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 /*
  61  * VM_FREELIST_DEFAULT is split into VM_NDOMAIN lists, one for each
  62  * domain.  These extra lists are stored at the end of the regular
  63  * free lists starting with VM_NFREELIST.
  64  */
  65 #define VM_RAW_NFREELIST        (VM_NFREELIST + VM_NDOMAIN - 1)
  66
  67 struct vm_freelist {
  68         struct pglist pl;
  69         int lcnt;
  70 };
  71
  72 struct vm_phys_seg {
  73         vm_paddr_t      start;
  74         vm_paddr_t      end;
  75         vm_page_t       first_page;
  76         int             domain;
  77         struct vm_freelist (*free_queues)[VM_NFREEPOOL][VM_NFREEORDER];
  78 };
  79
  80 struct mem_affinity *mem_affinity;
  81
  82 static struct vm_phys_seg vm_phys_segs[VM_PHYSSEG_MAX];
  83
  84 static int vm_phys_nsegs;
  85
  86 static struct vm_freelist
  87     vm_phys_free_queues[VM_RAW_NFREELIST][VM_NFREEPOOL][VM_NFREEORDER];
  88 static struct vm_freelist
  89 (*vm_phys_lookup_lists[VM_NDOMAIN][VM_RAW_NFREELIST])[VM_NFREEPOOL][VM_NFREEORDER];
  90
  91 static int vm_nfreelists = VM_FREELIST_DEFAULT + 1;
  92
  93 static int cnt_prezero;
  94 SYSCTL_INT(_vm_stats_misc, OID_AUTO, cnt_prezero, CTLFLAG_RD,
  95     &cnt_prezero, 0, "The number of physical pages prezeroed at idle time");
  96
  97 static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS);
  98 SYSCTL_OID(_vm, OID_AUTO, phys_free, CTLTYPE_STRING | CTLFLAG_RD,
  99     NULL, 0, sysctl_vm_phys_free, "A", "Phys Free Info");
 100
 101 static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS);
 102 SYSCTL_OID(_vm, OID_AUTO, phys_segs, CTLTYPE_STRING | CTLFLAG_RD,
 103     NULL, 0, sysctl_vm_phys_segs, "A", "Phys Seg Info");
 104
 105 #if VM_NDOMAIN > 1
 106 static int sysctl_vm_phys_lookup_lists(SYSCTL_HANDLER_ARGS);
 107 SYSCTL_OID(_vm, OID_AUTO, phys_lookup_lists, CTLTYPE_STRING | CTLFLAG_RD,
 108     NULL, 0, sysctl_vm_phys_lookup_lists, "A", "Phys Lookup Lists");
 109 #endif
 110
 111 static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind,
 112     int domain);
 113 static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind);
 114 static int vm_phys_paddr_to_segind(vm_paddr_t pa);
 115 static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl,
 116     int order);
 117
 118 /*
 119  * Outputs the state of the physical memory allocator, specifically,
 120  * the amount of physical memory in each free list.
 121  */
 122 static int
 123 sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS)
 124 {
 125         struct sbuf sbuf;
 126         struct vm_freelist *fl;
 127         int error, flind, oind, pind;
 128
 129         error = sysctl_wire_old_buffer(req, 0);
 130         if (error != 0)
 131                 return (error);
 132         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 133         for (flind = 0; flind < vm_nfreelists; flind++) {
 134                 sbuf_printf(&sbuf, "\nFREE LIST %d:\n"
 135                     "\n  ORDER (SIZE)  |  NUMBER"
 136                     "\n              ", flind);
 137                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 138                         sbuf_printf(&sbuf, "  |  POOL %d", pind);
 139                 sbuf_printf(&sbuf, "\n--            ");
 140                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 141                         sbuf_printf(&sbuf, "-- --      ");
 142                 sbuf_printf(&sbuf, "--\n");
 143                 for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
 144                         sbuf_printf(&sbuf, "  %2d (%6dK)", oind,
 145                             1 << (PAGE_SHIFT - 10 + oind));
 146                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 147                                 fl = vm_phys_free_queues[flind][pind];
 148                                 sbuf_printf(&sbuf, "  |  %6d", fl[oind].lcnt);
 149                         }
 150                         sbuf_printf(&sbuf, "\n");
 151                 }
 152         }
 153         error = sbuf_finish(&sbuf);
 154         sbuf_delete(&sbuf);
 155         return (error);
 156 }
 157
 158 /*
 159  * Outputs the set of physical memory segments.
 160  */
 161 static int
 162 sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS)
 163 {
 164         struct sbuf sbuf;
 165         struct vm_phys_seg *seg;
 166         int error, segind;
 167
 168         error = sysctl_wire_old_buffer(req, 0);
 169         if (error != 0)
 170                 return (error);
 171         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 172         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 173                 sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind);
 174                 seg = &vm_phys_segs[segind];
 175                 sbuf_printf(&sbuf, "start:     %#jx\n",
 176                     (uintmax_t)seg->start);
 177                 sbuf_printf(&sbuf, "end:       %#jx\n",
 178                     (uintmax_t)seg->end);
 179                 sbuf_printf(&sbuf, "domain:    %d\n", seg->domain);
 180                 sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues);
 181         }
 182         error = sbuf_finish(&sbuf);
 183         sbuf_delete(&sbuf);
 184         return (error);
 185 }
 186
 187 #if VM_NDOMAIN > 1
 188 /*
 189  * Outputs the set of free list lookup lists.
 190  */
 191 static int
 192 sysctl_vm_phys_lookup_lists(SYSCTL_HANDLER_ARGS)
 193 {
 194         struct sbuf sbuf;
 195         int domain, error, flind, ndomains;
 196
 197         error = sysctl_wire_old_buffer(req, 0);
 198         if (error != 0)
 199                 return (error);
 200         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 201         ndomains = vm_nfreelists - VM_NFREELIST + 1;
 202         for (domain = 0; domain < ndomains; domain++) {
 203                 sbuf_printf(&sbuf, "\nDOMAIN %d:\n\n", domain);
 204                 for (flind = 0; flind < vm_nfreelists; flind++)
 205                         sbuf_printf(&sbuf, "  [%d]:\t%p\n", flind,
 206                             vm_phys_lookup_lists[domain][flind]);
 207         }
 208         error = sbuf_finish(&sbuf);
 209         sbuf_delete(&sbuf);
 210         return (error);
 211 }
 212 #endif
 213
 214 /*
 215  * Create a physical memory segment.
 216  */
 217 static void
 218 _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind, int domain)
 219 {
 220         struct vm_phys_seg *seg;
 221 #ifdef VM_PHYSSEG_SPARSE
 222         long pages;
 223         int segind;
 224
 225         pages = 0;
 226         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 227                 seg = &vm_phys_segs[segind];
 228                 pages += atop(seg->end - seg->start);
 229         }
 230 #endif
 231         KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX,
 232             ("vm_phys_create_seg: increase VM_PHYSSEG_MAX"));
 233         seg = &vm_phys_segs[vm_phys_nsegs++];
 234         seg->start = start;
 235         seg->end = end;
 236         seg->domain = domain;
 237 #ifdef VM_PHYSSEG_SPARSE
 238         seg->first_page = &vm_page_array[pages];
 239 #else
 240         seg->first_page = PHYS_TO_VM_PAGE(start);
 241 #endif
 242 #if VM_NDOMAIN > 1
 243         if (flind == VM_FREELIST_DEFAULT && domain != 0) {
 244                 flind = VM_NFREELIST + (domain - 1);
 245                 if (flind >= vm_nfreelists)
 246                         vm_nfreelists = flind + 1;
 247         }
 248 #endif
 249         seg->free_queues = &vm_phys_free_queues[flind];
 250 }
 251
 252 static void
 253 vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind)
 254 {
 255         int i;
 256
 257         if (mem_affinity == NULL) {
 258                 _vm_phys_create_seg(start, end, flind, 0);
 259                 return;
 260         }
 261
 262         for (i = 0;; i++) {
 263                 if (mem_affinity[i].end == 0)
 264                         panic("Reached end of affinity info");
 265                 if (mem_affinity[i].end <= start)
 266                         continue;
 267                 if (mem_affinity[i].start > start)
 268                         panic("No affinity info for start %jx",
 269                             (uintmax_t)start);
 270                 if (mem_affinity[i].end >= end) {
 271                         _vm_phys_create_seg(start, end, flind,
 272                             mem_affinity[i].domain);
 273                         break;
 274                 }
 275                 _vm_phys_create_seg(start, mem_affinity[i].end, flind,
 276                     mem_affinity[i].domain);
 277                 start = mem_affinity[i].end;
 278         }
 279 }
 280
 281 /*
 282  * Initialize the physical memory allocator.
 283  */
 284 void
 285 vm_phys_init(void)
 286 {
 287         struct vm_freelist *fl;
 288         int flind, i, oind, pind;
 289 #if VM_NDOMAIN > 1
 290         int ndomains, j;
 291 #endif
 292
 293         for (i = 0; phys_avail[i + 1] != 0; i += 2) {
 294 #ifdef  VM_FREELIST_ISADMA
 295                 if (phys_avail[i] < 16777216) {
 296                         if (phys_avail[i + 1] > 16777216) {
 297                                 vm_phys_create_seg(phys_avail[i], 16777216,
 298                                     VM_FREELIST_ISADMA);
 299                                 vm_phys_create_seg(16777216, phys_avail[i + 1],
 300                                     VM_FREELIST_DEFAULT);
 301                         } else {
 302                                 vm_phys_create_seg(phys_avail[i],
 303                                     phys_avail[i + 1], VM_FREELIST_ISADMA);
 304                         }
 305                         if (VM_FREELIST_ISADMA >= vm_nfreelists)
 306                                 vm_nfreelists = VM_FREELIST_ISADMA + 1;
 307                 } else
 308 #endif
 309 #ifdef  VM_FREELIST_HIGHMEM
 310                 if (phys_avail[i + 1] > VM_HIGHMEM_ADDRESS) {
 311                         if (phys_avail[i] < VM_HIGHMEM_ADDRESS) {
 312                                 vm_phys_create_seg(phys_avail[i],
 313                                     VM_HIGHMEM_ADDRESS, VM_FREELIST_DEFAULT);
 314                                 vm_phys_create_seg(VM_HIGHMEM_ADDRESS,
 315                                     phys_avail[i + 1], VM_FREELIST_HIGHMEM);
 316                         } else {
 317                                 vm_phys_create_seg(phys_avail[i],
 318                                     phys_avail[i + 1], VM_FREELIST_HIGHMEM);
 319                         }
 320                         if (VM_FREELIST_HIGHMEM >= vm_nfreelists)
 321                                 vm_nfreelists = VM_FREELIST_HIGHMEM + 1;
 322                 } else
 323 #endif
 324                 vm_phys_create_seg(phys_avail[i], phys_avail[i + 1],
 325                     VM_FREELIST_DEFAULT);
 326         }
 327         for (flind = 0; flind < vm_nfreelists; flind++) {
 328                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 329                         fl = vm_phys_free_queues[flind][pind];
 330                         for (oind = 0; oind < VM_NFREEORDER; oind++)
 331                                 TAILQ_INIT(&fl[oind].pl);
 332                 }
 333         }
 334 #if VM_NDOMAIN > 1
 335         /*
 336          * Build a free list lookup list for each domain.  All of the
 337          * memory domain lists are inserted at the VM_FREELIST_DEFAULT
 338          * index in a round-robin order starting with the current
 339          * domain.
 340          */
 341         ndomains = vm_nfreelists - VM_NFREELIST + 1;
 342         for (flind = 0; flind < VM_FREELIST_DEFAULT; flind++)
 343                 for (i = 0; i < ndomains; i++)
 344                         vm_phys_lookup_lists[i][flind] =
 345                             &vm_phys_free_queues[flind];
 346         for (i = 0; i < ndomains; i++)
 347                 for (j = 0; j < ndomains; j++) {
 348                         flind = (i + j) % ndomains;
 349                         if (flind == 0)
 350                                 flind = VM_FREELIST_DEFAULT;
 351                         else
 352                                 flind += VM_NFREELIST - 1;
 353                         vm_phys_lookup_lists[i][VM_FREELIST_DEFAULT + j] =
 354                             &vm_phys_free_queues[flind];
 355                 }
 356         for (flind = VM_FREELIST_DEFAULT + 1; flind < VM_NFREELIST;
 357              flind++)
 358                 for (i = 0; i < ndomains; i++)
 359                         vm_phys_lookup_lists[i][flind + ndomains - 1] =
 360                             &vm_phys_free_queues[flind];
 361 #else
 362         for (flind = 0; flind < vm_nfreelists; flind++)
 363                 vm_phys_lookup_lists[0][flind] = &vm_phys_free_queues[flind];
 364 #endif
 365 }
 366
 367 /*
 368  * Split a contiguous, power of two-sized set of physical pages.
 369  */
 370 static __inline void
 371 vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order)
 372 {
 373         vm_page_t m_buddy;
 374
 375         while (oind > order) {
 376                 oind--;
 377                 m_buddy = &m[1 << oind];
 378                 KASSERT(m_buddy->order == VM_NFREEORDER,
 379                     ("vm_phys_split_pages: page %p has unexpected order %d",
 380                     m_buddy, m_buddy->order));
 381                 m_buddy->order = oind;
 382                 TAILQ_INSERT_HEAD(&fl[oind].pl, m_buddy, pageq);
 383                 fl[oind].lcnt++;
 384         }
 385 }
 386
 387 /*
 388  * Initialize a physical page and add it to the free lists.
 389  */
 390 void
 391 vm_phys_add_page(vm_paddr_t pa)
 392 {
 393         vm_page_t m;
 394
 395         cnt.v_page_count++;
 396         m = vm_phys_paddr_to_vm_page(pa);
 397         m->phys_addr = pa;
 398         m->queue = PQ_NONE;
 399         m->segind = vm_phys_paddr_to_segind(pa);
 400         m->flags = PG_FREE;
 401         KASSERT(m->order == VM_NFREEORDER,
 402             ("vm_phys_add_page: page %p has unexpected order %d",
 403             m, m->order));
 404         m->pool = VM_FREEPOOL_DEFAULT;
 405         pmap_page_init(m);
 406         mtx_lock(&vm_page_queue_free_mtx);
 407         cnt.v_free_count++;
 408         vm_phys_free_pages(m, 0);
 409         mtx_unlock(&vm_page_queue_free_mtx);
 410 }
 411
 412 /*
 413  * Allocate a contiguous, power of two-sized set of physical pages
 414  * from the free lists.
 415  *
 416  * The free page queues must be locked.
 417  */
 418 vm_page_t
 419 vm_phys_alloc_pages(int pool, int order)
 420 {
 421         vm_page_t m;
 422         int flind;
 423
 424         for (flind = 0; flind < vm_nfreelists; flind++) {
 425                 m = vm_phys_alloc_freelist_pages(flind, pool, order);
 426                 if (m != NULL)
 427                         return (m);
 428         }
 429         return (NULL);
 430 }
 431
 432 /*
 433  * Find and dequeue a free page on the given free list, with the
 434  * specified pool and order
 435  */
 436 vm_page_t
 437 vm_phys_alloc_freelist_pages(int flind, int pool, int order)
 438 {
 439         struct vm_freelist *fl;
 440         struct vm_freelist *alt;
 441         int domain, oind, pind;
 442         vm_page_t m;
 443
 444         KASSERT(flind < VM_NFREELIST,
 445             ("vm_phys_alloc_freelist_pages: freelist %d is out of range", flind));
 446         KASSERT(pool < VM_NFREEPOOL,
 447             ("vm_phys_alloc_freelist_pages: pool %d is out of range", pool));
 448         KASSERT(order < VM_NFREEORDER,
 449             ("vm_phys_alloc_freelist_pages: order %d is out of range", order));
 450
 451 #if VM_NDOMAIN > 1
 452         domain = PCPU_GET(domain);
 453 #else
 454         domain = 0;
 455 #endif
 456         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 457         fl = (*vm_phys_lookup_lists[domain][flind])[pool];
 458         for (oind = order; oind < VM_NFREEORDER; oind++) {
 459                 m = TAILQ_FIRST(&fl[oind].pl);
 460                 if (m != NULL) {
 461                         TAILQ_REMOVE(&fl[oind].pl, m, pageq);
 462                         fl[oind].lcnt--;
 463                         m->order = VM_NFREEORDER;
 464                         vm_phys_split_pages(m, oind, fl, order);
 465                         return (m);
 466                 }
 467         }
 468
 469         /*
 470          * The given pool was empty.  Find the largest
 471          * contiguous, power-of-two-sized set of pages in any
 472          * pool.  Transfer these pages to the given pool, and
 473          * use them to satisfy the allocation.
 474          */
 475         for (oind = VM_NFREEORDER - 1; oind >= order; oind--) {
 476                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 477                         alt = (*vm_phys_lookup_lists[domain][flind])[pind];
 478                         m = TAILQ_FIRST(&alt[oind].pl);
 479                         if (m != NULL) {
 480                                 TAILQ_REMOVE(&alt[oind].pl, m, pageq);
 481                                 alt[oind].lcnt--;
 482                                 m->order = VM_NFREEORDER;
 483                                 vm_phys_set_pool(pool, m, oind);
 484                                 vm_phys_split_pages(m, oind, fl, order);
 485                                 return (m);
 486                         }
 487                 }
 488         }
 489         return (NULL);
 490 }
 491
 492 /*
 493  * Allocate physical memory from phys_avail[].
 494  */
 495 vm_paddr_t
 496 vm_phys_bootstrap_alloc(vm_size_t size, unsigned long alignment)
 497 {
 498         vm_paddr_t pa;
 499         int i;
 500
 501         size = round_page(size);
 502         for (i = 0; phys_avail[i + 1] != 0; i += 2) {
 503                 if (phys_avail[i + 1] - phys_avail[i] < size)
 504                         continue;
 505                 pa = phys_avail[i];
 506                 phys_avail[i] += size;
 507                 return (pa);
 508         }
 509         panic("vm_phys_bootstrap_alloc");
 510 }
 511
 512 /*
 513  * Find the vm_page corresponding to the given physical address.
 514  */
 515 vm_page_t
 516 vm_phys_paddr_to_vm_page(vm_paddr_t pa)
 517 {
 518         struct vm_phys_seg *seg;
 519         int segind;
 520
 521         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 522                 seg = &vm_phys_segs[segind];
 523                 if (pa >= seg->start && pa < seg->end)
 524                         return (&seg->first_page[atop(pa - seg->start)]);
 525         }
 526         return (NULL);
 527 }
 528
 529 /*
 530  * Find the segment containing the given physical address.
 531  */
 532 static int
 533 vm_phys_paddr_to_segind(vm_paddr_t pa)
 534 {
 535         struct vm_phys_seg *seg;
 536         int segind;
 537
 538         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 539                 seg = &vm_phys_segs[segind];
 540                 if (pa >= seg->start && pa < seg->end)
 541                         return (segind);
 542         }
 543         panic("vm_phys_paddr_to_segind: paddr %#jx is not in any segment" ,
 544             (uintmax_t)pa);
 545 }
 546
 547 /*
 548  * Free a contiguous, power of two-sized set of physical pages.
 549  *
 550  * The free page queues must be locked.
 551  */
 552 void
 553 vm_phys_free_pages(vm_page_t m, int order)
 554 {
 555         struct vm_freelist *fl;
 556         struct vm_phys_seg *seg;
 557         vm_paddr_t pa, pa_buddy;
 558         vm_page_t m_buddy;
 559
 560         KASSERT(m->order == VM_NFREEORDER,
 561             ("vm_phys_free_pages: page %p has unexpected order %d",
 562             m, m->order));
 563         KASSERT(m->pool < VM_NFREEPOOL,
 564             ("vm_phys_free_pages: page %p has unexpected pool %d",
 565             m, m->pool));
 566         KASSERT(order < VM_NFREEORDER,
 567             ("vm_phys_free_pages: order %d is out of range", order));
 568         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 569         pa = VM_PAGE_TO_PHYS(m);
 570         seg = &vm_phys_segs[m->segind];
 571         while (order < VM_NFREEORDER - 1) {
 572                 pa_buddy = pa ^ (1 << (PAGE_SHIFT + order));
 573                 if (pa_buddy < seg->start ||
 574                     pa_buddy >= seg->end)
 575                         break;
 576                 m_buddy = &seg->first_page[atop(pa_buddy - seg->start)];
 577                 if (m_buddy->order != order)
 578                         break;
 579                 fl = (*seg->free_queues)[m_buddy->pool];
 580                 TAILQ_REMOVE(&fl[m_buddy->order].pl, m_buddy, pageq);
 581                 fl[m_buddy->order].lcnt--;
 582                 m_buddy->order = VM_NFREEORDER;
 583                 if (m_buddy->pool != m->pool)
 584                         vm_phys_set_pool(m->pool, m_buddy, order);
 585                 order++;
 586                 pa &= ~((1 << (PAGE_SHIFT + order)) - 1);
 587                 m = &seg->first_page[atop(pa - seg->start)];
 588         }
 589         m->order = order;
 590         fl = (*seg->free_queues)[m->pool];
 591         TAILQ_INSERT_TAIL(&fl[order].pl, m, pageq);
 592         fl[order].lcnt++;
 593 }
 594
 595 /*
 596  * Set the pool for a contiguous, power of two-sized set of physical pages.
 597  */
 598 void
 599 vm_phys_set_pool(int pool, vm_page_t m, int order)
 600 {
 601         vm_page_t m_tmp;
 602
 603         for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++)
 604                 m_tmp->pool = pool;
 605 }
 606
 607 /*
 608  * Search for the given physical page "m" in the free lists.  If the search
 609  * succeeds, remove "m" from the free lists and return TRUE.  Otherwise, return
 610  * FALSE, indicating that "m" is not in the free lists.
 611  *
 612  * The free page queues must be locked.
 613  */
 614 boolean_t
 615 vm_phys_unfree_page(vm_page_t m)
 616 {
 617         struct vm_freelist *fl;
 618         struct vm_phys_seg *seg;
 619         vm_paddr_t pa, pa_half;
 620         vm_page_t m_set, m_tmp;
 621         int order;
 622
 623         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 624
 625         /*
 626          * First, find the contiguous, power of two-sized set of free
 627          * physical pages containing the given physical page "m" and
 628          * assign it to "m_set".
 629          */
 630         seg = &vm_phys_segs[m->segind];
 631         for (m_set = m, order = 0; m_set->order == VM_NFREEORDER &&
 632             order < VM_NFREEORDER - 1; ) {
 633                 order++;
 634                 pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order));
 635                 if (pa >= seg->start)
 636                         m_set = &seg->first_page[atop(pa - seg->start)];
 637                 else
 638                         return (FALSE);
 639         }
 640         if (m_set->order < order)
 641                 return (FALSE);
 642         if (m_set->order == VM_NFREEORDER)
 643                 return (FALSE);
 644         KASSERT(m_set->order < VM_NFREEORDER,
 645             ("vm_phys_unfree_page: page %p has unexpected order %d",
 646             m_set, m_set->order));
 647
 648         /*
 649          * Next, remove "m_set" from the free lists.  Finally, extract
 650          * "m" from "m_set" using an iterative algorithm: While "m_set"
 651          * is larger than a page, shrink "m_set" by returning the half
 652          * of "m_set" that does not contain "m" to the free lists.
 653          */
 654         fl = (*seg->free_queues)[m_set->pool];
 655         order = m_set->order;
 656         TAILQ_REMOVE(&fl[order].pl, m_set, pageq);
 657         fl[order].lcnt--;
 658         m_set->order = VM_NFREEORDER;
 659         while (order > 0) {
 660                 order--;
 661                 pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order));
 662                 if (m->phys_addr < pa_half)
 663                         m_tmp = &seg->first_page[atop(pa_half - seg->start)];
 664                 else {
 665                         m_tmp = m_set;
 666                         m_set = &seg->first_page[atop(pa_half - seg->start)];
 667                 }
 668                 m_tmp->order = order;
 669                 TAILQ_INSERT_HEAD(&fl[order].pl, m_tmp, pageq);
 670                 fl[order].lcnt++;
 671         }
 672         KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency"));
 673         return (TRUE);
 674 }
 675
 676 /*
 677  * Try to zero one physical page.  Used by an idle priority thread.
 678  */
 679 boolean_t
 680 vm_phys_zero_pages_idle(void)
 681 {
 682         static struct vm_freelist *fl = vm_phys_free_queues[0][0];
 683         static int flind, oind, pind;
 684         vm_page_t m, m_tmp;
 685
 686         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 687         for (;;) {
 688                 TAILQ_FOREACH_REVERSE(m, &fl[oind].pl, pglist, pageq) {
 689                         for (m_tmp = m; m_tmp < &m[1 << oind]; m_tmp++) {
 690                                 if ((m_tmp->flags & (PG_CACHED | PG_ZERO)) == 0) {
 691                                         vm_phys_unfree_page(m_tmp);
 692                                         cnt.v_free_count--;
 693                                         mtx_unlock(&vm_page_queue_free_mtx);
 694                                         pmap_zero_page_idle(m_tmp);
 695                                         m_tmp->flags |= PG_ZERO;
 696                                         mtx_lock(&vm_page_queue_free_mtx);
 697                                         cnt.v_free_count++;
 698                                         vm_phys_free_pages(m_tmp, 0);
 699                                         vm_page_zero_count++;
 700                                         cnt_prezero++;
 701                                         return (TRUE);
 702                                 }
 703                         }
 704                 }
 705                 oind++;
 706                 if (oind == VM_NFREEORDER) {
 707                         oind = 0;
 708                         pind++;
 709                         if (pind == VM_NFREEPOOL) {
 710                                 pind = 0;
 711                                 flind++;
 712                                 if (flind == vm_nfreelists)
 713                                         flind = 0;
 714                         }
 715                         fl = vm_phys_free_queues[flind][pind];
 716                 }
 717         }
 718 }
 719
 720 /*
 721  * Allocate a contiguous set of physical pages of the given size
 722  * "npages" from the free lists.  All of the physical pages must be at
 723  * or above the given physical address "low" and below the given
 724  * physical address "high".  The given value "alignment" determines the
 725  * alignment of the first physical page in the set.  If the given value
 726  * "boundary" is non-zero, then the set of physical pages cannot cross
 727  * any physical address boundary that is a multiple of that value.  Both
 728  * "alignment" and "boundary" must be a power of two.
 729  */
 730 vm_page_t
 731 vm_phys_alloc_contig(unsigned long npages, vm_paddr_t low, vm_paddr_t high,
 732     unsigned long alignment, unsigned long boundary)
 733 {
 734         struct vm_freelist *fl;
 735         struct vm_phys_seg *seg;
 736         struct vnode *vp;
 737         vm_paddr_t pa, pa_last, size;
 738         vm_page_t deferred_vdrop_list, m, m_ret;
 739         int domain, flind, i, oind, order, pind;
 740
 741 #if VM_NDOMAIN > 1
 742         domain = PCPU_GET(domain);
 743 #else
 744         domain = 0;
 745 #endif
 746         size = npages << PAGE_SHIFT;
 747         KASSERT(size != 0,
 748             ("vm_phys_alloc_contig: size must not be 0"));
 749         KASSERT((alignment & (alignment - 1)) == 0,
 750             ("vm_phys_alloc_contig: alignment must be a power of 2"));
 751         KASSERT((boundary & (boundary - 1)) == 0,
 752             ("vm_phys_alloc_contig: boundary must be a power of 2"));
 753         deferred_vdrop_list = NULL;
 754         /* Compute the queue that is the best fit for npages. */
 755         for (order = 0; (1 << order) < npages; order++);
 756         mtx_lock(&vm_page_queue_free_mtx);
 757 #if VM_NRESERVLEVEL > 0
 758 retry:
 759 #endif
 760         for (flind = 0; flind < vm_nfreelists; flind++) {
 761                 for (oind = min(order, VM_NFREEORDER - 1); oind < VM_NFREEORDER; oind++) {
 762                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 763                                 fl = (*vm_phys_lookup_lists[domain][flind])
 764                                     [pind];
 765                                 TAILQ_FOREACH(m_ret, &fl[oind].pl, pageq) {
 766                                         /*
 767                                          * A free list may contain physical pages
 768                                          * from one or more segments.
 769                                          */
 770                                         seg = &vm_phys_segs[m_ret->segind];
 771                                         if (seg->start > high ||
 772                                             low >= seg->end)
 773                                                 continue;
 774
 775                                         /*
 776                                          * Is the size of this allocation request
 777                                          * larger than the largest block size?
 778                                          */
 779                                         if (order >= VM_NFREEORDER) {
 780                                                 /*
 781                                                  * Determine if a sufficient number
 782                                                  * of subsequent blocks to satisfy
 783                                                  * the allocation request are free.
 784                                                  */
 785                                                 pa = VM_PAGE_TO_PHYS(m_ret);
 786                                                 pa_last = pa + size;
 787                                                 for (;;) {
 788                                                         pa += 1 << (PAGE_SHIFT + VM_NFREEORDER - 1);
 789                                                         if (pa >= pa_last)
 790                                                                 break;
 791                                                         if (pa < seg->start ||
 792                                                             pa >= seg->end)
 793                                                                 break;
 794                                                         m = &seg->first_page[atop(pa - seg->start)];
 795                                                         if (m->order != VM_NFREEORDER - 1)
 796                                                                 break;
 797                                                 }
 798                                                 /* If not, continue to the next block. */
 799                                                 if (pa < pa_last)
 800                                                         continue;
 801                                         }
 802
 803                                         /*
 804                                          * Determine if the blocks are within the given range,
 805                                          * satisfy the given alignment, and do not cross the
 806                                          * given boundary.
 807                                          */
 808                                         pa = VM_PAGE_TO_PHYS(m_ret);
 809                                         if (pa >= low &&
 810                                             pa + size <= high &&
 811                                             (pa & (alignment - 1)) == 0 &&
 812                                             ((pa ^ (pa + size - 1)) & ~(boundary - 1)) == 0)
 813                                                 goto done;
 814                                 }
 815                         }
 816                 }
 817         }
 818 #if VM_NRESERVLEVEL > 0
 819         if (vm_reserv_reclaim_contig(size, low, high, alignment, boundary))
 820                 goto retry;
 821 #endif
 822         mtx_unlock(&vm_page_queue_free_mtx);
 823         return (NULL);
 824 done:
 825         for (m = m_ret; m < &m_ret[npages]; m = &m[1 << oind]) {
 826                 fl = (*seg->free_queues)[m->pool];
 827                 TAILQ_REMOVE(&fl[m->order].pl, m, pageq);
 828                 fl[m->order].lcnt--;
 829                 m->order = VM_NFREEORDER;
 830         }
 831         if (m_ret->pool != VM_FREEPOOL_DEFAULT)
 832                 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m_ret, oind);
 833         fl = (*seg->free_queues)[m_ret->pool];
 834         vm_phys_split_pages(m_ret, oind, fl, order);
 835         for (i = 0; i < npages; i++) {
 836                 m = &m_ret[i];
 837                 vp = vm_page_alloc_init(m);
 838                 if (vp != NULL) {
 839                         /*
 840                          * Enqueue the vnode for deferred vdrop().
 841                          *
 842                          * Unmanaged pages don't use "pageq", so it
 843                          * can be safely abused to construct a short-
 844                          * lived queue of vnodes.
 845                          */
 846                         m->pageq.tqe_prev = (void *)vp;
 847                         m->pageq.tqe_next = deferred_vdrop_list;
 848                         deferred_vdrop_list = m;
 849                 }
 850         }
 851         for (; i < roundup2(npages, 1 << imin(oind, order)); i++) {
 852                 m = &m_ret[i];
 853                 KASSERT(m->order == VM_NFREEORDER,
 854                     ("vm_phys_alloc_contig: page %p has unexpected order %d",
 855                     m, m->order));
 856                 vm_phys_free_pages(m, 0);
 857         }
 858         mtx_unlock(&vm_page_queue_free_mtx);
 859         while (deferred_vdrop_list != NULL) {
 860                 vdrop((struct vnode *)deferred_vdrop_list->pageq.tqe_prev);
 861                 deferred_vdrop_list = deferred_vdrop_list->pageq.tqe_next;
 862         }
 863         return (m_ret);
 864 }
 865
 866 #ifdef DDB
 867 /*
 868  * Show the number of physical pages in each of the free lists.
 869  */
 870 DB_SHOW_COMMAND(freepages, db_show_freepages)
 871 {
 872         struct vm_freelist *fl;
 873         int flind, oind, pind;
 874
 875         for (flind = 0; flind < vm_nfreelists; flind++) {
 876                 db_printf("FREE LIST %d:\n"
 877                     "\n  ORDER (SIZE)  |  NUMBER"
 878                     "\n              ", flind);
 879                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 880                         db_printf("  |  POOL %d", pind);
 881                 db_printf("\n--            ");
 882                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 883                         db_printf("-- --      ");
 884                 db_printf("--\n");
 885                 for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
 886                         db_printf("  %2.2d (%6.6dK)", oind,
 887                             1 << (PAGE_SHIFT - 10 + oind));
 888                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 889                                 fl = vm_phys_free_queues[flind][pind];
 890                                 db_printf("  |  %6.6d", fl[oind].lcnt);
 891                         }
 892                         db_printf("\n");
 893                 }
 894                 db_printf("\n");
 895         }
 896 }
 897 #endif