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 1.10 2003/09/26 19:23:34 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 55 #ifdef INVARIANTS 56 if (z == NULL) 57 zerror(ZONE_ERROR_INVALID); 58 #endif 59 lwkt_gettoken(&z->zlock); 60 if (z->zfreecnt <= z->zfreemin) { 61 item = zget(z); 62 /* 63 * PANICFAIL allows the caller to assume that the zalloc() 64 * will always succeed. If it doesn't, we panic here. 65 */ 66 if (item == NULL && (z->zflags & ZONE_PANICFAIL)) 67 panic("zalloc(%s) failed", z->zname); 68 } else { 69 item = z->zitems; 70 z->zitems = ((void **) item)[0]; 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->zfreecnt--; 78 z->znalloc++; 79 } 80 lwkt_reltoken(&z->zlock); 81 return item; 82 } 83 84 /* 85 * Free an item to the specified zone. This function is interrupt/MP 86 * thread safe, but might block. 87 */ 88 void 89 zfree(vm_zone_t z, void *item) 90 { 91 lwkt_gettoken(&z->zlock); 92 ((void **) item)[0] = z->zitems; 93 #ifdef INVARIANTS 94 if (((void **) item)[1] == (void *) ZENTRY_FREE) 95 zerror(ZONE_ERROR_ALREADYFREE); 96 ((void **) item)[1] = (void *) ZENTRY_FREE; 97 #endif 98 z->zitems = item; 99 z->zfreecnt++; 100 lwkt_reltoken(&z->zlock); 101 } 102 103 /* 104 * This file comprises a very simple zone allocator. This is used 105 * in lieu of the malloc allocator, where needed or more optimal. 106 * 107 * Note that the initial implementation of this had coloring, and 108 * absolutely no improvement (actually perf degradation) occurred. 109 * 110 * Note also that the zones are type stable. The only restriction is 111 * that the first two longwords of a data structure can be changed 112 * between allocations. Any data that must be stable between allocations 113 * must reside in areas after the first two longwords. 114 * 115 * zinitna, zinit, zbootinit are the initialization routines. 116 * zalloc, zfree, are the allocation/free routines. 117 */ 118 119 static struct vm_zone *zlist; 120 static int sysctl_vm_zone(SYSCTL_HANDLER_ARGS); 121 static int zone_kmem_pages, zone_kern_pages, zone_kmem_kvaspace; 122 123 /* 124 * Create a zone, but don't allocate the zone structure. If the 125 * zone had been previously created by the zone boot code, initialize 126 * various parts of the zone code. 127 * 128 * If waits are not allowed during allocation (e.g. during interrupt 129 * code), a-priori allocate the kernel virtual space, and allocate 130 * only pages when needed. 131 * 132 * Arguments: 133 * z pointer to zone structure. 134 * obj pointer to VM object (opt). 135 * name name of zone. 136 * size size of zone entries. 137 * nentries number of zone entries allocated (only ZONE_INTERRUPT.) 138 * flags ZONE_INTERRUPT -- items can be allocated at interrupt time. 139 * zalloc number of pages allocated when memory is needed. 140 * 141 * Note that when using ZONE_INTERRUPT, the size of the zone is limited 142 * by the nentries argument. The size of the memory allocatable is 143 * unlimited if ZONE_INTERRUPT is not set. 144 * 145 */ 146 int 147 zinitna(vm_zone_t z, vm_object_t obj, char *name, int size, 148 int nentries, int flags, int zalloc) 149 { 150 int totsize; 151 152 if ((z->zflags & ZONE_BOOT) == 0) { 153 z->zsize = (size + ZONE_ROUNDING - 1) & ~(ZONE_ROUNDING - 1); 154 lwkt_inittoken(&z->zlock); 155 z->zfreecnt = 0; 156 z->ztotal = 0; 157 z->zmax = 0; 158 z->zname = name; 159 z->znalloc = 0; 160 z->zitems = NULL; 161 162 z->znext = zlist; 163 zlist = z; 164 } 165 166 z->zflags |= flags; 167 168 /* 169 * If we cannot wait, allocate KVA space up front, and we will fill 170 * in pages as needed. This is particularly required when creating 171 * an allocation space for map entries in kernel_map, because we 172 * do not want to go into a recursion deadlock with 173 * vm_map_entry_reserve(). 174 */ 175 if (z->zflags & ZONE_INTERRUPT) { 176 177 totsize = round_page(z->zsize * nentries); 178 zone_kmem_kvaspace += totsize; 179 180 z->zkva = kmem_alloc_pageable(kernel_map, totsize); 181 if (z->zkva == 0) { 182 zlist = z->znext; 183 return 0; 184 } 185 186 z->zpagemax = totsize / PAGE_SIZE; 187 if (obj == NULL) { 188 z->zobj = vm_object_allocate(OBJT_DEFAULT, z->zpagemax); 189 } else { 190 z->zobj = obj; 191 _vm_object_allocate(OBJT_DEFAULT, z->zpagemax, obj); 192 } 193 z->zallocflag = VM_ALLOC_INTERRUPT; 194 z->zmax += nentries; 195 } else { 196 z->zallocflag = VM_ALLOC_SYSTEM; 197 z->zmax = 0; 198 } 199 200 201 if (z->zsize > PAGE_SIZE) 202 z->zfreemin = 1; 203 else 204 z->zfreemin = PAGE_SIZE / z->zsize; 205 206 z->zpagecount = 0; 207 if (zalloc) 208 z->zalloc = zalloc; 209 else 210 z->zalloc = 1; 211 212 return 1; 213 } 214 215 /* 216 * Subroutine same as zinitna, except zone data structure is allocated 217 * automatically by malloc. This routine should normally be used, except 218 * in certain tricky startup conditions in the VM system -- then 219 * zbootinit and zinitna can be used. Zinit is the standard zone 220 * initialization call. 221 */ 222 vm_zone_t 223 zinit(char *name, int size, int nentries, int flags, int zalloc) 224 { 225 vm_zone_t z; 226 227 z = (vm_zone_t) malloc(sizeof (struct vm_zone), M_ZONE, M_NOWAIT); 228 if (z == NULL) 229 return NULL; 230 231 z->zflags = 0; 232 if (zinitna(z, NULL, name, size, nentries, flags, zalloc) == 0) { 233 free(z, M_ZONE); 234 return NULL; 235 } 236 237 return z; 238 } 239 240 /* 241 * Initialize a zone before the system is fully up. This routine should 242 * only be called before full VM startup. 243 */ 244 void 245 zbootinit(vm_zone_t z, char *name, int size, void *item, int nitems) 246 { 247 int i; 248 249 z->zname = name; 250 z->zsize = size; 251 z->zpagemax = 0; 252 z->zobj = NULL; 253 z->zflags = ZONE_BOOT; 254 z->zfreemin = 0; 255 z->zallocflag = 0; 256 z->zpagecount = 0; 257 z->zalloc = 0; 258 z->znalloc = 0; 259 lwkt_inittoken(&z->zlock); 260 261 bzero(item, nitems * z->zsize); 262 z->zitems = NULL; 263 for (i = 0; i < nitems; i++) { 264 ((void **) item)[0] = z->zitems; 265 #ifdef INVARIANTS 266 ((void **) item)[1] = (void *) ZENTRY_FREE; 267 #endif 268 z->zitems = item; 269 (char *) item += z->zsize; 270 } 271 z->zfreecnt = nitems; 272 z->zmax = nitems; 273 z->ztotal = nitems; 274 275 if (zlist == 0) { 276 zlist = z; 277 } else { 278 z->znext = zlist; 279 zlist = z; 280 } 281 } 282 283 /* 284 * void *zalloc(vm_zone_t zone) -- 285 * Returns an item from a specified zone. May not be called from a 286 * FAST interrupt or IPI function. 287 * 288 * void zfree(vm_zone_t zone, void *item) -- 289 * Frees an item back to a specified zone. May not be called from a 290 * FAST interrupt or IPI function. 291 */ 292 293 /* 294 * Internal zone routine. Not to be called from external (non vm_zone) code. 295 */ 296 static void * 297 zget(vm_zone_t z) 298 { 299 int i; 300 vm_page_t m; 301 int nitems, nbytes; 302 void *item; 303 304 if (z == NULL) 305 panic("zget: null zone"); 306 307 if (z->zflags & ZONE_INTERRUPT) { 308 nbytes = z->zpagecount * PAGE_SIZE; 309 nbytes -= nbytes % z->zsize; 310 item = (char *) z->zkva + nbytes; 311 for (i = 0; ((i < z->zalloc) && (z->zpagecount < z->zpagemax)); 312 i++) { 313 vm_offset_t zkva; 314 315 m = vm_page_alloc(z->zobj, z->zpagecount, 316 z->zallocflag); 317 if (m == NULL) 318 break; 319 lwkt_regettoken(&z->zlock); 320 321 zkva = z->zkva + z->zpagecount * PAGE_SIZE; 322 pmap_kenter(zkva, VM_PAGE_TO_PHYS(m)); /* YYY */ 323 bzero((caddr_t) zkva, PAGE_SIZE); 324 z->zpagecount++; 325 zone_kmem_pages++; 326 vmstats.v_wire_count++; 327 } 328 nitems = ((z->zpagecount * PAGE_SIZE) - nbytes) / z->zsize; 329 } else { 330 nbytes = z->zalloc * PAGE_SIZE; 331 332 #if defined(USE_KMEM_MAP) 333 /* 334 * Check to see if the kernel map is already locked. 335 * We could allow for recursive locks, but that eliminates 336 * a valuable debugging mechanism, and opens up the kernel 337 * map for potential corruption by inconsistent data structure 338 * manipulation. We could also use the interrupt allocation 339 * mechanism, but that has size limitations. Luckily, we 340 * have kmem_map that is a submap of kernel map available 341 * for memory allocation, and manipulation of that map doesn't 342 * affect the kernel map structures themselves. 343 * 344 * We can wait, so just do normal map allocation in the 345 * appropriate map. 346 */ 347 if (lockstatus(&kernel_map->lock, NULL)) { 348 int s; 349 s = splvm(); 350 item = (void *) kmem_malloc(kmem_map, nbytes, M_WAITOK); 351 lwkt_regettoken(&z->zlock); 352 if (item != NULL) 353 zone_kmem_pages += z->zalloc; 354 splx(s); 355 } else 356 #endif 357 { 358 item = (void *) kmem_alloc(kernel_map, nbytes); 359 lwkt_regettoken(&z->zlock); 360 if (item != NULL) 361 zone_kern_pages += z->zalloc; 362 } 363 if (item != NULL) { 364 bzero(item, nbytes); 365 } else { 366 nbytes = 0; 367 } 368 nitems = nbytes / z->zsize; 369 } 370 z->ztotal += nitems; 371 372 /* 373 * Save one for immediate allocation 374 */ 375 if (nitems != 0) { 376 nitems -= 1; 377 for (i = 0; i < nitems; i++) { 378 ((void **) item)[0] = z->zitems; 379 #ifdef INVARIANTS 380 ((void **) item)[1] = (void *) ZENTRY_FREE; 381 #endif 382 z->zitems = item; 383 (char *) item += z->zsize; 384 } 385 z->zfreecnt += nitems; 386 z->znalloc++; 387 } else if (z->zfreecnt > 0) { 388 item = z->zitems; 389 z->zitems = ((void **) item)[0]; 390 #ifdef INVARIANTS 391 if (((void **) item)[1] != (void *) ZENTRY_FREE) 392 zerror(ZONE_ERROR_NOTFREE); 393 ((void **) item)[1] = 0; 394 #endif 395 z->zfreecnt--; 396 z->znalloc++; 397 } else { 398 item = NULL; 399 } 400 401 /* 402 * Recover any reserve missing due to a zalloc/kreserve/krelease 403 * recursion. 404 */ 405 vm_map_entry_reserve(0); 406 407 return item; 408 } 409 410 static int 411 sysctl_vm_zone(SYSCTL_HANDLER_ARGS) 412 { 413 int error=0; 414 vm_zone_t curzone, nextzone; 415 char tmpbuf[128]; 416 char tmpname[14]; 417 418 snprintf(tmpbuf, sizeof(tmpbuf), 419 "\nITEM SIZE LIMIT USED FREE REQUESTS\n"); 420 error = SYSCTL_OUT(req, tmpbuf, strlen(tmpbuf)); 421 if (error) 422 return (error); 423 424 for (curzone = zlist; curzone; curzone = nextzone) { 425 int i; 426 int len; 427 int offset; 428 429 nextzone = curzone->znext; 430 len = strlen(curzone->zname); 431 if (len >= (sizeof(tmpname) - 1)) 432 len = (sizeof(tmpname) - 1); 433 for(i = 0; i < sizeof(tmpname) - 1; i++) 434 tmpname[i] = ' '; 435 tmpname[i] = 0; 436 memcpy(tmpname, curzone->zname, len); 437 tmpname[len] = ':'; 438 offset = 0; 439 if (curzone == zlist) { 440 offset = 1; 441 tmpbuf[0] = '\n'; 442 } 443 444 snprintf(tmpbuf + offset, sizeof(tmpbuf) - offset, 445 "%s %6.6u, %8.8u, %6.6u, %6.6u, %8.8u\n", 446 tmpname, curzone->zsize, curzone->zmax, 447 (curzone->ztotal - curzone->zfreecnt), 448 curzone->zfreecnt, curzone->znalloc); 449 450 len = strlen((char *)tmpbuf); 451 if (nextzone == NULL) 452 tmpbuf[len - 1] = 0; 453 454 error = SYSCTL_OUT(req, tmpbuf, len); 455 456 if (error) 457 return (error); 458 } 459 return (0); 460 } 461 462 #if defined(INVARIANTS) 463 void 464 zerror(int error) 465 { 466 char *msg; 467 468 switch (error) { 469 case ZONE_ERROR_INVALID: 470 msg = "zone: invalid zone"; 471 break; 472 case ZONE_ERROR_NOTFREE: 473 msg = "zone: entry not free"; 474 break; 475 case ZONE_ERROR_ALREADYFREE: 476 msg = "zone: freeing free entry"; 477 break; 478 default: 479 msg = "zone: invalid error"; 480 break; 481 } 482 panic(msg); 483 } 484 #endif 485 486 SYSCTL_OID(_vm, OID_AUTO, zone, CTLTYPE_STRING|CTLFLAG_RD, \ 487 NULL, 0, sysctl_vm_zone, "A", "Zone Info"); 488 489 SYSCTL_INT(_vm, OID_AUTO, zone_kmem_pages, 490 CTLFLAG_RD, &zone_kmem_pages, 0, "Number of interrupt safe pages allocated by zone"); 491 SYSCTL_INT(_vm, OID_AUTO, zone_kmem_kvaspace, 492 CTLFLAG_RD, &zone_kmem_kvaspace, 0, "KVA space allocated by zone"); 493 SYSCTL_INT(_vm, OID_AUTO, zone_kern_pages, 494 CTLFLAG_RD, &zone_kern_pages, 0, "Number of non-interrupt safe pages allocated by zone"); 495