMOVE(&fl[order].pl, m_set, pageq);
 516         fl[order].lcnt--;
 517         m_set->order = VM_NFREEORDER;
 518         while (order > 0) {
 519                 order--;
 520                 pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order));
 521                 if (m->phys_addr < pa_half)
 522                         m_tmp = &seg->first_page[atop(pa_half - seg->start)];
 523                 else {
 524                         m_tmp = m_set;
 525                         m_set = &seg->first_page[atop(pa_half - seg->start)];
 526                 }
 527                 m_tmp->order = order;
 528                 TAILQ_INSERT_HEAD(&fl[order].pl, m_tmp, pageq);
 529                 fl[order].lcnt++;
 530         }
 531         KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency"));
 532         return (TRUE);
 533 }
 534
 535 /*
 536  * Try to zero one physical page.  Used by an idle priority thread.
 537  */
 538 boolean_t
 539 vm_phys_zero_pages_idle(void)
 540 {
 541         static struct vm_freelist *fl = vm_phys_free_queues[0][0];
 542         static int flind, oind, pind;
 543         vm_page_t m, m_tmp;
 544
 545         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 546         for (;;) {
 547                 TAILQ_FOREACH_REVERSE(m, &fl[oind].pl, pglist, pageq) {
 548                         for (m_tmp = m; m_tmp < &m[1 << oind]; m_tmp++) {
 549                                 if ((m_tmp->flags & (PG_CACHED | PG_ZERO)) == 0) {
 550                                         vm_phys_unfree_page(m_tmp);
 551                                         cnt.v_free_count--;
 552                                         mtx_unlock(&vm_page_queue_free_mtx);
 553                                         pmap_zero_page_idle(m_tmp);
 554                                         m_tmp->flags |= PG_ZERO;
 555                                         mtx_lock(&vm_page_queue_free_mtx);
 556                                         cnt.v_free_count++;
 557                                         vm_phys_free_pages(m_tmp, 0);
 558                                         vm_page_zero_count++;
 559                                         cnt_prezero++;
 560                                         return (TRUE);
 561                                 }
 562                         }
 563                 }
 564                 oind++;
 565                 if (oind == VM_NFREEORDER) {
 566                         oind = 0;
 567                         pind++;
 568                         if (pind == VM_NFREEPOOL) {
 569                                 pind = 0;
 570                                 flind++;
 571                                 if (flind == vm_nfreelists)
 572                                         flind = 0;
 573                         }
 574                         fl = vm_phys_free_queues[flind][pind];
 575                 }
 576         }
 577 }
 578
 579 /*
 580  * Allocate a contiguous set of physical pages of the given size
 581  * "npages" from the free lists.  All of the physical pages must be at
 582  * or above the given physical address "low" and below the given
 583  * physical address "high".  The given value "alignment" determines the
 584  * alignment of the first physical page in the set.  If the given value
 585  * "boundary" is non-zero, then the set of physical pages cannot cross
 586  * any physical address boundary that is a multiple of that value.  Both
 587  * "alignment" and "boundary" must be a power of two.
 588  */
 589 vm_page_t
 590 vm_phys_alloc_contig(unsigned long npages, vm_paddr_t low, vm_paddr_t high,
 591     unsigned long alignment, unsigned long boundary)
 592 {
 593         struct vm_freelist *fl;
 594         struct vm_phys_seg *seg;
 595         vm_object_t m_object;
 596         vm_paddr_t pa, pa_last, size;
 597         vm_page_t deferred_vdrop_list, m, m_ret;
 598         int flind, i, oind, order, pind;
 599
 600         size = npages << PAGE_SHIFT;
 601         KASSERT(size != 0,
 602             ("vm_phys_alloc_contig: size must not be 0"));
 603         KASSERT((alignment & (alignment - 1)) == 0,
 604             ("vm_phys_alloc_contig: alignment must be a power of 2"));
 605         KASSERT((boundary & (boundary - 1)) == 0,
 606             ("vm_phys_alloc_contig: boundary must be a power of 2"));
 607         deferred_vdrop_list = NULL;
 608         /* Compute the queue that is the best fit for npages. */
 609         for (order = 0; (1 << order) < npages; order++);
 610         mtx_lock(&vm_page_queue_free_mtx);
 611 #if VM_NRESERVLEVEL > 0
 612 retry:
 613 #endif
 614         for (flind = 0; flind < vm_nfreelists; flind++) {
 615                 for (oind = min(order, VM_NFREEORDER - 1); oind < VM_NFREEORDER; oind++) {
 616                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 617                                 fl = vm_phys_free_queues[flind][pind];
 618                                 TAILQ_FOREACH(m_ret, &fl[oind].pl, pageq) {
 619                                         /*
 620                                          * A free list may contain physical pages
 621                                          * from one or more segments.
 622                                          */
 623                                         seg = &vm_phys_segs[m_ret->segind];
 624                                         if (seg->start > high ||
 625                                             low >= seg->end)
 626                                                 continue;
 627
 628                                         /*
 629                                          * Is the size of this allocation request
 630                                          * larger than the largest block size?
 631                                          */
 632                                         if (order >= VM_NFREEORDER) {
 633                                                 /*
 634                                                  * Determine if a sufficient number
 635                                                  * of subsequent blocks to satisfy
 636                                                  * the allocation request are free.
 637                                                  */
 638                                                 pa = VM_PAGE_TO_PHYS(m_ret);
 639                                                 pa_last = pa + size;
 640                                                 for (;;) {
 641                                                         pa += 1 << (PAGE_SHIFT + VM_NFREEORDER - 1);
 642                                                         if (pa >= pa_last)
 643                                                                 break;
 644                                                         if (pa < seg->start ||
 645                                                             pa >= seg->end)
 646                                                                 break;
 647                                                         m = &seg->first_page[atop(pa - seg->start)];
 648                                                         if (m->order != VM_NFREEORDER - 1)
 649                                                                 break;
 650                                                 }
 651                                                 /* If not, continue to the next block. */
 652                                                 if (pa < pa_last)
 653                                                         continue;
 654                                         }
 655
 656                                         /*
 657                                          * Determine if the blocks are within the given range,
 658                                          * satisfy the given alignment, and do not cross the
 659                                          * given boundary.
 660                                          */
 661                                         pa = VM_PAGE_TO_PHYS(m_ret);
 662                                         if (pa >= low &&
 663                                             pa + size <= high &&
 664                                             (pa & (alignment - 1)) == 0 &&
 665                                             ((pa ^ (pa + size - 1)) & ~(boundary - 1)) == 0)
 666                                                 goto done;
 667                                 }
 668                         }
 669                 }
 670         }
 671 #if VM_NRESERVLEVEL > 0
 672         if (vm_reserv_reclaim_contig(size, low, high, alignment, boundary))
 673                 goto retry;
 674 #endif
 675         mtx_unlock(&vm_page_queue_free_mtx);
 676         return (NULL);
 677 done:
 678         for (m = m_ret; m < &m_ret[npages]; m = &m[1 << oind]) {
 679                 fl = (*seg->free_queues)[m->pool];
 680                 TAILQ_REMOVE(&fl[m->order].pl, m, pageq);
 681                 fl[m->order].lcnt--;
 682                 m->order = VM_NFREEORDER;
 683         }
 684         if (m_ret->pool != VM_FREEPOOL_DEFAULT)
 685                 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m_ret, oind);
 686         fl = (*seg->free_queues)[m_ret->pool];
 687         vm_phys_split_pages(m_ret, oind, fl, order);
 688         for (i = 0; i < npages; i++) {
 689                 m = &m_ret[i];
 690                 KASSERT(m->queue == PQ_NONE,
 691                     ("vm_phys_alloc_contig: page %p has unexpected queue %d",
 692                     m, m->queue));
 693                 KASSERT(m->wire_count == 0,
 694                     ("vm_phys_alloc_contig: page %p is wired", m));
 695                 KASSERT(m->hold_count == 0,
 696                     ("vm_phys_alloc_contig: page %p is held", m));
 697                 KASSERT(m->busy == 0,
 698                     ("vm_phys_alloc_contig: page %p is busy", m));
 699                 KASSERT(m->dirty == 0,
 700                     ("vm_phys_alloc_contig: page %p is dirty", m));
 701                 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
 702                     ("vm_phys_alloc_contig: page %p has unexpected memattr %d",
 703                     m, pmap_page_get_memattr(m)));
 704                 if ((m->flags & PG_CACHED) != 0) {
 705                         m->valid = 0;
 706                         m_object = m->object;
 707                         vm_page_cache_remove(m);
 708                         if (m_object->type == OBJT_VNODE &&
 709                             m_object->cache == 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 = m_object->handle;
 718                                 m->pageq.tqe_next = deferred_vdrop_list;
 719                                 deferred_vdrop_list = m;
 720                         }
 721                 } else {
 722                         KASSERT(VM_PAGE_IS_FREE(m),
 723                             ("vm_phys_alloc_contig: page %p is not free", m));
 724                         KASSERT(m->valid == 0,
 725                             ("vm_phys_alloc_contig: free page %p is valid", m));
 726                         cnt.v_free_count--;
 727                 }
 728                 if (m->flags & PG_ZERO)
 729                         vm_page_zero_count--;
 730                 /* Don't clear the PG_ZERO flag; we'll need it later. */
 731                 m->flags = PG_UNMANAGED | (m->flags & PG_ZERO);
 732                 m->oflags = 0;
 733                 /* Unmanaged pages don't use "act_count". */
 734         }
 735         for (; i < roundup2(npages, 1 << imin(oind, order)); i++) {
 736                 m = &m_ret[i];
 737                 KASSERT(m->order == VM_NFREEORDER,
 738                     ("vm_phys_alloc_contig: page %p has unexpected order %d",
 739                     m, m->order));
 740                 vm_phys_free_pages(m, 0);
 741         }
 742         mtx_unlock(&vm_page_queue_free_mtx);
 743         while (deferred_vdrop_list != NULL) {
 744                 vdrop((struct vnode *)deferred_vdrop_list->pageq.tqe_prev);
 745                 deferred_vdrop_list = deferred_vdrop_list->pageq.tqe_next;
 746         }
 747         return (m_ret);
 748 }
 749
 750 #ifdef DDB
 751 /*
 752  * Show the number of physical pages in each of the free lists.
 753  */
 754 DB_SHOW_COMMAND(freepages, db_show_freepages)
 755 {
 756         struct vm_freelist *fl;
 757         int flind, oind, pind;
 758
 759         for (flind = 0; flind < vm_nfreelists; flind++) {
 760                 db_printf("FREE LIST %d:\n"
 761                     "\n  ORDER (SIZE)  |  NUMBER"
 762                     "\n              ", flind);
 763                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 764                         db_printf("  |  POOL %d", pind);
 765                 db_printf("\n--            ");
 766                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 767                         db_printf("-- --      ");
 768                 db_printf("--\n");
 769                 for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
 770                         db_printf("  %2.2d (%6.6dK)", oind,
 771                             1 << (PAGE_SHIFT - 10 + oind));
 772                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 773                                 fl = vm_phys_free_queues[flind][pind];
 774                                 db_printf("  |  %6.6d", fl[oind].lcnt);
 775                         }
 776                         db_printf("\n");
 777                 }
 778                 db_printf("\n");
 779         }
 780 }
 781 #endif