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