1 /* $NetBSD: uvm_aobj.c,v 1.64 2004/04/25 16:42:44 simonb 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 1.64 2004/04/25 16:42:44 simonb 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 * NOTE: Pages for this pool must not come from a pageable kernel map! 141 */ 142 POOL_INIT(uao_swhash_elt_pool, sizeof(struct uao_swhash_elt), 0, 0, 0, 143 "uaoeltpl", NULL); 144 145 /* 146 * uvm_aobj: the actual anon-backed uvm_object 147 * 148 * => the uvm_object is at the top of the structure, this allows 149 * (struct uvm_aobj *) == (struct uvm_object *) 150 * => only one of u_swslots and u_swhash is used in any given aobj 151 */ 152 153 struct uvm_aobj { 154 struct uvm_object u_obj; /* has: lock, pgops, memq, #pages, #refs */ 155 int u_pages; /* number of pages in entire object */ 156 int u_flags; /* the flags (see uvm_aobj.h) */ 157 int *u_swslots; /* array of offset->swapslot mappings */ 158 /* 159 * hashtable of offset->swapslot mappings 160 * (u_swhash is an array of bucket heads) 161 */ 162 struct uao_swhash *u_swhash; 163 u_long u_swhashmask; /* mask for hashtable */ 164 LIST_ENTRY(uvm_aobj) u_list; /* global list of aobjs */ 165 }; 166 167 /* 168 * uvm_aobj_pool: pool of uvm_aobj structures 169 */ 170 POOL_INIT(uvm_aobj_pool, sizeof(struct uvm_aobj), 0, 0, 0, "aobjpl", 171 &pool_allocator_nointr); 172 173 MALLOC_DEFINE(M_UVMAOBJ, "UVM aobj", "UVM aobj and related structures"); 174 175 /* 176 * local functions 177 */ 178 179 static struct uao_swhash_elt *uao_find_swhash_elt 180 (struct uvm_aobj *, int, boolean_t); 181 182 static void uao_free(struct uvm_aobj *); 183 static int uao_get(struct uvm_object *, voff_t, struct vm_page **, 184 int *, int, vm_prot_t, int, int); 185 static boolean_t uao_put(struct uvm_object *, voff_t, voff_t, int); 186 static boolean_t uao_pagein(struct uvm_aobj *, int, int); 187 static boolean_t uao_pagein_page(struct uvm_aobj *, int); 188 189 /* 190 * aobj_pager 191 * 192 * note that some functions (e.g. put) are handled elsewhere 193 */ 194 195 struct uvm_pagerops aobj_pager = { 196 NULL, /* init */ 197 uao_reference, /* reference */ 198 uao_detach, /* detach */ 199 NULL, /* fault */ 200 uao_get, /* get */ 201 uao_put, /* flush */ 202 }; 203 204 /* 205 * uao_list: global list of active aobjs, locked by uao_list_lock 206 */ 207 208 static LIST_HEAD(aobjlist, uvm_aobj) uao_list; 209 static struct simplelock uao_list_lock; 210 211 /* 212 * functions 213 */ 214 215 /* 216 * hash table/array related functions 217 */ 218 219 /* 220 * uao_find_swhash_elt: find (or create) a hash table entry for a page 221 * offset. 222 * 223 * => the object should be locked by the caller 224 */ 225 226 static struct uao_swhash_elt * 227 uao_find_swhash_elt(aobj, pageidx, create) 228 struct uvm_aobj *aobj; 229 int pageidx; 230 boolean_t create; 231 { 232 struct uao_swhash *swhash; 233 struct uao_swhash_elt *elt; 234 voff_t page_tag; 235 236 swhash = UAO_SWHASH_HASH(aobj, pageidx); 237 page_tag = UAO_SWHASH_ELT_TAG(pageidx); 238 239 /* 240 * now search the bucket for the requested tag 241 */ 242 243 LIST_FOREACH(elt, swhash, list) { 244 if (elt->tag == page_tag) { 245 return elt; 246 } 247 } 248 if (!create) { 249 return NULL; 250 } 251 252 /* 253 * allocate a new entry for the bucket and init/insert it in 254 */ 255 256 elt = pool_get(&uao_swhash_elt_pool, PR_NOWAIT); 257 if (elt == NULL) { 258 return NULL; 259 } 260 LIST_INSERT_HEAD(swhash, elt, list); 261 elt->tag = page_tag; 262 elt->count = 0; 263 memset(elt->slots, 0, sizeof(elt->slots)); 264 return elt; 265 } 266 267 /* 268 * uao_find_swslot: find the swap slot number for an aobj/pageidx 269 * 270 * => object must be locked by caller 271 */ 272 273 int 274 uao_find_swslot(uobj, pageidx) 275 struct uvm_object *uobj; 276 int pageidx; 277 { 278 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 279 struct uao_swhash_elt *elt; 280 281 /* 282 * if noswap flag is set, then we never return a slot 283 */ 284 285 if (aobj->u_flags & UAO_FLAG_NOSWAP) 286 return(0); 287 288 /* 289 * if hashing, look in hash table. 290 */ 291 292 if (UAO_USES_SWHASH(aobj)) { 293 elt = uao_find_swhash_elt(aobj, pageidx, FALSE); 294 if (elt) 295 return(UAO_SWHASH_ELT_PAGESLOT(elt, pageidx)); 296 else 297 return(0); 298 } 299 300 /* 301 * otherwise, look in the array 302 */ 303 304 return(aobj->u_swslots[pageidx]); 305 } 306 307 /* 308 * uao_set_swslot: set the swap slot for a page in an aobj. 309 * 310 * => setting a slot to zero frees the slot 311 * => object must be locked by caller 312 * => we return the old slot number, or -1 if we failed to allocate 313 * memory to record the new slot number 314 */ 315 316 int 317 uao_set_swslot(uobj, pageidx, slot) 318 struct uvm_object *uobj; 319 int pageidx, slot; 320 { 321 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 322 struct uao_swhash_elt *elt; 323 int oldslot; 324 UVMHIST_FUNC("uao_set_swslot"); UVMHIST_CALLED(pdhist); 325 UVMHIST_LOG(pdhist, "aobj %p pageidx %d slot %d", 326 aobj, pageidx, slot, 0); 327 328 /* 329 * if noswap flag is set, then we can't set a non-zero slot. 330 */ 331 332 if (aobj->u_flags & UAO_FLAG_NOSWAP) { 333 if (slot == 0) 334 return(0); 335 336 printf("uao_set_swslot: uobj = %p\n", uobj); 337 panic("uao_set_swslot: NOSWAP object"); 338 } 339 340 /* 341 * are we using a hash table? if so, add it in the hash. 342 */ 343 344 if (UAO_USES_SWHASH(aobj)) { 345 346 /* 347 * Avoid allocating an entry just to free it again if 348 * the page had not swap slot in the first place, and 349 * we are freeing. 350 */ 351 352 elt = uao_find_swhash_elt(aobj, pageidx, slot != 0); 353 if (elt == NULL) { 354 return slot ? -1 : 0; 355 } 356 357 oldslot = UAO_SWHASH_ELT_PAGESLOT(elt, pageidx); 358 UAO_SWHASH_ELT_PAGESLOT(elt, pageidx) = slot; 359 360 /* 361 * now adjust the elt's reference counter and free it if we've 362 * dropped it to zero. 363 */ 364 365 if (slot) { 366 if (oldslot == 0) 367 elt->count++; 368 } else { 369 if (oldslot) 370 elt->count--; 371 372 if (elt->count == 0) { 373 LIST_REMOVE(elt, list); 374 pool_put(&uao_swhash_elt_pool, elt); 375 } 376 } 377 } else { 378 /* we are using an array */ 379 oldslot = aobj->u_swslots[pageidx]; 380 aobj->u_swslots[pageidx] = slot; 381 } 382 return (oldslot); 383 } 384 385 /* 386 * end of hash/array functions 387 */ 388 389 /* 390 * uao_free: free all resources held by an aobj, and then free the aobj 391 * 392 * => the aobj should be dead 393 */ 394 395 static void 396 uao_free(aobj) 397 struct uvm_aobj *aobj; 398 { 399 int swpgonlydelta = 0; 400 401 simple_unlock(&aobj->u_obj.vmobjlock); 402 if (UAO_USES_SWHASH(aobj)) { 403 int i, hashbuckets = aobj->u_swhashmask + 1; 404 405 /* 406 * free the swslots from each hash bucket, 407 * then the hash bucket, and finally the hash table itself. 408 */ 409 410 for (i = 0; i < hashbuckets; i++) { 411 struct uao_swhash_elt *elt, *next; 412 413 for (elt = LIST_FIRST(&aobj->u_swhash[i]); 414 elt != NULL; 415 elt = next) { 416 int j; 417 418 for (j = 0; j < UAO_SWHASH_CLUSTER_SIZE; j++) { 419 int slot = elt->slots[j]; 420 421 if (slot > 0) { 422 uvm_swap_free(slot, 1); 423 swpgonlydelta++; 424 } 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 > 0) { 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 /* 585 * uao_reference: add a ref to an aobj 586 * 587 * => aobj must be unlocked 588 * => just lock it and call the locked version 589 */ 590 591 void 592 uao_reference(uobj) 593 struct uvm_object *uobj; 594 { 595 simple_lock(&uobj->vmobjlock); 596 uao_reference_locked(uobj); 597 simple_unlock(&uobj->vmobjlock); 598 } 599 600 /* 601 * uao_reference_locked: add a ref to an aobj that is already locked 602 * 603 * => aobj must be locked 604 * this needs to be separate from the normal routine 605 * since sometimes we need to add a reference to an aobj when 606 * it's already locked. 607 */ 608 609 void 610 uao_reference_locked(uobj) 611 struct uvm_object *uobj; 612 { 613 UVMHIST_FUNC("uao_reference"); UVMHIST_CALLED(maphist); 614 615 /* 616 * kernel_object already has plenty of references, leave it alone. 617 */ 618 619 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) 620 return; 621 622 uobj->uo_refs++; 623 UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)", 624 uobj, uobj->uo_refs,0,0); 625 } 626 627 /* 628 * uao_detach: drop a reference to an aobj 629 * 630 * => aobj must be unlocked 631 * => just lock it and call the locked version 632 */ 633 634 void 635 uao_detach(uobj) 636 struct uvm_object *uobj; 637 { 638 simple_lock(&uobj->vmobjlock); 639 uao_detach_locked(uobj); 640 } 641 642 /* 643 * uao_detach_locked: drop a reference to an aobj 644 * 645 * => aobj must be locked, and is unlocked (or freed) upon return. 646 * this needs to be separate from the normal routine 647 * since sometimes we need to detach from an aobj when 648 * it's already locked. 649 */ 650 651 void 652 uao_detach_locked(uobj) 653 struct uvm_object *uobj; 654 { 655 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 656 struct vm_page *pg; 657 UVMHIST_FUNC("uao_detach"); UVMHIST_CALLED(maphist); 658 659 /* 660 * detaching from kernel_object is a noop. 661 */ 662 663 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) { 664 simple_unlock(&uobj->vmobjlock); 665 return; 666 } 667 668 UVMHIST_LOG(maphist," (uobj=0x%x) ref=%d", uobj,uobj->uo_refs,0,0); 669 uobj->uo_refs--; 670 if (uobj->uo_refs) { 671 simple_unlock(&uobj->vmobjlock); 672 UVMHIST_LOG(maphist, "<- done (rc>0)", 0,0,0,0); 673 return; 674 } 675 676 /* 677 * remove the aobj from the global list. 678 */ 679 680 simple_lock(&uao_list_lock); 681 LIST_REMOVE(aobj, u_list); 682 simple_unlock(&uao_list_lock); 683 684 /* 685 * free all the pages left in the aobj. for each page, 686 * when the page is no longer busy (and thus after any disk i/o that 687 * it's involved in is complete), release any swap resources and 688 * free the page itself. 689 */ 690 691 uvm_lock_pageq(); 692 while ((pg = TAILQ_FIRST(&uobj->memq)) != NULL) { 693 pmap_page_protect(pg, VM_PROT_NONE); 694 if (pg->flags & PG_BUSY) { 695 pg->flags |= PG_WANTED; 696 uvm_unlock_pageq(); 697 UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, FALSE, 698 "uao_det", 0); 699 simple_lock(&uobj->vmobjlock); 700 uvm_lock_pageq(); 701 continue; 702 } 703 uao_dropswap(&aobj->u_obj, pg->offset >> PAGE_SHIFT); 704 uvm_pagefree(pg); 705 } 706 uvm_unlock_pageq(); 707 708 /* 709 * finally, free the aobj itself. 710 */ 711 712 uao_free(aobj); 713 } 714 715 /* 716 * uao_put: flush pages out of a uvm object 717 * 718 * => object should be locked by caller. we may _unlock_ the object 719 * if (and only if) we need to clean a page (PGO_CLEANIT). 720 * XXXJRT Currently, however, we don't. In the case of cleaning 721 * XXXJRT a page, we simply just deactivate it. Should probably 722 * XXXJRT handle this better, in the future (although "flushing" 723 * XXXJRT anonymous memory isn't terribly important). 724 * => if PGO_CLEANIT is not set, then we will neither unlock the object 725 * or block. 726 * => if PGO_ALLPAGE is set, then all pages in the object are valid targets 727 * for flushing. 728 * => NOTE: we rely on the fact that the object's memq is a TAILQ and 729 * that new pages are inserted on the tail end of the list. thus, 730 * we can make a complete pass through the object in one go by starting 731 * at the head and working towards the tail (new pages are put in 732 * front of us). 733 * => NOTE: we are allowed to lock the page queues, so the caller 734 * must not be holding the lock on them [e.g. pagedaemon had 735 * better not call us with the queues locked] 736 * => we return TRUE unless we encountered some sort of I/O error 737 * XXXJRT currently never happens, as we never directly initiate 738 * XXXJRT I/O 739 * 740 * note on page traversal: 741 * we can traverse the pages in an object either by going down the 742 * linked list in "uobj->memq", or we can go over the address range 743 * by page doing hash table lookups for each address. depending 744 * on how many pages are in the object it may be cheaper to do one 745 * or the other. we set "by_list" to true if we are using memq. 746 * if the cost of a hash lookup was equal to the cost of the list 747 * traversal we could compare the number of pages in the start->stop 748 * range to the total number of pages in the object. however, it 749 * seems that a hash table lookup is more expensive than the linked 750 * list traversal, so we multiply the number of pages in the 751 * start->stop range by a penalty which we define below. 752 */ 753 754 int 755 uao_put(uobj, start, stop, flags) 756 struct uvm_object *uobj; 757 voff_t start, stop; 758 int flags; 759 { 760 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 761 struct vm_page *pg, *nextpg, curmp, endmp; 762 boolean_t by_list; 763 voff_t curoff; 764 UVMHIST_FUNC("uao_put"); UVMHIST_CALLED(maphist); 765 766 curoff = 0; 767 if (flags & PGO_ALLPAGES) { 768 start = 0; 769 stop = aobj->u_pages << PAGE_SHIFT; 770 by_list = TRUE; /* always go by the list */ 771 } else { 772 start = trunc_page(start); 773 stop = round_page(stop); 774 if (stop > (aobj->u_pages << PAGE_SHIFT)) { 775 printf("uao_flush: strange, got an out of range " 776 "flush (fixed)\n"); 777 stop = aobj->u_pages << PAGE_SHIFT; 778 } 779 by_list = (uobj->uo_npages <= 780 ((stop - start) >> PAGE_SHIFT) * UVM_PAGE_HASH_PENALTY); 781 } 782 UVMHIST_LOG(maphist, 783 " flush start=0x%lx, stop=0x%x, by_list=%d, flags=0x%x", 784 start, stop, by_list, flags); 785 786 /* 787 * Don't need to do any work here if we're not freeing 788 * or deactivating pages. 789 */ 790 791 if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) { 792 simple_unlock(&uobj->vmobjlock); 793 return 0; 794 } 795 796 /* 797 * Initialize the marker pages. See the comment in 798 * genfs_putpages() also. 799 */ 800 801 curmp.uobject = uobj; 802 curmp.offset = (voff_t)-1; 803 curmp.flags = PG_BUSY; 804 endmp.uobject = uobj; 805 endmp.offset = (voff_t)-1; 806 endmp.flags = PG_BUSY; 807 808 /* 809 * now do it. note: we must update nextpg in the body of loop or we 810 * will get stuck. we need to use nextpg if we'll traverse the list 811 * because we may free "pg" before doing the next loop. 812 */ 813 814 if (by_list) { 815 TAILQ_INSERT_TAIL(&uobj->memq, &endmp, listq); 816 nextpg = TAILQ_FIRST(&uobj->memq); 817 PHOLD(curlwp); 818 } else { 819 curoff = start; 820 nextpg = NULL; /* Quell compiler warning */ 821 } 822 823 uvm_lock_pageq(); 824 825 /* locked: both page queues and uobj */ 826 for (;;) { 827 if (by_list) { 828 pg = nextpg; 829 if (pg == &endmp) 830 break; 831 nextpg = TAILQ_NEXT(pg, listq); 832 if (pg->offset < start || pg->offset >= stop) 833 continue; 834 } else { 835 if (curoff < stop) { 836 pg = uvm_pagelookup(uobj, curoff); 837 curoff += PAGE_SIZE; 838 } else 839 break; 840 if (pg == NULL) 841 continue; 842 } 843 switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) { 844 845 /* 846 * XXX In these first 3 cases, we always just 847 * XXX deactivate the page. We may want to 848 * XXX handle the different cases more specifically 849 * XXX in the future. 850 */ 851 852 case PGO_CLEANIT|PGO_FREE: 853 case PGO_CLEANIT|PGO_DEACTIVATE: 854 case PGO_DEACTIVATE: 855 deactivate_it: 856 /* skip the page if it's loaned or wired */ 857 if (pg->loan_count != 0 || pg->wire_count != 0) 858 continue; 859 860 /* ...and deactivate the page. */ 861 pmap_clear_reference(pg); 862 uvm_pagedeactivate(pg); 863 continue; 864 865 case PGO_FREE: 866 867 /* 868 * If there are multiple references to 869 * the object, just deactivate the page. 870 */ 871 872 if (uobj->uo_refs > 1) 873 goto deactivate_it; 874 875 /* XXX skip the page if it's loaned or wired */ 876 if (pg->loan_count != 0 || pg->wire_count != 0) 877 continue; 878 879 /* 880 * wait and try again if the page is busy. 881 * otherwise free the swap slot and the page. 882 */ 883 884 pmap_page_protect(pg, VM_PROT_NONE); 885 if (pg->flags & PG_BUSY) { 886 if (by_list) { 887 TAILQ_INSERT_BEFORE(pg, &curmp, listq); 888 } 889 pg->flags |= PG_WANTED; 890 uvm_unlock_pageq(); 891 UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, 0, 892 "uao_put", 0); 893 simple_lock(&uobj->vmobjlock); 894 uvm_lock_pageq(); 895 if (by_list) { 896 nextpg = TAILQ_NEXT(&curmp, listq); 897 TAILQ_REMOVE(&uobj->memq, &curmp, 898 listq); 899 } else 900 curoff -= PAGE_SIZE; 901 continue; 902 } 903 uao_dropswap(uobj, pg->offset >> PAGE_SHIFT); 904 uvm_pagefree(pg); 905 continue; 906 } 907 } 908 uvm_unlock_pageq(); 909 if (by_list) { 910 TAILQ_REMOVE(&uobj->memq, &endmp, listq); 911 PRELE(curlwp); 912 } 913 simple_unlock(&uobj->vmobjlock); 914 return 0; 915 } 916 917 /* 918 * uao_get: fetch me a page 919 * 920 * we have three cases: 921 * 1: page is resident -> just return the page. 922 * 2: page is zero-fill -> allocate a new page and zero it. 923 * 3: page is swapped out -> fetch the page from swap. 924 * 925 * cases 1 and 2 can be handled with PGO_LOCKED, case 3 cannot. 926 * so, if the "center" page hits case 3 (or any page, with PGO_ALLPAGES), 927 * then we will need to return EBUSY. 928 * 929 * => prefer map unlocked (not required) 930 * => object must be locked! we will _unlock_ it before starting any I/O. 931 * => flags: PGO_ALLPAGES: get all of the pages 932 * PGO_LOCKED: fault data structures are locked 933 * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx] 934 * => NOTE: caller must check for released pages!! 935 */ 936 937 static int 938 uao_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags) 939 struct uvm_object *uobj; 940 voff_t offset; 941 struct vm_page **pps; 942 int *npagesp; 943 int centeridx, advice, flags; 944 vm_prot_t access_type; 945 { 946 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 947 voff_t current_offset; 948 struct vm_page *ptmp = NULL; /* Quell compiler warning */ 949 int lcv, gotpages, maxpages, swslot, error, pageidx; 950 boolean_t done; 951 UVMHIST_FUNC("uao_get"); UVMHIST_CALLED(pdhist); 952 953 UVMHIST_LOG(pdhist, "aobj=%p offset=%d, flags=%d", 954 aobj, offset, flags,0); 955 956 /* 957 * get number of pages 958 */ 959 960 maxpages = *npagesp; 961 962 /* 963 * step 1: handled the case where fault data structures are locked. 964 */ 965 966 if (flags & PGO_LOCKED) { 967 968 /* 969 * step 1a: get pages that are already resident. only do 970 * this if the data structures are locked (i.e. the first 971 * time through). 972 */ 973 974 done = TRUE; /* be optimistic */ 975 gotpages = 0; /* # of pages we got so far */ 976 for (lcv = 0, current_offset = offset ; lcv < maxpages ; 977 lcv++, current_offset += PAGE_SIZE) { 978 /* do we care about this page? if not, skip it */ 979 if (pps[lcv] == PGO_DONTCARE) 980 continue; 981 ptmp = uvm_pagelookup(uobj, current_offset); 982 983 /* 984 * if page is new, attempt to allocate the page, 985 * zero-fill'd. 986 */ 987 988 if (ptmp == NULL && uao_find_swslot(&aobj->u_obj, 989 current_offset >> PAGE_SHIFT) == 0) { 990 ptmp = uvm_pagealloc(uobj, current_offset, 991 NULL, UVM_PGA_ZERO); 992 if (ptmp) { 993 /* new page */ 994 ptmp->flags &= ~(PG_FAKE); 995 ptmp->pqflags |= PQ_AOBJ; 996 goto gotpage; 997 } 998 } 999 1000 /* 1001 * to be useful must get a non-busy page 1002 */ 1003 1004 if (ptmp == NULL || (ptmp->flags & PG_BUSY) != 0) { 1005 if (lcv == centeridx || 1006 (flags & PGO_ALLPAGES) != 0) 1007 /* need to do a wait or I/O! */ 1008 done = FALSE; 1009 continue; 1010 } 1011 1012 /* 1013 * useful page: busy/lock it and plug it in our 1014 * result array 1015 */ 1016 1017 /* caller must un-busy this page */ 1018 ptmp->flags |= PG_BUSY; 1019 UVM_PAGE_OWN(ptmp, "uao_get1"); 1020 gotpage: 1021 pps[lcv] = ptmp; 1022 gotpages++; 1023 } 1024 1025 /* 1026 * step 1b: now we've either done everything needed or we 1027 * to unlock and do some waiting or I/O. 1028 */ 1029 1030 UVMHIST_LOG(pdhist, "<- done (done=%d)", done, 0,0,0); 1031 *npagesp = gotpages; 1032 if (done) 1033 return 0; 1034 else 1035 return EBUSY; 1036 } 1037 1038 /* 1039 * step 2: get non-resident or busy pages. 1040 * object is locked. data structures are unlocked. 1041 */ 1042 1043 for (lcv = 0, current_offset = offset ; lcv < maxpages ; 1044 lcv++, current_offset += PAGE_SIZE) { 1045 1046 /* 1047 * - skip over pages we've already gotten or don't want 1048 * - skip over pages we don't _have_ to get 1049 */ 1050 1051 if (pps[lcv] != NULL || 1052 (lcv != centeridx && (flags & PGO_ALLPAGES) == 0)) 1053 continue; 1054 1055 pageidx = current_offset >> PAGE_SHIFT; 1056 1057 /* 1058 * we have yet to locate the current page (pps[lcv]). we 1059 * first look for a page that is already at the current offset. 1060 * if we find a page, we check to see if it is busy or 1061 * released. if that is the case, then we sleep on the page 1062 * until it is no longer busy or released and repeat the lookup. 1063 * if the page we found is neither busy nor released, then we 1064 * busy it (so we own it) and plug it into pps[lcv]. this 1065 * 'break's the following while loop and indicates we are 1066 * ready to move on to the next page in the "lcv" loop above. 1067 * 1068 * if we exit the while loop with pps[lcv] still set to NULL, 1069 * then it means that we allocated a new busy/fake/clean page 1070 * ptmp in the object and we need to do I/O to fill in the data. 1071 */ 1072 1073 /* top of "pps" while loop */ 1074 while (pps[lcv] == NULL) { 1075 /* look for a resident page */ 1076 ptmp = uvm_pagelookup(uobj, current_offset); 1077 1078 /* not resident? allocate one now (if we can) */ 1079 if (ptmp == NULL) { 1080 1081 ptmp = uvm_pagealloc(uobj, current_offset, 1082 NULL, 0); 1083 1084 /* out of RAM? */ 1085 if (ptmp == NULL) { 1086 simple_unlock(&uobj->vmobjlock); 1087 UVMHIST_LOG(pdhist, 1088 "sleeping, ptmp == NULL\n",0,0,0,0); 1089 uvm_wait("uao_getpage"); 1090 simple_lock(&uobj->vmobjlock); 1091 continue; 1092 } 1093 1094 /* 1095 * safe with PQ's unlocked: because we just 1096 * alloc'd the page 1097 */ 1098 1099 ptmp->pqflags |= PQ_AOBJ; 1100 1101 /* 1102 * got new page ready for I/O. break pps while 1103 * loop. pps[lcv] is still NULL. 1104 */ 1105 1106 break; 1107 } 1108 1109 /* page is there, see if we need to wait on it */ 1110 if ((ptmp->flags & PG_BUSY) != 0) { 1111 ptmp->flags |= PG_WANTED; 1112 UVMHIST_LOG(pdhist, 1113 "sleeping, ptmp->flags 0x%x\n", 1114 ptmp->flags,0,0,0); 1115 UVM_UNLOCK_AND_WAIT(ptmp, &uobj->vmobjlock, 1116 FALSE, "uao_get", 0); 1117 simple_lock(&uobj->vmobjlock); 1118 continue; 1119 } 1120 1121 /* 1122 * if we get here then the page has become resident and 1123 * unbusy between steps 1 and 2. we busy it now (so we 1124 * own it) and set pps[lcv] (so that we exit the while 1125 * loop). 1126 */ 1127 1128 /* we own it, caller must un-busy */ 1129 ptmp->flags |= PG_BUSY; 1130 UVM_PAGE_OWN(ptmp, "uao_get2"); 1131 pps[lcv] = ptmp; 1132 } 1133 1134 /* 1135 * if we own the valid page at the correct offset, pps[lcv] will 1136 * point to it. nothing more to do except go to the next page. 1137 */ 1138 1139 if (pps[lcv]) 1140 continue; /* next lcv */ 1141 1142 /* 1143 * we have a "fake/busy/clean" page that we just allocated. 1144 * do the needed "i/o", either reading from swap or zeroing. 1145 */ 1146 1147 swslot = uao_find_swslot(&aobj->u_obj, pageidx); 1148 1149 /* 1150 * just zero the page if there's nothing in swap. 1151 */ 1152 1153 if (swslot == 0) { 1154 1155 /* 1156 * page hasn't existed before, just zero it. 1157 */ 1158 1159 uvm_pagezero(ptmp); 1160 } else { 1161 UVMHIST_LOG(pdhist, "pagein from swslot %d", 1162 swslot, 0,0,0); 1163 1164 /* 1165 * page in the swapped-out page. 1166 * unlock object for i/o, relock when done. 1167 */ 1168 1169 simple_unlock(&uobj->vmobjlock); 1170 error = uvm_swap_get(ptmp, swslot, PGO_SYNCIO); 1171 simple_lock(&uobj->vmobjlock); 1172 1173 /* 1174 * I/O done. check for errors. 1175 */ 1176 1177 if (error != 0) { 1178 UVMHIST_LOG(pdhist, "<- done (error=%d)", 1179 error,0,0,0); 1180 if (ptmp->flags & PG_WANTED) 1181 wakeup(ptmp); 1182 1183 /* 1184 * remove the swap slot from the aobj 1185 * and mark the aobj as having no real slot. 1186 * don't free the swap slot, thus preventing 1187 * it from being used again. 1188 */ 1189 1190 swslot = uao_set_swslot(&aobj->u_obj, pageidx, 1191 SWSLOT_BAD); 1192 if (swslot > 0) { 1193 uvm_swap_markbad(swslot, 1); 1194 } 1195 1196 uvm_lock_pageq(); 1197 uvm_pagefree(ptmp); 1198 uvm_unlock_pageq(); 1199 simple_unlock(&uobj->vmobjlock); 1200 return error; 1201 } 1202 } 1203 1204 /* 1205 * we got the page! clear the fake flag (indicates valid 1206 * data now in page) and plug into our result array. note 1207 * that page is still busy. 1208 * 1209 * it is the callers job to: 1210 * => check if the page is released 1211 * => unbusy the page 1212 * => activate the page 1213 */ 1214 1215 ptmp->flags &= ~PG_FAKE; 1216 pps[lcv] = ptmp; 1217 } 1218 1219 /* 1220 * finally, unlock object and return. 1221 */ 1222 1223 simple_unlock(&uobj->vmobjlock); 1224 UVMHIST_LOG(pdhist, "<- done (OK)",0,0,0,0); 1225 return 0; 1226 } 1227 1228 /* 1229 * uao_dropswap: release any swap resources from this aobj page. 1230 * 1231 * => aobj must be locked or have a reference count of 0. 1232 */ 1233 1234 void 1235 uao_dropswap(uobj, pageidx) 1236 struct uvm_object *uobj; 1237 int pageidx; 1238 { 1239 int slot; 1240 1241 slot = uao_set_swslot(uobj, pageidx, 0); 1242 if (slot) { 1243 uvm_swap_free(slot, 1); 1244 } 1245 } 1246 1247 /* 1248 * page in every page in every aobj that is paged-out to a range of swslots. 1249 * 1250 * => nothing should be locked. 1251 * => returns TRUE if pagein was aborted due to lack of memory. 1252 */ 1253 1254 boolean_t 1255 uao_swap_off(startslot, endslot) 1256 int startslot, endslot; 1257 { 1258 struct uvm_aobj *aobj, *nextaobj; 1259 boolean_t rv; 1260 1261 /* 1262 * walk the list of all aobjs. 1263 */ 1264 1265 restart: 1266 simple_lock(&uao_list_lock); 1267 for (aobj = LIST_FIRST(&uao_list); 1268 aobj != NULL; 1269 aobj = nextaobj) { 1270 1271 /* 1272 * try to get the object lock, start all over if we fail. 1273 * most of the time we'll get the aobj lock, 1274 * so this should be a rare case. 1275 */ 1276 1277 if (!simple_lock_try(&aobj->u_obj.vmobjlock)) { 1278 simple_unlock(&uao_list_lock); 1279 goto restart; 1280 } 1281 1282 /* 1283 * add a ref to the aobj so it doesn't disappear 1284 * while we're working. 1285 */ 1286 1287 uao_reference_locked(&aobj->u_obj); 1288 1289 /* 1290 * now it's safe to unlock the uao list. 1291 */ 1292 1293 simple_unlock(&uao_list_lock); 1294 1295 /* 1296 * page in any pages in the swslot range. 1297 * if there's an error, abort and return the error. 1298 */ 1299 1300 rv = uao_pagein(aobj, startslot, endslot); 1301 if (rv) { 1302 uao_detach_locked(&aobj->u_obj); 1303 return rv; 1304 } 1305 1306 /* 1307 * we're done with this aobj. 1308 * relock the list and drop our ref on the aobj. 1309 */ 1310 1311 simple_lock(&uao_list_lock); 1312 nextaobj = LIST_NEXT(aobj, u_list); 1313 uao_detach_locked(&aobj->u_obj); 1314 } 1315 1316 /* 1317 * done with traversal, unlock the list 1318 */ 1319 simple_unlock(&uao_list_lock); 1320 return FALSE; 1321 } 1322 1323 1324 /* 1325 * page in any pages from aobj in the given range. 1326 * 1327 * => aobj must be locked and is returned locked. 1328 * => returns TRUE if pagein was aborted due to lack of memory. 1329 */ 1330 static boolean_t 1331 uao_pagein(aobj, startslot, endslot) 1332 struct uvm_aobj *aobj; 1333 int startslot, endslot; 1334 { 1335 boolean_t rv; 1336 1337 if (UAO_USES_SWHASH(aobj)) { 1338 struct uao_swhash_elt *elt; 1339 int bucket; 1340 1341 restart: 1342 for (bucket = aobj->u_swhashmask; bucket >= 0; bucket--) { 1343 for (elt = LIST_FIRST(&aobj->u_swhash[bucket]); 1344 elt != NULL; 1345 elt = LIST_NEXT(elt, list)) { 1346 int i; 1347 1348 for (i = 0; i < UAO_SWHASH_CLUSTER_SIZE; i++) { 1349 int slot = elt->slots[i]; 1350 1351 /* 1352 * if the slot isn't in range, skip it. 1353 */ 1354 1355 if (slot < startslot || 1356 slot >= endslot) { 1357 continue; 1358 } 1359 1360 /* 1361 * process the page, 1362 * the start over on this object 1363 * since the swhash elt 1364 * may have been freed. 1365 */ 1366 1367 rv = uao_pagein_page(aobj, 1368 UAO_SWHASH_ELT_PAGEIDX_BASE(elt) + i); 1369 if (rv) { 1370 return rv; 1371 } 1372 goto restart; 1373 } 1374 } 1375 } 1376 } else { 1377 int i; 1378 1379 for (i = 0; i < aobj->u_pages; i++) { 1380 int slot = aobj->u_swslots[i]; 1381 1382 /* 1383 * if the slot isn't in range, skip it 1384 */ 1385 1386 if (slot < startslot || slot >= endslot) { 1387 continue; 1388 } 1389 1390 /* 1391 * process the page. 1392 */ 1393 1394 rv = uao_pagein_page(aobj, i); 1395 if (rv) { 1396 return rv; 1397 } 1398 } 1399 } 1400 1401 return FALSE; 1402 } 1403 1404 /* 1405 * page in a page from an aobj. used for swap_off. 1406 * returns TRUE if pagein was aborted due to lack of memory. 1407 * 1408 * => aobj must be locked and is returned locked. 1409 */ 1410 1411 static boolean_t 1412 uao_pagein_page(aobj, pageidx) 1413 struct uvm_aobj *aobj; 1414 int pageidx; 1415 { 1416 struct vm_page *pg; 1417 int rv, npages; 1418 1419 pg = NULL; 1420 npages = 1; 1421 /* locked: aobj */ 1422 rv = uao_get(&aobj->u_obj, pageidx << PAGE_SHIFT, 1423 &pg, &npages, 0, VM_PROT_READ|VM_PROT_WRITE, 0, 0); 1424 /* unlocked: aobj */ 1425 1426 /* 1427 * relock and finish up. 1428 */ 1429 1430 simple_lock(&aobj->u_obj.vmobjlock); 1431 switch (rv) { 1432 case 0: 1433 break; 1434 1435 case EIO: 1436 case ERESTART: 1437 1438 /* 1439 * nothing more to do on errors. 1440 * ERESTART can only mean that the anon was freed, 1441 * so again there's nothing to do. 1442 */ 1443 1444 return FALSE; 1445 1446 default: 1447 return TRUE; 1448 } 1449 1450 /* 1451 * ok, we've got the page now. 1452 * mark it as dirty, clear its swslot and un-busy it. 1453 */ 1454 uao_dropswap(&aobj->u_obj, pageidx); 1455 1456 /* 1457 * deactivate the page (to make sure it's on a page queue). 1458 */ 1459 uvm_lock_pageq(); 1460 if (pg->wire_count == 0) 1461 uvm_pagedeactivate(pg); 1462 uvm_unlock_pageq(); 1463 1464 if (pg->flags & PG_WANTED) { 1465 wakeup(pg); 1466 } 1467 pg->flags &= ~(PG_WANTED|PG_BUSY|PG_CLEAN|PG_FAKE); 1468 UVM_PAGE_OWN(pg, NULL); 1469 1470 return FALSE; 1471 } 1472