1 /* $NetBSD: uvm_anon.c,v 1.31 2004/09/01 11:53:38 yamt Exp $ */ 2 3 /* 4 * 5 * Copyright (c) 1997 Charles D. Cranor and Washington University. 6 * All rights reserved. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by Charles D. Cranor and 19 * Washington University. 20 * 4. The name of the author may not be used to endorse or promote products 21 * derived from this software without specific prior written permission. 22 * 23 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 24 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 25 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 26 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 27 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 28 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 29 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 30 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 31 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 32 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 33 */ 34 35 /* 36 * uvm_anon.c: uvm anon ops 37 */ 38 39 #include <sys/cdefs.h> 40 __KERNEL_RCSID(0, "$NetBSD: uvm_anon.c,v 1.31 2004/09/01 11:53:38 yamt Exp $"); 41 42 #include "opt_uvmhist.h" 43 44 #include <sys/param.h> 45 #include <sys/systm.h> 46 #include <sys/proc.h> 47 #include <sys/malloc.h> 48 #include <sys/pool.h> 49 #include <sys/kernel.h> 50 51 #include <uvm/uvm.h> 52 #include <uvm/uvm_swap.h> 53 54 /* 55 * anonblock_list: global list of anon blocks, 56 * locked by swap_syscall_lock (since we never remove 57 * anything from this list and we only add to it via swapctl(2)). 58 */ 59 60 struct uvm_anonblock { 61 LIST_ENTRY(uvm_anonblock) list; 62 int count; 63 struct vm_anon *anons; 64 }; 65 static LIST_HEAD(anonlist, uvm_anonblock) anonblock_list; 66 67 68 static boolean_t anon_pagein(struct vm_anon *); 69 70 71 /* 72 * allocate anons 73 */ 74 void 75 uvm_anon_init() 76 { 77 int nanon = uvmexp.free - (uvmexp.free / 16); /* XXXCDC ??? */ 78 79 simple_lock_init(&uvm.afreelock); 80 LIST_INIT(&anonblock_list); 81 82 /* 83 * Allocate the initial anons. 84 */ 85 uvm_anon_add(nanon); 86 } 87 88 /* 89 * add some more anons to the free pool. called when we add 90 * more swap space. 91 * 92 * => swap_syscall_lock should be held (protects anonblock_list). 93 */ 94 int 95 uvm_anon_add(count) 96 int count; 97 { 98 struct uvm_anonblock *anonblock; 99 struct vm_anon *anon; 100 int lcv, needed; 101 102 simple_lock(&uvm.afreelock); 103 uvmexp.nanonneeded += count; 104 needed = uvmexp.nanonneeded - uvmexp.nanon; 105 simple_unlock(&uvm.afreelock); 106 107 if (needed <= 0) { 108 return 0; 109 } 110 anon = (void *)uvm_km_alloc(kernel_map, sizeof(*anon) * needed); 111 if (anon == NULL) { 112 simple_lock(&uvm.afreelock); 113 uvmexp.nanonneeded -= count; 114 simple_unlock(&uvm.afreelock); 115 return ENOMEM; 116 } 117 MALLOC(anonblock, void *, sizeof(*anonblock), M_UVMAMAP, M_WAITOK); 118 119 anonblock->count = needed; 120 anonblock->anons = anon; 121 LIST_INSERT_HEAD(&anonblock_list, anonblock, list); 122 memset(anon, 0, sizeof(*anon) * needed); 123 124 simple_lock(&uvm.afreelock); 125 uvmexp.nanon += needed; 126 uvmexp.nfreeanon += needed; 127 for (lcv = 0; lcv < needed; lcv++) { 128 simple_lock_init(&anon[lcv].an_lock); 129 anon[lcv].u.an_nxt = uvm.afree; 130 uvm.afree = &anon[lcv]; 131 } 132 simple_unlock(&uvm.afreelock); 133 return 0; 134 } 135 136 /* 137 * remove anons from the free pool. 138 */ 139 void 140 uvm_anon_remove(count) 141 int count; 142 { 143 /* 144 * we never actually free any anons, to avoid allocation overhead. 145 * XXX someday we might want to try to free anons. 146 */ 147 148 simple_lock(&uvm.afreelock); 149 uvmexp.nanonneeded -= count; 150 simple_unlock(&uvm.afreelock); 151 } 152 153 /* 154 * allocate an anon 155 * 156 * => new anon is returned locked! 157 */ 158 struct vm_anon * 159 uvm_analloc() 160 { 161 struct vm_anon *a; 162 163 simple_lock(&uvm.afreelock); 164 a = uvm.afree; 165 if (a) { 166 uvm.afree = a->u.an_nxt; 167 uvmexp.nfreeanon--; 168 a->an_ref = 1; 169 a->an_swslot = 0; 170 a->u.an_page = NULL; /* so we can free quickly */ 171 LOCK_ASSERT(simple_lock_held(&a->an_lock) == 0); 172 simple_lock(&a->an_lock); 173 } 174 simple_unlock(&uvm.afreelock); 175 return(a); 176 } 177 178 /* 179 * uvm_anfree: free a single anon structure 180 * 181 * => caller must remove anon from its amap before calling (if it was in 182 * an amap). 183 * => anon must be unlocked and have a zero reference count. 184 * => we may lock the pageq's. 185 */ 186 187 void 188 uvm_anfree(anon) 189 struct vm_anon *anon; 190 { 191 struct vm_page *pg; 192 UVMHIST_FUNC("uvm_anfree"); UVMHIST_CALLED(maphist); 193 UVMHIST_LOG(maphist,"(anon=0x%x)", anon, 0,0,0); 194 195 KASSERT(anon->an_ref == 0); 196 LOCK_ASSERT(!simple_lock_held(&anon->an_lock)); 197 198 /* 199 * get page 200 */ 201 202 pg = anon->u.an_page; 203 204 /* 205 * if there is a resident page and it is loaned, then anon may not 206 * own it. call out to uvm_anon_lockpage() to ensure the real owner 207 * of the page has been identified and locked. 208 */ 209 210 if (pg && pg->loan_count) { 211 simple_lock(&anon->an_lock); 212 pg = uvm_anon_lockloanpg(anon); 213 simple_unlock(&anon->an_lock); 214 } 215 216 /* 217 * if we have a resident page, we must dispose of it before freeing 218 * the anon. 219 */ 220 221 if (pg) { 222 223 /* 224 * if the page is owned by a uobject (now locked), then we must 225 * kill the loan on the page rather than free it. 226 */ 227 228 if (pg->uobject) { 229 uvm_lock_pageq(); 230 KASSERT(pg->loan_count > 0); 231 pg->loan_count--; 232 pg->uanon = NULL; 233 uvm_unlock_pageq(); 234 simple_unlock(&pg->uobject->vmobjlock); 235 } else { 236 237 /* 238 * page has no uobject, so we must be the owner of it. 239 */ 240 241 KASSERT((pg->flags & PG_RELEASED) == 0); 242 simple_lock(&anon->an_lock); 243 pmap_page_protect(pg, VM_PROT_NONE); 244 245 /* 246 * if the page is busy, mark it as PG_RELEASED 247 * so that uvm_anon_release will release it later. 248 */ 249 250 if (pg->flags & PG_BUSY) { 251 pg->flags |= PG_RELEASED; 252 simple_unlock(&anon->an_lock); 253 return; 254 } 255 uvm_lock_pageq(); 256 uvm_pagefree(pg); 257 uvm_unlock_pageq(); 258 simple_unlock(&anon->an_lock); 259 UVMHIST_LOG(maphist, "anon 0x%x, page 0x%x: " 260 "freed now!", anon, pg, 0, 0); 261 } 262 } 263 if (pg == NULL && anon->an_swslot > 0) { 264 /* this page is no longer only in swap. */ 265 simple_lock(&uvm.swap_data_lock); 266 KASSERT(uvmexp.swpgonly > 0); 267 uvmexp.swpgonly--; 268 simple_unlock(&uvm.swap_data_lock); 269 } 270 271 /* 272 * free any swap resources. 273 */ 274 275 uvm_anon_dropswap(anon); 276 277 /* 278 * now that we've stripped the data areas from the anon, 279 * free the anon itself. 280 */ 281 282 KASSERT(anon->u.an_page == NULL); 283 KASSERT(anon->an_swslot == 0); 284 285 simple_lock(&uvm.afreelock); 286 anon->u.an_nxt = uvm.afree; 287 uvm.afree = anon; 288 uvmexp.nfreeanon++; 289 simple_unlock(&uvm.afreelock); 290 UVMHIST_LOG(maphist,"<- done!",0,0,0,0); 291 } 292 293 /* 294 * uvm_anon_dropswap: release any swap resources from this anon. 295 * 296 * => anon must be locked or have a reference count of 0. 297 */ 298 void 299 uvm_anon_dropswap(anon) 300 struct vm_anon *anon; 301 { 302 UVMHIST_FUNC("uvm_anon_dropswap"); UVMHIST_CALLED(maphist); 303 304 if (anon->an_swslot == 0) 305 return; 306 307 UVMHIST_LOG(maphist,"freeing swap for anon %p, paged to swslot 0x%x", 308 anon, anon->an_swslot, 0, 0); 309 uvm_swap_free(anon->an_swslot, 1); 310 anon->an_swslot = 0; 311 } 312 313 /* 314 * uvm_anon_lockloanpg: given a locked anon, lock its resident page 315 * 316 * => anon is locked by caller 317 * => on return: anon is locked 318 * if there is a resident page: 319 * if it has a uobject, it is locked by us 320 * if it is ownerless, we take over as owner 321 * we return the resident page (it can change during 322 * this function) 323 * => note that the only time an anon has an ownerless resident page 324 * is if the page was loaned from a uvm_object and the uvm_object 325 * disowned it 326 * => this only needs to be called when you want to do an operation 327 * on an anon's resident page and that page has a non-zero loan 328 * count. 329 */ 330 struct vm_page * 331 uvm_anon_lockloanpg(anon) 332 struct vm_anon *anon; 333 { 334 struct vm_page *pg; 335 boolean_t locked = FALSE; 336 337 LOCK_ASSERT(simple_lock_held(&anon->an_lock)); 338 339 /* 340 * loop while we have a resident page that has a non-zero loan count. 341 * if we successfully get our lock, we will "break" the loop. 342 * note that the test for pg->loan_count is not protected -- this 343 * may produce false positive results. note that a false positive 344 * result may cause us to do more work than we need to, but it will 345 * not produce an incorrect result. 346 */ 347 348 while (((pg = anon->u.an_page) != NULL) && pg->loan_count != 0) { 349 350 /* 351 * quickly check to see if the page has an object before 352 * bothering to lock the page queues. this may also produce 353 * a false positive result, but that's ok because we do a real 354 * check after that. 355 */ 356 357 if (pg->uobject) { 358 uvm_lock_pageq(); 359 if (pg->uobject) { 360 locked = 361 simple_lock_try(&pg->uobject->vmobjlock); 362 } else { 363 /* object disowned before we got PQ lock */ 364 locked = TRUE; 365 } 366 uvm_unlock_pageq(); 367 368 /* 369 * if we didn't get a lock (try lock failed), then we 370 * toggle our anon lock and try again 371 */ 372 373 if (!locked) { 374 simple_unlock(&anon->an_lock); 375 376 /* 377 * someone locking the object has a chance to 378 * lock us right now 379 */ 380 381 simple_lock(&anon->an_lock); 382 continue; 383 } 384 } 385 386 /* 387 * if page is un-owned [i.e. the object dropped its ownership], 388 * then we can take over as owner! 389 */ 390 391 if (pg->uobject == NULL && (pg->pqflags & PQ_ANON) == 0) { 392 uvm_lock_pageq(); 393 pg->pqflags |= PQ_ANON; 394 pg->loan_count--; 395 uvm_unlock_pageq(); 396 } 397 break; 398 } 399 return(pg); 400 } 401 402 403 404 /* 405 * page in every anon that is paged out to a range of swslots. 406 * 407 * swap_syscall_lock should be held (protects anonblock_list). 408 */ 409 410 boolean_t 411 anon_swap_off(startslot, endslot) 412 int startslot, endslot; 413 { 414 struct uvm_anonblock *anonblock; 415 416 LIST_FOREACH(anonblock, &anonblock_list, list) { 417 int i; 418 419 /* 420 * loop thru all the anons in the anonblock, 421 * paging in where needed. 422 */ 423 424 for (i = 0; i < anonblock->count; i++) { 425 struct vm_anon *anon = &anonblock->anons[i]; 426 int slot; 427 428 /* 429 * lock anon to work on it. 430 */ 431 432 simple_lock(&anon->an_lock); 433 434 /* 435 * is this anon's swap slot in range? 436 */ 437 438 slot = anon->an_swslot; 439 if (slot >= startslot && slot < endslot) { 440 boolean_t rv; 441 442 /* 443 * yup, page it in. 444 */ 445 446 /* locked: anon */ 447 rv = anon_pagein(anon); 448 /* unlocked: anon */ 449 450 if (rv) { 451 return rv; 452 } 453 } else { 454 455 /* 456 * nope, unlock and proceed. 457 */ 458 459 simple_unlock(&anon->an_lock); 460 } 461 } 462 } 463 return FALSE; 464 } 465 466 467 /* 468 * fetch an anon's page. 469 * 470 * => anon must be locked, and is unlocked upon return. 471 * => returns TRUE if pagein was aborted due to lack of memory. 472 */ 473 474 static boolean_t 475 anon_pagein(anon) 476 struct vm_anon *anon; 477 { 478 struct vm_page *pg; 479 struct uvm_object *uobj; 480 int rv; 481 482 /* locked: anon */ 483 LOCK_ASSERT(simple_lock_held(&anon->an_lock)); 484 485 rv = uvmfault_anonget(NULL, NULL, anon); 486 487 /* 488 * if rv == 0, anon is still locked, else anon 489 * is unlocked 490 */ 491 492 switch (rv) { 493 case 0: 494 break; 495 496 case EIO: 497 case ERESTART: 498 499 /* 500 * nothing more to do on errors. 501 * ERESTART can only mean that the anon was freed, 502 * so again there's nothing to do. 503 */ 504 505 return FALSE; 506 507 default: 508 return TRUE; 509 } 510 511 /* 512 * ok, we've got the page now. 513 * mark it as dirty, clear its swslot and un-busy it. 514 */ 515 516 pg = anon->u.an_page; 517 uobj = pg->uobject; 518 if (anon->an_swslot > 0) 519 uvm_swap_free(anon->an_swslot, 1); 520 anon->an_swslot = 0; 521 pg->flags &= ~(PG_CLEAN); 522 523 /* 524 * deactivate the page (to put it on a page queue) 525 */ 526 527 pmap_clear_reference(pg); 528 uvm_lock_pageq(); 529 if (pg->wire_count == 0) 530 uvm_pagedeactivate(pg); 531 uvm_unlock_pageq(); 532 533 if (pg->flags & PG_WANTED) { 534 wakeup(pg); 535 pg->flags &= ~(PG_WANTED); 536 } 537 538 /* 539 * unlock the anon and we're done. 540 */ 541 542 simple_unlock(&anon->an_lock); 543 if (uobj) { 544 simple_unlock(&uobj->vmobjlock); 545 } 546 return FALSE; 547 } 548 549 /* 550 * uvm_anon_release: release an anon and its page. 551 * 552 * => caller must lock the anon. 553 */ 554 555 void 556 uvm_anon_release(anon) 557 struct vm_anon *anon; 558 { 559 struct vm_page *pg = anon->u.an_page; 560 561 LOCK_ASSERT(simple_lock_held(&anon->an_lock)); 562 563 KASSERT(pg != NULL); 564 KASSERT((pg->flags & PG_RELEASED) != 0); 565 KASSERT((pg->flags & PG_BUSY) != 0); 566 KASSERT(pg->uobject == NULL); 567 KASSERT(pg->uanon == anon); 568 KASSERT(pg->loan_count == 0); 569 KASSERT(anon->an_ref == 0); 570 571 uvm_lock_pageq(); 572 uvm_pagefree(pg); 573 uvm_unlock_pageq(); 574 simple_unlock(&anon->an_lock); 575 576 KASSERT(anon->u.an_page == NULL); 577 578 uvm_anfree(anon); 579 } 580