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 /*
  33  *      Physical memory system implementation
  34  *
  35  * Any external functions defined by this module are only to be used by the
  36  * virtual memory system.
  37  */
  38
  39 #include <sys/cdefs.h>
  40 __FBSDID("$FreeBSD$");
  41
  42 #include "opt_ddb.h"
  43 #include "opt_vm.h"
  44
  45 #include <sys/param.h>
  46 #include <sys/systm.h>
  47 #include <sys/lock.h>
  48 #include <sys/kernel.h>
  49 #include <sys/malloc.h>
  50 #include <sys/mutex.h>
  51 #if MAXMEMDOM > 1
  52 #include <sys/proc.h>
  53 #endif
  54 #include <sys/queue.h>
  55 #include <sys/sbuf.h>
  56 #include <sys/sysctl.h>
  57 #include <sys/vmmeter.h>
  58
  59 #include <ddb/ddb.h>
  60
  61 #include <vm/vm.h>
  62 #include <vm/vm_param.h>
  63 #include <vm/vm_kern.h>
  64 #include <vm/vm_object.h>
  65 #include <vm/vm_page.h>
  66 #include <vm/vm_phys.h>
  67
  68 _Static_assert(sizeof(long) * NBBY >= VM_PHYSSEG_MAX,
  69     "Too many physsegs.");
  70
  71 struct mem_affinity *mem_affinity;
  72
  73 int vm_ndomains = 1;
  74
  75 struct vm_phys_seg vm_phys_segs[VM_PHYSSEG_MAX];
  76 int vm_phys_nsegs;
  77
  78 #define VM_PHYS_FICTITIOUS_NSEGS        8
  79 static struct vm_phys_fictitious_seg {
  80         vm_paddr_t      start;
  81         vm_paddr_t      end;
  82         vm_page_t       first_page;
  83 } vm_phys_fictitious_segs[VM_PHYS_FICTITIOUS_NSEGS];
  84 static struct mtx vm_phys_fictitious_reg_mtx;
  85 MALLOC_DEFINE(M_FICT_PAGES, "vm_fictitious", "Fictitious VM pages");
  86
  87 static struct vm_freelist
  88     vm_phys_free_queues[MAXMEMDOM][VM_NFREELIST][VM_NFREEPOOL][VM_NFREEORDER];
  89
  90 static int vm_nfreelists;
  91
  92 /*
  93  * Provides the mapping from VM_FREELIST_* to free list indices (flind).
  94  */
  95 static int vm_freelist_to_flind[VM_NFREELIST];
  96
  97 CTASSERT(VM_FREELIST_DEFAULT == 0);
  98
  99 #ifdef VM_FREELIST_ISADMA
 100 #define VM_ISADMA_BOUNDARY      16777216
 101 #endif
 102 #ifdef VM_FREELIST_DMA32
 103 #define VM_DMA32_BOUNDARY       ((vm_paddr_t)1 << 32)
 104 #endif
 105
 106 /*
 107  * Enforce the assumptions made by vm_phys_add_seg() and vm_phys_init() about
 108  * the ordering of the free list boundaries.
 109  */
 110 #if defined(VM_ISADMA_BOUNDARY) && defined(VM_LOWMEM_BOUNDARY)
 111 CTASSERT(VM_ISADMA_BOUNDARY < VM_LOWMEM_BOUNDARY);
 112 #endif
 113 #if defined(VM_LOWMEM_BOUNDARY) && defined(VM_DMA32_BOUNDARY)
 114 CTASSERT(VM_LOWMEM_BOUNDARY < VM_DMA32_BOUNDARY);
 115 #endif
 116
 117 static int cnt_prezero;
 118 SYSCTL_INT(_vm_stats_misc, OID_AUTO, cnt_prezero, CTLFLAG_RD,
 119     &cnt_prezero, 0, "The number of physical pages prezeroed at idle time");
 120
 121 static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS);
 122 SYSCTL_OID(_vm, OID_AUTO, phys_free, CTLTYPE_STRING | CTLFLAG_RD,
 123     NULL, 0, sysctl_vm_phys_free, "A", "Phys Free Info");
 124
 125 static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS);
 126 SYSCTL_OID(_vm, OID_AUTO, phys_segs, CTLTYPE_STRING | CTLFLAG_RD,
 127     NULL, 0, sysctl_vm_phys_segs, "A", "Phys Seg Info");
 128
 129 SYSCTL_INT(_vm, OID_AUTO, ndomains, CTLFLAG_RD,
 130     &vm_ndomains, 0, "Number of physical memory domains available.");
 131
 132 static vm_page_t vm_phys_alloc_domain_pages(int domain, int flind, int pool,
 133     int order);
 134 static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain);
 135 static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end);
 136 static int vm_phys_paddr_to_segind(vm_paddr_t pa);
 137 static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl,
 138     int order);
 139
 140 static __inline int
 141 vm_rr_selectdomain(void)
 142 {
 143 #if MAXMEMDOM > 1
 144         struct thread *td;
 145
 146         td = curthread;
 147
 148         td->td_dom_rr_idx++;
 149         td->td_dom_rr_idx %= vm_ndomains;
 150         return (td->td_dom_rr_idx);
 151 #else
 152         return (0);
 153 #endif
 154 }
 155
 156 boolean_t
 157 vm_phys_domain_intersects(long mask, vm_paddr_t low, vm_paddr_t high)
 158 {
 159         struct vm_phys_seg *s;
 160         int idx;
 161
 162         while ((idx = ffsl(mask)) != 0) {
 163                 idx--;  /* ffsl counts from 1 */
 164                 mask &= ~(1UL << idx);
 165                 s = &vm_phys_segs[idx];
 166                 if (low < s->end && high > s->start)
 167                         return (TRUE);
 168         }
 169         return (FALSE);
 170 }
 171
 172 /*
 173  * Outputs the state of the physical memory allocator, specifically,
 174  * the amount of physical memory in each free list.
 175  */
 176 static int
 177 sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS)
 178 {
 179         struct sbuf sbuf;
 180         struct vm_freelist *fl;
 181         int dom, error, flind, oind, pind;
 182
 183         error = sysctl_wire_old_buffer(req, 0);
 184         if (error != 0)
 185                 return (error);
 186         sbuf_new_for_sysctl(&sbuf, NULL, 128 * vm_ndomains, req);
 187         for (dom = 0; dom < vm_ndomains; dom++) {
 188                 sbuf_printf(&sbuf,"\nDOMAIN %d:\n", dom);
 189                 for (flind = 0; flind < vm_nfreelists; flind++) {
 190                         sbuf_printf(&sbuf, "\nFREE LIST %d:\n"
 191                             "\n  ORDER (SIZE)  |  NUMBER"
 192                             "\n              ", flind);
 193                         for (pind = 0; pind < VM_NFREEPOOL; pind++)
 194                                 sbuf_printf(&sbuf, "  |  POOL %d", pind);
 195                         sbuf_printf(&sbuf, "\n--            ");
 196                         for (pind = 0; pind < VM_NFREEPOOL; pind++)
 197                                 sbuf_printf(&sbuf, "-- --      ");
 198                         sbuf_printf(&sbuf, "--\n");
 199                         for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
 200                                 sbuf_printf(&sbuf, "  %2d (%6dK)", oind,
 201                                     1 << (PAGE_SHIFT - 10 + oind));
 202                                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 203                                 fl = vm_phys_free_queues[dom][flind][pind];
 204                                         sbuf_printf(&sbuf, "  |  %6d",
 205                                             fl[oind].lcnt);
 206                                 }
 207                                 sbuf_printf(&sbuf, "\n");
 208                         }
 209                 }
 210         }
 211         error = sbuf_finish(&sbuf);
 212         sbuf_delete(&sbuf);
 213         return (error);
 214 }
 215
 216 /*
 217  * Outputs the set of physical memory segments.
 218  */
 219 static int
 220 sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS)
 221 {
 222         struct sbuf sbuf;
 223         struct vm_phys_seg *seg;
 224         int error, segind;
 225
 226         error = sysctl_wire_old_buffer(req, 0);
 227         if (error != 0)
 228                 return (error);
 229         sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
 230         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 231                 sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind);
 232                 seg = &vm_phys_segs[segind];
 233                 sbuf_printf(&sbuf, "start:     %#jx\n",
 234                     (uintmax_t)seg->start);
 235                 sbuf_printf(&sbuf, "end:       %#jx\n",
 236                     (uintmax_t)seg->end);
 237                 sbuf_printf(&sbuf, "domain:    %d\n", seg->domain);
 238                 sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues);
 239         }
 240         error = sbuf_finish(&sbuf);
 241         sbuf_delete(&sbuf);
 242         return (error);
 243 }
 244
 245 static void
 246 vm_freelist_add(struct vm_freelist *fl, vm_page_t m, int order, int tail)
 247 {
 248
 249         m->order = order;
 250         if (tail)
 251                 TAILQ_INSERT_TAIL(&fl[order].pl, m, plinks.q);
 252         else
 253                 TAILQ_INSERT_HEAD(&fl[order].pl, m, plinks.q);
 254         fl[order].lcnt++;
 255 }
 256
 257 static void
 258 vm_freelist_rem(struct vm_freelist *fl, vm_page_t m, int order)
 259 {
 260
 261         TAILQ_REMOVE(&fl[order].pl, m, plinks.q);
 262         fl[order].lcnt--;
 263         m->order = VM_NFREEORDER;
 264 }
 265
 266 /*
 267  * Create a physical memory segment.
 268  */
 269 static void
 270 _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain)
 271 {
 272         struct vm_phys_seg *seg;
 273
 274         KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX,
 275             ("vm_phys_create_seg: increase VM_PHYSSEG_MAX"));
 276         KASSERT(domain < vm_ndomains,
 277             ("vm_phys_create_seg: invalid domain provided"));
 278         seg = &vm_phys_segs[vm_phys_nsegs++];
 279         while (seg > vm_phys_segs && (seg - 1)->start >= end) {
 280                 *seg = *(seg - 1);
 281                 seg--;
 282         }
 283         seg->start = start;
 284         seg->end = end;
 285         seg->domain = domain;
 286 }
 287
 288 static void
 289 vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end)
 290 {
 291         int i;
 292
 293         if (mem_affinity == NULL) {
 294                 _vm_phys_create_seg(start, end, 0);
 295                 return;
 296         }
 297
 298         for (i = 0;; i++) {
 299                 if (mem_affinity[i].end == 0)
 300                         panic("Reached end of affinity info");
 301                 if (mem_affinity[i].end <= start)
 302                         continue;
 303                 if (mem_affinity[i].start > start)
 304                         panic("No affinity info for start %jx",
 305                             (uintmax_t)start);
 306                 if (mem_affinity[i].end >= end) {
 307                         _vm_phys_create_seg(start, end,
 308                             mem_affinity[i].domain);
 309                         break;
 310                 }
 311                 _vm_phys_create_seg(start, mem_affinity[i].end,
 312                     mem_affinity[i].domain);
 313                 start = mem_affinity[i].end;
 314         }
 315 }
 316
 317 /*
 318  * Add a physical memory segment.
 319  */
 320 void
 321 vm_phys_add_seg(vm_paddr_t start, vm_paddr_t end)
 322 {
 323         vm_paddr_t paddr;
 324
 325         KASSERT((start & PAGE_MASK) == 0,
 326             ("vm_phys_define_seg: start is not page aligned"));
 327         KASSERT((end & PAGE_MASK) == 0,
 328             ("vm_phys_define_seg: end is not page aligned"));
 329
 330         /*
 331          * Split the physical memory segment if it spans two or more free
 332          * list boundaries.
 333          */
 334         paddr = start;
 335 #ifdef  VM_FREELIST_ISADMA
 336         if (paddr < VM_ISADMA_BOUNDARY && end > VM_ISADMA_BOUNDARY) {
 337                 vm_phys_create_seg(paddr, VM_ISADMA_BOUNDARY);
 338                 paddr = VM_ISADMA_BOUNDARY;
 339         }
 340 #endif
 341 #ifdef  VM_FREELIST_LOWMEM
 342         if (paddr < VM_LOWMEM_BOUNDARY && end > VM_LOWMEM_BOUNDARY) {
 343                 vm_phys_create_seg(paddr, VM_LOWMEM_BOUNDARY);
 344                 paddr = VM_LOWMEM_BOUNDARY;
 345         }
 346 #endif
 347 #ifdef  VM_FREELIST_DMA32
 348         if (paddr < VM_DMA32_BOUNDARY && end > VM_DMA32_BOUNDARY) {
 349                 vm_phys_create_seg(paddr, VM_DMA32_BOUNDARY);
 350                 paddr = VM_DMA32_BOUNDARY;
 351         }
 352 #endif
 353         vm_phys_create_seg(paddr, end);
 354 }
 355
 356 /*
 357  * Initialize the physical memory allocator.
 358  *
 359  * Requires that vm_page_array is initialized!
 360  */
 361 void
 362 vm_phys_init(void)
 363 {
 364         struct vm_freelist *fl;
 365         struct vm_phys_seg *seg;
 366         u_long npages;
 367         int dom, flind, freelist, oind, pind, segind;
 368
 369         /*
 370          * Compute the number of free lists, and generate the mapping from the
 371          * manifest constants VM_FREELIST_* to the free list indices.
 372          *
 373          * Initially, the entries of vm_freelist_to_flind[] are set to either
 374          * 0 or 1 to indicate which free lists should be created.
 375          */
 376         npages = 0;
 377         for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
 378                 seg = &vm_phys_segs[segind];
 379 #ifdef  VM_FREELIST_ISADMA
 380                 if (seg->end <= VM_ISADMA_BOUNDARY)
 381                         vm_freelist_to_flind[VM_FREELIST_ISADMA] = 1;
 382                 else
 383 #endif
 384 #ifdef  VM_FREELIST_LOWMEM
 385                 if (seg->end <= VM_LOWMEM_BOUNDARY)
 386                         vm_freelist_to_flind[VM_FREELIST_LOWMEM] = 1;
 387                 else
 388 #endif
 389 #ifdef  VM_FREELIST_DMA32
 390                 if (
 391 #ifdef  VM_DMA32_NPAGES_THRESHOLD
 392                     /*
 393                      * Create the DMA32 free list only if the amount of
 394                      * physical memory above physical address 4G exceeds the
 395                      * given threshold.
 396                      */
 397                     npages > VM_DMA32_NPAGES_THRESHOLD &&
 398 #endif
 399                     seg->end <= VM_DMA32_BOUNDARY)
 400                         vm_freelist_to_flind[VM_FREELIST_DMA32] = 1;
 401                 else
 402 #endif
 403                 {
 404                         npages += atop(seg->end - seg->start);
 405                         vm_freelist_to_flind[VM_FREELIST_DEFAULT] = 1;
 406                 }
 407         }
 408         /* Change each entry into a running total of the free lists. */
 409         for (freelist = 1; freelist < VM_NFREELIST; freelist++) {
 410                 vm_freelist_to_flind[freelist] +=
 411                     vm_freelist_to_flind[freelist - 1];
 412         }
 413         vm_nfreelists = vm_freelist_to_flind[VM_NFREELIST - 1];
 414         KASSERT(vm_nfreelists > 0, ("vm_phys_init: no free lists"));
 415         /* Change each entry into a free list index. */
 416         for (freelist = 0; freelist < VM_NFREELIST; freelist++)
 417                 vm_freelist_to_flind[freelist]--;
 418
 419         /*
 420          * Initialize the first_page and free_queues fields of each physical
 421          * memory segment.
 422          */
 423 #ifdef VM_PHYSSEG_SPARSE
 424         npages = 0;
 425 #endif
 426         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 427                 seg = &vm_phys_segs[segind];
 428 #ifdef VM_PHYSSEG_SPARSE
 429                 seg->first_page = &vm_page_array[npages];
 430                 npages += atop(seg->end - seg->start);
 431 #else
 432                 seg->first_page = PHYS_TO_VM_PAGE(seg->start);
 433 #endif
 434 #ifdef  VM_FREELIST_ISADMA
 435                 if (seg->end <= VM_ISADMA_BOUNDARY) {
 436                         flind = vm_freelist_to_flind[VM_FREELIST_ISADMA];
 437                         KASSERT(flind >= 0,
 438                             ("vm_phys_init: ISADMA flind < 0"));
 439                 } else
 440 #endif
 441 #ifdef  VM_FREELIST_LOWMEM
 442                 if (seg->end <= VM_LOWMEM_BOUNDARY) {
 443                         flind = vm_freelist_to_flind[VM_FREELIST_LOWMEM];
 444                         KASSERT(flind >= 0,
 445                             ("vm_phys_init: LOWMEM flind < 0"));
 446                 } else
 447 #endif
 448 #ifdef  VM_FREELIST_DMA32
 449                 if (seg->end <= VM_DMA32_BOUNDARY) {
 450                         flind = vm_freelist_to_flind[VM_FREELIST_DMA32];
 451                         KASSERT(flind >= 0,
 452                             ("vm_phys_init: DMA32 flind < 0"));
 453                 } else
 454 #endif
 455                 {
 456                         flind = vm_freelist_to_flind[VM_FREELIST_DEFAULT];
 457                         KASSERT(flind >= 0,
 458                             ("vm_phys_init: DEFAULT flind < 0"));
 459                 }
 460                 seg->free_queues = &vm_phys_free_queues[seg->domain][flind];
 461         }
 462
 463         /*
 464          * Initialize the free queues.
 465          */
 466         for (dom = 0; dom < vm_ndomains; dom++) {
 467                 for (flind = 0; flind < vm_nfreelists; flind++) {
 468                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 469                                 fl = vm_phys_free_queues[dom][flind][pind];
 470                                 for (oind = 0; oind < VM_NFREEORDER; oind++)
 471                                         TAILQ_INIT(&fl[oind].pl);
 472                         }
 473                 }
 474         }
 475         mtx_init(&vm_phys_fictitious_reg_mtx, "vmfctr", NULL, MTX_DEF);
 476 }
 477
 478 /*
 479  * Split a contiguous, power of two-sized set of physical pages.
 480  */
 481 static __inline void
 482 vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order)
 483 {
 484         vm_page_t m_buddy;
 485
 486         while (oind > order) {
 487                 oind--;
 488                 m_buddy = &m[1 << oind];
 489                 KASSERT(m_buddy->order == VM_NFREEORDER,
 490                     ("vm_phys_split_pages: page %p has unexpected order %d",
 491                     m_buddy, m_buddy->order));
 492                 vm_freelist_add(fl, m_buddy, oind, 0);
 493         }
 494 }
 495
 496 /*
 497  * Initialize a physical page and add it to the free lists.
 498  */
 499 void
 500 vm_phys_add_page(vm_paddr_t pa)
 501 {
 502         vm_page_t m;
 503         struct vm_domain *vmd;
 504
 505         cnt.v_page_count++;
 506         m = vm_phys_paddr_to_vm_page(pa);
 507         m->phys_addr = pa;
 508         m->queue = PQ_NONE;
 509         m->segind = vm_phys_paddr_to_segind(pa);
 510         vmd = vm_phys_domain(m);
 511         vmd->vmd_page_count++;
 512         vmd->vmd_segs |= 1UL << m->segind;
 513         m->flags = PG_FREE;
 514         KASSERT(m->order == VM_NFREEORDER,
 515             ("vm_phys_add_page: page %p has unexpected order %d",
 516             m, m->order));
 517         m->pool = VM_FREEPOOL_DEFAULT;
 518         pmap_page_init(m);
 519         mtx_lock(&vm_page_queue_free_mtx);
 520         vm_phys_freecnt_adj(m, 1);
 521         vm_phys_free_pages(m, 0);
 522         mtx_unlock(&vm_page_queue_free_mtx);
 523 }
 524
 525 /*
 526  * Allocate a contiguous, power of two-sized set of physical pages
 527  * from the free lists.
 528  *
 529  * The free page queues must be locked.
 530  */
 531 vm_page_t
 532 vm_phys_alloc_pages(int pool, int order)
 533 {
 534         vm_page_t m;
 535         int dom, domain, flind;
 536
 537         KASSERT(pool < VM_NFREEPOOL,
 538             ("vm_phys_alloc_pages: pool %d is out of range", pool));
 539         KASSERT(order < VM_NFREEORDER,
 540             ("vm_phys_alloc_pages: order %d is out of range", order));
 541
 542         for (dom = 0; dom < vm_ndomains; dom++) {
 543                 domain = vm_rr_selectdomain();
 544                 for (flind = 0; flind < vm_nfreelists; flind++) {
 545                         m = vm_phys_alloc_domain_pages(domain, flind, pool,
 546                             order);
 547                         if (m != NULL)
 548                                 return (m);
 549                 }
 550         }
 551         return (NULL);
 552 }
 553
 554 /*
 555  * Allocate a contiguous, power of two-sized set of physical pages from the
 556  * specified free list.  The free list must be specified using one of the
 557  * manifest constants VM_FREELIST_*.
 558  *
 559  * The free page queues must be locked.
 560  */
 561 vm_page_t
 562 vm_phys_alloc_freelist_pages(int freelist, int pool, int order)
 563 {
 564         vm_page_t m;
 565         int dom, domain;
 566
 567         KASSERT(freelist < VM_NFREELIST,
 568             ("vm_phys_alloc_freelist_pages: freelist %d is out of range",
 569             freelist));
 570         KASSERT(pool < VM_NFREEPOOL,
 571             ("vm_phys_alloc_freelist_pages: pool %d is out of range", pool));
 572         KASSERT(order < VM_NFREEORDER,
 573             ("vm_phys_alloc_freelist_pages: order %d is out of range", order));
 574         for (dom = 0; dom < vm_ndomains; dom++) {
 575                 domain = vm_rr_selectdomain();
 576                 m = vm_phys_alloc_domain_pages(domain,
 577                     vm_freelist_to_flind[freelist], pool, order);
 578                 if (m != NULL)
 579                         return (m);
 580         }
 581         return (NULL);
 582 }
 583
 584 static vm_page_t
 585 vm_phys_alloc_domain_pages(int domain, int flind, int pool, int order)
 586 {
 587         struct vm_freelist *fl;
 588         struct vm_freelist *alt;
 589         int oind, pind;
 590         vm_page_t m;
 591
 592         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 593         fl = &vm_phys_free_queues[domain][flind][pool][0];
 594         for (oind = order; oind < VM_NFREEORDER; oind++) {
 595                 m = TAILQ_FIRST(&fl[oind].pl);
 596                 if (m != NULL) {
 597                         vm_freelist_rem(fl, m, oind);
 598                         vm_phys_split_pages(m, oind, fl, order);
 599                         return (m);
 600                 }
 601         }
 602
 603         /*
 604          * The given pool was empty.  Find the largest
 605          * contiguous, power-of-two-sized set of pages in any
 606          * pool.  Transfer these pages to the given pool, and
 607          * use them to satisfy the allocation.
 608          */
 609         for (oind = VM_NFREEORDER - 1; oind >= order; oind--) {
 610                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
 611                         alt = &vm_phys_free_queues[domain][flind][pind][0];
 612                         m = TAILQ_FIRST(&alt[oind].pl);
 613                         if (m != NULL) {
 614                                 vm_freelist_rem(alt, m, oind);
 615                                 vm_phys_set_pool(pool, m, oind);
 616                                 vm_phys_split_pages(m, oind, fl, order);
 617                                 return (m);
 618                         }
 619                 }
 620         }
 621         return (NULL);
 622 }
 623
 624 /*
 625  * Find the vm_page corresponding to the given physical address.
 626  */
 627 vm_page_t
 628 vm_phys_paddr_to_vm_page(vm_paddr_t pa)
 629 {
 630         struct vm_phys_seg *seg;
 631         int segind;
 632
 633         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 634                 seg = &vm_phys_segs[segind];
 635                 if (pa >= seg->start && pa < seg->end)
 636                         return (&seg->first_page[atop(pa - seg->start)]);
 637         }
 638         return (NULL);
 639 }
 640
 641 vm_page_t
 642 vm_phys_fictitious_to_vm_page(vm_paddr_t pa)
 643 {
 644         struct vm_phys_fictitious_seg *seg;
 645         vm_page_t m;
 646         int segind;
 647
 648         m = NULL;
 649         for (segind = 0; segind < VM_PHYS_FICTITIOUS_NSEGS; segind++) {
 650                 seg = &vm_phys_fictitious_segs[segind];
 651                 if (pa >= seg->start && pa < seg->end) {
 652                         m = &seg->first_page[atop(pa - seg->start)];
 653                         KASSERT((m->flags & PG_FICTITIOUS) != 0,
 654                             ("%p not fictitious", m));
 655                         break;
 656                 }
 657         }
 658         return (m);
 659 }
 660
 661 int
 662 vm_phys_fictitious_reg_range(vm_paddr_t start, vm_paddr_t end,
 663     vm_memattr_t memattr)
 664 {
 665         struct vm_phys_fictitious_seg *seg;
 666         vm_page_t fp;
 667         long i, page_count;
 668         int segind;
 669 #ifdef VM_PHYSSEG_DENSE
 670         long pi;
 671         boolean_t malloced;
 672 #endif
 673
 674         page_count = (end - start) / PAGE_SIZE;
 675
 676 #ifdef VM_PHYSSEG_DENSE
 677         pi = atop(start);
 678         if (pi >= first_page && pi < vm_page_array_size + first_page) {
 679                 if (atop(end) >= vm_page_array_size + first_page)
 680                         return (EINVAL);
 681                 fp = &vm_page_array[pi - first_page];
 682                 malloced = FALSE;
 683         } else
 684 #endif
 685         {
 686                 fp = malloc(page_count * sizeof(struct vm_page), M_FICT_PAGES,
 687                     M_WAITOK | M_ZERO);
 688 #ifdef VM_PHYSSEG_DENSE
 689                 malloced = TRUE;
 690 #endif
 691         }
 692         for (i = 0; i < page_count; i++) {
 693                 vm_page_initfake(&fp[i], start + PAGE_SIZE * i, memattr);
 694                 fp[i].oflags &= ~VPO_UNMANAGED;
 695                 fp[i].busy_lock = VPB_UNBUSIED;
 696         }
 697         mtx_lock(&vm_phys_fictitious_reg_mtx);
 698         for (segind = 0; segind < VM_PHYS_FICTITIOUS_NSEGS; segind++) {
 699                 seg = &vm_phys_fictitious_segs[segind];
 700                 if (seg->start == 0 && seg->end == 0) {
 701                         seg->start = start;
 702                         seg->end = end;
 703                         seg->first_page = fp;
 704                         mtx_unlock(&vm_phys_fictitious_reg_mtx);
 705                         return (0);
 706                 }
 707         }
 708         mtx_unlock(&vm_phys_fictitious_reg_mtx);
 709 #ifdef VM_PHYSSEG_DENSE
 710         if (malloced)
 711 #endif
 712                 free(fp, M_FICT_PAGES);
 713         return (EBUSY);
 714 }
 715
 716 void
 717 vm_phys_fictitious_unreg_range(vm_paddr_t start, vm_paddr_t end)
 718 {
 719         struct vm_phys_fictitious_seg *seg;
 720         vm_page_t fp;
 721         int segind;
 722 #ifdef VM_PHYSSEG_DENSE
 723         long pi;
 724 #endif
 725
 726 #ifdef VM_PHYSSEG_DENSE
 727         pi = atop(start);
 728 #endif
 729
 730         mtx_lock(&vm_phys_fictitious_reg_mtx);
 731         for (segind = 0; segind < VM_PHYS_FICTITIOUS_NSEGS; segind++) {
 732                 seg = &vm_phys_fictitious_segs[segind];
 733                 if (seg->start == start && seg->end == end) {
 734                         seg->start = seg->end = 0;
 735                         fp = seg->first_page;
 736                         seg->first_page = NULL;
 737                         mtx_unlock(&vm_phys_fictitious_reg_mtx);
 738 #ifdef VM_PHYSSEG_DENSE
 739                         if (pi < first_page || atop(end) >= vm_page_array_size)
 740 #endif
 741                                 free(fp, M_FICT_PAGES);
 742                         return;
 743                 }
 744         }
 745         mtx_unlock(&vm_phys_fictitious_reg_mtx);
 746         KASSERT(0, ("Unregistering not registered fictitious range"));
 747 }
 748
 749 /*
 750  * Find the segment containing the given physical address.
 751  */
 752 static int
 753 vm_phys_paddr_to_segind(vm_paddr_t pa)
 754 {
 755         struct vm_phys_seg *seg;
 756         int segind;
 757
 758         for (segind = 0; segind < vm_phys_nsegs; segind++) {
 759                 seg = &vm_phys_segs[segind];
 760                 if (pa >= seg->start && pa < seg->end)
 761                         return (segind);
 762         }
 763         panic("vm_phys_paddr_to_segind: paddr %#jx is not in any segment" ,
 764             (uintmax_t)pa);
 765 }
 766
 767 /*
 768  * Free a contiguous, power of two-sized set of physical pages.
 769  *
 770  * The free page queues must be locked.
 771  */
 772 void
 773 vm_phys_free_pages(vm_page_t m, int order)
 774 {
 775         struct vm_freelist *fl;
 776         struct vm_phys_seg *seg;
 777         vm_paddr_t pa;
 778         vm_page_t m_buddy;
 779
 780         KASSERT(m->order == VM_NFREEORDER,
 781             ("vm_phys_free_pages: page %p has unexpected order %d",
 782             m, m->order));
 783         KASSERT(m->pool < VM_NFREEPOOL,
 784             ("vm_phys_free_pages: page %p has unexpected pool %d",
 785             m, m->pool));
 786         KASSERT(order < VM_NFREEORDER,
 787             ("vm_phys_free_pages: order %d is out of range", order));
 788         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 789         seg = &vm_phys_segs[m->segind];
 790         if (order < VM_NFREEORDER - 1) {
 791                 pa = VM_PAGE_TO_PHYS(m);
 792                 do {
 793                         pa ^= ((vm_paddr_t)1 << (PAGE_SHIFT + order));
 794                         if (pa < seg->start || pa >= seg->end)
 795                                 break;
 796                         m_buddy = &seg->first_page[atop(pa - seg->start)];
 797                         if (m_buddy->order != order)
 798                                 break;
 799                         fl = (*seg->free_queues)[m_buddy->pool];
 800                         vm_freelist_rem(fl, m_buddy, order);
 801                         if (m_buddy->pool != m->pool)
 802                                 vm_phys_set_pool(m->pool, m_buddy, order);
 803                         order++;
 804                         pa &= ~(((vm_paddr_t)1 << (PAGE_SHIFT + order)) - 1);
 805                         m = &seg->first_page[atop(pa - seg->start)];
 806                 } while (order < VM_NFREEORDER - 1);
 807         }
 808         fl = (*seg->free_queues)[m->pool];
 809         vm_freelist_add(fl, m, order, 1);
 810 }
 811
 812 /*
 813  * Free a contiguous, arbitrarily sized set of physical pages.
 814  *
 815  * The free page queues must be locked.
 816  */
 817 void
 818 vm_phys_free_contig(vm_page_t m, u_long npages)
 819 {
 820         u_int n;
 821         int order;
 822
 823         /*
 824          * Avoid unnecessary coalescing by freeing the pages in the largest
 825          * possible power-of-two-sized subsets.
 826          */
 827         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 828         for (;; npages -= n) {
 829                 /*
 830                  * Unsigned "min" is used here so that "order" is assigned
 831                  * "VM_NFREEORDER - 1" when "m"'s physical address is zero
 832                  * or the low-order bits of its physical address are zero
 833                  * because the size of a physical address exceeds the size of
 834                  * a long.
 835                  */
 836                 order = min(ffsl(VM_PAGE_TO_PHYS(m) >> PAGE_SHIFT) - 1,
 837                     VM_NFREEORDER - 1);
 838                 n = 1 << order;
 839                 if (npages < n)
 840                         break;
 841                 vm_phys_free_pages(m, order);
 842                 m += n;
 843         }
 844         /* The residual "npages" is less than "1 << (VM_NFREEORDER - 1)". */
 845         for (; npages > 0; npages -= n) {
 846                 order = flsl(npages) - 1;
 847                 n = 1 << order;
 848                 vm_phys_free_pages(m, order);
 849                 m += n;
 850         }
 851 }
 852
 853 /*
 854  * Set the pool for a contiguous, power of two-sized set of physical pages.
 855  */
 856 void
 857 vm_phys_set_pool(int pool, vm_page_t m, int order)
 858 {
 859         vm_page_t m_tmp;
 860
 861         for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++)
 862                 m_tmp->pool = pool;
 863 }
 864
 865 /*
 866  * Search for the given physical page "m" in the free lists.  If the search
 867  * succeeds, remove "m" from the free lists and return TRUE.  Otherwise, return
 868  * FALSE, indicating that "m" is not in the free lists.
 869  *
 870  * The free page queues must be locked.
 871  */
 872 boolean_t
 873 vm_phys_unfree_page(vm_page_t m)
 874 {
 875         struct vm_freelist *fl;
 876         struct vm_phys_seg *seg;
 877         vm_paddr_t pa, pa_half;
 878         vm_page_t m_set, m_tmp;
 879         int order;
 880
 881         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 882
 883         /*
 884          * First, find the contiguous, power of two-sized set of free
 885          * physical pages containing the given physical page "m" and
 886          * assign it to "m_set".
 887          */
 888         seg = &vm_phys_segs[m->segind];
 889         for (m_set = m, order = 0; m_set->order == VM_NFREEORDER &&
 890             order < VM_NFREEORDER - 1; ) {
 891                 order++;
 892                 pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order));
 893                 if (pa >= seg->start)
 894                         m_set = &seg->first_page[atop(pa - seg->start)];
 895                 else
 896                         return (FALSE);
 897         }
 898         if (m_set->order < order)
 899                 return (FALSE);
 900         if (m_set->order == VM_NFREEORDER)
 901                 return (FALSE);
 902         KASSERT(m_set->order < VM_NFREEORDER,
 903             ("vm_phys_unfree_page: page %p has unexpected order %d",
 904             m_set, m_set->order));
 905
 906         /*
 907          * Next, remove "m_set" from the free lists.  Finally, extract
 908          * "m" from "m_set" using an iterative algorithm: While "m_set"
 909          * is larger than a page, shrink "m_set" by returning the half
 910          * of "m_set" that does not contain "m" to the free lists.
 911          */
 912         fl = (*seg->free_queues)[m_set->pool];
 913         order = m_set->order;
 914         vm_freelist_rem(fl, m_set, order);
 915         while (order > 0) {
 916                 order--;
 917                 pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order));
 918                 if (m->phys_addr < pa_half)
 919                         m_tmp = &seg->first_page[atop(pa_half - seg->start)];
 920                 else {
 921                         m_tmp = m_set;
 922                         m_set = &seg->first_page[atop(pa_half - seg->start)];
 923                 }
 924                 vm_freelist_add(fl, m_tmp, order, 0);
 925         }
 926         KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency"));
 927         return (TRUE);
 928 }
 929
 930 /*
 931  * Try to zero one physical page.  Used by an idle priority thread.
 932  */
 933 boolean_t
 934 vm_phys_zero_pages_idle(void)
 935 {
 936         static struct vm_freelist *fl;
 937         static int flind, oind, pind;
 938         vm_page_t m, m_tmp;
 939         int domain;
 940
 941         domain = vm_rr_selectdomain();
 942         fl = vm_phys_free_queues[domain][0][0];
 943         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 944         for (;;) {
 945                 TAILQ_FOREACH_REVERSE(m, &fl[oind].pl, pglist, plinks.q) {
 946                         for (m_tmp = m; m_tmp < &m[1 << oind]; m_tmp++) {
 947                                 if ((m_tmp->flags & (PG_CACHED | PG_ZERO)) == 0) {
 948                                         vm_phys_unfree_page(m_tmp);
 949                                         vm_phys_freecnt_adj(m, -1);
 950                                         mtx_unlock(&vm_page_queue_free_mtx);
 951                                         pmap_zero_page_idle(m_tmp);
 952                                         m_tmp->flags |= PG_ZERO;
 953                                         mtx_lock(&vm_page_queue_free_mtx);
 954                                         vm_phys_freecnt_adj(m, 1);
 955                                         vm_phys_free_pages(m_tmp, 0);
 956                                         vm_page_zero_count++;
 957                                         cnt_prezero++;
 958                                         return (TRUE);
 959                                 }
 960                         }
 961                 }
 962                 oind++;
 963                 if (oind == VM_NFREEORDER) {
 964                         oind = 0;
 965                         pind++;
 966                         if (pind == VM_NFREEPOOL) {
 967                                 pind = 0;
 968                                 flind++;
 969                                 if (flind == vm_nfreelists)
 970                                         flind = 0;
 971                         }
 972                         fl = vm_phys_free_queues[domain][flind][pind];
 973                 }
 974         }
 975 }
 976
 977 /*
 978  * Allocate a contiguous set of physical pages of the given size
 979  * "npages" from the free lists.  All of the physical pages must be at
 980  * or above the given physical address "low" and below the given
 981  * physical address "high".  The given value "alignment" determines the
 982  * alignment of the first physical page in the set.  If the given value
 983  * "boundary" is non-zero, then the set of physical pages cannot cross
 984  * any physical address boundary that is a multiple of that value.  Both
 985  * "alignment" and "boundary" must be a power of two.
 986  */
 987 vm_page_t
 988 vm_phys_alloc_contig(u_long npages, vm_paddr_t low, vm_paddr_t high,
 989     u_long alignment, vm_paddr_t boundary)
 990 {
 991         struct vm_freelist *fl;
 992         struct vm_phys_seg *seg;
 993         vm_paddr_t pa, pa_last, size;
 994         vm_page_t m, m_ret;
 995         u_long npages_end;
 996         int dom, domain, flind, oind, order, pind;
 997
 998         mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
 999         size = npages << PAGE_SHIFT;
1000         KASSERT(size != 0,
1001             ("vm_phys_alloc_contig: size must not be 0"));
1002         KASSERT((alignment & (alignment - 1)) == 0,
1003             ("vm_phys_alloc_contig: alignment must be a power of 2"));
1004         KASSERT((boundary & (boundary - 1)) == 0,
1005             ("vm_phys_alloc_contig: boundary must be a power of 2"));
1006         /* Compute the queue that is the best fit for npages. */
1007         for (order = 0; (1 << order) < npages; order++);
1008         dom = 0;
1009 restartdom:
1010         domain = vm_rr_selectdomain();
1011         for (flind = 0; flind < vm_nfreelists; flind++) {
1012                 for (oind = min(order, VM_NFREEORDER - 1); oind < VM_NFREEORDER; oind++) {
1013                         for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1014                                 fl = &vm_phys_free_queues[domain][flind][pind][0];
1015                                 TAILQ_FOREACH(m_ret, &fl[oind].pl, plinks.q) {
1016                                         /*
1017                                          * A free list may contain physical pages
1018                                          * from one or more segments.
1019                                          */
1020                                         seg = &vm_phys_segs[m_ret->segind];
1021                                         if (seg->start > high ||
1022                                             low >= seg->end)
1023                                                 continue;
1024
1025                                         /*
1026                                          * Is the size of this allocation request
1027                                          * larger than the largest block size?
1028                                          */
1029                                         if (order >= VM_NFREEORDER) {
1030                                                 /*
1031                                                  * Determine if a sufficient number
1032                                                  * of subsequent blocks to satisfy
1033                                                  * the allocation request are free.
1034                                                  */
1035                                                 pa = VM_PAGE_TO_PHYS(m_ret);
1036                                                 pa_last = pa + size;
1037                                                 for (;;) {
1038                                                         pa += 1 << (PAGE_SHIFT + VM_NFREEORDER - 1);
1039                                                         if (pa >= pa_last)
1040                                                                 break;
1041                                                         if (pa < seg->start ||
1042                                                             pa >= seg->end)
1043                                                                 break;
1044                                                         m = &seg->first_page[atop(pa - seg->start)];
1045                                                         if (m->order != VM_NFREEORDER - 1)
1046                                                                 break;
1047                                                 }
1048                                                 /* If not, continue to the next block. */
1049                                                 if (pa < pa_last)
1050                                                         continue;
1051                                         }
1052
1053                                         /*
1054                                          * Determine if the blocks are within the given range,
1055                                          * satisfy the given alignment, and do not cross the
1056                                          * given boundary.
1057                                          */
1058                                         pa = VM_PAGE_TO_PHYS(m_ret);
1059                                         if (pa >= low &&
1060                                             pa + size <= high &&
1061                                             (pa & (alignment - 1)) == 0 &&
1062                                             ((pa ^ (pa + size - 1)) & ~(boundary - 1)) == 0)
1063                                                 goto done;
1064                                 }
1065                         }
1066                 }
1067         }
1068         if (++dom < vm_ndomains)
1069                 goto restartdom;
1070         return (NULL);
1071 done:
1072         for (m = m_ret; m < &m_ret[npages]; m = &m[1 << oind]) {
1073                 fl = (*seg->free_queues)[m->pool];
1074                 vm_freelist_rem(fl, m, m->order);
1075         }
1076         if (m_ret->pool != VM_FREEPOOL_DEFAULT)
1077                 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m_ret, oind);
1078         fl = (*seg->free_queues)[m_ret->pool];
1079         vm_phys_split_pages(m_ret, oind, fl, order);
1080         /* Return excess pages to the free lists. */
1081         npages_end = roundup2(npages, 1 << imin(oind, order));
1082         if (npages < npages_end)
1083                 vm_phys_free_contig(&m_ret[npages], npages_end - npages);
1084         return (m_ret);
1085 }
1086
1087 #ifdef DDB
1088 /*
1089  * Show the number of physical pages in each of the free lists.
1090  */
1091 DB_SHOW_COMMAND(freepages, db_show_freepages)
1092 {
1093         struct vm_freelist *fl;
1094         int flind, oind, pind, dom;
1095
1096         for (dom = 0; dom < vm_ndomains; dom++) {
1097                 db_printf("DOMAIN: %d\n", dom);
1098                 for (flind = 0; flind < vm_nfreelists; flind++) {
1099                         db_printf("FREE LIST %d:\n"
1100                             "\n  ORDER (SIZE)  |  NUMBER"
1101                             "\n              ", flind);
1102                         for (pind = 0; pind < VM_NFREEPOOL; pind++)
1103                                 db_printf("  |  POOL %d", pind);
1104                         db_printf("\n--            ");
1105                         for (pind = 0; pind < VM_NFREEPOOL; pind++)
1106                                 db_printf("-- --      ");
1107                         db_printf("--\n");
1108                         for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
1109                                 db_printf("  %2.2d (%6.6dK)", oind,
1110                                     1 << (PAGE_SHIFT - 10 + oind));
1111                                 for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1112                                 fl = vm_phys_free_queues[dom][flind][pind];
1113                                         db_printf("  |  %6.6d", fl[oind].lcnt);
1114                                 }
1115                                 db_printf("\n");
1116                         }
1117                         db_printf("\n");
1118                 }
1119                 db_printf("\n");
1120         }
1121 }
1122 #endif