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 #define VM_PHYS_FICTITIOUS_NSEGS        8
  87 static struct vm_phys_fictitious_seg {
  88         vm_paddr_t      start;
  89         vm_paddr_t      end;
  90         vm_page_t       first_page;
  91 } vm_phys_fictitious_segs[VM_PHYS_FICTITIOUS_NSEGS];
  92 static struct mtx vm_phys_fictitious_reg_mtx;
  93 MALLOC_DEFINE(M_FICT_PAGES, "", "");
  94
  95 static struct vm_freelist
  96     vm_phys_free_queues[VM_RAW_NFREELIST][VM_NFREEPOOL][VM_NFREEORDER];
  97 static struct vm_freelist
  98 (*vm_phys_lookup_lists[VM_NDOMAIN][VM_RAW_NFREELIST])[VM_NFREEPOOL][VM_NFREEORDER];
  99
 100 static int vm_nfreelists = VM_FREELIST_DEFAULT + 1;
 101
 102 static int cnt_prezero;
 103 SYSCTL_INT(_vm_stats_misc, OID_AUTO, cnt_prezero, CTLFLAG_RD,
 104     &cnt_prezero, 0, "The number of physical pages prezeroed at idle time");
 105
 106 static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS);
 107 SYSCTL_OID(_vm, OID_AUTO, phys_free, CTLTYPE_STRING | CTLFLAG_RD,
 108     NULL, 0, sysctl_vm_phys_free, "A", "Phys Free Info");
 109
 110 static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS);
 111 SYSCTL_OID(_vm, OID_AUTO, phys_segs, CTLTYPE_STRING | CTLFLAG_RD,
 112     NULL, 0, sysctl_vm_phys_segs, "A", "Phys Seg Info");
 113
 114 #if VM_NDOMAIN > 1
 115 static int sysctl_vm_phys_lookup_lists(SYSCTL_HANDLER_ARGS);
 116 SYSCTL_OID(_vm, OID_AUTO, phys_lookup_lists, CTLTYPE_STRING | CTLFLAG_RD,
 117     NULL, 0, sysctl_vm_phys_lookup_lists, "A", "Phys Lookup Lists");
 118 #endif
 119
 120 static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind,
 121     int domain);
 122 static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind);
 123 static int vm_phys_paddr_to_segind(vm_paddr_t pa);
 124 static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl,
 125     int order);
 126
 127 /*
 128  * Outputs the state of the physical memory allocator, specifically,
 129  * the amount of physical memory in each free list.
 130  */
 131 static int
 132 sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS)
 133 {
 134         struct sbuf sbuf;
 135         struct vm_freelist *fl;
 136         int error, flind, oind, pind;
 137
 138         error = sysctl_wire_old_buffer(req, 0);
 139         if (error != 0)
 140                 return (error);
 141         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 142         for (flind = 0; flind < vm_nfreelists; flind++) {
 143                 sbuf_printf(&sbuf, "\nFREE LIST %d:\n"
 144                     "\n  ORDER (SIZE)  |  NUMBER"
 145                     "\n              ", flind);
 146                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 147                         sbuf_printf(&sbuf, "  |  POOL %d", pind);
 148                 sbuf_printf(&sbuf, "\n--            ");
 149                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 150                         sbuf_printf(&sbuf, "-- --      ");
 151                 sbuf_printf(&sbuf, "--\n");
 152                 for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
 153                         sbuf_printf(&sbuf, "  %2d (%6dK)", oind,
 154                             1 << (PAGE_SHIFT - 10 + oind));
 155                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 156                                 fl = vm_phys_free_queues[flind][pind];
 157                                 sbuf_printf(&sbuf, "  |  %6d", fl[oind].lcnt);
 158                         }
 159                         sbuf_printf(&sbuf, "\n");
 160                 }
 161         }
 162         error = sbuf_finish(&sbuf);
 163         sbuf_delete(&sbuf);
 164         return (error);
 165 }
 166
 167 /*
 168  * Outputs the set of physical memory segments.
 169  */
 170 static int
 171 sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS)
 172 {
 173         struct sbuf sbuf;
 174         struct vm_phys_seg *seg;
 175         int error, segind;
 176
 177         error = sysctl_wire_old_buffer(req, 0);
 178         if (error != 0)
 179                 return (error);
 180         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 181         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 182                 sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind);
 183                 seg = &vm_phys_segs[segind];
 184                 sbuf_printf(&sbuf, "start:     %#jx\n",
 185                     (uintmax_t)seg->start);
 186                 sbuf_printf(&sbuf, "end:       %#jx\n",
 187                     (uintmax_t)seg->end);
 188                 sbuf_printf(&sbuf, "domain:    %d\n", seg->domain);
 189                 sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues);
 190         }
 191         error = sbuf_finish(&sbuf);
 192         sbuf_delete(&sbuf);
 193         return (error);
 194 }
 195
 196 #if VM_NDOMAIN > 1
 197 /*
 198  * Outputs the set of free list lookup lists.
 199  */
 200 static int
 201 sysctl_vm_phys_lookup_lists(SYSCTL_HANDLER_ARGS)
 202 {
 203         struct sbuf sbuf;
 204         int domain, error, flind, ndomains;
 205
 206         error = sysctl_wire_old_buffer(req, 0);
 207         if (error != 0)
 208                 return (error);
 209         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 210         ndomains = vm_nfreelists - VM_NFREELIST + 1;
 211         for (domain = 0; domain < ndomains; domain++) {
 212                 sbuf_printf(&sbuf, "\nDOMAIN %d:\n\n", domain);
 213                 for (flind = 0; flind < vm_nfreelists; flind++)
 214                         sbuf_printf(&sbuf, "  [%d]:\t%p\n", flind,
 215                             vm_phys_lookup_lists[domain][flind]);
 216         }
 217         error = sbuf_finish(&sbuf);
 218         sbuf_delete(&sbuf);
 219         return (error);
 220 }
 221 #endif
 222
 223 /*
 224  * Create a physical memory segment.
 225  */
 226 static void
 227 _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind, int domain)
 228 {
 229         struct vm_phys_seg *seg;
 230 #ifdef VM_PHYSSEG_SPARSE
 231         long pages;
 232         int segind;
 233
 234         pages = 0;
 235         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 236                 seg = &vm_phys_segs[segind];
 237                 pages += atop(seg->end - seg->start);
 238         }
 239 #endif
 240         KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX,
 241             ("vm_phys_create_seg: increase VM_PHYSSEG_MAX"));
 242         seg = &vm_phys_segs[vm_phys_nsegs++];
 243         seg->start = start;
 244         seg->end = end;
 245         seg->domain = domain;
 246 #ifdef VM_PHYSSEG_SPARSE
 247         seg->first_page = &vm_page_array[pages];
 248 #else
 249         seg->first_page = PHYS_TO_VM_PAGE(start);
 250 #endif
 251 #if VM_NDOMAIN > 1
 252         if (flind == VM_FREELIST_DEFAULT && domain != 0) {
 253                 flind = VM_NFREELIST + (domain - 1);
 254                 if (flind >= vm_nfreelists)
 255                         vm_nfreelists = flind + 1;
 256         }
 257 #endif
 258         seg->free_queues = &vm_phys_free_queues[flind];
 259 }
 260
 261 static void
 262 vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int flind)
 263 {
 264         int i;
 265
 266         if (mem_affinity == NULL) {
 267                 _vm_phys_create_seg(start, end, flind, 0);
 268                 return;
 269         }
 270
 271         for (i = 0;; i++) {
 272                 if (mem_affinity[i].end == 0)
 273                         panic("Reached end of affinity info");
 274                 if (mem_affinity[i].end <= start)
 275                         continue;
 276                 if (mem_affinity[i].start > start)
 277                         panic("No affinity info for start %jx",
 278                             (uintmax_t)start);
 279                 if (mem_affinity[i].end >= end) {
 280                         _vm_phys_create_seg(start, end, flind,
 281                             mem_affinity[i].domain);
 282                         break;
 283                 }
 284                 _vm_phys_create_seg(start, mem_affinity[i].end, flind,
 285                     mem_affinity[i].domain);
 286                 start = mem_affinity[i].end;
 287         }
 288 }
 289
 290 /*
 291  * Initialize the physical memory allocator.
 292  */
 293 void
 294 vm_phys_init(void)
 295 {
 296         struct vm_freelist *fl;
 297         int flind, i, oind, pind;
 298 #if VM_NDOMAIN > 1
 299         int ndomains, j;
 300 #endif
 301
 302         for (i = 0; phys_avail[i + 1] != 0; i += 2) {
 303 #ifdef  VM_FREELIST_ISADMA
 304                 if (phys_avail[i] < 16777216) {
 305                         if (phys_avail[i + 1] > 16777216) {
 306                                 vm_phys_create_seg(phys_avail[i], 16777216,
 307                                     VM_FREELIST_ISADMA);
 308                                 vm_phys_create_seg(16777216, phys_avail[i + 1],
 309                                     VM_FREELIST_DEFAULT);
 310                         } else {
 311                                 vm_phys_create_seg(phys_avail[i],
 312                                     phys_avail[i + 1], VM_FREELIST_ISADMA);
 313                         }
 314                         if (VM_FREELIST_ISADMA >= vm_nfreelists)
 315                                 vm_nfreelists = VM_FREELIST_ISADMA + 1;
 316                 } else
 317 #endif
 318 #ifdef  VM_FREELIST_HIGHMEM
 319                 if (phys_avail[i + 1] > VM_HIGHMEM_ADDRESS) {
 320                         if (phys_avail[i] < VM_HIGHMEM_ADDRESS) {
 321                                 vm_phys_create_seg(phys_avail[i],
 322                                     VM_HIGHMEM_ADDRESS, VM_FREELIST_DEFAULT);
 323                                 vm_phys_create_seg(VM_HIGHMEM_ADDRESS,
 324                                     phys_avail[i + 1], VM_FREELIST_HIGHMEM);
 325                         } else {
 326                                 vm_phys_create_seg(phys_avail[i],
 327                                     phys_avail[i + 1], VM_FREELIST_HIGHMEM);
 328                         }
 329                         if (VM_FREELIST_HIGHMEM >= vm_nfreelists)
 330                                 vm_nfreelists = VM_FREELIST_HIGHMEM + 1;
 331                 } else
 332 #endif
 333                 vm_phys_create_seg(phys_avail[i], phys_avail[i + 1],
 334                     VM_FREELIST_DEFAULT);
 335         }
 336         for (flind = 0; flind < vm_nfreelists; flind++) {
 337                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 338                         fl = vm_phys_free_queues[flind][pind];
 339                         for (oind = 0; oind < VM_NFREEORDER; oind++)
 340                                 TAILQ_INIT(&fl[oind].pl);
 341                 }
 342         }
 343 #if VM_NDOMAIN > 1
 344         /*
 345          * Build a free list lookup list for each domain.  All of the
 346          * memory domain lists are inserted at the VM_FREELIST_DEFAULT
 347          * index in a round-robin order starting with the current
 348          * domain.
 349          */
 350         ndomains = vm_nfreelists - VM_NFREELIST + 1;
 351         for (flind = 0; flind < VM_FREELIST_DEFAULT; flind++)
 352                 for (i = 0; i < ndomains; i++)
 353                         vm_phys_lookup_lists[i][flind] =
 354                             &vm_phys_free_queues[flind];
 355         for (i = 0; i < ndomains; i++)
 356                 for (j = 0; j < ndomains; j++) {
 357                         flind = (i + j) % ndomains;
 358                         if (flind == 0)
 359                                 flind = VM_FREELIST_DEFAULT;
 360                         else
 361                                 flind += VM_NFREELIST - 1;
 362                         vm_phys_lookup_lists[i][VM_FREELIST_DEFAULT + j] =
 363                             &vm_phys_free_queues[flind];
 364                 }
 365         for (flind = VM_FREELIST_DEFAULT + 1; flind < VM_NFREELIST;
 366              flind++)
 367                 for (i = 0; i < ndomains; i++)
 368                         vm_phys_lookup_lists[i][flind + ndomains - 1] =
 369                             &vm_phys_free_queues[flind];
 370 #else
 371         for (flind = 0; flind < vm_nfreelists; flind++)
 372                 vm_phys_lookup_lists[0][flind] = &vm_phys_free_queues[flind];
 373 #endif
 374
 375         mtx_init(&vm_phys_fictitious_reg_mtx, "vmfctr", NULL, MTX_DEF);
 376 }
 377
 378 /*
 379  * Split a contiguous, power of two-sized set of physical pages.
 380  */
 381 static __inline void
 382 vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order)
 383 {
 384         vm_page_t m_buddy;
 385
 386         while (oind > order) {
 387                 oind--;
 388                 m_buddy = &m[1 << oind];
 389                 KASSERT(m_buddy->order == VM_NFREEORDER,
 390                     ("vm_phys_split_pages: page %p has unexpected order %d",
 391                     m_buddy, m_buddy->order));
 392                 m_buddy->order = oind;
 393                 TAILQ_INSERT_HEAD(&fl[oind].pl, m_buddy, pageq);
 394                 fl[oind].lcnt++;
 395         }
 396 }
 397
 398 /*
 399  * Initialize a physical page and add it to the free lists.
 400  */
 401 void
 402 vm_phys_add_page(vm_paddr_t pa)
 403 {
 404         vm_page_t m;
 405
 406         cnt.v_page_count++;
 407         m = vm_phys_paddr_to_vm_page(pa);
 408         m->phys_addr = pa;
 409         m->queue = PQ_NONE;
 410         m->segind = vm_phys_paddr_to_segind(pa);
 411         m->flags = PG_FREE;
 412         KASSERT(m->order == VM_NFREEORDER,
 413             ("vm_phys_add_page: page %p has unexpected order %d",
 414             m, m->order));
 415         m->pool = VM_FREEPOOL_DEFAULT;
 416         pmap_page_init(m);
 417         mtx_lock(&vm_page_queue_free_mtx);
 418         cnt.v_free_count++;
 419         vm_phys_free_pages(m, 0);
 420         mtx_unlock(&vm_page_queue_free_mtx);
 421 }
 422
 423 /*
 424  * Allocate a contiguous, power of two-sized set of physical pages
 425  * from the free lists.
 426  *
 427  * The free page queues must be locked.
 428  */
 429 vm_page_t
 430 vm_phys_alloc_pages(int pool, int order)
 431 {
 432         vm_page_t m;
 433         int flind;
 434
 435         for (flind = 0; flind < vm_nfreelists; flind++) {
 436                 m = vm_phys_alloc_freelist_pages(flind, pool, order);
 437                 if (m != NULL)
 438                         return (m);
 439         }
 440         return (NULL);
 441 }
 442
 443 /*
 444  * Find and dequeue a free page on the given free list, with the
 445  * specified pool and order
 446  */
 447 vm_page_t
 448 vm_phys_alloc_freelist_pages(int flind, int pool, int order)
 449 {
 450         struct vm_freelist *fl;
 451         struct vm_freelist *alt;
 452         int domain, oind, pind;
 453         vm_page_t m;
 454
 455         KASSERT(flind < VM_NFREELIST,
 456             ("vm_phys_alloc_freelist_pages: freelist %d is out of range", flind));
 457         KASSERT(pool < VM_NFREEPOOL,
 458             ("vm_phys_alloc_freelist_pages: pool %d is out of range", pool));
 459         KASSERT(order < VM_NFREEORDER,
 460             ("vm_phys_alloc_freelist_pages: order %d is out of range", order));
 461
 462 #if VM_NDOMAIN > 1
 463         domain = PCPU_GET(domain);
 464 #else
 465         domain = 0;
 466 #endif
 467         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 468         fl = (*vm_phys_lookup_lists[domain][flind])[pool];
 469         for (oind = order; oind < VM_NFREEORDER; oind++) {
 470                 m = TAILQ_FIRST(&fl[oind].pl);
 471                 if (m != NULL) {
 472                         TAILQ_REMOVE(&fl[oind].pl, m, pageq);
 473                         fl[oind].lcnt--;
 474                         m->order = VM_NFREEORDER;
 475                         vm_phys_split_pages(m, oind, fl, order);
 476                         return (m);
 477                 }
 478         }
 479
 480         /*
 481          * The given pool was empty.  Find the largest
 482          * contiguous, power-of-two-sized set of pages in any
 483          * pool.  Transfer these pages to the given pool, and
 484          * use them to satisfy the allocation.
 485          */
 486         for (oind = VM_NFREEORDER - 1; oind >= order; oind--) {
 487                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 488                         alt = (*vm_phys_lookup_lists[domain][flind])[pind];
 489                         m = TAILQ_FIRST(&alt[oind].pl);
 490                         if (m != NULL) {
 491                                 TAILQ_REMOVE(&alt[oind].pl, m, pageq);
 492                                 alt[oind].lcnt--;
 493                                 m->order = VM_NFREEORDER;
 494                                 vm_phys_set_pool(pool, m, oind);
 495                                 vm_phys_split_pages(m, oind, fl, order);
 496                                 return (m);
 497                         }
 498                 }
 499         }
 500         return (NULL);
 501 }
 502
 503 /*
 504  * Allocate physical memory from phys_avail[].
 505  */
 506 vm_paddr_t
 507 vm_phys_bootstrap_alloc(vm_size_t size, unsigned long alignment)
 508 {
 509         vm_paddr_t pa;
 510         int i;
 511
 512         size = round_page(size);
 513         for (i = 0; phys_avail[i + 1] != 0; i += 2) {
 514                 if (phys_avail[i + 1] - phys_avail[i] < size)
 515                         continue;
 516                 pa = phys_avail[i];
 517                 phys_avail[i] += size;
 518                 return (pa);
 519         }
 520         panic("vm_phys_bootstrap_alloc");
 521 }
 522
 523 /*
 524  * Find the vm_page corresponding to the given physical address.
 525  */
 526 vm_page_t
 527 vm_phys_paddr_to_vm_page(vm_paddr_t pa)
 528 {
 529         struct vm_phys_seg *seg;
 530         int segind;
 531
 532         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 533                 seg = &vm_phys_segs[segind];
 534                 if (pa >= seg->start && pa < seg->end)
 535                         return (&seg->first_page[atop(pa - seg->start)]);
 536         }
 537         return (NULL);
 538 }
 539
 540 vm_page_t
 541 vm_phys_fictitious_to_vm_page(vm_paddr_t pa)
 542 {
 543         struct vm_phys_fictitious_seg *seg;
 544         vm_page_t m;
 545         int segind;
 546
 547         m = NULL;
 548         for (segind = 0; segind < VM_PHYS_FICTITIOUS_NSEGS; segind++) {
 549                 seg = &vm_phys_fictitious_segs[segind];
 550                 if (pa >= seg->start && pa < seg->end) {
 551                         m = &seg->first_page[atop(pa - seg->start)];
 552                         KASSERT((m->flags & PG_FICTITIOUS) != 0,
 553                             ("%p not fictitious", m));
 554                         break;
 555                 }
 556         }
 557         return (m);
 558 }
 559
 560 int
 561 vm_phys_fictitious_reg_range(vm_paddr_t start, vm_paddr_t end,
 562     vm_memattr_t memattr)
 563 {
 564         struct vm_phys_fictitious_seg *seg;
 565         vm_page_t fp;
 566         long i, page_count;
 567         int segind;
 568 #ifdef VM_PHYSSEG_DENSE
 569         long pi;
 570         boolean_t malloced;
 571 #endif
 572
 573         page_count = (end - start) / PAGE_SIZE;
 574
 575 #ifdef VM_PHYSSEG_DENSE
 576         pi = atop(start);
 577         if (pi >= first_page && atop(end) < vm_page_array_size) {
 578                 fp = &vm_page_array[pi - first_page];
 579                 malloced = FALSE;
 580         } else
 581 #endif
 582         {
 583                 fp = malloc(page_count * sizeof(struct vm_page), M_FICT_PAGES,
 584                     M_WAITOK | M_ZERO);
 585 #ifdef VM_PHYSSEG_DENSE
 586                 malloced = TRUE;
 587 #endif
 588         }
 589         for (i = 0; i < page_count; i++) {
 590                 vm_page_initfake(&fp[i], start + PAGE_SIZE * i, memattr);
 591                 pmap_page_init(&fp[i]);
 592                 fp[i].oflags &= ~(VPO_BUSY | VPO_UNMANAGED);
 593         }
 594         mtx_lock(&vm_phys_fictitious_reg_mtx);
 595         for (segind = 0; segind < VM_PHYS_FICTITIOUS_NSEGS; segind++) {
 596                 seg = &vm_phys_fictitious_segs[segind];
 597                 if (seg->start == 0 && seg->end == 0) {
 598                         seg->start = start;
 599                         seg->end = end;
 600                         seg->first_page = fp;
 601                         mtx_unlock(&vm_phys_fictitious_reg_mtx);
 602                         return (0);
 603                 }
 604         }
 605         mtx_unlock(&vm_phys_fictitious_reg_mtx);
 606 #ifdef VM_PHYSSEG_DENSE
 607         if (malloced)
 608 #endif
 609                 free(fp, M_FICT_PAGES);
 610         return (EBUSY);
 611 }
 612
 613 void
 614 vm_phys_fictitious_unreg_range(vm_paddr_t start, vm_paddr_t end)
 615 {
 616         struct vm_phys_fictitious_seg *seg;
 617         vm_page_t fp;
 618         int segind;
 619 #ifdef VM_PHYSSEG_DENSE
 620         long pi;
 621 #endif
 622
 623 #ifdef VM_PHYSSEG_DENSE
 624         pi = atop(start);
 625 #endif
 626
 627         mtx_lock(&vm_phys_fictitious_reg_mtx);
 628         for (segind = 0; segind < VM_PHYS_FICTITIOUS_NSEGS; segind++) {
 629                 seg = &vm_phys_fictitious_segs[segind];
 630                 if (seg->start == start && seg->end == end) {
 631                         seg->start = seg->end = 0;
 632                         fp = seg->first_page;
 633                         seg->first_page = NULL;
 634                         mtx_unlock(&vm_phys_fictitious_reg_mtx);
 635 #ifdef VM_PHYSSEG_DENSE
 636                         if (pi < first_page || atop(end) >= vm_page_array_size)
 637 #endif
 638                                 free(fp, M_FICT_PAGES);
 639                         return;
 640                 }
 641         }
 642         mtx_unlock(&vm_phys_fictitious_reg_mtx);
 643         KASSERT(0, ("Unregistering not registered fictitious range"));
 644 }
 645
 646 /*
 647  * Find the segment containing the given physical address.
 648  */
 649 static int
 650 vm_phys_paddr_to_segind(vm_paddr_t pa)
 651 {
 652         struct vm_phys_seg *seg;
 653         int segind;
 654
 655         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 656                 seg = &vm_phys_segs[segind];
 657                 if (pa >= seg->start && pa < seg->end)
 658                         return (segind);
 659         }
 660         panic("vm_phys_paddr_to_segind: paddr %#jx is not in any segment" ,
 661             (uintmax_t)pa);
 662 }
 663
 664 /*
 665  * Free a contiguous, power of two-sized set of physical pages.
 666  *
 667  * The free page queues must be locked.
 668  */
 669 void
 670 vm_phys_free_pages(vm_page_t m, int order)
 671 {
 672         struct vm_freelist *fl;
 673         struct vm_phys_seg *seg;
 674         vm_paddr_t pa, pa_buddy;
 675         vm_page_t m_buddy;
 676
 677         KASSERT(m->order == VM_NFREEORDER,
 678             ("vm_phys_free_pages: page %p has unexpected order %d",
 679             m, m->order));
 680         KASSERT(m->pool < VM_NFREEPOOL,
 681             ("vm_phys_free_pages: page %p has unexpected pool %d",
 682             m, m->pool));
 683         KASSERT(order < VM_NFREEORDER,
 684             ("vm_phys_free_pages: order %d is out of range", order));
 685         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 686         pa = VM_PAGE_TO_PHYS(m);
 687         seg = &vm_phys_segs[m->segind];
 688         while (order < VM_NFREEORDER - 1) {
 689                 pa_buddy = pa ^ (1 << (PAGE_SHIFT + order));
 690                 if (pa_buddy < seg->start ||
 691                     pa_buddy >= seg->end)
 692                         break;
 693                 m_buddy = &seg->first_page[atop(pa_buddy - seg->start)];
 694                 if (m_buddy->order != order)
 695                         break;
 696                 fl = (*seg->free_queues)[m_buddy->pool];
 697                 TAILQ_REMOVE(&fl[m_buddy->order].pl, m_buddy, pageq);
 698                 fl[m_buddy->order].lcnt--;
 699                 m_buddy->order = VM_NFREEORDER;
 700                 if (m_buddy->pool != m->pool)
 701                         vm_phys_set_pool(m->pool, m_buddy, order);
 702                 order++;
 703                 pa &= ~((1 << (PAGE_SHIFT + order)) - 1);
 704                 m = &seg->first_page[atop(pa - seg->start)];
 705         }
 706         m->order = order;
 707         fl = (*seg->free_queues)[m->pool];
 708         TAILQ_INSERT_TAIL(&fl[order].pl, m, pageq);
 709         fl[order].lcnt++;
 710 }
 711
 712 /*
 713  * Set the pool for a contiguous, power of two-sized set of physical pages.
 714  */
 715 void
 716 vm_phys_set_pool(int pool, vm_page_t m, int order)
 717 {
 718         vm_page_t m_tmp;
 719
 720         for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++)
 721                 m_tmp->pool = pool;
 722 }
 723
 724 /*
 725  * Search for the given physical page "m" in the free lists.  If the search
 726  * succeeds, remove "m" from the free lists and return TRUE.  Otherwise, return
 727  * FALSE, indicating that "m" is not in the free lists.
 728  *
 729  * The free page queues must be locked.
 730  */
 731 boolean_t
 732 vm_phys_unfree_page(vm_page_t m)
 733 {
 734         struct vm_freelist *fl;
 735         struct vm_phys_seg *seg;
 736         vm_paddr_t pa, pa_half;
 737         vm_page_t m_set, m_tmp;
 738         int order;
 739
 740         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 741
 742         /*
 743          * First, find the contiguous, power of two-sized set of free
 744          * physical pages containing the given physical page "m" and
 745          * assign it to "m_set".
 746          */
 747         seg = &vm_phys_segs[m->segind];
 748         for (m_set = m, order = 0; m_set->order == VM_NFREEORDER &&
 749             order < VM_NFREEORDER - 1; ) {
 750                 order++;
 751                 pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order));
 752                 if (pa >= seg->start)
 753                         m_set = &seg->first_page[atop(pa - seg->start)];
 754                 else
 755                         return (FALSE);
 756         }
 757         if (m_set->order < order)
 758                 return (FALSE);
 759         if (m_set->order == VM_NFREEORDER)
 760                 return (FALSE);
 761         KASSERT(m_set->order < VM_NFREEORDER,
 762             ("vm_phys_unfree_page: page %p has unexpected order %d",
 763             m_set, m_set->order));
 764
 765         /*
 766          * Next, remove "m_set" from the free lists.  Finally, extract
 767          * "m" from "m_set" using an iterative algorithm: While "m_set"
 768          * is larger than a page, shrink "m_set" by returning the half
 769          * of "m_set" that does not contain "m" to the free lists.
 770          */
 771         fl = (*seg->free_queues)[m_set->pool];
 772         order = m_set->order;
 773         TAILQ_REMOVE(&fl[order].pl, m_set, pageq);
 774         fl[order].lcnt--;
 775         m_set->order = VM_NFREEORDER;
 776         while (order > 0) {
 777                 order--;
 778                 pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order));
 779                 if (m->phys_addr < pa_half)
 780                         m_tmp = &seg->first_page[atop(pa_half - seg->start)];
 781                 else {
 782                         m_tmp = m_set;
 783                         m_set = &seg->first_page[atop(pa_half - seg->start)];
 784                 }
 785                 m_tmp->order = order;
 786                 TAILQ_INSERT_HEAD(&fl[order].pl, m_tmp, pageq);
 787                 fl[order].lcnt++;
 788         }
 789         KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency"));
 790         return (TRUE);
 791 }
 792
 793 /*
 794  * Try to zero one physical page.  Used by an idle priority thread.
 795  */
 796 boolean_t
 797 vm_phys_zero_pages_idle(void)
 798 {
 799         static struct vm_freelist *fl = vm_phys_free_queues[0][0];
 800         static int flind, oind, pind;
 801         vm_page_t m, m_tmp;
 802
 803         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 804         for (;;) {
 805                 TAILQ_FOREACH_REVERSE(m, &fl[oind].pl, pglist, pageq) {
 806                         for (m_tmp = m; m_tmp < &m[1 << oind]; m_tmp++) {
 807                                 if ((m_tmp->flags & (PG_CACHED | PG_ZERO)) == 0) {
 808                                         vm_phys_unfree_page(m_tmp);
 809                                         cnt.v_free_count--;
 810                                         mtx_unlock(&vm_page_queue_free_mtx);
 811                                         pmap_zero_page_idle(m_tmp);
 812                                         m_tmp->flags |= PG_ZERO;
 813                                         mtx_lock(&vm_page_queue_free_mtx);
 814                                         cnt.v_free_count++;
 815                                         vm_phys_free_pages(m_tmp, 0);
 816                                         vm_page_zero_count++;
 817                                         cnt_prezero++;
 818                                         return (TRUE);
 819                                 }
 820                         }
 821                 }
 822                 oind++;
 823                 if (oind == VM_NFREEORDER) {
 824                         oind = 0;
 825                         pind++;
 826                         if (pind == VM_NFREEPOOL) {
 827                                 pind = 0;
 828                                 flind++;
 829                                 if (flind == vm_nfreelists)
 830                                         flind = 0;
 831                         }
 832                         fl = vm_phys_free_queues[flind][pind];
 833                 }
 834         }
 835 }
 836
 837 /*
 838  * Allocate a contiguous set of physical pages of the given size
 839  * "npages" from the free lists.  All of the physical pages must be at
 840  * or above the given physical address "low" and below the given
 841  * physical address "high".  The given value "alignment" determines the
 842  * alignment of the first physical page in the set.  If the given value
 843  * "boundary" is non-zero, then the set of physical pages cannot cross
 844  * any physical address boundary that is a multiple of that value.  Both
 845  * "alignment" and "boundary" must be a power of two.
 846  */
 847 vm_page_t
 848 vm_phys_alloc_contig(unsigned long npages, vm_paddr_t low, vm_paddr_t high,
 849     unsigned long alignment, unsigned long boundary)
 850 {
 851         struct vm_freelist *fl;
 852         struct vm_phys_seg *seg;
 853         struct vnode *vp;
 854         vm_paddr_t pa, pa_last, size;
 855         vm_page_t deferred_vdrop_list, m, m_ret;
 856         int domain, flind, i, oind, order, pind;
 857
 858 #if VM_NDOMAIN > 1
 859         domain = PCPU_GET(domain);
 860 #else
 861         domain = 0;
 862 #endif
 863         size = npages << PAGE_SHIFT;
 864         KASSERT(size != 0,
 865             ("vm_phys_alloc_contig: size must not be 0"));
 866         KASSERT((alignment & (alignment - 1)) == 0,
 867             ("vm_phys_alloc_contig: alignment must be a power of 2"));
 868         KASSERT((boundary & (boundary - 1)) == 0,
 869             ("vm_phys_alloc_contig: boundary must be a power of 2"));
 870         deferred_vdrop_list = NULL;
 871         /* Compute the queue that is the best fit for npages. */
 872         for (order = 0; (1 << order) < npages; order++);
 873         mtx_lock(&vm_page_queue_free_mtx);
 874 #if VM_NRESERVLEVEL > 0
 875 retry:
 876 #endif
 877         for (flind = 0; flind < vm_nfreelists; flind++) {
 878                 for (oind = min(order, VM_NFREEORDER - 1); oind < VM_NFREEORDER; oind++) {
 879                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 880                                 fl = (*vm_phys_lookup_lists[domain][flind])
 881                                     [pind];
 882                                 TAILQ_FOREACH(m_ret, &fl[oind].pl, pageq) {
 883                                         /*
 884                                          * A free list may contain physical pages
 885                                          * from one or more segments.
 886                                          */
 887                                         seg = &vm_phys_segs[m_ret->segind];
 888                                         if (seg->start > high ||
 889                                             low >= seg->end)
 890                                                 continue;
 891
 892                                         /*
 893                                          * Is the size of this allocation request
 894                                          * larger than the largest block size?
 895                                          */
 896                                         if (order >= VM_NFREEORDER) {
 897                                                 /*
 898                                                  * Determine if a sufficient number
 899                                                  * of subsequent blocks to satisfy
 900                                                  * the allocation request are free.
 901                                                  */
 902                                                 pa = VM_PAGE_TO_PHYS(m_ret);
 903                                                 pa_last = pa + size;
 904                                                 for (;;) {
 905                                                         pa += 1 << (PAGE_SHIFT + VM_NFREEORDER - 1);
 906                                                         if (pa >= pa_last)
 907                                                                 break;
 908                                                         if (pa < seg->start ||
 909                                                             pa >= seg->end)
 910                                                                 break;
 911                                                         m = &seg->first_page[atop(pa - seg->start)];
 912                                                         if (m->order != VM_NFREEORDER - 1)
 913                                                                 break;
 914                                                 }
 915                                                 /* If not, continue to the next block. */
 916                                                 if (pa < pa_last)
 917                                                         continue;
 918                                         }
 919
 920                                         /*
 921                                          * Determine if the blocks are within the given range,
 922                                          * satisfy the given alignment, and do not cross the
 923                                          * given boundary.
 924                                          */
 925                                         pa = VM_PAGE_TO_PHYS(m_ret);
 926                                         if (pa >= low &&
 927                                             pa + size <= high &&
 928                                             (pa & (alignment - 1)) == 0 &&
 929                                             ((pa ^ (pa + size - 1)) & ~(boundary - 1)) == 0)
 930                                                 goto done;
 931                                 }
 932                         }
 933                 }
 934         }
 935 #if VM_NRESERVLEVEL > 0
 936         if (vm_reserv_reclaim_contig(size, low, high, alignment, boundary))
 937                 goto retry;
 938 #endif
 939         mtx_unlock(&vm_page_queue_free_mtx);
 940         return (NULL);
 941 done:
 942         for (m = m_ret; m < &m_ret[npages]; m = &m[1 << oind]) {
 943                 fl = (*seg->free_queues)[m->pool];
 944                 TAILQ_REMOVE(&fl[m->order].pl, m, pageq);
 945                 fl[m->order].lcnt--;
 946                 m->order = VM_NFREEORDER;
 947         }
 948         if (m_ret->pool != VM_FREEPOOL_DEFAULT)
 949                 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m_ret, oind);
 950         fl = (*seg->free_queues)[m_ret->pool];
 951         vm_phys_split_pages(m_ret, oind, fl, order);
 952         for (i = 0; i < npages; i++) {
 953                 m = &m_ret[i];
 954                 vp = vm_page_alloc_init(m);
 955                 if (vp != NULL) {
 956                         /*
 957                          * Enqueue the vnode for deferred vdrop().
 958                          *
 959                          * Unmanaged pages don't use "pageq", so it
 960                          * can be safely abused to construct a short-
 961                          * lived queue of vnodes.
 962                          */
 963                         m->pageq.tqe_prev = (void *)vp;
 964                         m->pageq.tqe_next = deferred_vdrop_list;
 965                         deferred_vdrop_list = m;
 966                 }
 967         }
 968         for (; i < roundup2(npages, 1 << imin(oind, order)); i++) {
 969                 m = &m_ret[i];
 970                 KASSERT(m->order == VM_NFREEORDER,
 971                     ("vm_phys_alloc_contig: page %p has unexpected order %d",
 972                     m, m->order));
 973                 vm_phys_free_pages(m, 0);
 974         }
 975         mtx_unlock(&vm_page_queue_free_mtx);
 976         while (deferred_vdrop_list != NULL) {
 977                 vdrop((struct vnode *)deferred_vdrop_list->pageq.tqe_prev);
 978                 deferred_vdrop_list = deferred_vdrop_list->pageq.tqe_next;
 979         }
 980         return (m_ret);
 981 }
 982
 983 #ifdef DDB
 984 /*
 985  * Show the number of physical pages in each of the free lists.
 986  */
 987 DB_SHOW_COMMAND(freepages, db_show_freepages)
 988 {
 989         struct vm_freelist *fl;
 990         int flind, oind, pind;
 991
 992         for (flind = 0; flind < vm_nfreelists; flind++) {
 993                 db_printf("FREE LIST %d:\n"
 994                     "\n  ORDER (SIZE)  |  NUMBER"
 995                     "\n              ", flind);
 996                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
 997                         db_printf("  |  POOL %d", pind);
 998                 db_printf("\n--            ");
 999                 for (pind = 0; pind < VM_NFREEPOOL; pind++)
1000                         db_printf("-- --      ");
1001                 db_printf("--\n");
1002                 for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
1003                         db_printf("  %2.2d (%6.6dK)", oind,
1004                             1 << (PAGE_SHIFT - 10 + oind));
1005                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1006                                 fl = vm_phys_free_queues[flind][pind];
1007                                 db_printf("  |  %6.6d", fl[oind].lcnt);
1008                         }
1009                         db_printf("\n");
1010                 }
1011                 db_printf("\n");
1012         }
1013 }
1014 #endif