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