1 /* $NetBSD: uvm_aobj.c,v 1.61 2003/09/18 13:48:05 drochner 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 1.61 2003/09/18 13:48:05 drochner 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 uvm_swap_free(slot, 1); 422 swpgonlydelta++; 423 } 424 } 425 426 next = LIST_NEXT(elt, list); 427 pool_put(&uao_swhash_elt_pool, elt); 428 } 429 } 430 free(aobj->u_swhash, M_UVMAOBJ); 431 } else { 432 int i; 433 434 /* 435 * free the array 436 */ 437 438 for (i = 0; i < aobj->u_pages; i++) { 439 int slot = aobj->u_swslots[i]; 440 441 if (slot > 0) { 442 uvm_swap_free(slot, 1); 443 swpgonlydelta++; 444 } 445 } 446 free(aobj->u_swslots, M_UVMAOBJ); 447 } 448 449 /* 450 * finally free the aobj itself 451 */ 452 453 pool_put(&uvm_aobj_pool, aobj); 454 455 /* 456 * adjust the counter of pages only in swap for all 457 * the swap slots we've freed. 458 */ 459 460 if (swpgonlydelta > 0) { 461 simple_lock(&uvm.swap_data_lock); 462 KASSERT(uvmexp.swpgonly >= swpgonlydelta); 463 uvmexp.swpgonly -= swpgonlydelta; 464 simple_unlock(&uvm.swap_data_lock); 465 } 466 } 467 468 /* 469 * pager functions 470 */ 471 472 /* 473 * uao_create: create an aobj of the given size and return its uvm_object. 474 * 475 * => for normal use, flags are always zero 476 * => for the kernel object, the flags are: 477 * UAO_FLAG_KERNOBJ - allocate the kernel object (can only happen once) 478 * UAO_FLAG_KERNSWAP - enable swapping of kernel object (" ") 479 */ 480 481 struct uvm_object * 482 uao_create(size, flags) 483 vsize_t size; 484 int flags; 485 { 486 static struct uvm_aobj kernel_object_store; 487 static int kobj_alloced = 0; 488 int pages = round_page(size) >> PAGE_SHIFT; 489 struct uvm_aobj *aobj; 490 491 /* 492 * malloc a new aobj unless we are asked for the kernel object 493 */ 494 495 if (flags & UAO_FLAG_KERNOBJ) { 496 KASSERT(!kobj_alloced); 497 aobj = &kernel_object_store; 498 aobj->u_pages = pages; 499 aobj->u_flags = UAO_FLAG_NOSWAP; 500 aobj->u_obj.uo_refs = UVM_OBJ_KERN; 501 kobj_alloced = UAO_FLAG_KERNOBJ; 502 } else if (flags & UAO_FLAG_KERNSWAP) { 503 KASSERT(kobj_alloced == UAO_FLAG_KERNOBJ); 504 aobj = &kernel_object_store; 505 kobj_alloced = UAO_FLAG_KERNSWAP; 506 } else { 507 aobj = pool_get(&uvm_aobj_pool, PR_WAITOK); 508 aobj->u_pages = pages; 509 aobj->u_flags = 0; 510 aobj->u_obj.uo_refs = 1; 511 } 512 513 /* 514 * allocate hash/array if necessary 515 * 516 * note: in the KERNSWAP case no need to worry about locking since 517 * we are still booting we should be the only thread around. 518 */ 519 520 if (flags == 0 || (flags & UAO_FLAG_KERNSWAP) != 0) { 521 int mflags = (flags & UAO_FLAG_KERNSWAP) != 0 ? 522 M_NOWAIT : M_WAITOK; 523 524 /* allocate hash table or array depending on object size */ 525 if (UAO_USES_SWHASH(aobj)) { 526 aobj->u_swhash = hashinit(UAO_SWHASH_BUCKETS(aobj), 527 HASH_LIST, M_UVMAOBJ, mflags, &aobj->u_swhashmask); 528 if (aobj->u_swhash == NULL) 529 panic("uao_create: hashinit swhash failed"); 530 } else { 531 aobj->u_swslots = malloc(pages * sizeof(int), 532 M_UVMAOBJ, mflags); 533 if (aobj->u_swslots == NULL) 534 panic("uao_create: malloc swslots failed"); 535 memset(aobj->u_swslots, 0, pages * sizeof(int)); 536 } 537 538 if (flags) { 539 aobj->u_flags &= ~UAO_FLAG_NOSWAP; /* clear noswap */ 540 return(&aobj->u_obj); 541 } 542 } 543 544 /* 545 * init aobj fields 546 */ 547 548 simple_lock_init(&aobj->u_obj.vmobjlock); 549 aobj->u_obj.pgops = &aobj_pager; 550 TAILQ_INIT(&aobj->u_obj.memq); 551 aobj->u_obj.uo_npages = 0; 552 553 /* 554 * now that aobj is ready, add it to the global list 555 */ 556 557 simple_lock(&uao_list_lock); 558 LIST_INSERT_HEAD(&uao_list, aobj, u_list); 559 simple_unlock(&uao_list_lock); 560 return(&aobj->u_obj); 561 } 562 563 564 565 /* 566 * uao_init: set up aobj pager subsystem 567 * 568 * => called at boot time from uvm_pager_init() 569 */ 570 571 void 572 uao_init(void) 573 { 574 static int uao_initialized; 575 576 if (uao_initialized) 577 return; 578 uao_initialized = TRUE; 579 LIST_INIT(&uao_list); 580 simple_lock_init(&uao_list_lock); 581 582 /* 583 * NOTE: Pages fror this pool must not come from a pageable 584 * kernel map! 585 */ 586 587 pool_init(&uao_swhash_elt_pool, sizeof(struct uao_swhash_elt), 588 0, 0, 0, "uaoeltpl", NULL); 589 pool_init(&uvm_aobj_pool, sizeof(struct uvm_aobj), 0, 0, 0, 590 "aobjpl", &pool_allocator_nointr); 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, curmp, endmp; 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 * Initialize the marker pages. See the comment in 807 * genfs_putpages() also. 808 */ 809 810 curmp.uobject = uobj; 811 curmp.offset = (voff_t)-1; 812 curmp.flags = PG_BUSY; 813 endmp.uobject = uobj; 814 endmp.offset = (voff_t)-1; 815 endmp.flags = PG_BUSY; 816 817 /* 818 * now do it. note: we must update nextpg in the body of loop or we 819 * will get stuck. we need to use nextpg if we'll traverse the list 820 * because we may free "pg" before doing the next loop. 821 */ 822 823 if (by_list) { 824 TAILQ_INSERT_TAIL(&uobj->memq, &endmp, listq); 825 nextpg = TAILQ_FIRST(&uobj->memq); 826 PHOLD(curlwp); 827 } else { 828 curoff = start; 829 nextpg = NULL; /* Quell compiler warning */ 830 } 831 832 uvm_lock_pageq(); 833 834 /* locked: both page queues and uobj */ 835 for (;;) { 836 if (by_list) { 837 pg = nextpg; 838 if (pg == &endmp) 839 break; 840 nextpg = TAILQ_NEXT(pg, listq); 841 if (pg->offset < start || pg->offset >= stop) 842 continue; 843 } else { 844 if (curoff < stop) { 845 pg = uvm_pagelookup(uobj, curoff); 846 curoff += PAGE_SIZE; 847 } else 848 break; 849 if (pg == NULL) 850 continue; 851 } 852 switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) { 853 854 /* 855 * XXX In these first 3 cases, we always just 856 * XXX deactivate the page. We may want to 857 * XXX handle the different cases more specifically 858 * XXX in the future. 859 */ 860 861 case PGO_CLEANIT|PGO_FREE: 862 case PGO_CLEANIT|PGO_DEACTIVATE: 863 case PGO_DEACTIVATE: 864 deactivate_it: 865 /* skip the page if it's loaned or wired */ 866 if (pg->loan_count != 0 || pg->wire_count != 0) 867 continue; 868 869 /* ...and deactivate the page. */ 870 pmap_clear_reference(pg); 871 uvm_pagedeactivate(pg); 872 continue; 873 874 case PGO_FREE: 875 876 /* 877 * If there are multiple references to 878 * the object, just deactivate the page. 879 */ 880 881 if (uobj->uo_refs > 1) 882 goto deactivate_it; 883 884 /* XXX skip the page if it's loaned or wired */ 885 if (pg->loan_count != 0 || pg->wire_count != 0) 886 continue; 887 888 /* 889 * wait and try again if the page is busy. 890 * otherwise free the swap slot and the page. 891 */ 892 893 pmap_page_protect(pg, VM_PROT_NONE); 894 if (pg->flags & PG_BUSY) { 895 if (by_list) { 896 TAILQ_INSERT_BEFORE(pg, &curmp, listq); 897 } 898 pg->flags |= PG_WANTED; 899 uvm_unlock_pageq(); 900 UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, 0, 901 "uao_put", 0); 902 simple_lock(&uobj->vmobjlock); 903 uvm_lock_pageq(); 904 if (by_list) { 905 nextpg = TAILQ_NEXT(&curmp, listq); 906 TAILQ_REMOVE(&uobj->memq, &curmp, 907 listq); 908 } else 909 curoff -= PAGE_SIZE; 910 continue; 911 } 912 uao_dropswap(uobj, pg->offset >> PAGE_SHIFT); 913 uvm_pagefree(pg); 914 continue; 915 } 916 } 917 uvm_unlock_pageq(); 918 if (by_list) { 919 TAILQ_REMOVE(&uobj->memq, &endmp, listq); 920 PRELE(curlwp); 921 } 922 simple_unlock(&uobj->vmobjlock); 923 return 0; 924 } 925 926 /* 927 * uao_get: fetch me a page 928 * 929 * we have three cases: 930 * 1: page is resident -> just return the page. 931 * 2: page is zero-fill -> allocate a new page and zero it. 932 * 3: page is swapped out -> fetch the page from swap. 933 * 934 * cases 1 and 2 can be handled with PGO_LOCKED, case 3 cannot. 935 * so, if the "center" page hits case 3 (or any page, with PGO_ALLPAGES), 936 * then we will need to return EBUSY. 937 * 938 * => prefer map unlocked (not required) 939 * => object must be locked! we will _unlock_ it before starting any I/O. 940 * => flags: PGO_ALLPAGES: get all of the pages 941 * PGO_LOCKED: fault data structures are locked 942 * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx] 943 * => NOTE: caller must check for released pages!! 944 */ 945 946 static int 947 uao_get(uobj, offset, pps, npagesp, centeridx, access_type, advice, flags) 948 struct uvm_object *uobj; 949 voff_t offset; 950 struct vm_page **pps; 951 int *npagesp; 952 int centeridx, advice, flags; 953 vm_prot_t access_type; 954 { 955 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 956 voff_t current_offset; 957 struct vm_page *ptmp = NULL; /* Quell compiler warning */ 958 int lcv, gotpages, maxpages, swslot, error, pageidx; 959 boolean_t done; 960 UVMHIST_FUNC("uao_get"); UVMHIST_CALLED(pdhist); 961 962 UVMHIST_LOG(pdhist, "aobj=%p offset=%d, flags=%d", 963 aobj, offset, flags,0); 964 965 /* 966 * get number of pages 967 */ 968 969 maxpages = *npagesp; 970 971 /* 972 * step 1: handled the case where fault data structures are locked. 973 */ 974 975 if (flags & PGO_LOCKED) { 976 977 /* 978 * step 1a: get pages that are already resident. only do 979 * this if the data structures are locked (i.e. the first 980 * time through). 981 */ 982 983 done = TRUE; /* be optimistic */ 984 gotpages = 0; /* # of pages we got so far */ 985 for (lcv = 0, current_offset = offset ; lcv < maxpages ; 986 lcv++, current_offset += PAGE_SIZE) { 987 /* do we care about this page? if not, skip it */ 988 if (pps[lcv] == PGO_DONTCARE) 989 continue; 990 ptmp = uvm_pagelookup(uobj, current_offset); 991 992 /* 993 * if page is new, attempt to allocate the page, 994 * zero-fill'd. 995 */ 996 997 if (ptmp == NULL && uao_find_swslot(&aobj->u_obj, 998 current_offset >> PAGE_SHIFT) == 0) { 999 ptmp = uvm_pagealloc(uobj, current_offset, 1000 NULL, UVM_PGA_ZERO); 1001 if (ptmp) { 1002 /* new page */ 1003 ptmp->flags &= ~(PG_FAKE); 1004 ptmp->pqflags |= PQ_AOBJ; 1005 goto gotpage; 1006 } 1007 } 1008 1009 /* 1010 * to be useful must get a non-busy page 1011 */ 1012 1013 if (ptmp == NULL || (ptmp->flags & PG_BUSY) != 0) { 1014 if (lcv == centeridx || 1015 (flags & PGO_ALLPAGES) != 0) 1016 /* need to do a wait or I/O! */ 1017 done = FALSE; 1018 continue; 1019 } 1020 1021 /* 1022 * useful page: busy/lock it and plug it in our 1023 * result array 1024 */ 1025 1026 /* caller must un-busy this page */ 1027 ptmp->flags |= PG_BUSY; 1028 UVM_PAGE_OWN(ptmp, "uao_get1"); 1029 gotpage: 1030 pps[lcv] = ptmp; 1031 gotpages++; 1032 } 1033 1034 /* 1035 * step 1b: now we've either done everything needed or we 1036 * to unlock and do some waiting or I/O. 1037 */ 1038 1039 UVMHIST_LOG(pdhist, "<- done (done=%d)", done, 0,0,0); 1040 *npagesp = gotpages; 1041 if (done) 1042 return 0; 1043 else 1044 return EBUSY; 1045 } 1046 1047 /* 1048 * step 2: get non-resident or busy pages. 1049 * object is locked. data structures are unlocked. 1050 */ 1051 1052 for (lcv = 0, current_offset = offset ; lcv < maxpages ; 1053 lcv++, current_offset += PAGE_SIZE) { 1054 1055 /* 1056 * - skip over pages we've already gotten or don't want 1057 * - skip over pages we don't _have_ to get 1058 */ 1059 1060 if (pps[lcv] != NULL || 1061 (lcv != centeridx && (flags & PGO_ALLPAGES) == 0)) 1062 continue; 1063 1064 pageidx = current_offset >> PAGE_SHIFT; 1065 1066 /* 1067 * we have yet to locate the current page (pps[lcv]). we 1068 * first look for a page that is already at the current offset. 1069 * if we find a page, we check to see if it is busy or 1070 * released. if that is the case, then we sleep on the page 1071 * until it is no longer busy or released and repeat the lookup. 1072 * if the page we found is neither busy nor released, then we 1073 * busy it (so we own it) and plug it into pps[lcv]. this 1074 * 'break's the following while loop and indicates we are 1075 * ready to move on to the next page in the "lcv" loop above. 1076 * 1077 * if we exit the while loop with pps[lcv] still set to NULL, 1078 * then it means that we allocated a new busy/fake/clean page 1079 * ptmp in the object and we need to do I/O to fill in the data. 1080 */ 1081 1082 /* top of "pps" while loop */ 1083 while (pps[lcv] == NULL) { 1084 /* look for a resident page */ 1085 ptmp = uvm_pagelookup(uobj, current_offset); 1086 1087 /* not resident? allocate one now (if we can) */ 1088 if (ptmp == NULL) { 1089 1090 ptmp = uvm_pagealloc(uobj, current_offset, 1091 NULL, 0); 1092 1093 /* out of RAM? */ 1094 if (ptmp == NULL) { 1095 simple_unlock(&uobj->vmobjlock); 1096 UVMHIST_LOG(pdhist, 1097 "sleeping, ptmp == NULL\n",0,0,0,0); 1098 uvm_wait("uao_getpage"); 1099 simple_lock(&uobj->vmobjlock); 1100 continue; 1101 } 1102 1103 /* 1104 * safe with PQ's unlocked: because we just 1105 * alloc'd the page 1106 */ 1107 1108 ptmp->pqflags |= PQ_AOBJ; 1109 1110 /* 1111 * got new page ready for I/O. break pps while 1112 * loop. pps[lcv] is still NULL. 1113 */ 1114 1115 break; 1116 } 1117 1118 /* page is there, see if we need to wait on it */ 1119 if ((ptmp->flags & PG_BUSY) != 0) { 1120 ptmp->flags |= PG_WANTED; 1121 UVMHIST_LOG(pdhist, 1122 "sleeping, ptmp->flags 0x%x\n", 1123 ptmp->flags,0,0,0); 1124 UVM_UNLOCK_AND_WAIT(ptmp, &uobj->vmobjlock, 1125 FALSE, "uao_get", 0); 1126 simple_lock(&uobj->vmobjlock); 1127 continue; 1128 } 1129 1130 /* 1131 * if we get here then the page has become resident and 1132 * unbusy between steps 1 and 2. we busy it now (so we 1133 * own it) and set pps[lcv] (so that we exit the while 1134 * loop). 1135 */ 1136 1137 /* we own it, caller must un-busy */ 1138 ptmp->flags |= PG_BUSY; 1139 UVM_PAGE_OWN(ptmp, "uao_get2"); 1140 pps[lcv] = ptmp; 1141 } 1142 1143 /* 1144 * if we own the valid page at the correct offset, pps[lcv] will 1145 * point to it. nothing more to do except go to the next page. 1146 */ 1147 1148 if (pps[lcv]) 1149 continue; /* next lcv */ 1150 1151 /* 1152 * we have a "fake/busy/clean" page that we just allocated. 1153 * do the needed "i/o", either reading from swap or zeroing. 1154 */ 1155 1156 swslot = uao_find_swslot(&aobj->u_obj, pageidx); 1157 1158 /* 1159 * just zero the page if there's nothing in swap. 1160 */ 1161 1162 if (swslot == 0) { 1163 1164 /* 1165 * page hasn't existed before, just zero it. 1166 */ 1167 1168 uvm_pagezero(ptmp); 1169 } else { 1170 UVMHIST_LOG(pdhist, "pagein from swslot %d", 1171 swslot, 0,0,0); 1172 1173 /* 1174 * page in the swapped-out page. 1175 * unlock object for i/o, relock when done. 1176 */ 1177 1178 simple_unlock(&uobj->vmobjlock); 1179 error = uvm_swap_get(ptmp, swslot, PGO_SYNCIO); 1180 simple_lock(&uobj->vmobjlock); 1181 1182 /* 1183 * I/O done. check for errors. 1184 */ 1185 1186 if (error != 0) { 1187 UVMHIST_LOG(pdhist, "<- done (error=%d)", 1188 error,0,0,0); 1189 if (ptmp->flags & PG_WANTED) 1190 wakeup(ptmp); 1191 1192 /* 1193 * remove the swap slot from the aobj 1194 * and mark the aobj as having no real slot. 1195 * don't free the swap slot, thus preventing 1196 * it from being used again. 1197 */ 1198 1199 swslot = uao_set_swslot(&aobj->u_obj, pageidx, 1200 SWSLOT_BAD); 1201 if (swslot > 0) { 1202 uvm_swap_markbad(swslot, 1); 1203 } 1204 1205 uvm_lock_pageq(); 1206 uvm_pagefree(ptmp); 1207 uvm_unlock_pageq(); 1208 simple_unlock(&uobj->vmobjlock); 1209 return error; 1210 } 1211 } 1212 1213 /* 1214 * we got the page! clear the fake flag (indicates valid 1215 * data now in page) and plug into our result array. note 1216 * that page is still busy. 1217 * 1218 * it is the callers job to: 1219 * => check if the page is released 1220 * => unbusy the page 1221 * => activate the page 1222 */ 1223 1224 ptmp->flags &= ~PG_FAKE; 1225 pps[lcv] = ptmp; 1226 } 1227 1228 /* 1229 * finally, unlock object and return. 1230 */ 1231 1232 simple_unlock(&uobj->vmobjlock); 1233 UVMHIST_LOG(pdhist, "<- done (OK)",0,0,0,0); 1234 return 0; 1235 } 1236 1237 /* 1238 * uao_dropswap: release any swap resources from this aobj page. 1239 * 1240 * => aobj must be locked or have a reference count of 0. 1241 */ 1242 1243 void 1244 uao_dropswap(uobj, pageidx) 1245 struct uvm_object *uobj; 1246 int pageidx; 1247 { 1248 int slot; 1249 1250 slot = uao_set_swslot(uobj, pageidx, 0); 1251 if (slot) { 1252 uvm_swap_free(slot, 1); 1253 } 1254 } 1255 1256 /* 1257 * page in every page in every aobj that is paged-out to a range of swslots. 1258 * 1259 * => nothing should be locked. 1260 * => returns TRUE if pagein was aborted due to lack of memory. 1261 */ 1262 1263 boolean_t 1264 uao_swap_off(startslot, endslot) 1265 int startslot, endslot; 1266 { 1267 struct uvm_aobj *aobj, *nextaobj; 1268 boolean_t rv; 1269 1270 /* 1271 * walk the list of all aobjs. 1272 */ 1273 1274 restart: 1275 simple_lock(&uao_list_lock); 1276 for (aobj = LIST_FIRST(&uao_list); 1277 aobj != NULL; 1278 aobj = nextaobj) { 1279 1280 /* 1281 * try to get the object lock, start all over if we fail. 1282 * most of the time we'll get the aobj lock, 1283 * so this should be a rare case. 1284 */ 1285 1286 if (!simple_lock_try(&aobj->u_obj.vmobjlock)) { 1287 simple_unlock(&uao_list_lock); 1288 goto restart; 1289 } 1290 1291 /* 1292 * add a ref to the aobj so it doesn't disappear 1293 * while we're working. 1294 */ 1295 1296 uao_reference_locked(&aobj->u_obj); 1297 1298 /* 1299 * now it's safe to unlock the uao list. 1300 */ 1301 1302 simple_unlock(&uao_list_lock); 1303 1304 /* 1305 * page in any pages in the swslot range. 1306 * if there's an error, abort and return the error. 1307 */ 1308 1309 rv = uao_pagein(aobj, startslot, endslot); 1310 if (rv) { 1311 uao_detach_locked(&aobj->u_obj); 1312 return rv; 1313 } 1314 1315 /* 1316 * we're done with this aobj. 1317 * relock the list and drop our ref on the aobj. 1318 */ 1319 1320 simple_lock(&uao_list_lock); 1321 nextaobj = LIST_NEXT(aobj, u_list); 1322 uao_detach_locked(&aobj->u_obj); 1323 } 1324 1325 /* 1326 * done with traversal, unlock the list 1327 */ 1328 simple_unlock(&uao_list_lock); 1329 return FALSE; 1330 } 1331 1332 1333 /* 1334 * page in any pages from aobj in the given range. 1335 * 1336 * => aobj must be locked and is returned locked. 1337 * => returns TRUE if pagein was aborted due to lack of memory. 1338 */ 1339 static boolean_t 1340 uao_pagein(aobj, startslot, endslot) 1341 struct uvm_aobj *aobj; 1342 int startslot, endslot; 1343 { 1344 boolean_t rv; 1345 1346 if (UAO_USES_SWHASH(aobj)) { 1347 struct uao_swhash_elt *elt; 1348 int bucket; 1349 1350 restart: 1351 for (bucket = aobj->u_swhashmask; bucket >= 0; bucket--) { 1352 for (elt = LIST_FIRST(&aobj->u_swhash[bucket]); 1353 elt != NULL; 1354 elt = LIST_NEXT(elt, list)) { 1355 int i; 1356 1357 for (i = 0; i < UAO_SWHASH_CLUSTER_SIZE; i++) { 1358 int slot = elt->slots[i]; 1359 1360 /* 1361 * if the slot isn't in range, skip it. 1362 */ 1363 1364 if (slot < startslot || 1365 slot >= endslot) { 1366 continue; 1367 } 1368 1369 /* 1370 * process the page, 1371 * the start over on this object 1372 * since the swhash elt 1373 * may have been freed. 1374 */ 1375 1376 rv = uao_pagein_page(aobj, 1377 UAO_SWHASH_ELT_PAGEIDX_BASE(elt) + i); 1378 if (rv) { 1379 return rv; 1380 } 1381 goto restart; 1382 } 1383 } 1384 } 1385 } else { 1386 int i; 1387 1388 for (i = 0; i < aobj->u_pages; i++) { 1389 int slot = aobj->u_swslots[i]; 1390 1391 /* 1392 * if the slot isn't in range, skip it 1393 */ 1394 1395 if (slot < startslot || slot >= endslot) { 1396 continue; 1397 } 1398 1399 /* 1400 * process the page. 1401 */ 1402 1403 rv = uao_pagein_page(aobj, i); 1404 if (rv) { 1405 return rv; 1406 } 1407 } 1408 } 1409 1410 return FALSE; 1411 } 1412 1413 /* 1414 * page in a page from an aobj. used for swap_off. 1415 * returns TRUE if pagein was aborted due to lack of memory. 1416 * 1417 * => aobj must be locked and is returned locked. 1418 */ 1419 1420 static boolean_t 1421 uao_pagein_page(aobj, pageidx) 1422 struct uvm_aobj *aobj; 1423 int pageidx; 1424 { 1425 struct vm_page *pg; 1426 int rv, npages; 1427 1428 pg = NULL; 1429 npages = 1; 1430 /* locked: aobj */ 1431 rv = uao_get(&aobj->u_obj, pageidx << PAGE_SHIFT, 1432 &pg, &npages, 0, VM_PROT_READ|VM_PROT_WRITE, 0, 0); 1433 /* unlocked: aobj */ 1434 1435 /* 1436 * relock and finish up. 1437 */ 1438 1439 simple_lock(&aobj->u_obj.vmobjlock); 1440 switch (rv) { 1441 case 0: 1442 break; 1443 1444 case EIO: 1445 case ERESTART: 1446 1447 /* 1448 * nothing more to do on errors. 1449 * ERESTART can only mean that the anon was freed, 1450 * so again there's nothing to do. 1451 */ 1452 1453 return FALSE; 1454 1455 default: 1456 return TRUE; 1457 } 1458 1459 /* 1460 * ok, we've got the page now. 1461 * mark it as dirty, clear its swslot and un-busy it. 1462 */ 1463 uao_dropswap(&aobj->u_obj, pageidx); 1464 1465 /* 1466 * deactivate the page (to make sure it's on a page queue). 1467 */ 1468 uvm_lock_pageq(); 1469 if (pg->wire_count == 0) 1470 uvm_pagedeactivate(pg); 1471 uvm_unlock_pageq(); 1472 1473 if (pg->flags & PG_WANTED) { 1474 wakeup(pg); 1475 } 1476 pg->flags &= ~(PG_WANTED|PG_BUSY|PG_CLEAN|PG_FAKE); 1477 UVM_PAGE_OWN(pg, NULL); 1478 1479 return FALSE; 1480 } 1481