1 /* $NetBSD: uvm_aobj.c,v 1.46 2001/09/15 20:36:45 chs Exp $ */ 2 3 /* 4 * Copyright (c) 1998 Chuck Silvers, Charles D. Cranor and 5 * 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 * from: Id: uvm_aobj.c,v 1.1.2.5 1998/02/06 05:14:38 chs Exp 35 */ 36 /* 37 * uvm_aobj.c: anonymous memory uvm_object pager 38 * 39 * author: Chuck Silvers <chuq@chuq.com> 40 * started: Jan-1998 41 * 42 * - design mostly from Chuck Cranor 43 */ 44 45 #include "opt_uvmhist.h" 46 47 #include <sys/param.h> 48 #include <sys/systm.h> 49 #include <sys/proc.h> 50 #include <sys/malloc.h> 51 #include <sys/kernel.h> 52 #include <sys/pool.h> 53 #include <sys/kernel.h> 54 55 #include <uvm/uvm.h> 56 57 /* 58 * an aobj manages anonymous-memory backed uvm_objects. in addition 59 * to keeping the list of resident pages, it also keeps a list of 60 * allocated swap blocks. depending on the size of the aobj this list 61 * of allocated swap blocks is either stored in an array (small objects) 62 * or in a hash table (large objects). 63 */ 64 65 /* 66 * local structures 67 */ 68 69 /* 70 * for hash tables, we break the address space of the aobj into blocks 71 * of UAO_SWHASH_CLUSTER_SIZE pages. we require the cluster size to 72 * be a power of two. 73 */ 74 75 #define UAO_SWHASH_CLUSTER_SHIFT 4 76 #define UAO_SWHASH_CLUSTER_SIZE (1 << UAO_SWHASH_CLUSTER_SHIFT) 77 78 /* get the "tag" for this page index */ 79 #define UAO_SWHASH_ELT_TAG(PAGEIDX) \ 80 ((PAGEIDX) >> UAO_SWHASH_CLUSTER_SHIFT) 81 82 /* given an ELT and a page index, find the swap slot */ 83 #define UAO_SWHASH_ELT_PAGESLOT(ELT, PAGEIDX) \ 84 ((ELT)->slots[(PAGEIDX) & (UAO_SWHASH_CLUSTER_SIZE - 1)]) 85 86 /* given an ELT, return its pageidx base */ 87 #define UAO_SWHASH_ELT_PAGEIDX_BASE(ELT) \ 88 ((ELT)->tag << UAO_SWHASH_CLUSTER_SHIFT) 89 90 /* 91 * the swhash hash function 92 */ 93 94 #define UAO_SWHASH_HASH(AOBJ, PAGEIDX) \ 95 (&(AOBJ)->u_swhash[(((PAGEIDX) >> UAO_SWHASH_CLUSTER_SHIFT) \ 96 & (AOBJ)->u_swhashmask)]) 97 98 /* 99 * the swhash threshhold determines if we will use an array or a 100 * hash table to store the list of allocated swap blocks. 101 */ 102 103 #define UAO_SWHASH_THRESHOLD (UAO_SWHASH_CLUSTER_SIZE * 4) 104 #define UAO_USES_SWHASH(AOBJ) \ 105 ((AOBJ)->u_pages > UAO_SWHASH_THRESHOLD) /* use hash? */ 106 107 /* 108 * the number of buckets in a swhash, with an upper bound 109 */ 110 111 #define UAO_SWHASH_MAXBUCKETS 256 112 #define UAO_SWHASH_BUCKETS(AOBJ) \ 113 (MIN((AOBJ)->u_pages >> UAO_SWHASH_CLUSTER_SHIFT, \ 114 UAO_SWHASH_MAXBUCKETS)) 115 116 117 /* 118 * uao_swhash_elt: when a hash table is being used, this structure defines 119 * the format of an entry in the bucket list. 120 */ 121 122 struct uao_swhash_elt { 123 LIST_ENTRY(uao_swhash_elt) list; /* the hash list */ 124 voff_t tag; /* our 'tag' */ 125 int count; /* our number of active slots */ 126 int slots[UAO_SWHASH_CLUSTER_SIZE]; /* the slots */ 127 }; 128 129 /* 130 * uao_swhash: the swap hash table structure 131 */ 132 133 LIST_HEAD(uao_swhash, uao_swhash_elt); 134 135 /* 136 * uao_swhash_elt_pool: pool of uao_swhash_elt structures 137 */ 138 139 struct pool uao_swhash_elt_pool; 140 141 /* 142 * uvm_aobj: the actual anon-backed uvm_object 143 * 144 * => the uvm_object is at the top of the structure, this allows 145 * (struct uvm_aobj *) == (struct uvm_object *) 146 * => only one of u_swslots and u_swhash is used in any given aobj 147 */ 148 149 struct uvm_aobj { 150 struct uvm_object u_obj; /* has: lock, pgops, memq, #pages, #refs */ 151 int u_pages; /* number of pages in entire object */ 152 int u_flags; /* the flags (see uvm_aobj.h) */ 153 int *u_swslots; /* array of offset->swapslot mappings */ 154 /* 155 * hashtable of offset->swapslot mappings 156 * (u_swhash is an array of bucket heads) 157 */ 158 struct uao_swhash *u_swhash; 159 u_long u_swhashmask; /* mask for hashtable */ 160 LIST_ENTRY(uvm_aobj) u_list; /* global list of aobjs */ 161 }; 162 163 /* 164 * uvm_aobj_pool: pool of uvm_aobj structures 165 */ 166 167 struct pool uvm_aobj_pool; 168 169 /* 170 * local functions 171 */ 172 173 static struct uao_swhash_elt *uao_find_swhash_elt 174 __P((struct uvm_aobj *, int, boolean_t)); 175 176 static void uao_free __P((struct uvm_aobj *)); 177 static int uao_get __P((struct uvm_object *, voff_t, struct vm_page **, 178 int *, int, vm_prot_t, int, int)); 179 static boolean_t uao_put __P((struct uvm_object *, voff_t, voff_t, int)); 180 static boolean_t uao_pagein __P((struct uvm_aobj *, int, int)); 181 static boolean_t uao_pagein_page __P((struct uvm_aobj *, int)); 182 183 /* 184 * aobj_pager 185 * 186 * note that some functions (e.g. put) are handled elsewhere 187 */ 188 189 struct uvm_pagerops aobj_pager = { 190 NULL, /* init */ 191 uao_reference, /* reference */ 192 uao_detach, /* detach */ 193 NULL, /* fault */ 194 uao_get, /* get */ 195 uao_put, /* flush */ 196 }; 197 198 /* 199 * uao_list: global list of active aobjs, locked by uao_list_lock 200 */ 201 202 static LIST_HEAD(aobjlist, uvm_aobj) uao_list; 203 static struct simplelock uao_list_lock; 204 205 /* 206 * functions 207 */ 208 209 /* 210 * hash table/array related functions 211 */ 212 213 /* 214 * uao_find_swhash_elt: find (or create) a hash table entry for a page 215 * offset. 216 * 217 * => the object should be locked by the caller 218 */ 219 220 static struct uao_swhash_elt * 221 uao_find_swhash_elt(aobj, pageidx, create) 222 struct uvm_aobj *aobj; 223 int pageidx; 224 boolean_t create; 225 { 226 struct uao_swhash *swhash; 227 struct uao_swhash_elt *elt; 228 voff_t page_tag; 229 230 swhash = UAO_SWHASH_HASH(aobj, pageidx); 231 page_tag = UAO_SWHASH_ELT_TAG(pageidx); 232 233 /* 234 * now search the bucket for the requested tag 235 */ 236 237 LIST_FOREACH(elt, swhash, list) { 238 if (elt->tag == page_tag) { 239 return elt; 240 } 241 } 242 if (!create) { 243 return NULL; 244 } 245 246 /* 247 * allocate a new entry for the bucket and init/insert it in 248 */ 249 250 elt = pool_get(&uao_swhash_elt_pool, PR_NOWAIT); 251 if (elt == NULL) { 252 return NULL; 253 } 254 LIST_INSERT_HEAD(swhash, elt, list); 255 elt->tag = page_tag; 256 elt->count = 0; 257 memset(elt->slots, 0, sizeof(elt->slots)); 258 return elt; 259 } 260 261 /* 262 * uao_find_swslot: find the swap slot number for an aobj/pageidx 263 * 264 * => object must be locked by caller 265 */ 266 267 int 268 uao_find_swslot(uobj, pageidx) 269 struct uvm_object *uobj; 270 int pageidx; 271 { 272 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 273 struct uao_swhash_elt *elt; 274 275 /* 276 * if noswap flag is set, then we never return a slot 277 */ 278 279 if (aobj->u_flags & UAO_FLAG_NOSWAP) 280 return(0); 281 282 /* 283 * if hashing, look in hash table. 284 */ 285 286 if (UAO_USES_SWHASH(aobj)) { 287 elt = uao_find_swhash_elt(aobj, pageidx, FALSE); 288 if (elt) 289 return(UAO_SWHASH_ELT_PAGESLOT(elt, pageidx)); 290 else 291 return(0); 292 } 293 294 /* 295 * otherwise, look in the array 296 */ 297 298 return(aobj->u_swslots[pageidx]); 299 } 300 301 /* 302 * uao_set_swslot: set the swap slot for a page in an aobj. 303 * 304 * => setting a slot to zero frees the slot 305 * => object must be locked by caller 306 * => we return the old slot number, or -1 if we failed to allocate 307 * memory to record the new slot number 308 */ 309 310 int 311 uao_set_swslot(uobj, pageidx, slot) 312 struct uvm_object *uobj; 313 int pageidx, slot; 314 { 315 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 316 struct uao_swhash_elt *elt; 317 int oldslot; 318 UVMHIST_FUNC("uao_set_swslot"); UVMHIST_CALLED(pdhist); 319 UVMHIST_LOG(pdhist, "aobj %p pageidx %d slot %d", 320 aobj, pageidx, slot, 0); 321 322 /* 323 * if noswap flag is set, then we can't set a non-zero slot. 324 */ 325 326 if (aobj->u_flags & UAO_FLAG_NOSWAP) { 327 if (slot == 0) 328 return(0); 329 330 printf("uao_set_swslot: uobj = %p\n", uobj); 331 panic("uao_set_swslot: NOSWAP object"); 332 } 333 334 /* 335 * are we using a hash table? if so, add it in the hash. 336 */ 337 338 if (UAO_USES_SWHASH(aobj)) { 339 340 /* 341 * Avoid allocating an entry just to free it again if 342 * the page had not swap slot in the first place, and 343 * we are freeing. 344 */ 345 346 elt = uao_find_swhash_elt(aobj, pageidx, slot != 0); 347 if (elt == NULL) { 348 return slot ? -1 : 0; 349 } 350 351 oldslot = UAO_SWHASH_ELT_PAGESLOT(elt, pageidx); 352 UAO_SWHASH_ELT_PAGESLOT(elt, pageidx) = slot; 353 354 /* 355 * now adjust the elt's reference counter and free it if we've 356 * dropped it to zero. 357 */ 358 359 if (slot) { 360 if (oldslot == 0) 361 elt->count++; 362 } else { 363 if (oldslot) 364 elt->count--; 365 366 if (elt->count == 0) { 367 LIST_REMOVE(elt, list); 368 pool_put(&uao_swhash_elt_pool, elt); 369 } 370 } 371 } else { 372 /* we are using an array */ 373 oldslot = aobj->u_swslots[pageidx]; 374 aobj->u_swslots[pageidx] = slot; 375 } 376 return (oldslot); 377 } 378 379 /* 380 * end of hash/array functions 381 */ 382 383 /* 384 * uao_free: free all resources held by an aobj, and then free the aobj 385 * 386 * => the aobj should be dead 387 */ 388 389 static void 390 uao_free(aobj) 391 struct uvm_aobj *aobj; 392 { 393 int swpgonlydelta = 0; 394 395 simple_unlock(&aobj->u_obj.vmobjlock); 396 if (UAO_USES_SWHASH(aobj)) { 397 int i, hashbuckets = aobj->u_swhashmask + 1; 398 399 /* 400 * free the swslots from each hash bucket, 401 * then the hash bucket, and finally the hash table itself. 402 */ 403 404 for (i = 0; i < hashbuckets; i++) { 405 struct uao_swhash_elt *elt, *next; 406 407 for (elt = LIST_FIRST(&aobj->u_swhash[i]); 408 elt != NULL; 409 elt = next) { 410 int j; 411 412 for (j = 0; j < UAO_SWHASH_CLUSTER_SIZE; j++) { 413 int slot = elt->slots[j]; 414 415 if (slot == 0) { 416 continue; 417 } 418 uvm_swap_free(slot, 1); 419 swpgonlydelta++; 420 } 421 422 next = LIST_NEXT(elt, list); 423 pool_put(&uao_swhash_elt_pool, elt); 424 } 425 } 426 free(aobj->u_swhash, M_UVMAOBJ); 427 } else { 428 int i; 429 430 /* 431 * free the array 432 */ 433 434 for (i = 0; i < aobj->u_pages; i++) { 435 int slot = aobj->u_swslots[i]; 436 437 if (slot) { 438 uvm_swap_free(slot, 1); 439 swpgonlydelta++; 440 } 441 } 442 free(aobj->u_swslots, M_UVMAOBJ); 443 } 444 445 /* 446 * finally free the aobj itself 447 */ 448 449 pool_put(&uvm_aobj_pool, aobj); 450 451 /* 452 * adjust the counter of pages only in swap for all 453 * the swap slots we've freed. 454 */ 455 456 simple_lock(&uvm.swap_data_lock); 457 uvmexp.swpgonly -= swpgonlydelta; 458 simple_unlock(&uvm.swap_data_lock); 459 } 460 461 /* 462 * pager functions 463 */ 464 465 /* 466 * uao_create: create an aobj of the given size and return its uvm_object. 467 * 468 * => for normal use, flags are always zero 469 * => for the kernel object, the flags are: 470 * UAO_FLAG_KERNOBJ - allocate the kernel object (can only happen once) 471 * UAO_FLAG_KERNSWAP - enable swapping of kernel object (" ") 472 */ 473 474 struct uvm_object * 475 uao_create(size, flags) 476 vsize_t size; 477 int flags; 478 { 479 static struct uvm_aobj kernel_object_store; 480 static int kobj_alloced = 0; 481 int pages = round_page(size) >> PAGE_SHIFT; 482 struct uvm_aobj *aobj; 483 484 /* 485 * malloc a new aobj unless we are asked for the kernel object 486 */ 487 488 if (flags & UAO_FLAG_KERNOBJ) { 489 KASSERT(!kobj_alloced); 490 aobj = &kernel_object_store; 491 aobj->u_pages = pages; 492 aobj->u_flags = UAO_FLAG_NOSWAP; 493 aobj->u_obj.uo_refs = UVM_OBJ_KERN; 494 kobj_alloced = UAO_FLAG_KERNOBJ; 495 } else if (flags & UAO_FLAG_KERNSWAP) { 496 KASSERT(kobj_alloced == UAO_FLAG_KERNOBJ); 497 aobj = &kernel_object_store; 498 kobj_alloced = UAO_FLAG_KERNSWAP; 499 } else { 500 aobj = pool_get(&uvm_aobj_pool, PR_WAITOK); 501 aobj->u_pages = pages; 502 aobj->u_flags = 0; 503 aobj->u_obj.uo_refs = 1; 504 } 505 506 /* 507 * allocate hash/array if necessary 508 * 509 * note: in the KERNSWAP case no need to worry about locking since 510 * we are still booting we should be the only thread around. 511 */ 512 513 if (flags == 0 || (flags & UAO_FLAG_KERNSWAP) != 0) { 514 int mflags = (flags & UAO_FLAG_KERNSWAP) != 0 ? 515 M_NOWAIT : M_WAITOK; 516 517 /* allocate hash table or array depending on object size */ 518 if (UAO_USES_SWHASH(aobj)) { 519 aobj->u_swhash = hashinit(UAO_SWHASH_BUCKETS(aobj), 520 HASH_LIST, M_UVMAOBJ, mflags, &aobj->u_swhashmask); 521 if (aobj->u_swhash == NULL) 522 panic("uao_create: hashinit swhash failed"); 523 } else { 524 aobj->u_swslots = malloc(pages * sizeof(int), 525 M_UVMAOBJ, mflags); 526 if (aobj->u_swslots == NULL) 527 panic("uao_create: malloc swslots failed"); 528 memset(aobj->u_swslots, 0, pages * sizeof(int)); 529 } 530 531 if (flags) { 532 aobj->u_flags &= ~UAO_FLAG_NOSWAP; /* clear noswap */ 533 return(&aobj->u_obj); 534 } 535 } 536 537 /* 538 * init aobj fields 539 */ 540 541 simple_lock_init(&aobj->u_obj.vmobjlock); 542 aobj->u_obj.pgops = &aobj_pager; 543 TAILQ_INIT(&aobj->u_obj.memq); 544 aobj->u_obj.uo_npages = 0; 545 546 /* 547 * now that aobj is ready, add it to the global list 548 */ 549 550 simple_lock(&uao_list_lock); 551 LIST_INSERT_HEAD(&uao_list, aobj, u_list); 552 simple_unlock(&uao_list_lock); 553 return(&aobj->u_obj); 554 } 555 556 557 558 /* 559 * uao_init: set up aobj pager subsystem 560 * 561 * => called at boot time from uvm_pager_init() 562 */ 563 564 void 565 uao_init(void) 566 { 567 static int uao_initialized; 568 569 if (uao_initialized) 570 return; 571 uao_initialized = TRUE; 572 LIST_INIT(&uao_list); 573 simple_lock_init(&uao_list_lock); 574 575 /* 576 * NOTE: Pages fror this pool must not come from a pageable 577 * kernel map! 578 */ 579 580 pool_init(&uao_swhash_elt_pool, sizeof(struct uao_swhash_elt), 581 0, 0, 0, "uaoeltpl", 0, NULL, NULL, M_UVMAOBJ); 582 pool_init(&uvm_aobj_pool, sizeof(struct uvm_aobj), 0, 0, 0, 583 "aobjpl", 0, 584 pool_page_alloc_nointr, pool_page_free_nointr, M_UVMAOBJ); 585 } 586 587 /* 588 * uao_reference: add a ref to an aobj 589 * 590 * => aobj must be unlocked 591 * => just lock it and call the locked version 592 */ 593 594 void 595 uao_reference(uobj) 596 struct uvm_object *uobj; 597 { 598 simple_lock(&uobj->vmobjlock); 599 uao_reference_locked(uobj); 600 simple_unlock(&uobj->vmobjlock); 601 } 602 603 /* 604 * uao_reference_locked: add a ref to an aobj that is already locked 605 * 606 * => aobj must be locked 607 * this needs to be separate from the normal routine 608 * since sometimes we need to add a reference to an aobj when 609 * it's already locked. 610 */ 611 612 void 613 uao_reference_locked(uobj) 614 struct uvm_object *uobj; 615 { 616 UVMHIST_FUNC("uao_reference"); UVMHIST_CALLED(maphist); 617 618 /* 619 * kernel_object already has plenty of references, leave it alone. 620 */ 621 622 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) 623 return; 624 625 uobj->uo_refs++; 626 UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)", 627 uobj, uobj->uo_refs,0,0); 628 } 629 630 /* 631 * uao_detach: drop a reference to an aobj 632 * 633 * => aobj must be unlocked 634 * => just lock it and call the locked version 635 */ 636 637 void 638 uao_detach(uobj) 639 struct uvm_object *uobj; 640 { 641 simple_lock(&uobj->vmobjlock); 642 uao_detach_locked(uobj); 643 } 644 645 /* 646 * uao_detach_locked: drop a reference to an aobj 647 * 648 * => aobj must be locked, and is unlocked (or freed) upon return. 649 * this needs to be separate from the normal routine 650 * since sometimes we need to detach from an aobj when 651 * it's already locked. 652 */ 653 654 void 655 uao_detach_locked(uobj) 656 struct uvm_object *uobj; 657 { 658 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 659 struct vm_page *pg; 660 UVMHIST_FUNC("uao_detach"); UVMHIST_CALLED(maphist); 661 662 /* 663 * detaching from kernel_object is a noop. 664 */ 665 666 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) { 667 simple_unlock(&uobj->vmobjlock); 668 return; 669 } 670 671 UVMHIST_LOG(maphist," (uobj=0x%x) ref=%d", uobj,uobj->uo_refs,0,0); 672 uobj->uo_refs--; 673 if (uobj->uo_refs) { 674 simple_unlock(&uobj->vmobjlock); 675 UVMHIST_LOG(maphist, "<- done (rc>0)", 0,0,0,0); 676 return; 677 } 678 679 /* 680 * remove the aobj from the global list. 681 */ 682 683 simple_lock(&uao_list_lock); 684 LIST_REMOVE(aobj, u_list); 685 simple_unlock(&uao_list_lock); 686 687 /* 688 * free all the pages left in the aobj. for each page, 689 * when the page is no longer busy (and thus after any disk i/o that 690 * it's involved in is complete), release any swap resources and 691 * free the page itself. 692 */ 693 694 uvm_lock_pageq(); 695 while ((pg = TAILQ_FIRST(&uobj->memq)) != NULL) { 696 pmap_page_protect(pg, VM_PROT_NONE); 697 if (pg->flags & PG_BUSY) { 698 pg->flags |= PG_WANTED; 699 uvm_unlock_pageq(); 700 UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, FALSE, 701 "uao_det", 0); 702 simple_lock(&uobj->vmobjlock); 703 uvm_lock_pageq(); 704 continue; 705 } 706 uao_dropswap(&aobj->u_obj, pg->offset >> PAGE_SHIFT); 707 uvm_pagefree(pg); 708 } 709 uvm_unlock_pageq(); 710 711 /* 712 * finally, free the aobj itself. 713 */ 714 715 uao_free(aobj); 716 } 717 718 /* 719 * uao_put: flush pages out of a uvm object 720 * 721 * => object should be locked by caller. we may _unlock_ the object 722 * if (and only if) we need to clean a page (PGO_CLEANIT). 723 * XXXJRT Currently, however, we don't. In the case of cleaning 724 * XXXJRT a page, we simply just deactivate it. Should probably 725 * XXXJRT handle this better, in the future (although "flushing" 726 * XXXJRT anonymous memory isn't terribly important). 727 * => if PGO_CLEANIT is not set, then we will neither unlock the object 728 * or block. 729 * => if PGO_ALLPAGE is set, then all pages in the object are valid targets 730 * for flushing. 731 * => NOTE: we rely on the fact that the object's memq is a TAILQ and 732 * that new pages are inserted on the tail end of the list. thus, 733 * we can make a complete pass through the object in one go by starting 734 * at the head and working towards the tail (new pages are put in 735 * front of us). 736 * => NOTE: we are allowed to lock the page queues, so the caller 737 * must not be holding the lock on them [e.g. pagedaemon had 738 * better not call us with the queues locked] 739 * => we return TRUE unless we encountered some sort of I/O error 740 * XXXJRT currently never happens, as we never directly initiate 741 * XXXJRT I/O 742 * 743 * note on page traversal: 744 * we can traverse the pages in an object either by going down the 745 * linked list in "uobj->memq", or we can go over the address range 746 * by page doing hash table lookups for each address. depending 747 * on how many pages are in the object it may be cheaper to do one 748 * or the other. we set "by_list" to true if we are using memq. 749 * if the cost of a hash lookup was equal to the cost of the list 750 * traversal we could compare the number of pages in the start->stop 751 * range to the total number of pages in the object. however, it 752 * seems that a hash table lookup is more expensive than the linked 753 * list traversal, so we multiply the number of pages in the 754 * start->stop range by a penalty which we define below. 755 */ 756 757 int 758 uao_put(uobj, start, stop, flags) 759 struct uvm_object *uobj; 760 voff_t start, stop; 761 int flags; 762 { 763 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 764 struct vm_page *pg, *nextpg; 765 boolean_t by_list; 766 voff_t curoff; 767 UVMHIST_FUNC("uao_put"); UVMHIST_CALLED(maphist); 768 769 curoff = 0; 770 if (flags & PGO_ALLPAGES) { 771 start = 0; 772 stop = aobj->u_pages << PAGE_SHIFT; 773 by_list = TRUE; /* always go by the list */ 774 } else { 775 start = trunc_page(start); 776 stop = round_page(stop); 777 if (stop > (aobj->u_pages << PAGE_SHIFT)) { 778 printf("uao_flush: strange, got an out of range " 779 "flush (fixed)\n"); 780 stop = aobj->u_pages << PAGE_SHIFT; 781 } 782 by_list = (uobj->uo_npages <= 783 ((stop - start) >> PAGE_SHIFT) * UVM_PAGE_HASH_PENALTY); 784 } 785 UVMHIST_LOG(maphist, 786 " flush start=0x%lx, stop=0x%x, by_list=%d, flags=0x%x", 787 start, stop, by_list, flags); 788 789 /* 790 * Don't need to do any work here if we're not freeing 791 * or deactivating pages. 792 */ 793 794 if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) { 795 simple_unlock(&uobj->vmobjlock); 796 return 0; 797 } 798 799 /* 800 * now do it. note: we must update nextpg in the body of loop or we 801 * will get stuck. we need to use nextpg because we may free "pg" 802 * before doing the next loop. 803 */ 804 805 if (by_list) { 806 pg = TAILQ_FIRST(&uobj->memq); 807 } else { 808 curoff = start; 809 pg = uvm_pagelookup(uobj, curoff); 810 } 811 812 nextpg = NULL; 813 uvm_lock_pageq(); 814 815 /* locked: both page queues and uobj */ 816 for ( ; (by_list && pg != NULL) || 817 (!by_list && curoff < stop) ; pg = nextpg) { 818 if (by_list) { 819 nextpg = TAILQ_NEXT(pg, listq); 820 if (pg->offset < start || pg->offset >= stop) 821 continue; 822 } else { 823 curoff += PAGE_SIZE; 824 if (curoff < stop) 825 nextpg = uvm_pagelookup(uobj, curoff); 826 if (pg == NULL) 827 continue; 828 } 829 switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) { 830 831 /* 832 * XXX In these first 3 cases, we always just 833 * XXX deactivate the page. We may want to 834 * XXX handle the different cases more specifically 835 * XXX in the future. 836 */ 837 838 case PGO_CLEANIT|PGO_FREE: 839 case PGO_CLEANIT|PGO_DEACTIVATE: 840 case PGO_DEACTIVATE: 841 deactivate_it: 842 /* skip the page if it's loaned or wired */ 843 if (pg->loan_count != 0 || pg->wire_count != 0) 844 continue; 845 846 /* ...and deactivate the page. */ 847 pmap_clear_reference(pg); 848 uvm_pagedeactivate(pg); 849 continue; 850 851 case PGO_FREE: 852 853 /* 854 * If there are multiple references to 855 * the object, just deactivate the page. 856 */ 857 858 if (uobj->uo_refs > 1) 859 goto deactivate_it; 860 861 /* XXX skip the page if it's loaned or wired */ 862 if (pg->loan_count != 0 || pg->wire_count != 0) 863 continue; 864 865 /* 866 * wait if the page is busy, then free the swap slot 867 * and the page. 868 */ 869 870 pmap_page_protect(pg, VM_PROT_NONE); 871 while (pg->flags & PG_BUSY) { 872 pg->flags |= PG_WANTED; 873 uvm_unlock_pageq(); 874 UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, 0, 875 "uao_put", 0); 876 simple_lock(&uobj->vmobjlock); 877 uvm_lock_pageq(); 878 } 879 uao_dropswap(uobj, pg->offset >> PAGE_SHIFT); 880 uvm_pagefree(pg); 881 continue; 882 } 883 } 884 uvm_unlock_pageq(); 885 simple_unlock(&uobj->vmobjlock); 886 return 0; 887 } 888 889 /* 890 * uao_get: fetch me a page 891 * 892 * we have three cases: 893 * 1: page is resident -> just return the page. 894 * 2: page is zero-fill -> allocate a new page and zero it. 895 * 3: page is swapped out -> fetch the page from swap. 896 * 897 * cases 1 and 2 can be handled with PGO_LOCKED, case 3 cannot. 898 * so, if the "center" page hits case 3 (or any page, with PGO_ALLPAGES), 899 * then we will need to return EBUSY. 900 * 901 * => prefer map unlocked (not required) 902 * => object must be locked! we will _unlock_ it before starting any I/O. 903 * => flags: PGO_ALLPAGES: get all of the pages 904 * PGO_LOCKED: fault data structures are locked 905 * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx] 906 * => NOTE: caller must check for released pages!! 907 */ 908 909 static int 910 uao_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags) 911 struct uvm_object *uobj; 912 voff_t offset; 913 struct vm_page **pps; 914 int *npagesp; 915 int centeridx, advice, flags; 916 vm_prot_t access_type; 917 { 918 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 919 voff_t current_offset; 920 struct vm_page *ptmp; 921 int lcv, gotpages, maxpages, swslot, error, pageidx; 922 boolean_t done; 923 UVMHIST_FUNC("uao_get"); UVMHIST_CALLED(pdhist); 924 925 UVMHIST_LOG(pdhist, "aobj=%p offset=%d, flags=%d", 926 aobj, offset, flags,0); 927 928 /* 929 * get number of pages 930 */ 931 932 maxpages = *npagesp; 933 934 /* 935 * step 1: handled the case where fault data structures are locked. 936 */ 937 938 if (flags & PGO_LOCKED) { 939 940 /* 941 * step 1a: get pages that are already resident. only do 942 * this if the data structures are locked (i.e. the first 943 * time through). 944 */ 945 946 done = TRUE; /* be optimistic */ 947 gotpages = 0; /* # of pages we got so far */ 948 for (lcv = 0, current_offset = offset ; lcv < maxpages ; 949 lcv++, current_offset += PAGE_SIZE) { 950 /* do we care about this page? if not, skip it */ 951 if (pps[lcv] == PGO_DONTCARE) 952 continue; 953 ptmp = uvm_pagelookup(uobj, current_offset); 954 955 /* 956 * if page is new, attempt to allocate the page, 957 * zero-fill'd. 958 */ 959 960 if (ptmp == NULL && uao_find_swslot(&aobj->u_obj, 961 current_offset >> PAGE_SHIFT) == 0) { 962 ptmp = uvm_pagealloc(uobj, current_offset, 963 NULL, UVM_PGA_ZERO); 964 if (ptmp) { 965 /* new page */ 966 ptmp->flags &= ~(PG_BUSY|PG_FAKE); 967 ptmp->pqflags |= PQ_AOBJ; 968 UVM_PAGE_OWN(ptmp, NULL); 969 } 970 } 971 972 /* 973 * to be useful must get a non-busy page 974 */ 975 976 if (ptmp == NULL || (ptmp->flags & PG_BUSY) != 0) { 977 if (lcv == centeridx || 978 (flags & PGO_ALLPAGES) != 0) 979 /* need to do a wait or I/O! */ 980 done = FALSE; 981 continue; 982 } 983 984 /* 985 * useful page: busy/lock it and plug it in our 986 * result array 987 */ 988 989 /* caller must un-busy this page */ 990 ptmp->flags |= PG_BUSY; 991 UVM_PAGE_OWN(ptmp, "uao_get1"); 992 pps[lcv] = ptmp; 993 gotpages++; 994 } 995 996 /* 997 * step 1b: now we've either done everything needed or we 998 * to unlock and do some waiting or I/O. 999 */ 1000 1001 UVMHIST_LOG(pdhist, "<- done (done=%d)", done, 0,0,0); 1002 *npagesp = gotpages; 1003 if (done) 1004 return 0; 1005 else 1006 return EBUSY; 1007 } 1008 1009 /* 1010 * step 2: get non-resident or busy pages. 1011 * object is locked. data structures are unlocked. 1012 */ 1013 1014 for (lcv = 0, current_offset = offset ; lcv < maxpages ; 1015 lcv++, current_offset += PAGE_SIZE) { 1016 1017 /* 1018 * - skip over pages we've already gotten or don't want 1019 * - skip over pages we don't _have_ to get 1020 */ 1021 1022 if (pps[lcv] != NULL || 1023 (lcv != centeridx && (flags & PGO_ALLPAGES) == 0)) 1024 continue; 1025 1026 pageidx = current_offset >> PAGE_SHIFT; 1027 1028 /* 1029 * we have yet to locate the current page (pps[lcv]). we 1030 * first look for a page that is already at the current offset. 1031 * if we find a page, we check to see if it is busy or 1032 * released. if that is the case, then we sleep on the page 1033 * until it is no longer busy or released and repeat the lookup. 1034 * if the page we found is neither busy nor released, then we 1035 * busy it (so we own it) and plug it into pps[lcv]. this 1036 * 'break's the following while loop and indicates we are 1037 * ready to move on to the next page in the "lcv" loop above. 1038 * 1039 * if we exit the while loop with pps[lcv] still set to NULL, 1040 * then it means that we allocated a new busy/fake/clean page 1041 * ptmp in the object and we need to do I/O to fill in the data. 1042 */ 1043 1044 /* top of "pps" while loop */ 1045 while (pps[lcv] == NULL) { 1046 /* look for a resident page */ 1047 ptmp = uvm_pagelookup(uobj, current_offset); 1048 1049 /* not resident? allocate one now (if we can) */ 1050 if (ptmp == NULL) { 1051 1052 ptmp = uvm_pagealloc(uobj, current_offset, 1053 NULL, 0); 1054 1055 /* out of RAM? */ 1056 if (ptmp == NULL) { 1057 simple_unlock(&uobj->vmobjlock); 1058 UVMHIST_LOG(pdhist, 1059 "sleeping, ptmp == NULL\n",0,0,0,0); 1060 uvm_wait("uao_getpage"); 1061 simple_lock(&uobj->vmobjlock); 1062 continue; 1063 } 1064 1065 /* 1066 * safe with PQ's unlocked: because we just 1067 * alloc'd the page 1068 */ 1069 1070 ptmp->pqflags |= PQ_AOBJ; 1071 1072 /* 1073 * got new page ready for I/O. break pps while 1074 * loop. pps[lcv] is still NULL. 1075 */ 1076 1077 break; 1078 } 1079 1080 /* page is there, see if we need to wait on it */ 1081 if ((ptmp->flags & PG_BUSY) != 0) { 1082 ptmp->flags |= PG_WANTED; 1083 UVMHIST_LOG(pdhist, 1084 "sleeping, ptmp->flags 0x%x\n", 1085 ptmp->flags,0,0,0); 1086 UVM_UNLOCK_AND_WAIT(ptmp, &uobj->vmobjlock, 1087 FALSE, "uao_get", 0); 1088 simple_lock(&uobj->vmobjlock); 1089 continue; 1090 } 1091 1092 /* 1093 * if we get here then the page has become resident and 1094 * unbusy between steps 1 and 2. we busy it now (so we 1095 * own it) and set pps[lcv] (so that we exit the while 1096 * loop). 1097 */ 1098 1099 /* we own it, caller must un-busy */ 1100 ptmp->flags |= PG_BUSY; 1101 UVM_PAGE_OWN(ptmp, "uao_get2"); 1102 pps[lcv] = ptmp; 1103 } 1104 1105 /* 1106 * if we own the valid page at the correct offset, pps[lcv] will 1107 * point to it. nothing more to do except go to the next page. 1108 */ 1109 1110 if (pps[lcv]) 1111 continue; /* next lcv */ 1112 1113 /* 1114 * we have a "fake/busy/clean" page that we just allocated. 1115 * do the needed "i/o", either reading from swap or zeroing. 1116 */ 1117 1118 swslot = uao_find_swslot(&aobj->u_obj, pageidx); 1119 1120 /* 1121 * just zero the page if there's nothing in swap. 1122 */ 1123 1124 if (swslot == 0) { 1125 1126 /* 1127 * page hasn't existed before, just zero it. 1128 */ 1129 1130 uvm_pagezero(ptmp); 1131 } else { 1132 UVMHIST_LOG(pdhist, "pagein from swslot %d", 1133 swslot, 0,0,0); 1134 1135 /* 1136 * page in the swapped-out page. 1137 * unlock object for i/o, relock when done. 1138 */ 1139 1140 simple_unlock(&uobj->vmobjlock); 1141 error = uvm_swap_get(ptmp, swslot, PGO_SYNCIO); 1142 simple_lock(&uobj->vmobjlock); 1143 1144 /* 1145 * I/O done. check for errors. 1146 */ 1147 1148 if (error != 0) { 1149 UVMHIST_LOG(pdhist, "<- done (error=%d)", 1150 error,0,0,0); 1151 if (ptmp->flags & PG_WANTED) 1152 wakeup(ptmp); 1153 1154 /* 1155 * remove the swap slot from the aobj 1156 * and mark the aobj as having no real slot. 1157 * don't free the swap slot, thus preventing 1158 * it from being used again. 1159 */ 1160 1161 swslot = uao_set_swslot(&aobj->u_obj, pageidx, 1162 SWSLOT_BAD); 1163 if (swslot != -1) { 1164 uvm_swap_markbad(swslot, 1); 1165 } 1166 1167 uvm_lock_pageq(); 1168 uvm_pagefree(ptmp); 1169 uvm_unlock_pageq(); 1170 simple_unlock(&uobj->vmobjlock); 1171 return error; 1172 } 1173 } 1174 1175 /* 1176 * we got the page! clear the fake flag (indicates valid 1177 * data now in page) and plug into our result array. note 1178 * that page is still busy. 1179 * 1180 * it is the callers job to: 1181 * => check if the page is released 1182 * => unbusy the page 1183 * => activate the page 1184 */ 1185 1186 ptmp->flags &= ~PG_FAKE; 1187 pps[lcv] = ptmp; 1188 } 1189 1190 /* 1191 * finally, unlock object and return. 1192 */ 1193 1194 simple_unlock(&uobj->vmobjlock); 1195 UVMHIST_LOG(pdhist, "<- done (OK)",0,0,0,0); 1196 return 0; 1197 } 1198 1199 /* 1200 * uao_dropswap: release any swap resources from this aobj page. 1201 * 1202 * => aobj must be locked or have a reference count of 0. 1203 */ 1204 1205 void 1206 uao_dropswap(uobj, pageidx) 1207 struct uvm_object *uobj; 1208 int pageidx; 1209 { 1210 int slot; 1211 1212 slot = uao_set_swslot(uobj, pageidx, 0); 1213 if (slot) { 1214 uvm_swap_free(slot, 1); 1215 } 1216 } 1217 1218 1219 /* 1220 * page in every page in every aobj that is paged-out to a range of swslots. 1221 * 1222 * => nothing should be locked. 1223 * => returns TRUE if pagein was aborted due to lack of memory. 1224 */ 1225 1226 boolean_t 1227 uao_swap_off(startslot, endslot) 1228 int startslot, endslot; 1229 { 1230 struct uvm_aobj *aobj, *nextaobj; 1231 boolean_t rv; 1232 1233 /* 1234 * walk the list of all aobjs. 1235 */ 1236 1237 restart: 1238 simple_lock(&uao_list_lock); 1239 for (aobj = LIST_FIRST(&uao_list); 1240 aobj != NULL; 1241 aobj = nextaobj) { 1242 1243 /* 1244 * try to get the object lock, start all over if we fail. 1245 * most of the time we'll get the aobj lock, 1246 * so this should be a rare case. 1247 */ 1248 1249 if (!simple_lock_try(&aobj->u_obj.vmobjlock)) { 1250 simple_unlock(&uao_list_lock); 1251 goto restart; 1252 } 1253 1254 /* 1255 * add a ref to the aobj so it doesn't disappear 1256 * while we're working. 1257 */ 1258 1259 uao_reference_locked(&aobj->u_obj); 1260 1261 /* 1262 * now it's safe to unlock the uao list. 1263 */ 1264 1265 simple_unlock(&uao_list_lock); 1266 1267 /* 1268 * page in any pages in the swslot range. 1269 * if there's an error, abort and return the error. 1270 */ 1271 1272 rv = uao_pagein(aobj, startslot, endslot); 1273 if (rv) { 1274 uao_detach_locked(&aobj->u_obj); 1275 return rv; 1276 } 1277 1278 /* 1279 * we're done with this aobj. 1280 * relock the list and drop our ref on the aobj. 1281 */ 1282 1283 simple_lock(&uao_list_lock); 1284 nextaobj = LIST_NEXT(aobj, u_list); 1285 uao_detach_locked(&aobj->u_obj); 1286 } 1287 1288 /* 1289 * done with traversal, unlock the list 1290 */ 1291 simple_unlock(&uao_list_lock); 1292 return FALSE; 1293 } 1294 1295 1296 /* 1297 * page in any pages from aobj in the given range. 1298 * 1299 * => aobj must be locked and is returned locked. 1300 * => returns TRUE if pagein was aborted due to lack of memory. 1301 */ 1302 static boolean_t 1303 uao_pagein(aobj, startslot, endslot) 1304 struct uvm_aobj *aobj; 1305 int startslot, endslot; 1306 { 1307 boolean_t rv; 1308 1309 if (UAO_USES_SWHASH(aobj)) { 1310 struct uao_swhash_elt *elt; 1311 int bucket; 1312 1313 restart: 1314 for (bucket = aobj->u_swhashmask; bucket >= 0; bucket--) { 1315 for (elt = LIST_FIRST(&aobj->u_swhash[bucket]); 1316 elt != NULL; 1317 elt = LIST_NEXT(elt, list)) { 1318 int i; 1319 1320 for (i = 0; i < UAO_SWHASH_CLUSTER_SIZE; i++) { 1321 int slot = elt->slots[i]; 1322 1323 /* 1324 * if the slot isn't in range, skip it. 1325 */ 1326 1327 if (slot < startslot || 1328 slot >= endslot) { 1329 continue; 1330 } 1331 1332 /* 1333 * process the page, 1334 * the start over on this object 1335 * since the swhash elt 1336 * may have been freed. 1337 */ 1338 1339 rv = uao_pagein_page(aobj, 1340 UAO_SWHASH_ELT_PAGEIDX_BASE(elt) + i); 1341 if (rv) { 1342 return rv; 1343 } 1344 goto restart; 1345 } 1346 } 1347 } 1348 } else { 1349 int i; 1350 1351 for (i = 0; i < aobj->u_pages; i++) { 1352 int slot = aobj->u_swslots[i]; 1353 1354 /* 1355 * if the slot isn't in range, skip it 1356 */ 1357 1358 if (slot < startslot || slot >= endslot) { 1359 continue; 1360 } 1361 1362 /* 1363 * process the page. 1364 */ 1365 1366 rv = uao_pagein_page(aobj, i); 1367 if (rv) { 1368 return rv; 1369 } 1370 } 1371 } 1372 1373 return FALSE; 1374 } 1375 1376 /* 1377 * page in a page from an aobj. used for swap_off. 1378 * returns TRUE if pagein was aborted due to lack of memory. 1379 * 1380 * => aobj must be locked and is returned locked. 1381 */ 1382 1383 static boolean_t 1384 uao_pagein_page(aobj, pageidx) 1385 struct uvm_aobj *aobj; 1386 int pageidx; 1387 { 1388 struct vm_page *pg; 1389 int rv, slot, npages; 1390 1391 pg = NULL; 1392 npages = 1; 1393 /* locked: aobj */ 1394 rv = uao_get(&aobj->u_obj, pageidx << PAGE_SHIFT, 1395 &pg, &npages, 0, VM_PROT_READ|VM_PROT_WRITE, 0, 0); 1396 /* unlocked: aobj */ 1397 1398 /* 1399 * relock and finish up. 1400 */ 1401 1402 simple_lock(&aobj->u_obj.vmobjlock); 1403 switch (rv) { 1404 case 0: 1405 break; 1406 1407 case EIO: 1408 case ERESTART: 1409 1410 /* 1411 * nothing more to do on errors. 1412 * ERESTART can only mean that the anon was freed, 1413 * so again there's nothing to do. 1414 */ 1415 1416 return FALSE; 1417 } 1418 1419 /* 1420 * ok, we've got the page now. 1421 * mark it as dirty, clear its swslot and un-busy it. 1422 */ 1423 1424 slot = uao_set_swslot(&aobj->u_obj, pageidx, 0); 1425 uvm_swap_free(slot, 1); 1426 pg->flags &= ~(PG_BUSY|PG_CLEAN|PG_FAKE); 1427 UVM_PAGE_OWN(pg, NULL); 1428 1429 /* 1430 * deactivate the page (to make sure it's on a page queue). 1431 */ 1432 1433 uvm_lock_pageq(); 1434 uvm_pagedeactivate(pg); 1435 uvm_unlock_pageq(); 1436 return FALSE; 1437 } 1438