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 1.17 2004/10/26 04:33:11 dillon Exp $
  16  */
  17
  18 #include <sys/param.h>
  19 #include <sys/systm.h>
  20 #include <sys/kernel.h>
  21 #include <sys/lock.h>
  22 #include <sys/malloc.h>
  23 #include <sys/sysctl.h>
  24 #include <sys/vmmeter.h>
  25
  26 #include <vm/vm.h>
  27 #include <vm/vm_object.h>
  28 #include <vm/vm_page.h>
  29 #include <vm/vm_map.h>
  30 #include <vm/vm_kern.h>
  31 #include <vm/vm_extern.h>
  32 #include <vm/vm_zone.h>
  33
  34 static MALLOC_DEFINE(M_ZONE, "ZONE", "Zone header");
  35
  36 #define ZONE_ERROR_INVALID 0
  37 #define ZONE_ERROR_NOTFREE 1
  38 #define ZONE_ERROR_ALREADYFREE 2
  39
  40 #define ZONE_ROUNDING   32
  41
  42 #define ZENTRY_FREE     0x12342378
  43
  44 static void *zget(vm_zone_t z);
  45
  46 /*
  47  * Return an item from the specified zone.   This function is interrupt/MP
  48  * thread safe, but might block.
  49  */
  50 void *
  51 zalloc(vm_zone_t z)
  52 {
  53         void *item;
  54         lwkt_tokref ilock;
  55
  56 #ifdef INVARIANTS
  57         if (z == NULL)
  58                 zerror(ZONE_ERROR_INVALID);
  59 #endif
  60         lwkt_gettoken(&ilock, &z->zlock);
  61         if (z->zfreecnt <= z->zfreemin) {
  62                 item = zget(z);
  63                 /*
  64                  * PANICFAIL allows the caller to assume that the zalloc()
  65                  * will always succeed.  If it doesn't, we panic here.
  66                  */
  67                 if (item == NULL && (z->zflags & ZONE_PANICFAIL))
  68                         panic("zalloc(%s) failed", z->zname);
  69         } else {
  70                 item = z->zitems;
  71 #ifdef INVARIANTS
  72                 KASSERT(item != NULL, ("zitems unexpectedly NULL"));
  73                 if (((void **) item)[1] != (void *) ZENTRY_FREE)
  74                         zerror(ZONE_ERROR_NOTFREE);
  75                 ((void **) item)[1] = 0;
  76 #endif
  77                 z->zitems = ((void **) item)[0];
  78                 z->zfreecnt--;
  79                 z->znalloc++;
  80         }
  81         lwkt_reltoken(&ilock);
  82         return item;
  83 }
  84
  85 /*
  86  * Free an item to the specified zone.   This function is interrupt/MP
  87  * thread safe, but might block.
  88  */
  89 void
  90 zfree(vm_zone_t z, void *item)
  91 {
  92         lwkt_tokref ilock;
  93
  94         lwkt_gettoken(&ilock, &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         lwkt_reltoken(&ilock);
 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 static struct vm_zone *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         if ((z->zflags & ZONE_BOOT) == 0) {
 156                 z->zsize = (size + ZONE_ROUNDING - 1) & ~(ZONE_ROUNDING - 1);
 157                 lwkt_token_init(&z->zlock);
 158                 z->zfreecnt = 0;
 159                 z->ztotal = 0;
 160                 z->zmax = 0;
 161                 z->zname = name;
 162                 z->znalloc = 0;
 163                 z->zitems = NULL;
 164
 165                 z->znext = zlist;
 166                 zlist = z;
 167         }
 168
 169         z->zflags |= flags;
 170
 171         /*
 172          * If we cannot wait, allocate KVA space up front, and we will fill
 173          * in pages as needed.  This is particularly required when creating
 174          * an allocation space for map entries in kernel_map, because we
 175          * do not want to go into a recursion deadlock with
 176          * vm_map_entry_reserve().
 177          */
 178         if (z->zflags & ZONE_INTERRUPT) {
 179
 180                 totsize = round_page(z->zsize * nentries);
 181                 zone_kmem_kvaspace += totsize;
 182
 183                 z->zkva = kmem_alloc_pageable(kernel_map, totsize);
 184                 if (z->zkva == 0) {
 185                         zlist = z->znext;
 186                         return 0;
 187                 }
 188
 189                 z->zpagemax = totsize / PAGE_SIZE;
 190                 if (obj == NULL) {
 191                         z->zobj = vm_object_allocate(OBJT_DEFAULT, z->zpagemax);
 192                 } else {
 193                         z->zobj = obj;
 194                         _vm_object_allocate(OBJT_DEFAULT, z->zpagemax, obj);
 195                 }
 196                 z->zallocflag = VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT;
 197                 z->zmax += nentries;
 198         } else {
 199                 z->zallocflag = VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM;
 200                 z->zmax = 0;
 201         }
 202
 203
 204         if (z->zsize > PAGE_SIZE)
 205                 z->zfreemin = 1;
 206         else
 207                 z->zfreemin = PAGE_SIZE / z->zsize;
 208
 209         z->zpagecount = 0;
 210         if (zalloc)
 211                 z->zalloc = zalloc;
 212         else
 213                 z->zalloc = 1;
 214
 215         return 1;
 216 }
 217
 218 /*
 219  * Subroutine same as zinitna, except zone data structure is allocated
 220  * automatically by malloc.  This routine should normally be used, except
 221  * in certain tricky startup conditions in the VM system -- then
 222  * zbootinit and zinitna can be used.  Zinit is the standard zone
 223  * initialization call.
 224  */
 225 vm_zone_t
 226 zinit(char *name, int size, int nentries, int flags, int zalloc)
 227 {
 228         vm_zone_t z;
 229
 230         z = (vm_zone_t) malloc(sizeof (struct vm_zone), M_ZONE, M_NOWAIT);
 231         if (z == NULL)
 232                 return NULL;
 233
 234         z->zflags = 0;
 235         if (zinitna(z, NULL, name, size, nentries, flags, zalloc) == 0) {
 236                 free(z, M_ZONE);
 237                 return NULL;
 238         }
 239
 240         return z;
 241 }
 242
 243 /*
 244  * Initialize a zone before the system is fully up.  This routine should
 245  * only be called before full VM startup.
 246  */
 247 void
 248 zbootinit(vm_zone_t z, char *name, int size, void *item, int nitems)
 249 {
 250         int i;
 251
 252         z->zname = name;
 253         z->zsize = size;
 254         z->zpagemax = 0;
 255         z->zobj = NULL;
 256         z->zflags = ZONE_BOOT;
 257         z->zfreemin = 0;
 258         z->zallocflag = 0;
 259         z->zpagecount = 0;
 260         z->zalloc = 0;
 261         z->znalloc = 0;
 262         lwkt_token_init(&z->zlock);
 263
 264         bzero(item, nitems * z->zsize);
 265         z->zitems = NULL;
 266         for (i = 0; i < nitems; i++) {
 267                 ((void **) item)[0] = z->zitems;
 268 #ifdef INVARIANTS
 269                 ((void **) item)[1] = (void *) ZENTRY_FREE;
 270 #endif
 271                 z->zitems = item;
 272                 item = (uint8_t *)item + z->zsize;
 273         }
 274         z->zfreecnt = nitems;
 275         z->zmax = nitems;
 276         z->ztotal = nitems;
 277
 278         if (zlist == 0) {
 279                 zlist = z;
 280         } else {
 281                 z->znext = zlist;
 282                 zlist = z;
 283         }
 284 }
 285
 286 /*
 287  * void *zalloc(vm_zone_t zone) --
 288  *      Returns an item from a specified zone.  May not be called from a
 289  *      FAST interrupt or IPI function.
 290  *
 291  * void zfree(vm_zone_t zone, void *item) --
 292  *      Frees an item back to a specified zone.  May not be called from a
 293  *      FAST interrupt or IPI function.
 294  */
 295
 296 /*
 297  * Internal zone routine.  Not to be called from external (non vm_zone) code.
 298  */
 299 static void *
 300 zget(vm_zone_t z)
 301 {
 302         int i;
 303         vm_page_t m;
 304         int nitems, nbytes;
 305         void *item;
 306
 307         if (z == NULL)
 308                 panic("zget: null zone");
 309
 310         if (z->zflags & ZONE_INTERRUPT) {
 311                 /*
 312                  * Interrupt zones do not mess with the kernel_map, they
 313                  * simply populate an existing mapping.
 314                  */
 315                 nbytes = z->zpagecount * PAGE_SIZE;
 316                 nbytes -= nbytes % z->zsize;
 317                 item = (char *) z->zkva + nbytes;
 318                 for (i = 0; ((i < z->zalloc) && (z->zpagecount < z->zpagemax));
 319                      i++) {
 320                         vm_offset_t zkva;
 321
 322                         m = vm_page_alloc(z->zobj, z->zpagecount,
 323                                           z->zallocflag);
 324                         /* note: z might be modified due to blocking */
 325                         if (m == NULL)
 326                                 break;
 327
 328                         zkva = z->zkva + z->zpagecount * PAGE_SIZE;
 329                         pmap_kenter(zkva, VM_PAGE_TO_PHYS(m)); /* YYY */
 330                         bzero((caddr_t) zkva, PAGE_SIZE);
 331                         z->zpagecount++;
 332                         zone_kmem_pages++;
 333                         vmstats.v_wire_count++;
 334                 }
 335                 nitems = ((z->zpagecount * PAGE_SIZE) - nbytes) / z->zsize;
 336         } else if (z->zflags & ZONE_SPECIAL) {
 337                 /*
 338                  * The special zone is the one used for vm_map_entry_t's.
 339                  * We have to avoid an infinite recursion in
 340                  * vm_map_entry_reserve() by using vm_map_entry_kreserve()
 341                  * instead.  The map entries are pre-reserved by the kernel
 342                  * by vm_map_entry_reserve_cpu_init().
 343                  */
 344                 nbytes = z->zalloc * PAGE_SIZE;
 345
 346                 item = (void *)kmem_alloc3(kernel_map, nbytes, KM_KRESERVE);
 347
 348                 /* note: z might be modified due to blocking */
 349                 if (item != NULL) {
 350                         zone_kern_pages += z->zalloc;
 351                         bzero(item, nbytes);
 352                 } else {
 353                         nbytes = 0;
 354                 }
 355                 nitems = nbytes / z->zsize;
 356         } else {
 357                 /*
 358                  * Otherwise allocate KVA from the kernel_map.
 359                  */
 360                 nbytes = z->zalloc * PAGE_SIZE;
 361
 362                 item = (void *)kmem_alloc3(kernel_map, nbytes, 0);
 363
 364                 /* note: z might be modified due to blocking */
 365                 if (item != NULL) {
 366                         zone_kern_pages += z->zalloc;
 367                         bzero(item, nbytes);
 368                 } else {
 369                         nbytes = 0;
 370                 }
 371                 nitems = nbytes / z->zsize;
 372         }
 373         z->ztotal += nitems;
 374
 375         /*
 376          * Save one for immediate allocation
 377          */
 378         if (nitems != 0) {
 379                 nitems -= 1;
 380                 for (i = 0; i < nitems; i++) {
 381                         ((void **) item)[0] = z->zitems;
 382 #ifdef INVARIANTS
 383                         ((void **) item)[1] = (void *) ZENTRY_FREE;
 384 #endif
 385                         z->zitems = item;
 386                         item = (uint8_t *)item + z->zsize;
 387                 }
 388                 z->zfreecnt += nitems;
 389                 z->znalloc++;
 390         } else if (z->zfreecnt > 0) {
 391                 item = z->zitems;
 392                 z->zitems = ((void **) item)[0];
 393 #ifdef INVARIANTS
 394                 if (((void **) item)[1] != (void *) ZENTRY_FREE)
 395                         zerror(ZONE_ERROR_NOTFREE);
 396                 ((void **) item)[1] = 0;
 397 #endif
 398                 z->zfreecnt--;
 399                 z->znalloc++;
 400         } else {
 401                 item = NULL;
 402         }
 403
 404         /*
 405          * A special zone may have used a kernel-reserved vm_map_entry.  If
 406          * so we have to be sure to recover our reserve so we don't run out.
 407          * We will panic if we run out.
 408          */
 409         if (z->zflags & ZONE_SPECIAL)
 410                 vm_map_entry_reserve(0);
 411
 412         return item;
 413 }
 414
 415 static int
 416 sysctl_vm_zone(SYSCTL_HANDLER_ARGS)
 417 {
 418         int error=0;
 419         vm_zone_t curzone, nextzone;
 420         char tmpbuf[128];
 421         char tmpname[14];
 422
 423         snprintf(tmpbuf, sizeof(tmpbuf),
 424             "\nITEM            SIZE     LIMIT    USED    FREE  REQUESTS\n");
 425         error = SYSCTL_OUT(req, tmpbuf, strlen(tmpbuf));
 426         if (error)
 427                 return (error);
 428
 429         for (curzone = zlist; curzone; curzone = nextzone) {
 430                 int i;
 431                 int len;
 432                 int offset;
 433
 434                 nextzone = curzone->znext;
 435                 len = strlen(curzone->zname);
 436                 if (len >= (sizeof(tmpname) - 1))
 437                         len = (sizeof(tmpname) - 1);
 438                 for(i = 0; i < sizeof(tmpname) - 1; i++)
 439                         tmpname[i] = ' ';
 440                 tmpname[i] = 0;
 441                 memcpy(tmpname, curzone->zname, len);
 442                 tmpname[len] = ':';
 443                 offset = 0;
 444                 if (curzone == zlist) {
 445                         offset = 1;
 446                         tmpbuf[0] = '\n';
 447                 }
 448
 449                 snprintf(tmpbuf + offset, sizeof(tmpbuf) - offset,
 450                         "%s %6.6u, %8.8u, %6.6u, %6.6u, %8.8u\n",
 451                         tmpname, curzone->zsize, curzone->zmax,
 452                         (curzone->ztotal - curzone->zfreecnt),
 453                         curzone->zfreecnt, curzone->znalloc);
 454
 455                 len = strlen((char *)tmpbuf);
 456                 if (nextzone == NULL)
 457                         tmpbuf[len - 1] = 0;
 458
 459                 error = SYSCTL_OUT(req, tmpbuf, len);
 460
 461                 if (error)
 462                         return (error);
 463         }
 464         return (0);
 465 }
 466
 467 #if defined(INVARIANTS)
 468 void
 469 zerror(int error)
 470 {
 471         char *msg;
 472
 473         switch (error) {
 474         case ZONE_ERROR_INVALID:
 475                 msg = "zone: invalid zone";
 476                 break;
 477         case ZONE_ERROR_NOTFREE:
 478                 msg = "zone: entry not free";
 479                 break;
 480         case ZONE_ERROR_ALREADYFREE:
 481                 msg = "zone: freeing free entry";
 482                 break;
 483         default:
 484                 msg = "zone: invalid error";
 485                 break;
 486         }
 487         panic(msg);
 488 }
 489 #endif
 490
 491 SYSCTL_OID(_vm, OID_AUTO, zone, CTLTYPE_STRING|CTLFLAG_RD, \
 492         NULL, 0, sysctl_vm_zone, "A", "Zone Info");
 493
 494 SYSCTL_INT(_vm, OID_AUTO, zone_kmem_pages,
 495         CTLFLAG_RD, &zone_kmem_pages, 0, "Number of interrupt safe pages allocated by zone");
 496 SYSCTL_INT(_vm, OID_AUTO, zone_kmem_kvaspace,
 497         CTLFLAG_RD, &zone_kmem_kvaspace, 0, "KVA space allocated by zone");
 498 SYSCTL_INT(_vm, OID_AUTO, zone_kern_pages,
 499         CTLFLAG_RD, &zone_kern_pages, 0, "Number of non-interrupt safe pages allocated by zone");