1 /* $NetBSD: uvm_aobj.c,v 1.106 2009/02/18 13:16:58 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 1.106 2009/02/18 13:16:58 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/kernel.h> 54 #include <sys/kmem.h> 55 #include <sys/pool.h> 56 57 #include <uvm/uvm.h> 58 59 /* 60 * an aobj manages anonymous-memory backed uvm_objects. in addition 61 * to keeping the list of resident pages, it also keeps a list of 62 * allocated swap blocks. depending on the size of the aobj this list 63 * of allocated swap blocks is either stored in an array (small objects) 64 * or in a hash table (large objects). 65 */ 66 67 /* 68 * local structures 69 */ 70 71 /* 72 * for hash tables, we break the address space of the aobj into blocks 73 * of UAO_SWHASH_CLUSTER_SIZE pages. we require the cluster size to 74 * be a power of two. 75 */ 76 77 #define UAO_SWHASH_CLUSTER_SHIFT 4 78 #define UAO_SWHASH_CLUSTER_SIZE (1 << UAO_SWHASH_CLUSTER_SHIFT) 79 80 /* get the "tag" for this page index */ 81 #define UAO_SWHASH_ELT_TAG(PAGEIDX) \ 82 ((PAGEIDX) >> UAO_SWHASH_CLUSTER_SHIFT) 83 84 #define UAO_SWHASH_ELT_PAGESLOT_IDX(PAGEIDX) \ 85 ((PAGEIDX) & (UAO_SWHASH_CLUSTER_SIZE - 1)) 86 87 /* given an ELT and a page index, find the swap slot */ 88 #define UAO_SWHASH_ELT_PAGESLOT(ELT, PAGEIDX) \ 89 ((ELT)->slots[UAO_SWHASH_ELT_PAGESLOT_IDX(PAGEIDX)]) 90 91 /* given an ELT, return its pageidx base */ 92 #define UAO_SWHASH_ELT_PAGEIDX_BASE(ELT) \ 93 ((ELT)->tag << UAO_SWHASH_CLUSTER_SHIFT) 94 95 /* 96 * the swhash hash function 97 */ 98 99 #define UAO_SWHASH_HASH(AOBJ, PAGEIDX) \ 100 (&(AOBJ)->u_swhash[(((PAGEIDX) >> UAO_SWHASH_CLUSTER_SHIFT) \ 101 & (AOBJ)->u_swhashmask)]) 102 103 /* 104 * the swhash threshhold determines if we will use an array or a 105 * hash table to store the list of allocated swap blocks. 106 */ 107 108 #define UAO_SWHASH_THRESHOLD (UAO_SWHASH_CLUSTER_SIZE * 4) 109 #define UAO_USES_SWHASH(AOBJ) \ 110 ((AOBJ)->u_pages > UAO_SWHASH_THRESHOLD) /* use hash? */ 111 112 /* 113 * the number of buckets in a swhash, with an upper bound 114 */ 115 116 #define UAO_SWHASH_MAXBUCKETS 256 117 #define UAO_SWHASH_BUCKETS(AOBJ) \ 118 (MIN((AOBJ)->u_pages >> UAO_SWHASH_CLUSTER_SHIFT, \ 119 UAO_SWHASH_MAXBUCKETS)) 120 121 /* 122 * uao_swhash_elt: when a hash table is being used, this structure defines 123 * the format of an entry in the bucket list. 124 */ 125 126 struct uao_swhash_elt { 127 LIST_ENTRY(uao_swhash_elt) list; /* the hash list */ 128 voff_t tag; /* our 'tag' */ 129 int count; /* our number of active slots */ 130 int slots[UAO_SWHASH_CLUSTER_SIZE]; /* the slots */ 131 }; 132 133 /* 134 * uao_swhash: the swap hash table structure 135 */ 136 137 LIST_HEAD(uao_swhash, uao_swhash_elt); 138 139 /* 140 * uao_swhash_elt_pool: pool of uao_swhash_elt structures 141 * NOTE: Pages for this pool must not come from a pageable kernel map! 142 */ 143 static POOL_INIT(uao_swhash_elt_pool, sizeof(struct uao_swhash_elt), 0, 0, 0, 144 "uaoeltpl", NULL, IPL_VM); 145 146 static struct pool_cache uvm_aobj_cache; 147 148 /* 149 * uvm_aobj: the actual anon-backed uvm_object 150 * 151 * => the uvm_object is at the top of the structure, this allows 152 * (struct uvm_aobj *) == (struct uvm_object *) 153 * => only one of u_swslots and u_swhash is used in any given aobj 154 */ 155 156 struct uvm_aobj { 157 struct uvm_object u_obj; /* has: lock, pgops, memq, #pages, #refs */ 158 pgoff_t u_pages; /* number of pages in entire object */ 159 int u_flags; /* the flags (see uvm_aobj.h) */ 160 int *u_swslots; /* array of offset->swapslot mappings */ 161 /* 162 * hashtable of offset->swapslot mappings 163 * (u_swhash is an array of bucket heads) 164 */ 165 struct uao_swhash *u_swhash; 166 u_long u_swhashmask; /* mask for hashtable */ 167 LIST_ENTRY(uvm_aobj) u_list; /* global list of aobjs */ 168 }; 169 170 /* 171 * local functions 172 */ 173 174 static void uao_free(struct uvm_aobj *); 175 static int uao_get(struct uvm_object *, voff_t, struct vm_page **, 176 int *, int, vm_prot_t, int, int); 177 static int uao_put(struct uvm_object *, voff_t, voff_t, int); 178 179 static void uao_detach_locked(struct uvm_object *); 180 static void uao_reference_locked(struct uvm_object *); 181 182 #if defined(VMSWAP) 183 static struct uao_swhash_elt *uao_find_swhash_elt 184 (struct uvm_aobj *, int, bool); 185 186 static bool uao_pagein(struct uvm_aobj *, int, int); 187 static bool uao_pagein_page(struct uvm_aobj *, int); 188 static void uao_dropswap_range1(struct uvm_aobj *, voff_t, voff_t); 189 #endif /* defined(VMSWAP) */ 190 191 /* 192 * aobj_pager 193 * 194 * note that some functions (e.g. put) are handled elsewhere 195 */ 196 197 const struct uvm_pagerops aobj_pager = { 198 .pgo_reference = uao_reference, 199 .pgo_detach = uao_detach, 200 .pgo_get = uao_get, 201 .pgo_put = uao_put, 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 kmutex_t uao_list_lock; 210 211 /* 212 * functions 213 */ 214 215 /* 216 * hash table/array related functions 217 */ 218 219 #if defined(VMSWAP) 220 221 /* 222 * uao_find_swhash_elt: find (or create) a hash table entry for a page 223 * offset. 224 * 225 * => the object should be locked by the caller 226 */ 227 228 static struct uao_swhash_elt * 229 uao_find_swhash_elt(struct uvm_aobj *aobj, int pageidx, bool 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(struct uvm_object *uobj, int pageidx) 274 { 275 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 276 struct uao_swhash_elt *elt; 277 278 /* 279 * if noswap flag is set, then we never return a slot 280 */ 281 282 if (aobj->u_flags & UAO_FLAG_NOSWAP) 283 return(0); 284 285 /* 286 * if hashing, look in hash table. 287 */ 288 289 if (UAO_USES_SWHASH(aobj)) { 290 elt = uao_find_swhash_elt(aobj, pageidx, false); 291 if (elt) 292 return(UAO_SWHASH_ELT_PAGESLOT(elt, pageidx)); 293 else 294 return(0); 295 } 296 297 /* 298 * otherwise, look in the array 299 */ 300 301 return(aobj->u_swslots[pageidx]); 302 } 303 304 /* 305 * uao_set_swslot: set the swap slot for a page in an aobj. 306 * 307 * => setting a slot to zero frees the slot 308 * => object must be locked by caller 309 * => we return the old slot number, or -1 if we failed to allocate 310 * memory to record the new slot number 311 */ 312 313 int 314 uao_set_swslot(struct uvm_object *uobj, int pageidx, int slot) 315 { 316 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 317 struct uao_swhash_elt *elt; 318 int oldslot; 319 UVMHIST_FUNC("uao_set_swslot"); UVMHIST_CALLED(pdhist); 320 UVMHIST_LOG(pdhist, "aobj %p pageidx %d slot %d", 321 aobj, pageidx, slot, 0); 322 323 /* 324 * if noswap flag is set, then we can't set a non-zero slot. 325 */ 326 327 if (aobj->u_flags & UAO_FLAG_NOSWAP) { 328 if (slot == 0) 329 return(0); 330 331 printf("uao_set_swslot: uobj = %p\n", uobj); 332 panic("uao_set_swslot: NOSWAP object"); 333 } 334 335 /* 336 * are we using a hash table? if so, add it in the hash. 337 */ 338 339 if (UAO_USES_SWHASH(aobj)) { 340 341 /* 342 * Avoid allocating an entry just to free it again if 343 * the page had not swap slot in the first place, and 344 * we are freeing. 345 */ 346 347 elt = uao_find_swhash_elt(aobj, pageidx, slot != 0); 348 if (elt == NULL) { 349 return slot ? -1 : 0; 350 } 351 352 oldslot = UAO_SWHASH_ELT_PAGESLOT(elt, pageidx); 353 UAO_SWHASH_ELT_PAGESLOT(elt, pageidx) = slot; 354 355 /* 356 * now adjust the elt's reference counter and free it if we've 357 * dropped it to zero. 358 */ 359 360 if (slot) { 361 if (oldslot == 0) 362 elt->count++; 363 } else { 364 if (oldslot) 365 elt->count--; 366 367 if (elt->count == 0) { 368 LIST_REMOVE(elt, list); 369 pool_put(&uao_swhash_elt_pool, elt); 370 } 371 } 372 } else { 373 /* we are using an array */ 374 oldslot = aobj->u_swslots[pageidx]; 375 aobj->u_swslots[pageidx] = slot; 376 } 377 return (oldslot); 378 } 379 380 #endif /* defined(VMSWAP) */ 381 382 /* 383 * end of hash/array functions 384 */ 385 386 /* 387 * uao_free: free all resources held by an aobj, and then free the aobj 388 * 389 * => the aobj should be dead 390 */ 391 392 static void 393 uao_free(struct uvm_aobj *aobj) 394 { 395 int swpgonlydelta = 0; 396 397 398 #if defined(VMSWAP) 399 uao_dropswap_range1(aobj, 0, 0); 400 #endif /* defined(VMSWAP) */ 401 402 mutex_exit(&aobj->u_obj.vmobjlock); 403 404 #if defined(VMSWAP) 405 if (UAO_USES_SWHASH(aobj)) { 406 407 /* 408 * free the hash table itself. 409 */ 410 411 hashdone(aobj->u_swhash, HASH_LIST, aobj->u_swhashmask); 412 } else { 413 414 /* 415 * free the array itsself. 416 */ 417 418 kmem_free(aobj->u_swslots, aobj->u_pages * sizeof(int)); 419 } 420 #endif /* defined(VMSWAP) */ 421 422 /* 423 * finally free the aobj itself 424 */ 425 426 UVM_OBJ_DESTROY(&aobj->u_obj); 427 pool_cache_put(&uvm_aobj_cache, aobj); 428 429 /* 430 * adjust the counter of pages only in swap for all 431 * the swap slots we've freed. 432 */ 433 434 if (swpgonlydelta > 0) { 435 mutex_enter(&uvm_swap_data_lock); 436 KASSERT(uvmexp.swpgonly >= swpgonlydelta); 437 uvmexp.swpgonly -= swpgonlydelta; 438 mutex_exit(&uvm_swap_data_lock); 439 } 440 } 441 442 /* 443 * pager functions 444 */ 445 446 /* 447 * uao_create: create an aobj of the given size and return its uvm_object. 448 * 449 * => for normal use, flags are always zero 450 * => for the kernel object, the flags are: 451 * UAO_FLAG_KERNOBJ - allocate the kernel object (can only happen once) 452 * UAO_FLAG_KERNSWAP - enable swapping of kernel object (" ") 453 */ 454 455 struct uvm_object * 456 uao_create(vsize_t size, int flags) 457 { 458 static struct uvm_aobj kernel_object_store; 459 static int kobj_alloced = 0; 460 pgoff_t pages = round_page(size) >> PAGE_SHIFT; 461 struct uvm_aobj *aobj; 462 int refs; 463 464 /* 465 * malloc a new aobj unless we are asked for the kernel object 466 */ 467 468 if (flags & UAO_FLAG_KERNOBJ) { 469 KASSERT(!kobj_alloced); 470 aobj = &kernel_object_store; 471 aobj->u_pages = pages; 472 aobj->u_flags = UAO_FLAG_NOSWAP; 473 refs = UVM_OBJ_KERN; 474 kobj_alloced = UAO_FLAG_KERNOBJ; 475 } else if (flags & UAO_FLAG_KERNSWAP) { 476 KASSERT(kobj_alloced == UAO_FLAG_KERNOBJ); 477 aobj = &kernel_object_store; 478 kobj_alloced = UAO_FLAG_KERNSWAP; 479 refs = 0xdeadbeaf; /* XXX: gcc */ 480 } else { 481 aobj = pool_cache_get(&uvm_aobj_cache, PR_WAITOK); 482 aobj->u_pages = pages; 483 aobj->u_flags = 0; 484 refs = 1; 485 } 486 487 /* 488 * allocate hash/array if necessary 489 * 490 * note: in the KERNSWAP case no need to worry about locking since 491 * we are still booting we should be the only thread around. 492 */ 493 494 if (flags == 0 || (flags & UAO_FLAG_KERNSWAP) != 0) { 495 #if defined(VMSWAP) 496 const int kernswap = (flags & UAO_FLAG_KERNSWAP) != 0; 497 498 /* allocate hash table or array depending on object size */ 499 if (UAO_USES_SWHASH(aobj)) { 500 aobj->u_swhash = hashinit(UAO_SWHASH_BUCKETS(aobj), 501 HASH_LIST, kernswap ? false : true, 502 &aobj->u_swhashmask); 503 if (aobj->u_swhash == NULL) 504 panic("uao_create: hashinit swhash failed"); 505 } else { 506 aobj->u_swslots = kmem_zalloc(pages * sizeof(int), 507 kernswap ? KM_NOSLEEP : KM_SLEEP); 508 if (aobj->u_swslots == NULL) 509 panic("uao_create: malloc swslots failed"); 510 } 511 #endif /* defined(VMSWAP) */ 512 513 if (flags) { 514 aobj->u_flags &= ~UAO_FLAG_NOSWAP; /* clear noswap */ 515 return(&aobj->u_obj); 516 } 517 } 518 519 /* 520 * init aobj fields 521 */ 522 523 UVM_OBJ_INIT(&aobj->u_obj, &aobj_pager, refs); 524 525 /* 526 * now that aobj is ready, add it to the global list 527 */ 528 529 mutex_enter(&uao_list_lock); 530 LIST_INSERT_HEAD(&uao_list, aobj, u_list); 531 mutex_exit(&uao_list_lock); 532 return(&aobj->u_obj); 533 } 534 535 536 537 /* 538 * uao_init: set up aobj pager subsystem 539 * 540 * => called at boot time from uvm_pager_init() 541 */ 542 543 void 544 uao_init(void) 545 { 546 static int uao_initialized; 547 548 if (uao_initialized) 549 return; 550 uao_initialized = true; 551 LIST_INIT(&uao_list); 552 mutex_init(&uao_list_lock, MUTEX_DEFAULT, IPL_NONE); 553 pool_cache_bootstrap(&uvm_aobj_cache, sizeof(struct uvm_aobj), 0, 0, 554 0, "aobj", NULL, IPL_NONE, NULL, NULL, NULL); 555 } 556 557 /* 558 * uao_reference: add a ref to an aobj 559 * 560 * => aobj must be unlocked 561 * => just lock it and call the locked version 562 */ 563 564 void 565 uao_reference(struct uvm_object *uobj) 566 { 567 568 /* 569 * kernel_object already has plenty of references, leave it alone. 570 */ 571 572 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) 573 return; 574 575 mutex_enter(&uobj->vmobjlock); 576 uao_reference_locked(uobj); 577 mutex_exit(&uobj->vmobjlock); 578 } 579 580 /* 581 * uao_reference_locked: add a ref to an aobj that is already locked 582 * 583 * => aobj must be locked 584 * this needs to be separate from the normal routine 585 * since sometimes we need to add a reference to an aobj when 586 * it's already locked. 587 */ 588 589 static void 590 uao_reference_locked(struct uvm_object *uobj) 591 { 592 UVMHIST_FUNC("uao_reference"); UVMHIST_CALLED(maphist); 593 594 /* 595 * kernel_object already has plenty of references, leave it alone. 596 */ 597 598 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) 599 return; 600 601 uobj->uo_refs++; 602 UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)", 603 uobj, uobj->uo_refs,0,0); 604 } 605 606 /* 607 * uao_detach: drop a reference to an aobj 608 * 609 * => aobj must be unlocked 610 * => just lock it and call the locked version 611 */ 612 613 void 614 uao_detach(struct uvm_object *uobj) 615 { 616 617 /* 618 * detaching from kernel_object is a noop. 619 */ 620 621 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) 622 return; 623 624 mutex_enter(&uobj->vmobjlock); 625 uao_detach_locked(uobj); 626 } 627 628 /* 629 * uao_detach_locked: drop a reference to an aobj 630 * 631 * => aobj must be locked, and is unlocked (or freed) upon return. 632 * this needs to be separate from the normal routine 633 * since sometimes we need to detach from an aobj when 634 * it's already locked. 635 */ 636 637 static void 638 uao_detach_locked(struct uvm_object *uobj) 639 { 640 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 641 struct vm_page *pg; 642 UVMHIST_FUNC("uao_detach"); UVMHIST_CALLED(maphist); 643 644 /* 645 * detaching from kernel_object is a noop. 646 */ 647 648 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) { 649 mutex_exit(&uobj->vmobjlock); 650 return; 651 } 652 653 UVMHIST_LOG(maphist," (uobj=0x%x) ref=%d", uobj,uobj->uo_refs,0,0); 654 uobj->uo_refs--; 655 if (uobj->uo_refs) { 656 mutex_exit(&uobj->vmobjlock); 657 UVMHIST_LOG(maphist, "<- done (rc>0)", 0,0,0,0); 658 return; 659 } 660 661 /* 662 * remove the aobj from the global list. 663 */ 664 665 mutex_enter(&uao_list_lock); 666 LIST_REMOVE(aobj, u_list); 667 mutex_exit(&uao_list_lock); 668 669 /* 670 * free all the pages left in the aobj. for each page, 671 * when the page is no longer busy (and thus after any disk i/o that 672 * it's involved in is complete), release any swap resources and 673 * free the page itself. 674 */ 675 676 mutex_enter(&uvm_pageqlock); 677 while ((pg = TAILQ_FIRST(&uobj->memq)) != NULL) { 678 pmap_page_protect(pg, VM_PROT_NONE); 679 if (pg->flags & PG_BUSY) { 680 pg->flags |= PG_WANTED; 681 mutex_exit(&uvm_pageqlock); 682 UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, false, 683 "uao_det", 0); 684 mutex_enter(&uobj->vmobjlock); 685 mutex_enter(&uvm_pageqlock); 686 continue; 687 } 688 uao_dropswap(&aobj->u_obj, pg->offset >> PAGE_SHIFT); 689 uvm_pagefree(pg); 690 } 691 mutex_exit(&uvm_pageqlock); 692 693 /* 694 * finally, free the aobj itself. 695 */ 696 697 uao_free(aobj); 698 } 699 700 /* 701 * uao_put: flush pages out of a uvm object 702 * 703 * => object should be locked by caller. we may _unlock_ the object 704 * if (and only if) we need to clean a page (PGO_CLEANIT). 705 * XXXJRT Currently, however, we don't. In the case of cleaning 706 * XXXJRT a page, we simply just deactivate it. Should probably 707 * XXXJRT handle this better, in the future (although "flushing" 708 * XXXJRT anonymous memory isn't terribly important). 709 * => if PGO_CLEANIT is not set, then we will neither unlock the object 710 * or block. 711 * => if PGO_ALLPAGE is set, then all pages in the object are valid targets 712 * for flushing. 713 * => NOTE: we rely on the fact that the object's memq is a TAILQ and 714 * that new pages are inserted on the tail end of the list. thus, 715 * we can make a complete pass through the object in one go by starting 716 * at the head and working towards the tail (new pages are put in 717 * front of us). 718 * => NOTE: we are allowed to lock the page queues, so the caller 719 * must not be holding the lock on them [e.g. pagedaemon had 720 * better not call us with the queues locked] 721 * => we return 0 unless we encountered some sort of I/O error 722 * XXXJRT currently never happens, as we never directly initiate 723 * XXXJRT I/O 724 * 725 * note on page traversal: 726 * we can traverse the pages in an object either by going down the 727 * linked list in "uobj->memq", or we can go over the address range 728 * by page doing hash table lookups for each address. depending 729 * on how many pages are in the object it may be cheaper to do one 730 * or the other. we set "by_list" to true if we are using memq. 731 * if the cost of a hash lookup was equal to the cost of the list 732 * traversal we could compare the number of pages in the start->stop 733 * range to the total number of pages in the object. however, it 734 * seems that a hash table lookup is more expensive than the linked 735 * list traversal, so we multiply the number of pages in the 736 * start->stop range by a penalty which we define below. 737 */ 738 739 static int 740 uao_put(struct uvm_object *uobj, voff_t start, voff_t stop, int flags) 741 { 742 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 743 struct vm_page *pg, *nextpg, curmp, endmp; 744 bool by_list; 745 voff_t curoff; 746 UVMHIST_FUNC("uao_put"); UVMHIST_CALLED(maphist); 747 748 KASSERT(mutex_owned(&uobj->vmobjlock)); 749 750 curoff = 0; 751 if (flags & PGO_ALLPAGES) { 752 start = 0; 753 stop = aobj->u_pages << PAGE_SHIFT; 754 by_list = true; /* always go by the list */ 755 } else { 756 start = trunc_page(start); 757 if (stop == 0) { 758 stop = aobj->u_pages << PAGE_SHIFT; 759 } else { 760 stop = round_page(stop); 761 } 762 if (stop > (aobj->u_pages << PAGE_SHIFT)) { 763 printf("uao_flush: strange, got an out of range " 764 "flush (fixed)\n"); 765 stop = aobj->u_pages << PAGE_SHIFT; 766 } 767 by_list = (uobj->uo_npages <= 768 ((stop - start) >> PAGE_SHIFT) * UVM_PAGE_TREE_PENALTY); 769 } 770 UVMHIST_LOG(maphist, 771 " flush start=0x%lx, stop=0x%x, by_list=%d, flags=0x%x", 772 start, stop, by_list, flags); 773 774 /* 775 * Don't need to do any work here if we're not freeing 776 * or deactivating pages. 777 */ 778 779 if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) { 780 mutex_exit(&uobj->vmobjlock); 781 return 0; 782 } 783 784 /* 785 * Initialize the marker pages. See the comment in 786 * genfs_putpages() also. 787 */ 788 789 curmp.uobject = uobj; 790 curmp.offset = (voff_t)-1; 791 curmp.flags = PG_BUSY; 792 endmp.uobject = uobj; 793 endmp.offset = (voff_t)-1; 794 endmp.flags = PG_BUSY; 795 796 /* 797 * now do it. note: we must update nextpg in the body of loop or we 798 * will get stuck. we need to use nextpg if we'll traverse the list 799 * because we may free "pg" before doing the next loop. 800 */ 801 802 if (by_list) { 803 TAILQ_INSERT_TAIL(&uobj->memq, &endmp, listq.queue); 804 nextpg = TAILQ_FIRST(&uobj->memq); 805 uvm_lwp_hold(curlwp); 806 } else { 807 curoff = start; 808 nextpg = NULL; /* Quell compiler warning */ 809 } 810 811 /* locked: uobj */ 812 for (;;) { 813 if (by_list) { 814 pg = nextpg; 815 if (pg == &endmp) 816 break; 817 nextpg = TAILQ_NEXT(pg, listq.queue); 818 if (pg->offset < start || pg->offset >= stop) 819 continue; 820 } else { 821 if (curoff < stop) { 822 pg = uvm_pagelookup(uobj, curoff); 823 curoff += PAGE_SIZE; 824 } else 825 break; 826 if (pg == NULL) 827 continue; 828 } 829 830 /* 831 * wait and try again if the page is busy. 832 */ 833 834 if (pg->flags & PG_BUSY) { 835 if (by_list) { 836 TAILQ_INSERT_BEFORE(pg, &curmp, listq.queue); 837 } 838 pg->flags |= PG_WANTED; 839 UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, 0, 840 "uao_put", 0); 841 mutex_enter(&uobj->vmobjlock); 842 if (by_list) { 843 nextpg = TAILQ_NEXT(&curmp, listq.queue); 844 TAILQ_REMOVE(&uobj->memq, &curmp, 845 listq.queue); 846 } else 847 curoff -= PAGE_SIZE; 848 continue; 849 } 850 851 switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) { 852 853 /* 854 * XXX In these first 3 cases, we always just 855 * XXX deactivate the page. We may want to 856 * XXX handle the different cases more specifically 857 * XXX in the future. 858 */ 859 860 case PGO_CLEANIT|PGO_FREE: 861 case PGO_CLEANIT|PGO_DEACTIVATE: 862 case PGO_DEACTIVATE: 863 deactivate_it: 864 mutex_enter(&uvm_pageqlock); 865 /* skip the page if it's wired */ 866 if (pg->wire_count == 0) { 867 uvm_pagedeactivate(pg); 868 } 869 mutex_exit(&uvm_pageqlock); 870 break; 871 872 case PGO_FREE: 873 /* 874 * If there are multiple references to 875 * the object, just deactivate the page. 876 */ 877 878 if (uobj->uo_refs > 1) 879 goto deactivate_it; 880 881 /* 882 * free the swap slot and the page. 883 */ 884 885 pmap_page_protect(pg, VM_PROT_NONE); 886 887 /* 888 * freeing swapslot here is not strictly necessary. 889 * however, leaving it here doesn't save much 890 * because we need to update swap accounting anyway. 891 */ 892 893 uao_dropswap(uobj, pg->offset >> PAGE_SHIFT); 894 mutex_enter(&uvm_pageqlock); 895 uvm_pagefree(pg); 896 mutex_exit(&uvm_pageqlock); 897 break; 898 899 default: 900 panic("%s: impossible", __func__); 901 } 902 } 903 if (by_list) { 904 TAILQ_REMOVE(&uobj->memq, &endmp, listq.queue); 905 uvm_lwp_rele(curlwp); 906 } 907 mutex_exit(&uobj->vmobjlock); 908 return 0; 909 } 910 911 /* 912 * uao_get: fetch me a page 913 * 914 * we have three cases: 915 * 1: page is resident -> just return the page. 916 * 2: page is zero-fill -> allocate a new page and zero it. 917 * 3: page is swapped out -> fetch the page from swap. 918 * 919 * cases 1 and 2 can be handled with PGO_LOCKED, case 3 cannot. 920 * so, if the "center" page hits case 3 (or any page, with PGO_ALLPAGES), 921 * then we will need to return EBUSY. 922 * 923 * => prefer map unlocked (not required) 924 * => object must be locked! we will _unlock_ it before starting any I/O. 925 * => flags: PGO_ALLPAGES: get all of the pages 926 * PGO_LOCKED: fault data structures are locked 927 * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx] 928 * => NOTE: caller must check for released pages!! 929 */ 930 931 static int 932 uao_get(struct uvm_object *uobj, voff_t offset, struct vm_page **pps, 933 int *npagesp, int centeridx, vm_prot_t access_type, int advice, int flags) 934 { 935 #if defined(VMSWAP) 936 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 937 #endif /* defined(VMSWAP) */ 938 voff_t current_offset; 939 struct vm_page *ptmp = NULL; /* Quell compiler warning */ 940 int lcv, gotpages, maxpages, swslot, pageidx; 941 bool done; 942 UVMHIST_FUNC("uao_get"); UVMHIST_CALLED(pdhist); 943 944 UVMHIST_LOG(pdhist, "aobj=%p offset=%d, flags=%d", 945 (struct uvm_aobj *)uobj, offset, flags,0); 946 947 /* 948 * get number of pages 949 */ 950 951 maxpages = *npagesp; 952 953 /* 954 * step 1: handled the case where fault data structures are locked. 955 */ 956 957 if (flags & PGO_LOCKED) { 958 959 /* 960 * step 1a: get pages that are already resident. only do 961 * this if the data structures are locked (i.e. the first 962 * time through). 963 */ 964 965 done = true; /* be optimistic */ 966 gotpages = 0; /* # of pages we got so far */ 967 for (lcv = 0, current_offset = offset ; lcv < maxpages ; 968 lcv++, current_offset += PAGE_SIZE) { 969 /* do we care about this page? if not, skip it */ 970 if (pps[lcv] == PGO_DONTCARE) 971 continue; 972 ptmp = uvm_pagelookup(uobj, current_offset); 973 974 /* 975 * if page is new, attempt to allocate the page, 976 * zero-fill'd. 977 */ 978 979 if (ptmp == NULL && uao_find_swslot(&aobj->u_obj, 980 current_offset >> PAGE_SHIFT) == 0) { 981 ptmp = uvm_pagealloc(uobj, current_offset, 982 NULL, UVM_PGA_ZERO); 983 if (ptmp) { 984 /* new page */ 985 ptmp->flags &= ~(PG_FAKE); 986 ptmp->pqflags |= PQ_AOBJ; 987 goto gotpage; 988 } 989 } 990 991 /* 992 * to be useful must get a non-busy page 993 */ 994 995 if (ptmp == NULL || (ptmp->flags & PG_BUSY) != 0) { 996 if (lcv == centeridx || 997 (flags & PGO_ALLPAGES) != 0) 998 /* need to do a wait or I/O! */ 999 done = false; 1000 continue; 1001 } 1002 1003 /* 1004 * useful page: busy/lock it and plug it in our 1005 * result array 1006 */ 1007 1008 /* caller must un-busy this page */ 1009 ptmp->flags |= PG_BUSY; 1010 UVM_PAGE_OWN(ptmp, "uao_get1"); 1011 gotpage: 1012 pps[lcv] = ptmp; 1013 gotpages++; 1014 } 1015 1016 /* 1017 * step 1b: now we've either done everything needed or we 1018 * to unlock and do some waiting or I/O. 1019 */ 1020 1021 UVMHIST_LOG(pdhist, "<- done (done=%d)", done, 0,0,0); 1022 *npagesp = gotpages; 1023 if (done) 1024 return 0; 1025 else 1026 return EBUSY; 1027 } 1028 1029 /* 1030 * step 2: get non-resident or busy pages. 1031 * object is locked. data structures are unlocked. 1032 */ 1033 1034 if ((flags & PGO_SYNCIO) == 0) { 1035 goto done; 1036 } 1037 1038 for (lcv = 0, current_offset = offset ; lcv < maxpages ; 1039 lcv++, current_offset += PAGE_SIZE) { 1040 1041 /* 1042 * - skip over pages we've already gotten or don't want 1043 * - skip over pages we don't _have_ to get 1044 */ 1045 1046 if (pps[lcv] != NULL || 1047 (lcv != centeridx && (flags & PGO_ALLPAGES) == 0)) 1048 continue; 1049 1050 pageidx = current_offset >> PAGE_SHIFT; 1051 1052 /* 1053 * we have yet to locate the current page (pps[lcv]). we 1054 * first look for a page that is already at the current offset. 1055 * if we find a page, we check to see if it is busy or 1056 * released. if that is the case, then we sleep on the page 1057 * until it is no longer busy or released and repeat the lookup. 1058 * if the page we found is neither busy nor released, then we 1059 * busy it (so we own it) and plug it into pps[lcv]. this 1060 * 'break's the following while loop and indicates we are 1061 * ready to move on to the next page in the "lcv" loop above. 1062 * 1063 * if we exit the while loop with pps[lcv] still set to NULL, 1064 * then it means that we allocated a new busy/fake/clean page 1065 * ptmp in the object and we need to do I/O to fill in the data. 1066 */ 1067 1068 /* top of "pps" while loop */ 1069 while (pps[lcv] == NULL) { 1070 /* look for a resident page */ 1071 ptmp = uvm_pagelookup(uobj, current_offset); 1072 1073 /* not resident? allocate one now (if we can) */ 1074 if (ptmp == NULL) { 1075 1076 ptmp = uvm_pagealloc(uobj, current_offset, 1077 NULL, 0); 1078 1079 /* out of RAM? */ 1080 if (ptmp == NULL) { 1081 mutex_exit(&uobj->vmobjlock); 1082 UVMHIST_LOG(pdhist, 1083 "sleeping, ptmp == NULL\n",0,0,0,0); 1084 uvm_wait("uao_getpage"); 1085 mutex_enter(&uobj->vmobjlock); 1086 continue; 1087 } 1088 1089 /* 1090 * safe with PQ's unlocked: because we just 1091 * alloc'd the page 1092 */ 1093 1094 ptmp->pqflags |= PQ_AOBJ; 1095 1096 /* 1097 * got new page ready for I/O. break pps while 1098 * loop. pps[lcv] is still NULL. 1099 */ 1100 1101 break; 1102 } 1103 1104 /* page is there, see if we need to wait on it */ 1105 if ((ptmp->flags & PG_BUSY) != 0) { 1106 ptmp->flags |= PG_WANTED; 1107 UVMHIST_LOG(pdhist, 1108 "sleeping, ptmp->flags 0x%x\n", 1109 ptmp->flags,0,0,0); 1110 UVM_UNLOCK_AND_WAIT(ptmp, &uobj->vmobjlock, 1111 false, "uao_get", 0); 1112 mutex_enter(&uobj->vmobjlock); 1113 continue; 1114 } 1115 1116 /* 1117 * if we get here then the page has become resident and 1118 * unbusy between steps 1 and 2. we busy it now (so we 1119 * own it) and set pps[lcv] (so that we exit the while 1120 * loop). 1121 */ 1122 1123 /* we own it, caller must un-busy */ 1124 ptmp->flags |= PG_BUSY; 1125 UVM_PAGE_OWN(ptmp, "uao_get2"); 1126 pps[lcv] = ptmp; 1127 } 1128 1129 /* 1130 * if we own the valid page at the correct offset, pps[lcv] will 1131 * point to it. nothing more to do except go to the next page. 1132 */ 1133 1134 if (pps[lcv]) 1135 continue; /* next lcv */ 1136 1137 /* 1138 * we have a "fake/busy/clean" page that we just allocated. 1139 * do the needed "i/o", either reading from swap or zeroing. 1140 */ 1141 1142 swslot = uao_find_swslot(&aobj->u_obj, pageidx); 1143 1144 /* 1145 * just zero the page if there's nothing in swap. 1146 */ 1147 1148 if (swslot == 0) { 1149 1150 /* 1151 * page hasn't existed before, just zero it. 1152 */ 1153 1154 uvm_pagezero(ptmp); 1155 } else { 1156 #if defined(VMSWAP) 1157 int error; 1158 1159 UVMHIST_LOG(pdhist, "pagein from swslot %d", 1160 swslot, 0,0,0); 1161 1162 /* 1163 * page in the swapped-out page. 1164 * unlock object for i/o, relock when done. 1165 */ 1166 1167 mutex_exit(&uobj->vmobjlock); 1168 error = uvm_swap_get(ptmp, swslot, PGO_SYNCIO); 1169 mutex_enter(&uobj->vmobjlock); 1170 1171 /* 1172 * I/O done. check for errors. 1173 */ 1174 1175 if (error != 0) { 1176 UVMHIST_LOG(pdhist, "<- done (error=%d)", 1177 error,0,0,0); 1178 if (ptmp->flags & PG_WANTED) 1179 wakeup(ptmp); 1180 1181 /* 1182 * remove the swap slot from the aobj 1183 * and mark the aobj as having no real slot. 1184 * don't free the swap slot, thus preventing 1185 * it from being used again. 1186 */ 1187 1188 swslot = uao_set_swslot(&aobj->u_obj, pageidx, 1189 SWSLOT_BAD); 1190 if (swslot > 0) { 1191 uvm_swap_markbad(swslot, 1); 1192 } 1193 1194 mutex_enter(&uvm_pageqlock); 1195 uvm_pagefree(ptmp); 1196 mutex_exit(&uvm_pageqlock); 1197 mutex_exit(&uobj->vmobjlock); 1198 return error; 1199 } 1200 #else /* defined(VMSWAP) */ 1201 panic("%s: pagein", __func__); 1202 #endif /* defined(VMSWAP) */ 1203 } 1204 1205 if ((access_type & VM_PROT_WRITE) == 0) { 1206 ptmp->flags |= PG_CLEAN; 1207 pmap_clear_modify(ptmp); 1208 } 1209 1210 /* 1211 * we got the page! clear the fake flag (indicates valid 1212 * data now in page) and plug into our result array. note 1213 * that page is still busy. 1214 * 1215 * it is the callers job to: 1216 * => check if the page is released 1217 * => unbusy the page 1218 * => activate the page 1219 */ 1220 1221 ptmp->flags &= ~PG_FAKE; 1222 pps[lcv] = ptmp; 1223 } 1224 1225 /* 1226 * finally, unlock object and return. 1227 */ 1228 1229 done: 1230 mutex_exit(&uobj->vmobjlock); 1231 UVMHIST_LOG(pdhist, "<- done (OK)",0,0,0,0); 1232 return 0; 1233 } 1234 1235 #if defined(VMSWAP) 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(struct uvm_object *uobj, int pageidx) 1245 { 1246 int slot; 1247 1248 slot = uao_set_swslot(uobj, pageidx, 0); 1249 if (slot) { 1250 uvm_swap_free(slot, 1); 1251 } 1252 } 1253 1254 /* 1255 * page in every page in every aobj that is paged-out to a range of swslots. 1256 * 1257 * => nothing should be locked. 1258 * => returns true if pagein was aborted due to lack of memory. 1259 */ 1260 1261 bool 1262 uao_swap_off(int startslot, int endslot) 1263 { 1264 struct uvm_aobj *aobj, *nextaobj; 1265 bool rv; 1266 1267 /* 1268 * walk the list of all aobjs. 1269 */ 1270 1271 restart: 1272 mutex_enter(&uao_list_lock); 1273 for (aobj = LIST_FIRST(&uao_list); 1274 aobj != NULL; 1275 aobj = nextaobj) { 1276 1277 /* 1278 * try to get the object lock, start all over if we fail. 1279 * most of the time we'll get the aobj lock, 1280 * so this should be a rare case. 1281 */ 1282 1283 if (!mutex_tryenter(&aobj->u_obj.vmobjlock)) { 1284 mutex_exit(&uao_list_lock); 1285 /* XXX Better than yielding but inadequate. */ 1286 kpause("livelock", false, 1, NULL); 1287 goto restart; 1288 } 1289 1290 /* 1291 * add a ref to the aobj so it doesn't disappear 1292 * while we're working. 1293 */ 1294 1295 uao_reference_locked(&aobj->u_obj); 1296 1297 /* 1298 * now it's safe to unlock the uao list. 1299 */ 1300 1301 mutex_exit(&uao_list_lock); 1302 1303 /* 1304 * page in any pages in the swslot range. 1305 * if there's an error, abort and return the error. 1306 */ 1307 1308 rv = uao_pagein(aobj, startslot, endslot); 1309 if (rv) { 1310 uao_detach_locked(&aobj->u_obj); 1311 return rv; 1312 } 1313 1314 /* 1315 * we're done with this aobj. 1316 * relock the list and drop our ref on the aobj. 1317 */ 1318 1319 mutex_enter(&uao_list_lock); 1320 nextaobj = LIST_NEXT(aobj, u_list); 1321 uao_detach_locked(&aobj->u_obj); 1322 } 1323 1324 /* 1325 * done with traversal, unlock the list 1326 */ 1327 mutex_exit(&uao_list_lock); 1328 return false; 1329 } 1330 1331 1332 /* 1333 * page in any pages from aobj in the given range. 1334 * 1335 * => aobj must be locked and is returned locked. 1336 * => returns true if pagein was aborted due to lack of memory. 1337 */ 1338 static bool 1339 uao_pagein(struct uvm_aobj *aobj, int startslot, int endslot) 1340 { 1341 bool rv; 1342 1343 if (UAO_USES_SWHASH(aobj)) { 1344 struct uao_swhash_elt *elt; 1345 int buck; 1346 1347 restart: 1348 for (buck = aobj->u_swhashmask; buck >= 0; buck--) { 1349 for (elt = LIST_FIRST(&aobj->u_swhash[buck]); 1350 elt != NULL; 1351 elt = LIST_NEXT(elt, list)) { 1352 int i; 1353 1354 for (i = 0; i < UAO_SWHASH_CLUSTER_SIZE; i++) { 1355 int slot = elt->slots[i]; 1356 1357 /* 1358 * if the slot isn't in range, skip it. 1359 */ 1360 1361 if (slot < startslot || 1362 slot >= endslot) { 1363 continue; 1364 } 1365 1366 /* 1367 * process the page, 1368 * the start over on this object 1369 * since the swhash elt 1370 * may have been freed. 1371 */ 1372 1373 rv = uao_pagein_page(aobj, 1374 UAO_SWHASH_ELT_PAGEIDX_BASE(elt) + i); 1375 if (rv) { 1376 return rv; 1377 } 1378 goto restart; 1379 } 1380 } 1381 } 1382 } else { 1383 int i; 1384 1385 for (i = 0; i < aobj->u_pages; i++) { 1386 int slot = aobj->u_swslots[i]; 1387 1388 /* 1389 * if the slot isn't in range, skip it 1390 */ 1391 1392 if (slot < startslot || slot >= endslot) { 1393 continue; 1394 } 1395 1396 /* 1397 * process the page. 1398 */ 1399 1400 rv = uao_pagein_page(aobj, i); 1401 if (rv) { 1402 return rv; 1403 } 1404 } 1405 } 1406 1407 return false; 1408 } 1409 1410 /* 1411 * page in a page from an aobj. used for swap_off. 1412 * returns true if pagein was aborted due to lack of memory. 1413 * 1414 * => aobj must be locked and is returned locked. 1415 */ 1416 1417 static bool 1418 uao_pagein_page(struct uvm_aobj *aobj, int pageidx) 1419 { 1420 struct vm_page *pg; 1421 int rv, npages; 1422 1423 pg = NULL; 1424 npages = 1; 1425 /* locked: aobj */ 1426 rv = uao_get(&aobj->u_obj, pageidx << PAGE_SHIFT, 1427 &pg, &npages, 0, VM_PROT_READ|VM_PROT_WRITE, 0, PGO_SYNCIO); 1428 /* unlocked: aobj */ 1429 1430 /* 1431 * relock and finish up. 1432 */ 1433 1434 mutex_enter(&aobj->u_obj.vmobjlock); 1435 switch (rv) { 1436 case 0: 1437 break; 1438 1439 case EIO: 1440 case ERESTART: 1441 1442 /* 1443 * nothing more to do on errors. 1444 * ERESTART can only mean that the anon was freed, 1445 * so again there's nothing to do. 1446 */ 1447 1448 return false; 1449 1450 default: 1451 return true; 1452 } 1453 1454 /* 1455 * ok, we've got the page now. 1456 * mark it as dirty, clear its swslot and un-busy it. 1457 */ 1458 uao_dropswap(&aobj->u_obj, pageidx); 1459 1460 /* 1461 * make sure it's on a page queue. 1462 */ 1463 mutex_enter(&uvm_pageqlock); 1464 if (pg->wire_count == 0) 1465 uvm_pageenqueue(pg); 1466 mutex_exit(&uvm_pageqlock); 1467 1468 if (pg->flags & PG_WANTED) { 1469 wakeup(pg); 1470 } 1471 pg->flags &= ~(PG_WANTED|PG_BUSY|PG_CLEAN|PG_FAKE); 1472 UVM_PAGE_OWN(pg, NULL); 1473 1474 return false; 1475 } 1476 1477 /* 1478 * uao_dropswap_range: drop swapslots in the range. 1479 * 1480 * => aobj must be locked and is returned locked. 1481 * => start is inclusive. end is exclusive. 1482 */ 1483 1484 void 1485 uao_dropswap_range(struct uvm_object *uobj, voff_t start, voff_t end) 1486 { 1487 struct uvm_aobj *aobj = (struct uvm_aobj *)uobj; 1488 1489 KASSERT(mutex_owned(&uobj->vmobjlock)); 1490 1491 uao_dropswap_range1(aobj, start, end); 1492 } 1493 1494 static void 1495 uao_dropswap_range1(struct uvm_aobj *aobj, voff_t start, voff_t end) 1496 { 1497 int swpgonlydelta = 0; 1498 1499 if (end == 0) { 1500 end = INT64_MAX; 1501 } 1502 1503 if (UAO_USES_SWHASH(aobj)) { 1504 int i, hashbuckets = aobj->u_swhashmask + 1; 1505 voff_t taghi; 1506 voff_t taglo; 1507 1508 taglo = UAO_SWHASH_ELT_TAG(start); 1509 taghi = UAO_SWHASH_ELT_TAG(end); 1510 1511 for (i = 0; i < hashbuckets; i++) { 1512 struct uao_swhash_elt *elt, *next; 1513 1514 for (elt = LIST_FIRST(&aobj->u_swhash[i]); 1515 elt != NULL; 1516 elt = next) { 1517 int startidx, endidx; 1518 int j; 1519 1520 next = LIST_NEXT(elt, list); 1521 1522 if (elt->tag < taglo || taghi < elt->tag) { 1523 continue; 1524 } 1525 1526 if (elt->tag == taglo) { 1527 startidx = 1528 UAO_SWHASH_ELT_PAGESLOT_IDX(start); 1529 } else { 1530 startidx = 0; 1531 } 1532 1533 if (elt->tag == taghi) { 1534 endidx = 1535 UAO_SWHASH_ELT_PAGESLOT_IDX(end); 1536 } else { 1537 endidx = UAO_SWHASH_CLUSTER_SIZE; 1538 } 1539 1540 for (j = startidx; j < endidx; j++) { 1541 int slot = elt->slots[j]; 1542 1543 KASSERT(uvm_pagelookup(&aobj->u_obj, 1544 (UAO_SWHASH_ELT_PAGEIDX_BASE(elt) 1545 + j) << PAGE_SHIFT) == NULL); 1546 if (slot > 0) { 1547 uvm_swap_free(slot, 1); 1548 swpgonlydelta++; 1549 KASSERT(elt->count > 0); 1550 elt->slots[j] = 0; 1551 elt->count--; 1552 } 1553 } 1554 1555 if (elt->count == 0) { 1556 LIST_REMOVE(elt, list); 1557 pool_put(&uao_swhash_elt_pool, elt); 1558 } 1559 } 1560 } 1561 } else { 1562 int i; 1563 1564 if (aobj->u_pages < end) { 1565 end = aobj->u_pages; 1566 } 1567 for (i = start; i < end; i++) { 1568 int slot = aobj->u_swslots[i]; 1569 1570 if (slot > 0) { 1571 uvm_swap_free(slot, 1); 1572 swpgonlydelta++; 1573 } 1574 } 1575 } 1576 1577 /* 1578 * adjust the counter of pages only in swap for all 1579 * the swap slots we've freed. 1580 */ 1581 1582 if (swpgonlydelta > 0) { 1583 mutex_enter(&uvm_swap_data_lock); 1584 KASSERT(uvmexp.swpgonly >= swpgonlydelta); 1585 uvmexp.swpgonly -= swpgonlydelta; 1586 mutex_exit(&uvm_swap_data_lock); 1587 } 1588 } 1589 1590 #endif /* defined(VMSWAP) */ 1591