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 1.24 2007/01/12 03:05:49 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 #include <sys/spinlock2.h>              /* XXX */
  34
  35 static MALLOC_DEFINE(M_ZONE, "ZONE", "Zone header");
  36
  37 #define ZONE_ERROR_INVALID 0
  38 #define ZONE_ERROR_NOTFREE 1
  39 #define ZONE_ERROR_ALREADYFREE 2
  40
  41 #define ZONE_ROUNDING   32
  42
  43 #define ZENTRY_FREE     0x12342378
  44
  45 static void *zget(vm_zone_t z);
  46
  47 /*
  48  * Return an item from the specified zone.   This function is interrupt/MP
  49  * thread safe and is non-blocking for ZONE_INTERRUPT zones.
  50  */
  51 void *
  52 zalloc(vm_zone_t z)
  53 {
  54         void *item;
  55
  56 #ifdef INVARIANTS
  57         if (z == NULL)
  58                 zerror(ZONE_ERROR_INVALID);
  59 #endif
  60         spin_lock_wr(&z->zlock);
  61         if (z->zfreecnt > z->zfreemin) {
  62                 item = z->zitems;
  63 #ifdef INVARIANTS
  64                 KASSERT(item != NULL, ("zitems unexpectedly NULL"));
  65                 if (((void **) item)[1] != (void *) ZENTRY_FREE)
  66                         zerror(ZONE_ERROR_NOTFREE);
  67                 ((void **) item)[1] = 0;
  68 #endif
  69                 z->zitems = ((void **) item)[0];
  70                 z->zfreecnt--;
  71                 z->znalloc++;
  72                 spin_unlock_wr(&z->zlock);
  73         } else {
  74                 spin_unlock_wr(&z->zlock);
  75                 item = zget(z);
  76                 /*
  77                  * PANICFAIL allows the caller to assume that the zalloc()
  78                  * will always succeed.  If it doesn't, we panic here.
  79                  */
  80                 if (item == NULL && (z->zflags & ZONE_PANICFAIL))
  81                         panic("zalloc(%s) failed", z->zname);
  82         }
  83         return item;
  84 }
  85
  86 /*
  87  * Free an item to the specified zone.   This function is interrupt/MP
  88  * thread safe and is non-blocking.
  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 struct vm_zone *zlist;  /* exported to vmstat */
 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                 spin_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                 totsize = round_page(z->zsize * nentries);
 180                 zone_kmem_kvaspace += totsize;
 181
 182                 z->zkva = kmem_alloc_pageable(&kernel_map, totsize);
 183                 if (z->zkva == 0) {
 184                         zlist = z->znext;
 185                         return 0;
 186                 }
 187
 188                 z->zpagemax = totsize / PAGE_SIZE;
 189                 if (obj == NULL) {
 190                         z->zobj = vm_object_allocate(OBJT_DEFAULT, z->zpagemax);
 191                 } else {
 192                         z->zobj = obj;
 193                         _vm_object_allocate(OBJT_DEFAULT, z->zpagemax, obj);
 194                 }
 195                 z->zallocflag = VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT;
 196                 z->zmax += nentries;
 197         } else {
 198                 z->zallocflag = VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM;
 199                 z->zmax = 0;
 200         }
 201
 202
 203         if (z->zsize > PAGE_SIZE)
 204                 z->zfreemin = 1;
 205         else
 206                 z->zfreemin = PAGE_SIZE / z->zsize;
 207
 208         z->zpagecount = 0;
 209         if (zalloc)
 210                 z->zalloc = zalloc;
 211         else
 212                 z->zalloc = 1;
 213
 214         /*
 215          * Populate the interrrupt zone at creation time rather than
 216          * on first allocation, as this is a potentially long operation.
 217          */
 218         if (z->zflags & ZONE_INTERRUPT) {
 219                 void *buf;
 220
 221                 buf = zget(z);
 222                 zfree(z, buf);
 223         }
 224
 225         return 1;
 226 }
 227
 228 /*
 229  * Subroutine same as zinitna, except zone data structure is allocated
 230  * automatically by malloc.  This routine should normally be used, except
 231  * in certain tricky startup conditions in the VM system -- then
 232  * zbootinit and zinitna can be used.  Zinit is the standard zone
 233  * initialization call.
 234  */
 235 vm_zone_t
 236 zinit(char *name, int size, int nentries, int flags, int zalloc)
 237 {
 238         vm_zone_t z;
 239
 240         z = (vm_zone_t) kmalloc(sizeof (struct vm_zone), M_ZONE, M_NOWAIT);
 241         if (z == NULL)
 242                 return NULL;
 243
 244         z->zflags = 0;
 245         if (zinitna(z, NULL, name, size, nentries, flags, zalloc) == 0) {
 246                 kfree(z, M_ZONE);
 247                 return NULL;
 248         }
 249
 250         return z;
 251 }
 252
 253 /*
 254  * Initialize a zone before the system is fully up.  This routine should
 255  * only be called before full VM startup.
 256  */
 257 void
 258 zbootinit(vm_zone_t z, char *name, int size, void *item, int nitems)
 259 {
 260         int i;
 261
 262         z->zname = name;
 263         z->zsize = size;
 264         z->zpagemax = 0;
 265         z->zobj = NULL;
 266         z->zflags = ZONE_BOOT;
 267         z->zfreemin = 0;
 268         z->zallocflag = 0;
 269         z->zpagecount = 0;
 270         z->zalloc = 0;
 271         z->znalloc = 0;
 272         spin_init(&z->zlock);
 273
 274         bzero(item, nitems * z->zsize);
 275         z->zitems = NULL;
 276         for (i = 0; i < nitems; i++) {
 277                 ((void **) item)[0] = z->zitems;
 278 #ifdef INVARIANTS
 279                 ((void **) item)[1] = (void *) ZENTRY_FREE;
 280 #endif
 281                 z->zitems = item;
 282                 item = (uint8_t *)item + z->zsize;
 283         }
 284         z->zfreecnt = nitems;
 285         z->zmax = nitems;
 286         z->ztotal = nitems;
 287
 288         if (zlist == 0) {
 289                 zlist = z;
 290         } else {
 291                 z->znext = zlist;
 292                 zlist = z;
 293         }
 294 }
 295
 296 /*
 297  * void *zalloc(vm_zone_t zone) --
 298  *      Returns an item from a specified zone.  May not be called from a
 299  *      FAST interrupt or IPI function.
 300  *
 301  * void zfree(vm_zone_t zone, void *item) --
 302  *      Frees an item back to a specified zone.  May not be called from a
 303  *      FAST interrupt or IPI function.
 304  */
 305
 306 /*
 307  * Internal zone routine.  Not to be called from external (non vm_zone) code.
 308  */
 309 static void *
 310 zget(vm_zone_t z)
 311 {
 312         int i;
 313         vm_page_t m;
 314         int nitems, nbytes;
 315         void *item;
 316
 317         if (z == NULL)
 318                 panic("zget: null zone");
 319
 320         if (z->zflags & ZONE_INTERRUPT) {
 321                 /*
 322                  * Interrupt zones do not mess with the kernel_map, they
 323                  * simply populate an existing mapping.
 324                  */
 325                 nbytes = z->zpagecount * PAGE_SIZE;
 326                 nbytes -= nbytes % z->zsize;
 327                 item = (char *) z->zkva + nbytes;
 328                 for (i = 0; ((i < z->zalloc) && (z->zpagecount < z->zpagemax));
 329                      i++) {
 330                         vm_offset_t zkva;
 331
 332                         m = vm_page_alloc(z->zobj, z->zpagecount,
 333                                           z->zallocflag);
 334                         /* note: z might be modified due to blocking */
 335                         if (m == NULL)
 336                                 break;
 337
 338                         zkva = z->zkva + z->zpagecount * PAGE_SIZE;
 339                         pmap_kenter(zkva, VM_PAGE_TO_PHYS(m)); /* YYY */
 340                         bzero((void *)zkva, PAGE_SIZE);
 341                         z->zpagecount++;
 342                         zone_kmem_pages++;
 343                         vmstats.v_wire_count++;
 344                 }
 345                 nitems = ((z->zpagecount * PAGE_SIZE) - nbytes) / z->zsize;
 346         } else if (z->zflags & ZONE_SPECIAL) {
 347                 /*
 348                  * The special zone is the one used for vm_map_entry_t's.
 349                  * We have to avoid an infinite recursion in
 350                  * vm_map_entry_reserve() by using vm_map_entry_kreserve()
 351                  * instead.  The map entries are pre-reserved by the kernel
 352                  * by vm_map_entry_reserve_cpu_init().
 353                  */
 354                 nbytes = z->zalloc * PAGE_SIZE;
 355
 356                 item = (void *)kmem_alloc3(&kernel_map, nbytes, KM_KRESERVE);
 357
 358                 /* note: z might be modified due to blocking */
 359                 if (item != NULL) {
 360                         zone_kern_pages += z->zalloc;   /* not MP-safe XXX */
 361                         bzero(item, nbytes);
 362                 } else {
 363                         nbytes = 0;
 364                 }
 365                 nitems = nbytes / z->zsize;
 366         } else {
 367                 /*
 368                  * Otherwise allocate KVA from the kernel_map.
 369                  */
 370                 nbytes = z->zalloc * PAGE_SIZE;
 371
 372                 item = (void *)kmem_alloc3(&kernel_map, nbytes, 0);
 373
 374                 /* note: z might be modified due to blocking */
 375                 if (item != NULL) {
 376                         zone_kern_pages += z->zalloc;   /* not MP-safe XXX */
 377                         bzero(item, nbytes);
 378                 } else {
 379                         nbytes = 0;
 380                 }
 381                 nitems = nbytes / z->zsize;
 382         }
 383
 384         spin_lock_wr(&z->zlock);
 385         z->ztotal += nitems;
 386         /*
 387          * Save one for immediate allocation
 388          */
 389         if (nitems != 0) {
 390                 nitems -= 1;
 391                 for (i = 0; i < nitems; i++) {
 392                         ((void **) item)[0] = z->zitems;
 393 #ifdef INVARIANTS
 394                         ((void **) item)[1] = (void *) ZENTRY_FREE;
 395 #endif
 396                         z->zitems = item;
 397                         item = (uint8_t *)item + z->zsize;
 398                 }
 399                 z->zfreecnt += nitems;
 400                 z->znalloc++;
 401         } else if (z->zfreecnt > 0) {
 402                 item = z->zitems;
 403                 z->zitems = ((void **) item)[0];
 404 #ifdef INVARIANTS
 405                 if (((void **) item)[1] != (void *) ZENTRY_FREE)
 406                         zerror(ZONE_ERROR_NOTFREE);
 407                 ((void **) item)[1] = 0;
 408 #endif
 409                 z->zfreecnt--;
 410                 z->znalloc++;
 411         } else {
 412                 item = NULL;
 413         }
 414         spin_unlock_wr(&z->zlock);
 415
 416         /*
 417          * A special zone may have used a kernel-reserved vm_map_entry.  If
 418          * so we have to be sure to recover our reserve so we don't run out.
 419          * We will panic if we run out.
 420          */
 421         if (z->zflags & ZONE_SPECIAL)
 422                 vm_map_entry_reserve(0);
 423
 424         return item;
 425 }
 426
 427 static int
 428 sysctl_vm_zone(SYSCTL_HANDLER_ARGS)
 429 {
 430         int error=0;
 431         vm_zone_t curzone, nextzone;
 432         char tmpbuf[128];
 433         char tmpname[14];
 434
 435         ksnprintf(tmpbuf, sizeof(tmpbuf),
 436             "\nITEM            SIZE     LIMIT    USED    FREE  REQUESTS\n");
 437         error = SYSCTL_OUT(req, tmpbuf, strlen(tmpbuf));
 438         if (error)
 439                 return (error);
 440
 441         for (curzone = zlist; curzone; curzone = nextzone) {
 442                 int i;
 443                 int len;
 444                 int offset;
 445
 446                 nextzone = curzone->znext;
 447                 len = strlen(curzone->zname);
 448                 if (len >= (sizeof(tmpname) - 1))
 449                         len = (sizeof(tmpname) - 1);
 450                 for(i = 0; i < sizeof(tmpname) - 1; i++)
 451                         tmpname[i] = ' ';
 452                 tmpname[i] = 0;
 453                 memcpy(tmpname, curzone->zname, len);
 454                 tmpname[len] = ':';
 455                 offset = 0;
 456                 if (curzone == zlist) {
 457                         offset = 1;
 458                         tmpbuf[0] = '\n';
 459                 }
 460
 461                 ksnprintf(tmpbuf + offset, sizeof(tmpbuf) - offset,
 462                         "%s %6.6u, %8.8u, %6.6u, %6.6u, %8.8u\n",
 463                         tmpname, curzone->zsize, curzone->zmax,
 464                         (curzone->ztotal - curzone->zfreecnt),
 465                         curzone->zfreecnt, curzone->znalloc);
 466
 467                 len = strlen((char *)tmpbuf);
 468                 if (nextzone == NULL)
 469                         tmpbuf[len - 1] = 0;
 470
 471                 error = SYSCTL_OUT(req, tmpbuf, len);
 472
 473                 if (error)
 474                         return (error);
 475         }
 476         return (0);
 477 }
 478
 479 #if defined(INVARIANTS)
 480 void
 481 zerror(int error)
 482 {
 483         char *msg;
 484
 485         switch (error) {
 486         case ZONE_ERROR_INVALID:
 487                 msg = "zone: invalid zone";
 488                 break;
 489         case ZONE_ERROR_NOTFREE:
 490                 msg = "zone: entry not free";
 491                 break;
 492         case ZONE_ERROR_ALREADYFREE:
 493                 msg = "zone: freeing free entry";
 494                 break;
 495         default:
 496                 msg = "zone: invalid error";
 497                 break;
 498         }
 499         panic(msg);
 500 }
 501 #endif
 502
 503 SYSCTL_OID(_vm, OID_AUTO, zone, CTLTYPE_STRING|CTLFLAG_RD, \
 504         NULL, 0, sysctl_vm_zone, "A", "Zone Info");
 505
 506 SYSCTL_INT(_vm, OID_AUTO, zone_kmem_pages,
 507         CTLFLAG_RD, &zone_kmem_pages, 0, "Number of interrupt safe pages allocated by zone");
 508 SYSCTL_INT(_vm, OID_AUTO, zone_kmem_kvaspace,
 509         CTLFLAG_RD, &zone_kmem_kvaspace, 0, "KVA space allocated by zone");
 510 SYSCTL_INT(_vm, OID_AUTO, zone_kern_pages,
 511         CTLFLAG_RD, &zone_kern_pages, 0, "Number of non-interrupt safe pages allocated by zone");