1 /* $NetBSD: uvm_aobj.c,v 1.49 2001/11/10 07:36:59 lukem 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 <sys/cdefs.h> 46 __KERNEL_RCSID(0, "$NetBSD: uvm_aobj.c,v 1.49 2001/11/10 07:36:59 lukem Exp $"); 47 48 #include "opt_uvmhist.h" 49 50 #include <sys/param.h> 51 #include <sys/systm.h> 52 #include <sys/proc.h> 53 #include <sys/malloc.h> 54 #include <sys/kernel.h> 55 #include <sys/pool.h> 56 #include <sys/kernel.h> 57 58 #include <uvm/uvm.h> 59 60 /* 61 * an aobj manages anonymous-memory backed uvm_objects. in addition 62 * to keeping the list of resident pages, it also keeps a list of 63 * allocated swap blocks. depending on the size of the aobj this list 64 * of allocated swap blocks is either stored in an array (small objects) 65 * or in a hash table (large objects). 66 */ 67 68 /* 69 * local structures 70 */ 71 72 /* 73 * for hash tables, we break the address space of the aobj into blocks 74 * of UAO_SWHASH_CLUSTER_SIZE pages. we require the cluster size to 75 * be a power of two. 76 */ 77 78 #define UAO_SWHASH_CLUSTER_SHIFT 4 79 #define UAO_SWHASH_CLUSTER_SIZE (1 << UAO_SWHASH_CLUSTER_SHIFT) 80 81 /* get the "tag" for this page index */ 82 #define UAO_SWHASH_ELT_TAG(PAGEIDX) \ 83 ((PAGEIDX) >> UAO_SWHASH_CLUSTER_SHIFT) 84 85 /* given an ELT and a page index, find the swap slot */ 86 #define UAO_SWHASH_ELT_PAGESLOT(ELT, PAGEIDX) \ 87 ((ELT)->slots[(PAGEIDX) & (UAO_SWHASH_CLUSTER_SIZE - 1)]) 88 89 /* given an ELT, return its pageidx base */ 90 #define UAO_SWHASH_ELT_PAGEIDX_BASE(ELT) \ 91 ((ELT)->tag << UAO_SWHASH_CLUSTER_SHIFT) 92 93 /* 94 * the swhash hash function 95 */ 96 97 #define UAO_SWHASH_HASH(AOBJ, PAGEIDX) \ 98 (&(AOBJ)->u_swhash[(((PAGEIDX) >> UAO_SWHASH_CLUSTER_SHIFT) \ 99 & (AOBJ)->u_swhashmask)]) 100 101 /* 102 * the swhash threshhold determines if we will use an array or a 103 * hash table to store the list of allocated swap blocks. 104 */ 105 106 #define UAO_SWHASH_THRESHOLD (UAO_SWHASH_CLUSTER_SIZE * 4) 107 #define UAO_USES_SWHASH(AOBJ) \ 108 ((AOBJ)->u_pages > UAO_SWHASH_THRESHOLD) /* use hash? */ 109 110 /* 111 * the number of buckets in a swhash, with an upper bound 112 */ 113 114 #define UAO_SWHASH_MAXBUCKETS 256 115 #define UAO_SWHASH_BUCKETS(AOBJ) \ 116 (MIN((AOBJ)->u_pages >> UAO_SWHASH_CLUSTER_SHIFT, \ 117 UAO_SWHASH_MAXBUCKETS)) 118 119 120 /* 121 * uao_swhash_elt: when a hash table is being used, this structure defines 122 * the format of an entry in the bucket list. 123 */ 124 125 struct uao_swhash_elt { 126 LIST_ENTRY(uao_swhash_elt) list; /* the hash list */ 127 voff_t tag; /* our 'tag' */ 128 int count; /* our number of active slots */ 129 int slots[UAO_SWHASH_CLUSTER_SIZE]; /* the slots */ 130 }; 131 132 /* 133 * uao_swhash: the swap hash table structure 134 */ 135 136 LIST_HEAD(uao_swhash, uao_swhash_elt); 137 138 /* 139 * uao_swhash_elt_pool: pool of uao_swhash_elt structures 140 */ 141 142 struct pool uao_swhash_elt_pool; 143 144 /* 145 * uvm_aobj: the actual anon-backed uvm_object 146 * 147 * => the uvm_object is at the top of the structure, this allows 148 * (struct uvm_aobj *) == (struct uvm_object *) 149 * => only one of u_swslots and u_swhash is used in any given aobj 150 */ 151 152 struct uvm_aobj { 153 struct uvm_object u_obj; /* has: lock, pgops, memq, #pages, #refs */ 154 int u_pages; /* number of pages in entire object */ 155 int u_flags; /* the flags (see uvm_aobj.h) */ 156 int *u_swslots; /* array of offset->swapslot mappings */ 157 /* 158 * hashtable of offset->swapslot mappings 159 * (u_swhash is an array of bucket heads) 160 */ 161 struct uao_swhash *u_swhash; 162 u_long u_swhashmask; /* mask for hashtable */ 163 LIST_ENTRY(uvm_aobj) u_list; /* global list of aobjs */ 164 }; 165 166 /* 167 * uvm_aobj_pool: pool of uvm_aobj structures 168 */ 169 170 struct pool uvm_aobj_pool; 171 172 /* 173 * local functions 174 */ 175 176 static struct uao_swhash_elt *uao_find_swhash_elt 177 __P((struct uvm_aobj *, int, boolean_t)); 178 179 static void uao_free __P((struct uvm_aobj *)); 180 static int uao_get __P((struct uvm_object *, voff_t, struct vm_page **, 181 int *, int, vm_prot_t, int, int)); 182 static boolean_t uao_put __P((struct uvm_object *, voff_t, voff_t, int)); 183 static boolean_t uao_pagein __P((struct uvm_aobj *, int, int)); 184 static boolean_t uao_pagein_page __P((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_get, /* get */ 198 uao_put, /* flush */ 199 }; 200 201 /* 202 * uao_list: global list of active aobjs, locked by uao_list_lock 203 */ 204 205 static LIST_HEAD(aobjlist, uvm_aobj) uao_list; 206 static struct simplelock uao_list_lock; 207 208 /* 209 * functions 210 */ 211 212 /* 213 * hash table/array related functions 214 */ 215 216 /* 217 * uao_find_swhash_elt: find (or create) a hash table entry for a page 218 * offset. 219 * 220 * => the object should be locked by the caller 221 */ 222 223 static struct uao_swhash_elt * 224 uao_find_swhash_elt(aobj, pageidx, create) 225 struct uvm_aobj *aobj; 226 int pageidx; 227 boolean_t create; 228 { 229 struct uao_swhash *swhash; 230 struct uao_swhash_elt *elt; 231 voff_t page_tag; 232 233 swhash = UAO_SWHASH_HASH(aobj, pageidx); 234 page_tag = UAO_SWHASH_ELT_TAG(pageidx); 235 236 /* 237 * now search the bucket for the requested tag 238 */ 239 240 LIST_FOREACH(elt, swhash, list) { 241 if (elt->tag == page_tag) { 242 return elt; 243 } 244 } 245 if (!create) { 246 return NULL; 247 } 248 249 /* 250 * allocate a new entry for the bucket and init/insert it in 251 */ 252 253 elt = pool_get(&uao_swhash_elt_pool, PR_NOWAIT); 254 if (elt == NULL) { 255 return NULL; 256 } 257 LIST_INSERT_HEAD(swhash, elt, list); 258 elt->tag = page_tag; 259 elt->count = 0; 260 memset(elt->slots, 0, sizeof(elt->slots)); 261 return elt; 262 } 263 264 /* 265 * uao_find_swslot: find the swap slot number for an aobj/pageidx 266 * 267 * => object must be locked by caller 268 */ 269 270 int 271 uao_find_swslot(uobj, pageidx) 272 struct uvm_object *uobj; 273 int pageidx; 274 { 275 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 276 struct uao_swhash_elt *elt; 277 278 /* 279 * if noswap flag is set, then we never return a slot 280 */ 281 282 if (aobj->u_flags & UAO_FLAG_NOSWAP) 283 return(0); 284 285 /* 286 * if hashing, look in hash table. 287 */ 288 289 if (UAO_USES_SWHASH(aobj)) { 290 elt = uao_find_swhash_elt(aobj, pageidx, FALSE); 291 if (elt) 292 return(UAO_SWHASH_ELT_PAGESLOT(elt, pageidx)); 293 else 294 return(0); 295 } 296 297 /* 298 * otherwise, look in the array 299 */ 300 301 return(aobj->u_swslots[pageidx]); 302 } 303 304 /* 305 * uao_set_swslot: set the swap slot for a page in an aobj. 306 * 307 * => setting a slot to zero frees the slot 308 * => object must be locked by caller 309 * => we return the old slot number, or -1 if we failed to allocate 310 * memory to record the new slot number 311 */ 312 313 int 314 uao_set_swslot(uobj, pageidx, slot) 315 struct uvm_object *uobj; 316 int pageidx, slot; 317 { 318 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 319 struct uao_swhash_elt *elt; 320 int oldslot; 321 UVMHIST_FUNC("uao_set_swslot"); UVMHIST_CALLED(pdhist); 322 UVMHIST_LOG(pdhist, "aobj %p pageidx %d slot %d", 323 aobj, pageidx, slot, 0); 324 325 /* 326 * if noswap flag is set, then we can't set a non-zero slot. 327 */ 328 329 if (aobj->u_flags & UAO_FLAG_NOSWAP) { 330 if (slot == 0) 331 return(0); 332 333 printf("uao_set_swslot: uobj = %p\n", uobj); 334 panic("uao_set_swslot: NOSWAP object"); 335 } 336 337 /* 338 * are we using a hash table? if so, add it in the hash. 339 */ 340 341 if (UAO_USES_SWHASH(aobj)) { 342 343 /* 344 * Avoid allocating an entry just to free it again if 345 * the page had not swap slot in the first place, and 346 * we are freeing. 347 */ 348 349 elt = uao_find_swhash_elt(aobj, pageidx, slot != 0); 350 if (elt == NULL) { 351 return slot ? -1 : 0; 352 } 353 354 oldslot = UAO_SWHASH_ELT_PAGESLOT(elt, pageidx); 355 UAO_SWHASH_ELT_PAGESLOT(elt, pageidx) = slot; 356 357 /* 358 * now adjust the elt's reference counter and free it if we've 359 * dropped it to zero. 360 */ 361 362 if (slot) { 363 if (oldslot == 0) 364 elt->count++; 365 } else { 366 if (oldslot) 367 elt->count--; 368 369 if (elt->count == 0) { 370 LIST_REMOVE(elt, list); 371 pool_put(&uao_swhash_elt_pool, elt); 372 } 373 } 374 } else { 375 /* we are using an array */ 376 oldslot = aobj->u_swslots[pageidx]; 377 aobj->u_swslots[pageidx] = slot; 378 } 379 return (oldslot); 380 } 381 382 /* 383 * end of hash/array functions 384 */ 385 386 /* 387 * uao_free: free all resources held by an aobj, and then free the aobj 388 * 389 * => the aobj should be dead 390 */ 391 392 static void 393 uao_free(aobj) 394 struct uvm_aobj *aobj; 395 { 396 int swpgonlydelta = 0; 397 398 simple_unlock(&aobj->u_obj.vmobjlock); 399 if (UAO_USES_SWHASH(aobj)) { 400 int i, hashbuckets = aobj->u_swhashmask + 1; 401 402 /* 403 * free the swslots from each hash bucket, 404 * then the hash bucket, and finally the hash table itself. 405 */ 406 407 for (i = 0; i < hashbuckets; i++) { 408 struct uao_swhash_elt *elt, *next; 409 410 for (elt = LIST_FIRST(&aobj->u_swhash[i]); 411 elt != NULL; 412 elt = next) { 413 int j; 414 415 for (j = 0; j < UAO_SWHASH_CLUSTER_SIZE; j++) { 416 int slot = elt->slots[j]; 417 418 if (slot == 0) { 419 continue; 420 } 421 uvm_swap_free(slot, 1); 422 swpgonlydelta++; 423 } 424 425 next = LIST_NEXT(elt, list); 426 pool_put(&uao_swhash_elt_pool, elt); 427 } 428 } 429 free(aobj->u_swhash, M_UVMAOBJ); 430 } else { 431 int i; 432 433 /* 434 * free the array 435 */ 436 437 for (i = 0; i < aobj->u_pages; i++) { 438 int slot = aobj->u_swslots[i]; 439 440 if (slot) { 441 uvm_swap_free(slot, 1); 442 swpgonlydelta++; 443 } 444 } 445 free(aobj->u_swslots, M_UVMAOBJ); 446 } 447 448 /* 449 * finally free the aobj itself 450 */ 451 452 pool_put(&uvm_aobj_pool, aobj); 453 454 /* 455 * adjust the counter of pages only in swap for all 456 * the swap slots we've freed. 457 */ 458 459 if (swpgonlydelta > 0) { 460 simple_lock(&uvm.swap_data_lock); 461 KASSERT(uvmexp.swpgonly >= swpgonlydelta); 462 uvmexp.swpgonly -= swpgonlydelta; 463 simple_unlock(&uvm.swap_data_lock); 464 } 465 } 466 467 /* 468 * pager functions 469 */ 470 471 /* 472 * uao_create: create an aobj of the given size and return its uvm_object. 473 * 474 * => for normal use, flags are always zero 475 * => for the kernel object, the flags are: 476 * UAO_FLAG_KERNOBJ - allocate the kernel object (can only happen once) 477 * UAO_FLAG_KERNSWAP - enable swapping of kernel object (" ") 478 */ 479 480 struct uvm_object * 481 uao_create(size, flags) 482 vsize_t size; 483 int flags; 484 { 485 static struct uvm_aobj kernel_object_store; 486 static int kobj_alloced = 0; 487 int pages = round_page(size) >> PAGE_SHIFT; 488 struct uvm_aobj *aobj; 489 490 /* 491 * malloc a new aobj unless we are asked for the kernel object 492 */ 493 494 if (flags & UAO_FLAG_KERNOBJ) { 495 KASSERT(!kobj_alloced); 496 aobj = &kernel_object_store; 497 aobj->u_pages = pages; 498 aobj->u_flags = UAO_FLAG_NOSWAP; 499 aobj->u_obj.uo_refs = UVM_OBJ_KERN; 500 kobj_alloced = UAO_FLAG_KERNOBJ; 501 } else if (flags & UAO_FLAG_KERNSWAP) { 502 KASSERT(kobj_alloced == UAO_FLAG_KERNOBJ); 503 aobj = &kernel_object_store; 504 kobj_alloced = UAO_FLAG_KERNSWAP; 505 } else { 506 aobj = pool_get(&uvm_aobj_pool, PR_WAITOK); 507 aobj->u_pages = pages; 508 aobj->u_flags = 0; 509 aobj->u_obj.uo_refs = 1; 510 } 511 512 /* 513 * allocate hash/array if necessary 514 * 515 * note: in the KERNSWAP case no need to worry about locking since 516 * we are still booting we should be the only thread around. 517 */ 518 519 if (flags == 0 || (flags & UAO_FLAG_KERNSWAP) != 0) { 520 int mflags = (flags & UAO_FLAG_KERNSWAP) != 0 ? 521 M_NOWAIT : M_WAITOK; 522 523 /* allocate hash table or array depending on object size */ 524 if (UAO_USES_SWHASH(aobj)) { 525 aobj->u_swhash = hashinit(UAO_SWHASH_BUCKETS(aobj), 526 HASH_LIST, M_UVMAOBJ, mflags, &aobj->u_swhashmask); 527 if (aobj->u_swhash == NULL) 528 panic("uao_create: hashinit swhash failed"); 529 } else { 530 aobj->u_swslots = malloc(pages * sizeof(int), 531 M_UVMAOBJ, mflags); 532 if (aobj->u_swslots == NULL) 533 panic("uao_create: malloc swslots failed"); 534 memset(aobj->u_swslots, 0, pages * sizeof(int)); 535 } 536 537 if (flags) { 538 aobj->u_flags &= ~UAO_FLAG_NOSWAP; /* clear noswap */ 539 return(&aobj->u_obj); 540 } 541 } 542 543 /* 544 * init aobj fields 545 */ 546 547 simple_lock_init(&aobj->u_obj.vmobjlock); 548 aobj->u_obj.pgops = &aobj_pager; 549 TAILQ_INIT(&aobj->u_obj.memq); 550 aobj->u_obj.uo_npages = 0; 551 552 /* 553 * now that aobj is ready, add it to the global list 554 */ 555 556 simple_lock(&uao_list_lock); 557 LIST_INSERT_HEAD(&uao_list, aobj, u_list); 558 simple_unlock(&uao_list_lock); 559 return(&aobj->u_obj); 560 } 561 562 563 564 /* 565 * uao_init: set up aobj pager subsystem 566 * 567 * => called at boot time from uvm_pager_init() 568 */ 569 570 void 571 uao_init(void) 572 { 573 static int uao_initialized; 574 575 if (uao_initialized) 576 return; 577 uao_initialized = TRUE; 578 LIST_INIT(&uao_list); 579 simple_lock_init(&uao_list_lock); 580 581 /* 582 * NOTE: Pages fror this pool must not come from a pageable 583 * kernel map! 584 */ 585 586 pool_init(&uao_swhash_elt_pool, sizeof(struct uao_swhash_elt), 587 0, 0, 0, "uaoeltpl", 0, NULL, NULL, M_UVMAOBJ); 588 pool_init(&uvm_aobj_pool, sizeof(struct uvm_aobj), 0, 0, 0, 589 "aobjpl", 0, 590 pool_page_alloc_nointr, pool_page_free_nointr, M_UVMAOBJ); 591 } 592 593 /* 594 * uao_reference: add a ref to an aobj 595 * 596 * => aobj must be unlocked 597 * => just lock it and call the locked version 598 */ 599 600 void 601 uao_reference(uobj) 602 struct uvm_object *uobj; 603 { 604 simple_lock(&uobj->vmobjlock); 605 uao_reference_locked(uobj); 606 simple_unlock(&uobj->vmobjlock); 607 } 608 609 /* 610 * uao_reference_locked: add a ref to an aobj that is already locked 611 * 612 * => aobj must be locked 613 * this needs to be separate from the normal routine 614 * since sometimes we need to add a reference to an aobj when 615 * it's already locked. 616 */ 617 618 void 619 uao_reference_locked(uobj) 620 struct uvm_object *uobj; 621 { 622 UVMHIST_FUNC("uao_reference"); UVMHIST_CALLED(maphist); 623 624 /* 625 * kernel_object already has plenty of references, leave it alone. 626 */ 627 628 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) 629 return; 630 631 uobj->uo_refs++; 632 UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)", 633 uobj, uobj->uo_refs,0,0); 634 } 635 636 /* 637 * uao_detach: drop a reference to an aobj 638 * 639 * => aobj must be unlocked 640 * => just lock it and call the locked version 641 */ 642 643 void 644 uao_detach(uobj) 645 struct uvm_object *uobj; 646 { 647 simple_lock(&uobj->vmobjlock); 648 uao_detach_locked(uobj); 649 } 650 651 /* 652 * uao_detach_locked: drop a reference to an aobj 653 * 654 * => aobj must be locked, and is unlocked (or freed) upon return. 655 * this needs to be separate from the normal routine 656 * since sometimes we need to detach from an aobj when 657 * it's already locked. 658 */ 659 660 void 661 uao_detach_locked(uobj) 662 struct uvm_object *uobj; 663 { 664 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 665 struct vm_page *pg; 666 UVMHIST_FUNC("uao_detach"); UVMHIST_CALLED(maphist); 667 668 /* 669 * detaching from kernel_object is a noop. 670 */ 671 672 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) { 673 simple_unlock(&uobj->vmobjlock); 674 return; 675 } 676 677 UVMHIST_LOG(maphist," (uobj=0x%x) ref=%d", uobj,uobj->uo_refs,0,0); 678 uobj->uo_refs--; 679 if (uobj->uo_refs) { 680 simple_unlock(&uobj->vmobjlock); 681 UVMHIST_LOG(maphist, "<- done (rc>0)", 0,0,0,0); 682 return; 683 } 684 685 /* 686 * remove the aobj from the global list. 687 */ 688 689 simple_lock(&uao_list_lock); 690 LIST_REMOVE(aobj, u_list); 691 simple_unlock(&uao_list_lock); 692 693 /* 694 * free all the pages left in the aobj. for each page, 695 * when the page is no longer busy (and thus after any disk i/o that 696 * it's involved in is complete), release any swap resources and 697 * free the page itself. 698 */ 699 700 uvm_lock_pageq(); 701 while ((pg = TAILQ_FIRST(&uobj->memq)) != NULL) { 702 pmap_page_protect(pg, VM_PROT_NONE); 703 if (pg->flags & PG_BUSY) { 704 pg->flags |= PG_WANTED; 705 uvm_unlock_pageq(); 706 UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, FALSE, 707 "uao_det", 0); 708 simple_lock(&uobj->vmobjlock); 709 uvm_lock_pageq(); 710 continue; 711 } 712 uao_dropswap(&aobj->u_obj, pg->offset >> PAGE_SHIFT); 713 uvm_pagefree(pg); 714 } 715 uvm_unlock_pageq(); 716 717 /* 718 * finally, free the aobj itself. 719 */ 720 721 uao_free(aobj); 722 } 723 724 /* 725 * uao_put: flush pages out of a uvm object 726 * 727 * => object should be locked by caller. we may _unlock_ the object 728 * if (and only if) we need to clean a page (PGO_CLEANIT). 729 * XXXJRT Currently, however, we don't. In the case of cleaning 730 * XXXJRT a page, we simply just deactivate it. Should probably 731 * XXXJRT handle this better, in the future (although "flushing" 732 * XXXJRT anonymous memory isn't terribly important). 733 * => if PGO_CLEANIT is not set, then we will neither unlock the object 734 * or block. 735 * => if PGO_ALLPAGE is set, then all pages in the object are valid targets 736 * for flushing. 737 * => NOTE: we rely on the fact that the object's memq is a TAILQ and 738 * that new pages are inserted on the tail end of the list. thus, 739 * we can make a complete pass through the object in one go by starting 740 * at the head and working towards the tail (new pages are put in 741 * front of us). 742 * => NOTE: we are allowed to lock the page queues, so the caller 743 * must not be holding the lock on them [e.g. pagedaemon had 744 * better not call us with the queues locked] 745 * => we return TRUE unless we encountered some sort of I/O error 746 * XXXJRT currently never happens, as we never directly initiate 747 * XXXJRT I/O 748 * 749 * note on page traversal: 750 * we can traverse the pages in an object either by going down the 751 * linked list in "uobj->memq", or we can go over the address range 752 * by page doing hash table lookups for each address. depending 753 * on how many pages are in the object it may be cheaper to do one 754 * or the other. we set "by_list" to true if we are using memq. 755 * if the cost of a hash lookup was equal to the cost of the list 756 * traversal we could compare the number of pages in the start->stop 757 * range to the total number of pages in the object. however, it 758 * seems that a hash table lookup is more expensive than the linked 759 * list traversal, so we multiply the number of pages in the 760 * start->stop range by a penalty which we define below. 761 */ 762 763 int 764 uao_put(uobj, start, stop, flags) 765 struct uvm_object *uobj; 766 voff_t start, stop; 767 int flags; 768 { 769 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 770 struct vm_page *pg, *nextpg; 771 boolean_t by_list; 772 voff_t curoff; 773 UVMHIST_FUNC("uao_put"); UVMHIST_CALLED(maphist); 774 775 curoff = 0; 776 if (flags & PGO_ALLPAGES) { 777 start = 0; 778 stop = aobj->u_pages << PAGE_SHIFT; 779 by_list = TRUE; /* always go by the list */ 780 } else { 781 start = trunc_page(start); 782 stop = round_page(stop); 783 if (stop > (aobj->u_pages << PAGE_SHIFT)) { 784 printf("uao_flush: strange, got an out of range " 785 "flush (fixed)\n"); 786 stop = aobj->u_pages << PAGE_SHIFT; 787 } 788 by_list = (uobj->uo_npages <= 789 ((stop - start) >> PAGE_SHIFT) * UVM_PAGE_HASH_PENALTY); 790 } 791 UVMHIST_LOG(maphist, 792 " flush start=0x%lx, stop=0x%x, by_list=%d, flags=0x%x", 793 start, stop, by_list, flags); 794 795 /* 796 * Don't need to do any work here if we're not freeing 797 * or deactivating pages. 798 */ 799 800 if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) { 801 simple_unlock(&uobj->vmobjlock); 802 return 0; 803 } 804 805 /* 806 * now do it. note: we must update nextpg in the body of loop or we 807 * will get stuck. we need to use nextpg because we may free "pg" 808 * before doing the next loop. 809 */ 810 811 if (by_list) { 812 pg = TAILQ_FIRST(&uobj->memq); 813 } else { 814 curoff = start; 815 pg = uvm_pagelookup(uobj, curoff); 816 } 817 818 nextpg = NULL; 819 uvm_lock_pageq(); 820 821 /* locked: both page queues and uobj */ 822 for ( ; (by_list && pg != NULL) || 823 (!by_list && curoff < stop) ; pg = nextpg) { 824 if (by_list) { 825 nextpg = TAILQ_NEXT(pg, listq); 826 if (pg->offset < start || pg->offset >= stop) 827 continue; 828 } else { 829 curoff += PAGE_SIZE; 830 if (curoff < stop) 831 nextpg = uvm_pagelookup(uobj, curoff); 832 if (pg == NULL) 833 continue; 834 } 835 switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) { 836 837 /* 838 * XXX In these first 3 cases, we always just 839 * XXX deactivate the page. We may want to 840 * XXX handle the different cases more specifically 841 * XXX in the future. 842 */ 843 844 case PGO_CLEANIT|PGO_FREE: 845 case PGO_CLEANIT|PGO_DEACTIVATE: 846 case PGO_DEACTIVATE: 847 deactivate_it: 848 /* skip the page if it's loaned or wired */ 849 if (pg->loan_count != 0 || pg->wire_count != 0) 850 continue; 851 852 /* ...and deactivate the page. */ 853 pmap_clear_reference(pg); 854 uvm_pagedeactivate(pg); 855 continue; 856 857 case PGO_FREE: 858 859 /* 860 * If there are multiple references to 861 * the object, just deactivate the page. 862 */ 863 864 if (uobj->uo_refs > 1) 865 goto deactivate_it; 866 867 /* XXX skip the page if it's loaned or wired */ 868 if (pg->loan_count != 0 || pg->wire_count != 0) 869 continue; 870 871 /* 872 * wait if the page is busy, then free the swap slot 873 * and the page. 874 */ 875 876 pmap_page_protect(pg, VM_PROT_NONE); 877 while (pg->flags & PG_BUSY) { 878 pg->flags |= PG_WANTED; 879 uvm_unlock_pageq(); 880 UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, 0, 881 "uao_put", 0); 882 simple_lock(&uobj->vmobjlock); 883 uvm_lock_pageq(); 884 } 885 uao_dropswap(uobj, pg->offset >> PAGE_SHIFT); 886 uvm_pagefree(pg); 887 continue; 888 } 889 } 890 uvm_unlock_pageq(); 891 simple_unlock(&uobj->vmobjlock); 892 return 0; 893 } 894 895 /* 896 * uao_get: fetch me a page 897 * 898 * we have three cases: 899 * 1: page is resident -> just return the page. 900 * 2: page is zero-fill -> allocate a new page and zero it. 901 * 3: page is swapped out -> fetch the page from swap. 902 * 903 * cases 1 and 2 can be handled with PGO_LOCKED, case 3 cannot. 904 * so, if the "center" page hits case 3 (or any page, with PGO_ALLPAGES), 905 * then we will need to return EBUSY. 906 * 907 * => prefer map unlocked (not required) 908 * => object must be locked! we will _unlock_ it before starting any I/O. 909 * => flags: PGO_ALLPAGES: get all of the pages 910 * PGO_LOCKED: fault data structures are locked 911 * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx] 912 * => NOTE: caller must check for released pages!! 913 */ 914 915 static int 916 uao_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags) 917 struct uvm_object *uobj; 918 voff_t offset; 919 struct vm_page **pps; 920 int *npagesp; 921 int centeridx, advice, flags; 922 vm_prot_t access_type; 923 { 924 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 925 voff_t current_offset; 926 struct vm_page *ptmp; 927 int lcv, gotpages, maxpages, swslot, error, pageidx; 928 boolean_t done; 929 UVMHIST_FUNC("uao_get"); UVMHIST_CALLED(pdhist); 930 931 UVMHIST_LOG(pdhist, "aobj=%p offset=%d, flags=%d", 932 aobj, offset, flags,0); 933 934 /* 935 * get number of pages 936 */ 937 938 maxpages = *npagesp; 939 940 /* 941 * step 1: handled the case where fault data structures are locked. 942 */ 943 944 if (flags & PGO_LOCKED) { 945 946 /* 947 * step 1a: get pages that are already resident. only do 948 * this if the data structures are locked (i.e. the first 949 * time through). 950 */ 951 952 done = TRUE; /* be optimistic */ 953 gotpages = 0; /* # of pages we got so far */ 954 for (lcv = 0, current_offset = offset ; lcv < maxpages ; 955 lcv++, current_offset += PAGE_SIZE) { 956 /* do we care about this page? if not, skip it */ 957 if (pps[lcv] == PGO_DONTCARE) 958 continue; 959 ptmp = uvm_pagelookup(uobj, current_offset); 960 961 /* 962 * if page is new, attempt to allocate the page, 963 * zero-fill'd. 964 */ 965 966 if (ptmp == NULL && uao_find_swslot(&aobj->u_obj, 967 current_offset >> PAGE_SHIFT) == 0) { 968 ptmp = uvm_pagealloc(uobj, current_offset, 969 NULL, UVM_PGA_ZERO); 970 if (ptmp) { 971 /* new page */ 972 ptmp->flags &= ~(PG_FAKE); 973 ptmp->pqflags |= PQ_AOBJ; 974 goto gotpage; 975 } 976 } 977 978 /* 979 * to be useful must get a non-busy page 980 */ 981 982 if (ptmp == NULL || (ptmp->flags & PG_BUSY) != 0) { 983 if (lcv == centeridx || 984 (flags & PGO_ALLPAGES) != 0) 985 /* need to do a wait or I/O! */ 986 done = FALSE; 987 continue; 988 } 989 990 /* 991 * useful page: busy/lock it and plug it in our 992 * result array 993 */ 994 995 /* caller must un-busy this page */ 996 ptmp->flags |= PG_BUSY; 997 UVM_PAGE_OWN(ptmp, "uao_get1"); 998 gotpage: 999 pps[lcv] = ptmp; 1000 gotpages++; 1001 } 1002 1003 /* 1004 * step 1b: now we've either done everything needed or we 1005 * to unlock and do some waiting or I/O. 1006 */ 1007 1008 UVMHIST_LOG(pdhist, "<- done (done=%d)", done, 0,0,0); 1009 *npagesp = gotpages; 1010 if (done) 1011 return 0; 1012 else 1013 return EBUSY; 1014 } 1015 1016 /* 1017 * step 2: get non-resident or busy pages. 1018 * object is locked. data structures are unlocked. 1019 */ 1020 1021 for (lcv = 0, current_offset = offset ; lcv < maxpages ; 1022 lcv++, current_offset += PAGE_SIZE) { 1023 1024 /* 1025 * - skip over pages we've already gotten or don't want 1026 * - skip over pages we don't _have_ to get 1027 */ 1028 1029 if (pps[lcv] != NULL || 1030 (lcv != centeridx && (flags & PGO_ALLPAGES) == 0)) 1031 continue; 1032 1033 pageidx = current_offset >> PAGE_SHIFT; 1034 1035 /* 1036 * we have yet to locate the current page (pps[lcv]). we 1037 * first look for a page that is already at the current offset. 1038 * if we find a page, we check to see if it is busy or 1039 * released. if that is the case, then we sleep on the page 1040 * until it is no longer busy or released and repeat the lookup. 1041 * if the page we found is neither busy nor released, then we 1042 * busy it (so we own it) and plug it into pps[lcv]. this 1043 * 'break's the following while loop and indicates we are 1044 * ready to move on to the next page in the "lcv" loop above. 1045 * 1046 * if we exit the while loop with pps[lcv] still set to NULL, 1047 * then it means that we allocated a new busy/fake/clean page 1048 * ptmp in the object and we need to do I/O to fill in the data. 1049 */ 1050 1051 /* top of "pps" while loop */ 1052 while (pps[lcv] == NULL) { 1053 /* look for a resident page */ 1054 ptmp = uvm_pagelookup(uobj, current_offset); 1055 1056 /* not resident? allocate one now (if we can) */ 1057 if (ptmp == NULL) { 1058 1059 ptmp = uvm_pagealloc(uobj, current_offset, 1060 NULL, 0); 1061 1062 /* out of RAM? */ 1063 if (ptmp == NULL) { 1064 simple_unlock(&uobj->vmobjlock); 1065 UVMHIST_LOG(pdhist, 1066 "sleeping, ptmp == NULL\n",0,0,0,0); 1067 uvm_wait("uao_getpage"); 1068 simple_lock(&uobj->vmobjlock); 1069 continue; 1070 } 1071 1072 /* 1073 * safe with PQ's unlocked: because we just 1074 * alloc'd the page 1075 */ 1076 1077 ptmp->pqflags |= PQ_AOBJ; 1078 1079 /* 1080 * got new page ready for I/O. break pps while 1081 * loop. pps[lcv] is still NULL. 1082 */ 1083 1084 break; 1085 } 1086 1087 /* page is there, see if we need to wait on it */ 1088 if ((ptmp->flags & PG_BUSY) != 0) { 1089 ptmp->flags |= PG_WANTED; 1090 UVMHIST_LOG(pdhist, 1091 "sleeping, ptmp->flags 0x%x\n", 1092 ptmp->flags,0,0,0); 1093 UVM_UNLOCK_AND_WAIT(ptmp, &uobj->vmobjlock, 1094 FALSE, "uao_get", 0); 1095 simple_lock(&uobj->vmobjlock); 1096 continue; 1097 } 1098 1099 /* 1100 * if we get here then the page has become resident and 1101 * unbusy between steps 1 and 2. we busy it now (so we 1102 * own it) and set pps[lcv] (so that we exit the while 1103 * loop). 1104 */ 1105 1106 /* we own it, caller must un-busy */ 1107 ptmp->flags |= PG_BUSY; 1108 UVM_PAGE_OWN(ptmp, "uao_get2"); 1109 pps[lcv] = ptmp; 1110 } 1111 1112 /* 1113 * if we own the valid page at the correct offset, pps[lcv] will 1114 * point to it. nothing more to do except go to the next page. 1115 */ 1116 1117 if (pps[lcv]) 1118 continue; /* next lcv */ 1119 1120 /* 1121 * we have a "fake/busy/clean" page that we just allocated. 1122 * do the needed "i/o", either reading from swap or zeroing. 1123 */ 1124 1125 swslot = uao_find_swslot(&aobj->u_obj, pageidx); 1126 1127 /* 1128 * just zero the page if there's nothing in swap. 1129 */ 1130 1131 if (swslot == 0) { 1132 1133 /* 1134 * page hasn't existed before, just zero it. 1135 */ 1136 1137 uvm_pagezero(ptmp); 1138 } else { 1139 UVMHIST_LOG(pdhist, "pagein from swslot %d", 1140 swslot, 0,0,0); 1141 1142 /* 1143 * page in the swapped-out page. 1144 * unlock object for i/o, relock when done. 1145 */ 1146 1147 simple_unlock(&uobj->vmobjlock); 1148 error = uvm_swap_get(ptmp, swslot, PGO_SYNCIO); 1149 simple_lock(&uobj->vmobjlock); 1150 1151 /* 1152 * I/O done. check for errors. 1153 */ 1154 1155 if (error != 0) { 1156 UVMHIST_LOG(pdhist, "<- done (error=%d)", 1157 error,0,0,0); 1158 if (ptmp->flags & PG_WANTED) 1159 wakeup(ptmp); 1160 1161 /* 1162 * remove the swap slot from the aobj 1163 * and mark the aobj as having no real slot. 1164 * don't free the swap slot, thus preventing 1165 * it from being used again. 1166 */ 1167 1168 swslot = uao_set_swslot(&aobj->u_obj, pageidx, 1169 SWSLOT_BAD); 1170 if (swslot != -1) { 1171 uvm_swap_markbad(swslot, 1); 1172 } 1173 1174 uvm_lock_pageq(); 1175 uvm_pagefree(ptmp); 1176 uvm_unlock_pageq(); 1177 simple_unlock(&uobj->vmobjlock); 1178 return error; 1179 } 1180 } 1181 1182 /* 1183 * we got the page! clear the fake flag (indicates valid 1184 * data now in page) and plug into our result array. note 1185 * that page is still busy. 1186 * 1187 * it is the callers job to: 1188 * => check if the page is released 1189 * => unbusy the page 1190 * => activate the page 1191 */ 1192 1193 ptmp->flags &= ~PG_FAKE; 1194 pps[lcv] = ptmp; 1195 } 1196 1197 /* 1198 * finally, unlock object and return. 1199 */ 1200 1201 simple_unlock(&uobj->vmobjlock); 1202 UVMHIST_LOG(pdhist, "<- done (OK)",0,0,0,0); 1203 return 0; 1204 } 1205 1206 /* 1207 * uao_dropswap: release any swap resources from this aobj page. 1208 * 1209 * => aobj must be locked or have a reference count of 0. 1210 */ 1211 1212 void 1213 uao_dropswap(uobj, pageidx) 1214 struct uvm_object *uobj; 1215 int pageidx; 1216 { 1217 int slot; 1218 1219 slot = uao_set_swslot(uobj, pageidx, 0); 1220 if (slot) { 1221 uvm_swap_free(slot, 1); 1222 } 1223 } 1224 1225 /* 1226 * page in every page in every aobj that is paged-out to a range of swslots. 1227 * 1228 * => nothing should be locked. 1229 * => returns TRUE if pagein was aborted due to lack of memory. 1230 */ 1231 1232 boolean_t 1233 uao_swap_off(startslot, endslot) 1234 int startslot, endslot; 1235 { 1236 struct uvm_aobj *aobj, *nextaobj; 1237 boolean_t rv; 1238 1239 /* 1240 * walk the list of all aobjs. 1241 */ 1242 1243 restart: 1244 simple_lock(&uao_list_lock); 1245 for (aobj = LIST_FIRST(&uao_list); 1246 aobj != NULL; 1247 aobj = nextaobj) { 1248 1249 /* 1250 * try to get the object lock, start all over if we fail. 1251 * most of the time we'll get the aobj lock, 1252 * so this should be a rare case. 1253 */ 1254 1255 if (!simple_lock_try(&aobj->u_obj.vmobjlock)) { 1256 simple_unlock(&uao_list_lock); 1257 goto restart; 1258 } 1259 1260 /* 1261 * add a ref to the aobj so it doesn't disappear 1262 * while we're working. 1263 */ 1264 1265 uao_reference_locked(&aobj->u_obj); 1266 1267 /* 1268 * now it's safe to unlock the uao list. 1269 */ 1270 1271 simple_unlock(&uao_list_lock); 1272 1273 /* 1274 * page in any pages in the swslot range. 1275 * if there's an error, abort and return the error. 1276 */ 1277 1278 rv = uao_pagein(aobj, startslot, endslot); 1279 if (rv) { 1280 uao_detach_locked(&aobj->u_obj); 1281 return rv; 1282 } 1283 1284 /* 1285 * we're done with this aobj. 1286 * relock the list and drop our ref on the aobj. 1287 */ 1288 1289 simple_lock(&uao_list_lock); 1290 nextaobj = LIST_NEXT(aobj, u_list); 1291 uao_detach_locked(&aobj->u_obj); 1292 } 1293 1294 /* 1295 * done with traversal, unlock the list 1296 */ 1297 simple_unlock(&uao_list_lock); 1298 return FALSE; 1299 } 1300 1301 1302 /* 1303 * page in any pages from aobj in the given range. 1304 * 1305 * => aobj must be locked and is returned locked. 1306 * => returns TRUE if pagein was aborted due to lack of memory. 1307 */ 1308 static boolean_t 1309 uao_pagein(aobj, startslot, endslot) 1310 struct uvm_aobj *aobj; 1311 int startslot, endslot; 1312 { 1313 boolean_t rv; 1314 1315 if (UAO_USES_SWHASH(aobj)) { 1316 struct uao_swhash_elt *elt; 1317 int bucket; 1318 1319 restart: 1320 for (bucket = aobj->u_swhashmask; bucket >= 0; bucket--) { 1321 for (elt = LIST_FIRST(&aobj->u_swhash[bucket]); 1322 elt != NULL; 1323 elt = LIST_NEXT(elt, list)) { 1324 int i; 1325 1326 for (i = 0; i < UAO_SWHASH_CLUSTER_SIZE; i++) { 1327 int slot = elt->slots[i]; 1328 1329 /* 1330 * if the slot isn't in range, skip it. 1331 */ 1332 1333 if (slot < startslot || 1334 slot >= endslot) { 1335 continue; 1336 } 1337 1338 /* 1339 * process the page, 1340 * the start over on this object 1341 * since the swhash elt 1342 * may have been freed. 1343 */ 1344 1345 rv = uao_pagein_page(aobj, 1346 UAO_SWHASH_ELT_PAGEIDX_BASE(elt) + i); 1347 if (rv) { 1348 return rv; 1349 } 1350 goto restart; 1351 } 1352 } 1353 } 1354 } else { 1355 int i; 1356 1357 for (i = 0; i < aobj->u_pages; i++) { 1358 int slot = aobj->u_swslots[i]; 1359 1360 /* 1361 * if the slot isn't in range, skip it 1362 */ 1363 1364 if (slot < startslot || slot >= endslot) { 1365 continue; 1366 } 1367 1368 /* 1369 * process the page. 1370 */ 1371 1372 rv = uao_pagein_page(aobj, i); 1373 if (rv) { 1374 return rv; 1375 } 1376 } 1377 } 1378 1379 return FALSE; 1380 } 1381 1382 /* 1383 * page in a page from an aobj. used for swap_off. 1384 * returns TRUE if pagein was aborted due to lack of memory. 1385 * 1386 * => aobj must be locked and is returned locked. 1387 */ 1388 1389 static boolean_t 1390 uao_pagein_page(aobj, pageidx) 1391 struct uvm_aobj *aobj; 1392 int pageidx; 1393 { 1394 struct vm_page *pg; 1395 int rv, slot, npages; 1396 1397 pg = NULL; 1398 npages = 1; 1399 /* locked: aobj */ 1400 rv = uao_get(&aobj->u_obj, pageidx << PAGE_SHIFT, 1401 &pg, &npages, 0, VM_PROT_READ|VM_PROT_WRITE, 0, 0); 1402 /* unlocked: aobj */ 1403 1404 /* 1405 * relock and finish up. 1406 */ 1407 1408 simple_lock(&aobj->u_obj.vmobjlock); 1409 switch (rv) { 1410 case 0: 1411 break; 1412 1413 case EIO: 1414 case ERESTART: 1415 1416 /* 1417 * nothing more to do on errors. 1418 * ERESTART can only mean that the anon was freed, 1419 * so again there's nothing to do. 1420 */ 1421 1422 return FALSE; 1423 } 1424 1425 /* 1426 * ok, we've got the page now. 1427 * mark it as dirty, clear its swslot and un-busy it. 1428 */ 1429 1430 slot = uao_set_swslot(&aobj->u_obj, pageidx, 0); 1431 uvm_swap_free(slot, 1); 1432 pg->flags &= ~(PG_BUSY|PG_CLEAN|PG_FAKE); 1433 UVM_PAGE_OWN(pg, NULL); 1434 1435 /* 1436 * deactivate the page (to make sure it's on a page queue). 1437 */ 1438 1439 uvm_lock_pageq(); 1440 uvm_pagedeactivate(pg); 1441 uvm_unlock_pageq(); 1442 return FALSE; 1443 } 1444