| 1 | /* |
| 2 | * Copyright (c) 1997, 1998 John S. Dyson. All rights reserved. |
| 3 | * |
| 4 | * Redistribution and use in source and binary forms, with or without |
| 5 | * modification, are permitted provided that the following conditions |
| 6 | * are met: |
| 7 | * 1. Redistributions of source code must retain the above copyright |
| 8 | * notice immediately at the beginning of the file, without modification, |
| 9 | * this list of conditions, and the following disclaimer. |
| 10 | * 2. Absolutely no warranty of function or purpose is made by the author |
| 11 | * John S. Dyson. |
| 12 | * |
| 13 | * $FreeBSD: src/sys/vm/vm_zone.c,v 1.30.2.6 2002/10/10 19:50:16 dillon Exp $ |
| 14 | * |
| 15 | * Copyright (c) 2003-2017,2019 The DragonFly Project. All rights reserved. |
| 16 | * |
| 17 | * This code is derived from software contributed to The DragonFly Project |
| 18 | * by Matthew Dillon <dillon@backplane.com> |
| 19 | * |
| 20 | * Redistribution and use in source and binary forms, with or without |
| 21 | * modification, are permitted provided that the following conditions |
| 22 | * are met: |
| 23 | * |
| 24 | * 1. Redistributions of source code must retain the above copyright |
| 25 | * notice, this list of conditions and the following disclaimer. |
| 26 | * 2. Redistributions in binary form must reproduce the above copyright |
| 27 | * notice, this list of conditions and the following disclaimer in |
| 28 | * the documentation and/or other materials provided with the |
| 29 | * distribution. |
| 30 | * 3. Neither the name of The DragonFly Project nor the names of its |
| 31 | * contributors may be used to endorse or promote products derived |
| 32 | * from this software without specific, prior written permission. |
| 33 | * |
| 34 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS |
| 35 | * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT |
| 36 | * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS |
| 37 | * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE |
| 38 | * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, |
| 39 | * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, |
| 40 | * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; |
| 41 | * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED |
| 42 | * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, |
| 43 | * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT |
| 44 | * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF |
| 45 | * SUCH DAMAGE. |
| 46 | */ |
| 47 | |
| 48 | #include <sys/param.h> |
| 49 | #include <sys/queue.h> |
| 50 | #include <sys/systm.h> |
| 51 | #include <sys/kernel.h> |
| 52 | #include <sys/lock.h> |
| 53 | #include <sys/malloc.h> |
| 54 | #include <sys/sysctl.h> |
| 55 | #include <sys/vmmeter.h> |
| 56 | |
| 57 | #include <vm/vm.h> |
| 58 | #include <vm/vm_object.h> |
| 59 | #include <vm/vm_page.h> |
| 60 | #include <vm/vm_map.h> |
| 61 | #include <vm/vm_kern.h> |
| 62 | #include <vm/vm_extern.h> |
| 63 | #include <vm/vm_zone.h> |
| 64 | |
| 65 | #include <sys/spinlock2.h> |
| 66 | #include <vm/vm_page2.h> |
| 67 | |
| 68 | static MALLOC_DEFINE(M_ZONE, "ZONE", "Zone header"); |
| 69 | |
| 70 | #define ZONE_ERROR_INVALID 0 |
| 71 | #define ZONE_ERROR_NOTFREE 1 |
| 72 | #define ZONE_ERROR_ALREADYFREE 2 |
| 73 | |
| 74 | #define ZONE_ROUNDING 32 |
| 75 | |
| 76 | #define ZENTRY_FREE 0x12342378 |
| 77 | |
| 78 | long zone_burst = 128; |
| 79 | |
| 80 | static void *zget(vm_zone_t z, int *tryagainp); |
| 81 | |
| 82 | /* |
| 83 | * Return an item from the specified zone. This function is non-blocking for |
| 84 | * ZONE_INTERRUPT zones. |
| 85 | * |
| 86 | * No requirements. |
| 87 | */ |
| 88 | void * |
| 89 | zalloc(vm_zone_t z) |
| 90 | { |
| 91 | globaldata_t gd = mycpu; |
| 92 | vm_zpcpu_t *zpcpu; |
| 93 | void *item; |
| 94 | int tryagain; |
| 95 | long n; |
| 96 | |
| 97 | #ifdef INVARIANTS |
| 98 | if (z == NULL) |
| 99 | zerror(ZONE_ERROR_INVALID); |
| 100 | #endif |
| 101 | zpcpu = &z->zpcpu[gd->gd_cpuid]; |
| 102 | retry: |
| 103 | /* |
| 104 | * Avoid spinlock contention by allocating from a per-cpu queue |
| 105 | */ |
| 106 | if (zpcpu->zfreecnt > 0) { |
| 107 | crit_enter_gd(gd); |
| 108 | if (zpcpu->zfreecnt > 0) { |
| 109 | item = zpcpu->zitems; |
| 110 | #ifdef INVARIANTS |
| 111 | KASSERT(item != NULL, |
| 112 | ("zitems_pcpu unexpectedly NULL")); |
| 113 | if (((void **)item)[1] != (void *)ZENTRY_FREE) |
| 114 | zerror(ZONE_ERROR_NOTFREE); |
| 115 | ((void **)item)[1] = NULL; |
| 116 | #endif |
| 117 | zpcpu->zitems = ((void **) item)[0]; |
| 118 | --zpcpu->zfreecnt; |
| 119 | ++zpcpu->znalloc; |
| 120 | crit_exit_gd(gd); |
| 121 | |
| 122 | return item; |
| 123 | } |
| 124 | crit_exit_gd(gd); |
| 125 | } |
| 126 | |
| 127 | /* |
| 128 | * Per-zone spinlock for the remainder. Always load at least one |
| 129 | * item. |
| 130 | */ |
| 131 | spin_lock(&z->zspin); |
| 132 | if (z->zfreecnt > z->zfreemin) { |
| 133 | n = zone_burst; |
| 134 | do { |
| 135 | item = z->zitems; |
| 136 | #ifdef INVARIANTS |
| 137 | KASSERT(item != NULL, ("zitems unexpectedly NULL")); |
| 138 | if (((void **)item)[1] != (void *)ZENTRY_FREE) |
| 139 | zerror(ZONE_ERROR_NOTFREE); |
| 140 | #endif |
| 141 | z->zitems = ((void **)item)[0]; |
| 142 | --z->zfreecnt; |
| 143 | ((void **)item)[0] = zpcpu->zitems; |
| 144 | zpcpu->zitems = item; |
| 145 | ++zpcpu->zfreecnt; |
| 146 | } while (--n > 0 && z->zfreecnt > z->zfreemin); |
| 147 | spin_unlock(&z->zspin); |
| 148 | goto retry; |
| 149 | } else { |
| 150 | spin_unlock(&z->zspin); |
| 151 | tryagain = 0; |
| 152 | item = zget(z, &tryagain); |
| 153 | if (tryagain) |
| 154 | goto retry; |
| 155 | |
| 156 | /* |
| 157 | * PANICFAIL allows the caller to assume that the zalloc() |
| 158 | * will always succeed. If it doesn't, we panic here. |
| 159 | */ |
| 160 | if (item == NULL && (z->zflags & ZONE_PANICFAIL)) |
| 161 | panic("zalloc(%s) failed", z->zname); |
| 162 | } |
| 163 | return item; |
| 164 | } |
| 165 | |
| 166 | /* |
| 167 | * Free an item to the specified zone. |
| 168 | * |
| 169 | * No requirements. |
| 170 | */ |
| 171 | void |
| 172 | zfree(vm_zone_t z, void *item) |
| 173 | { |
| 174 | globaldata_t gd = mycpu; |
| 175 | vm_zpcpu_t *zpcpu; |
| 176 | void *tail_item; |
| 177 | long count; |
| 178 | long zmax; |
| 179 | |
| 180 | zpcpu = &z->zpcpu[gd->gd_cpuid]; |
| 181 | |
| 182 | /* |
| 183 | * Avoid spinlock contention by freeing into a per-cpu queue |
| 184 | */ |
| 185 | zmax = z->zmax_pcpu; |
| 186 | if (zmax < 1024) |
| 187 | zmax = 1024; |
| 188 | |
| 189 | /* |
| 190 | * Add to pcpu cache |
| 191 | */ |
| 192 | crit_enter_gd(gd); |
| 193 | ((void **)item)[0] = zpcpu->zitems; |
| 194 | #ifdef INVARIANTS |
| 195 | if (((void **)item)[1] == (void *)ZENTRY_FREE) |
| 196 | zerror(ZONE_ERROR_ALREADYFREE); |
| 197 | ((void **)item)[1] = (void *)ZENTRY_FREE; |
| 198 | #endif |
| 199 | zpcpu->zitems = item; |
| 200 | ++zpcpu->zfreecnt; |
| 201 | |
| 202 | if (zpcpu->zfreecnt < zmax) { |
| 203 | crit_exit_gd(gd); |
| 204 | return; |
| 205 | } |
| 206 | |
| 207 | /* |
| 208 | * Hystereis, move (zmax) (calculated below) items to the pool. |
| 209 | */ |
| 210 | zmax = zmax / 2; |
| 211 | if (zmax > zone_burst) |
| 212 | zmax = zone_burst; |
| 213 | tail_item = item; |
| 214 | count = 1; |
| 215 | |
| 216 | while (count < zmax) { |
| 217 | tail_item = ((void **)tail_item)[0]; |
| 218 | ++count; |
| 219 | } |
| 220 | zpcpu->zitems = ((void **)tail_item)[0]; |
| 221 | zpcpu->zfreecnt -= count; |
| 222 | |
| 223 | /* |
| 224 | * Per-zone spinlock for the remainder. |
| 225 | * |
| 226 | * Also implement hysteresis by freeing a number of pcpu |
| 227 | * entries. |
| 228 | */ |
| 229 | spin_lock(&z->zspin); |
| 230 | ((void **)tail_item)[0] = z->zitems; |
| 231 | z->zitems = item; |
| 232 | z->zfreecnt += count; |
| 233 | spin_unlock(&z->zspin); |
| 234 | |
| 235 | crit_exit_gd(gd); |
| 236 | } |
| 237 | |
| 238 | /* |
| 239 | * This file comprises a very simple zone allocator. This is used |
| 240 | * in lieu of the malloc allocator, where needed or more optimal. |
| 241 | * |
| 242 | * Note that the initial implementation of this had coloring, and |
| 243 | * absolutely no improvement (actually perf degradation) occurred. |
| 244 | * |
| 245 | * Note also that the zones are type stable. The only restriction is |
| 246 | * that the first two longwords of a data structure can be changed |
| 247 | * between allocations. Any data that must be stable between allocations |
| 248 | * must reside in areas after the first two longwords. |
| 249 | * |
| 250 | * zinitna, zinit, zbootinit are the initialization routines. |
| 251 | * zalloc, zfree, are the allocation/free routines. |
| 252 | */ |
| 253 | |
| 254 | LIST_HEAD(zlist, vm_zone) zlist = LIST_HEAD_INITIALIZER(zlist); |
| 255 | static int sysctl_vm_zone(SYSCTL_HANDLER_ARGS); |
| 256 | static vm_pindex_t zone_kmem_pages, zone_kern_pages; |
| 257 | static long zone_kmem_kvaspace; |
| 258 | |
| 259 | /* |
| 260 | * Create a zone, but don't allocate the zone structure. If the |
| 261 | * zone had been previously created by the zone boot code, initialize |
| 262 | * various parts of the zone code. |
| 263 | * |
| 264 | * If waits are not allowed during allocation (e.g. during interrupt |
| 265 | * code), a-priori allocate the kernel virtual space, and allocate |
| 266 | * only pages when needed. |
| 267 | * |
| 268 | * Arguments: |
| 269 | * z pointer to zone structure. |
| 270 | * obj pointer to VM object (opt). |
| 271 | * name name of zone. |
| 272 | * size size of zone entries. |
| 273 | * nentries number of zone entries allocated (only ZONE_INTERRUPT.) |
| 274 | * flags ZONE_INTERRUPT -- items can be allocated at interrupt time. |
| 275 | * zalloc number of pages allocated when memory is needed. |
| 276 | * |
| 277 | * Note that when using ZONE_INTERRUPT, the size of the zone is limited |
| 278 | * by the nentries argument. The size of the memory allocatable is |
| 279 | * unlimited if ZONE_INTERRUPT is not set. |
| 280 | * |
| 281 | * No requirements. |
| 282 | */ |
| 283 | int |
| 284 | zinitna(vm_zone_t z, char *name, size_t size, long nentries, uint32_t flags) |
| 285 | { |
| 286 | size_t totsize; |
| 287 | |
| 288 | /* |
| 289 | * Only zones created with zinit() are destroyable. |
| 290 | */ |
| 291 | if (z->zflags & ZONE_DESTROYABLE) |
| 292 | panic("zinitna: can't create destroyable zone"); |
| 293 | |
| 294 | /* |
| 295 | * NOTE: We can only adjust zsize if we previously did not |
| 296 | * use zbootinit(). |
| 297 | */ |
| 298 | if ((z->zflags & ZONE_BOOT) == 0) { |
| 299 | z->zsize = roundup2(size, ZONE_ROUNDING); |
| 300 | spin_init(&z->zspin, "zinitna"); |
| 301 | lockinit(&z->zgetlk, "zgetlk", 0, LK_CANRECURSE); |
| 302 | |
| 303 | z->zfreecnt = 0; |
| 304 | z->ztotal = 0; |
| 305 | z->zmax = 0; |
| 306 | z->zname = name; |
| 307 | z->zitems = NULL; |
| 308 | |
| 309 | lwkt_gettoken(&vm_token); |
| 310 | LIST_INSERT_HEAD(&zlist, z, zlink); |
| 311 | lwkt_reltoken(&vm_token); |
| 312 | |
| 313 | bzero(z->zpcpu, sizeof(z->zpcpu)); |
| 314 | } |
| 315 | |
| 316 | z->zkmvec = NULL; |
| 317 | z->zkmcur = z->zkmmax = 0; |
| 318 | z->zflags |= flags; |
| 319 | |
| 320 | /* |
| 321 | * If we cannot wait, allocate KVA space up front, and we will fill |
| 322 | * in pages as needed. This is particularly required when creating |
| 323 | * an allocation space for map entries in kernel_map, because we |
| 324 | * do not want to go into a recursion deadlock with |
| 325 | * vm_map_entry_reserve(). |
| 326 | */ |
| 327 | if (z->zflags & ZONE_INTERRUPT) { |
| 328 | totsize = round_page((size_t)z->zsize * nentries); |
| 329 | atomic_add_long(&zone_kmem_kvaspace, totsize); |
| 330 | |
| 331 | z->zkva = kmem_alloc_pageable(kernel_map, totsize, |
| 332 | VM_SUBSYS_ZALLOC); |
| 333 | if (z->zkva == 0) { |
| 334 | LIST_REMOVE(z, zlink); |
| 335 | return 0; |
| 336 | } |
| 337 | |
| 338 | z->zpagemax = totsize / PAGE_SIZE; |
| 339 | z->zallocflag = VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT | |
| 340 | VM_ALLOC_NORMAL | VM_ALLOC_RETRY; |
| 341 | z->zmax += nentries; |
| 342 | |
| 343 | /* |
| 344 | * Set reasonable pcpu cache bounds. Low-memory systems |
| 345 | * might try to cache too little, large-memory systems |
| 346 | * might try to cache more than necessarsy. |
| 347 | * |
| 348 | * In particular, pvzone can wind up being excessive and |
| 349 | * waste memory unnecessarily. |
| 350 | */ |
| 351 | z->zmax_pcpu = z->zmax / ncpus / 64; |
| 352 | if (z->zmax_pcpu < 1024) |
| 353 | z->zmax_pcpu = 1024; |
| 354 | if (z->zmax_pcpu * z->zsize > 16*1024*1024) |
| 355 | z->zmax_pcpu = 16*1024*1024 / z->zsize; |
| 356 | } else { |
| 357 | z->zallocflag = VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM; |
| 358 | z->zmax = 0; |
| 359 | z->zmax_pcpu = 8192; |
| 360 | } |
| 361 | |
| 362 | |
| 363 | if (z->zsize > PAGE_SIZE) |
| 364 | z->zfreemin = 1; |
| 365 | else |
| 366 | z->zfreemin = PAGE_SIZE / z->zsize; |
| 367 | |
| 368 | z->zpagecount = 0; |
| 369 | |
| 370 | /* |
| 371 | * Reduce kernel_map spam by allocating in chunks. |
| 372 | */ |
| 373 | z->zalloc = ZONE_MAXPGLOAD; |
| 374 | |
| 375 | /* |
| 376 | * Populate the interrrupt zone at creation time rather than |
| 377 | * on first allocation, as this is a potentially long operation. |
| 378 | */ |
| 379 | if (z->zflags & ZONE_INTERRUPT) { |
| 380 | void *buf; |
| 381 | |
| 382 | buf = zget(z, NULL); |
| 383 | if (buf) |
| 384 | zfree(z, buf); |
| 385 | } |
| 386 | |
| 387 | return 1; |
| 388 | } |
| 389 | |
| 390 | /* |
| 391 | * Subroutine same as zinitna, except zone data structure is allocated |
| 392 | * automatically by malloc. This routine should normally be used, except |
| 393 | * in certain tricky startup conditions in the VM system -- then |
| 394 | * zbootinit and zinitna can be used. Zinit is the standard zone |
| 395 | * initialization call. |
| 396 | * |
| 397 | * No requirements. |
| 398 | */ |
| 399 | vm_zone_t |
| 400 | zinit(char *name, size_t size, long nentries, uint32_t flags) |
| 401 | { |
| 402 | vm_zone_t z; |
| 403 | |
| 404 | z = (vm_zone_t) kmalloc(sizeof (struct vm_zone), M_ZONE, M_NOWAIT); |
| 405 | if (z == NULL) |
| 406 | return NULL; |
| 407 | |
| 408 | z->zflags = 0; |
| 409 | if (zinitna(z, name, size, nentries, flags & ~ZONE_DESTROYABLE) == 0) { |
| 410 | kfree(z, M_ZONE); |
| 411 | return NULL; |
| 412 | } |
| 413 | |
| 414 | if (flags & ZONE_DESTROYABLE) |
| 415 | z->zflags |= ZONE_DESTROYABLE; |
| 416 | |
| 417 | return z; |
| 418 | } |
| 419 | |
| 420 | /* |
| 421 | * Initialize a zone before the system is fully up. This routine should |
| 422 | * only be called before full VM startup. |
| 423 | * |
| 424 | * Called from the low level boot code only. |
| 425 | */ |
| 426 | void |
| 427 | zbootinit(vm_zone_t z, char *name, size_t size, void *item, long nitems) |
| 428 | { |
| 429 | long i; |
| 430 | |
| 431 | spin_init(&z->zspin, "zbootinit"); |
| 432 | lockinit(&z->zgetlk, "zgetlk", 0, LK_CANRECURSE); |
| 433 | bzero(z->zpcpu, sizeof(z->zpcpu)); |
| 434 | z->zname = name; |
| 435 | z->zsize = size; |
| 436 | z->zpagemax = 0; |
| 437 | z->zflags = ZONE_BOOT; |
| 438 | z->zfreemin = 0; |
| 439 | z->zallocflag = 0; |
| 440 | z->zpagecount = 0; |
| 441 | z->zalloc = 0; |
| 442 | |
| 443 | bzero(item, (size_t)nitems * z->zsize); |
| 444 | z->zitems = NULL; |
| 445 | for (i = 0; i < nitems; i++) { |
| 446 | ((void **)item)[0] = z->zitems; |
| 447 | #ifdef INVARIANTS |
| 448 | ((void **)item)[1] = (void *)ZENTRY_FREE; |
| 449 | #endif |
| 450 | z->zitems = item; |
| 451 | item = (uint8_t *)item + z->zsize; |
| 452 | } |
| 453 | z->zfreecnt = nitems; |
| 454 | z->zmax = nitems; |
| 455 | z->ztotal = nitems; |
| 456 | |
| 457 | lwkt_gettoken(&vm_token); |
| 458 | LIST_INSERT_HEAD(&zlist, z, zlink); |
| 459 | lwkt_reltoken(&vm_token); |
| 460 | } |
| 461 | |
| 462 | /* |
| 463 | * Release all resources owned by zone created with zinit(). |
| 464 | * |
| 465 | * No requirements. |
| 466 | */ |
| 467 | void |
| 468 | zdestroy(vm_zone_t z) |
| 469 | { |
| 470 | vm_pindex_t i; |
| 471 | |
| 472 | if (z == NULL) |
| 473 | panic("zdestroy: null zone"); |
| 474 | if ((z->zflags & ZONE_DESTROYABLE) == 0) |
| 475 | panic("zdestroy: undestroyable zone"); |
| 476 | |
| 477 | lwkt_gettoken(&vm_token); |
| 478 | LIST_REMOVE(z, zlink); |
| 479 | lwkt_reltoken(&vm_token); |
| 480 | |
| 481 | /* |
| 482 | * Release virtual mappings, physical memory and update sysctl stats. |
| 483 | */ |
| 484 | KKASSERT((z->zflags & ZONE_INTERRUPT) == 0); |
| 485 | for (i = 0; i < z->zkmcur; i++) { |
| 486 | kmem_free(kernel_map, z->zkmvec[i], |
| 487 | (size_t)z->zalloc * PAGE_SIZE); |
| 488 | atomic_subtract_long(&zone_kern_pages, z->zalloc); |
| 489 | } |
| 490 | if (z->zkmvec != NULL) |
| 491 | kfree(z->zkmvec, M_ZONE); |
| 492 | |
| 493 | spin_uninit(&z->zspin); |
| 494 | kfree(z, M_ZONE); |
| 495 | } |
| 496 | |
| 497 | |
| 498 | /* |
| 499 | * void *zalloc(vm_zone_t zone) -- |
| 500 | * Returns an item from a specified zone. May not be called from a |
| 501 | * FAST interrupt or IPI function. |
| 502 | * |
| 503 | * void zfree(vm_zone_t zone, void *item) -- |
| 504 | * Frees an item back to a specified zone. May not be called from a |
| 505 | * FAST interrupt or IPI function. |
| 506 | */ |
| 507 | |
| 508 | /* |
| 509 | * Internal zone routine. Not to be called from external (non vm_zone) code. |
| 510 | * |
| 511 | * This function may return NULL. |
| 512 | * |
| 513 | * No requirements. |
| 514 | */ |
| 515 | static void * |
| 516 | zget(vm_zone_t z, int *tryagainp) |
| 517 | { |
| 518 | vm_page_t pgs[ZONE_MAXPGLOAD]; |
| 519 | vm_page_t m; |
| 520 | long nitems; |
| 521 | long savezpc; |
| 522 | size_t nbytes; |
| 523 | size_t noffset; |
| 524 | void *item; |
| 525 | vm_pindex_t npages; |
| 526 | vm_pindex_t nalloc; |
| 527 | vm_pindex_t i; |
| 528 | |
| 529 | if (z == NULL) |
| 530 | panic("zget: null zone"); |
| 531 | |
| 532 | /* |
| 533 | * We need an encompassing per-zone lock for zget() refills. |
| 534 | * |
| 535 | * Without this we wind up locking on the vm_map inside kmem_alloc*() |
| 536 | * prior to any entries actually being added to the zone, potentially |
| 537 | * exhausting the per-cpu cache of vm_map_entry's when multiple threads |
| 538 | * are blocked on the same lock on the same cpu. |
| 539 | */ |
| 540 | if ((z->zflags & ZONE_INTERRUPT) == 0) { |
| 541 | if (lockmgr(&z->zgetlk, LK_EXCLUSIVE | LK_SLEEPFAIL)) { |
| 542 | *tryagainp = 1; |
| 543 | return NULL; |
| 544 | } |
| 545 | } |
| 546 | |
| 547 | if (z->zflags & ZONE_INTERRUPT) { |
| 548 | /* |
| 549 | * Interrupt zones do not mess with the kernel_map, they |
| 550 | * simply populate an existing mapping. |
| 551 | * |
| 552 | * First allocate as many pages as we can, stopping at |
| 553 | * our limit or if the page allocation fails. Try to |
| 554 | * avoid exhausting the interrupt free minimum by backing |
| 555 | * off to normal page allocations after a certain point. |
| 556 | */ |
| 557 | for (i = 0; i < ZONE_MAXPGLOAD && i < z->zalloc; ++i) { |
| 558 | if (i < 4) { |
| 559 | m = vm_page_alloc(NULL, |
| 560 | mycpu->gd_rand_incr++, |
| 561 | z->zallocflag); |
| 562 | } else { |
| 563 | m = vm_page_alloc(NULL, |
| 564 | mycpu->gd_rand_incr++, |
| 565 | VM_ALLOC_NORMAL | |
| 566 | VM_ALLOC_SYSTEM); |
| 567 | } |
| 568 | if (m == NULL) |
| 569 | break; |
| 570 | pgs[i] = m; |
| 571 | } |
| 572 | nalloc = i; |
| 573 | |
| 574 | /* |
| 575 | * Account for the pages. |
| 576 | * |
| 577 | * NOTE! Do not allow overlap with a prior page as it |
| 578 | * may still be undergoing allocation on another |
| 579 | * cpu. |
| 580 | */ |
| 581 | spin_lock(&z->zspin); |
| 582 | noffset = (size_t)z->zpagecount * PAGE_SIZE; |
| 583 | /* noffset -= noffset % z->zsize; */ |
| 584 | savezpc = z->zpagecount; |
| 585 | |
| 586 | /* |
| 587 | * Track total memory use and kmem offset. |
| 588 | */ |
| 589 | if (z->zpagecount + nalloc > z->zpagemax) |
| 590 | z->zpagecount = z->zpagemax; |
| 591 | else |
| 592 | z->zpagecount += nalloc; |
| 593 | |
| 594 | item = (char *)z->zkva + noffset; |
| 595 | npages = z->zpagecount - savezpc; |
| 596 | nitems = ((size_t)(savezpc + npages) * PAGE_SIZE - noffset) / |
| 597 | z->zsize; |
| 598 | atomic_add_long(&zone_kmem_pages, npages); |
| 599 | spin_unlock(&z->zspin); |
| 600 | |
| 601 | /* |
| 602 | * Enter the pages into the reserved KVA space. |
| 603 | */ |
| 604 | for (i = 0; i < npages; ++i) { |
| 605 | vm_offset_t zkva; |
| 606 | |
| 607 | m = pgs[i]; |
| 608 | KKASSERT(m->queue == PQ_NONE); |
| 609 | m->valid = VM_PAGE_BITS_ALL; |
| 610 | vm_page_wire(m); |
| 611 | vm_page_wakeup(m); |
| 612 | |
| 613 | zkva = z->zkva + (size_t)(savezpc + i) * PAGE_SIZE; |
| 614 | pmap_kenter(zkva, VM_PAGE_TO_PHYS(m)); |
| 615 | bzero((void *)zkva, PAGE_SIZE); |
| 616 | } |
| 617 | for (i = npages; i < nalloc; ++i) { |
| 618 | m = pgs[i]; |
| 619 | vm_page_free(m); |
| 620 | } |
| 621 | } else if (z->zflags & ZONE_SPECIAL) { |
| 622 | /* |
| 623 | * The special zone is the one used for vm_map_entry_t's. |
| 624 | * We have to avoid an infinite recursion in |
| 625 | * vm_map_entry_reserve() by using vm_map_entry_kreserve() |
| 626 | * instead. The map entries are pre-reserved by the kernel |
| 627 | * by vm_map_entry_reserve_cpu_init(). |
| 628 | */ |
| 629 | nbytes = (size_t)z->zalloc * PAGE_SIZE; |
| 630 | z->zpagecount += z->zalloc; /* Track total memory use */ |
| 631 | |
| 632 | item = (void *)kmem_alloc3(kernel_map, nbytes, |
| 633 | VM_SUBSYS_ZALLOC, KM_KRESERVE); |
| 634 | |
| 635 | /* note: z might be modified due to blocking */ |
| 636 | if (item != NULL) { |
| 637 | atomic_add_long(&zone_kern_pages, z->zalloc); |
| 638 | bzero(item, nbytes); |
| 639 | } else { |
| 640 | nbytes = 0; |
| 641 | } |
| 642 | nitems = nbytes / z->zsize; |
| 643 | } else { |
| 644 | /* |
| 645 | * Otherwise allocate KVA from the kernel_map. |
| 646 | */ |
| 647 | nbytes = (size_t)z->zalloc * PAGE_SIZE; |
| 648 | z->zpagecount += z->zalloc; /* Track total memory use */ |
| 649 | |
| 650 | item = (void *)kmem_alloc3(kernel_map, nbytes, |
| 651 | VM_SUBSYS_ZALLOC, 0); |
| 652 | |
| 653 | /* note: z might be modified due to blocking */ |
| 654 | if (item != NULL) { |
| 655 | atomic_add_long(&zone_kern_pages, z->zalloc); |
| 656 | bzero(item, nbytes); |
| 657 | |
| 658 | if (z->zflags & ZONE_DESTROYABLE) { |
| 659 | if (z->zkmcur == z->zkmmax) { |
| 660 | z->zkmmax = |
| 661 | z->zkmmax==0 ? 1 : z->zkmmax*2; |
| 662 | z->zkmvec = krealloc(z->zkmvec, |
| 663 | z->zkmmax * sizeof(z->zkmvec[0]), |
| 664 | M_ZONE, M_WAITOK); |
| 665 | } |
| 666 | z->zkmvec[z->zkmcur++] = (vm_offset_t)item; |
| 667 | } |
| 668 | } else { |
| 669 | nbytes = 0; |
| 670 | } |
| 671 | nitems = nbytes / z->zsize; |
| 672 | } |
| 673 | |
| 674 | /* |
| 675 | * Enter any new pages into the pool, reserving one, or get the |
| 676 | * item from the existing pool. |
| 677 | */ |
| 678 | spin_lock(&z->zspin); |
| 679 | z->ztotal += nitems; |
| 680 | |
| 681 | /* |
| 682 | * The zone code may need to allocate kernel memory, which can |
| 683 | * recurse zget() infinitely if we do not handle it properly. |
| 684 | * We deal with this by directly repopulating the pcpu vm_map_entry |
| 685 | * cache. |
| 686 | */ |
| 687 | if (nitems > 1 && (z->zflags & ZONE_SPECIAL)) { |
| 688 | struct globaldata *gd = mycpu; |
| 689 | vm_map_entry_t entry; |
| 690 | |
| 691 | /* |
| 692 | * Make sure we have enough structures in gd_vme_base to handle |
| 693 | * the reservation request. |
| 694 | * |
| 695 | * The critical section protects access to the per-cpu gd. |
| 696 | */ |
| 697 | crit_enter(); |
| 698 | while (gd->gd_vme_avail < 2 && nitems > 1) { |
| 699 | entry = item; |
| 700 | MAPENT_FREELIST(entry) = gd->gd_vme_base; |
| 701 | gd->gd_vme_base = entry; |
| 702 | atomic_add_int(&gd->gd_vme_avail, 1); |
| 703 | item = (uint8_t *)item + z->zsize; |
| 704 | --nitems; |
| 705 | } |
| 706 | crit_exit(); |
| 707 | } |
| 708 | |
| 709 | if (nitems != 0) { |
| 710 | /* |
| 711 | * Enter pages into the pool saving one for immediate |
| 712 | * allocation. |
| 713 | */ |
| 714 | nitems -= 1; |
| 715 | for (i = 0; i < nitems; i++) { |
| 716 | ((void **)item)[0] = z->zitems; |
| 717 | #ifdef INVARIANTS |
| 718 | ((void **)item)[1] = (void *)ZENTRY_FREE; |
| 719 | #endif |
| 720 | z->zitems = item; |
| 721 | item = (uint8_t *)item + z->zsize; |
| 722 | } |
| 723 | z->zfreecnt += nitems; |
| 724 | ++z->znalloc; |
| 725 | } else if (z->zfreecnt > 0) { |
| 726 | /* |
| 727 | * Get an item from the existing pool. |
| 728 | */ |
| 729 | item = z->zitems; |
| 730 | z->zitems = ((void **)item)[0]; |
| 731 | #ifdef INVARIANTS |
| 732 | if (((void **)item)[1] != (void *)ZENTRY_FREE) |
| 733 | zerror(ZONE_ERROR_NOTFREE); |
| 734 | ((void **) item)[1] = NULL; |
| 735 | #endif |
| 736 | --z->zfreecnt; |
| 737 | ++z->znalloc; |
| 738 | } else { |
| 739 | /* |
| 740 | * No items available. |
| 741 | */ |
| 742 | item = NULL; |
| 743 | } |
| 744 | spin_unlock(&z->zspin); |
| 745 | |
| 746 | /* |
| 747 | * Release the per-zone global lock after the items have been |
| 748 | * added. Any other threads blocked in zget()'s zgetlk will |
| 749 | * then retry rather than potentially exhaust the per-cpu cache |
| 750 | * of vm_map_entry structures doing their own kmem_alloc() calls, |
| 751 | * or allocating excessive amounts of space unnecessarily. |
| 752 | */ |
| 753 | if ((z->zflags & ZONE_INTERRUPT) == 0) |
| 754 | lockmgr(&z->zgetlk, LK_RELEASE); |
| 755 | |
| 756 | return item; |
| 757 | } |
| 758 | |
| 759 | /* |
| 760 | * No requirements. |
| 761 | */ |
| 762 | static int |
| 763 | sysctl_vm_zone(SYSCTL_HANDLER_ARGS) |
| 764 | { |
| 765 | vm_zone_t curzone; |
| 766 | char tmpbuf[128]; |
| 767 | char tmpname[14]; |
| 768 | int error = 0; |
| 769 | |
| 770 | ksnprintf(tmpbuf, sizeof(tmpbuf), |
| 771 | "\nITEM SIZE LIMIT USED FREE REQUESTS\n"); |
| 772 | error = SYSCTL_OUT(req, tmpbuf, strlen(tmpbuf)); |
| 773 | if (error) |
| 774 | return (error); |
| 775 | |
| 776 | lwkt_gettoken(&vm_token); |
| 777 | LIST_FOREACH(curzone, &zlist, zlink) { |
| 778 | size_t i; |
| 779 | size_t len; |
| 780 | int offset; |
| 781 | long freecnt; |
| 782 | long znalloc; |
| 783 | int n; |
| 784 | |
| 785 | len = strlen(curzone->zname); |
| 786 | if (len >= (sizeof(tmpname) - 1)) |
| 787 | len = (sizeof(tmpname) - 1); |
| 788 | for(i = 0; i < sizeof(tmpname) - 1; i++) |
| 789 | tmpname[i] = ' '; |
| 790 | tmpname[i] = 0; |
| 791 | memcpy(tmpname, curzone->zname, len); |
| 792 | tmpname[len] = ':'; |
| 793 | offset = 0; |
| 794 | if (curzone == LIST_FIRST(&zlist)) { |
| 795 | offset = 1; |
| 796 | tmpbuf[0] = '\n'; |
| 797 | } |
| 798 | freecnt = curzone->zfreecnt; |
| 799 | znalloc = curzone->znalloc; |
| 800 | for (n = 0; n < ncpus; ++n) { |
| 801 | freecnt += curzone->zpcpu[n].zfreecnt; |
| 802 | znalloc += curzone->zpcpu[n].znalloc; |
| 803 | } |
| 804 | |
| 805 | ksnprintf(tmpbuf + offset, sizeof(tmpbuf) - offset, |
| 806 | "%s %6.6lu, %8.8lu, %6.6lu, %6.6lu, %8.8lu\n", |
| 807 | tmpname, curzone->zsize, curzone->zmax, |
| 808 | (curzone->ztotal - freecnt), |
| 809 | freecnt, znalloc); |
| 810 | |
| 811 | len = strlen((char *)tmpbuf); |
| 812 | if (LIST_NEXT(curzone, zlink) == NULL) |
| 813 | tmpbuf[len - 1] = 0; |
| 814 | |
| 815 | error = SYSCTL_OUT(req, tmpbuf, len); |
| 816 | |
| 817 | if (error) |
| 818 | break; |
| 819 | } |
| 820 | lwkt_reltoken(&vm_token); |
| 821 | return (error); |
| 822 | } |
| 823 | |
| 824 | #if defined(INVARIANTS) |
| 825 | |
| 826 | /* |
| 827 | * Debugging only. |
| 828 | */ |
| 829 | void |
| 830 | zerror(int error) |
| 831 | { |
| 832 | char *msg; |
| 833 | |
| 834 | switch (error) { |
| 835 | case ZONE_ERROR_INVALID: |
| 836 | msg = "zone: invalid zone"; |
| 837 | break; |
| 838 | case ZONE_ERROR_NOTFREE: |
| 839 | msg = "zone: entry not free"; |
| 840 | break; |
| 841 | case ZONE_ERROR_ALREADYFREE: |
| 842 | msg = "zone: freeing free entry"; |
| 843 | break; |
| 844 | default: |
| 845 | msg = "zone: invalid error"; |
| 846 | break; |
| 847 | } |
| 848 | panic("%s", msg); |
| 849 | } |
| 850 | #endif |
| 851 | |
| 852 | SYSCTL_OID(_vm, OID_AUTO, zone, CTLTYPE_STRING|CTLFLAG_RD, \ |
| 853 | NULL, 0, sysctl_vm_zone, "A", "Zone Info"); |
| 854 | |
| 855 | SYSCTL_LONG(_vm, OID_AUTO, zone_kmem_pages, |
| 856 | CTLFLAG_RD, &zone_kmem_pages, 0, "Number of interrupt safe pages allocated by zone"); |
| 857 | SYSCTL_LONG(_vm, OID_AUTO, zone_burst, |
| 858 | CTLFLAG_RW, &zone_burst, 0, "Burst from depot to pcpu cache"); |
| 859 | SYSCTL_LONG(_vm, OID_AUTO, zone_kmem_kvaspace, |
| 860 | CTLFLAG_RD, &zone_kmem_kvaspace, 0, "KVA space allocated by zone"); |
| 861 | SYSCTL_LONG(_vm, OID_AUTO, zone_kern_pages, |
| 862 | CTLFLAG_RD, &zone_kern_pages, 0, "Number of non-interrupt safe pages allocated by zone"); |