1 /* $OpenBSD: uvm_aobj.c,v 1.54 2011/07/03 18:34:14 oga Exp $ */ 2 /* $NetBSD: uvm_aobj.c,v 1.39 2001/02/18 21:19:08 chs Exp $ */ 3 4 /* 5 * Copyright (c) 1998 Chuck Silvers, Charles D. Cranor and 6 * Washington University. 7 * All rights reserved. 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 3. All advertising materials mentioning features or use of this software 18 * must display the following acknowledgement: 19 * This product includes software developed by Charles D. Cranor and 20 * Washington University. 21 * 4. The name of the author may not be used to endorse or promote products 22 * derived from this software without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 25 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 26 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 27 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 28 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 29 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 30 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 31 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 32 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 33 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 34 * 35 * from: Id: uvm_aobj.c,v 1.1.2.5 1998/02/06 05:14:38 chs Exp 36 */ 37 /* 38 * uvm_aobj.c: anonymous memory uvm_object pager 39 * 40 * author: Chuck Silvers <chuq@chuq.com> 41 * started: Jan-1998 42 * 43 * - design mostly from Chuck Cranor 44 */ 45 46 #include <sys/param.h> 47 #include <sys/systm.h> 48 #include <sys/proc.h> 49 #include <sys/malloc.h> 50 #include <sys/kernel.h> 51 #include <sys/pool.h> 52 #include <sys/kernel.h> 53 54 #include <uvm/uvm.h> 55 56 /* 57 * an aobj manages anonymous-memory backed uvm_objects. in addition 58 * to keeping the list of resident pages, it also keeps a list of 59 * allocated swap blocks. depending on the size of the aobj this list 60 * of allocated swap blocks is either stored in an array (small objects) 61 * or in a hash table (large objects). 62 */ 63 64 /* 65 * local structures 66 */ 67 68 /* 69 * for hash tables, we break the address space of the aobj into blocks 70 * of UAO_SWHASH_CLUSTER_SIZE pages. we require the cluster size to 71 * be a power of two. 72 */ 73 74 #define UAO_SWHASH_CLUSTER_SHIFT 4 75 #define UAO_SWHASH_CLUSTER_SIZE (1 << UAO_SWHASH_CLUSTER_SHIFT) 76 77 /* get the "tag" for this page index */ 78 #define UAO_SWHASH_ELT_TAG(PAGEIDX) \ 79 ((PAGEIDX) >> UAO_SWHASH_CLUSTER_SHIFT) 80 81 /* given an ELT and a page index, find the swap slot */ 82 #define UAO_SWHASH_ELT_PAGESLOT(ELT, PAGEIDX) \ 83 ((ELT)->slots[(PAGEIDX) & (UAO_SWHASH_CLUSTER_SIZE - 1)]) 84 85 /* given an ELT, return its pageidx base */ 86 #define UAO_SWHASH_ELT_PAGEIDX_BASE(ELT) \ 87 ((ELT)->tag << UAO_SWHASH_CLUSTER_SHIFT) 88 89 /* 90 * the swhash hash function 91 */ 92 #define UAO_SWHASH_HASH(AOBJ, PAGEIDX) \ 93 (&(AOBJ)->u_swhash[(((PAGEIDX) >> UAO_SWHASH_CLUSTER_SHIFT) \ 94 & (AOBJ)->u_swhashmask)]) 95 96 /* 97 * the swhash threshold determines if we will use an array or a 98 * hash table to store the list of allocated swap blocks. 99 */ 100 101 #define UAO_SWHASH_THRESHOLD (UAO_SWHASH_CLUSTER_SIZE * 4) 102 #define UAO_USES_SWHASH(AOBJ) \ 103 ((AOBJ)->u_pages > UAO_SWHASH_THRESHOLD) /* use hash? */ 104 105 /* 106 * the number of buckets in a swhash, with an upper bound 107 */ 108 #define UAO_SWHASH_MAXBUCKETS 256 109 #define UAO_SWHASH_BUCKETS(AOBJ) \ 110 (min((AOBJ)->u_pages >> UAO_SWHASH_CLUSTER_SHIFT, \ 111 UAO_SWHASH_MAXBUCKETS)) 112 113 114 /* 115 * uao_swhash_elt: when a hash table is being used, this structure defines 116 * the format of an entry in the bucket list. 117 */ 118 119 struct uao_swhash_elt { 120 LIST_ENTRY(uao_swhash_elt) list; /* the hash list */ 121 voff_t tag; /* our 'tag' */ 122 int count; /* our number of active slots */ 123 int slots[UAO_SWHASH_CLUSTER_SIZE]; /* the slots */ 124 }; 125 126 /* 127 * uao_swhash: the swap hash table structure 128 */ 129 130 LIST_HEAD(uao_swhash, uao_swhash_elt); 131 132 /* 133 * uao_swhash_elt_pool: pool of uao_swhash_elt structures 134 */ 135 136 struct pool uao_swhash_elt_pool; 137 138 /* 139 * uvm_aobj: the actual anon-backed uvm_object 140 * 141 * => the uvm_object is at the top of the structure, this allows 142 * (struct uvm_aobj *) == (struct uvm_object *) 143 * => only one of u_swslots and u_swhash is used in any given aobj 144 */ 145 146 struct uvm_aobj { 147 struct uvm_object u_obj; /* has: lock, pgops, memt, #pages, #refs */ 148 int u_pages; /* number of pages in entire object */ 149 int u_flags; /* the flags (see uvm_aobj.h) */ 150 /* 151 * Either an array or hashtable (array of bucket heads) of 152 * offset -> swapslot mappings for the aobj. 153 */ 154 #define u_swslots u_swap.slot_array 155 #define u_swhash u_swap.slot_hash 156 union swslots { 157 int *slot_array; 158 struct uao_swhash *slot_hash; 159 } u_swap; 160 u_long u_swhashmask; /* mask for hashtable */ 161 LIST_ENTRY(uvm_aobj) u_list; /* global list of aobjs */ 162 }; 163 164 /* 165 * uvm_aobj_pool: pool of uvm_aobj structures 166 */ 167 168 struct pool uvm_aobj_pool; 169 170 /* 171 * local functions 172 */ 173 174 static struct uao_swhash_elt *uao_find_swhash_elt(struct uvm_aobj *, int, 175 boolean_t); 176 static int uao_find_swslot(struct uvm_aobj *, int); 177 static boolean_t uao_flush(struct uvm_object *, voff_t, 178 voff_t, int); 179 static void uao_free(struct uvm_aobj *); 180 static int uao_get(struct uvm_object *, voff_t, 181 vm_page_t *, int *, int, vm_prot_t, 182 int, int); 183 static boolean_t uao_pagein(struct uvm_aobj *, int, int); 184 static boolean_t uao_pagein_page(struct uvm_aobj *, int); 185 186 /* 187 * aobj_pager 188 * 189 * note that some functions (e.g. put) are handled elsewhere 190 */ 191 192 struct uvm_pagerops aobj_pager = { 193 NULL, /* init */ 194 uao_reference, /* reference */ 195 uao_detach, /* detach */ 196 NULL, /* fault */ 197 uao_flush, /* flush */ 198 uao_get, /* get */ 199 }; 200 201 /* 202 * uao_list: global list of active aobjs, locked by uao_list_lock 203 * 204 * Lock ordering: generally the locking order is object lock, then list lock. 205 * in the case of swap off we have to iterate over the list, and thus the 206 * ordering is reversed. In that case we must use trylocking to prevent 207 * deadlock. 208 */ 209 210 static LIST_HEAD(aobjlist, uvm_aobj) uao_list = LIST_HEAD_INITIALIZER(uao_list); 211 static struct mutex uao_list_lock = MUTEX_INITIALIZER(IPL_NONE); 212 213 214 /* 215 * functions 216 */ 217 218 /* 219 * hash table/array related functions 220 */ 221 222 /* 223 * uao_find_swhash_elt: find (or create) a hash table entry for a page 224 * offset. 225 * 226 * => the object should be locked by the caller 227 */ 228 229 static struct uao_swhash_elt * 230 uao_find_swhash_elt(struct uvm_aobj *aobj, int pageidx, boolean_t create) 231 { 232 struct uao_swhash *swhash; 233 struct uao_swhash_elt *elt; 234 voff_t page_tag; 235 236 swhash = UAO_SWHASH_HASH(aobj, pageidx); /* first hash to get bucket */ 237 page_tag = UAO_SWHASH_ELT_TAG(pageidx); /* tag to search for */ 238 239 /* 240 * now search the bucket for the requested tag 241 */ 242 LIST_FOREACH(elt, swhash, list) { 243 if (elt->tag == page_tag) 244 return(elt); 245 } 246 247 /* fail now if we are not allowed to create a new entry in the bucket */ 248 if (!create) 249 return NULL; 250 251 252 /* 253 * allocate a new entry for the bucket and init/insert it in 254 */ 255 elt = pool_get(&uao_swhash_elt_pool, PR_WAITOK | PR_ZERO); 256 LIST_INSERT_HEAD(swhash, elt, list); 257 elt->tag = page_tag; 258 259 return(elt); 260 } 261 262 /* 263 * uao_find_swslot: find the swap slot number for an aobj/pageidx 264 * 265 * => object must be locked by caller 266 */ 267 __inline static int 268 uao_find_swslot(struct uvm_aobj *aobj, int pageidx) 269 { 270 271 /* 272 * if noswap flag is set, then we never return a slot 273 */ 274 275 if (aobj->u_flags & UAO_FLAG_NOSWAP) 276 return(0); 277 278 /* 279 * if hashing, look in hash table. 280 */ 281 282 if (UAO_USES_SWHASH(aobj)) { 283 struct uao_swhash_elt *elt = 284 uao_find_swhash_elt(aobj, pageidx, FALSE); 285 286 if (elt) 287 return(UAO_SWHASH_ELT_PAGESLOT(elt, pageidx)); 288 else 289 return(0); 290 } 291 292 /* 293 * otherwise, look in the array 294 */ 295 return(aobj->u_swslots[pageidx]); 296 } 297 298 /* 299 * uao_set_swslot: set the swap slot for a page in an aobj. 300 * 301 * => setting a slot to zero frees the slot 302 * => object must be locked by caller 303 */ 304 int 305 uao_set_swslot(struct uvm_object *uobj, int pageidx, int slot) 306 { 307 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 308 int oldslot; 309 310 /* 311 * if noswap flag is set, then we can't set a slot 312 */ 313 314 if (aobj->u_flags & UAO_FLAG_NOSWAP) { 315 316 if (slot == 0) 317 return(0); /* a clear is ok */ 318 319 /* but a set is not */ 320 printf("uao_set_swslot: uobj = %p\n", uobj); 321 panic("uao_set_swslot: attempt to set a slot on a NOSWAP object"); 322 } 323 324 /* 325 * are we using a hash table? if so, add it in the hash. 326 */ 327 328 if (UAO_USES_SWHASH(aobj)) { 329 330 /* 331 * Avoid allocating an entry just to free it again if 332 * the page had not swap slot in the first place, and 333 * we are freeing. 334 */ 335 336 struct uao_swhash_elt *elt = 337 uao_find_swhash_elt(aobj, pageidx, slot ? TRUE : FALSE); 338 if (elt == NULL) { 339 KASSERT(slot == 0); 340 return (0); 341 } 342 343 oldslot = UAO_SWHASH_ELT_PAGESLOT(elt, pageidx); 344 UAO_SWHASH_ELT_PAGESLOT(elt, pageidx) = slot; 345 346 /* 347 * now adjust the elt's reference counter and free it if we've 348 * dropped it to zero. 349 */ 350 351 /* an allocation? */ 352 if (slot) { 353 if (oldslot == 0) 354 elt->count++; 355 } else { /* freeing slot ... */ 356 if (oldslot) /* to be safe */ 357 elt->count--; 358 359 if (elt->count == 0) { 360 LIST_REMOVE(elt, list); 361 pool_put(&uao_swhash_elt_pool, elt); 362 } 363 } 364 } else { 365 /* we are using an array */ 366 oldslot = aobj->u_swslots[pageidx]; 367 aobj->u_swslots[pageidx] = slot; 368 } 369 return (oldslot); 370 } 371 372 /* 373 * end of hash/array functions 374 */ 375 376 /* 377 * uao_free: free all resources held by an aobj, and then free the aobj 378 * 379 * => the aobj should be dead 380 */ 381 static void 382 uao_free(struct uvm_aobj *aobj) 383 { 384 385 simple_unlock(&aobj->u_obj.vmobjlock); 386 387 if (UAO_USES_SWHASH(aobj)) { 388 int i, hashbuckets = aobj->u_swhashmask + 1; 389 390 /* 391 * free the swslots from each hash bucket, 392 * then the hash bucket, and finally the hash table itself. 393 */ 394 for (i = 0; i < hashbuckets; i++) { 395 struct uao_swhash_elt *elt, *next; 396 397 for (elt = LIST_FIRST(&aobj->u_swhash[i]); 398 elt != NULL; 399 elt = next) { 400 int j; 401 402 for (j = 0; j < UAO_SWHASH_CLUSTER_SIZE; j++) { 403 int slot = elt->slots[j]; 404 405 if (slot == 0) { 406 continue; 407 } 408 uvm_swap_free(slot, 1); 409 410 /* 411 * this page is no longer 412 * only in swap. 413 */ 414 simple_lock(&uvm.swap_data_lock); 415 uvmexp.swpgonly--; 416 simple_unlock(&uvm.swap_data_lock); 417 } 418 419 next = LIST_NEXT(elt, list); 420 pool_put(&uao_swhash_elt_pool, elt); 421 } 422 } 423 free(aobj->u_swhash, M_UVMAOBJ); 424 } else { 425 int i; 426 427 /* 428 * free the array 429 */ 430 431 for (i = 0; i < aobj->u_pages; i++) { 432 int slot = aobj->u_swslots[i]; 433 434 if (slot) { 435 uvm_swap_free(slot, 1); 436 437 /* this page is no longer only in swap. */ 438 simple_lock(&uvm.swap_data_lock); 439 uvmexp.swpgonly--; 440 simple_unlock(&uvm.swap_data_lock); 441 } 442 } 443 free(aobj->u_swslots, M_UVMAOBJ); 444 } 445 446 /* 447 * finally free the aobj itself 448 */ 449 pool_put(&uvm_aobj_pool, aobj); 450 } 451 452 /* 453 * pager functions 454 */ 455 456 /* 457 * uao_create: create an aobj of the given size and return its uvm_object. 458 * 459 * => for normal use, flags are always zero 460 * => for the kernel object, the flags are: 461 * UAO_FLAG_KERNOBJ - allocate the kernel object (can only happen once) 462 * UAO_FLAG_KERNSWAP - enable swapping of kernel object (" ") 463 */ 464 struct uvm_object * 465 uao_create(vsize_t size, int flags) 466 { 467 static struct uvm_aobj kernel_object_store; /* home of kernel_object */ 468 static int kobj_alloced = 0; /* not allocated yet */ 469 int pages = round_page(size) >> PAGE_SHIFT; 470 int refs = UVM_OBJ_KERN; 471 struct uvm_aobj *aobj; 472 473 /* 474 * malloc a new aobj unless we are asked for the kernel object 475 */ 476 if (flags & UAO_FLAG_KERNOBJ) { /* want kernel object? */ 477 if (kobj_alloced) 478 panic("uao_create: kernel object already allocated"); 479 480 aobj = &kernel_object_store; 481 aobj->u_pages = pages; 482 aobj->u_flags = UAO_FLAG_NOSWAP; /* no swap to start */ 483 /* we are special, we never die */ 484 kobj_alloced = UAO_FLAG_KERNOBJ; 485 } else if (flags & UAO_FLAG_KERNSWAP) { 486 aobj = &kernel_object_store; 487 if (kobj_alloced != UAO_FLAG_KERNOBJ) 488 panic("uao_create: asked to enable swap on kernel object"); 489 kobj_alloced = UAO_FLAG_KERNSWAP; 490 } else { /* normal object */ 491 aobj = pool_get(&uvm_aobj_pool, PR_WAITOK); 492 aobj->u_pages = pages; 493 aobj->u_flags = 0; /* normal object */ 494 refs = 1; /* normal object so 1 ref */ 495 } 496 497 /* 498 * allocate hash/array if necessary 499 * 500 * note: in the KERNSWAP case no need to worry about locking since 501 * we are still booting we should be the only thread around. 502 */ 503 if (flags == 0 || (flags & UAO_FLAG_KERNSWAP) != 0) { 504 int mflags = (flags & UAO_FLAG_KERNSWAP) != 0 ? 505 M_NOWAIT : M_WAITOK; 506 507 /* allocate hash table or array depending on object size */ 508 if (UAO_USES_SWHASH(aobj)) { 509 aobj->u_swhash = hashinit(UAO_SWHASH_BUCKETS(aobj), 510 M_UVMAOBJ, mflags, &aobj->u_swhashmask); 511 if (aobj->u_swhash == NULL) 512 panic("uao_create: hashinit swhash failed"); 513 } else { 514 aobj->u_swslots = malloc(pages * sizeof(int), 515 M_UVMAOBJ, mflags|M_ZERO); 516 if (aobj->u_swslots == NULL) 517 panic("uao_create: malloc swslots failed"); 518 } 519 520 if (flags) { 521 aobj->u_flags &= ~UAO_FLAG_NOSWAP; /* clear noswap */ 522 return(&aobj->u_obj); 523 /* done! */ 524 } 525 } 526 527 uvm_objinit(&aobj->u_obj, &aobj_pager, refs); 528 529 /* 530 * now that aobj is ready, add it to the global list 531 */ 532 mtx_enter(&uao_list_lock); 533 LIST_INSERT_HEAD(&uao_list, aobj, u_list); 534 mtx_leave(&uao_list_lock); 535 536 /* 537 * done! 538 */ 539 return(&aobj->u_obj); 540 } 541 542 543 544 /* 545 * uao_init: set up aobj pager subsystem 546 * 547 * => called at boot time from uvm_pager_init() 548 */ 549 void 550 uao_init(void) 551 { 552 static int uao_initialized; 553 554 if (uao_initialized) 555 return; 556 uao_initialized = TRUE; 557 558 /* 559 * NOTE: Pages for this pool must not come from a pageable 560 * kernel map! 561 */ 562 pool_init(&uao_swhash_elt_pool, sizeof(struct uao_swhash_elt), 563 0, 0, 0, "uaoeltpl", &pool_allocator_nointr); 564 565 pool_init(&uvm_aobj_pool, sizeof(struct uvm_aobj), 0, 0, 0, 566 "aobjpl", &pool_allocator_nointr); 567 } 568 569 /* 570 * uao_reference: add a ref to an aobj 571 * 572 * => aobj must be unlocked 573 * => just lock it and call the locked version 574 */ 575 void 576 uao_reference(struct uvm_object *uobj) 577 { 578 simple_lock(&uobj->vmobjlock); 579 uao_reference_locked(uobj); 580 simple_unlock(&uobj->vmobjlock); 581 } 582 583 /* 584 * uao_reference_locked: add a ref to an aobj that is already locked 585 * 586 * => aobj must be locked 587 * this needs to be separate from the normal routine 588 * since sometimes we need to add a reference to an aobj when 589 * it's already locked. 590 */ 591 void 592 uao_reference_locked(struct uvm_object *uobj) 593 { 594 595 /* 596 * kernel_object already has plenty of references, leave it alone. 597 */ 598 599 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) 600 return; 601 602 uobj->uo_refs++; /* bump! */ 603 } 604 605 606 /* 607 * uao_detach: drop a reference to an aobj 608 * 609 * => aobj must be unlocked 610 * => just lock it and call the locked version 611 */ 612 void 613 uao_detach(struct uvm_object *uobj) 614 { 615 simple_lock(&uobj->vmobjlock); 616 uao_detach_locked(uobj); 617 } 618 619 620 /* 621 * uao_detach_locked: drop a reference to an aobj 622 * 623 * => aobj must be locked, and is unlocked (or freed) upon return. 624 * this needs to be separate from the normal routine 625 * since sometimes we need to detach from an aobj when 626 * it's already locked. 627 */ 628 void 629 uao_detach_locked(struct uvm_object *uobj) 630 { 631 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 632 struct vm_page *pg; 633 634 /* 635 * detaching from kernel_object is a noop. 636 */ 637 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) { 638 simple_unlock(&uobj->vmobjlock); 639 return; 640 } 641 642 uobj->uo_refs--; /* drop ref! */ 643 if (uobj->uo_refs) { /* still more refs? */ 644 simple_unlock(&uobj->vmobjlock); 645 return; 646 } 647 648 /* 649 * remove the aobj from the global list. 650 */ 651 mtx_enter(&uao_list_lock); 652 LIST_REMOVE(aobj, u_list); 653 mtx_leave(&uao_list_lock); 654 655 /* 656 * Free all pages left in the object. If they're busy, wait 657 * for them to become available before we kill it. 658 * Release swap resources then free the page. 659 */ 660 uvm_lock_pageq(); 661 while((pg = RB_ROOT(&uobj->memt)) != NULL) { 662 if (pg->pg_flags & PG_BUSY) { 663 atomic_setbits_int(&pg->pg_flags, PG_WANTED); 664 uvm_unlock_pageq(); 665 UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, 0, 666 "uao_det", 0); 667 simple_lock(&uobj->vmobjlock); 668 uvm_lock_pageq(); 669 continue; 670 } 671 pmap_page_protect(pg, VM_PROT_NONE); 672 uao_dropswap(&aobj->u_obj, pg->offset >> PAGE_SHIFT); 673 uvm_pagefree(pg); 674 } 675 uvm_unlock_pageq(); 676 677 /* 678 * finally, free the rest. 679 */ 680 uao_free(aobj); 681 } 682 683 /* 684 * uao_flush: "flush" pages out of a uvm object 685 * 686 * => object should be locked by caller. we may _unlock_ the object 687 * if (and only if) we need to clean a page (PGO_CLEANIT). 688 * XXXJRT Currently, however, we don't. In the case of cleaning 689 * XXXJRT a page, we simply just deactivate it. Should probably 690 * XXXJRT handle this better, in the future (although "flushing" 691 * XXXJRT anonymous memory isn't terribly important). 692 * => if PGO_CLEANIT is not set, then we will neither unlock the object 693 * or block. 694 * => if PGO_ALLPAGE is set, then all pages in the object are valid targets 695 * for flushing. 696 * => NOTE: we are allowed to lock the page queues, so the caller 697 * must not be holding the lock on them [e.g. pagedaemon had 698 * better not call us with the queues locked] 699 * => we return TRUE unless we encountered some sort of I/O error 700 * XXXJRT currently never happens, as we never directly initiate 701 * XXXJRT I/O 702 */ 703 704 #define UAO_HASH_PENALTY 4 /* XXX: a guess */ 705 706 boolean_t 707 uao_flush(struct uvm_object *uobj, voff_t start, voff_t stop, int flags) 708 { 709 struct uvm_aobj *aobj = (struct uvm_aobj *) uobj; 710 struct vm_page *pp; 711 voff_t curoff; 712 713 if (flags & PGO_ALLPAGES) { 714 start = 0; 715 stop = aobj->u_pages << PAGE_SHIFT; 716 } else { 717 start = trunc_page(start); 718 stop = round_page(stop); 719 if (stop > (aobj->u_pages << PAGE_SHIFT)) { 720 printf("uao_flush: strange, got an out of range " 721 "flush (fixed)\n"); 722 stop = aobj->u_pages << PAGE_SHIFT; 723 } 724 } 725 726 /* 727 * Don't need to do any work here if we're not freeing 728 * or deactivating pages. 729 */ 730 if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) 731 return (TRUE); 732 733 /* locked: uobj */ 734 curoff = start; 735 for (;;) { 736 if (curoff < stop) { 737 pp = uvm_pagelookup(uobj, curoff); 738 curoff += PAGE_SIZE; 739 if (pp == NULL) 740 continue; 741 } else { 742 break; 743 } 744 745 /* Make sure page is unbusy, else wait for it. */ 746 if (pp->pg_flags & PG_BUSY) { 747 atomic_setbits_int(&pp->pg_flags, PG_WANTED); 748 UVM_UNLOCK_AND_WAIT(pp, &uobj->vmobjlock, 0, 749 "uaoflsh", 0); 750 simple_lock(&uobj->vmobjlock); 751 curoff -= PAGE_SIZE; 752 continue; 753 } 754 755 switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) { 756 /* 757 * XXX In these first 3 cases, we always just 758 * XXX deactivate the page. We may want to 759 * XXX handle the different cases more specifically 760 * XXX in the future. 761 */ 762 case PGO_CLEANIT|PGO_FREE: 763 /* FALLTHROUGH */ 764 case PGO_CLEANIT|PGO_DEACTIVATE: 765 /* FALLTHROUGH */ 766 case PGO_DEACTIVATE: 767 deactivate_it: 768 /* skip the page if it's loaned or wired */ 769 if (pp->loan_count != 0 || 770 pp->wire_count != 0) 771 continue; 772 773 uvm_lock_pageq(); 774 /* zap all mappings for the page. */ 775 pmap_page_protect(pp, VM_PROT_NONE); 776 777 /* ...and deactivate the page. */ 778 uvm_pagedeactivate(pp); 779 uvm_unlock_pageq(); 780 781 continue; 782 783 case PGO_FREE: 784 /* 785 * If there are multiple references to 786 * the object, just deactivate the page. 787 */ 788 if (uobj->uo_refs > 1) 789 goto deactivate_it; 790 791 /* XXX skip the page if it's loaned or wired */ 792 if (pp->loan_count != 0 || 793 pp->wire_count != 0) 794 continue; 795 796 /* zap all mappings for the page. */ 797 pmap_page_protect(pp, VM_PROT_NONE); 798 799 uao_dropswap(uobj, pp->offset >> PAGE_SHIFT); 800 uvm_lock_pageq(); 801 uvm_pagefree(pp); 802 uvm_unlock_pageq(); 803 804 continue; 805 806 default: 807 panic("uao_flush: weird flags"); 808 } 809 } 810 811 return (TRUE); 812 } 813 814 /* 815 * uao_get: fetch me a page 816 * 817 * we have three cases: 818 * 1: page is resident -> just return the page. 819 * 2: page is zero-fill -> allocate a new page and zero it. 820 * 3: page is swapped out -> fetch the page from swap. 821 * 822 * cases 1 and 2 can be handled with PGO_LOCKED, case 3 cannot. 823 * so, if the "center" page hits case 3 (or any page, with PGO_ALLPAGES), 824 * then we will need to return VM_PAGER_UNLOCK. 825 * 826 * => prefer map unlocked (not required) 827 * => object must be locked! we will _unlock_ it before starting any I/O. 828 * => flags: PGO_ALLPAGES: get all of the pages 829 * PGO_LOCKED: fault data structures are locked 830 * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx] 831 * => NOTE: caller must check for released pages!! 832 */ 833 static int 834 uao_get(struct uvm_object *uobj, voff_t offset, struct vm_page **pps, 835 int *npagesp, int centeridx, vm_prot_t access_type, int advice, int flags) 836 { 837 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 838 voff_t current_offset; 839 vm_page_t ptmp; 840 int lcv, gotpages, maxpages, swslot, rv, pageidx; 841 boolean_t done; 842 843 /* 844 * get number of pages 845 */ 846 maxpages = *npagesp; 847 848 /* 849 * step 1: handled the case where fault data structures are locked. 850 */ 851 852 if (flags & PGO_LOCKED) { 853 /* 854 * step 1a: get pages that are already resident. only do 855 * this if the data structures are locked (i.e. the first 856 * time through). 857 */ 858 859 done = TRUE; /* be optimistic */ 860 gotpages = 0; /* # of pages we got so far */ 861 862 for (lcv = 0, current_offset = offset ; lcv < maxpages ; 863 lcv++, current_offset += PAGE_SIZE) { 864 /* do we care about this page? if not, skip it */ 865 if (pps[lcv] == PGO_DONTCARE) 866 continue; 867 868 ptmp = uvm_pagelookup(uobj, current_offset); 869 870 /* 871 * if page is new, attempt to allocate the page, 872 * zero-fill'd. 873 */ 874 if (ptmp == NULL && uao_find_swslot(aobj, 875 current_offset >> PAGE_SHIFT) == 0) { 876 ptmp = uvm_pagealloc(uobj, current_offset, 877 NULL, UVM_PGA_ZERO); 878 if (ptmp) { 879 /* new page */ 880 atomic_clearbits_int(&ptmp->pg_flags, 881 PG_BUSY|PG_FAKE); 882 atomic_setbits_int(&ptmp->pg_flags, 883 PQ_AOBJ); 884 UVM_PAGE_OWN(ptmp, NULL); 885 } 886 } 887 888 /* 889 * to be useful must get a non-busy page 890 */ 891 if (ptmp == NULL || 892 (ptmp->pg_flags & PG_BUSY) != 0) { 893 if (lcv == centeridx || 894 (flags & PGO_ALLPAGES) != 0) 895 /* need to do a wait or I/O! */ 896 done = FALSE; 897 continue; 898 } 899 900 /* 901 * useful page: busy/lock it and plug it in our 902 * result array 903 */ 904 /* caller must un-busy this page */ 905 atomic_setbits_int(&ptmp->pg_flags, PG_BUSY); 906 UVM_PAGE_OWN(ptmp, "uao_get1"); 907 pps[lcv] = ptmp; 908 gotpages++; 909 910 } /* "for" lcv loop */ 911 912 /* 913 * step 1b: now we've either done everything needed or we 914 * to unlock and do some waiting or I/O. 915 */ 916 917 *npagesp = gotpages; 918 if (done) 919 /* bingo! */ 920 return(VM_PAGER_OK); 921 else 922 /* EEK! Need to unlock and I/O */ 923 return(VM_PAGER_UNLOCK); 924 } 925 926 /* 927 * step 2: get non-resident or busy pages. 928 * object is locked. data structures are unlocked. 929 */ 930 931 for (lcv = 0, current_offset = offset ; lcv < maxpages ; 932 lcv++, current_offset += PAGE_SIZE) { 933 934 /* 935 * - skip over pages we've already gotten or don't want 936 * - skip over pages we don't _have_ to get 937 */ 938 939 if (pps[lcv] != NULL || 940 (lcv != centeridx && (flags & PGO_ALLPAGES) == 0)) 941 continue; 942 943 pageidx = current_offset >> PAGE_SHIFT; 944 945 /* 946 * we have yet to locate the current page (pps[lcv]). we 947 * first look for a page that is already at the current offset. 948 * if we find a page, we check to see if it is busy or 949 * released. if that is the case, then we sleep on the page 950 * until it is no longer busy or released and repeat the lookup. 951 * if the page we found is neither busy nor released, then we 952 * busy it (so we own it) and plug it into pps[lcv]. this 953 * 'break's the following while loop and indicates we are 954 * ready to move on to the next page in the "lcv" loop above. 955 * 956 * if we exit the while loop with pps[lcv] still set to NULL, 957 * then it means that we allocated a new busy/fake/clean page 958 * ptmp in the object and we need to do I/O to fill in the data. 959 */ 960 961 /* top of "pps" while loop */ 962 while (pps[lcv] == NULL) { 963 /* look for a resident page */ 964 ptmp = uvm_pagelookup(uobj, current_offset); 965 966 /* not resident? allocate one now (if we can) */ 967 if (ptmp == NULL) { 968 969 ptmp = uvm_pagealloc(uobj, current_offset, 970 NULL, 0); 971 972 /* out of RAM? */ 973 if (ptmp == NULL) { 974 simple_unlock(&uobj->vmobjlock); 975 uvm_wait("uao_getpage"); 976 simple_lock(&uobj->vmobjlock); 977 /* goto top of pps while loop */ 978 continue; 979 } 980 981 /* 982 * safe with PQ's unlocked: because we just 983 * alloc'd the page 984 */ 985 atomic_setbits_int(&ptmp->pg_flags, PQ_AOBJ); 986 987 /* 988 * got new page ready for I/O. break pps while 989 * loop. pps[lcv] is still NULL. 990 */ 991 break; 992 } 993 994 /* page is there, see if we need to wait on it */ 995 if ((ptmp->pg_flags & PG_BUSY) != 0) { 996 atomic_setbits_int(&ptmp->pg_flags, PG_WANTED); 997 UVM_UNLOCK_AND_WAIT(ptmp, &uobj->vmobjlock, 998 FALSE, "uao_get", 0); 999 simple_lock(&uobj->vmobjlock); 1000 continue; /* goto top of pps while loop */ 1001 } 1002 1003 /* 1004 * if we get here then the page has become resident and 1005 * unbusy between steps 1 and 2. we busy it now (so we 1006 * own it) and set pps[lcv] (so that we exit the while 1007 * loop). 1008 */ 1009 /* we own it, caller must un-busy */ 1010 atomic_setbits_int(&ptmp->pg_flags, PG_BUSY); 1011 UVM_PAGE_OWN(ptmp, "uao_get2"); 1012 pps[lcv] = ptmp; 1013 } 1014 1015 /* 1016 * if we own the valid page at the correct offset, pps[lcv] will 1017 * point to it. nothing more to do except go to the next page. 1018 */ 1019 if (pps[lcv]) 1020 continue; /* next lcv */ 1021 1022 /* 1023 * we have a "fake/busy/clean" page that we just allocated. 1024 * do the needed "i/o", either reading from swap or zeroing. 1025 */ 1026 swslot = uao_find_swslot(aobj, pageidx); 1027 1028 /* 1029 * just zero the page if there's nothing in swap. 1030 */ 1031 if (swslot == 0) { 1032 /* 1033 * page hasn't existed before, just zero it. 1034 */ 1035 uvm_pagezero(ptmp); 1036 } else { 1037 /* 1038 * page in the swapped-out page. 1039 * unlock object for i/o, relock when done. 1040 */ 1041 simple_unlock(&uobj->vmobjlock); 1042 rv = uvm_swap_get(ptmp, swslot, PGO_SYNCIO); 1043 simple_lock(&uobj->vmobjlock); 1044 1045 /* 1046 * I/O done. check for errors. 1047 */ 1048 if (rv != VM_PAGER_OK) 1049 { 1050 if (ptmp->pg_flags & PG_WANTED) 1051 wakeup(ptmp); 1052 1053 /* 1054 * remove the swap slot from the aobj 1055 * and mark the aobj as having no real slot. 1056 * don't free the swap slot, thus preventing 1057 * it from being used again. 1058 */ 1059 swslot = uao_set_swslot(&aobj->u_obj, pageidx, 1060 SWSLOT_BAD); 1061 uvm_swap_markbad(swslot, 1); 1062 1063 atomic_clearbits_int(&ptmp->pg_flags, 1064 PG_WANTED|PG_BUSY); 1065 UVM_PAGE_OWN(ptmp, NULL); 1066 uvm_lock_pageq(); 1067 uvm_pagefree(ptmp); 1068 uvm_unlock_pageq(); 1069 1070 simple_unlock(&uobj->vmobjlock); 1071 return (rv); 1072 } 1073 } 1074 1075 /* 1076 * we got the page! clear the fake flag (indicates valid 1077 * data now in page) and plug into our result array. note 1078 * that page is still busy. 1079 * 1080 * it is the callers job to: 1081 * => check if the page is released 1082 * => unbusy the page 1083 * => activate the page 1084 */ 1085 1086 /* data is valid ... */ 1087 atomic_clearbits_int(&ptmp->pg_flags, PG_FAKE); 1088 pmap_clear_modify(ptmp); /* ... and clean */ 1089 pps[lcv] = ptmp; 1090 1091 } /* lcv loop */ 1092 1093 /* 1094 * finally, unlock object and return. 1095 */ 1096 1097 simple_unlock(&uobj->vmobjlock); 1098 return(VM_PAGER_OK); 1099 } 1100 1101 /* 1102 * uao_dropswap: release any swap resources from this aobj page. 1103 * 1104 * => aobj must be locked or have a reference count of 0. 1105 */ 1106 1107 int 1108 uao_dropswap(struct uvm_object *uobj, int pageidx) 1109 { 1110 int slot; 1111 1112 slot = uao_set_swslot(uobj, pageidx, 0); 1113 if (slot) { 1114 uvm_swap_free(slot, 1); 1115 } 1116 return (slot); 1117 } 1118 1119 1120 /* 1121 * page in every page in every aobj that is paged-out to a range of swslots. 1122 * 1123 * => nothing should be locked. 1124 * => returns TRUE if pagein was aborted due to lack of memory. 1125 */ 1126 boolean_t 1127 uao_swap_off(int startslot, int endslot) 1128 { 1129 struct uvm_aobj *aobj, *nextaobj, *prevaobj = NULL; 1130 1131 /* 1132 * walk the list of all aobjs. 1133 */ 1134 1135 restart: 1136 mtx_enter(&uao_list_lock); 1137 1138 for (aobj = LIST_FIRST(&uao_list); 1139 aobj != NULL; 1140 aobj = nextaobj) { 1141 boolean_t rv; 1142 1143 /* 1144 * try to get the object lock, 1145 * start all over if we fail. 1146 * most of the time we'll get the aobj lock, 1147 * so this should be a rare case. 1148 */ 1149 if (!simple_lock_try(&aobj->u_obj.vmobjlock)) { 1150 mtx_leave(&uao_list_lock); 1151 if (prevaobj) { 1152 uao_detach_locked(&prevaobj->u_obj); 1153 prevaobj = NULL; 1154 } 1155 goto restart; 1156 } 1157 1158 /* 1159 * add a ref to the aobj so it doesn't disappear 1160 * while we're working. 1161 */ 1162 uao_reference_locked(&aobj->u_obj); 1163 1164 /* 1165 * now it's safe to unlock the uao list. 1166 * note that lock interleaving is alright with IPL_NONE mutexes. 1167 */ 1168 mtx_leave(&uao_list_lock); 1169 1170 if (prevaobj) { 1171 uao_detach_locked(&prevaobj->u_obj); 1172 prevaobj = NULL; 1173 } 1174 1175 /* 1176 * page in any pages in the swslot range. 1177 * if there's an error, abort and return the error. 1178 */ 1179 rv = uao_pagein(aobj, startslot, endslot); 1180 if (rv) { 1181 uao_detach_locked(&aobj->u_obj); 1182 return rv; 1183 } 1184 1185 /* 1186 * we're done with this aobj. 1187 * relock the list and drop our ref on the aobj. 1188 */ 1189 mtx_enter(&uao_list_lock); 1190 nextaobj = LIST_NEXT(aobj, u_list); 1191 /* 1192 * prevaobj means that we have an object that we need 1193 * to drop a reference for. We can't just drop it now with 1194 * the list locked since that could cause lock recursion in 1195 * the case where we reduce the refcount to 0. It will be 1196 * released the next time we drop the list lock. 1197 */ 1198 prevaobj = aobj; 1199 } 1200 1201 /* 1202 * done with traversal, unlock the list 1203 */ 1204 mtx_leave(&uao_list_lock); 1205 if (prevaobj) { 1206 uao_detach_locked(&prevaobj->u_obj); 1207 } 1208 return FALSE; 1209 } 1210 1211 1212 /* 1213 * page in any pages from aobj in the given range. 1214 * 1215 * => aobj must be locked and is returned locked. 1216 * => returns TRUE if pagein was aborted due to lack of memory. 1217 */ 1218 static boolean_t 1219 uao_pagein(struct uvm_aobj *aobj, int startslot, int endslot) 1220 { 1221 boolean_t rv; 1222 1223 if (UAO_USES_SWHASH(aobj)) { 1224 struct uao_swhash_elt *elt; 1225 int bucket; 1226 1227 restart: 1228 for (bucket = aobj->u_swhashmask; bucket >= 0; bucket--) { 1229 for (elt = LIST_FIRST(&aobj->u_swhash[bucket]); 1230 elt != NULL; 1231 elt = LIST_NEXT(elt, list)) { 1232 int i; 1233 1234 for (i = 0; i < UAO_SWHASH_CLUSTER_SIZE; i++) { 1235 int slot = elt->slots[i]; 1236 1237 /* 1238 * if the slot isn't in range, skip it. 1239 */ 1240 if (slot < startslot || 1241 slot >= endslot) { 1242 continue; 1243 } 1244 1245 /* 1246 * process the page, 1247 * the start over on this object 1248 * since the swhash elt 1249 * may have been freed. 1250 */ 1251 rv = uao_pagein_page(aobj, 1252 UAO_SWHASH_ELT_PAGEIDX_BASE(elt) + i); 1253 if (rv) { 1254 return rv; 1255 } 1256 goto restart; 1257 } 1258 } 1259 } 1260 } else { 1261 int i; 1262 1263 for (i = 0; i < aobj->u_pages; i++) { 1264 int slot = aobj->u_swslots[i]; 1265 1266 /* 1267 * if the slot isn't in range, skip it 1268 */ 1269 if (slot < startslot || slot >= endslot) { 1270 continue; 1271 } 1272 1273 /* 1274 * process the page. 1275 */ 1276 rv = uao_pagein_page(aobj, i); 1277 if (rv) { 1278 return rv; 1279 } 1280 } 1281 } 1282 1283 return FALSE; 1284 } 1285 1286 /* 1287 * page in a page from an aobj. used for swap_off. 1288 * returns TRUE if pagein was aborted due to lack of memory. 1289 * 1290 * => aobj must be locked and is returned locked. 1291 */ 1292 static boolean_t 1293 uao_pagein_page(struct uvm_aobj *aobj, int pageidx) 1294 { 1295 struct vm_page *pg; 1296 int rv, slot, npages; 1297 1298 pg = NULL; 1299 npages = 1; 1300 /* locked: aobj */ 1301 rv = uao_get(&aobj->u_obj, pageidx << PAGE_SHIFT, 1302 &pg, &npages, 0, VM_PROT_READ|VM_PROT_WRITE, 0, 0); 1303 /* unlocked: aobj */ 1304 1305 /* 1306 * relock and finish up. 1307 */ 1308 simple_lock(&aobj->u_obj.vmobjlock); 1309 1310 switch (rv) { 1311 case VM_PAGER_OK: 1312 break; 1313 1314 case VM_PAGER_ERROR: 1315 case VM_PAGER_REFAULT: 1316 /* 1317 * nothing more to do on errors. 1318 * VM_PAGER_REFAULT can only mean that the anon was freed, 1319 * so again there's nothing to do. 1320 */ 1321 return FALSE; 1322 1323 } 1324 1325 /* 1326 * ok, we've got the page now. 1327 * mark it as dirty, clear its swslot and un-busy it. 1328 */ 1329 slot = uao_set_swslot(&aobj->u_obj, pageidx, 0); 1330 uvm_swap_free(slot, 1); 1331 atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_CLEAN|PG_FAKE); 1332 UVM_PAGE_OWN(pg, NULL); 1333 1334 /* 1335 * deactivate the page (to put it on a page queue). 1336 */ 1337 pmap_clear_reference(pg); 1338 #ifndef UBC 1339 pmap_page_protect(pg, VM_PROT_NONE); 1340 #endif 1341 uvm_lock_pageq(); 1342 uvm_pagedeactivate(pg); 1343 uvm_unlock_pageq(); 1344 1345 return FALSE; 1346 } 1347