sys/vm/vm_phys.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2002-2006 Rice University
   5  * Copyright (c) 2007 Alan L. Cox <alc@cs.rice.edu>
   6  * All rights reserved.
   7  *
   8  * This software was developed for the FreeBSD Project by Alan L. Cox,
   9  * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro.
  10  *
  11  * Redistribution and use in source and binary forms, with or without
  12  * modification, are permitted provided that the following conditions
  13  * are met:
  14  * 1. Redistributions of source code must retain the above copyright
  15  *    notice, this list of conditions and the following disclaimer.
  16  * 2. Redistributions in binary form must reproduce the above copyright
  17  *    notice, this list of conditions and the following disclaimer in the
  18  *    documentation and/or other materials provided with the distribution.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  23  * A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT
  24  * HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  25  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
  26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
  27  * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY
  30  * WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
  31  * POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 /*
  35  *      Physical memory system implementation
  36  *
  37  * Any external functions defined by this module are only to be used by the
  38  * virtual memory system.
  39  */
  40
  41 #include <sys/cdefs.h>
  42 __FBSDID("$FreeBSD$");
  43
  44 #include "opt_ddb.h"
  45 #include "opt_vm.h"
  46
  47 #include <sys/param.h>
  48 #include <sys/systm.h>
  49 #include <sys/domainset.h>
  50 #include <sys/lock.h>
  51 #include <sys/kernel.h>
  52 #include <sys/malloc.h>
  53 #include <sys/mutex.h>
  54 #include <sys/proc.h>
  55 #include <sys/queue.h>
  56 #include <sys/rwlock.h>
  57 #include <sys/sbuf.h>
  58 #include <sys/sysctl.h>
  59 #include <sys/tree.h>
  60 #include <sys/vmmeter.h>
  61
  62 #include <ddb/ddb.h>
  63
  64 #include <vm/vm.h>
  65 #include <vm/vm_param.h>
  66 #include <vm/vm_kern.h>
  67 #include <vm/vm_object.h>
  68 #include <vm/vm_page.h>
  69 #include <vm/vm_phys.h>
  70 #include <vm/vm_pagequeue.h>
  71
  72 _Static_assert(sizeof(long) * NBBY >= VM_PHYSSEG_MAX,
  73     "Too many physsegs.");
  74
  75 #ifdef NUMA
  76 struct mem_affinity __read_mostly *mem_affinity;
  77 int __read_mostly *mem_locality;
  78 #endif
  79
  80 int __read_mostly vm_ndomains = 1;
  81 domainset_t __read_mostly all_domains = DOMAINSET_T_INITIALIZER(0x1);
  82
  83 struct vm_phys_seg __read_mostly vm_phys_segs[VM_PHYSSEG_MAX];
  84 int __read_mostly vm_phys_nsegs;
  85
  86 struct vm_phys_fictitious_seg;
  87 static int vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *,
  88     struct vm_phys_fictitious_seg *);
  89
  90 RB_HEAD(fict_tree, vm_phys_fictitious_seg) vm_phys_fictitious_tree =
  91     RB_INITIALIZER(_vm_phys_fictitious_tree);
  92
  93 struct vm_phys_fictitious_seg {
  94         RB_ENTRY(vm_phys_fictitious_seg) node;
  95         /* Memory region data */
  96         vm_paddr_t      start;
  97         vm_paddr_t      end;
  98         vm_page_t       first_page;
  99 };
 100
 101 RB_GENERATE_STATIC(fict_tree, vm_phys_fictitious_seg, node,
 102     vm_phys_fictitious_cmp);
 103
 104 static struct rwlock_padalign vm_phys_fictitious_reg_lock;
 105 MALLOC_DEFINE(M_FICT_PAGES, "vm_fictitious", "Fictitious VM pages");
 106
 107 static struct vm_freelist __aligned(CACHE_LINE_SIZE)
 108     vm_phys_free_queues[MAXMEMDOM][VM_NFREELIST][VM_NFREEPOOL]
 109     [VM_NFREEORDER_MAX];
 110
 111 static int __read_mostly vm_nfreelists;
 112
 113 /*
 114  * Provides the mapping from VM_FREELIST_* to free list indices (flind).
 115  */
 116 static int __read_mostly vm_freelist_to_flind[VM_NFREELIST];
 117
 118 CTASSERT(VM_FREELIST_DEFAULT == 0);
 119
 120 #ifdef VM_FREELIST_DMA32
 121 #define VM_DMA32_BOUNDARY       ((vm_paddr_t)1 << 32)
 122 #endif
 123
 124 /*
 125  * Enforce the assumptions made by vm_phys_add_seg() and vm_phys_init() about
 126  * the ordering of the free list boundaries.
 127  */
 128 #if defined(VM_LOWMEM_BOUNDARY) && defined(VM_DMA32_BOUNDARY)
 129 CTASSERT(VM_LOWMEM_BOUNDARY < VM_DMA32_BOUNDARY);
 130 #endif
 131
 132 static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS);
 133 SYSCTL_OID(_vm, OID_AUTO, phys_free, CTLTYPE_STRING | CTLFLAG_RD,
 134     NULL, 0, sysctl_vm_phys_free, "A", "Phys Free Info");
 135
 136 static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS);
 137 SYSCTL_OID(_vm, OID_AUTO, phys_segs, CTLTYPE_STRING | CTLFLAG_RD,
 138     NULL, 0, sysctl_vm_phys_segs, "A", "Phys Seg Info");
 139
 140 #ifdef NUMA
 141 static int sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS);
 142 SYSCTL_OID(_vm, OID_AUTO, phys_locality, CTLTYPE_STRING | CTLFLAG_RD,
 143     NULL, 0, sysctl_vm_phys_locality, "A", "Phys Locality Info");
 144 #endif
 145
 146 SYSCTL_INT(_vm, OID_AUTO, ndomains, CTLFLAG_RD,
 147     &vm_ndomains, 0, "Number of physical memory domains available.");
 148
 149 static vm_page_t vm_phys_alloc_seg_contig(struct vm_phys_seg *seg,
 150     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
 151     vm_paddr_t boundary);
 152 static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain);
 153 static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end);
 154 static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl,
 155     int order, int tail);
 156
 157 /*
 158  * Red-black tree helpers for vm fictitious range management.
 159  */
 160 static inline int
 161 vm_phys_fictitious_in_range(struct vm_phys_fictitious_seg *p,
 162     struct vm_phys_fictitious_seg *range)
 163 {
 164
 165         KASSERT(range->start != 0 && range->end != 0,
 166             ("Invalid range passed on search for vm_fictitious page"));
 167         if (p->start >= range->end)
 168                 return (1);
 169         if (p->start < range->start)
 170                 return (-1);
 171
 172         return (0);
 173 }
 174
 175 static int
 176 vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *p1,
 177     struct vm_phys_fictitious_seg *p2)
 178 {
 179
 180         /* Check if this is a search for a page */
 181         if (p1->end == 0)
 182                 return (vm_phys_fictitious_in_range(p1, p2));
 183
 184         KASSERT(p2->end != 0,
 185     ("Invalid range passed as second parameter to vm fictitious comparison"));
 186
 187         /* Searching to add a new range */
 188         if (p1->end <= p2->start)
 189                 return (-1);
 190         if (p1->start >= p2->end)
 191                 return (1);
 192
 193         panic("Trying to add overlapping vm fictitious ranges:\n"
 194             "[%#jx:%#jx] and [%#jx:%#jx]", (uintmax_t)p1->start,
 195             (uintmax_t)p1->end, (uintmax_t)p2->start, (uintmax_t)p2->end);
 196 }
 197
 198 int
 199 vm_phys_domain_match(int prefer, vm_paddr_t low, vm_paddr_t high)
 200 {
 201 #ifdef NUMA
 202         domainset_t mask;
 203         int i;
 204
 205         if (vm_ndomains == 1 || mem_affinity == NULL)
 206                 return (0);
 207
 208         DOMAINSET_ZERO(&mask);
 209         /*
 210          * Check for any memory that overlaps low, high.
 211          */
 212         for (i = 0; mem_affinity[i].end != 0; i++)
 213                 if (mem_affinity[i].start <= high &&
 214                     mem_affinity[i].end >= low)
 215                         DOMAINSET_SET(mem_affinity[i].domain, &mask);
 216         if (prefer != -1 && DOMAINSET_ISSET(prefer, &mask))
 217                 return (prefer);
 218         if (DOMAINSET_EMPTY(&mask))
 219                 panic("vm_phys_domain_match:  Impossible constraint");
 220         return (DOMAINSET_FFS(&mask) - 1);
 221 #else
 222         return (0);
 223 #endif
 224 }
 225
 226 /*
 227  * Outputs the state of the physical memory allocator, specifically,
 228  * the amount of physical memory in each free list.
 229  */
 230 static int
 231 sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS)
 232 {
 233         struct sbuf sbuf;
 234         struct vm_freelist *fl;
 235         int dom, error, flind, oind, pind;
 236
 237         error = sysctl_wire_old_buffer(req, 0);
 238         if (error != 0)
 239                 return (error);
 240         sbuf_new_for_sysctl(&sbuf, NULL, 128 * vm_ndomains, req);
 241         for (dom = 0; dom < vm_ndomains; dom++) {
 242                 sbuf_printf(&sbuf,"\nDOMAIN %d:\n", dom);
 243                 for (flind = 0; flind < vm_nfreelists; flind++) {
 244                         sbuf_printf(&sbuf, "\nFREE LIST %d:\n"
 245                             "\n  ORDER (SIZE)  |  NUMBER"
 246                             "\n              ", flind);
 247                         for (pind = 0; pind < VM_NFREEPOOL; pind++)
 248                                 sbuf_printf(&sbuf, "  |  POOL %d", pind);
 249                         sbuf_printf(&sbuf, "\n--            ");
 250                         for (pind = 0; pind < VM_NFREEPOOL; pind++)
 251                                 sbuf_printf(&sbuf, "-- --      ");
 252                         sbuf_printf(&sbuf, "--\n");
 253                         for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
 254                                 sbuf_printf(&sbuf, "  %2d (%6dK)", oind,
 255                                     1 << (PAGE_SHIFT - 10 + oind));
 256                                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 257                                 fl = vm_phys_free_queues[dom][flind][pind];
 258                                         sbuf_printf(&sbuf, "  |  %6d",
 259                                             fl[oind].lcnt);
 260                                 }
 261                                 sbuf_printf(&sbuf, "\n");
 262                         }
 263                 }
 264         }
 265         error = sbuf_finish(&sbuf);
 266         sbuf_delete(&sbuf);
 267         return (error);
 268 }
 269
 270 /*
 271  * Outputs the set of physical memory segments.
 272  */
 273 static int
 274 sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS)
 275 {
 276         struct sbuf sbuf;
 277         struct vm_phys_seg *seg;
 278         int error, segind;
 279
 280         error = sysctl_wire_old_buffer(req, 0);
 281         if (error != 0)
 282                 return (error);
 283         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 284         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 285                 sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind);
 286                 seg = &vm_phys_segs[segind];
 287                 sbuf_printf(&sbuf, "start:     %#jx\n",
 288                     (uintmax_t)seg->start);
 289                 sbuf_printf(&sbuf, "end:       %#jx\n",
 290                     (uintmax_t)seg->end);
 291                 sbuf_printf(&sbuf, "domain:    %d\n", seg->domain);
 292                 sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues);
 293         }
 294         error = sbuf_finish(&sbuf);
 295         sbuf_delete(&sbuf);
 296         return (error);
 297 }
 298
 299 /*
 300  * Return affinity, or -1 if there's no affinity information.
 301  */
 302 int
 303 vm_phys_mem_affinity(int f, int t)
 304 {
 305
 306 #ifdef NUMA
 307         if (mem_locality == NULL)
 308                 return (-1);
 309         if (f >= vm_ndomains || t >= vm_ndomains)
 310                 return (-1);
 311         return (mem_locality[f * vm_ndomains + t]);
 312 #else
 313         return (-1);
 314 #endif
 315 }
 316
 317 #ifdef NUMA
 318 /*
 319  * Outputs the VM locality table.
 320  */
 321 static int
 322 sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS)
 323 {
 324         struct sbuf sbuf;
 325         int error, i, j;
 326
 327         error = sysctl_wire_old_buffer(req, 0);
 328         if (error != 0)
 329                 return (error);
 330         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 331
 332         sbuf_printf(&sbuf, "\n");
 333
 334         for (i = 0; i < vm_ndomains; i++) {
 335                 sbuf_printf(&sbuf, "%d: ", i);
 336                 for (j = 0; j < vm_ndomains; j++) {
 337                         sbuf_printf(&sbuf, "%d ", vm_phys_mem_affinity(i, j));
 338                 }
 339                 sbuf_printf(&sbuf, "\n");
 340         }
 341         error = sbuf_finish(&sbuf);
 342         sbuf_delete(&sbuf);
 343         return (error);
 344 }
 345 #endif
 346
 347 static void
 348 vm_freelist_add(struct vm_freelist *fl, vm_page_t m, int order, int tail)
 349 {
 350
 351         m->order = order;
 352         if (tail)
 353                 TAILQ_INSERT_TAIL(&fl[order].pl, m, listq);
 354         else
 355                 TAILQ_INSERT_HEAD(&fl[order].pl, m, listq);
 356         fl[order].lcnt++;
 357 }
 358
 359 static void
 360 vm_freelist_rem(struct vm_freelist *fl, vm_page_t m, int order)
 361 {
 362
 363         TAILQ_REMOVE(&fl[order].pl, m, listq);
 364         fl[order].lcnt--;
 365         m->order = VM_NFREEORDER;
 366 }
 367
 368 /*
 369  * Create a physical memory segment.
 370  */
 371 static void
 372 _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain)
 373 {
 374         struct vm_phys_seg *seg;
 375
 376         KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX,
 377             ("vm_phys_create_seg: increase VM_PHYSSEG_MAX"));
 378         KASSERT(domain >= 0 && domain < vm_ndomains,
 379             ("vm_phys_create_seg: invalid domain provided"));
 380         seg = &vm_phys_segs[vm_phys_nsegs++];
 381         while (seg > vm_phys_segs && (seg - 1)->start >= end) {
 382                 *seg = *(seg - 1);
 383                 seg--;
 384         }
 385         seg->start = start;
 386         seg->end = end;
 387         seg->domain = domain;
 388 }
 389
 390 static void
 391 vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end)
 392 {
 393 #ifdef NUMA
 394         int i;
 395
 396         if (mem_affinity == NULL) {
 397                 _vm_phys_create_seg(start, end, 0);
 398                 return;
 399         }
 400
 401         for (i = 0;; i++) {
 402                 if (mem_affinity[i].end == 0)
 403                         panic("Reached end of affinity info");
 404                 if (mem_affinity[i].end <= start)
 405                         continue;
 406                 if (mem_affinity[i].start > start)
 407                         panic("No affinity info for start %jx",
 408                             (uintmax_t)start);
 409                 if (mem_affinity[i].end >= end) {
 410                         _vm_phys_create_seg(start, end,
 411                             mem_affinity[i].domain);
 412                         break;
 413                 }
 414                 _vm_phys_create_seg(start, mem_affinity[i].end,
 415                     mem_affinity[i].domain);
 416                 start = mem_affinity[i].end;
 417         }
 418 #else
 419         _vm_phys_create_seg(start, end, 0);
 420 #endif
 421 }
 422
 423 /*
 424  * Add a physical memory segment.
 425  */
 426 void
 427 vm_phys_add_seg(vm_paddr_t start, vm_paddr_t end)
 428 {
 429         vm_paddr_t paddr;
 430
 431         KASSERT((start & PAGE_MASK) == 0,
 432             ("vm_phys_define_seg: start is not page aligned"));
 433         KASSERT((end & PAGE_MASK) == 0,
 434             ("vm_phys_define_seg: end is not page aligned"));
 435
 436         /*
 437          * Split the physical memory segment if it spans two or more free
 438          * list boundaries.
 439          */
 440         paddr = start;
 441 #ifdef  VM_FREELIST_LOWMEM
 442         if (paddr < VM_LOWMEM_BOUNDARY && end > VM_LOWMEM_BOUNDARY) {
 443                 vm_phys_create_seg(paddr, VM_LOWMEM_BOUNDARY);
 444                 paddr = VM_LOWMEM_BOUNDARY;
 445         }
 446 #endif
 447 #ifdef  VM_FREELIST_DMA32
 448         if (paddr < VM_DMA32_BOUNDARY && end > VM_DMA32_BOUNDARY) {
 449                 vm_phys_create_seg(paddr, VM_DMA32_BOUNDARY);
 450                 paddr = VM_DMA32_BOUNDARY;
 451         }
 452 #endif
 453         vm_phys_create_seg(paddr, end);
 454 }
 455
 456 /*
 457  * Initialize the physical memory allocator.
 458  *
 459  * Requires that vm_page_array is initialized!
 460  */
 461 void
 462 vm_phys_init(void)
 463 {
 464         struct vm_freelist *fl;
 465         struct vm_phys_seg *end_seg, *prev_seg, *seg, *tmp_seg;
 466         u_long npages;
 467         int dom, flind, freelist, oind, pind, segind;
 468
 469         /*
 470          * Compute the number of free lists, and generate the mapping from the
 471          * manifest constants VM_FREELIST_* to the free list indices.
 472          *
 473          * Initially, the entries of vm_freelist_to_flind[] are set to either
 474          * 0 or 1 to indicate which free lists should be created.
 475          */
 476         npages = 0;
 477         for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
 478                 seg = &vm_phys_segs[segind];
 479 #ifdef  VM_FREELIST_LOWMEM
 480                 if (seg->end <= VM_LOWMEM_BOUNDARY)
 481                         vm_freelist_to_flind[VM_FREELIST_LOWMEM] = 1;
 482                 else
 483 #endif
 484 #ifdef  VM_FREELIST_DMA32
 485                 if (
 486 #ifdef  VM_DMA32_NPAGES_THRESHOLD
 487                     /*
 488                      * Create the DMA32 free list only if the amount of
 489                      * physical memory above physical address 4G exceeds the
 490                      * given threshold.
 491                      */
 492                     npages > VM_DMA32_NPAGES_THRESHOLD &&
 493 #endif
 494                     seg->end <= VM_DMA32_BOUNDARY)
 495                         vm_freelist_to_flind[VM_FREELIST_DMA32] = 1;
 496                 else
 497 #endif
 498                 {
 499                         npages += atop(seg->end - seg->start);
 500                         vm_freelist_to_flind[VM_FREELIST_DEFAULT] = 1;
 501                 }
 502         }
 503         /* Change each entry into a running total of the free lists. */
 504         for (freelist = 1; freelist < VM_NFREELIST; freelist++) {
 505                 vm_freelist_to_flind[freelist] +=
 506                     vm_freelist_to_flind[freelist - 1];
 507         }
 508         vm_nfreelists = vm_freelist_to_flind[VM_NFREELIST - 1];
 509         KASSERT(vm_nfreelists > 0, ("vm_phys_init: no free lists"));
 510         /* Change each entry into a free list index. */
 511         for (freelist = 0; freelist < VM_NFREELIST; freelist++)
 512                 vm_freelist_to_flind[freelist]--;
 513
 514         /*
 515          * Initialize the first_page and free_queues fields of each physical
 516          * memory segment.
 517          */
 518 #ifdef VM_PHYSSEG_SPARSE
 519         npages = 0;
 520 #endif
 521         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 522                 seg = &vm_phys_segs[segind];
 523 #ifdef VM_PHYSSEG_SPARSE
 524                 seg->first_page = &vm_page_array[npages];
 525                 npages += atop(seg->end - seg->start);
 526 #else
 527                 seg->first_page = PHYS_TO_VM_PAGE(seg->start);
 528 #endif
 529 #ifdef  VM_FREELIST_LOWMEM
 530                 if (seg->end <= VM_LOWMEM_BOUNDARY) {
 531                         flind = vm_freelist_to_flind[VM_FREELIST_LOWMEM];
 532                         KASSERT(flind >= 0,
 533                             ("vm_phys_init: LOWMEM flind < 0"));
 534                 } else
 535 #endif
 536 #ifdef  VM_FREELIST_DMA32
 537                 if (seg->end <= VM_DMA32_BOUNDARY) {
 538                         flind = vm_freelist_to_flind[VM_FREELIST_DMA32];
 539                         KASSERT(flind >= 0,
 540                             ("vm_phys_init: DMA32 flind < 0"));
 541                 } else
 542 #endif
 543                 {
 544                         flind = vm_freelist_to_flind[VM_FREELIST_DEFAULT];
 545                         KASSERT(flind >= 0,
 546                             ("vm_phys_init: DEFAULT flind < 0"));
 547                 }
 548                 seg->free_queues = &vm_phys_free_queues[seg->domain][flind];
 549         }
 550
 551         /*
 552          * Coalesce physical memory segments that are contiguous and share the
 553          * same per-domain free queues.
 554          */
 555         prev_seg = vm_phys_segs;
 556         seg = &vm_phys_segs[1];
 557         end_seg = &vm_phys_segs[vm_phys_nsegs];
 558         while (seg < end_seg) {
 559                 if (prev_seg->end == seg->start &&
 560                     prev_seg->free_queues == seg->free_queues) {
 561                         prev_seg->end = seg->end;
 562                         KASSERT(prev_seg->domain == seg->domain,
 563                             ("vm_phys_init: free queues cannot span domains"));
 564                         vm_phys_nsegs--;
 565                         end_seg--;
 566                         for (tmp_seg = seg; tmp_seg < end_seg; tmp_seg++)
 567                                 *tmp_seg = *(tmp_seg + 1);
 568                 } else {
 569                         prev_seg = seg;
 570                         seg++;
 571                 }
 572         }
 573
 574         /*
 575          * Initialize the free queues.
 576          */
 577         for (dom = 0; dom < vm_ndomains; dom++) {
 578                 for (flind = 0; flind < vm_nfreelists; flind++) {
 579                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 580                                 fl = vm_phys_free_queues[dom][flind][pind];
 581                                 for (oind = 0; oind < VM_NFREEORDER; oind++)
 582                                         TAILQ_INIT(&fl[oind].pl);
 583                         }
 584                 }
 585         }
 586
 587         rw_init(&vm_phys_fictitious_reg_lock, "vmfctr");
 588 }
 589
 590 /*
 591  * Register info about the NUMA topology of the system.
 592  *
 593  * Invoked by platform-dependent code prior to vm_phys_init().
 594  */
 595 void
 596 vm_phys_register_domains(int ndomains, struct mem_affinity *affinity,
 597     int *locality)
 598 {
 599 #ifdef NUMA
 600         int d, i;
 601
 602         /*
 603          * For now the only override value that we support is 1, which
 604          * effectively disables NUMA-awareness in the allocators.
 605          */
 606         d = 0;
 607         TUNABLE_INT_FETCH("vm.numa.disabled", &d);
 608         if (d)
 609                 ndomains = 1;
 610
 611         if (ndomains > 1) {
 612                 vm_ndomains = ndomains;
 613                 mem_affinity = affinity;
 614                 mem_locality = locality;
 615         }
 616
 617         for (i = 0; i < vm_ndomains; i++)
 618                 DOMAINSET_SET(i, &all_domains);
 619 #else
 620         (void)ndomains;
 621         (void)affinity;
 622         (void)locality;
 623 #endif
 624 }
 625
 626 int
 627 _vm_phys_domain(vm_paddr_t pa)
 628 {
 629 #ifdef NUMA
 630         int i;
 631
 632         if (vm_ndomains == 1 || mem_affinity == NULL)
 633                 return (0);
 634
 635         /*
 636          * Check for any memory that overlaps.
 637          */
 638         for (i = 0; mem_affinity[i].end != 0; i++)
 639                 if (mem_affinity[i].start <= pa &&
 640                     mem_affinity[i].end >= pa)
 641                         return (mem_affinity[i].domain);
 642 #endif
 643         return (0);
 644 }
 645
 646 /*
 647  * Split a contiguous, power of two-sized set of physical pages.
 648  *
 649  * When this function is called by a page allocation function, the caller
 650  * should request insertion at the head unless the order [order, oind) queues
 651  * are known to be empty.  The objective being to reduce the likelihood of
 652  * long-term fragmentation by promoting contemporaneous allocation and
 653  * (hopefully) deallocation.
 654  */
 655 static __inline void
 656 vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order,
 657     int tail)
 658 {
 659         vm_page_t m_buddy;
 660
 661         while (oind > order) {
 662                 oind--;
 663                 m_buddy = &m[1 << oind];
 664                 KASSERT(m_buddy->order == VM_NFREEORDER,
 665                     ("vm_phys_split_pages: page %p has unexpected order %d",
 666                     m_buddy, m_buddy->order));
 667                 vm_freelist_add(fl, m_buddy, oind, tail);
 668         }
 669 }
 670
 671 /*
 672  * Add the physical pages [m, m + npages) at the end of a power-of-two aligned
 673  * and sized set to the specified free list.
 674  *
 675  * When this function is called by a page allocation function, the caller
 676  * should request insertion at the head unless the lower-order queues are
 677  * known to be empty.  The objective being to reduce the likelihood of long-
 678  * term fragmentation by promoting contemporaneous allocation and (hopefully)
 679  * deallocation.
 680  *
 681  * The physical page m's buddy must not be free.
 682  */
 683 static void
 684 vm_phys_enq_range(vm_page_t m, u_int npages, struct vm_freelist *fl, int tail)
 685 {
 686         u_int n;
 687         int order;
 688
 689         KASSERT(npages > 0, ("vm_phys_enq_range: npages is 0"));
 690         KASSERT(((VM_PAGE_TO_PHYS(m) + npages * PAGE_SIZE) &
 691             ((PAGE_SIZE << (fls(npages) - 1)) - 1)) == 0,
 692             ("vm_phys_enq_range: page %p and npages %u are misaligned",
 693             m, npages));
 694         do {
 695                 KASSERT(m->order == VM_NFREEORDER,
 696                     ("vm_phys_enq_range: page %p has unexpected order %d",
 697                     m, m->order));
 698                 order = ffs(npages) - 1;
 699                 KASSERT(order < VM_NFREEORDER,
 700                     ("vm_phys_enq_range: order %d is out of range", order));
 701                 vm_freelist_add(fl, m, order, tail);
 702                 n = 1 << order;
 703                 m += n;
 704                 npages -= n;
 705         } while (npages > 0);
 706 }
 707
 708 /*
 709  * Tries to allocate the specified number of pages from the specified pool
 710  * within the specified domain.  Returns the actual number of allocated pages
 711  * and a pointer to each page through the array ma[].
 712  *
 713  * The returned pages may not be physically contiguous.  However, in contrast
 714  * to performing multiple, back-to-back calls to vm_phys_alloc_pages(..., 0),
 715  * calling this function once to allocate the desired number of pages will
 716  * avoid wasted time in vm_phys_split_pages().
 717  *
 718  * The free page queues for the specified domain must be locked.
 719  */
 720 int
 721 vm_phys_alloc_npages(int domain, int pool, int npages, vm_page_t ma[])
 722 {
 723         struct vm_freelist *alt, *fl;
 724         vm_page_t m;
 725         int avail, end, flind, freelist, i, need, oind, pind;
 726
 727         KASSERT(domain >= 0 && domain < vm_ndomains,
 728             ("vm_phys_alloc_npages: domain %d is out of range", domain));
 729         KASSERT(pool < VM_NFREEPOOL,
 730             ("vm_phys_alloc_npages: pool %d is out of range", pool));
 731         KASSERT(npages <= 1 << (VM_NFREEORDER - 1),
 732             ("vm_phys_alloc_npages: npages %d is out of range", npages));
 733         vm_domain_free_assert_locked(VM_DOMAIN(domain));
 734         i = 0;
 735         for (freelist = 0; freelist < VM_NFREELIST; freelist++) {
 736                 flind = vm_freelist_to_flind[freelist];
 737                 if (flind < 0)
 738                         continue;
 739                 fl = vm_phys_free_queues[domain][flind][pool];
 740                 for (oind = 0; oind < VM_NFREEORDER; oind++) {
 741                         while ((m = TAILQ_FIRST(&fl[oind].pl)) != NULL) {
 742                                 vm_freelist_rem(fl, m, oind);
 743                                 avail = 1 << oind;
 744                                 need = imin(npages - i, avail);
 745                                 for (end = i + need; i < end;)
 746                                         ma[i++] = m++;
 747                                 if (need < avail) {
 748                                         /*
 749                                          * Return excess pages to fl.  Its
 750                                          * order [0, oind) queues are empty.
 751                                          */
 752                                         vm_phys_enq_range(m, avail - need, fl,
 753                                             1);
 754                                         return (npages);
 755                                 } else if (i == npages)
 756                                         return (npages);
 757                         }
 758                 }
 759                 for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
 760                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 761                                 alt = vm_phys_free_queues[domain][flind][pind];
 762                                 while ((m = TAILQ_FIRST(&alt[oind].pl)) !=
 763                                     NULL) {
 764                                         vm_freelist_rem(alt, m, oind);
 765                                         vm_phys_set_pool(pool, m, oind);
 766                                         avail = 1 << oind;
 767                                         need = imin(npages - i, avail);
 768                                         for (end = i + need; i < end;)
 769                                                 ma[i++] = m++;
 770                                         if (need < avail) {
 771                                                 /*
 772                                                  * Return excess pages to fl.
 773                                                  * Its order [0, oind) queues
 774                                                  * are empty.
 775                                                  */
 776                                                 vm_phys_enq_range(m, avail -
 777                                                     need, fl, 1);
 778                                                 return (npages);
 779                                         } else if (i == npages)
 780                                                 return (npages);
 781                                 }
 782                         }
 783                 }
 784         }
 785         return (i);
 786 }
 787
 788 /*
 789  * Allocate a contiguous, power of two-sized set of physical pages
 790  * from the free lists.
 791  *
 792  * The free page queues must be locked.
 793  */
 794 vm_page_t
 795 vm_phys_alloc_pages(int domain, int pool, int order)
 796 {
 797         vm_page_t m;
 798         int freelist;
 799
 800         for (freelist = 0; freelist < VM_NFREELIST; freelist++) {
 801                 m = vm_phys_alloc_freelist_pages(domain, freelist, pool, order);
 802                 if (m != NULL)
 803                         return (m);
 804         }
 805         return (NULL);
 806 }
 807
 808 /*
 809  * Allocate a contiguous, power of two-sized set of physical pages from the
 810  * specified free list.  The free list must be specified using one of the
 811  * manifest constants VM_FREELIST_*.
 812  *
 813  * The free page queues must be locked.
 814  */
 815 vm_page_t
 816 vm_phys_alloc_freelist_pages(int domain, int freelist, int pool, int order)
 817 {
 818         struct vm_freelist *alt, *fl;
 819         vm_page_t m;
 820         int oind, pind, flind;
 821
 822         KASSERT(domain >= 0 && domain < vm_ndomains,
 823             ("vm_phys_alloc_freelist_pages: domain %d is out of range",
 824             domain));
 825         KASSERT(freelist < VM_NFREELIST,
 826             ("vm_phys_alloc_freelist_pages: freelist %d is out of range",
 827             freelist));
 828         KASSERT(pool < VM_NFREEPOOL,
 829             ("vm_phys_alloc_freelist_pages: pool %d is out of range", pool));
 830         KASSERT(order < VM_NFREEORDER,
 831             ("vm_phys_alloc_freelist_pages: order %d is out of range", order));
 832
 833         flind = vm_freelist_to_flind[freelist];
 834         /* Check if freelist is present */
 835         if (flind < 0)
 836                 return (NULL);
 837
 838         vm_domain_free_assert_locked(VM_DOMAIN(domain));
 839         fl = &vm_phys_free_queues[domain][flind][pool][0];
 840         for (oind = order; oind < VM_NFREEORDER; oind++) {
 841                 m = TAILQ_FIRST(&fl[oind].pl);
 842                 if (m != NULL) {
 843                         vm_freelist_rem(fl, m, oind);
 844                         /* The order [order, oind) queues are empty. */
 845                         vm_phys_split_pages(m, oind, fl, order, 1);
 846                         return (m);
 847                 }
 848         }
 849
 850         /*
 851          * The given pool was empty.  Find the largest
 852          * contiguous, power-of-two-sized set of pages in any
 853          * pool.  Transfer these pages to the given pool, and
 854          * use them to satisfy the allocation.
 855          */
 856         for (oind = VM_NFREEORDER - 1; oind >= order; oind--) {
 857                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 858                         alt = &vm_phys_free_queues[domain][flind][pind][0];
 859                         m = TAILQ_FIRST(&alt[oind].pl);
 860                         if (m != NULL) {
 861                                 vm_freelist_rem(alt, m, oind);
 862                                 vm_phys_set_pool(pool, m, oind);
 863                                 /* The order [order, oind) queues are empty. */
 864                                 vm_phys_split_pages(m, oind, fl, order, 1);
 865                                 return (m);
 866                         }
 867                 }
 868         }
 869         return (NULL);
 870 }
 871
 872 /*
 873  * Find the vm_page corresponding to the given physical address.
 874  */
 875 vm_page_t
 876 vm_phys_paddr_to_vm_page(vm_paddr_t pa)
 877 {
 878         struct vm_phys_seg *seg;
 879         int segind;
 880
 881         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 882                 seg = &vm_phys_segs[segind];
 883                 if (pa >= seg->start && pa < seg->end)
 884                         return (&seg->first_page[atop(pa - seg->start)]);
 885         }
 886         return (NULL);
 887 }
 888
 889 vm_page_t
 890 vm_phys_fictitious_to_vm_page(vm_paddr_t pa)
 891 {
 892         struct vm_phys_fictitious_seg tmp, *seg;
 893         vm_page_t m;
 894
 895         m = NULL;
 896         tmp.start = pa;
 897         tmp.end = 0;
 898
 899         rw_rlock(&vm_phys_fictitious_reg_lock);
 900         seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
 901         rw_runlock(&vm_phys_fictitious_reg_lock);
 902         if (seg == NULL)
 903                 return (NULL);
 904
 905         m = &seg->first_page[atop(pa - seg->start)];
 906         KASSERT((m->flags & PG_FICTITIOUS) != 0, ("%p not fictitious", m));
 907
 908         return (m);
 909 }
 910
 911 static inline void
 912 vm_phys_fictitious_init_range(vm_page_t range, vm_paddr_t start,
 913     long page_count, vm_memattr_t memattr)
 914 {
 915         long i;
 916
 917         bzero(range, page_count * sizeof(*range));
 918         for (i = 0; i < page_count; i++) {
 919                 vm_page_initfake(&range[i], start + PAGE_SIZE * i, memattr);
 920                 range[i].oflags &= ~VPO_UNMANAGED;
 921                 range[i].busy_lock = VPB_UNBUSIED;
 922         }
 923 }
 924
 925 int
 926 vm_phys_fictitious_reg_range(vm_paddr_t start, vm_paddr_t end,
 927     vm_memattr_t memattr)
 928 {
 929         struct vm_phys_fictitious_seg *seg;
 930         vm_page_t fp;
 931         long page_count;
 932 #ifdef VM_PHYSSEG_DENSE
 933         long pi, pe;
 934         long dpage_count;
 935 #endif
 936
 937         KASSERT(start < end,
 938             ("Start of segment isn't less than end (start: %jx end: %jx)",
 939             (uintmax_t)start, (uintmax_t)end));
 940
 941         page_count = (end - start) / PAGE_SIZE;
 942
 943 #ifdef VM_PHYSSEG_DENSE
 944         pi = atop(start);
 945         pe = atop(end);
 946         if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
 947                 fp = &vm_page_array[pi - first_page];
 948                 if ((pe - first_page) > vm_page_array_size) {
 949                         /*
 950                          * We have a segment that starts inside
 951                          * of vm_page_array, but ends outside of it.
 952                          *
 953                          * Use vm_page_array pages for those that are
 954                          * inside of the vm_page_array range, and
 955                          * allocate the remaining ones.
 956                          */
 957                         dpage_count = vm_page_array_size - (pi - first_page);
 958                         vm_phys_fictitious_init_range(fp, start, dpage_count,
 959                             memattr);
 960                         page_count -= dpage_count;
 961                         start += ptoa(dpage_count);
 962                         goto alloc;
 963                 }
 964                 /*
 965                  * We can allocate the full range from vm_page_array,
 966                  * so there's no need to register the range in the tree.
 967                  */
 968                 vm_phys_fictitious_init_range(fp, start, page_count, memattr);
 969                 return (0);
 970         } else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
 971                 /*
 972                  * We have a segment that ends inside of vm_page_array,
 973                  * but starts outside of it.
 974                  */
 975                 fp = &vm_page_array[0];
 976                 dpage_count = pe - first_page;
 977                 vm_phys_fictitious_init_range(fp, ptoa(first_page), dpage_count,
 978                     memattr);
 979                 end -= ptoa(dpage_count);
 980                 page_count -= dpage_count;
 981                 goto alloc;
 982         } else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
 983                 /*
 984                  * Trying to register a fictitious range that expands before
 985                  * and after vm_page_array.
 986                  */
 987                 return (EINVAL);
 988         } else {
 989 alloc:
 990 #endif
 991                 fp = malloc(page_count * sizeof(struct vm_page), M_FICT_PAGES,
 992                     M_WAITOK);
 993 #ifdef VM_PHYSSEG_DENSE
 994         }
 995 #endif
 996         vm_phys_fictitious_init_range(fp, start, page_count, memattr);
 997
 998         seg = malloc(sizeof(*seg), M_FICT_PAGES, M_WAITOK | M_ZERO);
 999         seg->start = start;
1000         seg->end = end;
1001         seg->first_page = fp;
1002
1003         rw_wlock(&vm_phys_fictitious_reg_lock);
1004         RB_INSERT(fict_tree, &vm_phys_fictitious_tree, seg);
1005         rw_wunlock(&vm_phys_fictitious_reg_lock);
1006
1007         return (0);
1008 }
1009
1010 void
1011 vm_phys_fictitious_unreg_range(vm_paddr_t start, vm_paddr_t end)
1012 {
1013         struct vm_phys_fictitious_seg *seg, tmp;
1014 #ifdef VM_PHYSSEG_DENSE
1015         long pi, pe;
1016 #endif
1017
1018         KASSERT(start < end,
1019             ("Start of segment isn't less than end (start: %jx end: %jx)",
1020             (uintmax_t)start, (uintmax_t)end));
1021
1022 #ifdef VM_PHYSSEG_DENSE
1023         pi = atop(start);
1024         pe = atop(end);
1025         if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1026                 if ((pe - first_page) <= vm_page_array_size) {
1027                         /*
1028                          * This segment was allocated using vm_page_array
1029                          * only, there's nothing to do since those pages
1030                          * were never added to the tree.
1031                          */
1032                         return;
1033                 }
1034                 /*
1035                  * We have a segment that starts inside
1036                  * of vm_page_array, but ends outside of it.
1037                  *
1038                  * Calculate how many pages were added to the
1039                  * tree and free them.
1040                  */
1041                 start = ptoa(first_page + vm_page_array_size);
1042         } else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
1043                 /*
1044                  * We have a segment that ends inside of vm_page_array,
1045                  * but starts outside of it.
1046                  */
1047                 end = ptoa(first_page);
1048         } else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
1049                 /* Since it's not possible to register such a range, panic. */
1050                 panic(
1051                     "Unregistering not registered fictitious range [%#jx:%#jx]",
1052                     (uintmax_t)start, (uintmax_t)end);
1053         }
1054 #endif
1055         tmp.start = start;
1056         tmp.end = 0;
1057
1058         rw_wlock(&vm_phys_fictitious_reg_lock);
1059         seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
1060         if (seg->start != start || seg->end != end) {
1061                 rw_wunlock(&vm_phys_fictitious_reg_lock);
1062                 panic(
1063                     "Unregistering not registered fictitious range [%#jx:%#jx]",
1064                     (uintmax_t)start, (uintmax_t)end);
1065         }
1066         RB_REMOVE(fict_tree, &vm_phys_fictitious_tree, seg);
1067         rw_wunlock(&vm_phys_fictitious_reg_lock);
1068         free(seg->first_page, M_FICT_PAGES);
1069         free(seg, M_FICT_PAGES);
1070 }
1071
1072 /*
1073  * Free a contiguous, power of two-sized set of physical pages.
1074  *
1075  * The free page queues must be locked.
1076  */
1077 void
1078 vm_phys_free_pages(vm_page_t m, int order)
1079 {
1080         struct vm_freelist *fl;
1081         struct vm_phys_seg *seg;
1082         vm_paddr_t pa;
1083         vm_page_t m_buddy;
1084
1085         KASSERT(m->order == VM_NFREEORDER,
1086             ("vm_phys_free_pages: page %p has unexpected order %d",
1087             m, m->order));
1088         KASSERT(m->pool < VM_NFREEPOOL,
1089             ("vm_phys_free_pages: page %p has unexpected pool %d",
1090             m, m->pool));
1091         KASSERT(order < VM_NFREEORDER,
1092             ("vm_phys_free_pages: order %d is out of range", order));
1093         seg = &vm_phys_segs[m->segind];
1094         vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1095         if (order < VM_NFREEORDER - 1) {
1096                 pa = VM_PAGE_TO_PHYS(m);
1097                 do {
1098                         pa ^= ((vm_paddr_t)1 << (PAGE_SHIFT + order));
1099                         if (pa < seg->start || pa >= seg->end)
1100                                 break;
1101                         m_buddy = &seg->first_page[atop(pa - seg->start)];
1102                         if (m_buddy->order != order)
1103                                 break;
1104                         fl = (*seg->free_queues)[m_buddy->pool];
1105                         vm_freelist_rem(fl, m_buddy, order);
1106                         if (m_buddy->pool != m->pool)
1107                                 vm_phys_set_pool(m->pool, m_buddy, order);
1108                         order++;
1109                         pa &= ~(((vm_paddr_t)1 << (PAGE_SHIFT + order)) - 1);
1110                         m = &seg->first_page[atop(pa - seg->start)];
1111                 } while (order < VM_NFREEORDER - 1);
1112         }
1113         fl = (*seg->free_queues)[m->pool];
1114         vm_freelist_add(fl, m, order, 1);
1115 }
1116
1117 /*
1118  * Return the largest possible order of a set of pages starting at m.
1119  */
1120 static int
1121 max_order(vm_page_t m)
1122 {
1123
1124         /*
1125          * Unsigned "min" is used here so that "order" is assigned
1126          * "VM_NFREEORDER - 1" when "m"'s physical address is zero
1127          * or the low-order bits of its physical address are zero
1128          * because the size of a physical address exceeds the size of
1129          * a long.
1130          */
1131         return (min(ffsl(VM_PAGE_TO_PHYS(m) >> PAGE_SHIFT) - 1,
1132             VM_NFREEORDER - 1));
1133 }
1134
1135 /*
1136  * Free a contiguous, arbitrarily sized set of physical pages, without
1137  * merging across set boundaries.
1138  *
1139  * The free page queues must be locked.
1140  */
1141 void
1142 vm_phys_enqueue_contig(vm_page_t m, u_long npages)
1143 {
1144         struct vm_freelist *fl;
1145         struct vm_phys_seg *seg;
1146         vm_page_t m_end;
1147         int order;
1148
1149         /*
1150          * Avoid unnecessary coalescing by freeing the pages in the largest
1151          * possible power-of-two-sized subsets.
1152          */
1153         vm_domain_free_assert_locked(vm_pagequeue_domain(m));
1154         seg = &vm_phys_segs[m->segind];
1155         fl = (*seg->free_queues)[m->pool];
1156         m_end = m + npages;
1157         /* Free blocks of increasing size. */
1158         while ((order = max_order(m)) < VM_NFREEORDER - 1 &&
1159             m + (1 << order) <= m_end) {
1160                 KASSERT(seg == &vm_phys_segs[m->segind],
1161                     ("%s: page range [%p,%p) spans multiple segments",
1162                     __func__, m_end - npages, m));
1163                 vm_freelist_add(fl, m, order, 1);
1164                 m += 1 << order;
1165         }
1166         /* Free blocks of maximum size. */
1167         while (m + (1 << order) <= m_end) {
1168                 KASSERT(seg == &vm_phys_segs[m->segind],
1169                     ("%s: page range [%p,%p) spans multiple segments",
1170                     __func__, m_end - npages, m));
1171                 vm_freelist_add(fl, m, order, 1);
1172                 m += 1 << order;
1173         }
1174         /* Free blocks of diminishing size. */
1175         while (m < m_end) {
1176                 KASSERT(seg == &vm_phys_segs[m->segind],
1177                     ("%s: page range [%p,%p) spans multiple segments",
1178                     __func__, m_end - npages, m));
1179                 order = flsl(m_end - m) - 1;
1180                 vm_freelist_add(fl, m, order, 1);
1181                 m += 1 << order;
1182         }
1183 }
1184
1185 /*
1186  * Free a contiguous, arbitrarily sized set of physical pages.
1187  *
1188  * The free page queues must be locked.
1189  */
1190 void
1191 vm_phys_free_contig(vm_page_t m, u_long npages)
1192 {
1193         int order_start, order_end;
1194         vm_page_t m_start, m_end;
1195
1196         vm_domain_free_assert_locked(vm_pagequeue_domain(m));
1197
1198         m_start = m;
1199         order_start = max_order(m_start);
1200         if (order_start < VM_NFREEORDER - 1)
1201                 m_start += 1 << order_start;
1202         m_end = m + npages;
1203         order_end = max_order(m_end);
1204         if (order_end < VM_NFREEORDER - 1)
1205                 m_end -= 1 << order_end;
1206         /*
1207          * Avoid unnecessary coalescing by freeing the pages at the start and
1208          * end of the range last.
1209          */
1210         if (m_start < m_end)
1211                 vm_phys_enqueue_contig(m_start, m_end - m_start);
1212         if (order_start < VM_NFREEORDER - 1)
1213                 vm_phys_free_pages(m, order_start);
1214         if (order_end < VM_NFREEORDER - 1)
1215                 vm_phys_free_pages(m_end, order_end);
1216 }
1217
1218 /*
1219  * Scan physical memory between the specified addresses "low" and "high" for a
1220  * run of contiguous physical pages that satisfy the specified conditions, and
1221  * return the lowest page in the run.  The specified "alignment" determines
1222  * the alignment of the lowest physical page in the run.  If the specified
1223  * "boundary" is non-zero, then the run of physical pages cannot span a
1224  * physical address that is a multiple of "boundary".
1225  *
1226  * "npages" must be greater than zero.  Both "alignment" and "boundary" must
1227  * be a power of two.
1228  */
1229 vm_page_t
1230 vm_phys_scan_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
1231     u_long alignment, vm_paddr_t boundary, int options)
1232 {
1233         vm_paddr_t pa_end;
1234         vm_page_t m_end, m_run, m_start;
1235         struct vm_phys_seg *seg;
1236         int segind;
1237
1238         KASSERT(npages > 0, ("npages is 0"));
1239         KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1240         KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1241         if (low >= high)
1242                 return (NULL);
1243         for (segind = 0; segind < vm_phys_nsegs; segind++) {
1244                 seg = &vm_phys_segs[segind];
1245                 if (seg->domain != domain)
1246                         continue;
1247                 if (seg->start >= high)
1248                         break;
1249                 if (low >= seg->end)
1250                         continue;
1251                 if (low <= seg->start)
1252                         m_start = seg->first_page;
1253                 else
1254                         m_start = &seg->first_page[atop(low - seg->start)];
1255                 if (high < seg->end)
1256                         pa_end = high;
1257                 else
1258                         pa_end = seg->end;
1259                 if (pa_end - VM_PAGE_TO_PHYS(m_start) < ptoa(npages))
1260                         continue;
1261                 m_end = &seg->first_page[atop(pa_end - seg->start)];
1262                 m_run = vm_page_scan_contig(npages, m_start, m_end,
1263                     alignment, boundary, options);
1264                 if (m_run != NULL)
1265                         return (m_run);
1266         }
1267         return (NULL);
1268 }
1269
1270 /*
1271  * Set the pool for a contiguous, power of two-sized set of physical pages.
1272  */
1273 void
1274 vm_phys_set_pool(int pool, vm_page_t m, int order)
1275 {
1276         vm_page_t m_tmp;
1277
1278         for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++)
1279                 m_tmp->pool = pool;
1280 }
1281
1282 /*
1283  * Search for the given physical page "m" in the free lists.  If the search
1284  * succeeds, remove "m" from the free lists and return TRUE.  Otherwise, return
1285  * FALSE, indicating that "m" is not in the free lists.
1286  *
1287  * The free page queues must be locked.
1288  */
1289 boolean_t
1290 vm_phys_unfree_page(vm_page_t m)
1291 {
1292         struct vm_freelist *fl;
1293         struct vm_phys_seg *seg;
1294         vm_paddr_t pa, pa_half;
1295         vm_page_t m_set, m_tmp;
1296         int order;
1297
1298         /*
1299          * First, find the contiguous, power of two-sized set of free
1300          * physical pages containing the given physical page "m" and
1301          * assign it to "m_set".
1302          */
1303         seg = &vm_phys_segs[m->segind];
1304         vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1305         for (m_set = m, order = 0; m_set->order == VM_NFREEORDER &&
1306             order < VM_NFREEORDER - 1; ) {
1307                 order++;
1308                 pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order));
1309                 if (pa >= seg->start)
1310                         m_set = &seg->first_page[atop(pa - seg->start)];
1311                 else
1312                         return (FALSE);
1313         }
1314         if (m_set->order < order)
1315                 return (FALSE);
1316         if (m_set->order == VM_NFREEORDER)
1317                 return (FALSE);
1318         KASSERT(m_set->order < VM_NFREEORDER,
1319             ("vm_phys_unfree_page: page %p has unexpected order %d",
1320             m_set, m_set->order));
1321
1322         /*
1323          * Next, remove "m_set" from the free lists.  Finally, extract
1324          * "m" from "m_set" using an iterative algorithm: While "m_set"
1325          * is larger than a page, shrink "m_set" by returning the half
1326          * of "m_set" that does not contain "m" to the free lists.
1327          */
1328         fl = (*seg->free_queues)[m_set->pool];
1329         order = m_set->order;
1330         vm_freelist_rem(fl, m_set, order);
1331         while (order > 0) {
1332                 order--;
1333                 pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order));
1334                 if (m->phys_addr < pa_half)
1335                         m_tmp = &seg->first_page[atop(pa_half - seg->start)];
1336                 else {
1337                         m_tmp = m_set;
1338                         m_set = &seg->first_page[atop(pa_half - seg->start)];
1339                 }
1340                 vm_freelist_add(fl, m_tmp, order, 0);
1341         }
1342         KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency"));
1343         return (TRUE);
1344 }
1345
1346 /*
1347  * Allocate a contiguous set of physical pages of the given size
1348  * "npages" from the free lists.  All of the physical pages must be at
1349  * or above the given physical address "low" and below the given
1350  * physical address "high".  The given value "alignment" determines the
1351  * alignment of the first physical page in the set.  If the given value
1352  * "boundary" is non-zero, then the set of physical pages cannot cross
1353  * any physical address boundary that is a multiple of that value.  Both
1354  * "alignment" and "boundary" must be a power of two.
1355  */
1356 vm_page_t
1357 vm_phys_alloc_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
1358     u_long alignment, vm_paddr_t boundary)
1359 {
1360         vm_paddr_t pa_end, pa_start;
1361         vm_page_t m_run;
1362         struct vm_phys_seg *seg;
1363         int segind;
1364
1365         KASSERT(npages > 0, ("npages is 0"));
1366         KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1367         KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1368         vm_domain_free_assert_locked(VM_DOMAIN(domain));
1369         if (low >= high)
1370                 return (NULL);
1371         m_run = NULL;
1372         for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
1373                 seg = &vm_phys_segs[segind];
1374                 if (seg->start >= high || seg->domain != domain)
1375                         continue;
1376                 if (low >= seg->end)
1377                         break;
1378                 if (low <= seg->start)
1379                         pa_start = seg->start;
1380                 else
1381                         pa_start = low;
1382                 if (high < seg->end)
1383                         pa_end = high;
1384                 else
1385                         pa_end = seg->end;
1386                 if (pa_end - pa_start < ptoa(npages))
1387                         continue;
1388                 m_run = vm_phys_alloc_seg_contig(seg, npages, low, high,
1389                     alignment, boundary);
1390                 if (m_run != NULL)
1391                         break;
1392         }
1393         return (m_run);
1394 }
1395
1396 /*
1397  * Allocate a run of contiguous physical pages from the free list for the
1398  * specified segment.
1399  */
1400 static vm_page_t
1401 vm_phys_alloc_seg_contig(struct vm_phys_seg *seg, u_long npages,
1402     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1403 {
1404         struct vm_freelist *fl;
1405         vm_paddr_t pa, pa_end, size;
1406         vm_page_t m, m_ret;
1407         u_long npages_end;
1408         int oind, order, pind;
1409
1410         KASSERT(npages > 0, ("npages is 0"));
1411         KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1412         KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1413         vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1414         /* Compute the queue that is the best fit for npages. */
1415         order = flsl(npages - 1);
1416         /* Search for a run satisfying the specified conditions. */
1417         size = npages << PAGE_SHIFT;
1418         for (oind = min(order, VM_NFREEORDER - 1); oind < VM_NFREEORDER;
1419             oind++) {
1420                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1421                         fl = (*seg->free_queues)[pind];
1422                         TAILQ_FOREACH(m_ret, &fl[oind].pl, listq) {
1423                                 /*
1424                                  * Is the size of this allocation request
1425                                  * larger than the largest block size?
1426                                  */
1427                                 if (order >= VM_NFREEORDER) {
1428                                         /*
1429                                          * Determine if a sufficient number of
1430                                          * subsequent blocks to satisfy the
1431                                          * allocation request are free.
1432                                          */
1433                                         pa = VM_PAGE_TO_PHYS(m_ret);
1434                                         pa_end = pa + size;
1435                                         if (pa_end < pa)
1436                                                 continue;
1437                                         for (;;) {
1438                                                 pa += 1 << (PAGE_SHIFT +
1439                                                     VM_NFREEORDER - 1);
1440                                                 if (pa >= pa_end ||
1441                                                     pa < seg->start ||
1442                                                     pa >= seg->end)
1443                                                         break;
1444                                                 m = &seg->first_page[atop(pa -
1445                                                     seg->start)];
1446                                                 if (m->order != VM_NFREEORDER -
1447                                                     1)
1448                                                         break;
1449                                         }
1450                                         /* If not, go to the next block. */
1451                                         if (pa < pa_end)
1452                                                 continue;
1453                                 }
1454
1455                                 /*
1456                                  * Determine if the blocks are within the
1457                                  * given range, satisfy the given alignment,
1458                                  * and do not cross the given boundary.
1459                                  */
1460                                 pa = VM_PAGE_TO_PHYS(m_ret);
1461                                 pa_end = pa + size;
1462                                 if (pa >= low && pa_end <= high &&
1463                                     (pa & (alignment - 1)) == 0 &&
1464                                     rounddown2(pa ^ (pa_end - 1), boundary) == 0)
1465                                         goto done;
1466                         }
1467                 }
1468         }
1469         return (NULL);
1470 done:
1471         for (m = m_ret; m < &m_ret[npages]; m = &m[1 << oind]) {
1472                 fl = (*seg->free_queues)[m->pool];
1473                 vm_freelist_rem(fl, m, oind);
1474                 if (m->pool != VM_FREEPOOL_DEFAULT)
1475                         vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, oind);
1476         }
1477         /* Return excess pages to the free lists. */
1478         npages_end = roundup2(npages, 1 << oind);
1479         if (npages < npages_end) {
1480                 fl = (*seg->free_queues)[VM_FREEPOOL_DEFAULT];
1481                 vm_phys_enq_range(&m_ret[npages], npages_end - npages, fl, 0);
1482         }
1483         return (m_ret);
1484 }
1485
1486 #ifdef DDB
1487 /*
1488  * Show the number of physical pages in each of the free lists.
1489  */
1490 DB_SHOW_COMMAND(freepages, db_show_freepages)
1491 {
1492         struct vm_freelist *fl;
1493         int flind, oind, pind, dom;
1494
1495         for (dom = 0; dom < vm_ndomains; dom++) {
1496                 db_printf("DOMAIN: %d\n", dom);
1497                 for (flind = 0; flind < vm_nfreelists; flind++) {
1498                         db_printf("FREE LIST %d:\n"
1499                             "\n  ORDER (SIZE)  |  NUMBER"
1500                             "\n              ", flind);
1501                         for (pind = 0; pind < VM_NFREEPOOL; pind++)
1502                                 db_printf("  |  POOL %d", pind);
1503                         db_printf("\n--            ");
1504                         for (pind = 0; pind < VM_NFREEPOOL; pind++)
1505                                 db_printf("-- --      ");
1506                         db_printf("--\n");
1507                         for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
1508                                 db_printf("  %2.2d (%6.6dK)", oind,
1509                                     1 << (PAGE_SHIFT - 10 + oind));
1510                                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1511                                 fl = vm_phys_free_queues[dom][flind][pind];
1512                                         db_printf("  |  %6.6d", fl[oind].lcnt);
1513                                 }
1514                                 db_printf("\n");
1515                         }
1516                         db_printf("\n");
1517                 }
1518                 db_printf("\n");
1519         }
1520 }
1521 #endif