sys/vm/vm_zone.c

   1 /*
   2  * Copyright (c) 1997, 1998 John S. Dyson
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *      notice immediately at the beginning of the file, without modification,
  10  *      this list of conditions, and the following disclaimer.
  11  * 2. Absolutely no warranty of function or purpose is made by the author
  12  *      John S. Dyson.
  13  *
  14  * $FreeBSD: src/sys/vm/vm_zone.c,v 1.30.2.6 2002/10/10 19:50:16 dillon Exp $
  15  * $DragonFly: src/sys/vm/vm_zone.c,v 1.28 2008/01/23 17:35:48 nth Exp $
  16  */
  17
  18 #include <sys/param.h>
  19 #include <sys/queue.h>
  20 #include <sys/systm.h>
  21 #include <sys/kernel.h>
  22 #include <sys/lock.h>
  23 #include <sys/malloc.h>
  24 #include <sys/sysctl.h>
  25 #include <sys/vmmeter.h>
  26
  27 #include <vm/vm.h>
  28 #include <vm/vm_object.h>
  29 #include <vm/vm_page.h>
  30 #include <vm/vm_map.h>
  31 #include <vm/vm_kern.h>
  32 #include <vm/vm_extern.h>
  33 #include <vm/vm_zone.h>
  34
  35 #include <sys/spinlock2.h>
  36 #include <sys/mplock2.h>
  37
  38 static MALLOC_DEFINE(M_ZONE, "ZONE", "Zone header");
  39
  40 #define ZONE_ERROR_INVALID 0
  41 #define ZONE_ERROR_NOTFREE 1
  42 #define ZONE_ERROR_ALREADYFREE 2
  43
  44 #define ZONE_ROUNDING   32
  45
  46 #define ZENTRY_FREE     0x12342378
  47
  48 static void *zget(vm_zone_t z);
  49
  50 /*
  51  * Return an item from the specified zone.   This function is non-blocking for
  52  * ZONE_INTERRUPT zones.
  53  */
  54 void *
  55 zalloc(vm_zone_t z)
  56 {
  57         void *item;
  58
  59 #ifdef INVARIANTS
  60         if (z == NULL)
  61                 zerror(ZONE_ERROR_INVALID);
  62 #endif
  63         spin_lock_wr(&z->zlock);
  64         if (z->zfreecnt > z->zfreemin) {
  65                 item = z->zitems;
  66 #ifdef INVARIANTS
  67                 KASSERT(item != NULL, ("zitems unexpectedly NULL"));
  68                 if (((void **) item)[1] != (void *) ZENTRY_FREE)
  69                         zerror(ZONE_ERROR_NOTFREE);
  70                 ((void **) item)[1] = 0;
  71 #endif
  72                 z->zitems = ((void **) item)[0];
  73                 z->zfreecnt--;
  74                 z->znalloc++;
  75                 spin_unlock_wr(&z->zlock);
  76         } else {
  77                 spin_unlock_wr(&z->zlock);
  78                 item = zget(z);
  79                 /*
  80                  * PANICFAIL allows the caller to assume that the zalloc()
  81                  * will always succeed.  If it doesn't, we panic here.
  82                  */
  83                 if (item == NULL && (z->zflags & ZONE_PANICFAIL))
  84                         panic("zalloc(%s) failed", z->zname);
  85         }
  86         return item;
  87 }
  88
  89 /*
  90  * Free an item to the specified zone.
  91  */
  92 void
  93 zfree(vm_zone_t z, void *item)
  94 {
  95
  96         spin_lock_wr(&z->zlock);
  97         ((void **) item)[0] = z->zitems;
  98 #ifdef INVARIANTS
  99         if (((void **) item)[1] == (void *) ZENTRY_FREE)
 100                 zerror(ZONE_ERROR_ALREADYFREE);
 101         ((void **) item)[1] = (void *) ZENTRY_FREE;
 102 #endif
 103         z->zitems = item;
 104         z->zfreecnt++;
 105         spin_unlock_wr(&z->zlock);
 106 }
 107
 108 /*
 109  * This file comprises a very simple zone allocator.  This is used
 110  * in lieu of the malloc allocator, where needed or more optimal.
 111  *
 112  * Note that the initial implementation of this had coloring, and
 113  * absolutely no improvement (actually perf degradation) occurred.
 114  *
 115  * Note also that the zones are type stable.  The only restriction is
 116  * that the first two longwords of a data structure can be changed
 117  * between allocations.  Any data that must be stable between allocations
 118  * must reside in areas after the first two longwords.
 119  *
 120  * zinitna, zinit, zbootinit are the initialization routines.
 121  * zalloc, zfree, are the allocation/free routines.
 122  */
 123
 124 LIST_HEAD(zlist, vm_zone) zlist = LIST_HEAD_INITIALIZER(zlist);
 125 static int sysctl_vm_zone(SYSCTL_HANDLER_ARGS);
 126 static int zone_kmem_pages, zone_kern_pages, zone_kmem_kvaspace;
 127
 128 /*
 129  * Create a zone, but don't allocate the zone structure.  If the
 130  * zone had been previously created by the zone boot code, initialize
 131  * various parts of the zone code.
 132  *
 133  * If waits are not allowed during allocation (e.g. during interrupt
 134  * code), a-priori allocate the kernel virtual space, and allocate
 135  * only pages when needed.
 136  *
 137  * Arguments:
 138  * z            pointer to zone structure.
 139  * obj          pointer to VM object (opt).
 140  * name         name of zone.
 141  * size         size of zone entries.
 142  * nentries     number of zone entries allocated (only ZONE_INTERRUPT.)
 143  * flags        ZONE_INTERRUPT -- items can be allocated at interrupt time.
 144  * zalloc       number of pages allocated when memory is needed.
 145  *
 146  * Note that when using ZONE_INTERRUPT, the size of the zone is limited
 147  * by the nentries argument.  The size of the memory allocatable is
 148  * unlimited if ZONE_INTERRUPT is not set.
 149  *
 150  */
 151 int
 152 zinitna(vm_zone_t z, vm_object_t obj, char *name, int size,
 153         int nentries, int flags, int zalloc)
 154 {
 155         int totsize;
 156
 157         /*
 158          * Only zones created with zinit() are destroyable.
 159          */
 160         if (z->zflags & ZONE_DESTROYABLE)
 161                 panic("zinitna: can't create destroyable zone");
 162
 163         /*
 164          * NOTE: We can only adjust zsize if we previously did not
 165          *       use zbootinit().
 166          */
 167         if ((z->zflags & ZONE_BOOT) == 0) {
 168                 z->zsize = (size + ZONE_ROUNDING - 1) & ~(ZONE_ROUNDING - 1);
 169                 spin_init(&z->zlock);
 170                 z->zfreecnt = 0;
 171                 z->ztotal = 0;
 172                 z->zmax = 0;
 173                 z->zname = name;
 174                 z->znalloc = 0;
 175                 z->zitems = NULL;
 176
 177                 LIST_INSERT_HEAD(&zlist, z, zlink);
 178         }
 179
 180         z->zkmvec = NULL;
 181         z->zkmcur = z->zkmmax = 0;
 182         z->zflags |= flags;
 183
 184         /*
 185          * If we cannot wait, allocate KVA space up front, and we will fill
 186          * in pages as needed.  This is particularly required when creating
 187          * an allocation space for map entries in kernel_map, because we
 188          * do not want to go into a recursion deadlock with
 189          * vm_map_entry_reserve().
 190          */
 191         if (z->zflags & ZONE_INTERRUPT) {
 192                 totsize = round_page(z->zsize * nentries);
 193                 zone_kmem_kvaspace += totsize;
 194
 195                 z->zkva = kmem_alloc_pageable(&kernel_map, totsize);
 196                 if (z->zkva == 0) {
 197                         LIST_REMOVE(z, zlink);
 198                         return 0;
 199                 }
 200
 201                 z->zpagemax = totsize / PAGE_SIZE;
 202                 if (obj == NULL) {
 203                         z->zobj = vm_object_allocate(OBJT_DEFAULT, z->zpagemax);
 204                 } else {
 205                         z->zobj = obj;
 206                         _vm_object_allocate(OBJT_DEFAULT, z->zpagemax, obj);
 207                 }
 208                 z->zallocflag = VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT;
 209                 z->zmax += nentries;
 210         } else {
 211                 z->zallocflag = VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM;
 212                 z->zmax = 0;
 213         }
 214
 215
 216         if (z->zsize > PAGE_SIZE)
 217                 z->zfreemin = 1;
 218         else
 219                 z->zfreemin = PAGE_SIZE / z->zsize;
 220
 221         z->zpagecount = 0;
 222         if (zalloc)
 223                 z->zalloc = zalloc;
 224         else
 225                 z->zalloc = 1;
 226
 227         /*
 228          * Populate the interrrupt zone at creation time rather than
 229          * on first allocation, as this is a potentially long operation.
 230          */
 231         if (z->zflags & ZONE_INTERRUPT) {
 232                 void *buf;
 233
 234                 buf = zget(z);
 235                 zfree(z, buf);
 236         }
 237
 238         return 1;
 239 }
 240
 241 /*
 242  * Subroutine same as zinitna, except zone data structure is allocated
 243  * automatically by malloc.  This routine should normally be used, except
 244  * in certain tricky startup conditions in the VM system -- then
 245  * zbootinit and zinitna can be used.  Zinit is the standard zone
 246  * initialization call.
 247  */
 248 vm_zone_t
 249 zinit(char *name, int size, int nentries, int flags, int zalloc)
 250 {
 251         vm_zone_t z;
 252
 253         z = (vm_zone_t) kmalloc(sizeof (struct vm_zone), M_ZONE, M_NOWAIT);
 254         if (z == NULL)
 255                 return NULL;
 256
 257         z->zflags = 0;
 258         if (zinitna(z, NULL, name, size, nentries,
 259                     flags & ~ZONE_DESTROYABLE, zalloc) == 0) {
 260                 kfree(z, M_ZONE);
 261                 return NULL;
 262         }
 263
 264         if (flags & ZONE_DESTROYABLE)
 265                 z->zflags |= ZONE_DESTROYABLE;
 266
 267         return z;
 268 }
 269
 270 /*
 271  * Initialize a zone before the system is fully up.  This routine should
 272  * only be called before full VM startup.
 273  */
 274 void
 275 zbootinit(vm_zone_t z, char *name, int size, void *item, int nitems)
 276 {
 277         int i;
 278
 279         z->zname = name;
 280         z->zsize = size;
 281         z->zpagemax = 0;
 282         z->zobj = NULL;
 283         z->zflags = ZONE_BOOT;
 284         z->zfreemin = 0;
 285         z->zallocflag = 0;
 286         z->zpagecount = 0;
 287         z->zalloc = 0;
 288         z->znalloc = 0;
 289         spin_init(&z->zlock);
 290
 291         bzero(item, nitems * z->zsize);
 292         z->zitems = NULL;
 293         for (i = 0; i < nitems; i++) {
 294                 ((void **) item)[0] = z->zitems;
 295 #ifdef INVARIANTS
 296                 ((void **) item)[1] = (void *) ZENTRY_FREE;
 297 #endif
 298                 z->zitems = item;
 299                 item = (uint8_t *)item + z->zsize;
 300         }
 301         z->zfreecnt = nitems;
 302         z->zmax = nitems;
 303         z->ztotal = nitems;
 304
 305         LIST_INSERT_HEAD(&zlist, z, zlink);
 306 }
 307
 308 /*
 309  * Release all resources owned by zone created with zinit().
 310  */
 311 void
 312 zdestroy(vm_zone_t z)
 313 {
 314         int i;
 315
 316         if (z == NULL)
 317                 panic("zdestroy: null zone");
 318         if ((z->zflags & ZONE_DESTROYABLE) == 0)
 319                 panic("zdestroy: undestroyable zone");
 320
 321         LIST_REMOVE(z, zlink);
 322
 323         /*
 324          * Release virtual mappings, physical memory and update sysctl stats.
 325          */
 326         if (z->zflags & ZONE_INTERRUPT) {
 327                 /*
 328                  * Free the mapping.
 329                  */
 330                 kmem_free(&kernel_map, z->zkva, z->zpagemax*PAGE_SIZE);
 331                 atomic_subtract_int(&zone_kmem_kvaspace, z->zpagemax*PAGE_SIZE);
 332                 /*
 333                  * Free the backing object and physical pages.
 334                  */
 335                 vm_object_deallocate(z->zobj);
 336                 atomic_subtract_int(&zone_kmem_pages, z->zpagecount);
 337         } else {
 338                 for (i=0; i < z->zkmcur; i++) {
 339                         kmem_free(&kernel_map, z->zkmvec[i],
 340                             z->zalloc*PAGE_SIZE);
 341                         atomic_subtract_int(&zone_kern_pages, z->zalloc);
 342                 }
 343                 if (z->zkmvec != NULL)
 344                         kfree(z->zkmvec, M_ZONE);
 345         }
 346
 347         spin_uninit(&z->zlock);
 348         kfree(z, M_ZONE);
 349 }
 350
 351
 352 /*
 353  * void *zalloc(vm_zone_t zone) --
 354  *      Returns an item from a specified zone.  May not be called from a
 355  *      FAST interrupt or IPI function.
 356  *
 357  * void zfree(vm_zone_t zone, void *item) --
 358  *      Frees an item back to a specified zone.  May not be called from a
 359  *      FAST interrupt or IPI function.
 360  */
 361
 362 /*
 363  * Internal zone routine.  Not to be called from external (non vm_zone) code.
 364  */
 365 static void *
 366 zget(vm_zone_t z)
 367 {
 368         int i;
 369         vm_page_t m;
 370         int nitems, nbytes;
 371         int savezpc;
 372         void *item;
 373
 374         if (z == NULL)
 375                 panic("zget: null zone");
 376
 377         if (z->zflags & ZONE_INTERRUPT) {
 378                 /*
 379                  * Interrupt zones do not mess with the kernel_map, they
 380                  * simply populate an existing mapping.
 381                  */
 382                 get_mplock();
 383                 savezpc = z->zpagecount;
 384                 nbytes = z->zpagecount * PAGE_SIZE;
 385                 nbytes -= nbytes % z->zsize;
 386                 item = (char *) z->zkva + nbytes;
 387                 for (i = 0; ((i < z->zalloc) && (z->zpagecount < z->zpagemax));
 388                      i++) {
 389                         vm_offset_t zkva;
 390
 391                         m = vm_page_alloc(z->zobj, z->zpagecount,
 392                                           z->zallocflag);
 393                         /* note: z might be modified due to blocking */
 394                         if (m == NULL)
 395                                 break;
 396
 397                         /*
 398                          * Unbusy page so it can freed in zdestroy().  Make
 399                          * sure it is not on any queue and so can not be
 400                          * recycled under our feet.
 401                          */
 402                         KKASSERT(m->queue == PQ_NONE);
 403                         vm_page_flag_clear(m, PG_BUSY);
 404
 405                         zkva = z->zkva + z->zpagecount * PAGE_SIZE;
 406                         pmap_kenter(zkva, VM_PAGE_TO_PHYS(m)); /* YYY */
 407                         bzero((void *)zkva, PAGE_SIZE);
 408                         KKASSERT(savezpc == z->zpagecount);
 409                         ++savezpc;
 410                         z->zpagecount++;
 411                         zone_kmem_pages++;
 412                         vmstats.v_wire_count++;
 413                 }
 414                 nitems = ((z->zpagecount * PAGE_SIZE) - nbytes) / z->zsize;
 415                 rel_mplock();
 416         } else if (z->zflags & ZONE_SPECIAL) {
 417                 /*
 418                  * The special zone is the one used for vm_map_entry_t's.
 419                  * We have to avoid an infinite recursion in
 420                  * vm_map_entry_reserve() by using vm_map_entry_kreserve()
 421                  * instead.  The map entries are pre-reserved by the kernel
 422                  * by vm_map_entry_reserve_cpu_init().
 423                  */
 424                 nbytes = z->zalloc * PAGE_SIZE;
 425
 426                 item = (void *)kmem_alloc3(&kernel_map, nbytes, KM_KRESERVE);
 427
 428                 /* note: z might be modified due to blocking */
 429                 if (item != NULL) {
 430                         zone_kern_pages += z->zalloc;   /* not MP-safe XXX */
 431                         bzero(item, nbytes);
 432                 } else {
 433                         nbytes = 0;
 434                 }
 435                 nitems = nbytes / z->zsize;
 436         } else {
 437                 /*
 438                  * Otherwise allocate KVA from the kernel_map.
 439                  */
 440                 nbytes = z->zalloc * PAGE_SIZE;
 441
 442                 item = (void *)kmem_alloc3(&kernel_map, nbytes, 0);
 443
 444                 /* note: z might be modified due to blocking */
 445                 if (item != NULL) {
 446                         zone_kern_pages += z->zalloc;   /* not MP-safe XXX */
 447                         bzero(item, nbytes);
 448
 449                         if (z->zflags & ZONE_DESTROYABLE) {
 450                                 if (z->zkmcur == z->zkmmax) {
 451                                         z->zkmmax =
 452                                                 z->zkmmax==0 ? 1 : z->zkmmax*2;
 453                                         z->zkmvec = krealloc(z->zkmvec,
 454                                             z->zkmmax * sizeof(z->zkmvec[0]),
 455                                             M_ZONE, M_WAITOK);
 456                                 }
 457                                 z->zkmvec[z->zkmcur++] = (vm_offset_t)item;
 458                         }
 459                 } else {
 460                         nbytes = 0;
 461                 }
 462                 nitems = nbytes / z->zsize;
 463         }
 464
 465         spin_lock_wr(&z->zlock);
 466         z->ztotal += nitems;
 467         /*
 468          * Save one for immediate allocation
 469          */
 470         if (nitems != 0) {
 471                 nitems -= 1;
 472                 for (i = 0; i < nitems; i++) {
 473                         ((void **) item)[0] = z->zitems;
 474 #ifdef INVARIANTS
 475                         ((void **) item)[1] = (void *) ZENTRY_FREE;
 476 #endif
 477                         z->zitems = item;
 478                         item = (uint8_t *)item + z->zsize;
 479                 }
 480                 z->zfreecnt += nitems;
 481                 z->znalloc++;
 482         } else if (z->zfreecnt > 0) {
 483                 item = z->zitems;
 484                 z->zitems = ((void **) item)[0];
 485 #ifdef INVARIANTS
 486                 if (((void **) item)[1] != (void *) ZENTRY_FREE)
 487                         zerror(ZONE_ERROR_NOTFREE);
 488                 ((void **) item)[1] = 0;
 489 #endif
 490                 z->zfreecnt--;
 491                 z->znalloc++;
 492         } else {
 493                 item = NULL;
 494         }
 495         spin_unlock_wr(&z->zlock);
 496
 497         /*
 498          * A special zone may have used a kernel-reserved vm_map_entry.  If
 499          * so we have to be sure to recover our reserve so we don't run out.
 500          * We will panic if we run out.
 501          */
 502         if (z->zflags & ZONE_SPECIAL)
 503                 vm_map_entry_reserve(0);
 504
 505         return item;
 506 }
 507
 508 static int
 509 sysctl_vm_zone(SYSCTL_HANDLER_ARGS)
 510 {
 511         int error=0;
 512         vm_zone_t curzone;
 513         char tmpbuf[128];
 514         char tmpname[14];
 515
 516         ksnprintf(tmpbuf, sizeof(tmpbuf),
 517             "\nITEM            SIZE     LIMIT    USED    FREE  REQUESTS\n");
 518         error = SYSCTL_OUT(req, tmpbuf, strlen(tmpbuf));
 519         if (error)
 520                 return (error);
 521
 522         LIST_FOREACH(curzone, &zlist, zlink) {
 523                 int i;
 524                 int len;
 525                 int offset;
 526
 527                 len = strlen(curzone->zname);
 528                 if (len >= (sizeof(tmpname) - 1))
 529                         len = (sizeof(tmpname) - 1);
 530                 for(i = 0; i < sizeof(tmpname) - 1; i++)
 531                         tmpname[i] = ' ';
 532                 tmpname[i] = 0;
 533                 memcpy(tmpname, curzone->zname, len);
 534                 tmpname[len] = ':';
 535                 offset = 0;
 536                 if (curzone == LIST_FIRST(&zlist)) {
 537                         offset = 1;
 538                         tmpbuf[0] = '\n';
 539                 }
 540
 541                 ksnprintf(tmpbuf + offset, sizeof(tmpbuf) - offset,
 542                         "%s %6.6u, %8.8u, %6.6u, %6.6u, %8.8u\n",
 543                         tmpname, curzone->zsize, curzone->zmax,
 544                         (curzone->ztotal - curzone->zfreecnt),
 545                         curzone->zfreecnt, curzone->znalloc);
 546
 547                 len = strlen((char *)tmpbuf);
 548                 if (LIST_NEXT(curzone, zlink) == NULL)
 549                         tmpbuf[len - 1] = 0;
 550
 551                 error = SYSCTL_OUT(req, tmpbuf, len);
 552
 553                 if (error)
 554                         return (error);
 555         }
 556         return (0);
 557 }
 558
 559 #if defined(INVARIANTS)
 560 void
 561 zerror(int error)
 562 {
 563         char *msg;
 564
 565         switch (error) {
 566         case ZONE_ERROR_INVALID:
 567                 msg = "zone: invalid zone";
 568                 break;
 569         case ZONE_ERROR_NOTFREE:
 570                 msg = "zone: entry not free";
 571                 break;
 572         case ZONE_ERROR_ALREADYFREE:
 573                 msg = "zone: freeing free entry";
 574                 break;
 575         default:
 576                 msg = "zone: invalid error";
 577                 break;
 578         }
 579         panic(msg);
 580 }
 581 #endif
 582
 583 SYSCTL_OID(_vm, OID_AUTO, zone, CTLTYPE_STRING|CTLFLAG_RD, \
 584         NULL, 0, sysctl_vm_zone, "A", "Zone Info");
 585
 586 SYSCTL_INT(_vm, OID_AUTO, zone_kmem_pages,
 587         CTLFLAG_RD, &zone_kmem_pages, 0, "Number of interrupt safe pages allocated by zone");
 588 SYSCTL_INT(_vm, OID_AUTO, zone_kmem_kvaspace,
 589         CTLFLAG_RD, &zone_kmem_kvaspace, 0, "KVA space allocated by zone");
 590 SYSCTL_INT(_vm, OID_AUTO, zone_kern_pages,
 591         CTLFLAG_RD, &zone_kern_pages, 0, "Number of non-interrupt safe pages allocated by zone");