1 /* $NetBSD: uvm_aobj.c,v 1.56 2003/04/12 14:36:43 yamt 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 1.56 2003/04/12 14:36:43 yamt 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 MALLOC_DEFINE(M_UVMAOBJ, "UVM aobj", "UVM aobj and related structures"); 173 174 /* 175 * local functions 176 */ 177 178 static struct uao_swhash_elt *uao_find_swhash_elt 179 __P((struct uvm_aobj *, int, boolean_t)); 180 181 static void uao_free __P((struct uvm_aobj *)); 182 static int uao_get __P((struct uvm_object *, voff_t, struct vm_page **, 183 int *, int, vm_prot_t, int, int)); 184 static boolean_t uao_put __P((struct uvm_object *, voff_t, voff_t, int)); 185 static boolean_t uao_pagein __P((struct uvm_aobj *, int, int)); 186 static boolean_t uao_pagein_page __P((struct uvm_aobj *, int)); 187 188 /* 189 * aobj_pager 190 * 191 * note that some functions (e.g. put) are handled elsewhere 192 */ 193 194 struct uvm_pagerops aobj_pager = { 195 NULL, /* init */ 196 uao_reference, /* reference */ 197 uao_detach, /* detach */ 198 NULL, /* fault */ 199 uao_get, /* get */ 200 uao_put, /* flush */ 201 }; 202 203 /* 204 * uao_list: global list of active aobjs, locked by uao_list_lock 205 */ 206 207 static LIST_HEAD(aobjlist, uvm_aobj) uao_list; 208 static struct simplelock uao_list_lock; 209 210 /* 211 * functions 212 */ 213 214 /* 215 * hash table/array related functions 216 */ 217 218 /* 219 * uao_find_swhash_elt: find (or create) a hash table entry for a page 220 * offset. 221 * 222 * => the object should be locked by the caller 223 */ 224 225 static struct uao_swhash_elt * 226 uao_find_swhash_elt(aobj, pageidx, create) 227 struct uvm_aobj *aobj; 228 int pageidx; 229 boolean_t create; 230 { 231 struct uao_swhash *swhash; 232 struct uao_swhash_elt *elt; 233 voff_t page_tag; 234 235 swhash = UAO_SWHASH_HASH(aobj, pageidx); 236 page_tag = UAO_SWHASH_ELT_TAG(pageidx); 237 238 /* 239 * now search the bucket for the requested tag 240 */ 241 242 LIST_FOREACH(elt, swhash, list) { 243 if (elt->tag == page_tag) { 244 return elt; 245 } 246 } 247 if (!create) { 248 return NULL; 249 } 250 251 /* 252 * allocate a new entry for the bucket and init/insert it in 253 */ 254 255 elt = pool_get(&uao_swhash_elt_pool, PR_NOWAIT); 256 if (elt == NULL) { 257 return NULL; 258 } 259 LIST_INSERT_HEAD(swhash, elt, list); 260 elt->tag = page_tag; 261 elt->count = 0; 262 memset(elt->slots, 0, sizeof(elt->slots)); 263 return elt; 264 } 265 266 /* 267 * uao_find_swslot: find the swap slot number for an aobj/pageidx 268 * 269 * => object must be locked by caller 270 */ 271 272 int 273 uao_find_swslot(uobj, pageidx) 274 struct uvm_object *uobj; 275 int pageidx; 276 { 277 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 278 struct uao_swhash_elt *elt; 279 280 /* 281 * if noswap flag is set, then we never return a slot 282 */ 283 284 if (aobj->u_flags & UAO_FLAG_NOSWAP) 285 return(0); 286 287 /* 288 * if hashing, look in hash table. 289 */ 290 291 if (UAO_USES_SWHASH(aobj)) { 292 elt = uao_find_swhash_elt(aobj, pageidx, FALSE); 293 if (elt) 294 return(UAO_SWHASH_ELT_PAGESLOT(elt, pageidx)); 295 else 296 return(0); 297 } 298 299 /* 300 * otherwise, look in the array 301 */ 302 303 return(aobj->u_swslots[pageidx]); 304 } 305 306 /* 307 * uao_set_swslot: set the swap slot for a page in an aobj. 308 * 309 * => setting a slot to zero frees the slot 310 * => object must be locked by caller 311 * => we return the old slot number, or -1 if we failed to allocate 312 * memory to record the new slot number 313 */ 314 315 int 316 uao_set_swslot(uobj, pageidx, slot) 317 struct uvm_object *uobj; 318 int pageidx, slot; 319 { 320 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 321 struct uao_swhash_elt *elt; 322 int oldslot; 323 UVMHIST_FUNC("uao_set_swslot"); UVMHIST_CALLED(pdhist); 324 UVMHIST_LOG(pdhist, "aobj %p pageidx %d slot %d", 325 aobj, pageidx, slot, 0); 326 327 /* 328 * if noswap flag is set, then we can't set a non-zero slot. 329 */ 330 331 if (aobj->u_flags & UAO_FLAG_NOSWAP) { 332 if (slot == 0) 333 return(0); 334 335 printf("uao_set_swslot: uobj = %p\n", uobj); 336 panic("uao_set_swslot: NOSWAP object"); 337 } 338 339 /* 340 * are we using a hash table? if so, add it in the hash. 341 */ 342 343 if (UAO_USES_SWHASH(aobj)) { 344 345 /* 346 * Avoid allocating an entry just to free it again if 347 * the page had not swap slot in the first place, and 348 * we are freeing. 349 */ 350 351 elt = uao_find_swhash_elt(aobj, pageidx, slot != 0); 352 if (elt == NULL) { 353 return slot ? -1 : 0; 354 } 355 356 oldslot = UAO_SWHASH_ELT_PAGESLOT(elt, pageidx); 357 UAO_SWHASH_ELT_PAGESLOT(elt, pageidx) = slot; 358 359 /* 360 * now adjust the elt's reference counter and free it if we've 361 * dropped it to zero. 362 */ 363 364 if (slot) { 365 if (oldslot == 0) 366 elt->count++; 367 } else { 368 if (oldslot) 369 elt->count--; 370 371 if (elt->count == 0) { 372 LIST_REMOVE(elt, list); 373 pool_put(&uao_swhash_elt_pool, elt); 374 } 375 } 376 } else { 377 /* we are using an array */ 378 oldslot = aobj->u_swslots[pageidx]; 379 aobj->u_swslots[pageidx] = slot; 380 } 381 return (oldslot); 382 } 383 384 /* 385 * end of hash/array functions 386 */ 387 388 /* 389 * uao_free: free all resources held by an aobj, and then free the aobj 390 * 391 * => the aobj should be dead 392 */ 393 394 static void 395 uao_free(aobj) 396 struct uvm_aobj *aobj; 397 { 398 int swpgonlydelta = 0; 399 400 simple_unlock(&aobj->u_obj.vmobjlock); 401 if (UAO_USES_SWHASH(aobj)) { 402 int i, hashbuckets = aobj->u_swhashmask + 1; 403 404 /* 405 * free the swslots from each hash bucket, 406 * then the hash bucket, and finally the hash table itself. 407 */ 408 409 for (i = 0; i < hashbuckets; i++) { 410 struct uao_swhash_elt *elt, *next; 411 412 for (elt = LIST_FIRST(&aobj->u_swhash[i]); 413 elt != NULL; 414 elt = next) { 415 int j; 416 417 for (j = 0; j < UAO_SWHASH_CLUSTER_SIZE; j++) { 418 int slot = elt->slots[j]; 419 420 if (slot == 0) { 421 continue; 422 } 423 uvm_swap_free(slot, 1); 424 swpgonlydelta++; 425 } 426 427 next = LIST_NEXT(elt, list); 428 pool_put(&uao_swhash_elt_pool, elt); 429 } 430 } 431 free(aobj->u_swhash, M_UVMAOBJ); 432 } else { 433 int i; 434 435 /* 436 * free the array 437 */ 438 439 for (i = 0; i < aobj->u_pages; i++) { 440 int slot = aobj->u_swslots[i]; 441 442 if (slot) { 443 uvm_swap_free(slot, 1); 444 swpgonlydelta++; 445 } 446 } 447 free(aobj->u_swslots, M_UVMAOBJ); 448 } 449 450 /* 451 * finally free the aobj itself 452 */ 453 454 pool_put(&uvm_aobj_pool, aobj); 455 456 /* 457 * adjust the counter of pages only in swap for all 458 * the swap slots we've freed. 459 */ 460 461 if (swpgonlydelta > 0) { 462 simple_lock(&uvm.swap_data_lock); 463 KASSERT(uvmexp.swpgonly >= swpgonlydelta); 464 uvmexp.swpgonly -= swpgonlydelta; 465 simple_unlock(&uvm.swap_data_lock); 466 } 467 } 468 469 /* 470 * pager functions 471 */ 472 473 /* 474 * uao_create: create an aobj of the given size and return its uvm_object. 475 * 476 * => for normal use, flags are always zero 477 * => for the kernel object, the flags are: 478 * UAO_FLAG_KERNOBJ - allocate the kernel object (can only happen once) 479 * UAO_FLAG_KERNSWAP - enable swapping of kernel object (" ") 480 */ 481 482 struct uvm_object * 483 uao_create(size, flags) 484 vsize_t size; 485 int flags; 486 { 487 static struct uvm_aobj kernel_object_store; 488 static int kobj_alloced = 0; 489 int pages = round_page(size) >> PAGE_SHIFT; 490 struct uvm_aobj *aobj; 491 492 /* 493 * malloc a new aobj unless we are asked for the kernel object 494 */ 495 496 if (flags & UAO_FLAG_KERNOBJ) { 497 KASSERT(!kobj_alloced); 498 aobj = &kernel_object_store; 499 aobj->u_pages = pages; 500 aobj->u_flags = UAO_FLAG_NOSWAP; 501 aobj->u_obj.uo_refs = UVM_OBJ_KERN; 502 kobj_alloced = UAO_FLAG_KERNOBJ; 503 } else if (flags & UAO_FLAG_KERNSWAP) { 504 KASSERT(kobj_alloced == UAO_FLAG_KERNOBJ); 505 aobj = &kernel_object_store; 506 kobj_alloced = UAO_FLAG_KERNSWAP; 507 } else { 508 aobj = pool_get(&uvm_aobj_pool, PR_WAITOK); 509 aobj->u_pages = pages; 510 aobj->u_flags = 0; 511 aobj->u_obj.uo_refs = 1; 512 } 513 514 /* 515 * allocate hash/array if necessary 516 * 517 * note: in the KERNSWAP case no need to worry about locking since 518 * we are still booting we should be the only thread around. 519 */ 520 521 if (flags == 0 || (flags & UAO_FLAG_KERNSWAP) != 0) { 522 int mflags = (flags & UAO_FLAG_KERNSWAP) != 0 ? 523 M_NOWAIT : M_WAITOK; 524 525 /* allocate hash table or array depending on object size */ 526 if (UAO_USES_SWHASH(aobj)) { 527 aobj->u_swhash = hashinit(UAO_SWHASH_BUCKETS(aobj), 528 HASH_LIST, M_UVMAOBJ, mflags, &aobj->u_swhashmask); 529 if (aobj->u_swhash == NULL) 530 panic("uao_create: hashinit swhash failed"); 531 } else { 532 aobj->u_swslots = malloc(pages * sizeof(int), 533 M_UVMAOBJ, mflags); 534 if (aobj->u_swslots == NULL) 535 panic("uao_create: malloc swslots failed"); 536 memset(aobj->u_swslots, 0, pages * sizeof(int)); 537 } 538 539 if (flags) { 540 aobj->u_flags &= ~UAO_FLAG_NOSWAP; /* clear noswap */ 541 return(&aobj->u_obj); 542 } 543 } 544 545 /* 546 * init aobj fields 547 */ 548 549 simple_lock_init(&aobj->u_obj.vmobjlock); 550 aobj->u_obj.pgops = &aobj_pager; 551 TAILQ_INIT(&aobj->u_obj.memq); 552 aobj->u_obj.uo_npages = 0; 553 554 /* 555 * now that aobj is ready, add it to the global list 556 */ 557 558 simple_lock(&uao_list_lock); 559 LIST_INSERT_HEAD(&uao_list, aobj, u_list); 560 simple_unlock(&uao_list_lock); 561 return(&aobj->u_obj); 562 } 563 564 565 566 /* 567 * uao_init: set up aobj pager subsystem 568 * 569 * => called at boot time from uvm_pager_init() 570 */ 571 572 void 573 uao_init(void) 574 { 575 static int uao_initialized; 576 577 if (uao_initialized) 578 return; 579 uao_initialized = TRUE; 580 LIST_INIT(&uao_list); 581 simple_lock_init(&uao_list_lock); 582 583 /* 584 * NOTE: Pages fror this pool must not come from a pageable 585 * kernel map! 586 */ 587 588 pool_init(&uao_swhash_elt_pool, sizeof(struct uao_swhash_elt), 589 0, 0, 0, "uaoeltpl", NULL); 590 pool_init(&uvm_aobj_pool, sizeof(struct uvm_aobj), 0, 0, 0, 591 "aobjpl", &pool_allocator_nointr); 592 } 593 594 /* 595 * uao_reference: add a ref to an aobj 596 * 597 * => aobj must be unlocked 598 * => just lock it and call the locked version 599 */ 600 601 void 602 uao_reference(uobj) 603 struct uvm_object *uobj; 604 { 605 simple_lock(&uobj->vmobjlock); 606 uao_reference_locked(uobj); 607 simple_unlock(&uobj->vmobjlock); 608 } 609 610 /* 611 * uao_reference_locked: add a ref to an aobj that is already locked 612 * 613 * => aobj must be locked 614 * this needs to be separate from the normal routine 615 * since sometimes we need to add a reference to an aobj when 616 * it's already locked. 617 */ 618 619 void 620 uao_reference_locked(uobj) 621 struct uvm_object *uobj; 622 { 623 UVMHIST_FUNC("uao_reference"); UVMHIST_CALLED(maphist); 624 625 /* 626 * kernel_object already has plenty of references, leave it alone. 627 */ 628 629 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) 630 return; 631 632 uobj->uo_refs++; 633 UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)", 634 uobj, uobj->uo_refs,0,0); 635 } 636 637 /* 638 * uao_detach: drop a reference to an aobj 639 * 640 * => aobj must be unlocked 641 * => just lock it and call the locked version 642 */ 643 644 void 645 uao_detach(uobj) 646 struct uvm_object *uobj; 647 { 648 simple_lock(&uobj->vmobjlock); 649 uao_detach_locked(uobj); 650 } 651 652 /* 653 * uao_detach_locked: drop a reference to an aobj 654 * 655 * => aobj must be locked, and is unlocked (or freed) upon return. 656 * this needs to be separate from the normal routine 657 * since sometimes we need to detach from an aobj when 658 * it's already locked. 659 */ 660 661 void 662 uao_detach_locked(uobj) 663 struct uvm_object *uobj; 664 { 665 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 666 struct vm_page *pg; 667 UVMHIST_FUNC("uao_detach"); UVMHIST_CALLED(maphist); 668 669 /* 670 * detaching from kernel_object is a noop. 671 */ 672 673 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) { 674 simple_unlock(&uobj->vmobjlock); 675 return; 676 } 677 678 UVMHIST_LOG(maphist," (uobj=0x%x) ref=%d", uobj,uobj->uo_refs,0,0); 679 uobj->uo_refs--; 680 if (uobj->uo_refs) { 681 simple_unlock(&uobj->vmobjlock); 682 UVMHIST_LOG(maphist, "<- done (rc>0)", 0,0,0,0); 683 return; 684 } 685 686 /* 687 * remove the aobj from the global list. 688 */ 689 690 simple_lock(&uao_list_lock); 691 LIST_REMOVE(aobj, u_list); 692 simple_unlock(&uao_list_lock); 693 694 /* 695 * free all the pages left in the aobj. for each page, 696 * when the page is no longer busy (and thus after any disk i/o that 697 * it's involved in is complete), release any swap resources and 698 * free the page itself. 699 */ 700 701 uvm_lock_pageq(); 702 while ((pg = TAILQ_FIRST(&uobj->memq)) != NULL) { 703 pmap_page_protect(pg, VM_PROT_NONE); 704 if (pg->flags & PG_BUSY) { 705 pg->flags |= PG_WANTED; 706 uvm_unlock_pageq(); 707 UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, FALSE, 708 "uao_det", 0); 709 simple_lock(&uobj->vmobjlock); 710 uvm_lock_pageq(); 711 continue; 712 } 713 uao_dropswap(&aobj->u_obj, pg->offset >> PAGE_SHIFT); 714 uvm_pagefree(pg); 715 } 716 uvm_unlock_pageq(); 717 718 /* 719 * finally, free the aobj itself. 720 */ 721 722 uao_free(aobj); 723 } 724 725 /* 726 * uao_put: flush pages out of a uvm object 727 * 728 * => object should be locked by caller. we may _unlock_ the object 729 * if (and only if) we need to clean a page (PGO_CLEANIT). 730 * XXXJRT Currently, however, we don't. In the case of cleaning 731 * XXXJRT a page, we simply just deactivate it. Should probably 732 * XXXJRT handle this better, in the future (although "flushing" 733 * XXXJRT anonymous memory isn't terribly important). 734 * => if PGO_CLEANIT is not set, then we will neither unlock the object 735 * or block. 736 * => if PGO_ALLPAGE is set, then all pages in the object are valid targets 737 * for flushing. 738 * => NOTE: we rely on the fact that the object's memq is a TAILQ and 739 * that new pages are inserted on the tail end of the list. thus, 740 * we can make a complete pass through the object in one go by starting 741 * at the head and working towards the tail (new pages are put in 742 * front of us). 743 * => NOTE: we are allowed to lock the page queues, so the caller 744 * must not be holding the lock on them [e.g. pagedaemon had 745 * better not call us with the queues locked] 746 * => we return TRUE unless we encountered some sort of I/O error 747 * XXXJRT currently never happens, as we never directly initiate 748 * XXXJRT I/O 749 * 750 * note on page traversal: 751 * we can traverse the pages in an object either by going down the 752 * linked list in "uobj->memq", or we can go over the address range 753 * by page doing hash table lookups for each address. depending 754 * on how many pages are in the object it may be cheaper to do one 755 * or the other. we set "by_list" to true if we are using memq. 756 * if the cost of a hash lookup was equal to the cost of the list 757 * traversal we could compare the number of pages in the start->stop 758 * range to the total number of pages in the object. however, it 759 * seems that a hash table lookup is more expensive than the linked 760 * list traversal, so we multiply the number of pages in the 761 * start->stop range by a penalty which we define below. 762 */ 763 764 int 765 uao_put(uobj, start, stop, flags) 766 struct uvm_object *uobj; 767 voff_t start, stop; 768 int flags; 769 { 770 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 771 struct vm_page *pg, *nextpg, curmp, endmp; 772 boolean_t by_list; 773 voff_t curoff; 774 UVMHIST_FUNC("uao_put"); UVMHIST_CALLED(maphist); 775 776 curoff = 0; 777 if (flags & PGO_ALLPAGES) { 778 start = 0; 779 stop = aobj->u_pages << PAGE_SHIFT; 780 by_list = TRUE; /* always go by the list */ 781 } else { 782 start = trunc_page(start); 783 stop = round_page(stop); 784 if (stop > (aobj->u_pages << PAGE_SHIFT)) { 785 printf("uao_flush: strange, got an out of range " 786 "flush (fixed)\n"); 787 stop = aobj->u_pages << PAGE_SHIFT; 788 } 789 by_list = (uobj->uo_npages <= 790 ((stop - start) >> PAGE_SHIFT) * UVM_PAGE_HASH_PENALTY); 791 } 792 UVMHIST_LOG(maphist, 793 " flush start=0x%lx, stop=0x%x, by_list=%d, flags=0x%x", 794 start, stop, by_list, flags); 795 796 /* 797 * Don't need to do any work here if we're not freeing 798 * or deactivating pages. 799 */ 800 801 if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) { 802 simple_unlock(&uobj->vmobjlock); 803 return 0; 804 } 805 806 /* 807 * Initialize the marker pages. See the comment in 808 * genfs_putpages() also. 809 */ 810 811 curmp.uobject = uobj; 812 curmp.offset = (voff_t)-1; 813 curmp.flags = PG_BUSY; 814 endmp.uobject = uobj; 815 endmp.offset = (voff_t)-1; 816 endmp.flags = PG_BUSY; 817 818 /* 819 * now do it. note: we must update nextpg in the body of loop or we 820 * will get stuck. we need to use nextpg if we'll traverse the list 821 * because we may free "pg" before doing the next loop. 822 */ 823 824 if (by_list) { 825 TAILQ_INSERT_TAIL(&uobj->memq, &endmp, listq); 826 nextpg = TAILQ_FIRST(&uobj->memq); 827 PHOLD(curlwp); 828 } else { 829 curoff = start; 830 nextpg = NULL; /* Quell compiler warning */ 831 } 832 833 uvm_lock_pageq(); 834 835 /* locked: both page queues and uobj */ 836 for (;;) { 837 if (by_list) { 838 pg = nextpg; 839 if (pg == &endmp) 840 break; 841 nextpg = TAILQ_NEXT(pg, listq); 842 if (pg->offset < start || pg->offset >= stop) 843 continue; 844 } else { 845 if (curoff < stop) { 846 pg = uvm_pagelookup(uobj, curoff); 847 curoff += PAGE_SIZE; 848 } else 849 break; 850 if (pg == NULL) 851 continue; 852 } 853 switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) { 854 855 /* 856 * XXX In these first 3 cases, we always just 857 * XXX deactivate the page. We may want to 858 * XXX handle the different cases more specifically 859 * XXX in the future. 860 */ 861 862 case PGO_CLEANIT|PGO_FREE: 863 case PGO_CLEANIT|PGO_DEACTIVATE: 864 case PGO_DEACTIVATE: 865 deactivate_it: 866 /* skip the page if it's loaned or wired */ 867 if (pg->loan_count != 0 || pg->wire_count != 0) 868 continue; 869 870 /* ...and deactivate the page. */ 871 pmap_clear_reference(pg); 872 uvm_pagedeactivate(pg); 873 continue; 874 875 case PGO_FREE: 876 877 /* 878 * If there are multiple references to 879 * the object, just deactivate the page. 880 */ 881 882 if (uobj->uo_refs > 1) 883 goto deactivate_it; 884 885 /* XXX skip the page if it's loaned or wired */ 886 if (pg->loan_count != 0 || pg->wire_count != 0) 887 continue; 888 889 /* 890 * wait and try again if the page is busy. 891 * otherwise free the swap slot and the page. 892 */ 893 894 pmap_page_protect(pg, VM_PROT_NONE); 895 if (pg->flags & PG_BUSY) { 896 if (by_list) { 897 TAILQ_INSERT_BEFORE(pg, &curmp, listq); 898 } 899 pg->flags |= PG_WANTED; 900 uvm_unlock_pageq(); 901 UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, 0, 902 "uao_put", 0); 903 simple_lock(&uobj->vmobjlock); 904 uvm_lock_pageq(); 905 if (by_list) { 906 nextpg = TAILQ_NEXT(&curmp, listq); 907 TAILQ_REMOVE(&uobj->memq, &curmp, 908 listq); 909 } else 910 curoff -= PAGE_SIZE; 911 continue; 912 } 913 uao_dropswap(uobj, pg->offset >> PAGE_SHIFT); 914 uvm_pagefree(pg); 915 continue; 916 } 917 } 918 uvm_unlock_pageq(); 919 if (by_list) { 920 TAILQ_REMOVE(&uobj->memq, &endmp, listq); 921 PRELE(curlwp); 922 } 923 simple_unlock(&uobj->vmobjlock); 924 return 0; 925 } 926 927 /* 928 * uao_get: fetch me a page 929 * 930 * we have three cases: 931 * 1: page is resident -> just return the page. 932 * 2: page is zero-fill -> allocate a new page and zero it. 933 * 3: page is swapped out -> fetch the page from swap. 934 * 935 * cases 1 and 2 can be handled with PGO_LOCKED, case 3 cannot. 936 * so, if the "center" page hits case 3 (or any page, with PGO_ALLPAGES), 937 * then we will need to return EBUSY. 938 * 939 * => prefer map unlocked (not required) 940 * => object must be locked! we will _unlock_ it before starting any I/O. 941 * => flags: PGO_ALLPAGES: get all of the pages 942 * PGO_LOCKED: fault data structures are locked 943 * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx] 944 * => NOTE: caller must check for released pages!! 945 */ 946 947 static int 948 uao_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags) 949 struct uvm_object *uobj; 950 voff_t offset; 951 struct vm_page **pps; 952 int *npagesp; 953 int centeridx, advice, flags; 954 vm_prot_t access_type; 955 { 956 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 957 voff_t current_offset; 958 struct vm_page *ptmp = NULL; /* Quell compiler warning */ 959 int lcv, gotpages, maxpages, swslot, error, pageidx; 960 boolean_t done; 961 UVMHIST_FUNC("uao_get"); UVMHIST_CALLED(pdhist); 962 963 UVMHIST_LOG(pdhist, "aobj=%p offset=%d, flags=%d", 964 aobj, offset, flags,0); 965 966 /* 967 * get number of pages 968 */ 969 970 maxpages = *npagesp; 971 972 /* 973 * step 1: handled the case where fault data structures are locked. 974 */ 975 976 if (flags & PGO_LOCKED) { 977 978 /* 979 * step 1a: get pages that are already resident. only do 980 * this if the data structures are locked (i.e. the first 981 * time through). 982 */ 983 984 done = TRUE; /* be optimistic */ 985 gotpages = 0; /* # of pages we got so far */ 986 for (lcv = 0, current_offset = offset ; lcv < maxpages ; 987 lcv++, current_offset += PAGE_SIZE) { 988 /* do we care about this page? if not, skip it */ 989 if (pps[lcv] == PGO_DONTCARE) 990 continue; 991 ptmp = uvm_pagelookup(uobj, current_offset); 992 993 /* 994 * if page is new, attempt to allocate the page, 995 * zero-fill'd. 996 */ 997 998 if (ptmp == NULL && uao_find_swslot(&aobj->u_obj, 999 current_offset >> PAGE_SHIFT) == 0) { 1000 ptmp = uvm_pagealloc(uobj, current_offset, 1001 NULL, UVM_PGA_ZERO); 1002 if (ptmp) { 1003 /* new page */ 1004 ptmp->flags &= ~(PG_FAKE); 1005 ptmp->pqflags |= PQ_AOBJ; 1006 goto gotpage; 1007 } 1008 } 1009 1010 /* 1011 * to be useful must get a non-busy page 1012 */ 1013 1014 if (ptmp == NULL || (ptmp->flags & PG_BUSY) != 0) { 1015 if (lcv == centeridx || 1016 (flags & PGO_ALLPAGES) != 0) 1017 /* need to do a wait or I/O! */ 1018 done = FALSE; 1019 continue; 1020 } 1021 1022 /* 1023 * useful page: busy/lock it and plug it in our 1024 * result array 1025 */ 1026 1027 /* caller must un-busy this page */ 1028 ptmp->flags |= PG_BUSY; 1029 UVM_PAGE_OWN(ptmp, "uao_get1"); 1030 gotpage: 1031 pps[lcv] = ptmp; 1032 gotpages++; 1033 } 1034 1035 /* 1036 * step 1b: now we've either done everything needed or we 1037 * to unlock and do some waiting or I/O. 1038 */ 1039 1040 UVMHIST_LOG(pdhist, "<- done (done=%d)", done, 0,0,0); 1041 *npagesp = gotpages; 1042 if (done) 1043 return 0; 1044 else 1045 return EBUSY; 1046 } 1047 1048 /* 1049 * step 2: get non-resident or busy pages. 1050 * object is locked. data structures are unlocked. 1051 */ 1052 1053 for (lcv = 0, current_offset = offset ; lcv < maxpages ; 1054 lcv++, current_offset += PAGE_SIZE) { 1055 1056 /* 1057 * - skip over pages we've already gotten or don't want 1058 * - skip over pages we don't _have_ to get 1059 */ 1060 1061 if (pps[lcv] != NULL || 1062 (lcv != centeridx && (flags & PGO_ALLPAGES) == 0)) 1063 continue; 1064 1065 pageidx = current_offset >> PAGE_SHIFT; 1066 1067 /* 1068 * we have yet to locate the current page (pps[lcv]). we 1069 * first look for a page that is already at the current offset. 1070 * if we find a page, we check to see if it is busy or 1071 * released. if that is the case, then we sleep on the page 1072 * until it is no longer busy or released and repeat the lookup. 1073 * if the page we found is neither busy nor released, then we 1074 * busy it (so we own it) and plug it into pps[lcv]. this 1075 * 'break's the following while loop and indicates we are 1076 * ready to move on to the next page in the "lcv" loop above. 1077 * 1078 * if we exit the while loop with pps[lcv] still set to NULL, 1079 * then it means that we allocated a new busy/fake/clean page 1080 * ptmp in the object and we need to do I/O to fill in the data. 1081 */ 1082 1083 /* top of "pps" while loop */ 1084 while (pps[lcv] == NULL) { 1085 /* look for a resident page */ 1086 ptmp = uvm_pagelookup(uobj, current_offset); 1087 1088 /* not resident? allocate one now (if we can) */ 1089 if (ptmp == NULL) { 1090 1091 ptmp = uvm_pagealloc(uobj, current_offset, 1092 NULL, 0); 1093 1094 /* out of RAM? */ 1095 if (ptmp == NULL) { 1096 simple_unlock(&uobj->vmobjlock); 1097 UVMHIST_LOG(pdhist, 1098 "sleeping, ptmp == NULL\n",0,0,0,0); 1099 uvm_wait("uao_getpage"); 1100 simple_lock(&uobj->vmobjlock); 1101 continue; 1102 } 1103 1104 /* 1105 * safe with PQ's unlocked: because we just 1106 * alloc'd the page 1107 */ 1108 1109 ptmp->pqflags |= PQ_AOBJ; 1110 1111 /* 1112 * got new page ready for I/O. break pps while 1113 * loop. pps[lcv] is still NULL. 1114 */ 1115 1116 break; 1117 } 1118 1119 /* page is there, see if we need to wait on it */ 1120 if ((ptmp->flags & PG_BUSY) != 0) { 1121 ptmp->flags |= PG_WANTED; 1122 UVMHIST_LOG(pdhist, 1123 "sleeping, ptmp->flags 0x%x\n", 1124 ptmp->flags,0,0,0); 1125 UVM_UNLOCK_AND_WAIT(ptmp, &uobj->vmobjlock, 1126 FALSE, "uao_get", 0); 1127 simple_lock(&uobj->vmobjlock); 1128 continue; 1129 } 1130 1131 /* 1132 * if we get here then the page has become resident and 1133 * unbusy between steps 1 and 2. we busy it now (so we 1134 * own it) and set pps[lcv] (so that we exit the while 1135 * loop). 1136 */ 1137 1138 /* we own it, caller must un-busy */ 1139 ptmp->flags |= PG_BUSY; 1140 UVM_PAGE_OWN(ptmp, "uao_get2"); 1141 pps[lcv] = ptmp; 1142 } 1143 1144 /* 1145 * if we own the valid page at the correct offset, pps[lcv] will 1146 * point to it. nothing more to do except go to the next page. 1147 */ 1148 1149 if (pps[lcv]) 1150 continue; /* next lcv */ 1151 1152 /* 1153 * we have a "fake/busy/clean" page that we just allocated. 1154 * do the needed "i/o", either reading from swap or zeroing. 1155 */ 1156 1157 swslot = uao_find_swslot(&aobj->u_obj, pageidx); 1158 1159 /* 1160 * just zero the page if there's nothing in swap. 1161 */ 1162 1163 if (swslot == 0) { 1164 1165 /* 1166 * page hasn't existed before, just zero it. 1167 */ 1168 1169 uvm_pagezero(ptmp); 1170 } else { 1171 UVMHIST_LOG(pdhist, "pagein from swslot %d", 1172 swslot, 0,0,0); 1173 1174 /* 1175 * page in the swapped-out page. 1176 * unlock object for i/o, relock when done. 1177 */ 1178 1179 simple_unlock(&uobj->vmobjlock); 1180 error = uvm_swap_get(ptmp, swslot, PGO_SYNCIO); 1181 simple_lock(&uobj->vmobjlock); 1182 1183 /* 1184 * I/O done. check for errors. 1185 */ 1186 1187 if (error != 0) { 1188 UVMHIST_LOG(pdhist, "<- done (error=%d)", 1189 error,0,0,0); 1190 if (ptmp->flags & PG_WANTED) 1191 wakeup(ptmp); 1192 1193 /* 1194 * remove the swap slot from the aobj 1195 * and mark the aobj as having no real slot. 1196 * don't free the swap slot, thus preventing 1197 * it from being used again. 1198 */ 1199 1200 swslot = uao_set_swslot(&aobj->u_obj, pageidx, 1201 SWSLOT_BAD); 1202 if (swslot != -1) { 1203 uvm_swap_markbad(swslot, 1); 1204 } 1205 1206 uvm_lock_pageq(); 1207 uvm_pagefree(ptmp); 1208 uvm_unlock_pageq(); 1209 simple_unlock(&uobj->vmobjlock); 1210 return error; 1211 } 1212 } 1213 1214 /* 1215 * we got the page! clear the fake flag (indicates valid 1216 * data now in page) and plug into our result array. note 1217 * that page is still busy. 1218 * 1219 * it is the callers job to: 1220 * => check if the page is released 1221 * => unbusy the page 1222 * => activate the page 1223 */ 1224 1225 ptmp->flags &= ~PG_FAKE; 1226 pps[lcv] = ptmp; 1227 } 1228 1229 /* 1230 * finally, unlock object and return. 1231 */ 1232 1233 simple_unlock(&uobj->vmobjlock); 1234 UVMHIST_LOG(pdhist, "<- done (OK)",0,0,0,0); 1235 return 0; 1236 } 1237 1238 /* 1239 * uao_dropswap: release any swap resources from this aobj page. 1240 * 1241 * => aobj must be locked or have a reference count of 0. 1242 */ 1243 1244 void 1245 uao_dropswap(uobj, pageidx) 1246 struct uvm_object *uobj; 1247 int pageidx; 1248 { 1249 int slot; 1250 1251 slot = uao_set_swslot(uobj, pageidx, 0); 1252 if (slot) { 1253 uvm_swap_free(slot, 1); 1254 } 1255 } 1256 1257 /* 1258 * page in every page in every aobj that is paged-out to a range of swslots. 1259 * 1260 * => nothing should be locked. 1261 * => returns TRUE if pagein was aborted due to lack of memory. 1262 */ 1263 1264 boolean_t 1265 uao_swap_off(startslot, endslot) 1266 int startslot, endslot; 1267 { 1268 struct uvm_aobj *aobj, *nextaobj; 1269 boolean_t rv; 1270 1271 /* 1272 * walk the list of all aobjs. 1273 */ 1274 1275 restart: 1276 simple_lock(&uao_list_lock); 1277 for (aobj = LIST_FIRST(&uao_list); 1278 aobj != NULL; 1279 aobj = nextaobj) { 1280 1281 /* 1282 * try to get the object lock, start all over if we fail. 1283 * most of the time we'll get the aobj lock, 1284 * so this should be a rare case. 1285 */ 1286 1287 if (!simple_lock_try(&aobj->u_obj.vmobjlock)) { 1288 simple_unlock(&uao_list_lock); 1289 goto restart; 1290 } 1291 1292 /* 1293 * add a ref to the aobj so it doesn't disappear 1294 * while we're working. 1295 */ 1296 1297 uao_reference_locked(&aobj->u_obj); 1298 1299 /* 1300 * now it's safe to unlock the uao list. 1301 */ 1302 1303 simple_unlock(&uao_list_lock); 1304 1305 /* 1306 * page in any pages in the swslot range. 1307 * if there's an error, abort and return the error. 1308 */ 1309 1310 rv = uao_pagein(aobj, startslot, endslot); 1311 if (rv) { 1312 uao_detach_locked(&aobj->u_obj); 1313 return rv; 1314 } 1315 1316 /* 1317 * we're done with this aobj. 1318 * relock the list and drop our ref on the aobj. 1319 */ 1320 1321 simple_lock(&uao_list_lock); 1322 nextaobj = LIST_NEXT(aobj, u_list); 1323 uao_detach_locked(&aobj->u_obj); 1324 } 1325 1326 /* 1327 * done with traversal, unlock the list 1328 */ 1329 simple_unlock(&uao_list_lock); 1330 return FALSE; 1331 } 1332 1333 1334 /* 1335 * page in any pages from aobj in the given range. 1336 * 1337 * => aobj must be locked and is returned locked. 1338 * => returns TRUE if pagein was aborted due to lack of memory. 1339 */ 1340 static boolean_t 1341 uao_pagein(aobj, startslot, endslot) 1342 struct uvm_aobj *aobj; 1343 int startslot, endslot; 1344 { 1345 boolean_t rv; 1346 1347 if (UAO_USES_SWHASH(aobj)) { 1348 struct uao_swhash_elt *elt; 1349 int bucket; 1350 1351 restart: 1352 for (bucket = aobj->u_swhashmask; bucket >= 0; bucket--) { 1353 for (elt = LIST_FIRST(&aobj->u_swhash[bucket]); 1354 elt != NULL; 1355 elt = LIST_NEXT(elt, list)) { 1356 int i; 1357 1358 for (i = 0; i < UAO_SWHASH_CLUSTER_SIZE; i++) { 1359 int slot = elt->slots[i]; 1360 1361 /* 1362 * if the slot isn't in range, skip it. 1363 */ 1364 1365 if (slot < startslot || 1366 slot >= endslot) { 1367 continue; 1368 } 1369 1370 /* 1371 * process the page, 1372 * the start over on this object 1373 * since the swhash elt 1374 * may have been freed. 1375 */ 1376 1377 rv = uao_pagein_page(aobj, 1378 UAO_SWHASH_ELT_PAGEIDX_BASE(elt) + i); 1379 if (rv) { 1380 return rv; 1381 } 1382 goto restart; 1383 } 1384 } 1385 } 1386 } else { 1387 int i; 1388 1389 for (i = 0; i < aobj->u_pages; i++) { 1390 int slot = aobj->u_swslots[i]; 1391 1392 /* 1393 * if the slot isn't in range, skip it 1394 */ 1395 1396 if (slot < startslot || slot >= endslot) { 1397 continue; 1398 } 1399 1400 /* 1401 * process the page. 1402 */ 1403 1404 rv = uao_pagein_page(aobj, i); 1405 if (rv) { 1406 return rv; 1407 } 1408 } 1409 } 1410 1411 return FALSE; 1412 } 1413 1414 /* 1415 * page in a page from an aobj. used for swap_off. 1416 * returns TRUE if pagein was aborted due to lack of memory. 1417 * 1418 * => aobj must be locked and is returned locked. 1419 */ 1420 1421 static boolean_t 1422 uao_pagein_page(aobj, pageidx) 1423 struct uvm_aobj *aobj; 1424 int pageidx; 1425 { 1426 struct vm_page *pg; 1427 int rv, slot, npages; 1428 1429 pg = NULL; 1430 npages = 1; 1431 /* locked: aobj */ 1432 rv = uao_get(&aobj->u_obj, pageidx << PAGE_SHIFT, 1433 &pg, &npages, 0, VM_PROT_READ|VM_PROT_WRITE, 0, 0); 1434 /* unlocked: aobj */ 1435 1436 /* 1437 * relock and finish up. 1438 */ 1439 1440 simple_lock(&aobj->u_obj.vmobjlock); 1441 switch (rv) { 1442 case 0: 1443 break; 1444 1445 case EIO: 1446 case ERESTART: 1447 1448 /* 1449 * nothing more to do on errors. 1450 * ERESTART can only mean that the anon was freed, 1451 * so again there's nothing to do. 1452 */ 1453 1454 return FALSE; 1455 } 1456 1457 /* 1458 * ok, we've got the page now. 1459 * mark it as dirty, clear its swslot and un-busy it. 1460 */ 1461 1462 slot = uao_set_swslot(&aobj->u_obj, pageidx, 0); 1463 uvm_swap_free(slot, 1); 1464 1465 /* 1466 * deactivate the page (to make sure it's on a page queue). 1467 */ 1468 1469 uvm_lock_pageq(); 1470 uvm_pagedeactivate(pg); 1471 uvm_unlock_pageq(); 1472 1473 pg->flags &= ~(PG_BUSY|PG_CLEAN|PG_FAKE); 1474 UVM_PAGE_OWN(pg, NULL); 1475 1476 return FALSE; 1477 } 1478